Advances in Smart Technologies Applications and Case Studies: Selected Papers from the First International Conference on Smart Information and Communication Technologies, SmartICT 2019, September 26-28, 2019, Saidia, Morocco [1st ed.] 9783030531867, 9783030531874

Table of contents :
Front Matter ....Pages i-xxii
Front Matter ....Pages 1-1
Heuristic for Network Planning Based on 5G Services (M. Umar Khan, Mostafa Azizi, Ana García Armada, J. J. Escudero Garzás)....Pages 3-14
Performance Improvement of OFDM-ROF System with Combined Adaptive Coded Modulation and Power Control (Mohammed Amine Azza, Moussa El Yahyaoui, Ali El Moussati)....Pages 15-22
Optical Architecture for 60 GHz 4 × 4 MIMO Signal Distribution over Optical Fiber (Moussa El Yahyaoui, Hachim Azzahhafi, Ali El Moussati)....Pages 23-29
Evaluation of Railway Communications System Based on 5G-RoF Technology and Millimeter Wave Band (Hachim Azzahhafi, Moussa El Yahyaoui, Ali El Moussati)....Pages 30-37
Digital Video Broadcasting - Satellite - Second Generation Measurement and Test for Database Simulation (Youssef Bikrat, Khalid Salmi, Ahmad Benlghazi, Abdelhamid Benali, Driss Moussaid)....Pages 38-46
Design of a Microstrip Patch UWB Antenna with Notch Band Characteristic (L. Aguni, S. Chabaa, S. Ibnyaich, A. Zeroual)....Pages 47-54
60 GHz RoF System Based on IR-MBOOK Transmitter and Non-coherent Receiver (Tarik Zarrouk, Ali El Moussati, Papa Alioune Fall, Ghaïs El Zein)....Pages 55-62
Impact of Human Morphology on Measurement Errors of a RF Exposimeter (Abdechafik Derkaoui, Rodrigues Kwate Kwate, Bachir Elmagroud, Dominique Picard, Abdelhak Ziyyat)....Pages 63-70
RF-Exposimeter Errors Measurement: Frequency and Distance Impact (Rodrigues Kwate, Bachir Elmagroud, Abdechafik Derkaoui, Chakib Taybi, Dominique Picard, Abdelhak Ziyyat)....Pages 71-79
Front Matter ....Pages 81-81
Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts (Peeter Lorents, Ahto Kuuseok, Erika Lorents)....Pages 83-93
Local Binary Pattern and Its Variants: Application to Face Analysis (Jade Lizé, Vincent Débordès, Hua Lu, Kidiyo Kpalma, Joseph Ronsin)....Pages 94-102
Reducing LBP Features for Facial Identification and Expression Recognition (Joseph Ronsin, Kidiyo Kpalma, Hua Lu)....Pages 103-111
Video Retrieval Using Query Images and CNN Features (Imane Hachchane, Abdelmajid Badri, Aïcha Sahel, Yassine Ruichek)....Pages 112-120
CUDA Accelerating of Fractal Texture Features for a Neuro-morphological Image Segmentation Approach (Khalid Salhi, El Miloud Jaara, Mohammed Talibi Alaoui)....Pages 121-128
Efficient Mapping Method for Elliptic Curve Cryptosystems Based on PWLCM (Salma Bendaoud, Fatima Amounas, El Hassan El Kinani)....Pages 129-136
3D Shape Recognition Based on Uncoded Structured Light Using ANN Classifier (Kaoutar Baibai, Mohamed Emharraf, Wafae Mrabti, Khalid Hachami, Benaissa Bellach)....Pages 137-143
Machine Vision-Based Cocoa Beans Fermentation Degree Assessment (Yro Aubain, N’Zi Édié Camille, Kpalma Kidiyo)....Pages 144-148
Plants Classification Using Neural Shifted Legendre-Fourier Moments (Abderrahmane Machhour, Amal Zouhri, Mostafa El Mallahi, Zakia Lakhliai, Ahmed Tahiri, Driss Chenouni)....Pages 149-153
Front Matter ....Pages 155-155
Criteria for Security Classification of Smart Home Energy Management Systems (Manish Shrestha, Christian Johansen, Josef Noll)....Pages 157-165
Secure Linear Regression Algorithms: A Comparison (Fida Dankar, Nisha Madathil)....Pages 166-174
Multi-agents Intrusion Detection System Using Ontology for Manets (Sara Chadli, Hajar Chadli, Mohammed Saber, Mohammed Ghaouth Belkasmi, Ilhame El Farissi, Mohamed Emharraf)....Pages 175-182
Analysis of KDD Dataset Categories to Design a Performing Intrusion Detection System (Ilhame El Farissi, Mohammed Saber, Sara Chadli, Zineb Bougroun, Mohamed Emharraf, Mohammed Ghaouth Belkasmi et al.)....Pages 183-191
A Comparative Performance Analysis of the Intrusion Detection Systems (Mohammed Saber, Zineb Bougroun, Ilhame El Farissi, Sara Chadli, Mohamed Emharraf, Saida Belouali et al.)....Pages 192-200
New Improvement of Malware-Attack Scenarios Modeling (Noureddine Rahmoun, Yassine Ayachi, Jamal Berrich, Mohammed Saber, Toumi Bouchentouf)....Pages 201-211
IoT Security Management: Model and Design Issues (Ghizlane Benzekri, Omar Moussaoui, Ali El Moussati)....Pages 212-219
Front Matter ....Pages 221-221
Comparison Between Constant and Variable Switching Frequency Strategies Based Direct Torque Control of Asynchronous Motor (Soukaina El Daoudi, Loubna Lazrak, Chirine Benzazah, Mustapha Ait Lafkih)....Pages 223-231
Simulation and Analysis of Enhanced Perturb and Observe MPPT Algorithm Based on an Adaline Neural Network for Standalone PV System (Ihssane Chtouki, Houssam Eddine Chakir, Patrice Wira, Malika Zazi, Bruno Collicchio)....Pages 232-243
Performance Assessment of Solar Dish-Stirling System for Electricity Generation in Eastern Morocco (Hanane Ait Lahoussine Ouali, Benyounes Raillani, Samir Amraqui, Mohammed Amine Moussaoui, Ahmed Mezrhab)....Pages 244-252
Real Time Implementation of SPWM Signal Generation Technique for a New Five Level Inverter Using Microcontroller (Hajar Chadli, Zakariae Jebroni, Sara Chadli, Mohammed Saber, Khalid Salmi, Abdechafik Derkaoui et al.)....Pages 253-262
Design of a PWM Sliding Mode Voltage Controller of a DC-DC Boost Converter in CCM at Variable Conditions (Weam El Merrassi, Abdelouahed Abounada, Mohamed Ramzi)....Pages 263-270
Design and Performance Analysis of Super-Twisting Algorithm Control for Direct-Drive PMSG Wind Turbine Feeding a Water Pumping System (Benzaouia Soufyane, Zouggar Smail, Rabhi Abdelhamid, Mohammed Larbi Elhafyani)....Pages 271-281
Electric System Cascade Analysis for Optimal Sizing of an Autonomous Photovoltaic Water Pumping System (Mohammed Chennaif, Mohamed Larbi Elhafyani, Hassan Zahboune, Smail Zouggar)....Pages 282-290
Techno-Economic Sizing of a Stand-Alone Hybrid Energy and Storage for Water Pumping System (Mohammed Chennaif, Hassan Zahboune, Mohammed Larbi Elhafyani, S. Zouggar)....Pages 291-299
Rotating Machines Energy Loss Due to Unbalance (Ali Elkihel, Bouchra Abouelanouar, Hassan Gziri)....Pages 300-308
Comparative Study Between PI Speed Control and Sliding Mode Control of BLDC Motor (Ahmed Loukmane El Idrissi, Jamal Bouchnaif, Mohammed Mokhtari, Anas Bensliman)....Pages 309-317
PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine Using a Fuzzy Logic Controller (Mhamed Fannakh, Mohamed Larbi Elhafyani, Hassan Zahboune, Smail Zouggar)....Pages 318-329
Contribution to Power Maximization of an Asynchronous Wind Electric Water Pumping System Using Single Input Fuzzy Logic Controller and Modified Enhanced Perturb and Observe (Mohammed Mokhtari, Smail Zouggar, Nacer K. M’sirdi, Mohamed Larbi Elhafyani)....Pages 330-342
Hybrid System Energy Management in a Low Power Isolated Site (Mohammed Larbi El Hafyani, Abdelmalek El Elmehdi, Smail Zouggar, Toufik Ouchbel)....Pages 343-354
Front Matter ....Pages 355-355
Citation Classification Using Natural Language Processing and Machine Learning Models (Syyab Rahi, Iqra Safder, Sehrish Iqbal, Saeed-Ul Hassan, Iain Reid, Raheel Nawaz)....Pages 357-365
Path Planning Algorithm for Initially Unknown Indoor Environment Navigation (Mohamed Emharraf, Mohammed Saber, Mohammed Ghaouth Belkasmi, Ilhame El Farissi, Sara Chadli, Mohammed Rahmoun)....Pages 366-376
Ontology-Based Reasoning for Collective Intelligence of Multi-agents System (Yman Chemlal, Hicham Medroumi)....Pages 377-385
Prediction of Direct Normal Irradiance Using Artificial Neural Networks Under Oujda Climate (Latifa El Boujdaini, Ahmed Mezrhab, Mohammed Amine Moussaoui)....Pages 386-394
Corpus Construction and Annotation Challenge for Language Identification and Sentiment Analysis (Ibtissam Touahri, Azzeddine Mazroui)....Pages 395-403
Feature Selection for Community Evolution Prediction in Location-Based Social Network: Gowalla and Brightkite (Loubna Boujlaleb, Ali Idarrou, Driss Mammass)....Pages 404-412
Front Matter ....Pages 413-413
CropSAT – A Decision Support System for Practical Use of Satellite Images in Precision Agriculture (Omran Alshihabi, Kristin Piikki, Mats Söderström)....Pages 415-421
Rice Yield Prediction Using On-Farm Data Sets and Machine Learning (Oscar Barrero, Sofiane Ouazaa, Camilo Ignacio Jaramillo-Barrios, Mauricio Quevedo, Nesrine Chaali, Sair Jaramillo et al.)....Pages 422-430
Inter-comparison Between Different Techniques for Evapotranspiration Partitioning: Eddy Covariance-, Sap Flow-, Lysimeter- and FAO-Based Methods (Zoubair Rafi, Olivier Merlin, Valérie Le Dantec, Saïd Khabba, Salah Er-Raki, Patrick Mordelet)....Pages 431-439
Effects of Climate Change at the 2040’s Horizon on the Hydrology of the Pluvio-Nival Rheraya Watershed Near Marrakesh, Morocco (Youssef Hajhouji, Younes Fakir, Vincent Simonneaux, Simon Gascoin, El Houssaine Bouras, Abdelghani Chehbouni)....Pages 440-450
Estimation of the Evapotranspiration over Heterogeneous Region Using Shuttleworth-Wallace Model (Jamal Elfarkh, Salah Er-Raki, Jamal Ezzahar, Lionel Jarlan, Said Khabba, Abdelghani Chehbouni)....Pages 451-459
Intelligent Agriculture Platform Based on Low Energy and Cost Wireless Sensors for Efficient Water Irrigation (Mohamed Emharraf, Hamza Taous, Wiame Benzekri, Ali El Moussati, Kamal Aberkani)....Pages 460-469
Front Matter ....Pages 471-471
The Effects of Neurofeedback on Event Related Potential (ERP) in Zikr Meditation (Nur Arina Ayuni Helman, Muhamad Kamal Mohammed Amin)....Pages 473-480
Exploring Eye-Gaze Behaviors on Emotional States for Human-Computer Interaction (HCI) Using Eye Tracking Technique (Sumita Thiyagarajan, Muhamad Kamal M. Amin)....Pages 481-488
Novel Alignment Approach of DNA Sequences (Wajih Rhalem, Jamel El Mhamdi, Mourad Raji, Ahmed Hammouch, Nassim Kharmoum, Sanae Raoui et al.)....Pages 489-497
Understanding the Study Experiences of Students in Low Agency Profile: Towards a Smart Education Approach (Ville Heilala, Päivikki Jääskelä, Tommi Kärkkäinen, Mirka Saarela)....Pages 498-508
Let Me Hack It: Teachers’ Perceptions About ‘Making’ in Education (Ville Heilala, Mirka Saarela, Sanna Reponen, Tommi Kärkkäinen)....Pages 509-518
Comparing the Effect Size of School Level Support on Teachers’ Technology Integration (Eloho Ifinedo)....Pages 519-526
Tracking Entrepreneurial Mind Among University Students Through Functional Near-Infrared Spectroscopy (fNIRS) Technology (Nur Izzati J. Sham, Muhamad Kamal Mohammed Amin)....Pages 527-534
Ethics of AI or Ethical AI, Topical Point of View (Saida Belouali, Anas Belouali, Mohammed Saber, Khalid Jaafar, Mohammed Ghaouth Belkasmi)....Pages 535-540
Open Access Publishing and Ethical Issues in the Moroccan Context (Nadia Benaissa, Saïda Belouali)....Pages 541-547
Intelligent Model for Evaluation Based on Expert System and Fuzzy Logic (Khalid Salmi, Hanane Sefraoui, Hamid Magrez, Abdechafik Derkaoui, Abdelaziz Elmoussaouy, Abdelhak Ziyyat)....Pages 548-553
Front Matter ....Pages 555-555
Artificial Neural Networks for Text-to-SQL Task: State of the Art (Youssef Mellah, Hassane El Ettifouri, Toumi Bouchentouf, Mohammed Ghaouth Belkasmi)....Pages 557-565
SVM on HPC Clouds: Choosing the Appropriate Classification Algorithm and Kernel Type According to the Data Set Characteristics (Mouncef Filali Bouami, Mohamed Benchat)....Pages 566-574
Global IT Project Management: An Agile Planning Assistance (Mohammed Ghaouth Belkasmi, Zineb Bougroun, Ilhame El Farissi, Mohamed Emharraf, Saida Belouali, Sara Chadli et al.)....Pages 575-582
A Survey for Validation Concepts to Measure Quality as Well Their Application on the Maintainability of ISO (Zineb Bougroun, Mohammed Saber, Ilhame El Farissi, Ghaouth Mohammed Belkasmi, Toumi Bouchentouf)....Pages 583-591
NL2Code: A Corpus and Semantic Parser for Natural Language to Code (El Hassane Ettifouri, Walid Dahhane, Achraf Berrajaa, Toumi Bouchentouf, Mohammed Rahmoun)....Pages 592-599
Model-Based Testing from Model Driven Architecture: A Novel Approach for Automatic Test Cases Generation (Imane Essebaa, Salima Chantit, Mohammed Ramdani)....Pages 600-609
Front Matter ....Pages 611-611
Optical Degradation of CSP Reflectors Under Moroccan-Eastern Climate: An Experimental Investigation (Mouatassim Charai, Latifa Elboujdaini, Ahmed Mezrhab, Abdelhamid Mezrhab, Mustapha Karkri)....Pages 613-620
Modeling Spring Impact on Durability of Welded Structure for Electric Vehicles Utilization (Imane Amarir, Hamid Mounir, Abdellatif El Marjani, Kaoutar Daoudi)....Pages 621-628
Accurate Evaluation of Solar Irradiation of a Satellite Dataset Under Ground Measurements (Latifa El Boujdaini, Ahmed Mezrhab, Abdelhamid Mezrhab, Mohammed Amine Moussaoui, Mouatassim Charai)....Pages 629-637
Validation and Numerical Study of an Earth-to-Air Heat Exchanger for Cooling and Preheating (Haitham Sghiouri, Mouatassim Charai, Ahmed Mezrhab, Mustapha Karkri)....Pages 638-646
NaI(Tl) Detector Response at Different Energies and a Validation with Monte Carlo Simulation (Abdelkarim Bazza, Abdelkader El Hamli, Mohammed Hamal, Abdellah Moussa, Mostapha Zerfaoui, Lahsen Hamam et al.)....Pages 647-655
Running GATE Software on Moroccan Cluster Computing to Simulate Particle Interactions Within Linear Accelerator System (Deae-eddine Krim, Abdeslem Rrhioua, Dikra Bakari, Mustapha Zerfaoui, Yassine Oulhouq)....Pages 656-665
Analysis and Optimization of SM and TES Hours of Central Receiver Concentrated Solar Thermal with Two-Tank Molten Salt Thermal Storage (Hanane Ait Lahoussine Ouali, Benyounes Raillani, Samir Amraqui, Mohammed Amine Moussaoui, Abdelhamid Mezrhab, Ahmed Mezrhab)....Pages 666-673
Front Matter ....Pages 675-675
Wireless Network Stability Enhancement Based on Spatial Dependency (Halim Berradi, Ahmed Habbani, Chaimae Benjbara, Nada Mouchfiq, Hicham Amraoui)....Pages 677-686
Modeling and Simulation of LoRaWAN for Smart Metering Network (Zakariae Jebroni, Hajar Chadli, Khalid Salmi, Mohammed Saber, Belkassem Tidhaf)....Pages 687-695
Early Forest Fire Detection with Low Power Wireless Sensors Networks (Wiame Benzekri, Ali El Moussati, Omar Moussaoui, Mohammed Berrajaa)....Pages 696-704
Study and Optimization of the System Energy in WSN with Global and Sequential Experiment Designs (Mohammed Jabri, Omar Moussaoui, Mimoun Moussaoui, Ali El Moussati)....Pages 705-715
LoRa Based Smart Electrical Energy Meter (Zakariae Jebroni, Hajar Chadli, Khalid Salmi, Mohammed Saber, Belkassem Tidhaf)....Pages 716-723
Towards Reliable and Timely Communications in Wireless Body Area Networks: A Comparative Study (Azdad Nabila, Elboukhari Mohamed)....Pages 724-735
Back Matter ....Pages 737-740

Citation preview

Lecture Notes in Electrical Engineering 684

Ali El Moussati · Kidiyo Kpalma · Mohammed Ghaouth Belkasmi · Mohammed Saber · Sylvain Guégan   Editors

Advances in Smart Technologies Applications and Case Studies Selected Papers from the First International Conference on Smart Information and Communication Technologies, SmartICT 2019, September 26–28, 2019, Saidia, Morocco

Lecture Notes in Electrical Engineering Volume 684

Series Editors Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli Federico II, Naples, Italy Marco Arteaga, Departament de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán, Mexico Bijaya Ketan Panigrahi, Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, Munich, Germany Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China Shanben Chen, Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore, Singapore, Singapore Rüdiger Dillmann, Humanoids and Intelligent Systems Laboratory, Karlsruhe Institute for Technology, Karlsruhe, Germany Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China Gianluigi Ferrari, Università di Parma, Parma, Italy Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid, Madrid, Spain Sandra Hirche, Department of Electrical Engineering and Information Science, Technische Universität München, Munich, Germany Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA, USA Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Alaa Khamis, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt Torsten Kroeger, Stanford University, Stanford, CA, USA Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra, Barcelona, Spain Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA Sebastian Möller, Quality and Usability Laboratory, TU Berlin, Berlin, Germany Subhas Mukhopadhyay, School of Engineering & Advanced Technology, Massey University, Palmerston North, Manawatu-Wanganui, New Zealand Cun-Zheng Ning, Electrical Engineering, Arizona State University, Tempe, AZ, USA Toyoaki Nishida, Graduate School of Informatics, Kyoto University, Kyoto, Japan Federica Pascucci, Dipartimento di Ingegneria, Università degli Studi “Roma Tre”, Rome, Italy Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Gan Woon Seng, School of Electrical & Electronic Engineering, Nanyang Technological University, Singapore, Singapore Joachim Speidel, Institute of Telecommunications, Universität Stuttgart, Stuttgart, Germany Germano Veiga, Campus da FEUP, INESC Porto, Porto, Portugal Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Beijing, China Junjie James Zhang, Charlotte, NC, USA

The book series Lecture Notes in Electrical Engineering (LNEE) publishes the latest developments in Electrical Engineering - quickly, informally and in high quality. While original research reported in proceedings and monographs has traditionally formed the core of LNEE, we also encourage authors to submit books devoted to supporting student education and professional training in the various fields and applications areas of electrical engineering. The series cover classical and emerging topics concerning:

• • • • • • • • • • • •

Communication Engineering, Information Theory and Networks Electronics Engineering and Microelectronics Signal, Image and Speech Processing Wireless and Mobile Communication Circuits and Systems Energy Systems, Power Electronics and Electrical Machines Electro-optical Engineering Instrumentation Engineering Avionics Engineering Control Systems Internet-of-Things and Cybersecurity Biomedical Devices, MEMS and NEMS

For general information about this book series, comments or suggestions, please contact leontina. [email protected]. To submit a proposal or request further information, please contact the Publishing Editor in your country: China Jasmine Dou, Associate Editor ([email protected]) India, Japan, Rest of Asia Swati Meherishi, Executive Editor ([email protected]) Southeast Asia, Australia, New Zealand Ramesh Nath Premnath, Editor ([email protected]) USA, Canada: Michael Luby, Senior Editor ([email protected]) All other Countries: Leontina Di Cecco, Senior Editor ([email protected]) ** Indexing: The books of this series are submitted to ISI Proceedings, EI-Compendex, SCOPUS, MetaPress, Web of Science and Springerlink **

More information about this series at http://www.springer.com/series/7818

Ali El Moussati Kidiyo Kpalma Mohammed Ghaouth Belkasmi Mohammed Saber Sylvain Guégan •

•

•

•

Editors

Advances in Smart Technologies Applications and Case Studies Selected Papers from the First International Conference on Smart Information and Communication Technologies, SmartICT 2019, September 26–28, 2019, Saidia, Morocco

123

Editors Ali El Moussati Intelligent and Connected Systems Team, National School of Applied Sciences Université Mohammed Premier Oujda, Morocco Mohammed Ghaouth Belkasmi SmartICT Lab National School of Applied Sciences Université Mohammed Premier Oujda, Morocco

Kidiyo Kpalma Univ Rennes, INSA Rennes CNRS, IETR - UMR 6164 Rennes, France Mohammed Saber SmartICT Lab National School of Applied Sciences Université Mohammed Premier Oujda, Morocco

Sylvain Guégan Univ Rennes, INSA Rennes LGCGM EA 3913 Rennes, France

ISSN 1876-1100 ISSN 1876-1119 (electronic) Lecture Notes in Electrical Engineering ISBN 978-3-030-53186-7 ISBN 978-3-030-53187-4 (eBook) https://doi.org/10.1007/978-3-030-53187-4 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

This book gathers together selected papers presented at The first international Conference on Smart Information & Communication Technologies (SmartICT 2019). This conference took place at Be Live Collection Saïdia Hotel in the Moroccan “Blue pearl” city of Saïdia, during 26–28 September 2019. It was organized by National School of Applied Sciences Oujda of the Mohammed First University (Morocco) and Institut National des Sciences Appliquées (INSA) de Rennes (France) within the partnership of both institutions. SmartICT 2019 has provided an exceptional forum for researchers, PhD students, industrials and decision-makers to exchange ideas, techniques and tools, raise awareness, and share experience related to all practical and theoretical aspects of smart technologies in: from healthcare and energy management, to digital education, agriculture and cybersecurity. This book results from more than 300 contributions of researchers from more than 38 countries worldwide. After a thorough peer-review process, the programme committee has accepted 145 papers, which have undergone a selection stage to retain 79 papers for this LNEE volume. This achieves overall acceptance rate of 26%. The book is organized into ten parts corresponding to respective topics presented at the conference: 1. 2. 3. 4. 5. 6. 7. 8.

Communications Systems & Applications Computer Vision & Data Processing Cybersecurity & Data Protection Energy and Multi-Source Systems Management Machine Learning, Intelligent Systems & Applications Precision Agriculture Smart Health, Digital Education & Humanities Software Engineering & Data Science

v

vi

Preface

9. Solar Thermal and Mechanics 10. Wireless Sensor Network, Internet of Things & Applications We would like to thank the organization committee, the members of the technical program committee and reviewers. They did a relevant job in reviewing the articles and provided feedback with valuable suggestions to the authors. We would also like to express our thanks to the authors for sharing the results of their research in this conference. We are very grateful to the keynote speakers who have accepted our invitation to come and share their work during the conference: Professor Abdellah Touhafi from Vrije Universiteit Brussel (Belgium); Professor Sanela Klarić from International Burch University Sarajevo (Bosnia and Herzegovina); Professor Abderrahim Benslimane from Avignon University, (France); Professor Vincent Frémont from Ecole Centrale de Nantes (France); and Professor M’Hamed Drissi from Institut National des Sciences Appliquées de Rennes (France). We would also like to thank Springer’s staff for their support and for allowing us to participate in LNEE series. We hope that you will find it useful, exciting and inspiring, and we look forward to meeting you in the next SmartICT conference. Ali El Moussati Kidiyo Kpalma Mohammed Ghaouth Belkasmi Mohammed Saber Sylvain Guégan

Organization

Committees General Chairs Ali El Moussati Mohammed Ghaouth Belkasmi

Mohammed First University, Morocco Mohammed First University, Morocco

General Co-chairs Mohammed Saber Kidiyo Kpalma

Mohammed First University, Morocco INSA Rennes, France

Technical Program Chairs Sylvain Guégan Joseph Ronsin

INSA Rennes, France INSA Rennes, France

Organizing Committee Abdelmalek El Mehdi Ilhame El Farissi Mohamed Emharref Saida Belouali Sara Chadli Wiame Benzekri Zineb Bougroun

Mohammed Mohammed Mohammed Mohammed Mohammed Mohammed Mohammed

First First First First First First First

University, University, University, University, University, University, University,

Morocco Morocco Morocco Morocco Morocco Morocco Morocco

vii

viii

Organization

Keynotes Abdellah Touhafi Abderrahim Benslimane M’Hamed Drissi Sanela Klaric Vincent Frémont

Universiteit Brussel, Belgium Avignon University, France Director of INSA Rennes, France International Burch University Sarajevo, Bosnia and Herzegovina LS2N, Ecole Centrale de Nantes, France

Technical Program Committee Abdechafik Derkaoui Abdelaziz Lawani Abdelghani El Ougli Abdelhak Aziz Abdelhak Lakhouaja Abdelhak Ziyyat Abdelhamid Benali Abdelhamid Daamouche Abdelkader El Kebir Abdellatif Kobbane Abdelmalek El Mehdi Abdelouahad Tahani Abderrahim Benslimane Abderrahim Medbouhi Abderrahim Saaidi Abdeslam En-Nouaary Abdulwahab Almazroi Abir Ben Ali Ahmed Benlghazi Ahmed El Oualkadi Ahmed Faize Ahmed Hammad Ahmed Mezrhab Ahmed Zellou Akram Sarhan Alessandro Matese Ali El Moussati Ali Idarrou Ali Boularbah Anas Belouali

Mohammed First University, Morocco Eastern Kentucky University, USA Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco University M’hamed Bougara of Boumerdes, Algeria University of Mascara, Algeria Mohammed V University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Avignon University, France Mohammed First University, Morocco Sidi Mohamed Ben Abdellah University, Morocco INPT, Morocco University of Jeddah, Saudi Arabia CRISTAL, Tunisia Mohammed First University, Morocco Abdelmalek Essaadi University, Morocco Mohammed First University, Morocco University of Franche-comté, France Mohammed First University, Morocco Mohammed V University, Morocco University of Jeddah, Saudi Arabia CNR, Italy Mohammed First University, Morocco Ibn Zohr University, Morocco Cadi Ayyad University, Morocco Georgetown University, USA

Organization

Anas Benslimane Andrew Diniz da Costa Andrina Granić Agustin Francisco Gutierrez Tornes Ayman Alzayed Aziz Salah Bachir Elmagroud Baddou Mohammed Amine Badre Bossoufi Bekkay Hajji Belkassem Tidhaf Bingshu Wang Borka Jerman-Blažič Brahim Aksasse Brahim Raouyane Chakib Taybi Chantal Charnet Chrissanthi Angeli Costin Badica Daniel Borrajo Millan Diana Hoyos-Valdés Dominique Courault Dorsaf Sebai Dris Bahia Elarbi Badidi El Hassane Ettifouri El Hassane Chadli El Miloud Jaara El Mostapha Aboulhamid Emharraf mohamed Eric Rojas Ernst Leiss Farid Bagui Farid Boushaba Farshad Badie Filipe Duarte dos Santos Cardoso Francisco José Garcìa Peñalvo

ix

Mohammed First University, Morocco Pontifìcia Universidade Catòlica do Rio de Janeiro, Brazil University of Split, Croatia Universidad Autónoma de Guerrero, Mexico Kuwait University, Kuwait University of Quebec in Montreal, Canada Mohamed First University, Morocco Mohamed First University, Morocco Sidi Mohamed Ben Abdellah University, Morocco Mohamed First University, Morocco Mohamed First University, Morocco Peking University, China University of Ljubljana, Slovenia Faculty of Sciences and Technics Errachidia, Morocco Hassan II University, Morocco Mohammed First University, Morocco Paul-Valéry University, Montpellier, France University of West Attica, Greece University of Craiova, Romania Universidad Carlos III de Madrid, Spain Caldas University, Colombia INRA, France National Institute of Applied Science and Technology, Tunisia Mohammed First University, Morocco Faculty of Information Technology, UAE Novelis, France Mohammed First University, Morocco Mohammed First University, Morocco University of Montreal, Canada Mohammed First University, Morocco Pontifical Catholic University of Chile, Chile University of Houston, USA Université Rouen Haute Normandie, CESI, France Mohammed First University, Morocco Aalborg University, Denmark Instituto Politécnico de Setúbal, Portugal University of Salamanca, Spain

x

Ghaïs El Zein Gheorghe Zaharia Hafedh Abid Hajar Chadli Hakim El Fadili Hammadi Nait Charif Hanane Benouda Hassan Rhinane Hassan Zahboune Ikram El Azami Ilham Slimani Ilhame El Farissi Jamal Berrich Ján Skalka Jose Fonseca Jose M. Molina Joseph Ronsin Juan M. Corchado Juarez Bento da Silva Kamal Aberkani Kamal Hirech Kamel Haddadi Kaoutar Lamrini Uahabi Karima Aissaoui Kidiyo Kpalma Khaled Bensid Khalid Salmi Lahoucine El Maimouni Larbi Touaf Leila Ismail Loubna Elmansouri Louis Longchamps Lubomir Benko M’hamed Bakrim Maria Tortorella Matthew Montebello Merahi Bouziani Michal Munk Miloud Chikr El Mezouar

Organization

INSA Rennes, France IETR-INSA Rennes, France University of Sfax, Tunisia Mohammed First University, Morocco Sidi Mohamed Ben Abdellah University, Morocco Bournemouth University, UK Mohammed First University, Morocco Hassan II University, Morocco Mohammed First University, Morocco Ibn Tofail University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Constantine the Philosopher University in Nitra, Slovakia Universidade NOVA de Lisboa, Portugal Universidad Carlos III de Madrid, Spain INSA Rennes, France Air Institute - University of Salamanca, Spain Federal University of Santa Catarina, Brazil Mohammed First University, Morocco Moulay Ismail University, Morocco Lille University, France Mohammed First University, Morocco Sidi Mohamed Ben Abdellah University, Morocco INSA Rennes, France University of Ouargla, Algeria Mohammed First University, Morocco Ibn Zohr University, Morocco Mohammed First University, Morocco College of IT, UAE University, UAE IAV Hassan II, Morocco AAFC, Canada Constantine the Philosopher University in Nitra, Slovakia Cadi Ayyad University, Morocco University of Sannio, Italy University of Malta, Malta University Djillali Liabès of Sidi Bel Abbés, Algeria Constantine the Philosopher University in Nitra, Slovakia Djillali Liabes University, Algeria

Organization

Mimoun El Hammouti Mirka Saarela Moez Krichen Mohamed Atounti Mohamed Bellouki Mohamed Elboukhari Mohamed Moughit Mohamed Nabil Saidi Mohammed Abbadi Mohammed Amine Kasmi Mohammed Amine Moussaoui Mohammed Benabdellah Mohammed Bourhaleb Mohammed Blej Mohammed El Koutbi Mohammed Fattah Mohammed Gabli Mohammed Ghammouri Mohammed Ghaouth Belkasmi Mohammed Guamguami Mohammed Hassine Mohammed Kamal Benhaoua Mohammed Saber Mohammed Rahmoune Mohd. Saifuzzaman Mouncef Filali Bouami Mourad Said Mostafa Azizi Mostafa Belkasmi Mostafa Bellafkih Mostafa El Ouariachi Mostafa El Mallahi Mostapha Badri Moussa El ayachi Moussa El Yahyaoui Mustapha Machkour Naceur Aounallah Najiba El Amrani El Idrissi Nasreddine Taleb

xi

Mohammed First University, Morocco University of Jyväskylä, Finland University of Sfax, Tunisia Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Sultan Moulay Slimane University, Morocco INSEA, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco CPR Oujda, Morocco Mohammed V University, Morocco Moulay Ismail University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Tisalabs, Morocco University Mustapha Stambouli of Mascara, Algeria Mohammed First University, Morocco Mohammed First University, Morocco Daffodil International University, Dhaka, Bangladesh Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed V University, Morocco INPT, Morocco Mohammed First University, Morocco Sidi Mohamed Ben Abdellah University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Ibn Zohr University, Morocco Kasdi Merbah University of Ouargla, Algeria Sidi Mohamed Ben Abdellah University, Morocco Djillali Liabes University, Algeria

xii

Nicolás Duque Buitrago Nicolas Ragot Noureddine Chikouche Omar Moussaoui Ouerdi Noura Ounsa Roudies Papa Alioune Fall Paul Temple Paulo Mendes Pedro Brandão Peter Bentley Peter Papcun Rachid Hadria Rachid El Alami Rachid El Bouayadi Rachida Dssouli Raúl Correia Reda Korikache Réda Yahiaoui Redouane Esbai Ricardo Costa Ronan Le Breton Roumen Kountchev Roumiana Kountcheva Roza Dumbraveanu Said Hamdioui Saida Belouali Sajid Anwar Salah Er-raki Salah Eddine Samri Sanae Mazouz Sanela Klarić Sankhanil Dey Sara Chadli Souad El Houssaini Soumia Boutkhil Sunil Karamchandani Sylvain Guégan Taoufik Ouchbel Taoufik Serraj Toumi Bouchentouf Vincent Frémont

Organization

Caldas University, Colombia ESIGELEC, France University of M’sila, Algeria Mohammed First University, Morocco Mohammed First University, Morocco Mohammed V University, Morocco Gaston Berger University, Senegal University of Namur, Belgium University of Minho, Portugal University of Porto, Portugal University College London, UK Technical University of Kosice, Slovakia INRA, Morocco Sidi Mohamed Ben Abdellah University, Morocco Ibn Tofail University, Morocco University of Concordia-Montreal, Canada Universidade NOVA de Lisboa, Portugal Mohammed First University, Morocco University of Franche-Comté, France Mohammed First University, Morocco Instituto Superior de Engenharia do Porto, Portugal INSA de Rennes, France Technical University of Sofia, Bulgaria T&K Engineering, Sophia, Bulgaria State Pedagogical University, Moldova Delft University of Technology, Netherlands Mohammed First University, Morocco Institute of Management Sciences, Pakistan Cadi Ayyad University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco International Burch University, Bosnia and Herzegovina University of Calcutta, India Mohammed First University, Morocco Chouaib Doukkali University, Morocco Mohammed First University, Morocco University of Mumbai, India INSA Rennes, France Mohammed First University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Ecole Centrale de Nantes, France

Organization

Vincent Simonneaux Walid Osamy Wassima Aitfares William Irvin Grosky Xin Cao Yamina Khlifi Youssef Hanyf Youssef Regad Zineb Bougroun Zohra Bakkoury Zoltán Balogh Zohra Zerdoumi

xiii

CESBIO, France Qassim University, Saudi Arabia Mohammed V University, Morocco University of Michigan-Dearborn, USA University of New South Wales, Australia Mohammed First University, Morocco Chouaib Doukkali University, Morocco Mohammed First University, Morocco Mohammed First University, Morocco Mohammed V University, Morocco Constantine the Philosopher University in Nitra, Slovakia M’sila University, Algeria

Contents

Communications Systems and Applications Heuristic for Network Planning Based on 5G Services . . . . . . . . . . . . . . M. Umar Khan, Mostafa Azizi, Ana García Armada, and J. J. Escudero Garzás

3

Performance Improvement of OFDM-ROF System with Combined Adaptive Coded Modulation and Power Control . . . . . . . . . . . . . . . . . . Mohammed Amine Azza, Moussa El Yahyaoui, and Ali El Moussati

15

Optical Architecture for 60 GHz 4 × 4 MIMO Signal Distribution over Optical Fiber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Moussa El Yahyaoui, Hachim Azzahhafi, and Ali El Moussati

23

Evaluation of Railway Communications System Based on 5G-RoF Technology and Millimeter Wave Band . . . . . . . . . . . . . . . . Hachim Azzahhafi, Moussa El Yahyaoui, and Ali El Moussati

30

Digital Video Broadcasting - Satellite - Second Generation Measurement and Test for Database Simulation . . . . . . . . . . . . . . . . . . Youssef Bikrat, Khalid Salmi, Ahmad Benlghazi, Abdelhamid Benali, and Driss Moussaid

38

Design of a Microstrip Patch UWB Antenna with Notch Band Characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . L. Aguni, S. Chabaa, S. Ibnyaich, and A. Zeroual

47

60 GHz RoF System Based on IR-MBOOK Transmitter and Non-coherent Receiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tarik Zarrouk, Ali El Moussati, Papa Alioune Fall, and Ghaïs El Zein

55

Impact of Human Morphology on Measurement Errors of a RF Exposimeter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abdechafik Derkaoui, Rodrigues Kwate Kwate, Bachir Elmagroud, Dominique Picard, and Abdelhak Ziyyat

63

xv

xvi

Contents

RF-Exposimeter Errors Measurement: Frequency and Distance Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rodrigues Kwate, Bachir Elmagroud, Abdechafik Derkaoui, Chakib Taybi, Dominique Picard, and Abdelhak Ziyyat

71

Computer Vision and Data Processing Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Peeter Lorents, Ahto Kuuseok, and Erika Lorents

83

Local Binary Pattern and Its Variants: Application to Face Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jade Lizé, Vincent Débordès, Hua Lu, Kidiyo Kpalma, and Joseph Ronsin

94

Reducing LBP Features for Facial Identification and Expression Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Joseph Ronsin, Kidiyo Kpalma, and Hua Lu Video Retrieval Using Query Images and CNN Features . . . . . . . . . . . . 112 Imane Hachchane, Abdelmajid Badri, Aïcha Sahel, and Yassine Ruichek CUDA Accelerating of Fractal Texture Features for a Neuro-morphological Image Segmentation Approach . . . . . . . . . . 121 Khalid Salhi, El Miloud Jaara, and Mohammed Talibi Alaoui Efficient Mapping Method for Elliptic Curve Cryptosystems Based on PWLCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 Salma Bendaoud, Fatima Amounas, and El Hassan El Kinani 3D Shape Recognition Based on Uncoded Structured Light Using ANN Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Kaoutar Baibai, Mohamed Emharraf, Wafae Mrabti, Khalid Hachami, and Benaissa Bellach Machine Vision-Based Cocoa Beans Fermentation Degree Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 Yro Aubain, N’Zi Édié Camille, and Kpalma Kidiyo Plants Classification Using Neural Shifted LegendreFourier Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Abderrahmane Machhour, Amal Zouhri, Mostafa El Mallahi, Zakia Lakhliai, Ahmed Tahiri, and Driss Chenouni Cybersecurity and Data Protection Criteria for Security Classification of Smart Home Energy Management Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Manish Shrestha, Christian Johansen, and Josef Noll

Contents

xvii

Secure Linear Regression Algorithms: A Comparison . . . . . . . . . . . . . . 166 Fida Dankar and Nisha Madathil Multi-agents Intrusion Detection System Using Ontology for Manets . . . 175 Sara Chadli, Hajar Chadli, Mohammed Saber, Mohammed Ghaouth Belkasmi, Ilhame El Farissi, and Mohamed Emharraf Analysis of KDD Dataset Categories to Design a Performing Intrusion Detection System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 Ilhame El Farissi, Mohammed Saber, Sara Chadli, Zineb Bougroun, Mohamed Emharraf, Mohammed Ghaouth Belkasmi, and Rachida El Mehdi A Comparative Performance Analysis of the Intrusion Detection Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 Mohammed Saber, Zineb Bougroun, Ilhame El Farissi, Sara Chadli, Mohamed Emharraf, Saida Belouali, Mohammed Ghaouth Belkasmi, and Ilham Slimani New Improvement of Malware-Attack Scenarios Modeling . . . . . . . . . . 201 Noureddine Rahmoun, Yassine Ayachi, Jamal Berrich, Mohammed Saber, and Toumi Bouchentouf IoT Security Management: Model and Design Issues . . . . . . . . . . . . . . . 212 Ghizlane Benzekri, Omar Moussaoui, and Ali El Moussati Energy and MultiSource Systems Management Comparison Between Constant and Variable Switching Frequency Strategies Based Direct Torque Control of Asynchronous Motor . . . . . . 223 Soukaina El Daoudi, Loubna Lazrak, Chirine Benzazah, and Mustapha Ait Lafkih Simulation and Analysis of Enhanced Perturb and Observe MPPT Algorithm Based on an Adaline Neural Network for Standalone PV System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 Ihssane Chtouki, Houssam Eddine Chakir, Patrice Wira, Malika Zazi, and Bruno Collicchio Performance Assessment of Solar Dish-Stirling System for Electricity Generation in Eastern Morocco . . . . . . . . . . . . . . . . . . . . 244 Hanane Ait Lahoussine Ouali, Benyounes Raillani, Samir Amraqui, Mohammed Amine Moussaoui, and Ahmed Mezrhab Real Time Implementation of SPWM Signal Generation Technique for a New Five Level Inverter Using Microcontroller . . . . . . . . . . . . . . . 253 Hajar Chadli, Zakariae Jebroni, Sara Chadli, Mohammed Saber, Khalid Salmi, Abdechafik Derkaoui, and Abdelwahed Tahani

xviii

Contents

Design of a PWM Sliding Mode Voltage Controller of a DC-DC Boost Converter in CCM at Variable Conditions . . . . . . . . . . . . . . . . . . 263 Weam El Merrassi, Abdelouahed Abounada, and Mohamed Ramzi Design and Performance Analysis of Super-Twisting Algorithm Control for Direct-Drive PMSG Wind Turbine Feeding a Water Pumping System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 Benzaouia Soufyane, Zouggar Smail, Rabhi Abdelhamid, and Mohammed Larbi Elhafyani Electric System Cascade Analysis for Optimal Sizing of an Autonomous Photovoltaic Water Pumping System . . . . . . . . . . . . 282 Mohammed Chennaif, Mohamed Larbi Elhafyani, Hassan Zahboune, and Smail Zouggar Techno-Economic Sizing of a Stand-Alone Hybrid Energy and Storage for Water Pumping System . . . . . . . . . . . . . . . . . . . . . . . . 291 Mohammed Chennaif, Hassan Zahboune, Mohammed Larbi Elhafyani, and S. Zouggar Rotating Machines Energy Loss Due to Unbalance . . . . . . . . . . . . . . . . 300 Ali Elkihel, Bouchra Abouelanouar, and Hassan Gziri Comparative Study Between PI Speed Control and Sliding Mode Control of BLDC Motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 Ahmed Loukmane El Idrissi, Jamal Bouchnaif, Mohammed Mokhtari, and Anas Bensliman PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine Using a Fuzzy Logic Controller . . . . . . . . . . . 318 Mhamed Fannakh, Mohamed Larbi Elhafyani, Hassan Zahboune, and Smail Zouggar Contribution to Power Maximization of an Asynchronous Wind Electric Water Pumping System Using Single Input Fuzzy Logic Controller and Modified Enhanced Perturb and Observe . . . . . . . . . . . 330 Mohammed Mokhtari, Smail Zouggar, Nacer K. M’sirdi, and Mohamed Larbi Elhafyani Hybrid System Energy Management in a Low Power Isolated Site . . . . 343 Mohammed Larbi El Hafyani, Abdelmalek El Elmehdi, Smail Zouggar, and Toufik Ouchbel Machine Learning, Intelligent Systems and Applications Citation Classification Using Natural Language Processing and Machine Learning Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357 Syyab Rahi, Iqra Safder, Sehrish Iqbal, Saeed-Ul Hassan, Iain Reid, and Raheel Nawaz

Contents

xix

Path Planning Algorithm for Initially Unknown Indoor Environment Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 Mohamed Emharraf, Mohammed Saber, Mohammed Ghaouth Belkasmi, Ilhame El Farissi, Sara Chadli, and Mohammed Rahmoun Ontology-Based Reasoning for Collective Intelligence of Multi-agents System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377 Yman Chemlal and Hicham Medroumi Prediction of Direct Normal Irradiance Using Artificial Neural Networks Under Oujda Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386 Latifa El Boujdaini, Ahmed Mezrhab, and Mohammed Amine Moussaoui Corpus Construction and Annotation Challenge for Language Identification and Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 395 Ibtissam Touahri and Azzeddine Mazroui Feature Selection for Community Evolution Prediction in Location-Based Social Network: Gowalla and Brightkite . . . . . . . . . . 404 Loubna Boujlaleb, Ali Idarrou, and Driss Mammass Precision Agriculture CropSAT – A Decision Support System for Practical Use of Satellite Images in Precision Agriculture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415 Omran Alshihabi, Kristin Piikki, and Mats Söderström Rice Yield Prediction Using On-Farm Data Sets and Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422 Oscar Barrero, Sofiane Ouazaa, Camilo Ignacio Jaramillo-Barrios, Mauricio Quevedo, Nesrine Chaali, Sair Jaramillo, Isidro Beltran, and Omar Montenegro Inter-comparison Between Different Techniques for Evapotranspiration Partitioning: Eddy Covariance-, Sap Flow-, Lysimeter- and FAO-Based Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 431 Zoubair Rafi, Olivier Merlin, Valérie Le Dantec, Saïd Khabba, Salah Er-Raki, and Patrick Mordelet Effects of Climate Change at the 2040’s Horizon on the Hydrology of the Pluvio-Nival Rheraya Watershed Near Marrakesh, Morocco . . . . 440 Youssef Hajhouji, Younes Fakir, Vincent Simonneaux, Simon Gascoin, El Houssaine Bouras, and Abdelghani Chehbouni Estimation of the Evapotranspiration over Heterogeneous Region Using Shuttleworth-Wallace Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451 Jamal Elfarkh, Salah Er-Raki, Jamal Ezzahar, Lionel Jarlan, Said Khabba, and Abdelghani Chehbouni

xx

Contents

Intelligent Agriculture Platform Based on Low Energy and Cost Wireless Sensors for Efficient Water Irrigation . . . . . . . . . . . . . . . . . . . 460 Mohamed Emharraf, Hamza Taous, Wiame Benzekri, Ali El Moussati, and Kamal Aberkani Smart Health, Digital Education and Humanities The Effects of Neurofeedback on Event Related Potential (ERP) in Zikr Meditation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473 Nur Arina Ayuni Helman and Muhamad Kamal Mohammed Amin Exploring Eye-Gaze Behaviors on Emotional States for HumanComputer Interaction (HCI) Using Eye Tracking Technique . . . . . . . . . 481 Sumita Thiyagarajan and Muhamad Kamal M. Amin Novel Alignment Approach of DNA Sequences . . . . . . . . . . . . . . . . . . . 489 Wajih Rhalem, Jamel El Mhamdi, Mourad Raji, Ahmed Hammouch, Nassim Kharmoum, Sanae Raoui, Saaid Amzazi, Salsabil Hamdi, and Hassan Ghazal Understanding the Study Experiences of Students in Low Agency Profile: Towards a Smart Education Approach . . . . . . . . . . . . . . . . . . . 498 Ville Heilala, Päivikki Jääskelä, Tommi Kärkkäinen, and Mirka Saarela Let Me Hack It: Teachers’ Perceptions About ‘Making’ in Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509 Ville Heilala, Mirka Saarela, Sanna Reponen, and Tommi Kärkkäinen Comparing the Effect Size of School Level Support on Teachers’ Technology Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519 Eloho Ifinedo Tracking Entrepreneurial Mind Among University Students Through Functional Near-Infrared Spectroscopy (fNIRS) Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527 Nur Izzati J. Sham and Muhamad Kamal Mohammed Amin Ethics of AI or Ethical AI, Topical Point of View . . . . . . . . . . . . . . . . . 535 Saida Belouali, Anas Belouali, Mohammed Saber, Khalid Jaafar, and Mohammed Ghaouth Belkasmi Open Access Publishing and Ethical Issues in the Moroccan Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541 Nadia Benaissa and Saïda Belouali Intelligent Model for Evaluation Based on Expert System and Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548 Khalid Salmi, Hanane Sefraoui, Hamid Magrez, Abdechafik Derkaoui, Abdelaziz Elmoussaouy, and Abdelhak Ziyyat

Contents

xxi

Software Engineering and Data Science Artificial Neural Networks for Text-to-SQL Task: State of the Art . . . . 557 Youssef Mellah, Hassane El Ettifouri, Toumi Bouchentouf, and Mohammed Ghaouth Belkasmi SVM on HPC Clouds: Choosing the Appropriate Classification Algorithm and Kernel Type According to the Data Set Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566 Mouncef Filali Bouami and Mohamed Benchat Global IT Project Management: An Agile Planning Assistance . . . . . . . 575 Mohammed Ghaouth Belkasmi, Zineb Bougroun, Ilhame El Farissi, Mohamed Emharraf, Saida Belouali, Sara Chadli, and Mohammed Saber A Survey for Validation Concepts to Measure Quality as Well Their Application on the Maintainability of ISO . . . . . . . . . . . . . . . . . . 583 Zineb Bougroun, Mohammed Saber, Ilhame El Farissi, Ghaouth Mohammed Belkasmi, and Toumi Bouchentouf NL2Code: A Corpus and Semantic Parser for Natural Language to Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592 El Hassane Ettifouri, Walid Dahhane, Achraf Berrajaa, Toumi Bouchentouf, and Mohammed Rahmoun Model-Based Testing from Model Driven Architecture: A Novel Approach for Automatic Test Cases Generation . . . . . . . . . . . 600 Imane Essebaa, Salima Chantit, and Mohammed Ramdani Solar Thermal and Mechanics Optical Degradation of CSP Reflectors Under Moroccan-Eastern Climate: An Experimental Investigation . . . . . . . . . . . . . . . . . . . . . . . . . 613 Mouatassim Charai, Latifa Elboujdaini, Ahmed Mezrhab, Abdelhamid Mezrhab, and Mustapha Karkri Modeling Spring Impact on Durability of Welded Structure for Electric Vehicles Utilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621 Imane Amarir, Hamid Mounir, Abdellatif El Marjani, and Kaoutar Daoudi Accurate Evaluation of Solar Irradiation of a Satellite Dataset Under Ground Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629 Latifa El Boujdaini, Ahmed Mezrhab, Abdelhamid Mezrhab, Mohammed Amine Moussaoui, and Mouatassim Charai Validation and Numerical Study of an Earth-to-Air Heat Exchanger for Cooling and Preheating . . . . . . . . . . . . . . . . . . . . . . . . . 638 Haitham Sghiouri, Mouatassim Charai, Ahmed Mezrhab, and Mustapha Karkri

xxii

Contents

NaI(Tl) Detector Response at Different Energies and a Validation with Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647 Abdelkarim Bazza, Abdelkader El Hamli, Mohammed Hamal, Abdellah Moussa, Mostapha Zerfaoui, Lahsen Hamam, Mohammed Ouchrif, and Yahya Taylati Running GATE Software on Moroccan Cluster Computing to Simulate Particle Interactions Within Linear Accelerator System . . . 656 Deae-eddine Krim, Abdeslem Rrhioua, Dikra Bakari, Mustapha Zerfaoui, and Yassine Oulhouq Analysis and Optimization of SM and TES Hours of Central Receiver Concentrated Solar Thermal with Two-Tank Molten Salt Thermal Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666 Hanane Ait Lahoussine Ouali, Benyounes Raillani, Samir Amraqui, Mohammed Amine Moussaoui, Abdelhamid Mezrhab, and Ahmed Mezrhab Wireless Sensor Network, Internet of Things and Applications Wireless Network Stability Enhancement Based on Spatial Dependency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677 Halim Berradi, Ahmed Habbani, Chaimae Benjbara, Nada Mouchfiq, and Hicham Amraoui Modeling and Simulation of LoRaWAN for Smart Metering Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687 Zakariae Jebroni, Hajar Chadli, Khalid Salmi, Mohammed Saber, and Belkassem Tidhaf Early Forest Fire Detection with Low Power Wireless Sensors Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696 Wiame Benzekri, Ali El Moussati, Omar Moussaoui, and Mohammed Berrajaa Study and Optimization of the System Energy in WSN with Global and Sequential Experiment Designs . . . . . . . . . . . . . . . . . . 705 Mohammed Jabri, Omar Moussaoui, Mimoun Moussaoui, and Ali El Moussati LoRa Based Smart Electrical Energy Meter . . . . . . . . . . . . . . . . . . . . . 716 Zakariae Jebroni, Hajar Chadli, Khalid Salmi, Mohammed Saber, and Belkassem Tidhaf Towards Reliable and Timely Communications in Wireless Body Area Networks: A Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . 724 Azdad Nabila and Elboukhari Mohamed Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 737

Communications Systems and Applications

Heuristic for Network Planning Based on 5G Services M. Umar Khan1,2(B) , Mostafa Azizi2 , Ana Garc´ıa Armada1 , and J. J. Escudero Garz´ as1,3 1

2

Department of Signal Theory and Communications, Universidad Carlos III de Madrid, Madrid, Spain [email protected] MATSI Laboratory, ESTO, Universit´e Mohammed Premier, Oujda, Morocco 3 Department of Industrial and Systems Engineering, Texas A&M University, College Station, USA

Abstract. Cellular communications have evolved to Fifth Generation (5G) to accommodate versatile use cases (UCs) requirements. These UCs are ensured to provide higher data rates, massive connection density and ultra-low latencies into a new era of Internet of Things (IoT) and smart cities. ITU-R has classified UCs into three categories: Enhanced mobile broadband (eMBB), Massive machine-type communication (mMTC) or mIoT (massive IoT), and Ultra-reliable and low-latency communications (URLLC). It is very important to plan the network based on different UC requirements using 5G new radio (NR). In this paper, we design the numerology and corresponding bandwidth parts (BWPs) to support the desired requirements. Exercise of coverage and capacity dimensioning of the network is performed to determine the required cell sites. We then propose mixed-integer linear programming (MILP) based cost optimization model and heuristic. Finally, we evaluate and compare MILP and heuristic topology solutions for network planning in the context of cost minimization. Keywords: 5G

1

· Heuristics · Optimization · Network planning

Introduction

One of the major differences of 5G compared to previously evolved cellular networks is that it is not primarily focused on human-centric services. But it also addresses machine-centric applications and services such as mMTC/m-IoT and URLLC. Earlier communication networks were designed and planned for telephony coverage, broadband internet at home and multimedia mobile internet services. Therefore, existing macrocell networks are not capable to fulfil the demands of the 5G user. Besides operational expenditure (OPEX) and capital expenditure (CAPEX), we must consider the diverse 5G UC requirements in the network planning model. Moreover, dense deployment of small cells is expected c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 3–14, 2020. https://doi.org/10.1007/978-3-030-53187-4_1

4

M. Umar Khan et al.

to increase capacity and coverage with the introduction of 5G new radio (NR). To explore potential sites for small cell installations in the context of cost, coverage and capacity for network planning should include backhaul considerations [1]. Authors in [2] have considered 5G UCs in a real geographical area to present fronthaul/backhaul network architecture to fulfil demands of users and operators perspective. Another backhaul based extensive MILP framework has been studied in [3] to provide optimal network planning in terms of cost and performance. Moreover, the effectiveness of network planning also depends on the information of city infrastructures like dimension of roads, traffic signals, lamppost locations, buildings and sewage infrastructure. This information can be easily made available by utilizing open-source geographical information system (GIS) like QGIS [4]. Authors in [5] have considered GIS information and presented network planning optimization using an open source network planning software. However, we believe that the efficacy of 5G network planning depends on UCs requirements and it should be incorporated as an integral part of the network planning model. This manuscript is structured in the following manner. First, an overview of the three UCs and their corresponding minimum service requirement by ITUR are presented in Sect. 2. In Sect. 3, we design the Numerology to support the desired requirements. In Sect. 4, 5G coverage and capacity dimensioning is performed based on numerology parameters from Sect. 3. In Sect. 5, coverage and capacity output from Sect. 4 is taken as input to the optimization model for the deployment of supported 5G network topology with minimum cost. Eventually, results from the proposed network planning heuristic and MILP are discussed in Sect. 6. We conclude in Sect. 7 providing future research directions.

2

ITU-R 5G Use Case Requirements

Many reference from different standardization bodies and stakeholders like ITU, 5G-PPP, ETSI, and NGMN analyze the role of 5G services and UCs in detail [6–8]. The UCs that support 5G services fall into three categories according to ITU-R [9]. The first category eMBB includes applications such as augmented reality (AR), 3D video, ultra-high definition (UHD) screens, work and play in the ground etc. Data rate requirements of this UC are from 100 Mbps–10 Gbps with low latency of 1 ms and 20 ms for user and control plane respectively. The second category mMTC/mIoT refers to any time, anything and anywhere communication model [10]. This UC has the minimum requirement for massive connection density of 106 /km2 according to IMT-2020 radio interface [11]. URLLC is the third category which is sensitive to latency and reliability as it involves missioncritical applications. Latency requirements of such applications are 1–2 ms and 10 ms for user and control planes respectively [12].

3

5G Numerology

To adopt the desired requirements in upcoming 5G spectrum, 3rd Generation Partnership Project (3GPPP) has proposed New Radio (NR) with new frequency

Heuristic for Network Planning Based on 5G Services

5

bands ranging from 1 GHz–100 Ghz [13]. Moreover, each 5G UC requires a different range of parameters such as cyclic prefix (CP) length, subcarrier spacing (Δf ), number of physical resource blocks (NRB ), number of slots per subframe subf rame,μ (NRB ), slot duration (Tslot ) and OFDM symbol durations. And these parameters can be configured through numerology to support the desired requirements [14] in the access network where the required number of next-generation node B (gNB) could be deployed to fulfil capacity and coverage in the target area. Numerology in 5G has been defined as the flexibility to use multiple values of Δf and is characterized by the parameter μ, as shown in Table 1. Table 1. 5G transmission numerologies μ Δf = 2μ ∗ 15 KHz Cyclic Prefix

subf rame,μ NRB Tslot = 1 ms/2μ

0

15

Normal

1

1 ms

1

20

Normal

2

0.5 ms

2

60

Normal

4

0.25 ms

8

0.12 ms

16

0.06 ms

3 120

Normal, Extended

4 240

Normal

The concept of numerology is to use suitable waveform parameters for the required UC bandwidth and latency requirements. We consider the maximum transmission bandwidth configuration from 3GPP release 15 [15] to design numerology parameters for 5G UCs. For eMBB latency of Tslot = 1 ms can be acceptable with high data rates requirements by choosing μ = 0 corresponding to Δf of 15 Khz and providing 270 NRB in 50 MHz bandwidth part. However, in the case of URLLC/mMTC low data rates with lower latencies are required which corresponds to μ = 1 with Δf of 30 Khz providing Tslot = 0.25 ms with 10 MHz. The values of Δf depends upon the selected frequency band. For instance, 5G NR supports frequency range one (FR1) FR1 < 6GHz and FR2 > 6 GHz. We have chosen FR1 which allows subcarrier spacings of 15 KHz and 30 Khz for μ = 0 and μ = 1 respectively. The designed numerology with two bandwidth parts (BWP) configuration can be described by Fig. 1. In overall available carrier, we have two numerology configurations for μ = 0 and μ = 1. BWP1 can provide high data transmission to eMBB and BWP2 is meant for mMTC/URLLC requiring lower latencies with low data payloads. In BWP1, contiguous PRBs are available ranging from 0 − N 1 such that N 1 = NRB − 1 and for BWP2 PRBs are from 0 − N 2 and N 2 = NRB − 1, where NRB is 270 and 24 respectively. Moreover, the maximum achievable data rate for the configured BWPs is based on numerology factor μ. Any 5G device supporting Multi-RAT (Radio Access Technology) Dual Connectivity (MR-DC) can have maximum data rates defined by 3GPPP release 15 version 15.2.0 [16] given in Table 2.

6

M. Umar Khan et al.

Fig. 1. Considered bandwidth part configuration Table 2. Maximum data rates Device Type: eMBB

DL: 1.1 - UL: 1.2 (Gbps)

Device Type: mMTC/URLLC DL: 205.45 - UL: 219.78 (Mbps)

4

5G Coverage and Capacity Dimensioning

The requirements of UCs in Sect. 2 uncover information to be considered for optimization models like backhaul and its technological limitations to meet the capacity required by the UCs. For coverage dimensioning in the desired area, we proceed to analyse the target area through link budget analysis. This analysis include all potential gains and losses from the transmitter (Tx) through channel to the receiver (Rx). Link budget analysis includes the attenuation of signal due to propagation losses, feeder losses, fading and interference margins and possible antenna gains, as shown in Table 3. We adopt here the channel model from [17] as it considers the path loss as the dominant factor in outdoor scenarios. The mean path loss between gNB and a UE is presented as P Lmean = K  − 10α ∗ log10 (R), where K  = 137dB is a path loss constant and α = 35.2dB is path loss exponent and R represents the range of cell [17]. 4.1

Coverage Dimensioning

Coverage dimensioning refers to compute the required number of cell sites to provide coverage in the target area A. We first determine the maximum path loss through link budget analysis with parameters of Table 3. The maximum path loss P Lmax can then be calculated by Eq. (3): P Lmax = EEIRPsc − (SN Rthr + T N C + M + I),

(1)

where as EEIRPsc is the effective power per subcarrier, SN Rthr = −2.1dB is the threshold SNR, T N C is the total thermal noise in the channel with fading, and interference margins denoted by M and I respectively. Satisfying the maximum path loss P Lmax , any point in the target area A is said to have coverage, if UE can make a successful communication link with gNB. Therefore, micro-cell coverage radius denoted by Cradius can be computed as Cradius = 10

K  −P Lmax −10η

.

Heuristic for Network Planning Based on 5G Services

7

With the determined coverage radius we can calculate the coverage area Acell of the hexagonal cell with area coefficient Acoef f = 1.95 [17] given in Table 4. can be determined to cover Finally, the number of micro cells cites Nsites = AA cell the entire target area A. Table 3. Parameters for micro cell link budget analysis Parameters

Equation

Value

Unit

Micro-cell transmit power

Ptx

33

dBm

Transmit antenna gain

Gtx

18

dBi

Cable connector loss

Ltx

2.5

dB

EIRP

EIRP = Ptx + Gtx − Ltx

47

dBm

Total subcarriers

Nsc

3300

100 MHz

EIRP per subcarrier

EIRPsc = EIRP − 10 ∗ log10 (Nsc )

48.5

dBm

Thermal noise

TN

−173.8

dBm

Subcarrier spacing

Δf = 2μ ∗ 15 KHz

15

KHz

Thermal noise in channel

T N C = T N + 10log10 (Δf )

−132.04

dBm

Rx noise figure

RN F

7

dB

Fading margin

M

5

dB

Interference margin

I

2

dB

Control channel overhead

COH

0.14

dB

Receiver antenna gain

Grx

0

dB

Penetration loss

Lpen

18

dB

Effective EIRPsc

EEIRPsc = EIRPsc + Grx − (RN F + Lpen + COH )

−11.8

dBm

Table 4. Parameters for coverage dimensioning Parameters

Equation

P Lmax

(3)

Cradius

Cradius = 10

Value

115.34 dB K  −P Lmax −10η

Site area coefficient Acoef f

4.2

Unit

.867

km

1.95

–

Cell area

2 Acell = Acof f ∗ Cradius 1.4

km2

Number of sites

Nsites = A/Acell

–

51

Capacity Dimensioning

Capacity dimensioning refers to determining the required number of cell sites supported on a given cell load, capacity cell range, gNB user capacity and traffic density in the area. For instance, we have considered the summation of eMBB and mMTC/URLLC user densities as γA = 550/km2 , gNB site capacity 2100 users [17]. Then considering cell load at 38.49%, a single gNB can simultaneously serve SCload = 808 users. Now capacity cell range can be determined by the

8

M. Umar Khan et al.

residing area of the users as SCload /γA = 1.46 km2 . From this residing area, we  and consequently the number of sites due to can calculate the cell range Cradius capacity dimensioning given in Table 5. Capacity cell range is a function of cell load, the more the cell load the more is the cell range. However, a significant part of planning is to look over the results obtained in the coverage dimensioning. In our case, cell load at 38.49% gives a cell range of 0.865 km which is very close to the value 0.867 km obtained in the coverage dimensioning. With the considered parameters of gNB transmitting power and user capacity our plan supports maximum cell load of 38.49%. Moreover, there exists a trade-off between coverage, capacity and efficiency in the network planning. For example, in our case, Nsites obtained through coverage dimensioning is greater in number than Nsites obtained through capacity dimensioning. Let us assume that we consider the Nsites obtained in the coverage dimensioning, the network plan for access nodes gNBs, will be less efficient because of underutilization of gNBs. On the other hand, choosing Nsites determined through capacity dimensioning will be a balanced approach in terms of efficiency and cost minimization. Table 5. Parameters for capacity dimensioning Parameters

Equation

Value Unit

User density

γA

550

/km2

Site capacity

Scapacity

2100

–

Cell load

Cload

38.49 %

Capacity in cell load SCload

5

808

–

.865

km

Capacity radius

 Cradius

Number of sites

Nsites = A/Acell 49

–

Optimization Model and Heuristic

In this section, we present a MILP based optimization model for the distribution and access network. We have developed this framework on the network segment containing distribution node (dn) and gNB also known as small cells in 5G terminology. This segment of the backhaul is of utmost importance as it decides the capacity and coverage requirements. Our aim is the minimization of the total network cost of the considered backhaul segment with coverage and capacity constraints, as such deployments involve a high cost and impact directly on the capacity and coverage of the network. The backhaul segment has a set D of distribution nodes dni ∈ D. These are the potential locations where the distribution nodes can be placed. The access network consists of gN Bj ∈ B, where B is a set of access nodes. The set B is taken as an input to the optimization model from

Heuristic for Network Planning Based on 5G Services

9

the previous section of coverage and capacity dimensioning. Each gN B must be directly linked with one and only one dn to the distribution network. These links in a set L are represented by lij if ith dn is linked with j th gN B. The backhaul segment has a cost associated with the deployment of dn and gN B nodes plus the links. We express Γij as the link cost of connecting ith dn to j th gN B, while the cost of installing dn plus the cost of the node itself is expressed as ηi . We formulate our objective function with two decision variables;  1; if ith dni is connected with j th gN Bj xij = (2) 0; otherwise  1; if ith dni has been selected dni ∈ D yi = 0; otherwise   Γij xij + ci ηi yi , min ij



(3) (4)

i

xij = 1, ∀j

(5)

xij ≤ ζyi , ∀i

(6)

i

 j

The objective function (4) sums the total   deployment cost, which includes link cost ( ij Γij xij ) and the node cost ( i ηi yi ). Whereas ci is the centrality metric is simply the measure of being in the middle of all the access nodes, the of the node. Centrality can be determined lower the ci the higher is the centrality  2  2   xdni − xgN Bj + ydni − ygN Bj , ∀j . by the Euclidean distance as ci = i Constraint (5) ensures the coverage, means that each gN B should be connected to exactly one dn. Eventually, constraint in (6) fulfils the capacity constraint of considered backhaul technology. This constraint has two meanings; first, if ith dn  is not selected (yi = 0), then none of the gN B can be connected which leads to j xij = 0. Secondly, if ith dn is selected (yi = 1), then ζ which is a capacity limitation factor restricts maximum number of gNBs that can be served by each dn. This factor is dependent on the technology being used in dn. For instance, if the backhaul being used is a passive optical network (PON), the use of optical splitter at dn allows to feed gNBs according to its capacity. However, in wireless backhaul network ζ is depended on the total resource blocks being available at dn for gNB. 5.1

Proposed Heuristic

Optimization model discussed in the previous section is linear in nature and can be solved in polynomial time. However, with the introduction of SDN/NFV (Software Defined Networking/Network Function Virtualization), SDN controllers constantly improve the network performance through the information of

10

M. Umar Khan et al.

network topology (or change in network topology), capacity and coverage details [18]. Therefore, our proposed heuristic serves as a practical, a fast and feasible solution for network topology planning problems and possible interconnect and input to the online network optimization problems in SDN/NFV [19].

Algorithm 1. Minimum Cost Network Planning Heuristic 1: procedure NetworkPlan(D, B, ζ) 2: Phase 1: Initiate Network Planning 3: Compute dnmin 4: Determine Network Cost using equation (3) 5: Select dnmin with high centralities 6: Set topology with nearby gN Bs such that the cost is Minimum 7: Return: Initial Topology 8: Phase 2: Fulfil Capacity Constraint 9: while ζ is not fulfilled for each dn do 10: Compute node degrees for each dn 11: if (node degree for dn > ζ) then 12: Compute nearest dns 13: Exchange distant gN Bs with nearest dns 14: end 15: end 16: Update: Initial Topology

The proposed heuristic consists of two phases. In the first phase, we compute dnmin = |B| ζ , the minimum number of distribution nodes to serve gNBs. We select dnmin distribution nodes from set D, such that they have a high score of centrality. This phase returns the initial network topology by solving the relaxed MILP problem not considering the capacity constraint in (6). In the second phase, we improve the topology by imposing the capacity constraint. We compute the Euclidean distance between each dn node to see which are the nearest nodes. We iterate the network topology and exchange the access nodes until the capacity constraint is fulfilled.

6

Results and Discussion

In this section, we show some analysis of the resulting topologies of the considered backhaul segment we obtained from our heuristic. Using random uniform distribution in MATLAB we have plotted available places for dns and possible potential locations for installing gNBs which can be seen in Fig. 2(a). For simplification, Γij is approximated to the Euclidean distance which means unit cost Γij = 1 and ηi is positive constant. When ηi is significantly high, heuristic selects a single dn with the highest centrality as shown in Fig. 2(b). While lowering ηi allow heuristic to reduce the link costs and places minimum required distribution nodes dnmin = |B| ζ with higher centralities as shown in Fig. 3(a), where |B| = Nsites obtained from the capacity dimensioning. Red colour dns are selected to be included in the topological solution by heuristic because of having a higher score of centrality. However, as can be seen in the figure those

Heuristic for Network Planning Based on 5G Services

Fig. 2. (a) Nodes distribution (b) Selected dn with highest centrality

Fig. 3. (a) Initial topology from Phase 1 (b) Current topology at Phase 2

Fig. 4. (a) Final topology from the Phase 2 (b) Optimal Topology from MILP

11

12

M. Umar Khan et al.

blue dns are left apart due to less score of centrality. This score hints the heuristic to adopt dns which will contribute to have minimum link cost to backhaul available gNBs. Phase 1, which initiates the network planning of our heuristic in Algorithm 1 involves the measurement of the centrality and initial network cost. Finally, Phase 2 iterates the current topology to fulfil the capacity constraint. For the evaluation of heuristic, we have taken the value of ζ as 12, which represents the number of connections supported at each potential dn location. In the initial phase 1 of the heuristic, dn 4 does not fulfil the capacity constraint and have connection with 17 gNBs as shown in Fig. 3(a). It can be noticed that dn 3, 8, 7 and 1 has the capacity to accommodate 3, 4, 6 and 3 more nodes respectively. Phase 2 of the heuristic will iterate this topology according to Algorithm 1 unless capacity constraint in (6) is fulfilled. Capacity constraint is fulfilled by exchanging nodes with nearby dns on the basis of Euclidean distance. The current topology is shown in Fig. 3(b) of phase 2, it is the result of the first iteration of the heuristic. The nearest node for dn 4 is dn 3 and it has the capacity to receive only three nodes from dn 4. Hence, three nodes which are closer to dn are being assigned to dn 3. Now, dn has two more nodes to give, at this stage the second nearest node for dn 4 is dn 1, which has the tendency to take three more nodes. Finally, dn 4 assign its last two nodes which are closer to dn 1 and resulting topology can be seen in Fig. 4(a). The final topology has fulfilled the upper bound of capacity constraint which was set to ζ = 12. We have also evaluated MILP problem for optimal topology solution through MATLAB and presented in Fig. 4(b). Topology comparison in terms of network deployment cost between the proposed heuristic and MILP evaluation has been shown in Fig. 5. We have determined the network cost from 100 random deployments of the network topologies. We observed that the heuristic technique on average has 11% additional cost compared with the MILP evaluations. 1

Normalized Network Cost

0.9

Heuristic Cost Optimal Cost

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

10

20

30

40

50

60

70

80

No of random deployments

Fig. 5. Total deployment cost

90

100

Heuristic for Network Planning Based on 5G Services

7

13

Conclusions

With the succeeding 5G, network planning optimization based on use case requirements concerning coverage, capacity and the optimal cost are essential. In this work, we presented an ILP based network planning framework and also proposed low complexity heuristic. Our topological results show that capacity and coverage constraints can be achieved by providing backhaul to all the gNBs by optimally choosing the dn with a high score of centrality. We have considered the backhaul access network to fulfil capacity and coverage constraints such that deployment cost is minimum. Obtained results show how the cost associated with links and nodes affect the resulting topology in the context of backhaul technological constraints of capacity and coverage. Moreover, comparison of optimal and heuristic solution reveals that heuristic solution incurs 11% additional cost of the network deployment. However, In future, we aim to explore coverage, capacity, quality of service (QoS) and cost trade-off in the network planning phase of different 5G scenarios. Acknowledgement. This work has been supported by the Spanish National Projects ELISA (TEC2014-59255-C3-3-R) and TERESA-ADA (TEC2017-90093-C32-R) (MINECO/AEI/R, UE).

References 1. Lim, C., Ranaweera, C., Wong, E., Nirmalathas, A.: Design and planning for fiberbased small cell backhauling. In: IEEE 20th International Conference on Transparent Optical Networks (ICTON) (2018). https://doi.org/10.1109/ICTON.2018. 8473967. ISSN 2161-2064 2. Mesogiti, I., Theodoropoulou, E., Ropondi, A.: Fiber-wireless fronthaul/backhaul network architecture for 5G. In: IEEE 23rd International Workshop on Computer Aided Modeling and Design of Communication Links and Networks (CAMAD) (2018). https://doi.org/10.1109/CAMAD.2018.8514991. ISSN 2378-4873 3. Yu, Y., Ranaweera, C.: Optimization and deployment of survivable fiber-wireless (FiWi) access networks with integrated small cell and WiFi. In: IEEE International Conference on Ubiquitous Wireless Broadband (ICUWB) (2015). https://doi.org/ 10.1109/ICUWB.2015.7324491. ISSN 2162-6588 4. QGIS – Open source GIS tool. https://www.qgis.org 5. Romero-G´ azquez, J.L., Bueno-Delgado, M.V., Moreno-Muro, F.J., Pavon-Marino, P.: Net2Plan-GIS: an open-source Net2Plan extension integrating GIS data for 5G network planning. In: IEEE 20th International Conference on Transparent Optical Networks (ICTON) (2018). https://doi.org/10.1109/ICTON.2018.8473784. ISSN 2161-2064 6. Ullah, H., Nair, N.G., Moore, A.: Net2Plan-GIS: 5G communication: an overview of vehicle-to-everything, drones, and healthcare use-cases. In: IEEE Access, pp. 37251–37268 (2019). https://doi.org/10.1109/ACCESS.2019.2905347. ISSN 21693536 7. Wang, Q., Alcaraz-Calero, J., Ricart-Sanchez, R.: Enable advanced QoS-aware network slicing in 5G networks for slice-based media use cases. In: IEEE Transactions on Broadcasting, pp. 1–10 (2019). https://doi.org/10.1109/TBC.2019.2901402. ISSN 0018-9316

14

M. Umar Khan et al.

8. Valtanen, K., Backman, J., Yrj¨ ol¨ a, S.: Blockchain-powered value creation in the 5G and smart grid use cases. In: IEEE Access, pp. 25690–25707 (2019). https:// doi.org/10.1109/ACCESS.2019.2900514. ISSN 2169-3536 9. IT-R, Report: Setting the Scene for 5G: Opportunities & Challenges (2018). https://www.itu.int/en/ITU-D/Documents/ITU 5G REPORT-2018.pdf 10. Dong, L., Mingyue, R., Guoying, M.: Application of internet of things technology on predictive maintenance system of coal equipment. In: 13th Global Congress on Manufacturing and Management, GCMM, vol. 174, pp. 885–889. Elsevier (2017). https://doi.org/10.1016/j.proeng.2017.01.237 11. IT-R, Report M.2410-0: Minimum requirements related to technical performance for IMT-2020 radio interface(s) (2017). https://www.itu.int/pub/R-REP-M.24102017 ˇ Industry 4.0 from 5G perspective: use-cases, requirements, chal12. Stefanovi´c, C.: lenges and approaches. In: 11th CMI International Conference: Prospects and Challenges Towards Developing a Digital Economy within the EU, pp. 44–48 (2018). https://doi.org/10.1109/PCTDDE.2018.8624728 13. Lien, S.Y., Shieh, S.L., Huang, Y., Su, B., Hsu, Y.L., Wei, H.Y.: 5G new radio: waveform, frame structure, multiple access, and initial access. IEEE Commun. Mag. 55(6) (2017). https://doi.org/10.1109/MCOM.2017.1601107 14. Zaidi, A.A., Baldemair, R., Tullberg, H.: Waveform and numerology to support 5G services and requirements. IEEE Commun. Mag. 54(11) (2016). https://doi. org/10.1109/MCOM.2016.1600336CM 15. Technical Specification: 5G NR- User Equipment (UE) radio transmission and reception. In: 3GPP TS 38.101-1 version 15.2.0 Release 15 (2018). https://www.etsi.org/deliver/etsi ts/138100 138199/13810101/15.02.00 60/ts 13810101v150200p.pdf 16. Technical Specification: 5G NR- User Equipment (UE) radio access capabilities. In: 3GPP TS 38.306 version 15.2.0 Release 15 (2018). https://www.etsi.org/deliver/ etsi ts/138200 138299/138214/15.02.00 60/ts 138214v150200p.pdf 17. Jaber, M., Dawy, Z., Akl, N., Yaacoub, E.: Tutorial on LTE/LTE-A cellular network dimensioning using iterative statistical analysis. IEEE Commun. Surv. Tutor. 18(2), 1355–1383 (2016). https://doi.org/10.1109/COMST.2015.2513440 18. Ochoa-Aday, L., Cervell´ o-Pastor, C., Fern´ andez-Fern´ andez, A.: Discovering the network topology: an efficient approach for SDN. Adv. Distrib. Comput. Artif. Intell. J. 5(101) (2016). https://doi.org/10.14201/ADCAIJ201652101108 19. Vassilaras, S., Vigneri, L., Liakopoulos, N.: Problem-Adapted Artificial Intelligence for Online Network Optimization (2018). https://arxiv.org/abs/1805.12090

Performance Improvement of OFDM-ROF System with Combined Adaptive Coded Modulation and Power Control Mohammed Amine Azza(B) , Moussa El Yahyaoui, and Ali El Moussati Department of Electronics, Informatics and Telecommunications, ENSAO, Oujda, Morocco {m.azza,a.elmoussati}@ump.ac.ma, [email protected] Abstract. Radio-Over-Fiber (RoF), the unique blend of both optical and wireless systems, is the last mile solution for increasing the capacity and mobility as well as decreasing the expenses. The idea behind this work is to adapt the RoF system parameters in terms of modulation order, code rate and optical transmitting power according to the channel conditions. This system seeks of optimum parameters in order to deliver high data rates as well as high reliability. Simulation results show better BER performance and optical power efficiency compared to conventional fixed RoF systems. Keywords: Adaptive modulation and coding Adaptive optical power

1

· OFDM · RoF ·

Introduction

Orthogonal Frequency Division Multiplexing (OFDM) technology is widely used in wireless communications and offers excellent performance and high spectral efficiency [1]. Due to its high tolerance against various fiber dispersion effects, OFDM technology is widely used in high-capacity optical applications and nextgeneration passive optical network [2]. For broadband and fixed-point wireless access systems, Radio over Fiber (RoF) appears as a solution for alternative network architecture which offer high flexibility on multiple services access and dynamic bandwidth allocation [3,4]. OFDM-RoF system has proven to be one of the best candidates for high speed wireless communications such as 60GHzWLAN, Wimax, 4G etc. However, the nonlinear effects in optical link as well as the multipath fading in wireless transmission can severely degrade the performance of OFDM-RoF system. In order to overcome these impairments in wireless communication systems, and to enhance the reliability and to increase the system capacity, we propose an RF signal adaptation system depending on the current channel status. Adaptive coding and modulation (ACM) is very effective technique to increase the c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 15–22, 2020. https://doi.org/10.1007/978-3-030-53187-4_2

16

M. A. Azza et al.

system throughput by changing the modulation order and/or the Forward Error Correction (FEC) code rate [5,6]. The main advantage of this adaptation technique is to maximize the Spectral Efficiency (SE) and link capacity using the highest modulation format allowed by the current channel state. By maximizing SE, adaptive transceivers allow to handle more data at a given total power consumption, improving the energy efficiency. Moreover, an adaptive optical transmitting power control has been deployed to reduce the optical power injected to the system, therefore, to reduce the nonlinearity of optical components. In this work, a limited feedback Channel State Information (CSI) is utilized to reduce the complexity of the control algorithm. We formulate the design of an adaptation algorithm as an optimization of average transmitting power to guaranty a targeted Bit Error Rate (BER) under peak power constraints. We have adopted a strategy that adapts the SE, i.e., the number of information bits sent in each polarization per symbol period, and the power consumption by adjusting the FEC code rate, the underlying modulation format and the optical transmission power in accordance with the variation of the channel conditions. The bit rate can be maximized while ensuring reliable transmission of information. In the literature, we can find many adaptive transmission techniques. Adaptive modulation for OFDM systems was studied recently in [7] and [8], which propose an algorithm that uses the average value of the instantaneous SNR of the subcarriers in the sub-band as the switching parameter. In [9], three different modulation mode allocation algorithms were discussed and compared. Further based on the application of turbo code in adaptive modulation and coding is conducted in [10], which contains an optimal approach based on prediction of the average BER over all subcarriers. In this work, we propose a flexible OFDM-RoF transceiver, with hybrid link adaption methods combining adaptive modulation, coding and optical transmitting power. The purpose of this adaptive transceiver is to provide users the best compromise between several factors, such as spectral efficiency, robustness against transmission errors, and the power consumption. This paper is organized as follows: The Sect. 2 contains a description of the system model, the procedure and adaptive switching thresholds used in the simulations. In the Sect. 3 we discuss the simulation results. Finally, conclusions are made in Sect. 4.

2

System Model

The model of the proposed system is shown in Fig. 1. This system uses LowDensity Parity-Check (LDPC) codes as forward error-correction channel coding and adaptive coded modulation technology based on CSI. The adaptive coded modulation schemes used in this system are selected according to the physical layer standard IEEE 802.15.3c high speed interface (HSI) OFDM developed in [11,12,14], as shown in Table 1. The generated data enters the channel encoder. The coding scheme adapted to the transmission is chosen according to the CSI. The transmitter receives the

Adaptive Coded Modulation and Power Control

17

CSI and chooses a suitable Modulation and Coding Scheme (MCS) according to a given set of thresholds in order to keep the BER below a certain fixed level BER0. The modulated baseband signals are then transmitted over the RF and optical channel (fading channel and 300 m of Multi-Mode Fiber (MMF)). Using the CSI information, obtained by estimation technique we presented in our work [12], the demodulator then recovers the components of the baseband signals. Finally the channel decoder restores the transmitted data.

Fig. 1. Block diagram of the adaptive coded modulation system

Denoting the uncoded and coded symbol rates by Ru and Rs , respectively, their relationship is expressed by Rs = Ru /rc symbols/s, where rc = be /b is the code rate of the employed FEC code. In order to keep the occupied bandwidth constant, the number of symbols per frame should be fixed. Consequently, the maximum uncoded symbol rate for transmission at Rs symbols/s with a 2be -ary constellation and FEC code rate of rc is given by Ru = (be /b)Rs symbols/s. The SE of the proposed coded transmission is thus given by η = 2rc be bits/symbol whereas that of the uncoded transmission is 2b bits/symbol. Table 1. HSI PHY MCS used dependent parameters MCS index Data rate (Mb/s) Modulation scheme FEC rate 1

1540

QPSK

1/2

2

2310

QPSK

3/4

3

3080

16-QAM

1/2

4

4620

16-QAM

3/4

5

1925

QPSK

1/2 3/4

Our adaptive algorithm refers to switch between different coded modulations schemes with different SNR threshold values. In this work, we use threshold value

18

M. A. Azza et al.

determination criteria based on system reliability requirements, since each modulation scheme and coding has different performance in terms of BER. For a transmission scheme, when channel condition becomes severe (Low SNR), the transmission BER will increase, and the system transmission is hindered. Therefore, different threshold criteria are set up to hold a certain bit error rate BER0 and determine the decision thresholds for various coded modulation schemes. When BER of a channel becomes lower than the threshold value, the corresponding coded modulation method will no longer be used for system transmission. The BER0 is set at 10−4 in this paper for switching thresholds optimization of various coded modulation schemes. The SNR range is divided into N + 1 intervals. (N is the number of modulation and coding schemes used by the ACM method, in our case N = 5). MCSn is selected as the coded modulation scheme for transmitting the next data frame. Figure 2 shows the BER performance of five coded modulation schemes over AWGN channel. The discrete points represent the simulated points, while the solid lines are fitted lines. At the intersection of the curves with the BER threshold (of 10−4 ), we get the switching SNR threshold values for various coded modulation schemes, as shown in Table 2. In the context of optical fiber communication, the optical power that is launched into the fiber can be adjusted through four different mechanisms: changing the optical power of the laser, changing the electrical swing of the modulators, changing the gain of the optical booster amplifier that follows the transmitter, or using a variable optical attenuator (VOA) [13]. Here, we introduce two adaptive algorithms to reduce

Fig. 2. BER Performance of different MCS in AWGN channel

Adaptive Coded Modulation and Power Control

19

Fig. 3. Simulation setup of the optical fiber part

average power consumption of the RoF system by adjusting optical power of the laser (Fig. 3). The ACM control block has the function to detect the crossing of decision thresholds for determining the method of modulation and coding to be used in the next frames sent by returning this information not only to the transmitter but also to the receiver.

3

Simulation Results and Discussion

In these simulations, the performance of adaptive modulation is investigated in terms of throughput and BER performance. To shed light on adaptive transceiver advantages compared to fixed modulation and coding transceivers. The studied usage model corresponds to the transfer of data and files between computer network components, for example in typical office environments. Table 2. HSI PHY MCS used dependent parameters MCS index Modulation scheme FEC rate SNR Thresholds 1

QPSK

1/2

4dB < SNR < 7dB

2

16-QAM

1/2

7dB < SNR < 22.6dB

3

64-QAM

5/8

SNR > 22.6dB

In terms of BER performance, the used adaptive modulation and coding technique was better than the most robust fixed modulation and coding mode systems. As shown in Fig. 5, the BER level was almost constant, around a threshold value of 10−4 since the scheme is designed to meet the BER target of 10−4 . The variation of SNR at the reception due to Channel SNR variation, causes a modification in the MCS to be used, hence a throughput variation as shown in Fig. 4. It can be noticed that for high SNR values, we get a high throughput. This is due to the use of a high modulation order. Adaptive modulation and coding system provides the best throughput without constant BER performance. It is also observed that the throughput achieved by ACM system is the highest compared to fixed modulation system. Therefore

20

M. A. Azza et al.

Fig. 4. BER Performance of adaptive modulation and coding with RoF architecture

it’s clear that ACM techniques allows to achieve a given target level of BER by choosing the appropriate modulation and coding scheme for the instantaneous channel conditions.

(a)

(b)

Fig. 5. BER (a) and throughtput (b) performance of ACM method with RoF architecture

Figure 6 illustrates the desired BER versus the average required optical transmitting power to guaranty this BER. As we can see, when we use 64QAM modulation, the received optical power, for 20% of EVM, is −2dB while is equal to −10dB in case of QPSK modulation. So, there is no need to transmit high optical power when we use low MCS order. Moreover, in order to avoid to work in nonlinearity regime of optical components and to reduce optical power consumption, we have adopted the injected power as a function of the used MCS. The power consumption depends on two main components: a constant component that is independent of the bit rate and a component that changes in proportion to bit rate [13]. The first is determined by the total leakage and also by the circuits that

Adaptive Coded Modulation and Power Control

21

Fig. 6. Variation of the received optical power in function of EVM and modulation schemes

consume static power (e.g., current- mode logic used in high-speed I/Os). The second part of the power consumption depends on the net bit rate of the system (i.e., information rate or throughput), the energy consumption (per bit) of operations performed before and after FEC decoding and the energy consumption of operations such as encoding/decoding. Therefore the adaptation “change” of the coding scheme and modulation that we have implemented will clearly cause a variation in power consumption. This makes the adjustment of the transmission power necessary even primordial.

4

Conclusion

In the present paper, a solution of adaptation in terms of modulation and coding for OFDM-RoF system has been proposed and studied. Different modulation and coding schemes were used with three different types of modulation (QPSK, 16QAM, 64QAM) and variable FEC rate. The system model and simulations were made on Matlab/Simulink. We presented also a power control algorithm designed through solving an optimization problem to minimize average transmitting power required for a targeted BER. The main conclusion drawn from the results obtained that the adaptive telecommunication systems have an advantage over the conventional system with fixed transceivers; this advantage is the spectral and power efficiency and the stability of the transmission quality against the channel variation.

References 1. Razibul Islam, A.H.M., Md. Imrul, H., Ju, B.S.: Adjacent channel power ratio of OFDM signals for broadband convergence networks. In: Joint International Conference on Optical Internet and Next-generation Network, Korea, pp. 180–182, July 2006

22

M. A. Azza et al.

2. Zhang, C., Duan, J., Li, J., Hu, W., Li, H., Wuand, H., Chen, Z.: Bidirectional 60-GHz radio-over-fiber systems with downstream OFDMA and wavelength reuse upstream SC-FDMA. Opt Express 18, 19429–19437 (2010) 3. Yu, J., Hu, J., Qian, D., Jia, Z., Chang, G.K., Wang, T.: Transmission of microwave-photonics generated 16Gbit/s super broadband OFDM signals in radioover-fiber system. In: Optical Fiber Communication Conference/National Fiber Optic Engineers Conference, OSA Technical Digest (CD) (Optical Society of America, 2008), paper OThP2 (2008) 4. El Yahyaoui, M., El Moussati, A., Haddadi, K.: Performance evaluation of 60-GHzWPAN system distributed over multi-mode fiber. Int. J. Electron. Telecommun. 63(4), 381–387 (2017) 5. Arabaci, M., Djordjevic, I.B., Xu, L., Wang, T.: Nonbinary LDPC-coded modulation for rate-adaptive optical fiber communication without bandwidth expansion. IEEE Photonics Technol. Lett. 24(16), 1402–1404 (2012) 6. Azza, M.A., El Moussati, A., Mekaoui, S., Ghoumid, K.: Spectral management for a cognitive radio application with adaptive modulation and coding. Int. J. Microw. Opt. Technol. 9(6), 445–452 (2014) 7. Harivikram, T.S., Harikumar, R., Ganesh Babu, C., Murugamanickam, P.: Adaptive modulation and coding rate for OFDM systems. Int. J. Emerg. Technol. Adv. Eng. 3(2) (2013) 8. Faezah, J., Sabira, K.: Adaptive modulation for OFDM systems. Int. J. Commun. Netw. Inf. Secur. (IJCNIS) 1(2) (2009) 9. Keller, T., Hanzo, L.: Adaptive modulation techniques for duplex OFDM transmission. IEEE Trans. Veh. Technol. 49(5), 1893–1906 (2000) 10. Lei, Y., Burr, A.: Adaptive modulation and code rate for turbo coded OFDM transmissions. In: Vehicular Technology Conference VTC2007, 22–25 April 2007, pp. 2702–2706 (2007) 11. Azza, M.A., El Moussati, A., El Yahyaoui, M.: Adaptive modulation and coding for the IEEE 802.15.3c high speed interface physical layer mode. In: Proceedings of the Mediterranean Conference on Information and Communication Technologies, pp. 365–371 (2015) 12. Azza, M.A., El Yahyaoui, M., El Moussati, A.: Throughput performance of adaptive modulation and coding schemes for WPAN transceiver. In: International Symposium on Advanced Electrical and Communication Technologies (ISAECT), Rabat, Morocco, pp. 1–4 (2018) 13. Sedighi, B., Khodakarami, H., Pillai, B.S.G., Shieh, W.: Power-efficiency considerations for adaptive long-haul optical transceivers. IEEE/OSA J. Opt. Commun. Netw. 6(12), 1093–1103 (2014) 14. Zarrouk, T., El Moussati, A., El Yakyaoui, M., El Oualkadi, A.: On 60 GHz wireless link performance in indoor and outdoor environments based on IEEE 802.15.3c. In: 2015 4th International Conference on Electrical Engineering (ICEE), Boumerdes, pp. 1–5 (2015)

Optical Architecture for 60 GHz 4
4 MIMO Signal Distribution over Optical Fiber Moussa El Yahyaoui(&), Hachim Azzahhafi, and Ali El Moussati Department of Electronics, Informatics and Telecommunications, National School of Applied Sciences Oujda, Oujda, Morocco [email protected], [email protected], [email protected]

Abstract. This paper proposes a solution to distribute 4
4 MIMO (Multiple Input Multiple Output) over optical fiber at 60 GHz. This solution is based on the subcarrier multiplexing and OCS (Optical Carrier Suppression) modulation. The 60 GHz signal is generated by using two parallel modulators MZM. The First modulator is set to realize the OCS modulation and the subcarrier multiplexing, while the second MZM is set to generate two frequencies of 24 GHz and 48 GHz by heterodyne at reception. This system is realized in OptiSystem software. The results show a 40 Gb/s of data rate transported through 25 km SMF (Single Mode Fiber). Keywords: MIMO

 RoF  Millimeter wave  60 GHz  OCS

1 Introduction The number of users is still growing fast as well as the telecommunication services which consume more bandwidth. To serve these users, the solution is to add more bandwidth by using the Millimeter Wave (mmWave) band, which in past was unused due to lack of advanced technology. This band becomes, recently, a promising solution to install new standard and to deliver high data rate. Currently 5G considered to work in this band. A lot of other standard work in this band, we can list 802.15.3c, 802.11ad [1, 2]. Moreover, to increase the spectral efficiency of these technologies, MIMO technique is the most promising solution that largely used. Combining MIMO with RoF allows the transport of multiple signals at mmWave with minimum attenuation and reduces the complexity and cost of base station. There are a lot of researchers interested by the distribution of MIMO signals over fiber [3, 4]. In our previous work [5], we have proposed a novel architecture to transport 2
2 MIMO at 60 GHz. In this work, we propose a solution to distribute 4
4 MIMO over fiber at 60 GHz. This solution is based on OCS modulation and subcarriers multiplexing which are performed by Dual-Parallel Mach-Zehnder Modulator (DP-MZM) and Dual-Drive Mach-Zehnder Modulator (DD-MZM). This paper is organized as follow; we present the proposed architecture in the second section and the theoretical principle in the third section. In the fourth section we present and discuss the obtained results. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 23–29, 2020. https://doi.org/10.1007/978-3-030-53187-4_3

24

M. El Yahyaoui et al.

2 Model Principle The proposed 2
2 MIMO-RoF architecture is presented in Fig. 1. The signals, Ij  data and Qj  data (with j = 1, 2, 3, 4) are obtained by mapped binary data into QPSK modulation.

Fig. 1. Schematic principle of the proposed system

The principle of the used RoF downlink, with an optical carrier (fc ) transmitting four RF signals (RF1 , RF2 ; RF3 ; and RF4 ), is based on two dual parallel MZM (DPMZM) and dual drive MZM (DD-MZM). The baseband signals are up-converted electrically to fLO1 ¼ 12 GHz and fLO2 ¼ 36 GHz using I/Q mixers. The converted signals are coupled through 90° hybrid couplers. DP-MZM1 and DP-MZM2 are polarized to there minimum transmission bias point to realize optical carrier suppression (OCS) modulation. These MZMs are driven by the combined signals to modulate the optical carrier and to multiplex the subcarriers as depicted in inset (b) [6]. In order to generate 60 GHz from optical heterodyne in photo-detector at reception, a DD-MZM, with an oscillator of 24 GHz, is used to generate four subcarriers (inset d), to be combined with the multiplexed RF subcarriers (inset c). The multiplexed subcarriers, depicted in inset (e), are transported from central station to base station through a single mode optical fiber. The received optical signal at base station is filtered to separate the four RF signals of 4
4 MIMO signal.

Optical Architecture for 60 GHz 4
4 MIMO Signal Distribution

25

3 Theoretical Principle The optical signal generated from the optical modulator DD-MZM (inset (d) of Fig. 1), has the optical field Eout ðtÞ defined as Eout ðtÞ ¼ Ein ðtÞexp½jbcosðxm t þ uÞ þ exp½jbcosxm t þ j/0 

ð1Þ

where, • Ein ðtÞ ¼ Ec expðjxc tÞ is the optical carrier, j2 ¼ 1, Ec represents the amplitude, xc is the angular frequency of the optical carrier ðxc ¼ 2pfc Þ; • b is the modulation index of the DD-MZM; • Du is the phase difference of RF signal applied to two arms of DD-MZM; • /0 is the phase obtained by adjusting the DC bias of DD-MZM; • xm is the angular frequency for the RF signal. From Jacobi-Anger expansion [7], Eq. (1) has the form of Eout ðtÞ ¼ Ec

þ1 X

jn Jn ðbÞexp½jðxc þ nxm ÞtexpðjnDuÞ þ expðj/0 Þ

ð2Þ

n¼1

where Jn ðbÞ is the nth Bessel function of the first kind. We list the optical fields the sidebands of order 3rd , 2nd , 1st , 0th , þ 1st , þ 2nd , and þ 3rd because they of the most importance for our modulation:   3p E3 ¼ Ec exp j ½J3 ðb1 Þexpðj3uÞ þ J3 ðb2 Þexpðj/0 Þexp½jðxc t  3xe tÞ 2 E2 ¼ Ec expðjpÞ½J2 ðb1 Þexpðj2uÞ þ J2 ðb2 Þexpðj/0 Þexp½jðxc t  2xe tÞ  p E1 ¼ Ec exp j ½J1 ðb1 ÞexpðjuÞ þ J1 ðb2 Þexpðj/0 Þexp½jðxc t  xe tÞ 2 E0 ¼ Ec ½expJ0 ðb1 Þ þ J0 ðb2 Þexpðj/0 Þexpðjxc tÞ  p E þ 1 ¼ Ec exp j ½J1 ðb1 ÞexpðjuÞ þ J1 ðb2 Þexpðj/0 Þexp½jðxc t  xe tÞ 2 E þ 2 ¼ Ec expðjpÞ½J2 ðb1 Þexpðj2uÞ þ J2 ðb2 Þexpðj/0 Þexp½jðxc t  2xe tÞ  p E þ 3 ¼ Ec exp j ½J3 ðb1 Þexpðj3uÞ þ J3 ðb2 Þexpðj/0 Þexp½jðxc t  3xe tÞ 2

ð3Þ

26

M. El Yahyaoui et al.

The optical power of these sidebands are:   P3 ¼ Ec2 J32 ðb1 Þ þ J32 ðb2 Þ þ 2J3 ðb1 ÞJ3 ðb2 Þcosð3u þ /0 Þ   P2 ¼ Ec2 J22 ðb1 Þ þ J22 ðb2 Þ þ 2J2 ðb1 ÞJ2 ðb2 Þcosð2u þ /0 Þ   P1 ¼ Ec2 J12 ðb1 Þ þ J12 ðb2 Þ þ 2J1 ðb1 ÞJ1 ðb2 Þcosðu þ /0 Þ   P0 ¼ Ec2 J02 ðb1 Þ þ J02 ðb2 Þ þ 2J0 ðb1 ÞJ0 ðb2 Þcosð/0 Þ   P þ 1 ¼ Ec2 J12 ðb1 Þ þ J12 ðb2 Þ þ 2J1 ðb1 ÞJ1 ðb2 Þcosðu  /0 Þ   P þ 2 ¼ Ec2 J22 ðb1 Þ þ J22 ðb2 Þ þ 2J2 ðb1 ÞJ2 ðb2 Þcosð2u  /0 Þ   P þ 3 ¼ Ec2 J32 ðb1 Þ þ J32 ðb2 Þ þ 2J3 ðb1 ÞJ3 ðb2 Þcosð3u  /0 Þ

ð4Þ

In order to realize the OCS modulation with the 3rd subcarriers removed (3rd , þ 3rd order sidebands), the conditions below should be verified cosð/0 Þ ¼ 1 cosð3u  /0 Þ ¼ 1 cosð3u þ /0 Þ ¼ 1

ð5Þ

b1 ¼ b2 This correspond to b1 ¼ b2 , u ¼ 120ðmod2pÞ, and /0 ¼ pðmod2pÞ.

4 Results and Discussion The realization of the proposed model is performed in OptiSystem software. The results obtained after simulations are presented in this section. The optical spectrum of 4
4 MIMO signal is shown in. Fig. 2. This signal is obtained by combining the multiplexed subcarriers (Fig. 4) and the generated optical subcarriers (Fig. 3).

Fig. 2. Combination of multiplexed 4
4 MIMO subcarriers and generated subcarriers

Optical Architecture for 60 GHz 4
4 MIMO Signal Distribution

27

Fig. 3. Generated subcarriers with DD-MZM

Fig. 4. Multiplexed 4
4 MIMO subcarriers

The eye diagram and BER at reception for 4
4 MIMO at 60 GHz with 25 km SMF for data rates 40, 32, 24, 16 Gb/s are shown in Fig. 5 and Fig. 6. As we can see the three data rates have BER below 10−7 and a good eye opening diagrams. Figure 7 shows the constellation diagram of received signal after demodulation with data rate of 16 Gb/s and SMF length of 25 km. These results show that our architecture is capable to transport 4
4 MIMO signal with minimum degradation. This architecture has advantage of using simple components such as modulator MZM and optical filters. The perspective of this work is to increase the MIMO order e.g. 8
8 MIMO and 16
16 MIMO.

28

M. El Yahyaoui et al.

Fig. 5. Eye diagrams at reception of 40 Gb/s and 32 Gb/s data rate

Fig. 6. Eye diagrams at reception of 24 Gb/s and 16 Gb/s data rate

Optical Architecture for 60 GHz 4
4 MIMO Signal Distribution

29

Fig. 7. Constellation diagram of received signal

5 Conclusion In this work we have proposed a novel optical architecture to distribute four 60 GHz signals of 4
4 MIMO system. We have modeled and simulated this system in OptiSystem. The results obtained in this work show that the proposed architecture can transport the 4
4 MIMO signal at 40 Gb/s through 25 km of SMF.

References 1. Bhusal, R., Moh, S.: Qualitative and quantitative comparison of IEEE 802.15.3c and IEEE 802.11ad for multi-Gbps local communications. Wirel. Pers. Commun. 75, 2135–2149 (2014) 2. El Yahyaoui, M., El Moussati, A., Ghoumid, K., Al-Mahdawi, B., Lepers, C.: IEEE 802.15.3C transmission over multimode fiber links: performance comparison of RF and IF over fiber architectures. Int. J. Microw. Opt. Technol. 11(5), 384–390 (2016) 3. Elmagzoub, M.A., Bakar Mohammad, A., Shaddad, R.Q., Al-Gailani, S.A.: Polarization multiplexing of two MIMO RoF signals and one baseband signal over a single wavelength. Opt. Laser Technol. 76, 70–78 (2016) 4. Huang, H., et al.: Direct-detection PDM-OFDM RoF system for 60-GHz 2
2 MIMO wireless transmission without polarization tracking. J. Lightwave Technol. 36, 3739–3745 (2018) 5. El Yahyaoui, M., El Moussati, A., El Zein, G., Haddadi, K.: New millimeter wave generation scheme for MIMO-OFDM based radio-over-fiber system. Opt. Commun. 442, 101–105 (2019) 6. El Yahyaoui, M., El Moussati, A., El Zein, G.: On the capacity of MIMO-OFDM based diversity and spatial multiplexing in radio-over-fiber system. Opt. Commun. 402, 252–259 (2017) 7. Zheng, Z., et al.: Optical single sideband millimeter-wave signal generation and transmission using 120° hybrid coupler. Opt. Commun. 411, 21–26 (2018)

Evaluation of Railway Communications System Based on 5G-RoF Technology and Millimeter Wave Band Hachim Azzahhafi(&), Moussa El Yahyaoui, and Ali El Moussati Department of Electronics, Informatics and Telecommunications, ENSAO, University Mohammed Premier, Oujda, Morocco [email protected], [email protected], [email protected]

Abstract. In this paper, we present an evaluation of railway communications system based on 5G-RoF technology and millimeter wave band. Filter Bank MultiCarrier (FBMC) modulation is used in this work to generate the 5G waveform, and Radio over Fiber (RoF) technology is used to transport the Radio Frequency (RF) signal from Central Station (CS) to Base Stations (BSs). RoF system is based on Single Side Band (SSB) and Optical Carrier Suppression (OCS) modulations. For railway environment, we have considered the mobile radio channel taking into consideration a high velocity of receiver (train), which is characterized by Doppler effect. We have evaluated, by simulation, the performance of this system in terms of Bit Error Rate (BER) and Error Vector Magnitude (EVM). The simulations show that the length of optical link decreases by increasing the velocity of receiver, and the maximum length of railway track, that can be covered by each CS, is approximately 76 km for a velocity up to 300 km/h. Keywords: 5G


FBMC
Millimeter wave
RoF
Railway communications

1 Introduction Communication systems in the railway sector have undergone a significant evolution in recent years. Several technologies are deployed to ensure passenger safety, traffic management and signage. However, these technologies remain limited and cannot support broadband communications, which is necessary to provide high data-rate communications for passengers on the train. The new generation of wireless communication, 5G is destined to increase the data rate of each user, in order of several Gb/s and to support different usage scenarios and applications including high mobility scenario [1]. Indeed, to achieve high data rate, 5G uses new modulation technologies such FBMC and exploits the millimeter wave band. FBMC is a promising multicarrier modulation waveform contender for air interface of 5G in the railway environment [2, 3]. This waveform is based on filtering approach using a bank of modulated filters, where each subcarrier is filtered by a frequencyshifted version of a prototype filter designed to reduce the side lobes in frequency

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 30–37, 2020. https://doi.org/10.1007/978-3-030-53187-4_4

Evaluation of Railway Communications System

31

domain, and consequently to provide a low level of Out-Of-Band (OOB) power leakage [4]. The millimeter wave band offers a large free bandwidth up to 9 GHz [5]. Moreover, 5G systems support RoF technology in order to distribute Radio Frequency (RF) signals, from Central Station (CS) to Base Stations (BSs), over optical fiber. An experimental demonstration of multi bands transmission in millimeter wave radio over a fiber system is proposed in [6]. In [7] and [8], the authors compared the performances FBMC and Orthogonal Frequency-Division Multiplexing (OFDM) signals at millimeter wave frequencies. In this paper, we use a new scheme, proposed in our previous work [9], to generate and transmit FBMC millimeter wave signal with RoF system, using SSB and OCS optical modulation techniques, to support railway communication. We have considered the railway environment presented in Fig. 1, which is characterized by mobile radio channel with Doppler Effect. This system is simulated using a co-simulation between Matlab and OptiSystem [10], and the performances are evaluated in terms of BER and optical fiber length.

5G Network

... ...

...

CS

BS

BS

...

...

CS

BS

BS

...

... Fig. 1. 5G-RoF railway communications system.

The rest of this paper is organized as follow: in second section, we present FBMCRoF model, in third section, we show the simulation setup, in the fourth section, we discuss the obtained results, and in the fifth section, we summarize and draw some conclusions.

32

H. Azzahhafi et al.

2 FBMC-RoF Model 2.1

FBMC Model

Up conversion

Pulse shaping

P/S conversion

IFFT

Filtering

S/P conversion

OQAM modulation

FBMC model is shown in Fig. 2. The sequence data is modulated using Offset Quadrature Amplitude Modulation (OQAM). The outputs are parallelized before performing Inverse Fast Fourier Transform (IFFT) and filtering process using a filter bank. A parallel to serial conversion is applied to the filtered signal before pulse shaping and up conversion blocks. At the receiver side, the inverse process is performed.

Fig. 2. Functional block diagram of FBMC signal generation.

The parameters used to generate base band FBMC signal with bandwidth of 4.2 GHz, are given in Table 1. In order to transpose this signal to Intermediate Frequency (IF) band, we have used a Local Oscillator (LO) of 10 GHz.

Table 1. Parameters of generated FBMC signal. Parameter Modulation format Bit rate Sampling frequency

Value OQAM 7,168 Gbps 28,672 GS/s

Figure 3 shows the frequency spectrum of the generated IF FBMC signal, which is centred at 10 GHz.

Evaluation of Railway Communications System

33

Fig. 3. Electrical spectrum of FBMC signal at 10 GHz.

2.2

Optical Link

The optical link of RoF System is presented in Fig. 4. The FBMC signal, generated with MATLAB/Simulink, is transposed to IF band at 10 GHz.

Fig. 4. The optical link of RoF system.

The optical laser source generates an optical carrier of 1550 nm. This carrier is modulated with the IF FBMC signal, by using two optical Phase Modulators (PM), which differ in phase by 90°. This phase difference, between two versions of the double

34

H. Azzahhafi et al.

side band of IF FBMC signal, can help to suppress one side band of this signal thanks to the coupler. An OCS modulation is performed in order to obtain 60 GHz signal at reception by heterodyne technique. This kind of modulation is realized by polarizing the dual drive Mach-Zehnder Modulator (MZM) to its minimum intensity and applying two signals, with a phase difference of 180°, to its arms. Finally, a Single Mode Fiber (SMF) is used to transport the optical FBMC signal to the receiver, where a photo detector is used to convert the optical signal to the corresponding electrical signal.

3 Simulation Setup We have simulated the complete system by using two software, Matlab/Simulink and OptiSystem. The baseband model and radio mobile channel are implemented in Matlab/Simulink software and the optical link of RoF system is implemented in OptiSystem software. The architecture of simulated system is depicted in Fig. 5. In order to model the mobile radio channel, we have used Rayleigh channel model and we have added a Doppler effect, which is, corresponds to velocity of 300 km/h at radio frequency of 60 GHz.

Training QAM to

QAM mod

Data

OQAM

S/P

Filtering

IFFT

Overlap and sum

P/S

Optical Link OptiSystem

BER calculation

QAM demod

Channel

OQAM to QAM

S/P

Filtering

Freq domain equalizer

Mobile Radio Channel Matlab/Simulink

FFT

S/P

Windowing

Training

Fig. 5. The complete RoF system implemented in Matlab/Simulink and OptiSystem.

In order to fix the optical power of the source laser, we have calculated EVM for different values of the injected optical power. The results are shown in Fig. 6. As we can see in this figure, the minimum values of EVM are between 12 and 20 dBm. In order to minimize the optical power injected, we have chosen the value of 12 dBm.

Evaluation of Railway Communications System

35

Fig. 6. EVM in function of injected optical power.

4 Results and Discussion Figure 7 shows the obtained signal spectrum after performing SSB and OCS optical modulations. As we can see, the FBMC signal is centered at 60 GHz from the optical carrier.

Fig. 7. Spectrum of the millimeter wave signal at the output of OCS modulator.

36

H. Azzahhafi et al.

In order to evaluate the performance of our system, we have calculated the BER performance in function of the length of optical link with velocity of 300 km/h, these results are presented in the graph of Fig. 8.

Fig. 8. BER performances vs. fiber length.

As we can see in this graph, the proposed system can transmit FBMC signal over an optical fiber for a length up to 38 km, with BER below 10−6.

5 Conclusion In this work, we have generated and transmitted FBMC millimeter wave signal over an optical fiber in order to cover the railway track by using several base stations controlled by few central stations. The optical architecture of RoF system is based on SSB and OCS modulations. Using this architecture, we have evaluated the performance of millimeter wave system for railway communications in terms of BER and EVM for different length of optical fiber. This proposed system can transmit FBMC signal over an optical fiber of length up to 38 km, with BER below 10 6 , for one side of railway track, which is, corresponds to 76 km covered by one central station (Two sides) Fig. 9.

Evaluation of Railway Communications System

37

5G Network

38 km BS

CS

...

38 km BS

... ...

...

CS

BS

BS

...

... ~ 76 km Fig. 9. The maximum railway length covered by each central station.

References 1. ITU-R Rec. M.2083-0: IMT Vision Framework and Overall Objectives of the Future Development of IMT for 2020 and Beyond (2015) 2. Saideh, M., Berbineau, M., Dayoub, I.: 5G waveforms for railway. In: 15th International Conference on ITS Telecommunications (ITST), Warsaw, pp. 1–5 (2017) 3. Azzahhafi, H., Yahyaoui, M.E., Moussati, A.E.: Performance analysis of frequency spreading FBMC in mobile radio channel. In: International Symposium on Advanced Electrical and Communication Technologies (ISAECT), Rabat, pp. 1–3 (2018) 4. Nissel, R., Schwarz, S., Rupp, M.: Filter bank multicarrier modulation schemes for future mobile communications. IEEE J. Sel. Areas Commun. 35(8), 1768–1782 (2017) 5. Yahyaoui, M.E., Moussati, A.E., Haddadi, K.: Performance evaluation of 60-GHz-WPAN system distributed over multi-mode fiber. Int. J. Electron. Telecommun. 63(14), 381–387 (2017) 6. Parajuli, H., Shams, H., Gonzalez, L., Udvary, E., Renaud, C., Mitchell, J.: Experimental demonstration of multi-Gbps multi sub-bands FBMC transmission in mm-wave radio over a fiber system. Opt. Express 26, 7306–7312 (2018) 7. Saljoghe, A., Gutiérrez, F.A., Perry, P., Venkitesh, D., Koipilla, R.D., Barry, L.P.: Experimental comparison of FBMC and OFDM for multiple access uplink PON. J. Lightwave Technol. 35(19), 1595–1604 (2017) 8. Xu, M., Zhang, J., Lu, F., Wang, J., Cheng, L., Cho, H.J., Khalil, M.I., Guidotti, D., Chang, G.K.: FBMC in next-generation mobile fronthaul networks with centralized preequalization. IEEE Photonics Technol. Lett. 28(118), 1912–1915 (2016) 9. Azzahhafi, H., Yahyaoui, M.E., Moussati, A.E.: Generation and transmission of FBMC signal at mmWave over fiber for 5G. In: International Symposium on Advanced Electrical and Communication Technologies (ISAECT), Rabat, pp. 1–4 (2018) 10. Yahyaoui, M.E., Moussati, A.E., Ghoumid, K., Mekaoui, S., Gharbi, T.: Performance evaluation of coherent optical OFDM communications using LDPC codes. Int. J. Microw. Opt. Technol. 11(1), 72–78 (2016)

Digital Video Broadcasting - Satellite - Second Generation Measurement and Test for Database Simulation Youssef Bikrat(&), Khalid Salmi, Ahmad Benlghazi, Abdelhamid Benali, and Driss Moussaid Laboratoire d’Electronique et Systèmes, Faculté des sciences, Université Mohamed Premier, Oujda, Morocco [email protected]

Abstract. Modeling and simulation of communications systems is an efficient and fast way to highlight the performance and the main design difficulties of the latter. The experience and the real measure are still indispensable tools for the validation of the simulation results and its improvement. Our work consists of real experience of satellite transmission/reception. The experience was done in the national society of radio and television, where we measured and compared the transmitted and the received signal by changing the parameters at the input such as power, frequency, type of coding and modulation. In these experiences, we used a GSERTEL tool as measurement system in order to show the performance according to gain and emission power. As result, we have different types of parameters (C/N, MER BCHBER, LDPCBER, EIRP, Constellation and Link margin). In this paper, we present the main characteristics of satellite transmission/ reception and the description of modulation, demodulation, and the encoding types. These characteristics have been experimented and measured by doing a transmission with different frequencies within the National Broadcasting and Television Company, where we have transmitter and receive some data with different parameters. Finally, obtained results are evaluated, compared and discussed. Keywords: Digital video broadcasting
DVB-S
DVB-S2 FEC
Link margin
BCH
LDPC
Constellation diagram


BER
MER


1 Introduction Satellite communication systems have become an important and complementary part of the 3rd and 4th generation telecommunication systems, with numerous applications such as telephony [1], monitoring [2, 3], navigation, and multimedia broadcasting [4]. Especially for broadcasting applications there is a worldwide deployment of digital video broadcasting (DVB) systems over satellite (DVB-S) and its later standard versions, such as DVB-S2, DVB satellite to-hand-held (DVB-SH) [5, 6].

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 38–46, 2020. https://doi.org/10.1007/978-3-030-53187-4_5

Digital Video Broadcasting

39

The huge demand of application such as internet and HDTV makes the development of satellite communication very important, high quality and quantity are required in present generation and next generation communication system. The second-generation digital video broadcasting for satellites (DVB-S2) adopts the amplitude phase shift keying modulation for enhanced performance over nonlinear channels. One of the Amplitude Phase Shift Keying (APSK) represents an attractive modulation scheme for digital transmission over nonlinear satellite channels due to its power and spectral efficiency combined with its inherent robustness against nonlinear distortion. APSK modulation scheme has lower peak average power ratio (PAPR) than QAM modulation scheme because the constellation of APSK modulation scheme has circular structure. Over the last years, much research has been carried out relate to satellite transmission, DVB-S and DVB-S2 Broadcasting. For example Perica et al. [7], designed a simulation model of DVB-S2 system, which was implemented in Simulink Matlab. Labsky and Kratochvil [8], evaluated and compared an experimental measurement of real DVB-S and DVB-S2 signal. JongKeun and Daelg [9], checked the performance evolution of DVB-S2 according to roll off factor. Junyu et al. [10], proposed a hybrid multicast scheme for the next generation satellite TV system. Kresimir et al. [11], performed a measurement of DVB-S and DVB-S2 parameters in different weather situations. To develop more satellite transmission, we propose in this paper to evaluate, compare and discuss the results of some experimental tests and measurement in order to design an optimal simulation of an adequate satellite transmission models. The paper is organized as follows: first chapter deals with DVB-S2 signal basic attributes, with new techniques used to transmit and processing and with typical applications. In the next chapter are itemized equipment used to measurement, and describe our experience. The chapter four shows the results obtained and the interpretation, which are graphically expressed, compared and discussed.

2 DVB-S2 (Digital Video Broadcasting - Satellite - Second Generation) Overview The specification of the DVB-S2 standard revolves around three key concepts: the best transmission performance, total flexibility and reasonable complexity of the receiver. To achieve the right balance of performance and flexibility, translating into a 30% increase in capacity over DVB-S, the DVB-S2 incorporates the latest advances in channel modulation and coding [12]. The DVB-S standard only defined QPSK modulation for the distribution of satellite broadcasting and data broadcasting, which imposed a limit on applications operating with larger antennas and lower symbol rates. The professional infrastructure that was already in place required higher bit rates and was able to accommodate more advanced modulation schemes with higher thresholds, which is why four modulation schemes are proposed by DVB-S2, these schemes are given by QPSK, 8PSK, 16 APSK and 32 APSK modulations. The QPSK and 8-PSK modes are used in broadcasting applications because these modulations are characterized by a substantially constant envelope and can be used in quasi-saturation non-linear satellite transponders. The 16-APSK and

40

Y. Bikrat et al.

32-APSK modes, which are primarily focused on professional applications, can also be used for broadcasting, but require a higher level of available C/N ratio. DVB-S2 is so flexible that it is compatible with all the features of satellite transponders, as it offers a wide range of spectral efficiency and carrier-to-noise (C/N) configurations. In addition, it is not limited to MPEG-2 video and audio coding, but is designed to handle a variety of advanced audio and video formats that the DVB project currently defines. The DVBS-2 supports any incoming format, including single or multiple MPEG transport streams [13]. The new, more powerful FEC (Forward Error Correction) system is based on concatenated BCH and LDPC codes. The performance of the LDPC internal coding is within 1 dB of the theoretical maximum performance of the Shannon limit, which is equivalent to an improvement of the threshold of 2 to 3 dB compared to the DVB-S standard for a flow of information given. DVB-S2 benefits from developments that are more recent and has the following key technical characteristics: Modulation Modes – There are four modes available, with QPSK and 8PSK intended for broadcast applications two higher-order modulation modes, 16APSK and 32APSK can be used. The 16APSK and 32APSK requiring a higher level of C/N are mainly targeted at professional applications. As it is shown in the Fig. 1, the constellation points for 16APSK and 32APSK reside on circles, which provide compensation for transponder non-linearity,

Fig. 1. DVB-S2 modulation modes [14]

• ACM (Adaptive Coding and Modulation) – Allows the transmission parameters to be changed on a frame by frame basis depending on the particular conditions of the delivery path for each individual user, • Forward Error Correction (FEC) – DVB-S2 uses a very powerful FEC, a key factor in allowing the achievement of excellent performance in the presence of high levels

Digital Video Broadcasting

41

of noise and interference. The FEC system is based on concatenation of BCH (Bose Chaudhuri-Hocquengham) with LDPC (Low Density Parity Check) inner coding.

3 Equipment Used for Measurement and Experience In this experience, we will measure several parameters from transmitting data (audio video) on several satellite transmission channels, which imposes the use of several frequencies. At this point, we will use three frequencies to transmit our data to the satellite. At the reception, we will retrieve our data on precise frequencies by measuring our parameters (Level power, C/N, BER, LM). Each time we change the input parameters to visualize the change of parameters at the output to extract the maximum information expected. The parameters recovered are (reception power, C/N, MER, BHCBER, LDPCBER and Link Margin) (Fig. 2).

Fig. 2. Schematic of DVB-S2 transmission/reception blocks

The satellite transmission frequency equals the base frequency that is at the L band plus the oscillation frequency added by up-converter to transfer it to Ku band. The receive frequency is equal to the frequency of the local oscillator LNB plus the frequency of the modulator. The frequencies used for transmission are: F1 = 13881 MHz = 1081 MHz + 12800 MHz. F2 = 13888 MHz = 1088 MHz + 12800 MHz. F3 = 13895 MHz = 1095 MHz + 12800 MHz. The frequencies used for reception are: f1 = 11331 MHz = 1581 MHz + 9750 MHz. f2 = 11338 MHz = 1587 MHz + 9750 MHz. f3 = 11345 MHz = 1595 MHz + 9750 MHz.

42

Y. Bikrat et al.

The experience was done in the national society of radio and television (SNRT). In order to carry out our experiment respecting the DVBs2 standard we used the following equipment: Transmission equipment • • • •

Multiplexer Harmonics Harmonic Ellipse 3000 [15]. Modulator Advantech’s AMT 75 High Speed Broadcast Modem [16]. Parabolic antenna 180 cm dimeter. LNB, 3 W Ku Band BUC NJRC NJT8302F [17] (RF Frequency: 13.75 to 14.5 GHz, IF Frequency: 950 to 1,700 MHz, LO Frequency: 12.8 GHz).

Reception equipment • Measure system GSERTEL Hexylon [18] is a new high-performance portable meter intended for advanced features and high measurement accuracy. • Parabolic antenna 240 cm dimeter. • LNB (Frequency: 950 to 1,700 MHz) (Fig. 3).

Fig. 3. Equipment used for Transmission, reception and measurement

4 Result of Measurements The experience was done in the national society of radio and television (SNRT) (Latitude = 34.6631815753594; Longitude = −1.908102035522461) on May 22th, 2019. In which we have got the following results below (Table 1).

Digital Video Broadcasting

43

Table 1. Measurement values for three channels. Frequency Em/Rc

Transmission power

Modulation Power dBuV

C/N dB

MER dB

LDPC BER

BCH BER

Link Margin dB

Channel 1 Fe = 13881 MHz Fr = 11331 MHz FEC = 3/4

−10 dB

8PSK

−15 dB

8PSK

−20 dB

8PSK

−10 dB

8PSK

−15 dB

8PSK

−7 dB

8PSK

−10 dB

8PSK

−15 dB

8PSK

−20 dB

8PSK

13.3 13.2 13.2 12.5 12.6 12.6 11.9 12 12 14.8 14.5 14.5 14.1 13.7 13.7 14.7 14.6 14.6 15.7 15.6 15.5 15.2 15.4 15.3 15.4 15.3 15.2

13.5 13.4 13.7 11.5 11.7 11.3 7.9 7.9 8.3 14.3 14.4 14.3 12.2 12.3 11.8 14.6 14.3 14.2 13.9 14.4 14.5 12 12.1 11.7 8.6 8.8 8.4

5.3 10−4 6.2 10−4 6.2 10−4 4.8 10−4 5.7 10−4 4.3 10−5 9.6 10−3 1 10−2 5 10−3 2.6 10−5 3.1 10−5 5.3 10−5 3.4 10−4 3.8 10−4 3 10−4 3.5 10−5 1.7 10−5 5.1 10−5 3.5 10−5 3.5 10−5 3.1 10−5 2.1 10−4 3.4 10−4 4.3 10−4 8.7 10−3 5 10−3 4.9 10−3

10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8 10−8

5.4 5.2 5.5 3.4 3.5 3.1 −0.2 −0.2 0.1 6.1 6.2 6.1 4 4.5 3.6 6.5 6.1 6.0 5.7 6.2 5.9 3.9 4.0 3.5 0.6 0.5 0.4

Channel 2 Fe = 13888 MHz Fr = 11338 MHz FEC = 3/4

Channel 3 Fe = 13895 MHz Fr = 11345 MHz FEC = 3/4

69.6 69.6 69.6 68.8 68.7 68.7 67.9 68 68 71 71 71 69.5 69 69 70.1 70.2 70.2 72.6 72.6 72.7 72.3 72.2 72.2 72.0 71.8 71.9

5 Interpretation The first measurement was related to the difference in link margin, carrier to noise and modulation error ratio (MER) for powers over −10 dB in canal 2. As shown in Fig. 4.a, the link margin for −7 dB is approximately equal to the link margin for −10 dB, which is also the same for powers over −7 dB. The same note could be said for MER and C/N. Since there are minimum differences in response for powers over −10 dB and in order to preserve power, we chose to start measurements from −10 dB to −20 dB. The use of greater power may cause a risk of antennas damage.

44

Y. Bikrat et al.

Fig. 4. Graphs of measured Link Margin (a), MER (b) and C/N ratio (c) of signal level

Figure 4.b shows the difference in C/N between canals. The higher frequency gives a better response than lower frequency. Channel 3 shows the best C/N. In order to quantify the performance of the digital radio transmitter and receiver in our communications system using 8QPSK digital modulation, we used the Modulation Error Ratio (MER). Figure 4.c shows the MER values for canal 3 with powers between −10 dB and −20 dB. As we can see in the graph, the higher values of MER correspond to the higher power, which is completely normal since MER is calculated from the average power of the signal. MER is also closely related to signal-to-noise ratio. MER includes all imperfections including deterministic amplitude imbalance, quadrature error and distortion. We also note that the adopted channel coding is a concatenation of a block code (of the BCH type) and an LDPC code with an iterative decoding process. In all channels and for all transmission powers, the BER is always lower than 10−8 once both codes are applied.

Digital Video Broadcasting

45

6 Conclusion In this paper, the basic technical features and principals of DVB-S2 system are described. Measurement of DVB-S2 transmission satellite was done by using three channel by which we have transmitted signal with different parameters. Results were analyzed and discussed to show the optimal parameters values, in order to improve the transmission. As shown in the interpretation part we concluded that the choice of the frequency and then emission power are very important to improving the transmission. The next future step will include a simulation for video signal transmission based on this experience of satellite transmission including a proposal of optimal combination of all parameters.

References 1. Fischer, W.: Digital Video and Audio Broadcasting Technology: A Practical Engineering Guide. Springer, Heidelberg (2010) 2. Bikrat, Y., Moussaid, D., Benali, A., Benlghazi, A.: Electronic and computer system for monitoring a photovoltaic station. In: ISCV (2018) 3. Bikrat, Y., Salmi, K., Benlghazi, A., Benali, A., Moussaid, D.: A photovoltaic wireless monitoring system. In: 2018 International Symposium on Advanced Electrical and Communication Technologies (ISAECT), pp. 1–5 (2018) 4. Montalban, J., et al.: Multimedia multicast services in 5G networks: subgrouping and nonorthogonal multiple access techniques. IEEE Commun. Mag. 56(3), 91–95 (2018) 5. Maini, A.K., Agrawal, V.: Satellite Technology: Principles and Applications, 3rd edn (2014) 6. Maral, G., Bousquet, M.: Satellite communications systems: systems, techniques and technology. Electron. Radio Eng. J. Inst. (2011) 7. Baotic, P., Draganic, M., Bundalo, D., Kesegic, I., Tralic, D., Grgic, S.: Simulation model of DVB-S2 system, pp. 25–27 (2013) 8. Lábsky, B., Kratochvíl, T.: DVB-S/S2 satellite television broadcasting measurement and comparison. In: Proceedings of the 20th International Conference on Radioelektronika 2010, pp. 61–64 (2010) 9. Lee, J., Chang, D.: Performance evaluation of DVB-S2X satellite transmission according to sharp roll off factors. In: International Conference on Advanced Communication Technology, ICACT, pp. 362–365 (2017) 10. Lai, J., Zhao, J., Zhang, W., Li, L., Wu, W.: A hybrid broadcast & multicast scheme for the next generation satellite TV systems. In: 2016 International Conference on Communication Problem-Solving, ICCP 2016 (2016) 11. Malarić, K., Suć, I., Bačić, I.: Measurement of DVB-S and DVB-S2 parameters (2015) 12. Digital Video Broadcasting (DVB); Second generation framing structure, channel coding and modulation systems for Broadcasting, Interactive Services, News Gathering and other broadband satellite applications (DVB-S2) European Broadcasting Union Européenne de Radio-Télévision (2009) 13. dvb.org: What is DVB-S2 (2018) 14. Fischer, W.: Digital Video and Audio Broadcasting Technology (2008) 15. Ellipse Encoders: Ellipse® 3000 Contribution Encoders 16. Modem AMT 75: Forward Error Correction (FEC)

46

Y. Bikrat et al.

17. NJT8302U series | Ku-band BUC | Satellite Communications - VSAT components | Microwave Products | NJR. https://www.njr.com/micro/vsat/ku-buc/universal/3w.html. Accessed 16 June 2019 18. HEXYLON: High measurement accuracy of radio & TV signals – Gsertel. https://www. gsertel.com/hexylon. Accessed 16 June 2019

Design of a Microstrip Patch UWB Antenna with Notch Band Characteristic L. Aguni1(&), S. Chabaa1,2, S. Ibnyaich1, and A. Zeroual1 1

2

Department of Physics, Faculty of Sciences, Cadi Ayyad University, Semlalia, Marrakesh, Morocco [email protected] Industrial Engineering Department, National School of Applied Sciences, Ibn Zohr University, Agadir, Morocco

Abstract. The purpose of this study is to design a patch antenna for ultra wideband (UWB) applications. The antenna consists of a circular patch printed on the FR4 epoxy dielectric substrate. The global dimension of the proposed antenna is 24
14
0.8 mm3. We inserted a split ring in the radiating patch in order to reject the band 5–6 GHz to avoid interferences with other applications working in the same band. To study the performance of this antenna and reinforce the results, the antenna parameters such as reflection coefficient, radiation pattern, voltage standing wave ratio (VSWR), current distribution, and gain have been simulated and analyzed using Ansoft HFSS (High frequency structural simulator) and CST (Microwave Studio MWS). From these results, we can conclude that the designed antenna operates in the frequency band 3.85–12.38 GHz with good radiation characteristic which makes the proposed antenna a better device for UWB technology. Keywords: Patch antenna

 UWB  Notch band

1 Introduction Patch antennas are becoming more and more used in many applications and in several fields especially in the field of telecommunications, because they have several advantages. Among the advantages of the patch antennas, they are less expensive to manufacture from printed circuits and compatible with hybrid circuits and MMIC (Microwave Monolithic Integrated Circuit) circuits. These antennas are also compatible with planar and non-planar surfaces. Another advantage of these antennas is that they can be mounted on any surface due to their low weight and low volume. The ultra-wide bandwidth antennas have attracted the attention of designers and researchers. In the literatures, one can find several works related to the design of antenna for UWB applications. In [1], a design of an UWB antenna with two slits for 5G/WLAN notched bands is presented, an UWB antenna designed using gap loading technology was reported in [2]. As presented in [3], a compact UWB antenna with stop band characteristic is designed. In reference [4], a 10
10 mm2 hook-shaped UWB antenna operating from 3 GHz to 10.7 GHz. However, the designed antennas have © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 47–54, 2020. https://doi.org/10.1007/978-3-030-53187-4_6

48

L. Aguni et al.

some inconvenient, either they have complex geometry or they do not present the notch band characteristic. The objective of this paper is to design a microstrip patch antenna which is low cost and working for UWB technology, with the challenge of rejecting the band 5 GHz– 6 GHz. The organization of this paper is summarized as follows. In Sect. 2, we get into the antenna design, geometrical details of the proposed antenna, and its fundamental parameters. Subsequently, in Sect. 3, the notch band characteristic is presented. Discussion and simulation results of VSWR, surface current distribution, radiation pattern, and antenna gain are given in Sect. 4. Finally, the conclusion remark is summarized in Sect. 5.

2 Designed Antenna The designed patch antenna is illustrated in Fig. 1, it is composed of a radiating element patch, placed at a height of h = 0.8 mm from the dielectric substrate. The dielectric material used in our antenna is FR4-epoxy having a dielectric permittivity of 4.4, a thickness of 0.8 mm, and a loss tangent value of 0.025. This material is very low cost and commercially available. A partial ground plane is etched in the bottom of the substrate with dimensions lg
wg as seen in Fig. 1b. The antenna feeding impedance is proposed to be 50 X. A good impedance matching is achieved using a width of feedline calculated by the given equations [5]:

 2h er  1 0:61 wf ¼ A  1  lnð2A  1Þ þ lnðA  1Þ þ 0:39  p 2er er p ffiffiffi Where A ¼ 2 Zg0 p and Z0 ¼ 50 X e 0

r

g0 = 377 X the free-space wave impedance.

Fig. 1. Proposed antenna structure.

Design of a Microstrip Patch UWB Antenna with Notch Band Characteristic

49

The antenna parameters are summarized in Table 1. Table 1. Antenna parameters. Parameters Length of substrate L Width of substrate W Length of ground plane lg Width of feed line wf Radius of circular patch c Radius a Radius b Gap g Height of substrate h

Dimensions [mm] 24 14 9 1.5 7 3.1 4.1 3.9 0.8

3 Notch Band Characteristic In order to avoid the interferences with other application operating in the frequency band 5–6 GHz, we inserted a split ring in the circular patch, where its dimensions have been accurately determined. The notch band frequency generated by the etched split ring in the circular patch is calculated by the following equation [6]: fNB ¼

c pffiffiffiffiffiffi 2LNB eeff

Where c is the speed of light, LNB is the length of the etched split ring in the circular patch which is (2pa þ 2ðb  aÞ  g), and eeff is the effective dielectric constant. The notch band is located at 5.5 GHz and is achieved when LNB ¼ 16.59 mm.

4 Results and Discussions The VSWR is simulated using HFSS for an initial design which starts with a circular patch (Fig. 2a), in order to achieve the notch band characteristic, we inserted a split ring in the radiating patch as shown in Fig. 2b. The voltage standing wave ratio (VSWR) is a function of the reflection coefficient, it describes how well the matching between the transmission line and the antenna. A good antenna is defined by an acceptable value of VSWR (1  VSWR  2).

50

L. Aguni et al.

Fig. 2. Geometric evolution of the proposed antenna.

The VSWR is obtained using HFSS simulator as shown in Fig. 3, the antenna bandwidth varies from 3.85 GHz to 12.38 GHz for VSWR  2 with the rejection of the band 5.1 GHz–6.00 GHz. This characteristic helps to avoid interference with applications working at this band such as WLAN, IEEE 802.11a, and HIPERLAN/2.

Fig. 3. VSWR vs Frequency of the proposed antenna.

Design of a Microstrip Patch UWB Antenna with Notch Band Characteristic

51

From the −10 dB bandwidth simulated results illustrated in Fig. 4, it is clear that the antenna covers the entire band from 3.85–12.38 GHz except the notched band 5.1– 6 GHz, this result indicates that the proposed antenna can be a better candidate for UWB applications. The difference between the HFSS and CST results is due to the fact that the methods used by the two simulators are different; HFSS is based on Finite Element Method (FEM) while CST is based upon Finite Integration Technique (FIT).

S11 HFSS S11 CST

0

S11 [dB]

-10

-20

-30

-40

-50 2

4

6

8

10

12

Freq [GHz] Fig. 4. S11 vs Frequency of the proposed antenna.

Fig. 5. Surface current distribution at different frequencies.

14

52

L. Aguni et al.

To better understand the notch band characteristic and the radiation mechanism, the surface current distribution is plotted at different frequencies in Fig. 5. As shown, at 3.5 GHz small currents flow around the etched split ring, while important currents are around the feedline. At 5.5 GHz, one can observe that more currents are concentrated around the inner and outside the split ring. At 10 GHz, the currents are concentrated at the end of the feedline. The simulated radiation pattern at E plane and H plane at different resonant frequencies 3.5 GHz, 5.5 GHz, 5.8 GHz, and 10 GHz is presented in Fig. 6. As plotted, the antenna exhibits a bidirectional radiation pattern in E plane, while an omnidirectional radiation pattern is attained in H plane.

Fig. 6. Far field radiation pattern at different frequencies.

Design of a Microstrip Patch UWB Antenna with Notch Band Characteristic

53

The antenna gain has been also determined as plotted in Fig. 7, it can be seen that the proposed antenna presents an acceptable gain except for the notch band (Table 2).

Fig. 7. Gain of the proposed antenna. Table 2. Comparison of the proposed antenna with existing UWB antennas. Article [7] [3] [8] [9] Proposed

Antenna dimension [mm
mm
mm] 18.7
17.6
1.5 20
18
1.6 39
35
0.8 35
30
1.5 24
14
0.8

Bandwidth [GHz] 2.9–13.7 3.04–20.22 3.1–10.6 2.55–12 3.85–12.38

Max. Gain [dBi] 7 3 – 6 3.08

Notch-band [GHz] 5.1–5.9 5–6 5.15–5.85 5.4–6.1 5.1–6

5 Conclusion In this paper, we designed a microstrip patch antenna for UWB technology, the patch antenna has a global size of 24
14
0.8 mm3 and it is printed on a low cost dielectric substrate (FR4 epoxy). The antenna operates in the frequency band 3.85– 12.38 GHz with good radiation characteristic and acceptable gain. The main advantage of the proposed antenna is its simpler structure and the property of rejecting the band

54

L. Aguni et al.

5.1–6 GHz to avoid interference with applications operating at this band such as WLAN, IEEE 802.11a, and HIPERLAN/2.

References 1. Bong, H.-U., Hussain, N., Rhee, S.-Y., Gil, S.-K., Kim, N.: Design of an UWB antenna with two slits for 5G/WLAN-notched bands. Microw. Opt. Technol. Lett. 61, 1–6 (2019) 2. Shi-Wei, Q., Ruan, C.-L., Xue, Q.: A planar folded ultrawideband antenna with gap-loading. IEEE Trans. Antennas Propag. 55(1), 216–220 (2007) 3. Mohammadi, S., Nourinia, J., Ghobadi, C., Majidzadeh, M.: Compact CPW-fed rotated square shaped patch slot antenna with band-notched function for UWB applications. Electron. Lett. 47(24), 1307–1308 (2011) 4. Liu, H.-W., Yang, C.-F.: Miniature hook-shaped monopole antenna for UWB applications. Electron. Lett. 46(4), 265–266 (2010) 5. Hammerstad, E.O.: Equations for microstrip circuit design. In: Proceedings of the Fifth European Microwave Conference, pp. 268–272 (1975) 6. Dissanayake, T., Esselle, K.P.: Prediction of the notch frequency of slot loaded printed UWB antennas. IEEE Trans. Antennas Propag. 55(11), 3320–3325 (2007) 7. Syed, A., Aldhaheri, R.W.: A very compact and low profile UWB planar antenna with WLAN band rejection. Sci. World J. 2016, 1–7 (2016) 8. Yazdi, M., Komjani, N.: Design of a band-notched UWB monopole antenna by means of an EBG structure. EEE Antennas Wirel. Propag. Lett. 10, 170–173 (2011) 9. Liu, Y., Chen, Z., Gong, S.: Triple band-notched aperture UWB antenna using hollow-crossloop resonator. Electron. Lett. 50, 728–730 (2014)

60 GHz RoF System Based on IR-MBOOK Transmitter and Non-coherent Receiver Tarik Zarrouk1(B) , Ali El Moussati1 , Papa Alioune Fall2 , and Gha¨ıs El Zein3 1

Department of Electronics, Informatics and Telecommunications, ENSAO, University Mohammed Premier, Oujda, Morocco {t.zarrouk,a.elmoussati}@ump.ac.ma 2 Department Applied Science and Technology, Gaston Berger University, Saint Louis, Senegal [email protected] 3 INSA Rennes, CNRS, UMR 6164 - IETR, 35000 Rennes, France [email protected]

Abstract. A new approach to implement a Radio over Fiber (RoF) system for millimeter wave (mm-wave) is proposed and investigated. At the Central Station (CS) the mm-wave signal is produced using Impulse Radio Multiband On-Off Keying (IR-MBOOK) architecture. Then, we use an external modulator to modulate the optical signal that propagates to the Base Station (BS) through the optical fiber. This system is proposed as a solution to deal with the demands of multi-Gbps data transmission in the 60 GHz band and beyond, for nomadic applications. Low complexity, cost reduction, and performance enhancement are achieved by simplifying the mm-wave generation method. The IR-MBOOK design and external modulation are jointly used in this work. The optical link is based on Single Mode Fiber (SMF) to reach a long distance. At the receiver, a non-coherent receiver has been used in order to down-convert the signal to the baseband. The system efficiency is evaluated and analyzed by quality factor (Q factor) performance. Simulation results show the efficiency of monocycle pulse compared with Gaussian pulse shape, and RoF system with transmission rate of 4 Gbps is successfully achieved up to 45 km. Keywords: mm-wave · IR-MBOOK Monocycle pulse · RoF

1

· 60 GHz · Gaussian pulse ·

Introduction

Nowadays, the impressive growth in wireless and mobile technology, and the demand for high-speed data transmission encouraged the researchers and standards organizations to deliver their services and to keep pace with telecommunication development. The need for powerful systems, which could support high c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 55–62, 2020. https://doi.org/10.1007/978-3-030-53187-4_7

56

T. Zarrouk et al.

throughput, envisaged for the future wireless networks are encouraged to benefit from the advantages of both mm-wave frequencies and optical fiber. At this stage, mm-RoF technology is becoming a potential option for the broadband access network [1]. In addition, mm-RoF is a key enabler in offering broadband wireless access services at Ultra-High Frequency (UHF) radio signal [2]. Bandwidth capacity can be increased by using optical fiber technology which can support a wide coverage area. In addition to that, the use of mm-wave frequency (60 GHz) provides an authorized of 7 GHz band (57–64 GHz) suitable for high capacity and at the same time allows to overcome the problem of spectral congestion at lower frequencies [3]. To transport the mm-wave from Central Station (CS) to Base Station (BS), several techniques are proposed, like Base Band over Fiber (BBoF), Intermediate Frequency over Fiber (IFoF), and Radio Frequency over Fiber (RFoF) [4,5]. Recently, numerous studies have been published. For instance in [6], the authors consider that the RoF system can reach 2 Gbps up to 40 km in the 40 GHz and in [7], 20 Gbps up to 20 km in the 220 GHz band. In this work, we propose a new approach to implement a RoF system operating in the mm-wave frequency band. This approach is based on IR-MBOOK to transport the radio signal after being modulated from CS to BS. Moreover, we used a Single Mode Fiber (SMF), which is more suitable for long transmission link. Energy Detection (ED) receiver is used to down-convert the signal to the baseband, thereby avoiding the use of any local oscillator, and significantly reducing costs while maintaining the simplicity of the system. The principle of the MBOOK transceiver architecture consists of dividing the spectrum of mm-wave (57–64 GHz) into several sub-bands [8]. Each of these sub-bands contains an OOK modulated pulse to carry a bit. The demodulation is based on non-coherent energy detection [9]. Figure 1 shows the transceiver block diagrams. This approach of mm wave transport can reduce the architecture cost, as well as having the possibility to strengthen immunity against noise and to increase the spectral efficiency. In addition, it could improve the mm wave RoF technology to deploy the required 60 GHz networks. The remainder of this paper is organized as follows: Sect. 2 describes the system principle and architecture. Sect. 3 presents the simulation setup and results. Finally, conclusions are given in Sect. 4.

2

System Principle and Architecture

In this section, we present the operating principle of millimeter wave approach before discussing the system performance. Figure 1 shows the block diagram of the proposed system design. Two main parts constitute the mm-wave RoF architecture: Central Station (CS) and Base Station (BS). CS consists of a pulse generator, which delivers a very short pulse to cover the allowed band (57– 64 GHz). A multiplexer (1XN) divides the signal in N sub-bands; each sub-band has a bandwidth of 1.65 GHz. A filter bank divides signal energy in each band. An On-Off Keying (OOK) modulator modulates the pulses according to the

60 GHz RoF System Based on IR-MBOOK Transmitter

57

sequence of bits to be transmitted. For the optical part, a Continuous Wave (CW) source is used with a power of 10 dBm at 1550 nm wavelength and linewidth of 0.1 MHz, which emits a continuous laser beam. An external Mach-Zender Modulator (MZM) is used to obtain a modulated optical carrier. The optical signal is then launched into SMF, which has a dispersion of 16.75 ps/n/km and an attenuation of 0.2 dB/km before propagating to the BS.

Fig. 1. Block diagram of a RoF system with mm-wave photonic generation.

At the BS, the optical signal is converted to an electrical signal by a Photo Detector (PD). Then it is amplified, filtered and demodulated. Signal demodulation is made using a non-coherent demodulator, which works as an energy detector. The information is carried by signal amplitude rather than its phase, which compares the received signal energy to a given energy threshold [10]. The maximum data rate obtained at the transmitter output depends on the number of sub-bands N and the pulse repetition time (Tr ). D=

N Tr

(1)

In this work, we use the four sub-band approach. The latter has a good average probability of error (Pe), for a given signal-to-noise ratio (SNR) and with low complexity. Moreover, most standards in the frequency band around 60 GHz divide the allocated band into 4 channels. The non-coherent receiver provides an alternative low complexity detector, which is ideal for low-power, and lowcost applications. It consists of a squarer and Low Pass Filter (LPF) to recover the signal energy. The block diagram of this demodulator is described in Fig. 2, where the input pulse Xin (t) is squared that results in a signal Xout (t), which is then low-pass filtered. Considering the impulse response of the filter h(t), its output y(t) is defined as:  +∞ y(t) = Xout (t) ⊗ h(t) = h(ρ)Xout (t − ρ)dρ (2) −∞

58

T. Zarrouk et al.

Fig. 2. Energy Detector topology.

3

Simulation Setup and Results

The transmission architecture is realized using the OptiSystem software, to follow the transmission process and analyze the overall performance of the architecture in high frequency and in baseband. Simulation results are represented using advanced optical simulation package with the ability to virtually plan, test, and simulate almost every type of optical link in the transmission layer [11]. This work investigates a RoF downlink system for the generation of the mm-wave signal based on an IR-MBOOK method. Figure 3 presents the configuration of the mm-wave generating method, where different simulation spectrums are presented.

Fig. 3. Experimental setup of mm-wave generation method with simulation results.

60 GHz RoF System Based on IR-MBOOK Transmitter

59

Point (A) shows the generated mm-wave carrier at 60 GHz. Point (B) shows the optical spectrum of the continuous wave, injected optical mean power is 10 dBm. Point (C) represents the optical spectrum of signal after propagating through the fiber. Finally, point (D) shows the optical non coherent receiver output. To implement the new approach for generating the mm-wave, two main issues are taken into consideration: Laser optical power injected, and pulse generator type. First, we investigated the performance of the system according to the optical power. We consider a data rate of 2 Gbps at a distance of 40 km through SMF, and we measure the Q factor at reception as a function of the optical power injected (Fig. 4). Q factor measures the signal quality, which takes into account physical alterations of the signal, such as noise, chromatic dispersion and non-linear effects. These phenomena can degrade the signal. In fact, the higher the Q factor value the better the SNR, therefore the lower the probability of bit error.

Fig. 4. Optimum injected power for IR-MBOOK.

Laser nonlinearity degrades the performance of the transmission, thus decreasing the quality of signal. The resulting optimal value is 10 dBm, which corresponds to the maximum value of Q factor. In order to determine the optimum pulse generator, we measure the Q factor of the signal at reception in function of the type of pulse generator. The IRMBOOK over fiber downlink system, reported in Fig. 1, has been simulated with OptiSystem software to investigate the shape of the pulse after all-optical

60

T. Zarrouk et al.

frequency up-conversion at 60 GHz. The simulated input pulse, reported in Fig. 5, has a width of 142 ps to cover a bandwidth of 7 GHz.

Fig. 5. Simulated waveforms: (a) Gaussian pulse; (b) Monocycle pulse.

The performance of the pulse generator is first evaluated. Figure 6 shows the variation of Q factor according to the distance (D). We have fixed the injected optical power at 10 dBm, and we have compared the value of Q factor at the receiver for a data rate of 2 Gbps.

Fig. 6. Variation of Q factor according to the distance through SMF.

60 GHz RoF System Based on IR-MBOOK Transmitter

61

The acceptable value of Q factor is greater than 7, which corresponds to a Bit Error Rate (BER) less than 10e−9 [11]. Concerning the architecture that uses the monocycle pulse generator, we can observe that the signal range can reach 100 km, where the Q factor is 8.01. In the same architecture using the Gaussian pulse generator, signal range does not exceed 90 km. Thus, the monocycle pulse can be transmitted more efficiently compared to the Gaussian pulse, which makes it the most used in ultra-wide band impulse signals. In order to assess the performance of the MBOOK design over optical fiber, we use the monocycle pulse generator, four sub-band approach, SMF and noncoherent receiver. Figure 7 shows Q factor performance according to the distance (D).

Fig. 7. Q factor vs. SMF length for MBOOK system.

It can be seen that reached distance is 45 km at 4 Gbps, 76 km at 3 Gbps and 102 km at 2 Gbps with Q factor more than 7. Finally, these results clearly show that the proposed architecture achieves higher distance with low complexity and less power consumption.

4

Conclusion

In this paper, a new approach for the generation and transport of mm-wave signals, which is based on MBOOK architecture and energy detection, has been presented and analyzed. First, the performance of the monocycle pulse generator has been evaluated. Then, the Q factor performance of the proposed method has been calculated for various data rates. The obtained results show that the proposed system can ensure optimal transmission, with a reached distance of 45 km at a rate of 4 Gbps.

62

T. Zarrouk et al.

References 1. Beas, T., Castanon, G., Aldaya, A., Aragon-Zavala, A.: Millimeter-wave frequency radio over fiber systems: a survey. IEEE Commun. Surv. Tutor. 15(4), 1593–1619. Fourth Quarter (2013). https://doi.org/10.1109/SURV.2013.013013.00135 2. Abdollahi, S.R., Al-Raweshidy, H.S., Fakhraie, S.M., Nilavalan, R.: Digital radio over fibre for future broadband wireless access network solution. In: Proceedings of the IEEE International Conference on Wireless and Mobile Communications (ICWMC 2010), pp. 504–508 (2010). https://doi.org/10.1109/ICWMC.2010.32 3. Cheng, L., Zhu, M., Muhammad Usman Gul, M., Ma, X., Chang, G.: Adaptive photonics-aided coordinated multipoint transmissions for next-generation mobile fronthaul. J. Lightwave Technol. 32(10), 1907–1914 (2014). https://doi.org/10. 1109/JLT.2014.2316090 4. Zarrouk, T.E., Yahyaoui, M.E., Moussati, A.E., Oualkadi, A.: Investigation of radio channel model in indoor environment at 60 GHz. J. Model. Ident. Control 29(4), 359–363 (2018). https://doi.org/10.1504/IJMIC.2018.092137 5. El Yahaoui, M., El Moussati, A., Ghoumid, K., Lepers, C.: IEEE802.15.3C transmission over multimode BER links: performance comparison of RF and IF over BER architectures. Int. J. Microw. Opt. Technol. 11 384–390 (2016) 6. Zhang, J.W., Yu, J.J., Chi, N., Li, F., Li, X.: Experimental demonstration of 24Gb/s CAP-64QAM radio-over-fiber system over 40-GHz mm-wave fiber-wireless transmission. Opt. Express 21(22), 26888–26895 (2013). https://doi.org/10.1364/ OE.21.026888 7. Das, S., Dutta, S., Ghorai, K.: Design of an RoF Downlink System for generation of dual frequency millimetre wave carrier signal by frequency multiplication. In: International Conference on Applied and Theoretical Computing and Communication Technology (iCATccT), pp. 363–367 (2016) 8. Koeng, S., Antes, J., Lopez-Diaz, D., Kallfass, I., Koos, C., Freude, W., Leuthold, J.: High-speed wireless bridge at 220 GHz connecting two fiber-optic links each spanning up to 20 Km. In: Optical Fiber Communication Conference (OFC), pp. 1–3, March 2012 9. Paquelet, S., Aubert, L.M., Uguen, B.: An impulse radio asynchronous transceiver for high data rate. In: International Workshop on Ultra Wideband Systems Joint with Conference on Ultra Wideband Systems and Technologies. Joint UWBST IWUWBS 2004 (2004). https://doi.org/10.1109/UWBST.2004.1320888 10. Zarrouk, T., El Moussati, A., Fall, P.A., El Maimouni, L.: Performance evaluation of multiband 60 GHz impulse radio for broadband. In: International Symposium on Advanced Electrical and Communication Technologies (ISAECT), Rabat, Morocco, pp. 1–4 (2018). https://doi.org/10.1109/ISAECT.2018.8618808 11. https://optiwave.com/optisystem-overview/ 12. Waseer, T.A., Halepoto, I.A., Joyo, M.A.: Quantifying the Q-factor and minimizing BER in 32-channel DWDM system design using EDFA and RAMAN amplifiers. Mehran Univ. Res. J. Eng. Technol. 33(1), 1–8 (2014)

Impact of Human Morphology on Measurement Errors of a RF Exposimeter Abdechafik Derkaoui1 , Rodrigues Kwate Kwate1 , Bachir Elmagroud1(B) , Dominique Picard2 , and Abdelhak Ziyyat1 1

2

Electronic and Systems Laboratory, Mohammed Premier University, Oujda, Morocco [email protected] DRE, Signals and Systems Laboratory, CentraleSupelec, Paris, France

Abstract. Calibration of RF exposimeter is a serious problem when conventional measurement methods are used, especially when the morphology of the user, workers or public, changes a lot. In this paper we present a study of the behavior of new measurement methods based on linear regression with respect to morphological change. For this we use three models: Child, Gustav and Emma of different sex, age and dimensions. The study is conducted in the near field for a base station antenna at the DCS band. The study showed a low dependence of the proposed methods on morphology. Keywords: RF dosimetry

1

· Human morphology · Exposimeter

Introduction

Today, questions about human exposure to electromagnetic waves no longer concern only the conformity of radiocommunication devices but also real exposure, i.e. in the presence of the human body. Accurate assessment of a person’s level exposure to electromagnetic waves is very complex, especially in the presence of the human body [1–3]. Indeed, in the article [4] we have shown the difficulties posed by this measurement, then we have given a description of the errors induced by the presence of the human body and finally we have proposed several techniques to reduce these errors in a significant way. This is made possible by a calculation algorithm based on linear regression analysis and which can be easily integrated into a real exposimeter. On the other hand, the analysis of the actual exposure requires taking into account the variability of the different parameters that can influence the levels of electromagnetic fields. In fact, the human body diffuses the incident field around it and creates areas of strong field and weak field [5]. Thus the field measured with a radiofrequency exposimeter can depend on various parameters such as the morphology, posture of the human body, the frequency and incidence of c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 63–70, 2020. https://doi.org/10.1007/978-3-030-53187-4_8

64

A. Derkaoui et al.

the wave, the position of the dosimeter relative to the human body and the characteristics of the sources of exposure [6]. The variability of these parameters does not allows us an exact evaluation but only the obtaining a value with a certain uncertainty (Fig. 1).

Fig. 1. Parameters that can influence the measurement of human exposure by an RF dosimeter.

In this paper, we propose to calibrate and optimize the measurement techniques proposed in the article [4]. Indeed, we performed a parametric study according to the size of the human body. The main objective is to characterize the behavior of the proposed methods and their coefficients in real situations as well as the modifications that can result from these coefficients, this in accordance with the procedures recommended in the norms and standards related to the exposure to electromagnetic waves. Thus, we can know if the dimensions of the human body (body size and height), sex, age have an effect on the coefficients of the measurement methods proposed in [4]. In order to carry out this work, we opted for a set of technical choices that represent both the different exposure environments and facilitate the complex numerical simulations.

2 2.1

EMF Exposure Scenarios and Methods EMF Exposure Scenarios

Numerical simulation were carried out using CST Studio based on finite integration technique (FIT) as numerical simulation model. Several models of the human body are used (Fig. 2a): a small human body (Child, female, height:

Impact of Human Morphology on Measurement Errors of a RF Exposimeter

65

115 cm, mass: 21.7 kg, age: 7 years) to a large human body (Emma, female, height: 170 cm, mass: 81 kg, age: 26 years) through a medium-sized human body (Gustav, man, height: 176 cm, mass: 69 kg, age: 38 years). The simulation frequency is 1842 MHz (DCS band). The separation distance between the human body and the antenna is 2 m ≈ 12λ at near field zone. The antenna presented in Fig. 2b is used and the dipoles 1, 12, 13 and 24 are powered by signals of amplitude 0.25 V, the dipoles 2, 11, 14, 23 are powered by 0.5 V amplitude signals, the dipoles 3, 10, 15, 22 are powered by signals of amplitude 0.75 V and the other dipoles by signals of amplitude 1 V. The characteristics obtained from this antenna (Fig. 2c) make it an almost perfect replica of the PCN base station antenna D065-19-2ASB, serial number 99131 designed by COMSAT RSI CSA Antenna Systems.

(a) Body Size

(b) Base Station Antenna

(c) Scattering Diagram

Fig. 2. Exposure system

66

A. Derkaoui et al.

In [4] we have already defined many of errors due to the presence of human body: Body Reflection Error (BRE), Body Shadowing Error (BSE), Body Diffraction Error (BDE) and finally the Body average Error (BE) that we will use in this paper because it includes all the previous errors, express as follows: N

BE =

− EP i 100  EPfree i N i=1 EPfree i

body

(1)

where EPbody is the strength of the electric field at point Pi (Fig. 3), with the i presence of body and EPfree the undisturbed electric field strength at the same i point in the absence of human body as recommended by the ICNIRP [7].

Fig. 3. Locations of field evaluation points. The magnitude of the electromagnetic fields (EPi and HPi ) are evaluated on N points Pi (i = 1, ..., N ) located on ellipses around the body, at three different heights: around the chest, abdomen and waist.

2.2

Methods

For this analysis we will use three of error mitigation techniques: – Multi-Coefficient Method : The multi-coefficient method is based on multiple linear regression analysis to find a linear equation between field values measured in the presence of the human body and values measured in free space. It will be a question of determining the regression coefficients (vector A) which minimize the residue (vector Ra ) thus the error. For each situation, the A vector is obtained from digital postprocessing using the codes developed with Matlab: (2) E free = X body A + Ra ,

Impact of Human Morphology on Measurement Errors of a RF Exposimeter

67

with

E free

1 .. . = i .. . N

⎞ EPfree 1 ⎜ .. ⎟ ⎜ . ⎟ ⎜ free ⎟ ⎜ EP ⎟ ; A = ⎜ i ⎟ ⎜ . ⎟ ⎝ .. ⎠ ⎛

EPfree N

⎛ ⎞ 1 a1 .. ⎜ .. ⎟ ⎟ . ⎜ ⎜ . ⎟ ⎟ k ⎜ a ⎜ k ⎟. .. ⎜ .. ⎟ . ⎝ . ⎠ K aK

(3)

The X body matrix is composed by N rows and K columns. The different methods are based on the possibility of measuring four different quantities, all in V /m. For any (i, k) position on the measuring ellipse (Fig. 3). So, the (i, k) entry xbody Li,k in this matrix is defined as: ELbody : M11 body i,k xLi,k = (4) body body (ELi,k + Z0 × HLi,k )/2 : M14 where M11, M14 in Eq. 4 and M22, M23 in Eq. 6 are the references of each methods given in [4], Z0 ≈ 120 × π ≈ 377 Ω is the wave-impedance of a plane wave in free space and ELbody and HLbody are respectively the electric field i,k i,k strength and the magnetic field strength measured at the Li,k position on the body. The residual vector Ra contains the N errors terms: the maximum error is a a = max(Ria ), and the average error is Rmean = mean(Ria ), i = 1, . . . , N . Rmax – Single-coefficient method with maximum value: The single-coefficient method is based on a simple linear regression: E free = β Y body + Rb

(5)

in the where β is the single correction factor and where the i-th entry yPbody i body vector Y is defined as: 

yPbody = i

max(Z0 × HLbody , ..., Z0 × HLbody ) i,1 i,K body body body max(max(ELi,1 , Z0 × HLi,1 ), ..., max(EL , Z0 i,K

: M22 × HLbody )) : M23 i,K

(6) In this method, we take maximum quantity for all the K points used in measuring ellipse. The residual vector Rb contains the N error terms. As for b = max(Rb ) and the single-coefficient method, the maximum error is Rmax b b average error is Rmean = mean(R ). – Single-coefficient method with arithmetic mean: To the methods presented above we add a third method which uses the arithmetic mean over the number K of the point of measurement considered. Here the vector Y body is defined as: ⎧ body body EL +...+EL ⎨ i,1 i,K : M31 body K (7) yP i = body body body body ⎩ (ELi,1 +Z0 ×HLi,1 )/2+...+(ELi,K +Z0 ×HLi,K )/2 : M34 K

68

A. Derkaoui et al.

where M31, M34 in Eq. 7 are the references of each methods. In addition to the methods described above, we also consider method M00, which represents a conventional measurement technique using a single sensor, as well as methods M10 and M20, which represent the simple cases when we just measure the electric field in a two-sensor setup, either by averaging the values (M10) or taking the maximum value (M20). M10 is also used in [8].

3

Results and Discussion

Fig. 4. Distribution of the electric field for different morphology: Child, Gustav and Emma.

Figure 4 shows that the distribution of the electric field is strongly affected by the morphology of the human body. Indeed, the three field zones, reflection zone SR , shadowing SS zone and the diffraction zone SD [4], vary strongly according to the morphology. The reflection zone is relatively larger for the Child and Emma models, but it is more limited for the Gustav model. The diffraction zone is inversely larger for the Gustave model and it is smaller for the Child and Emma models. Figure 5 shows the global behavior of the error rates obtained for each morphology (Child, Gustav and Emma) with the measurement techniques presented above: the conventional methods: M00, M10 and M20, the multiple linear regression methods: M11 and M14, the simple linear regression method with maximum extraction: M22 and M23 and finally the simple linear regression method with arithmetic mean: M31 and M34. b b and average error Rmean defined Tables 1 and 2 give the maximum error Rmax in Sect. 2.2 and the weighting coefficients used by each method. With the weighting coefficients: β in the Eq. 5, a1 and a2 the elements of the vector A in Eq. 2. After analyzing the results, we can make the following remarks: – Whatever the body size, the M22 and M23 methods are the most accurate. Indeed, the value of the error remains below 20%, which constitutes a very acceptable value in RF dosimetry. – The difference between the results obtained for the different morphologies is relatively weak, with the exception of the two conventional methods M00, M01 and M02. Indeed, if we take for example the method M23, it produces average error rates of 15.5%, 11% and 16.3% for the Emma, Gustav and Child

Impact of Human Morphology on Measurement Errors of a RF Exposimeter

69

Fig. 5. The Body Error BE for differents methods and morphology. b b Table 1. Maximum error Rmax , average error Rmean and the weighting coefficients for conventional methods and Single-coefficient methods.

Emma β M00 –

Gustav

Child

Rmax (%) Rmean (%) β

Rmax (%) Rmean (%) β

Rmax (%) Rmean (%)

106

98.5

40.5

–

94.2

37.6

–

M10 0.50 66.7

29.8

0.50 66.9

33.2

0.50 62.2

29.1

M20 1

22.5

1

58.0

16.0

1

66.1

21.8

M22 1.05 54.6

12.6

1.06 57.7

13.5

1.12 51.0

17.4

M23 0.82 92.4

15.5

0.95 43.1

11.0

0.86 47.4

16.3

M31 1.24 98.5

23.2

1.39 53.7

22.3

1.36 44.3

16.7

M34 1.21 65.7

19.4

1.42 51.6

20.3

1.47 40.7

16.0

137

43.0

b b Table 2. Maximum error Rmax , average error Rmean and the weighting coefficients for Multi-Coefficient Method.

Emma a1 a2

Rmax (%)

Rmean (%)

Gustav a1 a2

Rmax (%)

Rmean (%)

Child a1 a2

Rmax (%)

Rmean (%)

M11 0.69 0.55 88.4

22.8

0.75 0.64 51.8

22.3

0.70 0.65 43.7

16.6

M14 0.76 0.61 57.4

20.3

0.77 0.65 51.8

19.2

0.76 0.70 41.6

15.9

models. In the same order, we obtain 12.6%, 13.5% and 17.4% for the M22 technique. Overall, the variation of the error rates for the different models remains of the order of 5%. – The reduction of the error is always ensured by the techniques proposed for different people of different ages and morphologies.

70

4

A. Derkaoui et al.

Conclusion

The analysis carried out in this paper shows that the morphology does not have a great impact on the behavior of the presented methods contrary to conventional methods. In general, we find that the methods based on linear regression give interesting results when the dimensions of the model is reduced in our case it is the model of a girl (Child). The use of a single exposimeter (M00) gives a BE error of 7.9 dB. The proposed methods has reduced the error significantly as shown in the Table 1 and this corresponds to the results found in [8]. Noting that the number of measuring points used in [8] (22 node) remains very much higher than the two measurement point we took. The implementation of the proposed methods improves the accuracy of the exposimeter. On the other hand, they offer the possibility of calibration before use with the predetermined coefficients which gives it a low dependence the user.

References 1. Jokela, K.: Assessment of complex EMF exposure situations including inhomogeneous field distribution. Health Phys. 92(6), 531–540 (2007) 2. Miguel-Bilbao, S., Garca, J., Ramos, V., Blas, J.: Assessment of human body influence on exposure measurements of electric field in indoor enclosures. Bioelectromagnetics 36(2), 118–132 (2015) 3. Elmagroud, B., Kwate, K.R., Taybi, C., Picard, D., Ziyyat, A.: Electromagnetic exposure assessment for telecommunication equipments using RF dosimeter. In: IEEE International Conference on Information Technology for Organizations Development, pp. 1–6. IEEE Press, Fez (2016) 4. Kwate, R.K., Elmagroud, B., Taybi, C., Beauvois, V., Geuzaine, C., Picard, D., Ziyyat, A.: Measurement methodologies for reducing errors in the assessment of EMF by exposimeter. Progress Electromagn. Res. B 78, 31–46 (2017) 5. Kwate, K.R., Taybi, C., Elmagroud, B., Beauvois, V., Geuzaine, C., Picard, D., Ziyyat, A.: On calibration of correction law for EMF measurement errors due to the proximity of the human body. In: 15th IEEE Mediterranean Microwave Symposium, pp. 1–4. IEEE Press, Lecce (2015) 6. Iskra, S., McKenzie, R., Cosic, I.: Factors influencing uncertainty in measurement of electric fields close to the body in personal. Radiat. Prot. Dosim. 140(1), 25–33 (2010) 7. ICNIRP: Guidelines for limiting exposure to time-varying electric, magnetic and electromagnetic fields (up to 300 GHz). Health Phys. 74(4), 494–522 (1998) 8. Aminzadeh, R., Thielens, A., Agneessens, S., Torre, P.V., Bossche, M.V., Dongus, S., Eeftens, M., Huss, A., Vermeulen, R., Seze, R., Mazet, P., Cardis, E., Rogier, H., R¨ oo ¨sli, M., Martens, L., Joseph, W.: The effect of antenna polarization and body morphology on the measurement uncertainty of a wearable multi-band distributed exposure meter. Ann. Telecommun. 74(1–2), 67–77 (2019)

RF-Exposimeter Errors Measurement: Frequency and Distance Impact Rodrigues Kwate1 , Bachir Elmagroud1(B) , Abdechafik Derkaoui1 , Chakib Taybi1 , Dominique Picard2 , and Abdelhak Ziyyat1 1

2

Electronic and Systems Laboratory, Mohammed Premier University, Oujda, Morocco [email protected] DRE, Signals and Systems Laboratory, CentraleSupelec, Paris, France

Abstract. In this paper, we present a study of the behavior of new RFexposimeter measurement methods based on linear regression as a function of frequency and distance change. For this, we use a realistic human model (Gustav) exposed to the radiation of a base station antenna. The frequency bands used are GSM, DCS and LTE-Wimax. The distance varies from the near field zone to the far field zone. The errors induced by the human body have been evaluated. The study showed that errors increase with frequency and decrease in Far Field zone. But in general, the proposed methods have a good response to variations in frequency and distance.

Keywords: RF exposimeter error

1

· Near field · Far field · Body exposition

Introduction

The EMF assessment, measured with exposimeters, is subject to additional errors and uncertainties due to the presence of the human body [1–3]. The device is worn by the user, who can be a general public person or a worker, allows measurement of EMF levels in order to compare them to the reference levels indicated by standards [4]. These portable body-worn devices can be mounted at different places on the human body. The measurement of the expositometer strongly depends on various parameters such as the morphology, posture of the human body, the frequency and incidence of the wave, the position of the dosimeter relating to the human body and the characteristics of the sources of exposure [5]. Indeed, several studies aimed at quantifying the influence of these various parameters on the measurements of the exposimeter have been published. It is shown in [6] and [7] that the body can produce attenuations up to 30 dB at 900 MHz. The errors due to the position of the exposimeter on the body have proved significant in [8]. The study also showed the effect of the frequency on c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 71–79, 2020. https://doi.org/10.1007/978-3-030-53187-4_9

72

R. Kwate et al.

the measurement error. Finally, a detailed study of the influence of the body for several frequency bands is presented in [9]. A number of techniques have been proposed to mitigate the error of evaluation of electromagnetic fields related to the presence of the human body. The method proposed in [10] uses numerical simulations to estimate an average electric field and a specific absorption rate (SAR). The main problem with the RF-exposimeter is its calibration in a free space. The presence of the human body and its interactions with the previously mentioned parameters make the response of the expository widely varied. We have been led to propose in [3] different measurement techniques which made it possible to reduce the effect of the human body on the measurement of the exposimeter [11]. In this paper we present a parametric study on the effect of the frequency band, the distance between the body and the source (near field/far field), which will allow a calibration of the device by modifying the weighting coefficients for the proposed methods.

2

Materials and Methods

The study based on electromagnetic simulations was carried out using CST Studio. The model used for simulation is a realistic human body phantom named Gustav (Fig. 1a) for a man, height: 176 cm, mass: 69 kg and age: 38 years. The model Gustav is exposed to the radiation of a base station antenna. The antenna presented in Fig. 1b is used and the dipoles 1, 12, 13 and 24 are powered by signals of amplitude 0.25 V, the dipoles 2, 11, 14, 23 are powered by 0.5 V amplitude signals, the dipoles 3, 10, 15, 22 are powered by signals of amplitude 0.75 V and the other dipoles by signals of amplitude 1 V. The characteristics obtained from this antenna in 942 MHz (Fig. 1c), make it an almost perfect replica of the PCN base station antenna D065-19-2ASB, serial number 99131 designed by COMSAT RSI CSA Antenna Systems. To study the effect of frequency we fixed the distance between the model and the source. The frequencies operated are 942 MHz, 1842 MHz and 3500 MHz. Simulations are performed for a separation distance that depends on the frequency used and is worth 12λ (Fig. 1a). The feed matrix of the antenna is preserved. For the effect of the separation distance between the exposure source (BTS antenna) and the human body, the variable element here is the distance d and the fixed elements is the frequency and the altitude. The objective here is to characterize the effect of the distance d when it varies from the near-field region (in particular the Rayleigh region) of the antenna to the far-field region (Fraunhofer region). In [3] we have already defined many of errors due to the presence of human body: Body Reflection Error (BRE), Body Shadowing Error (BSE), Body Diffraction Error (BDE) and finally the Body average Error (BE) that we will

RF-Exposimeter Errors Measurement: Frequency and Distance Impact

(a) Body Size

(b) Base Station Antenna

(c) Scatterring Diagram

Fig. 1. Exposure system.

73

74

R. Kwate et al.

use in this paper because it includes all the previous errors, expressed as follows: N

BE =

− EP i 100  EPfree i N i=1 EPfree i

body

(1)

where EPbody is the strength of the electric field at point Pi (Fig. 2), with the i presence of body and EPfree the undisturbed electric field strength at the same i point in the absence of human body as recommended by the ICNIRP [4].

Fig. 2. Locations of field evaluation points. The magnitudes of the electromagnetic fields (EPi and HPi ) are evaluated on N points Pi (i = 1, ..., N ) located on ellipses around the body, at three different heights: around the chest, abdomen and waist.

For this analysis we will use the single-coefficient method based on linear regression analysis to find a linear equation between field values measured in the presence of the human body and values measured in free space. It will be question of determining the regression coefficients which minimize the residue (vector Rb ) thus the error: E free = β Y body + Rb ,

(2)

in the where β is the single correction factor and where the i-th entry yPbody i body vector Y is defined as:  max(Z0 × HLbody , ..., Z0 × HLbody ) : M3 i,1 i,K = yPbody body body body body i max(max(ELi,1 , Z0 × HLi,1 ), ..., max(ELi,K , Z0 × HLi,K )) : M4 (3) where M3, M4 are the references of each methods. The Y body matrix is composed by N rows and K columns. The methods is based on the possibility of measuring two different quantities, all in V /m. For any (i, k) position on the measuring ellipse (Fig. 2). So, the (i, k) entry xbody Li,k in this matrix is defined as:

RF-Exposimeter Errors Measurement: Frequency and Distance Impact

75

where Z0 120 π 376.99ffl Ω is the wave-impedance of a plane wave in and HLbody are respectively the electric field strength and free space and ELbody i,k i,k the magnetic field strength measured at the Li,k position on the body. In this method, we take maximum quantity for all the K points used in measuring ellipse. The residual vector Rb contains the N error terms. As for the b = max(Rb ) and average single-coefficient method, the maximum error is Rmax b b error is Rmean = mean(R ). In addition to the methods described above, we also consider method M0, which represents a conventional measurement technique using a single sensor, as well as methods M1 and M2, which represent the simple cases when we just measure the electric field in a two-sensor setup, either by averaging the values (M1) or taking the maximum value (M2). M1 is also used in [12].

3

Results and Discussion

3.1

Frequency Effect

Figure 3 shows the global behavior of the error rates obtained for the different frequencies: GSM 942, DCS 1842 and LTE-Wimax 3500, with the measurement techniques presented above: the conventional methods: M0, M1 and M2 and the single linear regression methods: M3 and M4.

Fig. 3. The Body Error BE for different methods and frequencies. b On the other hand, Table 1 gives the maximum error Rmax and average error defined in Sect. 2 and the weighting coefficients used by each method (β defined in the Eq. 2). We can see that:

b Rmean

76

R. Kwate et al.

b b , average error Rmean conventional methods and singleTable 1. Maximum error Rmax coefficient method.

GSM (942 MHz) β M0 –

DCS (1842 MHz)

LTE-Wimax (3500 MHz)

Rmax (%) Rmean (%) β

Rmax (%) Rmean (%) β

Rmax (%) Rmean (%)

92.8

106.4

95.8

35.85

–

40.5

–

M1 0.50 51.7

23.5

0.50 66.7

29.8

0.50 76.9

26.3

M2 1

53.3

18.0

1

22.5

1

111

24.4

M3 1.15 49.6

15.5

1.05 54.6

12.6

1

98.4

23.3

M4 0.87 38.4

14.2

0.82 92.4

15.5

0.7

75.2

20

137.6

39.6

– The error rates are higher when the frequency is high. For example, for the M3 method, the average error rates are 15.1% for the 942 MHz GSM band, 22.8% for the 1842 MHz DCS band and 24.9% for the 3500 MHz LTE - Wimax band. The same behavior can be done for the method M4. – Aberrant values are higher and scattered for high frequencies. 3.2

Distance Effect

To analyze the effect of distance we will consider the near-field area: d = 3 λ, 6 λ and 12 λ, and the far-field area. The frequency is 942 MHz. The measurements are performed at three levels: Chest, Abdomen and Waist. The three figures show the body errors for Chest, Abdomen and Waist respectively. After analyzing the results, we can make the following remarks (Figs. 4, 5 and 6):

Fig. 4. The Body Error BE for different distances at Chest.

RF-Exposimeter Errors Measurement: Frequency and Distance Impact

77

Fig. 5. The Body Error BE for different distances at Abdomen.

Fig. 6. The Body Error BE for different distances at Waist.

It is strongly inadvisable to perform measurements at the waist (Waist) because the errors are very high. Chest and Abdomen levels are recommended with a preference for the chest. – Overall, we notice that the average error rate (BE) decreases as the distance increases. – The M3 and M4 methods remain the most effective for correcting errors in estimating the exposure whether we are in the near field or in the far field of the base station.

78

R. Kwate et al.

– we must avoid to perform measurements at the waist because the errors are very high. Chest and Abdomen levels are recommended with a preference for the chest.

4

Conclusion

In this work, the frequency dependence of the RF exposimeter error and the effect of the distance between the sources and the human body are assessed. The results show that the error increases, with the frequency. Secondly, the more the distance between the exposure source and the person (far field) increases, the error decreases. In general, we find that the methods based on linear regression give interesting results compared with conventional methods. Finally, we notice that the increase of the error with the frequency remains reasonable, likewise with the diminution of the distance. However, it would be interesting to study for a large frequency play.

References 1. Jokela, K.: Assessment of complex EMF exposure situations including inhomogeneous field distribution. Health Phys. 92(6), 531–540 (2007) 2. Miguel-Bilbao, S., Garca, J., Ramos, V., Blas, J.: Assessment of human body influence on exposure measurements of electric field in indoor enclosures. Bioelectromagnetics 36(2), 118–132 (2015) 3. Kwate, R.K., Elmagroud, B., Taybi, C., Beauvois, V., Geuzaine, C., Picard, D., Ziyyat, A.: Measurement methodologies for reducing errors in the assessment of EMF by exposimeter. Progress Electromagn. Res. B 78, 31–46 (2017) 4. ICNIRP: Guidelines for limiting exposure to time-varying electric, magnetic and electromagnetic fields (up to 300 GHz). Health Phys. 74(4), 494–522 (1998) 5. Iskra, S., McKenzie, R., Cosic, I.: Factors influencing uncertainty in measurement of electric fields close to the body in personal. Radiat. Prot. Dosim. 140(1), 25–33 (2010) 6. Blas, J., Lago, F.A., Fernndez, P., Lorenzo, R.M., Abril, E.J.: Potential exposure assessment errors associated with body-worn RF dosimeters. Bioelectromagnetics 28(7), 573–576 (2007) 7. Bahillo, A., Blas, J., Fernndez, P., Lorenzo, R.M., Mazuelas, S., Abril, E.J.: E-field assessment errors associated with RF dosemeters for different angles of arrival. Radiat. Prot. Dosim. 132(1), 51–56 (2008) 8. Neubauer, G., Cecil, S., Giczi, W., Petric, B., Preiner, P., Frhlich, J., Rsli, M.: The association between exposure determined by radiofrequency personal exposimeters and human exposure: A simulation study. Bioelectromagnetics 31(7), 535–545 (2010) 9. Roblin, C., Sibille, A.: Measurement of a body-worn triaxial sensor for electromagnetic field and exposure assessment. In: 8th European Conference on Antennas and Propagation, pp. 2631–2635. IEEE Press, The Hague (2014) 10. Iskra, S., McKenzie, R., Cosic, I.: Personal, non-invasive dosimetry for radiofrequency human exposure assessment. In: 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 2319–2322. IEEE Press, Lyon (2007)

RF-Exposimeter Errors Measurement: Frequency and Distance Impact

79

11. Elmagroud, B., Kwate, K.R., Taybi, C., Picard, D., Ziyyat, A.: Electromagnetic exposure assessment for telecommunication equipments using RF dosimeter. In: IEEE International Conference on Information Technology for Organizations Development, pp. 1–6. IEEE Press, Fez (2016) 12. Thielens, A., De Clercq, H., Agneessens, S., Lecoutere, J., Verloock, L., Declercq, F., Vermeeren, G., Tanghe, E., Rogier, H., Puers, R., Martens, L., Joseph, W.: Personal distributed exposimeter for radio frequency exposure assessment in real environments. Bioelectromagnetics 34(7), 563–567 (2013)

Computer Vision and Data Processing

Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts Peeter Lorents1(&), Ahto Kuuseok2, and Erika Lorents3 1

Estonian Business School, Department of Information Technology, Tallinn, Estonia [email protected] 2 Estonian Police and Border Guard Board, Tallinn, Estonia 3 Lorents Machine Learning OU, Tallinn, Estonia

Abstract. An assessment of the similarity of situations and developments in decision-making processes is under consideration. Specifically, mathematical and IT tools that make it possible to assess plausibility and similarity based on the limited amount of information available. Under observation is the assessing the plausibility of occurrence of armed conflict. The data this study is based on is limited periodically and regionally: starting from the end of World War II, until the year 2008. Developments and situations will be treated as algebraic systems. Keywords: Systems
Similarity of systems
Structural similarity
Descriptive similarity
Numeral assessment of similarities
Plausibility
Plausibility of occurrence of armed conflicts

1 Introduction The outbreak of armed conflict is normally, for most, an unpleasant surprise. It is often said that it should have been foreseen. And then, if questioned, how? The answers often based on similarities. More specifically, if in two situations: a) the first situation is preceded by the other, b) there first situation is similar to the other, c) an armed conflict started from the first situation; then it is plausible that also the similar situation might start the armed conflict. Described links of plausibility and similarity is very common in case of human analysis (preceding to decisions) and justification. Decisions have been made in this way for thousands of years, are being made today and will be made in the future. If we want to support this type of decision-making process with IT solutions, the mathematical nature of things must first be clarified. Agreeing on inevitable imperfection of real life, it is worth investigating this method based on integration of similarity and plausibility. Especially for the situations where there is shortage of information and time, and necessity for decision or position. For example: assessing if the probability for armed conflict is imminent or not. In case of outbreak of armed conflicts, it is unfortunate that normally there is never enough specific information, particularly if there is no specific intelligence information. Most of useful data is classified for the prescribed period. And when the information is open, available, this might be considered incorrect or corrupted. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 83–93, 2020. https://doi.org/10.1007/978-3-030-53187-4_10

84

P. Lorents et al.

Herein, in this current research we study: a) previous situations before armed conflicts, b) definition of similarity describing situations, and tools of numerical evaluation, c) definition of plausibility treating algebraic systems, d) integration of similarity and plausibility, e) Phenomenon of similarity and plausibility assessing armed conflicts’ manifestations of the relations between similarity and plausibility in armed conflicts (occurred after World War II until today, in regions of Eurasia, Indonesia and Middle East). Essential research problem is: a) what kind of (allowing numeric assessment) forms of similarities do exists, and why specifically – have authors used at this research area, b) how is it possible to integrate numeric assessments of similarity to plausibility, c) how to select useful data from available sources. To solve these problems authors have used particular resources and tools: a) For handling similarity authors have used structural and descriptive similarities, b) For numeric assessment of similarities authors have used P. Lorents method [1], which has already used by other authors in several cases, c) For treating allegations contained in available texts, also to identifying them, have been used.

2 Treating Situations and Developments in the Framework of Algebraic Systems Theories and Mathematical Logic 2.1

Situations as Algebraic Systems

Algebraic systems are is one appropriate for assessing situations if one is seeking clear, accurate and strict results [1]. In order to achieve mentioned objectives it is necessary to highlight and fix the following: (1) objects and facts in systems are figurative and deserve to appoint: what are the objects that will figure prominently in this situation, and worthy of observation (hereinafter referred to as system elements); (Sect. 2.1) features that require attention about figurative things in situations; (Sect. 2.2) The ways (how, what way) objects in particular situations are related. Have a settled fixation and ranked a pair of elements we have obtained an ordered pair. In this pair the first position comprises the set of objects, figurative in situation: system elements. In the second position we have the fixated set of features and relations: the signature of system. Every particular ordered pair of elements that have set of elements in first position and signature in second position are known as algebraic systems [2]. If it’s expedient is allowed to split the set of elements systems into several parts and handle them as systems with several sets of elements [10, 11]. Important Note. Everyone can decide, how and what to fixate when handling a situation. For this reason it is taken for granted that the same situation can be handled with different systems. It shouldn’t be surprising remembering the example, that in the case of natural numbers a set of all natural numbers can be handled in at least three different ways: the multiplicative semigroup of natural numbers, or the additive semigroup, or the Peano arithmetic [3].

Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts

2.2

85

Developments as Algebraic Systems

It is important to highlight and fixate with developments and situations, to study developments as systems, like handling any other thing as a system: – there are two types of elements. The first type of elements are situations. The second type of elements are stages, or times, or stages of development, etc. – relationships (I) between stages (to make clear what precedes, what follows), (II) between situations, (III) relationships between stages and situations Note. As situations, the developments also can handled with different systems. 2.3

Description of Situations, Developments and System Theories

Descriptions that are usable practically should consist of statements. The statements in the natural language can be transformed into formulas by a certain procedure [4, 5]. We can call these formulas of the corresponding system theory. Formulas that prove to be absolutely right in some framework (for example in framework of classical first, or second order predicate calculation), will constitute the elementary theory of the particular system [2]. Formulas are probably the most compact, strictest and clearest way to make statements. That might not be most convenient and customary for most of people, therefore, we agreed: On the rules in this work we expect, that in the descriptions we are limited just with statement, both for situations and developments. For example, “we have q with feature P”, “we have the x and the y, and they are in relation R between them”; “stage e and situation s are related to each other with functional relationship F” etc.). Submissions using formulas remain in the background. Submissions of statements using formulas remain in the work of the so-called second plan. Knowing that, if necessary, we can implement text transformation procedure (which translates statements in texts to formulas).

3 Treatment of Similarities of Situations and Developments with the Help of Systems and Statements. Structural and Descriptive Similarity Quantitative numeric assessments presented originate from the works of Swiss botanist and plant physiologist P. Jaccard published in 1901 with the following modifications. The Jaccard coefficient is calculated based on the proportion between the intersection and union of two finite sets [6]. The Lorents coefficient is calculated using the ratio between the subset of equalized elements and the union of the two finite sets. The numerical value of the coefficient depends on the method used for determining the equality of elements in one set with elements from the other set [3]. When treating situations and developments as systems we can rely (using similarities) on systems homomorphism (hail from algebraic system theory) [2].

86

P. Lorents et al.

Remark. The existence of isomorphism we can note as occurrence of perfect similarity. Homomorphism is also pretty good form of similarity. Mentioned forms of similarity we call as a structural similarity. Also, if we have finite systems, it is possible to make a numerical assessment to similarity, based on isomorphism or homomorphism. As a note, the following book, dedicated to this topic, will be released shortly (Lorents P. Describing and comparing situations and developments with numbers and systems. Definitions, procedures, explanations. Examples, examples, examples.). Hereby we can mark that in the case of isomorphism, the appropriate value for isomorphism, is number 1. In case of homomorphism the value of numerical assessment must be between 0 and 1 (in some cases it may be equal to 1). Theory of algebraic systems contains thoughtful theorems: Theorem 1. (See [2]). If two systems are isomorphic, then for each formula that comes from the theory of the first system, and is correct, the corresponding formula from the theory of the second system is also correct. Or: in case of isomorphic systems every sapience wisdom about first system is also relevant to the other system, and otherwise. Theorem 2. (See [2]). If the two systems are homomorphic, then for each formula that comes from the theory of the first system, and is correct, and is positive (i.e. does not contain negations or implications), then the corresponding formula from the theory of the second system is also correct. Or, in case of homomorphic systems, every positive sapience wisdom for the first system, is compliance and also relevant to the other system, but not otherwise. Remembering, that the value of the homomorphic systems similarity can be, depended on the selection of compliance, assessed with one or the other rating h (where 0  h  1). On this basis now we should seek on possibility to implement as much is possible the sapience wisdom we had about first system, apply to handling the other system. Unfortunately, it is obvious that there exists endless amount of sapience wisdom for non-trivial systems (mathematical logic aspect explains that from the existing formulas is possible to prepare and create increasingly new formulas, with the help of logical operators and quantifiers). So, it is necessary to curb ambitions. But how? There are many possibilities. Here we look at one of them: (based on the equalization of the elements of the final sets) descriptive similarity. The basis of descriptive similarity treated in this research, are descriptions of systems. Specifically – systems consisting only of relevant statements that do not duplicate each other’s content. For example, we cannot use statements like: “not only men read this text”; “we can found women who read this text”. Detection and evaluation of similarity performed in two. Therefore comes to create a list A for the statements describing the first system, also list B for the second system. Therefore comes need to create third list C, containing elements of ordered pairs. Here on the first position situated statement from list A, and on the second position is element from list B. It is important to declare, that they are very similar and equalized by descriptor decision made. If necessary we can agree to use triplets instead of pairs, containing on the third position formulation similar equal (and suitable) to statements on first and second position. We also have to agree, to mark hereinafter the number of finite set M elements with symbol E(M). In such a case we can calculate index or coefficient of descriptive similarity of first and second system: SimC(A, B) = E(C)/(E(A) + E(B) − E(C)).

Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts

87

4 Plausibility and Similarity. Algebraic Definition In current work we define the concept of the plausibility as follows: the plausibility we call every specific algebraic system, which: (I) Set of elements is formed by chosen and fixated things, we hereinafter name as assessments of plausibility or degrees of plausibility or degrees of credibility. (II) Signature is formed binary comparison relationships, chosen and fixed by as in order to compare degrees of plausibility. Remark. Here applies the principle: who decides it is responsible. Or: plausibility determinant is free in his decisions. He/her pics what to use as degrees of plausibility, also how to compare them. Therefore, we have to accept several plausibility’s, as we hopefully have accepted possible truth-values and several different (mathematical) logics. The concept of plausibility through algebraic systems is not, of course, only conceivable. See e.g. [7–9].

5 Plausibility Induced by Similarities of Situations Included in Developments We observe three developments D, D′ and D″, (I) Which stages associate in two related to each other, accordingly ∠, ∠′ and ∠″ (II) Which stages associated with situations with functional relations f, f′ and f″. In such a case we mark with writing t1∠t2 fact that in development D, the stage t1, followed by stage t2; with writing t1′∠′t2′ we mark the fact that in development t1′ is followed by stage t2′; with writing t1″∠″t2″ we mark the fact that in development D″ stage t1″ is followed by stage t2″. More, with writing f(e) we mark the situation which occurs in development D in stage e; in writing f′(e′) we mark the situation where which occurs in development D′ in the stage e′; in writing f″(e″) we mark the situation where which occurs in development D″ in the stage e″. Because stages are connected with situations with functional relations, then its possible exist only one situation in every stage of development. It does not exclude possibility, that we can meet some situations in several stages of developments. Next we have to choose and fixate (at least) one and treatable as same for evaluation of similarity’s came from developments D, D′, D″, using way of identification T. Results of evaluations calculated mentioned way, we mark in writing SimT(D1, D2), where D1, D2, different from each other, came from developments under observation {D, D′, D″}. Also we use writings SimT(SI, SII), where situations SI and SII cannot be from the same development. Next we need plausibility conducted to situations similarity. To define, that we need to choose and fixate one set (for example some numbers, suitable words, acronyms etc.) of elements which we name as above degrees of plausibility. Also we need to choose and fixate functional relationship r, which allows conduct situations from development D: to one, own and only possible degree of plausibility. For example – r(f(p)) meaning: how plausible is fact that development D comes in stage p to situation f(p). Let’s set that in development D some stage t corresponds to (in one side) to stage from development D´ and from the other site to stage t″ from development D″.

88

P. Lorents et al.

Assumption 1. Development D have has passed to stage t, there are not occurred yet any stages following to stage t and not any corresponding situations. Assumption 2. Developments D′ and D″ have passed to stages t′ and t″ corresponding to the stage t, after what they both still have stages of development (for example stages a′ and a″, in which case t′∠′a′ and t″∠″a″). Assumption 3. We assume that observed development D may proceed from last stage, which means strictly speaking, we need to form new development D*, which we get from development D, if: – We add in stages of development D at least one new stage, we call it p, in which case: • It’s necessary renew comparison relationship between stage, meaning that from relations ∠ must arise relation ∠*, and in this case must be valid t∠*p and doing so: • Between stages t and p cannot occur in any else stages (there in no stage q, which is not t or q, at the same time t∠*q∠*p). • We need to “upgrade” a functional link f, which associates the stages with situations, or more precisely – in order to form a new connection f*, a new “input-output pair” inevitably has to be added, in which the input is a new stage p and the output is some situation f*(p). It was not required that the situation corresponding to the new stage should necessarily be new (compared to the development situations D so far). Assumption 4. There exists at least one (or more) such as situation: S1, and also maybe S2, and … and maybe else several situations. Form these situations will materialize just one. It can be Sn (where 1  n  m), in which case will apply equality f*(p) = Sn. Problem. How to understand the plausibility of the fact that development D (so far) will continue precisely this way that in stage we have exactly the situation Sn. Possible Solution. We have to rely on assessments of similarity’s SimT(D, D′) and SimT(D, D″), based on development D (so far) and at least two other developments D′ and D″. We need to make it clear, is SimT(D, D′)  SimT(D, D″), or SimT(D, D ′) < SimT(D, D″). Let’s assume that SimT(D, D′)  SimT(D, D″). In this case we will find, continued development D, occurred stage p′ (from development D′) corresponding to stage p. So we calculate relevant evaluations of similarity SimT(f′(p′), S1), SimT(f′(p′), S2), …, SimT(f′(p′), Sm). – If turns out to be, that SimT(f′(p′), S1) = SimT(f′(p′), S2) = … = SimT(f′(p′), Sm), so we have equally similar situations, then it is equally plausible that development D will continue in any situation S1, …, Sm, – If turns out to be, that in case of mentioned situations two, for example Sj and Sn, SimT(f′(p′), Sj) < SimT(f′(p′), Sn), then the chance that development D will continue in stage p with situation Sn, is more plausible that development D will continue in stage p with situation Sj. Otherwise, the chance that development D will continue in stage p with situation Sj, is less plausible that development D will continue in stage p with situation Sn.

Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts

89

With other words, simplified: If development D is more similar to the development D′ and less similar to (for example) development D″, then it is more plausible, compared to other possibilities that development will continue similar to the way as was with development D′. Remark. Proposed solution for problem expressed above, probably cannot probably be proved or otherwise - refuted. It can be believed or not believed. And, as long the situation is not cleared, we have to believe something. And until the point is clearer, we believe what is stated above. We name it: The principle of plausibility for continuation of Similarly progressed developments: If development D is more similar to development D′, and less similar to development D″, then it is more plausible continuation of development D similarly to development D′. Remark 1. We deal with plausibility and comparison of its degrees, not (for example) with “modality- certainly” or maybe with “probability 1” (1 one event will defiantly occur, the event will definitely happen). Remark 2. Smaller plausibility of something does not exclude itself particular preceding of development. Remark 3. There is no restriction, what do not aloud treat plausibility relied on similarity: how plausible it is that the situation A will be followed by exactly this situation B, which is similar to situation B′, which has known followed to situation A′. We know that A and A′ were similar. Going this way have been possible for credit companies, with help of relevant IT solutions, to have necessary evaluations to answer claim to have credit [10]. Similarly made observations of plausibility for outbreak of armed conflicts, relying on similarities of previous situation “before wars”. Remark 4. There is recommendation to have some doubts for relevance to description of situations (also developments), if occurs that described situations, but with suspiciously little similarities are, time and again, followed by quite similar situations. So, we should have serious doubts, are the statements related to success/fail of IT solutions, solid enough to believe [1]. This kind of doubting is proved with mathematical logic, according to that is not correct to deduce from the implication – “if A is true, then B is true”, the implication – “if A is not true, then B is not true”. More Detail: Let us have some situations S′, S″, …, S‴, where it turns out, that all of them have the same situation S, or situations that are very similar to situation S. However, it turns out that situations S′, S″, …, S‴, descriptions are very different perhaps their numerical estimates of descriptive similarity are remarkably small. In such a case, it would probably be necessary to ask whether the descriptions of the situations S′, S″, …, S‴ are relevant at all when considering the nature of the situation S. It is by no means excluded that situations S′, S″, …, S‴, descriptions are overloaded with unnecessary details that are of no importance in the S context. In this case, the descriptions of situations S′, S″, …, S‴, should be reviewed with a “critical look” to find more suitable ones where possible.

90

P. Lorents et al.

6 Implementation of Similarity Indexes and Automatized Linguistic Analysis Assessing Situations Immediately Preceded Armed Conflicts In the following, we will give a brief overview of how it would be possible to assess, on the basis of the similarity of the existing descriptions of the situations that it is plausible that the situation described may be followed by an armed conflict. Situations are described by their temporal, geographic and geopolitical characteristics. This study covers years 1945 to 2008. The geographic area is limited to Eurasia and parts of other regions to include allies of the Republic of Estonia (mainly in NATO and the EU) and potential adversaries. We have analyzed only state players including relevant predecessor and successor states as required by the nature of the conflict (for example the invasion of Warsaw Pact countries in Czechoslovakia, or the Korean conflict). Consequently the 32 conflicts selected and studied, suggests that in situations where decision making is of vital importance and the amount of information is limited the methodology using similarity and plausibility is well suited for decision making process in order to minimize the risk of having an armed conflict. The situations immediately preceding armed conflict have been studied. Each conflict is assigned a specific set of assumptions based on the description of the situation. Conditions of equality between the assumptions have been rigorously studied and subsets of equivalent assumptions identified. These subsets were then used to calculate similarity indexes in the descriptive similarity framework. The main research subject were situations immediately preceding the armed conflict. Specifically, claims that are the subject of descriptions of these situations. The number of statements is associated with each of the conflicts were examined. For these sets of similar claims were prepared and fixed sets of asserted claims. On the basis of were calculated descriptors of descriptive similarity, clearly and strictly associated with specific ways of identification. Sources we used: Wikipedia and notes made within interviews with informed peoples from services. Because of the reason that there is endless source of books and writings about armed conflicts, we needed to limit the basis of sources with something. Therefore, we chose Wikipedia as the most common and generalize available sources with hope that chosen materials are comparable. Results. Wikipedia as source is useless for researches based on similarity, from other source of data we used, did bring out trends and keywords we can use as solid “triggers” assessing plausibility of outbreak of the crisis. Because all armed conflicts turned out, had logically the same result- one state (or union of states) attacked another one, the mathematical conclusion process must be right. Therefore, we observed the processes what preceded those conflicts. The aim was to observe and sort out words or expressions in texts, which having certain features could be the basis to assess proximity and distance of plausibility of situations and developments. Observing chosen texts, which were supported with citation and references, we tried to compare texts using substantive claims, using method created by P Lorents [1] (Table 1).

Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts

91

Table 1. Results of comparison of armed conflicts conflict

in- conflict dex Israel_Arab Confl´56 - Suez conflict 0.5 Russo-Georgian war.- Isrl_Arab war-Soviet Hungarian Revol.- Warsaw Pact Inv to Praha 0.38 Hungarian Revol.- Suez Crisis Russo-Georgian war- Russo Crimerian inv. 0.33 Russo-Georgian war-Korean Conflict Soviet_Afg.War-Russo Crimerian inv. 0.33 Warsaw Pact Inv. to Praha- Korean Confl. Korean conflict-Vietnam War 0.5 Crimenian Inv.- Suez crisis Hungarian Rev- Warsaw Pact Invasion Praha 0.26 Falkland War-Pealr Harb. Attack Russo-Georgian war- Donbass_conflict 0.33 Crimenian Inv.- Suez crisis Russo Crimerian inv.-Donbass_conflict 0.72 Hungarian Revol.-Suez crisis 1.Chechen War- 2. Chechen war 0.07 Warsaw Pact Invasion to Praha Suez crisis Falkland War-1 Chechen War 0 Warsaw Pact Inv. to Praha -Chechen war Pealr Harb. Attack-Gulf War 0.36 Warsaw Pact Inv to Praha- Pealr Harb. Atc Crimenian Inv.- 2-Chechen war 0.28 Warsaw Pact Inv to Praha- Vietnam War Ungari-2.Tsetsh 0.05 Hungarian Revol- Vietnam War Isrl_Arab war-Soviet_Soviet_Afg.War 0.14 Donbas Conflict- Suez Crisis Warsaw Pact Invasion to Praha- Gulf War 0.08 Gulf War-Hungarian Revol.

in- conflict dex 0.17 Vietnam-War.-Russo-Georgian war 0.25 Vietnam-War.-Crimenian Inv. 0.38 Gulf War -Donbas Conflict 0.08 Vietnam-War.-Donbas Conflict 0.1 Suez crisis-Gulf War 0.18 Vietnam War-Falkland War 0.07 Hungarian Rev-Russo-Georgian war 0.33 Vietnam war-Suez crisis 0.09 Vietnam War-2.Chechen war 0.25 Vietnam War-Gulf War 0.28 Russo-Georgian war-48.-Gulf War 0.18 Pealr Harb. Attack -Suez crisis 0.2 Pealr Harb. Attack- Donbas Confl. 0.2 Donbas Conflict- 1.Chechen War 0.06 Donbas Conflict- 2.Chechen War

index 0.36 0.15 0.14 0.09 0.11 0.31 0.19 0.13 0.21 0.2 0.17 0.43 0.15 0.48 0.19

We can deduct from this: (I) Available texts normally do not include substantive triggers. Obviously even not afterwards and after safe period. It would worth to compare sorted claims with disclosed security publications, for example open CIA libraries https://www.cia.gov/libr, within same chosen limitations (temporal, geographic and geopolitical characteristics). This would be voluminous work and deserves separate article (-s). (II) Obviously it is unbearable to make preciously this kind comparisons manually- preparing descriptions, claims, analysis. Figuratively we found more similarities observing mentioned texts in general. For example, we did sort out that, most of conflicts turned out at the second half of the h year – 67%; in October even 28% of the cases. Therefore, we used another point of view. For analysis of the same conflicts or situations we use automatized search engine. Working with chosen method and data occurred whole different results, using special software “Wordsmith” with the help of program we prepared “concordance list” which did bring up words and expressions connecter to our used sources. The selected words are taken from the “crisis anatomy” development table [1]. These statements describe the possible situations in the phases of crisis development. It has to be noted that, by studying military conflicts with the previous method, when phases of pre-conflict between states has been searched manually, we saw that the statements in the table of the anatomy of the crisis, have been also appeared. We did group sentences by selected keywords, and we did sort out groups which did include keywords more than 5 times. Each sentence was written in corresponding column connected with one armed conflict, including these sentences in text. Next the similarity index was calculated. Fragment of results are presented in Table 2.

92

P. Lorents et al. Table 2. Results of emerged search words. Fragment. social

x x x x x

society

x x x x x

public

x x x x x

rebellion

x x 0 x x

military

x x x x x

Political

0 0 0 x 0

leaders

x x 0 x x

leading

internal

x x x x x

demand

x x 0 x x

equipment

affairs

x 0 0 x 0

against

interventi Russo Crimerian inv. Donbas Conflict Israel_Arab Confl´56 Suez Crisis Hungarian Revolution

x x x x x

0 0 0 x 0

x 0 0 x 0

x 0 0 x x

Donbas ConflictSuez Crisis Russo Crimerian inv.-Donbass conflict Hungarian Revol.-Suez crisis Russo Crimerian inv. -Hungarian Rev. Russo Crimerian inv. - Israel_Arab Confl´56 Suez Crisis - Israel_Arab Confl´56

0.58 0.75 0.42 0.83 0.5 0.43

7 Conclusions Method of Identification would be worth’s to developing, and it would be helpful tool for professional decision makers. This would be mainly for evaluation of currents situations and systematizing of information. This is needed for sorting treated situation to claims, and descriptions in order to make justified decisions. This presume high professionality of decision maker, extensive experience and awareness of the situation. With the aim to go back in time and search more triggers on texts (describing armed conflicts) through a longer period, it would presume usage of automatized search engines. The reason is huge amount of submitted texts about thus kind of situations, which would be overwhelming task for one investigator or small search group. Difficulties of using automatized search engines is choice of search terms, words and expressions. This topic worth’s close investigation in special articles.

References 1. Lorents, P., Matsak, E., Kuuseok, A., Harik, D.: Assessing the similarity of situations and developments by using metrics. In: Intelligent Decision Technologies (KES-IDT 2017), Vilamoura, Algarve, Portugal, 21–23 June 2017, pp. 184–196. Springer, Heidelberg (2017) 2. Maltsev, A.I., Maлъцeв, A.И.: Aлгeбpaичecкиe cиcтeмы. “Hayкa”. Mocквa (1970) 3. Kleene, S.: Mathematical Logic. Dover Publications (2002) 4. Lorents, P.: Keel ja loogika (Language and logic). EBS Print (2000) 5. Matsak, E.: Dialogue system for extracting logic constructions in natural language texts. In: IC-AI 2005, pp. 791–797 (2005) 6. Jaccard, P.: Étude comparative de la distribution florale dans une portion des Alpes et des Jura. Bulletin de la Société Vaudoise des Sciences Naturelles 37, 547–579 (1901) 7. Michalski, R., Collins, A.: The logic of plausible reasoning: a core theory. Cogn. Sci. 13(1), 1–49 (1989) 8. Sigarreta, J., Ruesga, P., Rodriguez, M.: On mathematical foundations of the plausibility theory. Int. Math. Forum 2(27), 1319–1328 (2007) 9. Jakobson, G.: Extending situation modeling with inference of plausible future cyber situations. In: The 1st IEEE Conference on Cognitive Methods in Situation Awareness and Decision Support (CogSIMA 2011), Miami Beach, FL, USA (2011) 10. Matsak, E.: Credit scoring and the creation of a generic predictive model using countries’ similarities based on European values study. In: FinanceCom 2016, pp. 114–123 (2016)

Applying Systems’ Similarities to Assess the Plausibility of Armed Conflicts

93

11. Matsak, E., Lorents, P.: Decision-support systems for situation management and communication through the language of algebraic systems. In: 2012 IEEE International MultiDisciplinary Conference on Cognitive Methods in Situation Awareness and Decision Support (CogSIMA 2012): CogSima 2012, pp. 301–307. IEEE Press (2012) 12. Lorents, P., Matsak, E.: Applying time-dependent algebraic systems for describing situations. In: 2011 IEEE Conference on Cognitive Methods in Situation Awareness and Decision Support (CogSIMA 2011), Miami Beach, FL (IEEE Catalog Number: CFP11COH-CDR), pp. 25–31. IEEE Press (2011)

Local Binary Pattern and Its Variants: Application to Face Analysis Jade Liz´e1 , Vincent D´ebord`es1 , Hua Lu2 , Kidiyo Kpalma1(B) , and Joseph Ronsin1 1

Univ Rennes, INSA Rennes, CNRS, IETR - UMR 6164, 35000 Rennes, France [email protected] 2 Normal University of Shandong, Jinan, China http://www.ietr.fr

Abstract. Local Binary Pattern is a descriptor whose purpose is to summarize the local structure of the images. The goal is to be able to discriminate different images. This method has gone through a large number of changes and adjustments in different types of applications. This paper reviews various LBP methods for facial expression analysis and proposes a new set of variants. Firstly, the principle of LBP is presented and the main variants for face recognition and facial expression analysis are described. Then, new variants are proposed and finally, a comparison with the referred existing ones is made and analyzed through experiments conducted on facial recognition databases YaleB and ORL, and on facial expressions recognition database JAFFE. Keywords: LBP descriptors · Face analysis · Classification · Recognition · Facial expression analysis · Gender verification

1

Introduction

Facial recognition is one of the most popular topic in visual recognition. It becomes increasingly present in our daily life due to its wide range of applications such as security, home automation, photo identification on social networks. Like face recognition, facial expression recognition follows the same principle. As pattern recognition problems, they consist of two important parts: features extraction and classification. Feature extraction plays a crucial role in the recognition stage: the richer the extracted features, the better the classification will succeed. Over the last few decades, many features have been developed and the most popular and successful is the Local Binary Patterns (LBP). Due to its simplicity and its efficiency, a large number of variants have been proposed, focusing on various configurations such as pixel neighborhood topology, thresholding and quantification, encoding and grouping complementary features. This paper focuses on the neighborhood topology of the LBP variants dedicated to face and facial expression recognition: five new LBP-like descriptors c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 94–102, 2020. https://doi.org/10.1007/978-3-030-53187-4_11

Local Binary Pattern and Its Variants

95

are thus introduced and analysed which are Doubled Local Binary Pattern (dLBP), Reduced Divided Local Binary Pattern (RedDLBP), Median Block Local Binary Pattern (MedBLBP), Divided Block Local Binary Pattern (DBLBP) and Divided Median Block Local Binary Pattern (DMedBLBP). The proposed descriptors are compared with some existing variants on facial expression and face recognition tasks, respectively, over JAFFE [1] and, YaleB [2] and ORL [3] databases. Moreover, the noise robustness of the methods is evaluated. The rest of the paper is as follows. Section 2 introduces the general concepts of the LBP process followed by the presentation of various existing variants adapted to facial recognition and facial expression recognition. Section 3 presents the new proposed variants. Experiments and the comparison of the proposals with the existing approaches are presented in Sect. 4. Finally, Sect. 5 discusses the study and concludes the paper.

2 2.1

Local Binary Pattern Features and Variants Basics of Local Binary Pattern

The original LBP was proposed by Ojala in 1996 [4]. It describes the pixels of an image by using a 3 × 3 neighborhood around each pixel. The central pixel is then subtracted from its eight neighbors. If the resulting value is negative, the pixel is set to ‘0’, otherwise it is set to ‘1’ which concatenate together to give an 8-bits code corresponding to an integer ranging from 0 to 255. The original LBP is defined by Eq. (1) and based on the principle of Fig. 1.

Fig. 1. Basic LBP operator

LBP8,1 =

7 

S(gp − gc )2p

(1)

p=0

where, S(x) = 0 if x < 0 or 1 if x ≥ 0 and gp corresponds to the value of the pth neighbor pixel and gc the central pixel. With the basic LBP operator, dominant features with a large scale structure cannot be captured due to its small neighborhood. Hence, a variant is introduced [4] which extends the neighborhood to (P, R) corresponding to P sampling points symmetrically arranged on a circle of radius R as illustrated on Fig. 2.

96

J. Liz´e et al.

Fig. 2. Examples of the extended LBP’s with different (P, R)

One of the greatest advantages of LBPP,R is its rotation invariance. Examples of images processed with LBP8,1 are shown in Fig. 3.

a)

b) LBP8,1 of a)

c)

d) LBP8,1 of c)

Fig. 3. Examples of LBP images

In order to reduce the size of the original LBPP,R descriptor, Ojala et al. introduced the concept of uniform patterns [5] which represent 90% of the patterns. A pattern is uniform if in its binary code (considered circular), the number of transitions between ‘0’ and ‘1’ is less than three. For example, 01110000 (2 transitions) is uniform whereas 11001001 (4 transitions). 2.2

Local Binary Pattern Variants for Face Analysis

Several other variants have been developed which improve the performance, focusing on different aspects: e.g. MB-LBP [6], MQLBP [7] and many others. In this subsection, we review 4 LBP variants specifically adapted for facial analysis. 2.2.1 Multi-Block Local Binary Pattern (MB-LBP) It was proposed in 2007 by L. Zhang et al. [6]. This method uses the LBP principle but applies to the mean value of the surrounding blocks. Blocks of size 2 × 3 are used in [8]: the mean value of each block is compared to that of the central block and assigned ‘0’ if it is lower; otherwise, it is assigned ‘1’. By using mean values, MB-LBP is more robust to noise and more stable for face analysis than LBP. 2.2.2 Median Local Binary Pattern (MBP) Proposed by Hafiane et al. [9], MBP compares each pixel of 3 × 3 neighborhood with the median value of the block and assigns ‘0’ it is lower and ‘1’ otherwise. It includes the central pixel into the code so that to generate a 9-bits code.

Local Binary Pattern and Its Variants

97

2.2.3 Divided Local Binary Pattern (DLBP) Hua et al. [10] have proposed the Divided Local Binary Pattern (DLBP) which breaks down the LBP into two parts: one for even indices of the neighborhood and the other for odd indices. This reduces the range of the data by reducing the code length from 8 to 4 bits. 2.2.4 Multi-quantized Local Binary Patterns (MQLBP) Proposed by Patel et al. [7], MQLBP extends Local Ternary Patterns (LTP) [11] with the main idea to split the pixels difference into 2 L levels instead of two levels. This splitting is done depending on a set of 2L − 1 thresholds. It utilizes both the sign and magnitude of the difference between the central pixel and the surrounding ones. Each quantized level is encoded separately to generate multiple local binary patterns. The basic LBP corresponding to the case of L = 1.

3

Proposed Variants

After analysing the strengths and weaknesses of existing methods, new variants are proposed. First, we observed that capturing more global features could give better information about the pixel environment. So the neighborhood is extended by enlarging the radius R or by adding another radius. To capture more global features, a group of pixels is used instead of a single pixel: this way, one can reduce the noise effect too. The proposed new variants are introduced hereafter. 3.1

Doubled Local Binary Pattern (d-LBP)

It extends the neighborhood of the basic LBP by using two rings of radius 1 and 3. Thus d-LBP uses two neighborhoods of 8 pixels as illustrated in Fig. 4(a). This leads to two LBP codes and two local histograms which will be concatenated. 3.2

Reduced Divided Local Binary Pattern (RedDLBP)

This variant uses the radius of 2 and the neighborhood of 6 pixels. Then, the neighborhood is cut down into two groups as can be seen on Fig. 4(b): this gives two 3-bits codes leading to 2 histograms which concatenate to give the descriptor.

Fig. 4. Operating principle of: a) d-LBP and b) RedDLBP

98

3.3

J. Liz´e et al.

Median Block Local Binary Pattern (MedBLBP)

Inspired by the Multi-Block Local Binary Pattern (MB-LBP) [6], it uses the median values of the surrounding blocks instead of mean values. This helps discard abnormal values and reduce the influence of noise. 3.4

Divided Block Local Binary Pattern (DBLBP) and Divided Median Block Local Binary Pattern (DMedBLBP)

DBLBP and DMedBLBP are derived from a combination of MB-LBP [6] and DLBP [10] operators. DBLBP exploits a block of size 3 × 3 pixels around the central pixel and 8 surrounding blocks of size 3 × 3 too. Then, the mean values of the blocks are used instead of the pixel values. Finally, the binary code is cut down into two parts to generate two values like for DLBP [10]. DMedBLBP operates exactly as DBLBP, except that it uses the median value instead of the mean value of the block.

4 4.1

Experiments Databases

JAFFE: this dataset consists of 213 frontal black and white photos of 10 posed Japanese women [1]. They were photographed with 7 different facial expressions: 6 basic facial expressions (happiness, sadness, fear, anger, surprise, disgust) and a neutral expression, see Fig. 5 for some examples. This publicly available database is used for facial expression recognition.

Fig. 5. Some sample images from JAFFE database

YaleB: this database [2] includes 5850 of face images of 10 human subjects in 9 poses and 65 illumination conditions. The photos were taken in laboratory controlled lighting conditions. This database is used for the evaluation of face recognition methods under variable lighting conditions. ORL: it consists of 10 different images of 40 distinct subjects representing 7 facial expressions under different conditions: open/closed eyes, w/ and w/o smiling and w/ and w/o glasses. All the images were taken with a dark homogeneous background with the subjects in frontal position with some side movement [3]. It is used for the evaluation of face recognition under variable lighting.

Local Binary Pattern and Its Variants

4.2

99

Experimental Setup

This section describes the methodology used to evaluate the performance of our proposed methods against existing methods. To facilitate the execution of our experiments, we coded a Matlab application, named LBP Studio. It simplifies the implementation and testing of new LBP methods. Indeed, one only needs to edit a single file appropriately to generate the target LBP code. The developed tool comprises SVM as the classification engine. The user can exploit the friendly graphical interface to setup the experiments as illustrated on Fig. 6.

Fig. 6. Graphical interface of LBP Studio

First, the input images are divided into local blocks after the image has undergone the selected LBP operation. Then, a histogram of the generated LBP values is built for each block. Finally, these histograms are concatenated together to provide the image descriptor. Experimental Evaluation: experiments are conducted to evaluate the proposed new variants and to compare their performance with those of the existing variants. To do this, each class of the database is randomly divided into a training set and testing set. The recognition rate is then computed as: Recognition rate =

N o. of images classif ied correctly T otal no. of test images

(2)

To take the variability of the performance into account, this procedure is repeated 100 times on the JAFFE database and 30 times on the YaleB and ORL databases. The comparison is then based on the average rate of all iterations. Implementation Parameters of LBP Descriptors: the basic LBP with the parameters P = 8, R = 1 and then R = 2 are tested. The MBP variant with

100

J. Liz´e et al.

P = 8 and R = 1, and the MBLBP, MedBLBP, DBLBP, DMedBLBP variants with blocks of size 3 × 3 are considered. For DLBP and RedDLBP, two layers of P = 4 and P = 3 are used respectively, at a radius of R = 1 and R = 2. Finally, the d-LBP variant with two layers of P = 8 neighbors and radius R1 = 1 and R2 = 3 is implemented. 4.3

Results

General Results: Table 1 shows the recognition rates of proposed and existing operators on JAFFE [1] and, YaleB [2] and ORL [3] databases, for facial expressions and face recognition, respectively. Based on Table 1, one can see the effectiveness of the proposed variants.

Table 1. Comparison of average recognition rates on JAFFE, YaleB and ORLdatabases. Red color indicates the highest rate, cyan the 2nd and blue the 3rd . LBP variant

Recognition rate (%) Expression Face JAFFE YaleB ORL

LBP8,1 LBP8,2 MBP MB-LBP DLBP MQLBP

83.73 85.37 71.94 92.53 91.24 83.89

98.13 98.36 90.58 98.83 91.76 91.90

98.35 98.72 96.98 98.62 98.43 98.52

d-LBP RedDLBP MedBLBP DBLBP DMedBLBP

91.47 93.43 91.22 88.43 89.90

98.40 97.57 92.64 98.28 90.43

98.60 97.88 98.08 99.07 98.73

The RedDLBP descriptor stands out by obtaining the best expression recognition rate and also has high performance in face recognition. These results show that the proposed variants perform well both for face and facial expression recognition. For the tests on the YaleB database, the d-LBP stands out by obtaining the 2nd best score while DBLBP performs the best on ORL database. Robustness Against Noise: robustness to noise is evaluated. To achieve this, JAFFE database is noised with Gaussian noise or with salt-and-pepper noise of standard deviation ranging from 0 to 0.5. Examples of image with Gaussian noise and salt-and-pepper noise are shown Fig. 7.

Local Binary Pattern and Its Variants

101

The results are summarized in Table 2 as the average recognition rates. Based on these results, we can observe that the basic LBP8,2 is very sensitive to noise. With noise range of 0 to 0.5, the recognition rate has decreased from 85.37% to 57.40% and to 63.64% for Gaussian noise and salt-and-pepper noise respectively. This table shows that the proposed DBLBP is the most robust to Gaussian noise. The DMedBLBP also performs well faced to Gaussian noise and performs the best against the salt-and-pepper noise. The DBLBP and MedLBP variants also stand out, obtaining the 2nd and the 3rd highest average recognition rate, behind DMedBLBP, respectively on salt and pepper noise.

Fig. 7. Original image (left) and image with a Gaussian noise (middle) and Salt-andPepper noise (right) of 0.2 standard deviation

Table 2. Average recognition rate for noised JAFFE database. Red color indicates the highest rate, cyan the 2nd and blue the 3rd .

5

Method

Gaussian noise (0–0.5) Gaussian noise Salt & pepper noise

LBP8,2 MBP MB-LBP DLBP MQLBP

57.40 50.00 65.45 43.37 49.86

63.64 50.00 71.65 59.76 69.70

d-LBP RedDLBP MedBLBP DBLBP DMedBLBP

48.83 47.53 55.19 66.75 62.46

61.19 66.62 71.95 75.20 81.17

Conclusion

After analysing several variants of the LBP features, this paper has introduced five new ones. The new proposed variants exploit some interesting properties that can help improve the performance. This consists of:

102

• • • • •

J. Liz´e et al.

taking more pixels in the neighborhood to catch large scale objects, capturing more distant pixels, using blocks of pixels instead of pixels, to reduce the effect of the noise, using the mean value and the median value of the surrounding blocks, splitting the binary code.

Evaluated over JAFFE, YaleB and ORL databases, the proposed variants perform satisfactorily compared to the state of the art variants. These results demonstrate that taking more global information into account reduces the noise effect while keeping a good description. The proposed variants are also evaluated upon Feret’s database for the challenging task of gender recognition [7]. The preliminary performance sounds satisfactory and appeals for further investigation for improvement. In order to confirm the efficiency of the proposed LBP-like variants, in face, facial expression and gender recognition, future work includes extensive experiments over more databases and against various noise types.

References 1. Lyons, M., Akamatsu, S., Kamachi, M., Gyoba, J.: Coding facial expressions with gabor wavelets. In: Proceedings Third IEEE International Conference on Automatic Face and Gesture Recognition, pp. 200–205, April 1998 2. Georghiades, A., Belhumeur, P., Kriegman, D.: From few to many: illumination cone models for face recognition under variable lighting and pose. IEEE Trans. Pattern Anal. Mach. Intell. 23(6), 643–660 (2001) 3. Samaria, F.S., Harter, A.C.: Parameterisation of a stochastic model for human face identification. In: Proceedings of 1994 IEEE Workshop on Applications of Computer Vision, pp. 138–142, December 1994 4. Ojala, T., Pietik¨ ainen, M., Harwood, D.: A comparative study of texture measures with classification based on featured distributions. Pattern Recogn. 29(1), 51–59 (1996) 5. Ojala, T., Pietik¨ ainen, M., M¨ aenp¨ aa ¨, T.: Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Trans. Pattern Anal. Mach. Intell. 24, 971–987 (2002) 6. Zhang, L., Chu, R., Xiang, S., Liao, S., Li, S.Z.: Face detection based on multi-block LBP representation, pp. 11–18. Springer, Heidelberg (2007) 7. Patel, B., Maheshwari, R., Balasubramanian, R.: Multi-quantized local binary patterns for facial gender classification. Comput. Electr. Eng. 54(Suppl. C), 271–284 (2016) 8. Yoanna, M., Heydi, V., Yenisel, P., Edel, G.: Dissimilarity representations based on multi-block LBP for face detection. In: Progress in Pattern Recognition, vol. 7441, pp. 106–113, September 2012 9. Hafiane, A., Seetharaman, G., Zavidovique, B.: Median binary pattern for textures classification. In: Kamel, M., Campilho, A. (eds.) Image Analysis and Recognition, pp. 387–398. Springer, Heidelberg (2007) 10. Lu, H., Yang, M., Ben, X., Zhang, P.: Divided local binary pattern (DLBP) features description method for facial expression recognition. J. Inf. Comput. Sci. 11, 2425– 2433 (2014) 11. Tan, X., Triggs, B.: Enhanced local texture feature sets for face recognition under difficult lighting conditions. IEEE Trans. Image Process. 19, 1635–1650 (2010)

Reducing LBP Features for Facial Identification and Expression Recognition Joseph Ronsin1(&), Kidiyo Kpalma1, and Hua Lu2 1

2

Univ Rennes, INSA Rennes, IETR VAADER - UMR 6164, 35000 Rennes, France [email protected] School of Physics and Electronics, Shandong Normal University, Jinan 250358, China

Abstract. The LBP (Local Binary Pattern) texture descriptor has demonstrated its superiority in several image applications: texture characterization, facial identification and macro-expression recognition. Featuring by LBP characterizes image by its local structures, observing its micro patterns and building a histogram. Each observed pixel is featured and then encoded into one byte. All these codes constitute bins for the histogram. For an efficient classification, encoded bytes can be divided into uniform and non-uniform codes. In standard applications, only uniform codes are used leading to 59 codes. The present work proposes an additional process for the reduction of these codes. The proposal is developed and comparatively evaluated with success to classical LBP. Experimental evaluations are performed on 2 different databases for facial identification and macro-expression recognition respectively, and this for different reduction of code length. Though for macro-expression recognition the proposed features can give lower but comparable performance with the traditional LBP, for facial identification they perform very well and keep excellent efficiency. This approach can be extended to most part of LBP variants while keeping the simplicity of LBP. Keywords: Facial identification Textural approach


LBP
Macro-expression recognition


1 Introduction Since the middle of the 90’s, the LBP (Local Binary Pattern) texture descriptor has been the object of a lot of research. Now there is a strong interest with rapid advances in image technology and large databases of digital images for facial identification, macro-expression recognition, and even micro-expression spotting and recognition. Moreover immediate applications exist: in security, in business, in medicine etc. [1]. LBP was firstly presented by Ojala and Harwood [2], and was proved a powerful mean of texture description where each pixel is characterized observing its local neighborhood and featuring its micro patterns. Observation of micro patterns for a pixel in a window leads to extraction of a byte as feature. For classification of the image content, the distribution of these bytes is classically observed in a histogram. The © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 103–111, 2020. https://doi.org/10.1007/978-3-030-53187-4_12

104

J. Ronsin et al.

resulting volume of these features or codes can be a problem at different levels: consequent volume on databases, problematic dimension (called curse of dimensionality) of features lowering the accuracies of the results, computing cost, power consuming for embedded applications. There is interest in a reduction of the number of features or bins inside the histograms [3]. This paper focuses on the reduction of the LBP codes and proposes a solution extensible to its numerous variants [4]. After this introduction, the adopted presentation for sections is following: featuring image by LBP, LBP encoding, reducing LBP codes, experimental evaluation context, facial identification and macroexpression identification results, and a final conclusion.

2 Featuring Image by LBP LBP belongs to textural approaches and was proved a powerful mean of texture description where each pixel is characterized observing differences with its local neighborhood in a window and featuring these differences as micro patterns. There, image is considered as monochrome and its content presenting only the face of one person facing the camera, moreover face extraction and registration are considered as ready performed. The extent of the window or observed neighborhood of this textural approach can be from 3*3, 5*5 and even more but there, for simplification of the presentation of our results, only a 3*3 window with its 8 neighbors will be considered, as this is often proposed for the classical application of LBP. Process is following, featuring one still image operates dividing N*N facial image into blocks, and its characterization is obtained from the different concatenated histograms inside these blocks. So, each image content consists of a whole face, see Fig. 1. For its characterization image is divided into N*N blocks, here N = 6 on Fig. 1. Inside each block, observing the features from its pixels, a characterization is obtained in a histogram. For each pixel in a window 3*3, micro patterns constitute the component parts or bins of the histogram. Inside each window, featuring process operates from central pixel, considering its local neighborhood and observing signs of differences with each of its 8 neighbors.

… … Fig. 1. Featuring image from its blocks and concatenating histograms

Reducing LBP Features for Facial Identification

105

There, classical LBP will be only presented and tested for evaluation of our proposal. Now considering more deeply LBP algorithm, for each pixel of a central window, local differences with its 8 neighbors (or elementary gradients), are computed, and their 8 respective signs positioned and concatenated together in a byte, finally describing an encoded micro pattern as a feature for this central pixel, see Fig. 2. Starting from central pixel gc, for neighboring pixel gi, is computed gi – gc and then, corresponding sign of this difference gives a bit constitutive of a circular binary code of 8 bits or the byte encoding the pattern formed by the 8 neighbors. Figure 3 illustrates encoding 3*3 by LBP. The corresponding signs of differences are encoded with the convention of 0 if negative, and 1 if positive or null. On this figure, the circular binary code is presented on one byte, with a starting bit corresponding to neighboring pixel g1 and consecutive bits developed in a clockwise direction.

g1

g2

g3

g8

gc

g4

g7

g6

g5

Fig. 2. Window 3*3.

86 142 140

80 95 104 block 3*3

75 60

0 1

0

0 0

58

1

1 signs

0

0 0 0 0 0 1 1 1

=>

encoded pattern for window

Fig. 3. LBP encoding micro pattern.

For each block, one can build a histogram of the values derived from the LBP codes. Thus for the whole image, divided into N*N blocks, a descriptor is generated by concatenating the N*N histograms. LBP encoding. The LBP method is very efficient due to its easy-to-compute feature extraction operation and simple matching strategy. Formalism of algorithm for encoding the LBP code for central pixel is following: with gc central pixel and gi, with i = 1 to 8, its neighboring pixels and s the threshold function equivalent to sign function: LBP ¼

X8 i¼1

sðgi  gc Þ  2i

with the thresholding function s(x) defined as:

ð1Þ

106

J. Ronsin et al.

 sðxÞ ¼

0; x\0 1; x  0

ð2Þ

For featuring the gc pixel, the 8 values of s(x) on the 8 neighbors, constitute a pattern encoded into one byte which is the local binary pattern or LBP. LBP operator gives 256 output values or patterns. If all these patterns are adopted, representation is rather heavy and some no-frequent patterns appear with a low frequency, and so cannot be considered to be statistically robust for classification. Therefore, for a better efficiency, it could be interesting to use only a subset of these 256 patterns. Often are considered 2 types of LBP patterns: uniform and non-uniform patterns [2]. Uniform ones are those with a spatial homogeneity of 1’s (or 0’s) inside the codes. This will correspond to continuous sets of 1’s (or 0’s) forming one chain inside the byte. One possible classical pattern reduction is the restriction to patterns called uniform local binary patterns (ULBP) [2]. Their definition is following according to a uniformity measure U (pattern), defined as the number of spatial transitions (bitwise 0/1 changes) in the connected binary string in a pattern: U ðLBPÞ ¼ jsðg8  gcÞ  sðg1  gcÞj þ

X8 i¼2

jsðgi  gc Þsðgi1  gc Þj

ð3Þ

Patterns keep a same characteristic as LBP output values if there are at most two changes from 0 to 1 or 1 to 0 in the circular binary code. Instead of the 256 possible patterns of LBP, there are only 58 possible uniform patterns in computing U(LBP). Remaining non-uniform patterns with uniformity measure greater than 2 are accumulated into a single bin, resulting to a histogram of 59 bins. In this study, we will consider this LBP, as it is classically done, as basic LBP, and take it as the reference method in our experimentations for the assessment of the performance.

3 Reducing Patterns Number The objective of our proposal targets the reduction of the number of LBP codes. Our solution applies to uniform patterns: the 58 codes issued from the observation of relation (3). Each code, as a circular binary string, represents a pattern constituted of a contiguous chain of 1’s surrounded by 0’s. For a better clarity of our presentation we will now call these codes as “patterns”, so making difference with the word “code” that will be reserved to final encoding of patterns once they have been reduced. Inside LBP uniform patterns (basic LBP codes), there are chains of 1’s corresponding to surrounding pixels higher than the central pixel. One chain of 1’s synthesizes an elementary pattern with its length and position, its observation authorizes possible simplifications. Our hypothesis is that for patterns with long chains and close positions, one can force them to the same shorter pattern because their lengths can be reduced by suppressing, for example, one 1 of the chain, and this on its extremity, without introduction of a consequent distortion of the initial encoded pattern. This authorizes fusion between patterns and so reduction of their number. It must be recalled that the 1’s are issued from a simple thresholding by central pixel and the peripheral 1’s of the

Reducing LBP Features for Facial Identification

107

chain constitute a border where probably some values could be very close to the used threshold (presently gc), and therefore set to 0. For example, considering 2 patterns of length K with close positions, suppressing of one 1 can lead to the same final reduced pattern of length K − 1. See upper part Fig. 4 where patterns A and B are reduced into pattern C with the closest possible position to A and B.

A 8 7 6 5 4 3 2 1

B 8 7 6 5 4 3 2 1

C

8 7 6 5 4 3 2 1

D 8 7 6 5 4 3 2 1

E 8 7 6 5 4 3 2 1

F 8 7 6 5 4 3 2 1

G 8 7 6 5 4 3 2 1

H 8 7 6 5 4 3 2 1

Fig. 4. Upper part: reducing pattern pair A and B (length 6) into pattern C (length 5); lower part: reducing pattern quadruplet D, E, F and G (length 5) into pattern H (length 4).

Once a pattern has been reduced by suppressing one 1 it keeps its uniform property and moreover becomes identical to some other uniform patterns of lesser length (D, E, F and G on Fig. 4). So final reduced patterns could be “fused” into other ready existing patterns or consider as a new “reduced” pattern, and then could be encoded by a new dedicated code. On one hand, Fig. 4 with A and B patterns of length 6, once reduced in a same pattern C with length 5, could be encoded by the same code allocated to pattern E. On the other hand, the reduced pattern C could also receive a new code independent of those of patterns of length 6. So, for these both processing’s a reduction of the number of patterns is effective. With the first processing the 2 fused patterns A and B have completely disappeared using a ready used code (code E), while with second solution, the 2 patterns are reduced into a new one which will subsist with its own code. So are offered different possibilities of reduction. For a given set of patterns of same length, it can be grouped together: pairs, triplets, quadruplets of 1’s. Moreover it is possible to go further in this reduction by additional suppressions of more 1’s on each pattern. Another reduction can be obtained going further with a fusion on groups of patterns of lesser lengths: after fusion, taking pairs of patterns on set of patterns with length 6, then observing the lower part of Fig. 4 a new fusion can be performed on patterns with length 5, taking now quadruplet of patterns and leading to patterns of

108

J. Ronsin et al.

length 4, and so on. To summarize, 8 patterns with length K can fuse and completely disappear into patterns with length K − 1 using their codes, or can be reduced and give other patterns subsisting with new codes. Moreover on patterns of a same length can be performed, at a time, fusion on some groups and reduction on other ones.

4 Context for Experimentations Two Databases are respectively used for face identification and macro-expression recognition. The database ORL contains a set of faces used for facial identification. There are 40 distinct subjects with 10 different images leading to a total of 400 images. The size of each image is 92  112, 8-bit grey levels. Other JAFFE database evaluates facial expressions recognition where the different classes are macro facial expressions. It consists of 10 Japanese female models posed for 7 facial expressions and 3 times for each expression so we have 210 images. Figure 5 illustrates the 7 expressions of the images in the database. Evaluation is performed using only one component (red component) of each image (so assimilated as monochrome image).

disgust

surprise

happiness

fear

anger

sadness

neutral

Fig. 5. JAFFE database facial expressions.

Evaluations: From these databases, our experimentations are driven performing supervised classification by SVM. Facial identification or macro-expression recognition performances are observed at different rates for learning on the databases. All the evaluations on both databases were performed resizing images at 120*120 pixels. Facial Identification is evaluated on ORL face database, performing classification of 40 distinct subjects at different percentages of learning on this database of 400 images. The percentages of images used for learning on each subject before a classification on the remaining part of database have been respectively set to: 10, 30, 50, 70 and 90%. Each identification on the database was observed and evaluated for a given percentage of learning, and with 12 runs (or random realizations). To facilitate the comparison of performance between the different approaches, the obtained classification accuracies on different learning percentages are summarized with 2 means values. The first value Central mean or the mean on the central percentages of learning into the interval [0 100] i.e. computing the mean at 30, 50 and 70%, reflecting the reached performance in the context of a classical learning where generally the learning set integrates a reasonable part (30, 50 or 70%) of the global database. While the second value Global mean covers the performance of all the percentages for which a classification can be applied (10, 30, 50, 70 and 90%).

Reducing LBP Features for Facial Identification

109

Central mean = mean recognition rate of 30, 50, 70% learning rates. Global mean = mean recognition rate of 10, 30, 50, 70, 90% learning rates. Facial Macro-expression recognition is similarly evaluated on JAFFE database, performing SVM classification of the faces into 7 different macro-expressions and this at different percentages of learning on this database. Macro-expressions are: disgust, surprise, happiness, fear, anger, neutral and sadness. The presented results are summarized with 2 previous means values of classifications: Central and Global. Image Blocking: experimentations have been performed for different number of blocks 4*4, 5*5 and 6*6. The last parameterization being the classical adopted blocking for LBP on selected databases.

5 Comparative Results Classifications are comparatively evaluated: classical LBP and reduced LBP noted LBPF. LBP operates with 59 bins, while LBPF restricts codes to essential 1’s: i.e. with a number of bins corresponding to a power of 2: 32 bins for LBPF_32, and 16 bins for LBPF_16. Another presentation is also proposed with the strongest reduction to 10 bins: LBPF_10. The performances of the proposed approaches are assessed from the 2 defined metrics: Central mean and Global mean. The experimentations are conducted with different blocking sizes: 4*4, 5*5 and 6*6. Experimental results follow in 2 subparagraphs dedicated respectively to facial identification and macro-expression recognition. Facial Identification Table 1 summarizes obtained accuracies for facial identification. LBP is the classical approach, while LBPF is the LBP approach with a reduction of its bins. Bin number is suffixed to names of approaches. Table 1. Facial identification classification accuracies on ORL database

block

LBP 59 LBPF 32 LBPF 16 LBPF 10

4*4 Central Global mean mean 96,98 96,91 96,13 95,31

93,01 92,93 91,96 91,17

ACCURACIES 5*5 Central Global mean mean 96,69 96,64 96,07 95,46

92,52 92,32 91,67 91,25

Central mean 96,38 96,18 95,44 94,98

6*6 Data Global reduction mean 91,83 91,60 91,14 90,54

46% 73% 83%

110

J. Ronsin et al.

Observed accuracies show an excellent preservation of performance for LBPF, compared to LBP, with only 32 bins or 16 bins. So, reduction from 59 bins to 32, or 46% of featuring volume, leads for Central mean with accuracies around 96%, to a lowering of 0.07 at 4*4 which is the optimal blocking. And, yet for 4*4 blocking, with a reduction to 16, or 73% of featuring volume, this leads for Central mean at a lowering of less than 1, presently 0.85. Considering LBPF_10, the mean difference for any observed blockings with the Central mean of LBP is 1.4 which can be considered as very acceptable for a data reduction of 83%. Macro-expression Recognition Observing results for macro-expression recognition in Table 2, the gaps with LBP approach make the results of our proposal more modest and therefore of questionable interest. For LBPF_32, with all observed blockings, the mean difference of Central means of LBP is 4.17. So, reduction leads to a rather consequent degradation of the accuracies but this reduction is also consequent.

Table 2. Macro-expression recognition classification accuracies on JAFFE

block

LBP 59 LBPF 32 LBPF 16 LBPF 10

4*4 Central Global mean mean 75,60 71,47 60,79 59,60

73,15 69,42 59,32 57,93

ACCURACIES 5*5 Central Global mean mean 75,76 72,00 65,11 67,06

73,26 69,78 62,83 65,29

Central mean

6*6 Global mean

80,54 75,92 68,51 68,11

77,57 72,78 66,87 65,74

Data reduction 46% 73% 83%

6 Conclusion A procedure is proposed for the reduction of LBP codes. For facial identification, the results show the efficiency of our proposal, and that, whatever the blocking of the image or the percentage of learning set. Performance on macro-expression recognition is of lesser interest. Obtained resulting reduction of data volume can be consequent. The complexity of LBPF is the same as for LBP; it’s a simple look up table behind the encoded byte of the LBP pattern. As a result, LBPF easily can replace LBP which nowadays is often considered as the simplest most efficient basic solution for main applications with facial featuring. Moreover LBPF can be substituted into LBP variants and also their different extensions (size, color, multi resolution, 3D…). It must be recalled that for some applications, size reduction of features can preempt on classification accuracies. With LBPF, when the simplicity is required for embedded applications, the number of bins can be reduced continuously.

Reducing LBP Features for Facial Identification

111

References 1. Mirmehdi, M., Xie, X., Suris, J. (eds.): Handbook of Texture Analysis. Imperial College Press, London (2008) 2. Ojala, T., Pietikäinen, M., Harwood, D.: A comparative study of texture measures with classification based on featured distributions. Pattern Recogn. 29(1), 51–59 (1996) 3. Lu, H., Yang, M., Ben, X., Zhang, P.: Divided Local Binary Pattern (DLBP) features description method for facial expression recognition. J. Inf. Comput. Sci. 11, 2425–2433 (2014) 4. Wang, Y., See, J., Phan, R.C.-W., Oh, Y.-H.: Efficient spatio-temporal local binary patterns for spontaneous facial micro-expression recognition, 19 May 2015. https://doi.org/10.1371/ journal.pone.0124674

Video Retrieval Using Query Images and CNN Features Imane Hachchane1(&), Abdelmajid Badri1, Aïcha Sahel1, and Yassine Ruichek2 1

Laboratoire d’Electronique, Energie, Automatique & Traitement de l’Information (EEA&TI), Faculté des Sciences et Techniques Mohammedia, Université Hassan II Casablanca, Mohammedia, Morocco [email protected], [email protected], [email protected] 2 IRTES-Laboratoire SET, Université de Technologie de Belfort Montbéliard, Belfort, France [email protected]

Abstract. We address the problem of image-to-video face retrieval. Given a query image of a person, the aim is to retrieve videos of that same person. The methods proposed so far are based on hand-crafted features. In this work we investigate the use of an off-the-shelf object detection network as a feature extractor by building an image-to-video face retrieval pipeline composed of an offline feature extractor and online filtering and re-ranking steps that use the object proposals learned by a Region Proposal Network (RPN) and their associated representations taken from a CNN. Moreover we study the relevance of features from a fine-tuned network. In addition to that we explore the use of face detection before extracting the features and we test the impact of different similarity metrics. The results obtained are promising. Keywords: Image processing
Classification
Object recognition
CNN Faster R- CNN
Image-to-video instance retrieval
Face retrieval
Video retrieval




1 Introduction Visual search applications, especially image retrieval, have recently gained a vast popularity due to the explosion of visual content. This increase led to a proliferation of visual search applications like image-based retrieval or more specifically instance search. This is used to retrieve images of a specific object from large databases, by comparing a query image against a database of other images. This issue has been widely used in product recognition [1] and in building identification [2]. This work addresses a variant of this problem, the task of image-to-video instance retrieval which is the task of identifying a video collection from a specific instance in a static image. Image-to-video retrieval is an asymmetric problem. Images only contain static information but videos have much richer visual information, like optical flow. Due to the lack of temporal information, standard techniques used for extracting video © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 112–120, 2020. https://doi.org/10.1007/978-3-030-53187-4_13

Video Retrieval Using Query Images and CNN Features

113

descriptors [3–6] cannot be directly used on static images. On the other hand, standard features for image retrieval [7–10] can be applied to video data by processing each frame as an independent image. Temporal information is usually compressed either by reducing the number of local features or by encoding multiple frames into a single global representation. The methods based on the aggregation of local features perform better than methods that encode several frames into a single global representation at the expense of a lower data compression and an increased search time. In large-scale datasets, memory requirements and search time are crucial, so techniques based on global representations are preferred. Traditionally, image-to-video retrieval methods [11–13] are based on hand-crafted features (SIFT [14], BRIEF [15], etc.) and not much effort has been put so far into the adaptation of deep learning techniques, such as convolutional neural networks (CNN). CNNs trained with large amounts of data can learn features generic enough to be used to solve tasks for which the network has not been trained [4]. For image retrieval, in particular, many works in the literature [9, 16] have adopted solutions based on standard features extracted from a pretrained CNN for image classification [17], achieving encouraging performances. In this paper we try to fill this gap by exploring the relevance of on-the-shelf and fine-tuned features of an object detection CNN for image-to-video face retrieval.

2 Related Work In general, visual search or retrieval is an issue of indexing and querying visual data, and can be categorized depending on the type of queries and databases we are using. Most work in visual search focuses on image to image retrieval, we use a query image and a database of images [18, 19]. In video to video retrieval, we use a query video and a database of videos which can be used in event retrieval [20]. Another variant is video to image retrieval, used in augmented reality, it refers to using a query video to search a database of images [21]. This work focus on image to video retrieval were we search a database of videos using query images. But more precisely, we are focusing on instance face retrieval. Face retrieval remains a challenging task because conventional image retrieval approaches, such as bag of words, are difficult to adapt to the face domain [22]. This is mainly the result of using traditional key point detection based descriptors like SIFT, that have a tendency to fail due to the smooth face surface. Early works, using a pretrained image classification convolutional neural network as a feature extractor, showed that fully connected layers for image retrieval were more suitable [7]. Razavian et al. [23] improved the results by combining fully connected layers extracted from different image sub-matches. Later on, new works found that convolutional layers significantly outperform fully connected layers during image recovery tasks [2, 23]. Many CNN-based object detection pipelines have been proposed, but we are more interested in the latest ones. The faster R-CNN [24] created by Ren et al. uses a Region Proposal Network (RPN) to remove the dependence of object proposals on older CNN object detection systems. In Faster R-CNN, RPN shares features with the objectdetection network in [25] to simultaneously learn prominent object propositions and

114

I. Hachchane et al.

their associated class probabilities. Although the Faster R-CNN is designed for generic object detection, Jiang et al. [26] Demonstrated that it can achieve impressive face detection performance especially when retrained on a suitable face detection training set [27]. In this work we exploit the features of a state of the art pre-trained object detection CNN Faster R-CNN. We use his end-to-end object detection architecture to extract global and local convolutional features in a single forward pass and test their relevance for image-to-video face retrieval.

3 Methodology 3.1

CNN-Based Representations

We explore the relevance of using CNN features for image-to-video face retrieval. Each query instance is defined by a bounding box above the query image. For both the query images and videos frames, we use the features extracted from Faster R-CNN pretrained models [24] as our global and local features. Faster R-CNN has an RPN that provides the locations in the image with a higher probability of having an object, and a classifier that labels each of those object proposals as one of the classes in the learning dataset [2]. We will extract a compact image representation built from the activations of a convolutional layer in a CNN [2, 28]. Faster R-CNN is faster on a global and local scale. We build a global frame descriptor by overlooking all the layers that work with object proposals and extracting the features from the last convolutional layer. Considering the extracted activations of a convolution layer for a frame, we group the activations of each filter response to create a frame descriptor with the same dimension as the number of filters in the convolution layer. Both max and sum pooling strategies are considered and compared [27]. 3.2

Video Retrieval

This section describes the three ranking strategies we used: Filtering Step. We create image descriptors for query and videos frames. At testing time, the descriptor of the query is compared to all items in the database, which are then ranked according to a similarity measure. At this stage, the entire frame is considered as a query. Spatial Re-ranking. After the filtering step, the N upper elements are analyzed locally and re-ranked. Query Expansion (QE). We average the frame descriptors of the N higher elements of the first ranking with a query descriptor to carry out a new search. 3.3

Fine-Tuning Faster R-CNN

Fine tuning the Faster R-CNN network allows us to obtain features specific to face retrieval and should help improve the performance of spatial analysis and re-ranking.

Video Retrieval Using Query Images and CNN Features

115

To achieve this, we choose to fine-tune Faster R-CNN to detect the query faces to be retrieved by our system. The resulting networks will be used to extract better local and global representations, and will be used to perform spatial reranking.

4 Experiments 4.1

Datasets Exploited

We evaluate our methodologies using the following datasets: • YouTube Celebrities Face Tracking and Recognition Data (Y-Celeb) [29]: The dataset contains 1910 sequences of 47 subjects. All videos are encoded in MPEG4 at 25 fps rate. • YouTube Faces Database [30]: The data set contains 3,425 videos of 1,595 different people. All the videos were downloaded from YouTube. An average of 2.15 videos are available for each subject. The shortest clip duration is 48 frames, the longest clip is 6,070 frames, and the average length of a video clip is 181.3 frames. The datasets used to fine-tune the network: • FERET [31]: 3528 images, including 55 Query images. A framing box surrounding the target face is provided for query images. • FACES94 [32]: 2809 images, including 55 Query images. A framing box surrounding the target face is provided for query images. • FaceScrub [33]: 55127 images. 4.2

Experimental Setup

We use the VGG16 architecture of Faster R-CNN to extract the global and local features because previous works [2, 27] showed that deeper networks achieve better performances. The global descriptors are extracted from the last convolution layer “conv5_3” and are of dimension 512. The local features are grouped from the Faster RCNN RoI clustering layer. All experiments were performed on a Nvidia GTX GPU. 4.3

Off-the-Shelf Faster R-CNN Features

We evaluate the use of Faster R-CNN features for face image to video face retrieval. We experimented with different similarity metrics. The results were similar and close, but overall cosine similarity performs better. Table 1 shows an example of our results when using features from an on the shelf network with VGG16 architecture trained on Pascal VOC dataset. We carried out a comparative study of the sum and max-pooling strategies of the image-wise and region-wise descriptors. Table 2 summarizes most of our results. According to our experiments, sum-pooling gives better performance than maxpooling. It also shows the performance of Faster R- CNN, with a VGG16 architecture, trained on two different datasets (Pascal VOC and COCO). VGG16 trained on COCO performed better because the dataset is bigger and more diverse. Moreover, it presents

116

I. Hachchane et al.

Table 1. Mean Average Precision (Map) of pretrained Faster R-CNN models with VGG16 architectures on Pascal VOC dataset using different similarity measures using Y-Celeb dataset. Similarity metric Pooling Cosine Max Sum Manhattan Max Sum Euclidian Max Sum CityBlock Max Sum L1 Max Sum L2 Max Sum

Ranking 0.888 0.915 0.900 0.905 0.888 0.915 0.900 0.905 0.900 0.905 0.888 0.915

Reranking 0.860 0.846 0.869 0.841 0.860 0.846 0.869 0.841 0.869 0.841 0.860 0.846

QE 0.550 0.600 0.570 0.428 0.550 0.600 0.570 0.578 0.570 0.578 0.550 0.603

the impact of spatial reranking and query expansion. Using the global features of Faster R-CNN on their own without any reranking strategy give the best results. Spatial reranking & QE had no positive impact on the results. 4.4

Fine-Tuning Faster R-CNN

We evaluate the impact of fine-tuning a pre-trained network on recovery performance with the query objects to retrieve. We chose to refine the model VGG16 Faster R-CNN, pre-trained with the objects of Pascal VOC, with two datasets. The first network was refined using FERET and Faces94 datasets. Because of their smaller sizes we combine them to create one bigger dataset. We modify the output layer in the network to return 422 class probabilities (269 people in the FERET dataset plus 152 people in the Faces94 dataset, plus one additional class for the background) and their corresponding bounding box coordinates [27]. This new refined network will be called VGG(F-F). The second network was refined using the FaceScrub dataset. We modify the output layer in the network to return 530 class probabilities (530 people, plus one additional class for the background) and their corresponding bounding box coordinates. Our second refined network will be called VGG(F-S) [27]. We kept the Faster R-CNN original parameters described in [18], but due to the smaller number of training samples we decreased the number of iterations from 80,000 to 20,000. We use the refined networks of the tuning strategy (VGG(F-S) & VGG(F-F)) on all datasets to extract image and region descriptors and perform a face retrieval. Those results are also presented in Table 2. The refined features slightly exceeded the raw features in the spatial reranking and the QE stages. But still, the global features of Faster R-CNN from VGG16 trained on COCO used without any reranking strategy gives the best results.

Video Retrieval Using Query Images and CNN Features

117

Table 2. Mean Average Precision (MAP) of pre-trained Faster R-CNN models with VGG16 architectures trained with Pascal VOC, Microsoft COCO, FaceScrub or Feret & Faces94 images, respectively. With a comparison between sum and max pooling strategies. When indicated, QE is applied with M = 5. Network VGG16 (Pascal VOC) VGG16 (Microsoft COCO) VGG16 (FaceScrub) VGG16 (Feret & Faces94)

4.5

Pooling Y-Celeb Ranking Max 0.888 Sum 0.915 Max 0.911 Sum 0.926 Max 0.809 Sum 0.917 Max 0.915 Sum 0.924

Reranking 0.860 0.846 0.888 0.807 0.777 0.843 0.874 0.899

QE 0.550 0.600 0.522 0.512 0.457 0.578 0.554 0.621

YouTube faces database Ranking Reranking QE 0.892 0.877 0.882 0.897 0.886 0.891 0.892 0.878 0.889 0.903 0.882 0.896 0.848 0.834 0.838 0.882 0.873 0.874 0.894 0.884 0.887 0.896 0.892 0.893

Face Detection

We evaluate the impact of using a face detection algorithm on our datasets and queries before using Faster R-CNN for feature extraction. Table 3 presents a comparison between the results obtained on the Y-Celeb dataset, with and without face detection. As we can see face detection did not improve the results. As before, using the similarity metric on the raw Faster R-CNN features provided the best results.

Table 3. Mean Average Precision (MAP) of pre-trained Faster R-CNN models with VGG16 architectures trained with Pascal VOC, Microsoft COCO, FaceScrub or Feret&Faces94 images, respectively. With a comparison between sum and max pooling strategies. When indicated, QE is applied with M = 5. Network VGG16 (Pascal VOC) VGG16 (Microsoft COCO) VGG16 (FaceScrub) VGG16 (Feret & Faces94)

Pooling Y-Celeb Ranking Max 0.888 Sum 0.915 Max 0.911 Sum 0.926 Max 0.809 Sum 0.917 Max 0.915 Sum 0.924

Reranking 0.860 0.846 0.888 0.807 0.777 0.843 0.874 0.899

QE 0.550 0.600 0.522 0.512 0.457 0.578 0.554 0.621

Y-Celeb Ranking 0.574 0.618 0.622 0.705 0.477 0.635 0.666 0.715

+ Faces detection Reranking QE 0.516 0.542 0.486 0.511 0.574 0.617 0.538 0.551 0.423 0.450 0.509 0.519 0.656 0.682 0.612 0.646

118

4.6

I. Hachchane et al.

Comparison

Finally, we compare our results with other techniques applied on Faster R-CNN features. Table 4 presents the results using fisher vector (FV) and bag of visual word (BOVW). The raw faster R-CNN features outperformed all the other techniques. Table 4. Comparison with other techniques. Results provided as mAP. Method Faster R-CNN features Faster R-CNN features + FV Faster R-CNN features + BOVW

Y-Celeb 0.926 0.097 0.032

YouTube faces database 0.903 0.006 0.001

5 Conclusion This article explores the use of features from an object detection CNN for image-tovideo face retrieval. It uses Faster R-CNN features as global and local descriptors. We have shown that the common similarity metric give similar results. We also found that sum-pooling performs better than max-pooling in this case, and contrary to our previous work [27] fine tuning does not improve the results. In general we found that applying the similarity measure on the CNN feature gave the best results. In future work we will work on reducing the feature extraction time.

References 1. Tsai, S., Chen, D., Chandrasekhar, V., et al.: Mobile product recognition. In: Proceedings of the International Conference on Multimedia – MM 2010, p. 1587 (2015) 2. Salvador, A., Giro-I-Nieto, X., Marques, F., Satoh, S.: Faster R-CNN features for instance search. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, pp. 394–401 (2016) 3. Filgueiras De Araujo, A.: Large-scale video retrieval using image queries. A dissertation submitted to the Department of Electrical Engineering and the Committee on Graduate Studies of Stanford University in partial fulfillment of the requirements for the Degree of Doctor of Philosophy (2016) 4. Ng, J.Y.-H., Hausknecht, M., Vijayanarasimhan, S., Vinyals, O., Monga, R., Toderici, G.: Beyond short snippets: deep networks for video classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4694–4702 (2015) 5. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. In: Advances in Neural Information Processing Systems, pp. 568–576 (2014) 6. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3D convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 4489–4497 (2015) 7. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: European Conference on Computer Vision, pp. 584–599. Springer, Cham, September 2014

Video Retrieval Using Query Images and CNN Features

119

8. Kalantidis, Y., Mellina, C., Osindero, S.: Cross-dimensional weighting for aggregated deep convolutional features. In: European Conference on Computer Vision, pp. 685–701. Springer, Cham (2016) 9. Razavian, A.S., Sullivan, J., Carlsson, S., Maki, A.: Visual instance retrieval with deep convolutional networks. ITE Trans. Media Technol. Appl. 4(3), 251–258 (2016) 10. Wu, L., Wang, Y., Ge, Z., Hu, Q., Li, X.: Structured deep hashing with convolutional neural networks for fast person re-identification. Comput. Vis. Image Underst. 167, 63–73 (2018) 11. Arandjelovic, R., Zisserman, A.: Three things everyone should know to improve object retrieval. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp. 2911–2918 (2012) 12. Pont-Tuset, J., Arbelaez, P., Barron, J.T., Marques, F., Malik, J.: Multiscale combinatorial grouping for image segmentation and object proposal generation. IEEE Trans. Pattern Anal. Mach. Intell. 39(1), 128–140 (2016) 13. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3431–3440 (2015) 14. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis. 60(2), 91–110 (2004) 15. Calonder, M., Lepetit, V., Strecha, C., Fua, P.: Brief: binary robust independent elementary features. In: European Conference on Computer Vision, pp. 778–792. Springer, Heidelberg, September 2010 16. Araujo, A., Girod, B.: Large-scale video retrieval using image queries. IEEE Trans. Circuits Syst. Video Technol. 28(6), 1406–1420 (2017) 17. De Oliveira Barra, G., Lux, M., Giro-I-Nieto, X.: Large scale content-based video retrieval with LIvRE. In: 2016 14th International Workshop on Content-Based Multimedia Indexing (CBMI), pp. 1–4. IEEE, June 2016 18. Zheng, L., Yang, Y., Tian, Q.: SIFT meets CNN: a decade survey of instance retrieval. IEEE Trans. Pattern Anal. Mach. Intell. 40(5), 1224–1244 (2017) 19. Zhang, S., Yang, M., Cour, T., Yu, K., Metaxas, D.N.: Query specific rank fusion for image retrieval. IEEE Trans. Pattern Anal. Mach. Intell. 37(4), 803–815 (2014) 20. Poullot, S., Tsukatani, S., Phuong Nguyen, A., Jégou, H., Satoh, S.: Temporal matching kernel with explicit feature maps. In: Proceedings of the 23rd ACM International Conference on Multimedia, pp. 381–390. ACM, October 2015 21. Chen, D.M., Girod, B.: A hybrid mobile visual search system with compact global signatures. IEEE Trans. Multimedia 17(7), 1019–1030 (2015) 22. Herrmann, C., Beyerer, J.: Fast face recognition by using an inverted index. In: Image Processing: Machine Vision Applications VIII, vol. 9405, p. 940507. International Society for Optics and Photonics, February 2015 23. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: CNN features off-the-shelf: an astounding baseline for recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 806–813 (2014) 24. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. In: Advances in Neural Information Processing Systems, pp. 91– 99 (2015) 25. Girshick, R.: Fast R-CNN. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 1440–1448 (2015) 26. Jiang, H., Learned-Miller, E.: Face detection with the faster R-CNN. In: 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017), pp. 650– 657. IEEE, May 2017

120

I. Hachchane et al.

27. Hachchane, I., Badri, A., Sahel, A., Ruichek, Y.: New faster R-CNN neuronal approach for face retrieval. In: International Conference on Advanced Information Technology, Services and Systems, pp. 113–120. Springer, Cham, October 2018 28. Tolias, G., Sicre, R., Jégou, H.: Particular object retrieval with integral max-pooling of CNN activations. arXiv preprint arXiv:1511.05879 (2015) 29. Kim, M., Kumar, S., Pavlovic, V., Rowley, H.: Face tracking and recognition with visual constraints in real-world videos. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–8. IEEE, June 2018 30. Wolf, L., Hassner, T., Maoz, I.: Face recognition in unconstrained videos with matched background similarity, pp. 529–534. IEEE (2011) 31. Phillips, P.J., Wechsler, H., Huang, J., Rauss, P.J.: The FERET database and evaluation procedure for face-recognition algorithms. Image Vis. Comput. 16(5), 295–306 (1998) 32. Hond, D., Spacek, L.: Distinctive descriptions for face processing. In: Proceedings of the 8th British Machine Vision Conference BMVC97, Colchester, England, pp. 320–329 (1997) 33. Ng, H.-W., Winkler, S.: A data-driven approach to cleaning large face datasets. In: 2014 IEEE International Conference on Image Processing (ICIP), pp. 343–347. IEEE, October 2014

CUDA Accelerating of Fractal Texture Features for a Neuro-morphological Image Segmentation Approach Khalid Salhi1(&), El Miloud Jaara1, and Mohammed Talibi Alaoui2 1

Lari Laboratory, Faculty of Sciences Oujda, Mohammed First University, Oujda, Morocco [email protected] 2 SIA Laboratory, FST, Sidi Mohammed Ben Abdellah University, Fez, Morocco

Abstract. Image segmentation is one of the main tasks for many computer vision problems. In this paper, a GPU acceleration for a Fractal features extraction method is proposed, followed by our neuro-morphological approach that will allow to segment the images based on the Fractal texture features. In the first step, we use the CUDA environment on an NVIDIA GPU to compute the Fractal features in parallel for each pixel of our image, this makes it possible to optimize the extraction phase before starting the image segmentation by using our approach which is divided into two main steps. Firstly, we train a Kohonen self-organized Map (KSOM) using the extracted features. In the final step, we use our watershed method to extract the modals regions from the KSOM, these regions define the final regions found in the segmented image. To highlight the effectiveness of our parallel implementation, the performance results of the GPU extraction method are compared to his sequential counterpart based on CPU. In addition, the segmentation rate of the proposed approach is compared to the K-means results. Keywords: Fractal texture features
Parallel image segmentation Watershed
Kohonen neural networks


CUDA


1 Introduction Texture is one of the most used information sources in several image analysis applications such as pattern recognition, multimedia indexing, compression and image segmentation. Despite the fact that image segmentation is an easy and trivial task for our visual system, it is the part that poses the most problems and is the most difficult to automate. Many methods have been developed and can be roughly grouped into three families of approaches: region-based, contour-based and pixel-based approaches [1–3]. In our previous study we used a pixel-based approach [4, 5], using the Fractal dimension of each pixel as a texture information. The use of combination of neuronal concept and morphological technique allow us to exploit the extracted textural information to classify each pixel from the image. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 121–128, 2020. https://doi.org/10.1007/978-3-030-53187-4_14

122

K. Salhi et al.

One of the most important problems that confront our approach is the computing time of the feature extraction phase, to accelerate this phase, we present in the second section of our paper a GPU based parallel implementation, applied on NVIDIA hardware using the CUDA interface. In the third section we propose a segmentation approach based on combination between neuronal Kohonen map, followed by proposed morphological watershed application. The last section of this paper is devoted to present the comparison between the sequential and accelerated implementation in term of computation time, this section also covers the segmentation rate of the proposed neuro-morphological approach by comparing them with the experimental results of kmeans.

2 Fractal Features Extraction Method In 1982, Fractal geometry saw the light of existence by Mandelbrot to describe and analyze some complex and irregular natural phenomena [6], this concept has many fields of application, including image analysis. Generally, when it come to image analysis application, fractal geometry is used by means of the fractal dimension (FD). In this study, we have chosen to work with the differential box counting method [7], as it can be computed automatically and can be applied to patterns either with or without self-similarity. The steps used by this counting method begins by dividing the image space into boxes with different sizes r, the next step consist in computing the probability N ðr Þ as the difference between the maximum and minimum gray levels for each box. Afterwards the fractal dimension can be estimated using the equation: FD ¼ lim

r!0

ln½N ðr Þ lnð1=r Þ

ð1Þ

The sequential algorithm that allows us to extract the Fractal features of each pixel from an Image I is presented in this form: 1. For various r 2 [0,1] a. Divide Wði;jÞ into ð1=r Þ2 boxes b. Divide the range of intensities ½0; 255 into 1=r levels numbered ½1; 1=r  c. For each box bðp; qÞ 2 Wði;jÞ do: (1) l ¼ minimumðbðp; qÞÞ (2) k ¼ maximumðbðp; qÞÞ (3) np;q ðP rÞ ¼ l  k þ 1 d. N ðr Þ ¼ p;q np;q ðr Þ 2. Do line-fit of N ðr Þ and lnð1=r Þ. 3. The fractal dimension is obtained by linear regression of this line-fit. The application of this algorithm on our original image I1 allow us to extract only one feature from each pixel, to enrich the texture information extracted from the image, this method is applied to four other images derived from the original:

CUDA Accelerating of Fractal Texture Features

123

• high gray valued image (I2 ): 

I1 ði; jÞ  L1 ; if I1 ði; jÞ [ L1 0; otherwise

ð2Þ

255  L2 ; if I1 ði; jÞ [ ð255  L2 Þ I1 ði; jÞ; otherwise

ð3Þ

I2 ði; jÞ ¼ • low gray valued image (I3 ):  I3 ði; jÞ ¼

• horizontally smoothed image (I4 ): I4 ði; jÞ ¼

1 Xx I ði; j þ kÞ k¼x 1 2x þ 1

ð4Þ

1 Xx I ði þ k; jÞ k¼x 1 2x þ 1

ð5Þ

• vertically smoothed image (I5 Þ: I5 ði; jÞ ¼

Finally, for each pixel I ði; jÞ, we have a Fractal feature vector Xq ¼ ff1 ; f2 ; f3 ; f4 ; f5 g where fk is the fractal dimension of the window MWk ði; jÞ from the image Ik .

3 Parallel Implementation of Features Extraction Method 3.1

CUDA Architecture

Initially created for graphics-intensive applications such as 3D software and video games, the Graphical Processor Unit (GPU) has become over time a powerful tool for other resource-intensive tasks unrelated to graphics like scientific calculations, analyzing data and computer vision. In this study, we choose to work with solution proposed by NVIDIA GPU named CUDA, which is an Application Programmable Interface (API) created for programmers, this interface is designed to support various application programming interfaces and languages like C and C++, that facilitates its integration [8, 9]. As shown in Fig. 1(b) the NVIDIA GPU architecture is organized into a set of Streaming Multiprocessors (SM), each SM contains several cores called CUDA cores. Unlike CPU cores that allows to execute different instructions in parallel, CUDA cores can only the same instruction simultaneously on different data. The functions performed on the GPU are called kernels and are executed by multiple threads in parallel, as shown in the Fig. 1(a), each thread group is organized in a threads block, which in turn is grouped in blocks grid, all thread in one block can be synchronized and share a type of memory called shared memory, in the execution phase CUDA manages to dispatch each block to a Streaming Multiprocessors.

124

K. Salhi et al.

Fig. 1. CUDA architecture: (a) organizations of blocks and threads, (b) streaming multiprocessors.

3.2

CUDA Implementation

In order to exploit the CUDA architecture for fractal feature extraction phase, we assign a thread for each image pixel, which means for an image I with size N  M, we execute in parallel a kernel on N threads grouped in M thread blocks, for each thread with index t id in a block with index b id will calculate the five fractal feature, extracted from the sliding window MWt id;b id , we present the proposed algorithm in this form: • • • •

Read the image I in the host memory. Transfer the image I in the device memory. Create M blocks with N thread each. For each pixel I ðt id; b id Þ do in parallel: – Extract the sliding window MWt id;b id – Compute the 5 derived images for the window MWt id;b id – Extract the 5 Fractal features, so each pixel I ðt id; b id Þ have a vector Xq ¼ ff1 ; f2 ; f3 ; f4 ; f5 g • Transfer the vectors computed in parallel X1 ; X2 ; . . .; XNM from device memory to host memory.

CUDA Accelerating of Fractal Texture Features

125

4 Neuro-morphological Segmentation Approach The extracted Fractal features allows us to segment the image using our unsupervised clustering approach [4]. This approach can be divided on two steps: firstly, we use the extracted features in the learning phase of KSOM. Then, we apply a morphological watershed to a 2D projection of the Kohonen map, that allow us to segment our image without having to fix in advance the number of regions. 4.1

Kohonen Self-organized Map   Let C ¼ X1 ; X2 ; . . .. . .; XQ be a sample of Q observations in a N-dimensional space  T such as Xq ¼ xq;1 ; xq;2 ; . . .. . .; xq;N ; q ¼ 1; 2; . . .; Q. The Kohonen network is made of two layers (Fig. 2), the first one is the input layer which is composed of N attributes of the observation Xq . The output layer is composed of M neural units regularly distributed on the map which elaborates prototypes of the data [10].

Fig. 2. KSOM architecture

The neural units of the first layer are connected to the units of the second layer. Each interconnection from an input unit j to an output unit m has a weight Wm;j . That  means that each output unit m has a corresponding weight vector Wm ¼ Wm;1 ; Wm;2 ; . . .. . .; Wm;N T . The followed steps define the learning algorithm of our KSOM: 1. Initializing the weights of the neurons in the Kohonen map layer by giving them random values with small variation. 2. Presenting an input Fractal feature vector Xq . 3. Finding the winning node m* using the Euclidean distance between the vector Xq and the nodes of the output layer. 4. Updating the weights Wi winner node, as well as those around him, using the Eq. 6. 5. Decreasing the size of the neighborhood area winners’ nodes.

126

K. Salhi et al.

6. Decreasing the learning coefficient aðtÞ. 7. Going back to Step 2, or else complete learning. 

  Wm ðtÞ ¼ Wm ðt  1Þ þ aðtÞ: Xq Wm ðt  1Þ if m ¼ m Wm ðtÞ ¼ Wm ðt  1Þ þ aðtÞ:hm ðtÞ: Xq  Wm ðt  1Þ if m 2 Vðm ; r ðtÞÞ

ð6Þ

To perform a 2D projection of the Kohonen map, we use the nonparametric Parzen estimate [11] defined by: pðWm Þ ¼

4.2

  Wm  X q 1 XQ 1 : X q1 V ½DðW Þ Q hQ m

ð7Þ

Morphological Modal Regions Extraction Approach

To automate the extraction of modal regions form this 2D projection, we use our proposed watershed technique [5]. We first apply a numerical morphological opening to the projection. Then, a consecutive homotopic thinning operation followed by sequential pruning operations until idempotence are performed to seek the different regions in the estimation, that will allow us to assign each pixel from the original image to an appropriate region (Fig. 3).

Fig. 3. Watershed transformation: (a) PDF projection, (b) extracted modal regions

CUDA Accelerating of Fractal Texture Features

127

5 Results In order to assess the performance of the proposed GPU acceleration of Fractal features extraction, we have tested both sequential and parallel methods on multiple images with different size (Table 1). For this experimental test, we have chosen to work with a 11  11 pixels sliding window. And with regard to the material we use an intel i7 7700k CPU with 4.20 GHz, and a Nvidia GTX 1060 6G GPU (Table 2). Despite that the CPU used is a top of the line hardware, the results show that the GPU implementation it allowed us to gain up to 30 in computing time. Finally, to validate our segmentation approach, we use two images generated by combining several textures from Brodatz album [12] Fig. 4, and we compare the proposed approach results with k-means results, using a misclassification error measure, based on the number of misclassified pixels:

Number of misclassified pixels Clustering accuracyð%Þ ¼ Total number of pixels



Fig. 4. Test images: (a) Image 1, (b) Image 2, (c) Image 3.

Table 1. CPU and GPU Fractal features extraction time comparison. Image size 100  100 200  200 300  300 400  400 500  500 600  600

CPU time 00 m 17 s 01 m 01 s 02 m 25 s 04 m 33 s 07 m 01 s 10 m 08 s

GPU time 01.01 s 03.19 s 06.25 s 10.45 s 14.87 s 19.74 s

Speed-up 16.7 19.1 23.2 26.1 28.3 30.7

ð8Þ

128

K. Salhi et al. Table 2. Segmentation rate results. Image Proposed neuro-morphological approach K-means Image 1 97.00% 96.31% Image 2 87.04% 87.20% Image 3 94.66% 94.31%

6 Conclusion In this study, we present a GPU acceleration method for the Fractal texture features extraction step in our unsupervised segmentation approach, which has been done with a CUDA environment in NVIDIA GPU hardware. The computation time results show that the proposed parallel method allowed us to optimize the Fractal features extraction phase, and the segmentation rate results of the proposed neuro-morphological approach are very encouraging comparing it with K-Means, even if it requires no a priori knowledge of the number of regions in our image. In our future research, we are looking to implement the GPU acceleration on other texture features extraction methods. Besides, we search to parallelize the neuro-morphological segmentation phase of the proposed approach.

References 1. Salem, M.A.M., Atef, A., Salah, A., Shams, M.: Recent survey on medical image segmentation. In: Computer Vision: Concepts, Methodologies, Tools, and Applications, pp. 129–169 (2018) 2. Dhanachandra, N., Chanu, Y.J.: A survey on image segmentation methods using clustering techniques. Eur. J. Eng. Res. Sci. 2(1), 15–20 (2017) 3. Zaitoun, N.M., Aqel, M.J.: Survey on image segmentation techniques. Procedia Comput. Sci. 65, 797–806 (2015) 4. Salhi, K., Jaara, E.M., Alaoui, M.T.: Texture image segmentation approach based on neural networks. Int. J. Recent Contrib. Eng. Sci. 6(1), 19–32 (2018) 5. Salhi, K., Jaara, E.M., Alaoui, M.T., Alaoui, Y.T.: Color-texture image clustering based on neuro-morphological approach. IAENG Int. J. Comput. Sci. 46(1), 134–140 (2019) 6. Mandelbrot, B.B.: The Fractal Geometry of Nature/Revised and Enlarged Edition. WH Freeman and Co., New York (1983). 495 p. 7. Chaudhuri, B.B., Sarkar, N.: Texture segmentation using fractal dimension. IEEE Trans. Pattern Anal. Mach. Intell. 17(1), 72–77 (1995) 8. Nvidia, C. U. D. A. Programming guide (2010) 9. Sanders, J., Kandrot, E.: CUDA by Example: An Introduction to General-Purpose GPU Programming. Addison-Wesley Professional, Boston (2010) 10. Kohonen, T.: Essentials of the self-organizing map. Neural Netw. 37, 52–65 (2013) 11. Parzen, E.: On estimation of a probability density function and mode. Ann. Math. Stat. 33 (3), 1065–1076 (1962) 12. Brodatz, P.: Textures: A Photographic Album for Artists and Designers. Dover Pubns, Mineola (1966)

Efficient Mapping Method for Elliptic Curve Cryptosystems Based on PWLCM Salma Bendaoud1(&), Fatima Amounas1, and El Hassan El Kinani2 1

2

R.O.I Group, Computer Sciences Department, Faculty of Sciences and Technics, Moulay Ismaïl University, Errachidia, Morocco [email protected], [email protected] M.M.S.C Group, ENSAM, Moulay Ismaïl University, Meknes, Morocco [email protected]

Abstract. The security of digital images is an important issue that has been receiving considerable attention in the recent past. Different image encryption techniques have been proposed in the literature. Among the fundamental theories in the number theory, we can find the elliptic curve (EC) which is widely used to construct cryptographic primitives. This paper investigates the security of image encryption schemes based on elliptic curves cryptography (ECC) and Chaos theory. More precisely, in this paper, we propose a new image encryption scheme that utilize a new mapping method based on Piecewise Linear Chaotic Map (PWLCM) that converts each pixel of plain image into a point on an elliptic curve. Encryption and decryption process are given in details. After applying encryption, security analysis is performed to show that our scheme cannot only achieve good encryption, but also resist the statistical attacks. Keywords: Image
Encryption
Decryption
Elliptic curve
Piecewise linear chaotic map

1 Introduction As long as humans live in a networked society, where data are wildly exchanged, security of confidential digital information becomes mandatory, specially image communication, and thus the encryption image has received much attention by cryptographers. Digital images have two intrinsic properties: bulk data capacity and strong correlation among adjacent pixels. For this reason, conventional cipher algorithms such as Data Encryption Standard (DES), Advanced Encryption Standard (AES), RivestShamir-Adleman (RSA) are not suitable for image encryption [1, 2]. Therefore, plenty of interesting theories, such as elliptic curve cryptography [3–7] and chaos algorithm [8–11], have been proposed and applied in the image encryption. Several contributions have been suggested in the literature that deals with image encryption. For instance, Xingyuan Wang et al. [11] proposed a novel and effective image encryption algorithm based on chaos and DNA encoding. The algorithm uses PWLCM to generate the key image and a specific DNA rules or operations decided by chaos to obtain cipher image. The experiment indicates that the proposed algorithm has fast encryption speed, high security and key sensitivity by including chaos mechanism. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 129–136, 2020. https://doi.org/10.1007/978-3-030-53187-4_15

130

S. Bendaoud et al.

Moreover, Kamlesh et al. [13] had proposed in 2010 an encryption scheme based on ECC with knapsack for image encryption. The analysis shows that the scheme had less encryption time, consumes low power and more reliable. Results also show that ECC applications have high security. Bidyut Jyoti Saha et al. [12] proposed a digital Image Encryption using ECC and DES with Chaotic Key Generator in 2013. The scheme encrypts the original image using DES with the help of key sequence which is generated from a chaotic key generator with the help of heron map and then mapped the encrypted image in the points of the elliptic curve. In 2017 Jiahui Wu et al. [7] proposed new asymmetric image encryption based on chaotic systems and elliptic curve ElGamal scheme. This algorithm had proposed for the advantages that the key groups and the number of keys are very small, and the key transmission mode is relatively simple and secure on what is already demonstrated by the security performance. In this paper a cryptosystem is proposed based on ECC to encrypt the image. The most important phase in encryption using ECC is mapping a message to a point on the curve and converts the encrypted point to the current message type. F. Amounas et al. in [14] proposed a new mapping technique based on matrix properties for alphanumeric characters. In this methodology, the message with length n divided by 3 and arranged in a matrix of M3r. After that, the message matrix will be multiplied by non-singular matrix of A33 that is A = ±1 and the result Q = M. A is a matrix of mapped points and encryption will be done on elements of matrix Q. To decode the decrypted points to the message, matrix Q is multiplied by inverse of A, M = A−1Q. In [15] Two mapping methods are proposed static and dynamic: the first one is a one-to-one it is very weak. But in the second method, for one character there are different options too choose as a point. In this case by having the mapped point, it is very difficult to find the corresponded character of the plain message. In this paper a new mapping scheme based on PWLCM is proposed to convert pixels of the input image to a point on an elliptic curve. This mapping scheme is on a map table which is created and used for both the encryption and decryption process. The rest of this paper is arranged as follows. In Sect. 2, preliminary works are presented. In Sect. 3, the proposed image encryption algorithm is described in detail. Security analyses are drawn in Sect. 4. conclusions are drawn in to the final Sect. 5.

2 A Review on Mathematical Basics 2.1

Elliptic Curve Arithmetic

An elliptic curve E over finite field Fp is defined by a cubic equation of the form: y2 ¼ x3 þ ax þ b ðmod p)

ð1Þ

Where a, b 2 Fp and satisfy 4a3 + 27b2 6¼ 0 (mod p), then the elliptic curve is noted by Ep (a, b). The set E(Fp) consist of all point (x, y) where x, y 2 Fp which satisfy the Eq. (1) along with a point at infinity noted X [16]. The addition operation for two points over an elliptic group follows specific rules indicated below

Efficient Mapping Method for Elliptic Curve Cryptosystems

131

1. X + X = X. 2. P + X = P for all values of P = (x, y) 2 E. Namely, E has X as its identity element. 3. P + Q = X for all values of P = (x, y) 2 E and Q = (x, −y) 2 E. In other words, the inverse of (x, y) is simply (x, −y). 4. Adding two distinct points: let P = (x1, y1) 2 E and Q = (x2, y2) 2 E with x1 6¼ x2, Adding P and Q giving a point R: R = P + Q = (x3, y3) where


x 3 ¼ a2  x 1  x 2 y 3 ¼ að x 1  x 3 Þ  y 1

with a ¼ ðy2  y1 Þ=ðx2  x1 Þ

5. Multiplication operation: kP ¼ P þ . . . þ P ðk copiesÞ where k is an integer:

2.2

Piecewise Linear Chaotic Map

Piecewise linear chaotic map (PWLCM) has gained more and more attention in chaos research recently for its simplicity in representation, efficiency in implementation, as well as good dynamical behavior [17]. The simplest PWLCM is defined by: xn þ 1 ¼ FP ðxn Þ ¼

8 < xn =p;

0\xn \p p  xn \0:5 FP ð1  xn Þ; 0:5  x\1

ðxn pÞ ; : ð0:5pÞ

ð2Þ

Where xn 2 (0, 1) and p 2 (0, 0.5). In our experiment we assign p = 0.25678900.

3 Proposed Algorithm In this section, we introduce a new elliptic curve cryptosystem for image based on PWLCM. The proposed algorithm requires the generation of the mapping table which contains all possible points and encrypts the plain image with ECC process. 3.1

Encryption Algorithm

The process of the proposed cryptosystem consists in the following steps: Step 1. Chooses a random integer k and compute the key Q Q ¼ k PB

132

S. Bendaoud et al.

where E is an elliptic curve defined over a finite field Fp and P is a point on E (Fp) that P has prime order. The curve E and the point P are publicly known. B chooses a random integer kB, and publishes the point PB = kB P. Step 2. Generate the map table by: • calculate the initial condition x0 using: x0 ¼ mod ðxQ þ yQ ; N Þ=ðN  1Þ

with

  N ¼ #E Fp

• Iterate the PWLCM according to Eq. (1) using the initial condition x0 and store the results in an array S. • Rearrange the array S in the form of a table by grouping S in 256 groups. Each group has length(S)/256 members. The row indexes are start from 0 and end with 255. Each row stands for a pixel intensity value but for same values there are multiple points. Step 3. Starting from the first pixel in the plain image, the corresponded point with the intensity value in the map table is mapped to this pixel and continues to the last pixel. Step 4. Encrypt the mapping points using ECC technique. Step 5. View the encrypted points as an image and send it to the receiver. The decryption algorithm is the reverse of the encryption algorithm process. Results of encryption and decryption are presented in Fig. 1.

(a) Plain image of House

(b) Cipher image of House

(c) Decrypted image of House

(d) Plain image of Peppers

(e) Cipher image of Peppers

(f) Decrypted image of Peppers

Fig. 1. Results of encryption and decryption.

Efficient Mapping Method for Elliptic Curve Cryptosystems

133

4 Security Analysis 4.1

Statistical Analysis

In theory, a good image encryption algorithm should have strong immunity against any kind of statistical attack. Statistical analysis methods such as Histograms and correlation analysis of two adjacent pixels on cipher images can help us to validate the uniformity property of the encrypted image and to check the independence property. The histogram shows how the pixels in an image are distributed by plotting on a graph the number of pixels corresponding to each color intensity. In our work, processed images are grayscale images whose pixel values vary in the range [0.255]. We have traced and analyzed the histograms of Peppers and its corresponding cipher image. The plots of the original and the histograms of the encrypted images are shown in Fig. 2. From these plots, we can see that the histogram of the encrypted image is uniformly distributed with respect to the histogram of the original image. The proposed encryption algorithm makes the dependence of the statistical properties of the encrypted image and the original image almost random. This makes cryptanalysis increasingly difficult because the encrypted image does not provide any element which relies on the exploitation of the histogram and which makes it possible to design a statistical attack on the proposed image encryption process.

(a) Plain image

(b) Cipher image

Fig. 2. Histograms of plain image and cipher image of peppers.

In addition to the histogram analysis, we also have checked the correlations of adjacent horizontal, vertical and diagonal adjacent pixels of the plain image and cipher images respectively. Figure 3a, c, e shows the correlation distributions of adjacent pixels of plain image “Peppers” along with horizontal, vertical and diagonal directions, Fig. 3b, d, f gives the corresponding distributions of the cipher image. It is apparent from Fig. 3 that in the case of the original images, the horizontal adjacent pixels have strong correlations and align with the first bisector. On the other hand, in the case of encrypted images, the adjacent horizontal pixels are scattered almost randomly. In general, the observation of highly scattered pixels refers to an algorithm robust to any statistical attack.

134

S. Bendaoud et al.

(a) Vertical direction in plain image

(b) Vertical direction in cipher image

(c) Horizontal direction in plain image

(d) Horizontal direction in cipher image

(e) Diagonal direction in plain image

(f) Diagonal direction in cipher image

Fig. 3. Correlation of two adjacent pixel of peppers image.

4.2

Information Entropy

Information entropy is a parameter that measures the level of the complexity of a system. It is used by researchers to weigh the performance of the encryption algorithm. Let m be the information source, and the formula for calculating information entropy is: H ðmÞ ¼ 

X2N 1 i¼0

pðmi Þ log2 pðmi Þ

ð3Þ

where p(mi) represents the probability of symbol m, and N is the total number of symbols. We know the closer it gets to 8, the less possible for the cryptosystem to

Efficient Mapping Method for Elliptic Curve Cryptosystems

135

divulge information. We use Eq. (3) to calculate the information entropy of the ciphered images of Fig. 1. Table 1 shows the entropy of the assessing results of different cipher images, which are all close to the ideal value 8. Table 1. Information entropy of plain and cipher images. Image Plain image Chiper image House 6.4971 7.9973 Peppers 7.5327 7.9973

5 Conclusions Encryption and decryption using Elliptic Curves Cryptography is based on points. A plain message should be converted to a point and the results of the encryption process are also points. Hence, message encoding and decoding is a principal phase. This paper proposes a novel mapping method to convert an image pixel value to a point on a predefined elliptic curve using a map table generated by PWLCM chaotic system. The results of experiments and security analysis indicate that the proposed image encryption scheme can achieve a good encryption result with excellent image quality and can resist common attacks.

References 1. Blakley, G.R., Borosh, I.: Rivest-Shamir-Adleman public key cryptosystems do not always conceal messages. Comput. Math. Appl. 5(3), 169–178 (1979) 2. Wang, X.Y., Wang, Q.: A novel image encryption algorithm based on dynamic S-boxes constructed by chaos. Nonlinear Dyn. 75((3), 567–576 (2014) 3. Tawalbeh, L., Mowafi, M., Aljoby, W.: Use of elliptic curve cryptography for multimedia encryption. IET Inf. Secur. 7(2), 67–74 (2012) 4. Amounas, F., El Kinani, E.H.: Security enhancement of image encryption based on matrix approach using elliptic curve. Int. J. Eng. Inventions 3(11), 8–16 (2014) 5. Singh, L.D., Singh, K.M.: Encryption using elliptic curve cryptography. In: International Multi-Conference on Information Processing (IMCIP-2015) (2015) 6. Bendaoud, S., Amounas, F., El Kinani, E.H.: A novel image encryption scheme based on Elliptic curve and Rubik’s cube. Int. Res. J. Adv. Eng. Sci. 2(2), 144–147 (2017) 7. Wu, J., Liao, X., Yang, B.: Color image encryption based on chaotic systems and elliptic curve ElGamal scheme. Sig. Process. 141, 109–124 (2017) 8. Pareek, N.K., Patidar, V., Sud, K.K.: Image encryption using chaotic logistic map. Image Vis. Comput. 24(9), 926–934 (2006) 9. Hussain, I., Shah, T., Gondal, M.A., Mahmood, H.: A novel image encryption algorithm based on chaotic maps and GF (28) exponent transformation. Nonlinear Dyn. 72(1), 399– 406 (2013) 10. Wang, X.Y., Liu, L.T., Zhang, Y.Q.: A novel chaotic block image encryption algorithm based on dynamic random growth technique. Opt. Lasers Eng. 66, 10–18 (2014)

136

S. Bendaoud et al.

11. Wang, X.Y., Liu, C.: A novel and effective image encryption algorithm based on chaos and DNA encoding. Multi. Tools Appl. 76(5), 6229–6245 (2017) 12. Saha, B.J., Kabi, K., Arun, K.: Digital image encryption using ECC and DES with chaotic key generator. Int. J. Eng. Res. Technol. (IJERT) 2(11), 2593–2597 (2013) 13. Gupta, K., Silakari, S.: Performance analysis for image encryption using ECC. In: First International Conference on Computational Intelligence, Communication Networks, pp. 79– 82 (2010) 14. Amounas, F., El Kinani, E.H.: Fast mapping method based on matrix approach for elliptic curve cryptography. Int. J. Inf. Netw. Secur. 1(2), 54–59 (2012) 15. Rao, O.S., Setty, S.P.: Efficient mapping method for elliptic curve cryptosystems. Int. J. Eng. Sci. Technol. 2, 3651–3656 (2010) 16. Silverman, J.H., Pipher, J., Hoffstein, J.: An Introduction to Mathematical Cryptography. Springer, Berlin (2008) 17. Baranovsky, A., Daems, D.: Design of one-dimensional chaotic maps with prescribed statistical properties. Int. J. Bifurcat. Chaos 5(6), 1585–1598 (1995)

3D Shape Recognition Based on Uncoded Structured Light Using ANN Classifier Kaoutar Baibai1(&), Mohamed Emharraf1, Wafae Mrabti2, Khalid Hachami1, and Benaissa Bellach1 1

LSE2I Laboratory, School of Applied Sciences Engineering, Mohamed I University, Oujda, Morocco {k.baibai,m.emharraf}@ump.ac.ma 2 IIAN Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah University, Fez, Morocco [email protected]

Abstract. In this article we present a suite of our research work on object shape recognition based on uncoded structured light that is used to acquire 3D information in the form of lines distorted by the object’s relief, from these lines we extract the 1D signals corresponding to the object. These signals are used to extract the features of the 3D object shape. In this article we propose a new approach to determine 3D shape descriptors using 1D signals. And to improve the performance of the recognition system based on 1D signal processing, we thought about implementing more information on the object shape by adding to the characteristic vector other descriptors calculated in frequency domain called frequency-based descriptors. Once the shape descriptors are calculated, we proceed to the classification of descriptor vectors in order to recognize the different shapes of 3D objects. The results of the proposed approach allow 3D object recognition with an accuracy of 99.6% using the ANN classifier on a database formed by 10 objects. We present a comparison between the results obtained by applying our approach to different databases made up of 6, 7 and 10 objects and treat these results according to two classifiers KNN (K Nearest Neighbors) and ANN (Artificial Neural Network). Keywords: 3D shape Recognition


Frequency descriptors
1D signal
ANN classifier


1 Introduction 3D shape recognition is an important research subject in computer vision, which has been applied to several fields, including robotics [1], computer vision applications such as multimedia games and medical diagnostics, industrial applications [2–4] such as: inspection [5], sorting of objects, etc. Although many object recognition algorithms have been proposed, most of them are designed for 2D object recognition. With the development of 3D capture devices, it will be much easier to capture 3D objects, and it is very common to access 3D objects in our daily lives, so we have an urgent need to design algorithms to recognize 3D objects, but these algorithms are complex and high computational time. Computer vision researchers pay more attention to 3D object © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 137–143, 2020. https://doi.org/10.1007/978-3-030-53187-4_16

138

K. Baibai et al.

recognition, but they often extend 2D object recognition algorithms directly to 3D object recognition without taking into account the characteristics of 3D objects. In this context, many studies have been conducted using different types of visual sensors as well as different pre-treatment techniques and classification algorithms. However, in all these studies, the performance of the real-time recognition system depends on accuracy, recognition time and reliability. In fact, these are the main criteria that can be used to validate visual sensors as well as algorithmic approaches in an object recognition mechanism. The retrieval and classification of 3D objects according to shape has received increased attention in recent years due to the increase in the number of available 3D objects. Several articles have been published on the subject [6, 7]. There are three main categories of 3D object representation: feature-based methods, graph-based methods and view-based methods. Feature-based methods, which are the most popular, can be subdivided into (1) global characteristics, (2) global distributions of characteristics, (3) spatial maps and (4) local characteristics. The proposed approach is designed to recognize 3D shapes from 1D signals. Therefore, the first phase of the work consists in the acquisition of information related to the shape of 3D objects in the form of 1D signals. To acquire this information we use the acquisition technique based on uncoded structured light. The general principle of this technique is the projection of a multi-line pattern with uniform color, width and spacing. The Fig. 1 presents an example of used pattern with green color. After the pattern projection we acquire the deformed by the 3D shape of the objet, we convert these deformed line to 1D signals using our approaches presented in the article [8].

Fig. 1. Pattern of uncoded structured light.

This recognition approach is based on the analysis and processing of shape-related information contained in 1D signals. From these signals, we try to find an optimal representation of the original data by keeping as much information as possible. In this article, we present the proposed approach to extract this information then we present the methodology flowed recognize the 3D objects using ANN classifiers.

3D Shape Recognition Based on Uncoded Structured Light

139

2 Feature Extraction for Object Classification To characterize the 3D shape of the object we use the descriptors of the 1D signals extracted from the image of this object. In our previous work [9] we used the category of time-based descriptors, i.e. calculated in time domain. These descriptors describe the global and local characteristics of the 1D signals corresponding to the 3D object shape. They are composed by a set of parameters such as the mean, local extremes, curvature and other signal characteristics. The characteristic vector of the 3D object shape is a collection of all these descriptors. To improve the performance of the recognition system based on 1D signal processing, we thought about implementing more information on the object shape by adding to the characteristic vector other descriptors calculated in frequency domain called frequency-based descriptors. One of the most commonly used methods to analyze non-periodic signals is the Fast Fourier Transform (FFT). To determine the frequency descriptors we will first have to calculate the FFT by the expression (1). XðkÞ ¼

n1 X

xðnÞej2p N

nk

ð1Þ

n¼0

Where xðnÞ is a sequence of N elements of the signal x and k is between 0 and N – 1. Among the frequency characteristics that can be extracted from the Fourier transform elements of the signal are: – – – – – – –

Fundamental frequency, Frequency components, The mean frequency, Spectral centroid, Spectral flow, Spectral density, Spectral attenuation, etc.

3 Classification Once the characteristics of the object shape are determined, the classification step is performed to make a decision that allows the object to be recognized. This step is the classification, where the decision is made based on the descriptive vector of the shape of the 3D object. Classification predicts the categorical label (discrete, ordered). The data classification is divided into two steps. The first step is the training step in which a classifier is constructed to describe a predetermined set of data classes. In the second step, the model constructed in the first step is used for the classification of unknown data, i.e. the test data are used to estimate the accuracy of the classifier. In the literature there are many classifiers such as SVM, Decision tree, KNN and ANN. Among these classifiers we worked in our previous work [9] with the classifier KNN where we obtained a high accuracy of 99.1% on a database composed by 6 objects. However,

140

K. Baibai et al.

accuracy decreases if we increase the size of the database by adding other objects. To do this, we thought about using another classifier to improve the performance of the recognition system, and as classifier we choose ANN classifier. ANN is a set of connected input/output networks in which the weight is associated with each connection. It is composed of an input layer, one or more intermediate layers and an output layer. It has the ability to implicitly detect complex non-linear relationships between dependent and independent variables; it is able to detect all possible interactions between predictive variables. The neural network is learned by adjusting the weight of the connection. By updating the weight iteratively the network performance is improved. The artificial neural network is applicable in various applications such as pattern recognition [10], robotics [11]. The classification procedure starts with a data set that is divided into two parts: the training sample and the test sample. Training sample is used for learning of network while test sample is used for measuring the accuracy of classifier. The division of the data set can be done by various methods such as the hold-out method, k-cross-validation, random sampling, in our case of study we use 10-cross-validation. To achieve an effective classification, the descriptors describing the signals must be selected in an optimal way. Each signal can be described by several descriptors, which can be correlated and therefore redundant. In addition, there must be enough descriptors to avoid losing information about the signals, but not too many to avoid making the calculation cumbersome. It is therefore necessary to choose these descriptors wisely.

4 Results To validate the proposed 3D shape recognition approach based on uncoded structured light using neural networks, we applied this approach to a database formed by 10 objects. This database is composed of two data sets for each object, the first is the training set formed by 40 feature vectors and the second is the test set formed by 10 vectors. For each object, each vector is composed by a set of shape descriptors of the 1D signals which are determined according to different view angles. This vector is formed by combination of temporal and frequency descriptors. Experimental tests on the different frequency descriptors cited in previous section have allowed us to select two descriptors, the mean frequency Eq. (2) and the peaks frequencies Eq. (3). These descriptors give the best results in terms of classification accuracy. The Fig. 2 illustrates an example of peaks frequencies extraction from 1D signal. N1 P

Fm ¼

XðkÞ

k¼0

N

Fp ¼ maxðXðkÞÞ

ð2Þ ð3Þ

3D Shape Recognition Based on Uncoded Structured Light

141

Fig. 2. Example of frequencies descriptors extraction, (a) 1D signal, (b) peaks frequencies.

Once the descriptors are determined, we proceed to the classification step. In this step we determine the architecture of the ANN classifier that is made up of multiple layers, see Fig. 3: input layers, a number of hidden layer and output layer. As input layer we used the number of descriptors of an object multiplied by number of used line layers and 10 hidden layers.

Output layers

Object

10 Hidden layers

…

Descriptive vector

Input

Fig. 3. ANN classifier architecture.

142

K. Baibai et al.

Table 1 represents the results obtained by applying the recognition approach to three different databases consisting respectively of 6, 5 and 7 e objects using two classifiers KNN and ANN. This table illustrates the results obtained by using a combination of the two types of temporal and frequency descriptors. Table 1. Classification results obtained using KNN and ANN classifiers. Classifier dataset 6 Objects 7 Objects 10 Objects

KNN 99% 94,86% 96,40%

ANN 100% 100% 99,6%

These results show that the classification accuracy decreases if the database size for the KNN classifier is increased. However, for the ANN classifier, the recognition system maintains its performance even if the size of the processed data is increased. In this article we have increased the size of the database to 10 objects where the classification accuracy to be decreased to 96.4% with the KNN classifier. On the other hand, by using the ANN classifier we obtain efficient results. As well, we were able to reduce the size of the learning database for this classifier by using only 40 learning vectors instead of 70 without affecting classification performance.

5 Conclusion In this article we have presented a suite of our research work on object shape recognition based on uncoded structured light that is used to acquire 3D information in the form of lines distorted by the object’s relief, from these lines we extract the 1D signals corresponding to the object. This approach consists on 3D shape recognition based on time-frequencies descriptors of extracted 1D signals using ANN classifier. The proposed approach increase the classification accuracy to 99,6% compared to our previous work that we obtained 96,4% for database contained 10 objects. As well, we were able to reduce the size of the learning database for this classifier by using only 40 learning vectors instead of 70 without affecting classification performance.

References 1. Ye, C., Qian, X.: 3-D object recognition of a robotic navigation aid for the visually impaired. IEEE Trans. Neural Syst. Rehabil. Eng. 26(2), 441–450 (2018) 2. Hirano, Y., Garcia, C., Sukthankar, R., et al.: Industry and object recognition: applications, applied research and challenges. In: Toward Category-Level Object Recognition. Springer, Berlin, pp. 49–64 (2006) 3. Benhimane, S., Najafi, H., Grundmann, M., et al.: Real-time object detection and tracking for industrial applications. In: VISAPP (2), pp. 337–345 (2008)

3D Shape Recognition Based on Uncoded Structured Light

143

4. Simon, M., Milz, S., Amende, K., et al.: Complex-YOLO: real-time 3D object detection on point clouds. arXiv 2018. arXiv preprint arXiv:1803.06199 (2018) 5. Madrigal, C., Branch, J., Restrepo, A., et al.: A method for automatic surface inspection using a model-based 3D descriptor. Sensors 17(10), 2262 (2017) 6. Zhou, Y., Zeng, F., Qian, J., et al.: Shape classification and retrieval based on polar view. Inf. Sci. 474, 205–220 (2019) 7. Wei, H., Yu, Q., Yang, C.: Shape-based object recognition via evidence accumulation inference. Pattern Recogn. Lett. 77, 42–49 (2016) 8. Baibai, K., Elfakhouri, N., Bellach, B.: 3D acquisition system for 3D forms recognition. In: International Conference on Advanced Technologies for Signal and Image Processing (ATSIP), pp. 1–6. IEEE (2017) 9. Baibai, K., Emharraf, M., Mrabti, W., et al.: 1D signals descriptors for 3D shape recognition. In: International Conference Europe Middle East & North Africa Information Systems and Technologies to Support Learning, pp. 687–693. Springer (2018) 10. Nabhani, F., Shaw, T.: Performance analysis and optimisation of shape recognition and classification using ANN. Robot. Comput. Integr. Manuf. 18(3), 177–185 (2002) 11. Choe, Y., Seunguk, A., Chung, M.J.: Online urban object recognition in point clouds using consecutive point information for urban robotic missions. Robot. Auton. Syst. 62(8), 1130– 1152 (2014)

Machine Vision-Based Cocoa Beans Fermentation Degree Assessment Aubain Yro1(&), Édié Camille N’Zi1, and Kidiyo Kpalma2 1

Institut National Polytechnique Félix Houphouët-Boigny, Yamoussoukro, Côte d’Ivoire [email protected] 2 Univ Rennes, INSA Rennes, CNRS, IETR - UMR 6164, 35000 Rennes, France

Abstract. Fermentation degree is one of the main important indicators of cocoa bean quality. Therefore, accurate estimation of fermentation degree is very important for ensuring the quality of final products. This paper presents a quantitative method for assessing the cocoa beans fermentation degree by image analysis. In this approach, the image of cocoa beans are acquired using a camera and processed to obtain the bean’s target. Then, the target’s pixels are clustered into red, green and blue regions where each region’s pixel presents respectively a maximal value of R, G and B in RGB color space. After that, the first three color moments of each region are calculated from RGB space and used to describe the fermentation degree of the bean. Finally, multi-class support vector machine (SVM) algorithm is used as classifier to discriminate cocoa beans sample into unfermented (UF), partly fermented (PF) and well fermented (WF) categories. Experimental results show that 99.17% of UF beans, 97.50% of PF beans and 100% of WF beans were detected successfully. This results revealed that the proposed method could be used as a fast, accurate and a reliable tool to discriminate cocoa beans according to their fermentation degree for quality assurance. Keywords: Cocoa bean vision
SVM


Fermentation degree
Image analysis
Machine

1 Introduction Cocoa beans are the seeds from fruit pods of the tree Theobroma L. [1]. They are well known as the main raw material of chocolate and of a vast range of products like cocoa powder, cocoa beverages, bakery products and ice cream [2]. Worldwide cocoa is an important agricultural crop and it is one of the most exported raw material. However, before cocoa beans can be traded and processed into final industrial they have to undergo post-harvest processing on farms and plantations [3–5]. The fermentation is the critical post-harvest step to improve the quality of the beans [6–8]. Indeed, the flavour and aroma precursors which reduces the astringency and bitterness of the beans are produced during fermentation [3]. Therefore, the

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 144–148, 2020. https://doi.org/10.1007/978-3-030-53187-4_17

Machine Vision-Based Cocoa Beans Fermentation Degree Assessment

145

fermentation degree of cocoa beans is an important quality indicator which determine the acceptability or rejection of beans on the international market [3, 5, 6, 9]. Traditionally, cut-test is used to assess fermentation degree of cocoa beans [3, 9, 10]. It is a simple visual method which involves cutting a number of beans lengthwise and analyzing their internal color [4]: unfermented beans are slaty, partly fermented beans are purple or violet and fully or well fermented beans are brown in color. However, this method is time-consuming, subjective and not reliable. This paper proposed a simple color image analysis method for assessing cocoa fermentation degree by using RGB color features and a multi-class support vector machine algorithm.

2 Materials and Methods 2.1

Cocoa Beans Sample

The cocoa beans samples used in this study were acquired by the National Polytechnic Institute of Yamoussoukro in Côte d’Ivoire (7° 32’ 23.96” N; 5° 32’ 49.488” W). A total of 600 beans samples are manually selected by a quality control expert and image of each bean was captured by a computer vision system. The cocoa beans were divided into three classes of fermentation degree: unfermented (Class 1), partly fermented (Class 2) and well fermented (Class 3) were identified in this research. The training set consists of 80 images per class forming a set of 240 images and the remaining images which consists of 120 images for each class were used as testing set. 2.2

Image Acquisition

Figure 1 shows the acquisition system developed in the laboratory. For image acquisition, a color coupled charge device (CCD) camera (SONY XCG-5005CR, Japan) with lens zoom 16 mm (Fujifilm corporation, model HF16HA-1B, Japan) was used. The image acquisition card (Mil Matrox) was used for transferring information from camera to computer (Core-i7 CPU: 2.5 GHz; RAM: 4 GB). To ensure a correct and consistent lighting throughout the acquisition process, two white LED with 18 watts were used. Once the camera has been set, images have been acquired in tiff format with 512  512 pixels resolution. 2.3

Image Preprocessing

This image preprocessing step aims to identify each cocoa bean target from the image background. Firstly, each image were resized to 256  256 pixels in order to reduce images processing time and then they were segmented. To isolate the target from the background, the b gray image of CIELAB color space were segmented using Otsu’s thresholding method. After that, median blur type smoothing, morphological opening and closing were used to remove single noise pixels. Finally, the binary image was

146

A. Yro et al.

Camera Sample mage LED Sample Fig. 1. The cocoa bean image acquisition system.

used as masking on the input image to obtain the desired image of cocoa beans. The preprocessing steps performed on the cocoa beans samples are shown in Fig. 2.

(a)

(b)

(c)

(d)

(e)

Fig. 2. The chart of processing steps for segmenting the cocoa beans image from a noisy background: (a) the raw beans sample image; (b) gray image of channel b of the Lab image; (c) after the noise cancellation process; (d) binary image after morphological opening and closing and filling process; and (e) after the segmentation of the beans image from the background.

2.4

Features Extraction

After the desired cocoa beans image was isolated from the background, the mean, the standard deviation and the skewness were extracted using RGB color space. A total of 27 color feature were extracted to describe each cocoa bean target. In order to obtain this color features, we identified three color regions of each bean surface. The color region namely red, green and blue regions are identified as the region in which each pixel presents the maximal value of R, G and B channels in the RGB color space, respectively. Let Pi = (R, G, B) be the ith pixel of the cocoa bean, with its three colors components in the RGB color space. If the maximal value of the three colors components is R, the pixel Pi = (R, G, B) is classified as the pixel of the red region of cocoa bean. For example, the red region is determined according to Eq. (1): CR ¼ fP ¼ ðR; G; BÞ=R [ maxðG; BÞ; i ¼ 1; . . .; Ng

ð1Þ

Machine Vision-Based Cocoa Beans Fermentation Degree Assessment

147

Were CR is the set of the pixels in the red region and N is the total number of the pixel which represents the cocoa bean target in the image. 2.5

Classification

Finally, a Support Vector Machine (SVM) algorithm was attempted to develop discrimination model which were applied to classify the cocoa beans in three categories fermentation degree: unfermented, partly fermented and well fermented. SVM is a nonlinear supervised learning method for classification and regression problems which was developed by Vapnik. The basic purpose of an SVM algorithm is to obtain the optimal boundary which separates the different classes in the training set. The optimal performance of SVM is dependent on the choice of the kernel function. For this work, Gaussian radial basis function (RBF) is chosen as the kernel function.

3 Results and Discussion In this study, we used the confusion matrix and the classification accuracy to evaluate the performance of the proposed method. Experimental results obtained using the proposed method is shown in Table 1. It can be seen, this algorithm could distinguish between unfermented, partly fermented and well fermented beans with 99.17%, 97.50% and 100% respectively. The highest error of classification was related to class 2 (partly fermented beans) with 97.50%. For this class, 1 bean was wrongly allocated into class 1 (unfermented beans) and 2 others beans were allocated wrongly into class 3 (well fermented beans). However, the beans samples in class 3 (well fermented beans) are successfully classified. Table 1. Classification performances of the proposed method. Predicted classes Performance Class 1 Class 2 Class 3 Num. beans Accuracy (%) Class 1 119 1 0 120 99,17 Class 2 1 117 2 120 97,50 Class 3 0 0 120 120 100 Total accuracy 98,89

Misclassification of the 1 partly fermented bean as the unfermented bean and the 2 partly fermented beans to well fermented beans class was related to the color properties overlap between the three classes. The rate of well-classified images obtained by this proposed approach using color features in RGB color space was 98.89%. Among 360 samples, only 4 beans sample were wrongly classified. This means that this classifier system had the lowest error and an appropriate accuracy.

148

A. Yro et al.

4 Conclusion In this work we proposed an automatic procedure for estimating the cocoa beans fermentation degree. Based on computer vision methods, this system is easy to use and gives repeatable results of images classification. The classification accuracy obtained using color feature for unfermented, partly fermented and well fermented cocoa beans was 98.89%. These results shows that the proposed system can achieve very high classification performance. Finally, this study have proven that the presented system can potentially be used as a routine analytical tool for cocoa beans fermentation degree assessment. In the future we plan to increase the number of images per training class and to investigate combining color and texture features for cocoa beans fermentation degree assessment process by machine vision. Acknowledgements. This work is supported by the Strategic Support Program for Scientific Research (PASRES-CI) of Côte d’Ivoire. The first authors gratefully acknowledge IETR/INSARennes laboratory for its valuable collaboration.

References 1. Teye, E., Huang, X., Han, F., Botchway, F.: Discrimination of cocoa beans according to geographical origin by electronic tongue and multivariate algorithms. Food Anal. Methods 7 (2), 360–365 (2014) 2. Beg, M.S., Ahmad, S., Jan, K., Bashir, K.: Status, supply chain and processing of cocoa - a review. Trends Food Sci. Technol. 66, 108–116 (2017) 3. Aculey, P.C., Snitkjaer, P.: Ghanaian cocoa bean fermentation characterized by spectroscopic and chromatographic methods and chemometrics. J. Food Sci. 75(6), 300–307 (2010) 4. León-Roque, N., Abderrahim, M., Nuñez-Alejos, L., Arribas, S.M., Condezo-Hoyos, L.: Prediction of fermentation index of cocoa beans (Theobroma cacao L.) based on color measurement and artificial neural networks. Talanta 161, 31–39 (2016) 5. Emmanuel, O.A., Jennifer, Q., Agnes, S.B., Jemmy, S.T., Firibu, K.S.: Influence of pulppreconditioning and fermentation on fermentative quality and appearance of Ghanaian cocoa (Theobroma cacao) beans. Int. Food Res. J. 19(1), 127 (2012) 6. Teye, E., et al.: Estimating cocoa bean parameters by FT-NIRS and chemometrics analysis. Food Chem. 176, 403–410 (2015) 7. Diomande, G.G.D.: Contribution des méthodes spectroscopiques et isotopiques à la caractérisation géographique et phénotypique du cacao (2014) 8. Bankoff, L., Ouattara, G.H., Karou, T.G., Guehi, S.T., Nemlin, J.G., Diopoh, J.K.: Impacts de la fermentation du cacao sur la croissance de la flore microbienne et la qualite des feves marchande. Agron. Afr 25(2), 159–170 (2013) 9. Kongor, J.E., Takrama, J.F., Budu, A.S., Mensah-Brown, H., Afoakwa, E.O.: Effects of fermentation and drying on the fermentation index and cut test of pulp pre-conditioned Ghanaian cocoa (Theobroma cacao) beans. J. Food Sci. Eng. 3(11), 625 (2013) 10. ISO: Fèves de cacao - Épreuve à la coupe (ISO 1114). Organisation internationale de normalisation, Genève (1977)

Plants Classification Using Neural Shifted Legendre-Fourier Moments Abderrahmane Machhour(&), Amal Zouhri, Mostafa El Mallahi, Zakia Lakhliai, Ahmed Tahiri, and Driss Chenouni Laboratory of Computer Science and Interdisciplinary Physics LIPI, Sidi Mohamed Ben Abdellah University, Fez, Morocco [email protected], {amal.zouhri,Mostafa.elmallahi, zakia.lakhliai}@usmba.ac.ma, [email protected], [email protected]

Abstract. Plants are the primary food source of humans. They are the raw material of most medicines. Therefore, it was necessary to use artificial intelligence to help those interested in plants to classify and identify various plants types quickly and accurately. In this article we present Neural Shifted LegendreFourier Moments, we used shifted Legendre-Fourier moments to extract features from leaves images and build descriptor vectors. These vectors are the inputs of the artificial neural network. We tested this model on MalayaKew (MK) Leaf dataset and we got important results. The validity of this proposed method has been provided under different transformations. Keywords: Image classification
Plants classification
Shifted LegendreFourier moments
Deep learning
Artificial neural network
Features extraction
Descriptor vector
Artificial intelligence

1 Introduction In recent years, image classification has become a major concern in various fields because of the significant role it plays in helping to make decisions. The automatic image classification process begins with features extraction. One of the methods used in features extraction is the computation of image moments, in particular the orthogonal moments since they retain relevant information about the image [3]. The first use of moments in image analysis and pattern recognition took place in 1961 by Hu et al. [2]. They used geometric moments to extract invariants that are then used in automatic character recognition. This set of moment invariants is translation, scale and rotation independent. But these moments are not orthogonal and as a consequence reconstructing the image from it is seemed to be a hard operation. The next important step was the introduction of orthogonal moments by Teague et al. in 1980 [3]. These moments are based on continuous orthogonal polynomials such as Zernike and Legendre polynomials [4]. Since then, many continuous orthogonal moments were consecutively presented, Fourier-Mellin moments [5], ChebyshevFourier moments [6] and Gaussian-Hermite moments [7]. There are other methods that © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 149–153, 2020. https://doi.org/10.1007/978-3-030-53187-4_18

150

A. Machhour et al.

are based on the discrete moments using discrete polynomials such as Krawtchouk polynomials [8], dual Hahn polynomials [9] and Racah polynomials [10] for image processing. Regarding classification of plants based on their images, Redha Almahdi et al. [11] proposed a plant classification method based on two types of texture features and a multiclass Support Vector Machine. Sue Han Lee et al. [12] presented a plant classification system using Convolutional Neural Networks. Luciano D. S. Pacifico et al. [13] used a Multi-Layer Perceptron (MLP) artificial neural network with Backpropagation algorithm training to classify plants based on their leaves shapes. In this paper, we give the definition of Shifted Legendre-Fourier Moments and the expression of the extracted invariants to rotation, scaling and translation transformations (Sect. 2). We used these invariants as inputs to the artificial neural network and we applied this system (Fig. 1) on MalayaKew Leaf dataset [12] and we have built a classification model with a high accuracy using only a few hundred images in the training phase (Sect. 3).

Artificial Neural network

Training Image SLFM invariants

Label Prediction Image SLFM invariants

Classifier model

Label

Fig. 1. Proposed classification system

2 Shifted Legendre-Fourier Moment’s Invariants In a unit circle, Shifted Legendre–Fourier moments (SLFMs) of order p and repetition q are defined as following:

Plants Classification Using Neural Shifted

Mpq ¼

2p þ 1 p

Z 0

2p

Z

1

^

f ðr; hÞPp ðr Þeiqh rdrdh;

151

ð1Þ

0

Where Pp ðr Þ are the orthogonal Shifted Legendre polynomials .  Based on the formula (1), Machhour et al. [1] succeeded in extractinginvariants  xpq  to rotation, scaling and translation simultaneously. The expression of xpq  is as following:   ! p p    X 2p þ 1 X  ði þ 1Þ xpq  ¼  M00 ð f ÞÞ Spi Tik Mkq ð f Þ   k¼0 2k þ 1 i¼k

ð2Þ

Where: Spi ¼ ð1Þp1

Tik ¼

ðp þ iÞ! ðp  iÞ!ði!Þ2

ð2k þ 1Þði!Þ2 ði þ k þ 1Þ!ði  kÞ!

ð3Þ

ð4Þ

These invariants of different orders are the images features that will help us to build the descriptor vector of the plants subsequently used as learning data.

3 Experimentation During the features extraction phase, we chose to work on the MalayaKew (MK) Leaf dataset (Fig. 2), which is a challenging dataset because it contains 2816 color images in 44 classes. The images display different leaves of plants at resolution 256 by 256 pixels. Each image is assigned to a single label. The labels are an array of integers, ranging from 1 to 44. These correspond to the class of plant the image represents.

Fig. 2. Six samples from MalayaKew leaf dataset

Image moments of first orders are ideal for the classification of objects based on their shapes. Then to work on dataset images, we carried out a color to grayscale conversion before the calculation of SLFM invariants. The calculation of LFM invariants is carried out in the polar coordinates to avoid the geometric errors [14]. And it is done in a recursive way, and this is a little expensive

152

A. Machhour et al.

for the calculation time oand memory. We computed 16 LFM invariants n n   xp;q =ðp; qÞ 2 1; 2; 3; 4g2 for every sample of the first 20 first classes of the dataset without any fast or accelerated computation method. To build our artificial neural network we used Keras, it is a model-level library for Python that provides an appropriate way to define and train machine learning models. The input data of our model is vectors with 16 dimensions (16 LFM invariants), and the labels (Classes) are scalars {1; 2; 3;…; 20}. We used a fully connected (Dense) layers ANN with ReLu activations. First, we trained the ANN to classify 80% of the items that belong to two categories (1 and 2) and we tested the model on the remaining 20%, the model’s test accuracy beats 100%. Then we made several classifications by increasing the number of categories each time in ascending order of labels (Table 1).

Table 1. Classification accuracy according to number of classes. Number of classes Labels Training samples Test samples Epochs Train accuracy Test accuracy

2 1; 2 83 21 2000 100% 100%

4 6 8 10 20 1; 2; 3; 4 1;…; 6 1;…; 8 1;…; 10 1;…; 20 166 250 333 416 832 42 62 83 104 208 3000 3000 3000 3000 7000 95% 94% 92% 86% 87% 92% 82% 80% 79% 72%

We have noticed that the accuracy has decreased in the case of 20 classes and this is due to the confusion that the neural network has encountered when classifying the elements that belong to the different classes but have almost the same shape and texture. This cause appeared when we trained the ANN to classify only the elements of the 4 categories: 1, 3, 5 and 10. After 3000 epochs, the train accuracy was about 90% but the test accuracy was about 78%. These results are acceptable but insufficient, which leads us to seek to improve the classification system by combing the method of moments with other methods in the next projects.

4 Conclusion In this paper, a new method for plant classification has been presented based on the invariants extracted from the shifted Legendre-Fourier moments and by using a fully connected artificial neural network. The calculation of the moments was carried out in the polar coordinates to avoid the geometric errors. Consequently the classification has achieved a good level of accuracy by using only moments of low orders. In the upcoming works, we aim to increase our models accuracy and to use an accelerated algorithm of SLFM invariants calculation in order to work on SLFM invariants of higher orders and classify color images of higher resolution.

Plants Classification Using Neural Shifted

153

References 1. Machhour, A., EL, Mallahi, M., Lakhliai, Z., Tahiri, A., Chenouni, D.: Image classification using Legendre-Fourier moments and artificial neural network. In: 1st International Conference on Embedded Systems and Artificial Intelligence. Springer, May 2019 (ESAI 2019) 2. Hu, M.K.: Visual pattern recognition by moment invariants. IRE Trans. Inf. Theory 8(2), 179–187 (1962) 3. Teague, M.R.: Image analysis via the general theory of moments. J. Opt. Soc. Am 70(8), 920–930 (1980) 4. El Mallahi, M., Zouhri, A., Amakdouf, H., Qjidaa, H.: Rotation scaling and translation invariants of 3d radial shifted Legendre moments. Int. J. Autom. Comput. 15(2), 169–180 (2018) 5. Sheng, Y., Shen, L.: Orthogonal Fourier-Mellin moments for invariant pattern recognition. JOSA A 11(6), 1748–1757 (1994) 6. Ping, Z., Wu, R., Sheng, Y.: Image description with Chebyshev-Fourier moments. J. Opt. Soc. Am. A Opt. Image Sci. Vis 19(9), 1748–1754 (2002) 7. Yang, B., Dai, M.: Image analysis by Gaussian-Hermite moments. Signal Process 91(10), 2290–2303 (2011) 8. El Mallahi, M., Zouhri, A., El Mekkaoui, J., Qjidaa, H.: Three dimensional radial Krawtchouk moment invariants for volumetric image recognition. Pattern Recogn. Image Anal. 27(4), 810–824 (2017) 9. El Mallahi, M., Zouhri, A., Mesbah, A., Qjidaa, H.: 3D radial invariant of dual Hahn moments. Neural Comput. Appl. 30(7), 2283–2294 (2018) 10. El Mallahi, M., Zouhri, A., Mesbah, A., El Affar, I., Qjidaa, H.: Radial invariant of 2D and 3D Racah moments. Multi. Tools Appl. Int. J. 77(6), 6583–6604 (2018) 11. Almahdi, R., Hardie, R., Essa, A.: A leaf recognition approach to plant classification using machine learning. In: IEEE Spectrum December (2018) 12. Lee, S.H., Chan, C.S., Wilkiny, P., Remagninoz, P.: DEEP-PLANT: plant identification with convolutional neural networks. In: IEEE International Conference on Image Processing (ICIP 2015), pp. 452–456 (2015) 13. Pacifico, L.D.S., Macario, V., Oliveira, J.F.L.: Plant classification using artificial neural networks. In: International Joint Conference on Neural Networks (IJCNN) (2018) 14. Xin, Y., Pawlak, M., Liao, S.: Accurate computation of Zernike moments in polar coordinates. IEEE Trans. Image Process. 16(2), 581–587 (2007)

Cybersecurity and Data Protection

Criteria for Security Classification of Smart Home Energy Management Systems Manish Shrestha1,2(B) , Christian Johansen1 , and Josef Noll1 1 University of Oslo, Oslo, Norway [email protected], [email protected], [email protected] 2 eSmart Systems AS, Halden, Norway

Abstract. Internet of Things (IoT) is a growing field and its use in home automation is one of the dominating application areas. However, the end-users lack security awareness, whereas the system designers lack the incentives for building secure IoT systems. To address this challenge, we propose the notion of security classes to assess and present the security of complex IoT systems both for the users and for developers. Furthermore, regulatory bodies can use our security classification method as a reference to derive requirements for adequate security. This paper presents a security classification methodology and extends it towards Smart Home Energy Management Systems (SHEMS). We demonstrate its applicability by performing a systematic security classification assessment of an industrial SHEMS. Our results show that the use of security classes is a good indication of the level of security, as well as a guide to improve the security of IoT system. Keywords: Security classification · Exposure · Security assessment Cybersecurity · Smart home · IoT · Energy Management System

1

·

Introduction

The proliferation of IoT has created new transformative opportunities, e.g., in smart homes [1,14]. Today, the applications inside smart homes are more than luxury, where, e.g., SHEMS can enable efficient utilization of energy [4]. Industrial IoT providers are normally driven by the development of functionalities, creating a range of communication and sensing capabilities integrated into small devices. However, from a customer point of view, security and privacy has been a major concern, and often hinders a wider adoption of IoT systems. This paper is an exemplification and adaptation of our previous work [12] on general security classification methodology for smart grid systems. Here we extend the security classes with details regarding connectivity classes and protection mechanisms suitable for SHEMS and show the application of our approach c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 157–165, 2020. https://doi.org/10.1007/978-3-030-53187-4_19

158

M. Shrestha et al.

to an existing commercial system. One motivation of the present work is to help companies to improve and maintain IoT security of their products guided by security classes and the protection mechanisms that they require. We describe in Sect. 2 the reference architecture for SHEMS that we follow, and briefly introduce the system from our case study. Our main contribution is presented in Sect. 3 where we extend the security classification method towards SHEMS. We show the application of this new methodology in Sect. 4 using a case study of existing SHEMS from Develco Products.

2

A Commercial Home Energy Management System

A SHEMS is a smart home system dedicated to save energy by monitoring and managing electrical appliances, which may include load (e.g., heat pumps), storage (e.g., car batteries), or generation resources (e.g., solar panels) [6,7, 15]. Functional modules of SHEMS may include monitoring, logging, control, management, or alarm services [15]. Ghirardello et al. [5] summarize a smart home reference architecture (see Fig. 1) by integrating three different viewpoints: functional, physical, and communication capabilities. Based on this architecture, we describe the major components of smart home systems as below. IoT Devices have as primary functions [5] to sense the environment, transfer data, and receive commands. As such, these have communication capabilities and can interact with other components of the Home Area Network (HAN) such as IoT hubs, residential gateways, or other IoT devices. In SHEMS, IoT devices may include metering (and sensing) devices and controllable loads. IoT Hub acts as a central controller of IoT devices as well as a bridge between these and the backend system. Sensor data are reported to the IoT hub, which translates and sends them to the backend system. Similarly, the IoT hub may receive control commands, which it can relay to the intended devices. Opposed to IoT devices, the IoT hub has considerably more computing capability and can make decisions to manage and control the IoT devices. Residential Gateway is a bridge connecting IoT devices to Internet [5], i.e., between the HAN and the Wide Area Networks (WAN). Quite often an IoT hub and residential gateway functionalities are integrated into one device. Communication Channels. A SHEMS consists of two types of networks: HAN and WAN. The HAN is formed of the sensors and the IoT hub, and utilize wireless communication links such as Zigbee, Z-Wave, Wireless M-Bus, Thread [3,5]. The IoT hub and devices may also utilize Wi-Fi or Ethernet to connect with the residential gateway. In a WAN, a SHEMS typically utilizes the home Internet to communicate with the backend system.

Criteria for Security Classification of SHEMS

IoT Device IoT Gateway

ResidenƟal Gateway

Internet

159

Backend

IoT Device

Fig. 1. Smart home system architecture.

Backend System is a centralized component, which manages several smart homes, and resides remotely, communicating with the IoT hub through the Internet and performing storage, monitoring, and control functionalities of IoT devices. Backend systems provide an interface to external applications through APIs, enabling communications with SHEMS [14]. Application and Network Data. The network data includes mainly information related to connectivity. Application data are those which actually have business value and include meter values, commands for controlling devices, log data, firmware image files, etc. Metered values are produced by IoT devices and sent to the IoT hub, which further sends these to the backend systems for storage and analysis. On the other hand, control commands are received by the IoT hub from the backend system and then sent to the IoT devices for execution. We apply our security classification to a commercial smart home solution offered by E2U Systems AS, using hardware provided by Develco Products, and implementing customized software solutions for smart homes. Develco offers an IoT hub (called Squid.link gateway) and a variety of IoT devices such as smart plugs, sensors, alarms, and meter interfaces. The IoT hub can act as a residential gateway using the cellular network, and it also provides an Ethernet and a WLAN interface for Internet connection as well as a USB interface for plugging in 3G/4G dongles. The Squid.link gateway is a modular platform capable of bridging multiple wireless platforms, like Zigbee, Z-wave, Wireless M-Bus, in the HAN. The wireless module on the main board of the IoT gateway communicates with the CPU using the SmartAMM protocol, which is the proprietary protocol that also facilitates communication between gateways and the backend.

3

Extended Security Classification Method

The Smart Grid Security Classification (SGSC) methodology [12] is based on the ANSSI classification method [2]. However, instead of estimating the exposure based on the complexity of the system and attacker model (as in ANSSI), the SGSC combines the connectivity (which captures the surface of a system exposed to attacks) with protection (which describes the mechanisms of the system used to protect the connectivity surface) [12]. Figure 2 summarizes how a security class

160

M. Shrestha et al.

is computed, first looking at the components of the system and then aggregating the results upwards until reaching a security class for the whole system. Notably, the SGSC does not focus on attackers, as classical risk-based methods do, but is concerned instead with how a system can be securely built from a design point of view. The benefit is that the SGSC helps system designers to choose the most appropriate security functionalities to meet the envisaged security class. In addition, the focus on secure-by-design systems is better suited for long term applicability, as threats and vulnerabilities only represent a snapshot, whereas security classes present and inherent view of an IoT system.

Impact Security Class

ConnecƟvity Exposure ProtecƟon Level

Fig. 2. Methodology of computing a security class [12] (Impact as in ANSSI).

We consider two types of exposures: IT Exposure and Physical Exposure. For both, we evaluate the connectivity into one of five levels as follows: C1: Includes completely closed/isolated systems. C2: Includes the system with wired Local Area Network and does not permit any operations from outside the network. C3: Includes all C2 systems that also use wireless technologies. C4: Includes the system with private or leased infrastructure, which may permit remote operations (e.g., VPN, APN, etc.). C5: Includes distributed systems with public infrastructure, i.e., like the C4 category except that the communication infrastructure is public. We have defined Protection Levels (P) to capture the strength of the security mechanisms implemented in a system. Protection Levels have been inspired by the Safety Integrity Levels (SIL) [10]. Instead of the attacker model, we consider the connectivity of the system when setting the required security mechanisms. Each security mechanism possesses a different strength level which can be ranked. We have defined five protection levels, where P1 represents no protection and P5 represents the strongest protection mechanisms. Table 1(a) shows the evaluation of the exposure level from connectivity and protection levels. The evaluation of the protection level is conducted by security experts. In [12] we have not considered protection mechanisms in detail (standards like ANSSI also do not). Here we detail this important part of our SGSC by ranking various security mechanisms, focusing on the Smart Home application domain. We extend [12] by extracting the security criteria (classified in Table 2) for evaluating protection levels based on the following standards and best practices:

Criteria for Security Classification of SHEMS

161

ISO 27002, which does not cover the IoT systems; CSA the IoT Working Group of the Cloud Security Alliance; IIC the Industrial Internet of Things Volume G4 Security Framework; OWASP “IoT Security Guidance”; and ETSI TS 103 645 “Cyber Security for Consumer Internet of Things”.

Table 1. Calculations of (a) exposure levels and (b) security classes. P1

E4 E4 E5 E5 E5

Catastrophic

A

C

E

F

F

P2

E3 E4 E4 E5 E5

Major

A

B

D

E

F

P3

E2 E3 E3 E4 E4

Moderate

A

B

C

E

E

P4

E1 E1 E2 E2 E3

Minor

A

A

B

D

D

P5

E1 E1 E1 E1 E2

Insignificant

A

A

A

C

C

Protection/Connectivity C1 C2 C3 C4 C5

Impact/Exposure E1 E2 E3 E4 E5

Table 2. Referred sources for the construction of security criteria. Protection criteria

Source

Data encryption

ISO 27002, OWASP, ETSI

Communication and connectivity protection IIC, ISO 27002, ETSI Software/Firmware security

ISO 27002, OWASP, ETSI

Hardware-based security controls

CSA

Access control

ISO 27002, OWASP, IIC, CSA, ETSI

Cryptography techniques

IIC, ISO 27002

Physical and environmental security

ISO 27002, OWASP, CSA

Monitoring and analysis

ISO 27002, OWASP, IIC, CSA, ETSI

We detail further the security criteria with security functionalities inspired by the IoT Security Compliance Framework proposed by the IoT Security Foundation (IoTSF), which is in the form of a checklist. Table 3 shows the mapping of security criteria to security functionalities and protection level.

4

Applying the Security Classification to SHEMS

The SHEMS in our case complies with the reference architecture from Sect. 2 and consists of a centralized IoT hub and smart plugs connected to controllable loads such as water heater, air conditioner, and floor heating. For simplicity, we do not include storage devices such as batteries, that can act as both load and generation device. We first identify the criticality (Impacts) of successful cyberattacks on SHEMS.

162

M. Shrestha et al. Table 3. Protection level requirements.

Protection criteria

Security functionality

Data encryption

Encryption of data between system components

x

x

x

Strong encryption mechanism

x

x

x x

Credentials should not be exposed in the network

x

x

End-to-end encryption

x

x

Should not use custom encryption algorithms

x

x

Sensitive stored data should be encrypted

x

x

x

x

x

x

x

x

x

x

x

Use only standard communication protocol

x

x

Updatability of device firmware

x

x

Updatability of the operating system

x

x

Automatic updates available

x

x

Encryption of update files

x

x

Signing update files before installing

x

x

Using Trusted Platform Modules (TPM)

x

x

Use of Memory Protection Units (MPUs)

x

x

Incorporate physically unclonable functions

x

x

Use of cryptographic modules

x

x

Communication and Have a minimal number of network ports open connectivity Devices should not be accessible from the Internet protection Only authorized components can join the network Software/firmware security

Hardware-based security controls

Access control

Cryptography techniques

Physical and environmental protection

Monitoring and analysis

P5 P4 P3 P2

Disable remote access functionality

x

Only authorized devices can join the network

x

x

x

Default and weak passwords should not be used

x

x

x

Secure bootstrapping

x

x

Secure key generation/storage/rotation

x

x

Secure key distribution

x

x

x

Message integrity

x

x

x

Tamper resistance

x

x

Minimal physical ports available

x

x

x

Physical security of connections

x

x

x

Ability to disable external ports and only minimal ports enabled

x

x

Only authorized physical access

x

x

Monitoring system components

x

x

Analysis of monitored data

x

x

Act on analysed data

x

x

x

Safety. Leakage of data from SHEMS may disclose the presence of people inside their house, which may result in burglary or other types of crime. Moreover, residents may feel unsafe (reducing trust in SHEMS) if they realize that their privacy is breached and strangers can follow their activities. Grid Imbalance. During the execution of a demand response program, devices that utilize higher energy are turned off to shave the peaks. If an attacker can switch on/off a large number of loads, these are capable of destabilizing the grid [8,13], possibly destroying equipment and physical infrastructure.

Criteria for Security Classification of SHEMS

163

Increased Electricity Bills. Compromising SHEMS may result in equipment being switched on without authorized persons noticing. Privacy. Data from SHEMS can be privacy sensitive, e.g., [9] has demonstrated that High-frequency consumption data can be exploited to derive private information such as number of people in the house, sleep routines, and the presence of babies at home. Compromised SHEMS data may contain even more detailed information. Stealing such data may result in the exposure of personal habits of the residents, which can impact social reputation. Agents for Other Cyberattacks. Typically, smart home gateways have connectivity to the Internet. A compromised gateway may act as a bot to launch several other attacks. Among these impacts, grid imbalance and agents for other cyberattacks can be considered major, as they may result in blackouts and damage of physical infrastructures. The remaining impacts could be considered moderate or minor. We limit the presentation of the application of security classification only to Application and Network Data. In particular, we assess the Command and Control (C&C) functionality for a demand-response program, which is one of the most critical components of SHEMS. We apply the classification method for the following two scenarios. Scenario I: Centralized Control. In this scenario, Distribution System Operators (DSO) have an agreement with consumers to control the SHEMS appliances to properly manage peaks of energy demand. In our system, each controllable device is plugged into the corresponding smart plug. Depending on the device and their maximum effect, rules for controlling the devices are defined, e.g., a water heater with a maximum capacity 3 kW can be controlled only between 8:00 AM to 6:00 PM during weekdays, and once turned off, it cannot be turned on for a minimum of 15 min. The DSOs forecast the energy demand in advance and, if reductions are needed, they optimally select the devices to be turned off. Control commands are sent to the selected devices from the DSO to meet the goal of targeted reduction of energy consumption. Class Evaluation. The connectivity between the IoT device and the hub is C3 (cf. Sect. 3) and between the hub and the backend system is C5. If an attacker is able to control the device only inside the HAN (C3), the impact is only Minor. However, if an attacker can trigger or manipulate the message for the demand control program from the backend (C5), several devices can be turned off, resulting in grid imbalance as discussed earlier. As a result, for this scenario, we evaluate the overall impact as Major. To evaluate the security class, we first select the relevant security criteria for C&C as Data Encryption, Communication and Connectivity Protection, Access Control, and Monitoring and Analysis. We then evaluate the protection level based on the strength of the mechanisms in these selected criteria. Due to space limitations, we do not discuss every detail of our specific evaluations (see technical report [11]. Using Table 3 we assign the overall protection level P4.

164

M. Shrestha et al.

Using Table 1(a) we determine from the computed values of connectivity (C5) and protection level (P4), the exposure E3. Using Table 1(b) we get the class D (Impact Major and Exposure E3), which is a score not suitable for SHEMS. To improve the security class, Table 1(b) indicates that either exposure or impacts need to be reduced. Similarly, exposure can be reduced either by increasing the protection level or by reducing the connectivity, cf. Table 1(a). Scenario II: Edge Control. In this scenario, the control signals are sent by the IoT hub autonomously, based on the time of peak demand or price of electricity, and thresholds set by the end-user. Users can also set priorities for the devices that can be controlled and rules to decide, e.g., when and how long the devices can be controlled. Thresholds and rules can also be persisted in the IoT gateway, allowing control of devices without requiring interaction with the backend. Class Evaluation. Similarly, if an attacker can manipulate the control message within the HAN network (C3), the impact is considered as Minor. However, since there is no flow of commands from the backend system, an attacker cannot influence many devices on a large scale. Since there are no changes in the protection mechanisms, we can consider it as P4. Moreover, using Table 1(a), we obtain the Exposure E2 and using Table 1(b) we computed the security class as A. The analyses of scenario I and II showed that by moving from the centralized control to the edge control for the demand control functionality, the security class of the demand control is significantly improved from class D to class A. In addition, scenario II may even be more efficient and have lower latency because the trigger of device control initiates locally rather than from the backend system to several IoT devices. Such improvements in the design of IoT systems should be considered to improve the security of the overall system.

5

Conclusion and Further Work

We have presented a security classification methodology for SHEMS. We have applied this methodology to the commercial SHEMS from E2U. The example focusses on the C&C part of SHEMS for demand-response programs. First, we evaluated the security class of the C&C for a centralized control architecture, resulting in a low and not acceptable security class D. Using our methodology, we are able to indicate how the system needs to be improved to achieve an acceptable security class. Using an edge controlling concept, our analysis demonstrated an achievable security class A. Further work will focus on aggregation mechanisms for calculating the overall system security class from its components.

Criteria for Security Classification of SHEMS

165

References 1. Aldrich, F.K.: Smart homes: past, present and future. In: Inside the Smart Home, pp. 17–39. Springer (2003) 2. ANSSI: Classification method and key measures (2014) 3. Celebucki, D., Lin, M.A., Graham, S.: A security evaluation of popular internet of things protocols for manufacturers. In: ICCE, pp. 1–6. IEEE (2018) 4. Fitriaty, P., Shen, Z., Sugihara, K.: How green is your smart house: looking back to the original concept of the smart house. In: Green City Planning and Practices in Asian Cities, pp. 39–76. Springer (2018) 5. Ghirardello, K., Maple, C., Ng, D., Kearney, P.: Cyber security of smart homes: development of a reference architecture for attack surface analysis (2018) 6. Lee, J.I., Choi, C.S., Park, W.K., Han, J.S., Lee, I.W.: A study on the use cases of the smart grid home energy management. In: ICTC, pp. 746–750. IEEE (2011) 7. Liu, Y., Qiu, B., Fan, X., Zhu, H., Han, B.: Review of smart home energy management systems. Energy Procedia 104, 504–508 (2016) 8. Mohsenian-Rad, A.H., Leon-Garcia, A.: Distributed internet-based load altering attacks against smart power grids. IEEE Trans. Smart Grid 2(4), 667–674 (2011) 9. Molina-Markham, A., Shenoy, P., Fu, K., Cecchet, E., Irwin, D.: Private memoirs of a smart meter. In: Proceedings of the 2nd ACM Workshop on Embedded Sensing Systems for Energy-Efficiency in Building, pp. 61–66. ACM (2010) 10. Redmill, F.: Understanding the use, misuse and abuse of safety integrity levels. In: 8th Safety-critical Systems Symposium, pp. 8–10 (2000) 11. Shrestha, M., Johansen, C.: Criteria for security classification of smart home energy management systems (long version). Technical report 476, University of Oslo (2019) 12. Shrestha, M., Johansen, C., Noll, J., Roverso, D.: A methodology for security classification applied to smart grid infrastructures. Int. J. Criti. Infrastruct. Protect. 28, 100342 (2020) 13. Soltan, S., Mittal, P., Poor, H.V.: BlackIoT: IoT botnet of high wattage devices can disrupt the power grid. In: 27th USENIX Security Symposium, pp. 15–32 (2018) 14. Stojkoska, B.L.R., Trivodaliev, K.V.: A review of internet of things for smart home: challenges and solutions. J. Clean. Prod. 140, 1454–1464 (2017) 15. Zhou, B., Li, W., Chan, K.W., Cao, Y., Kuang, Y., Liu, X., Wang, X.: Smart home energy management systems: concept, configurations, and scheduling strategies. Renew. Sustain. Energy Rev. 61, 30–40 (2016)

Secure Linear Regression Algorithms: A Comparison Fida Dankar(&) and Nisha Madathil College of IT, United Arab Emirates University, 15551, Al Ain, Abu Dhabi, UAE {fida.dankar,nisha.t}@uaeu.ac.ae

Abstract. The problem of secure linear regression calculation has been widely considered in the literature. It involves multiple parties, with a private dataset each, wanting to collectively carry out linear regression on the union of their datasets but are unable to combine the data due to privacy restrictions. The solutions suggested in the literature use different methods from cryptography to securely calculate the regression parameters while keeping the parties’ data private. In this paper, we compare the different algorithms in terms of security, efficiency and accuracy. Keywords: Data privacy


Secure multiparty computation
Linear regression

1 Introduction In many real-life settings, investigators wish to analyze data from multiple separate databases but are unable to combine the data due to restrictions such as privacy concerns, privacy laws or the sheer size of the data. The scenarios in which data sharing is desired are multiple, investigators may need to increase the statistical accuracy of a study and lower bias [1], they may wish to perform benchmarking studies [2], or to attain a required cohort particularly in studies involving rare diseases [3]. Such scenarios could benefit from secure multiparty computations (SMC), which allow multiple parties (or sites) to collectively carry out calculations on their datasets without having to reveal their own private data [4]. 1.1

Secure Multiparty Computation

SMC addresses the above problem by providing a mechanism for the joint computation over the private input of different parties without revealing anything apart from the final output and what can be derived from it. SMC protocols differ in their security guarantees (cryptographic security (CS) or information theoretic security (IS)) [4], as well as in the underlying security model used: the semi-honest security model assumes that all parties follow the protocol correctly, but some may be curious and (cooperatively) try to infer information about the private input of other parties. The malicious security model assumes that the semi-honest parties may also be malicious and deviate from the protocol. In both security models the protocol must guarantee that no knowledge about the private inputs is gained and that the correct output is propagated to all participating © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 166–174, 2020. https://doi.org/10.1007/978-3-030-53187-4_20

Secure Linear Regression Algorithms: A Comparison

167

parties. Protocols with malicious parties are harder to construct and less efficient than the semi-honest one. It is thus common in the data mining applications to adopt the semi-honest model, supported by the fact that data holders are usually genuinely interested in the result of the computation. 1.2

Contributions

In this paper, we consider the problem of linear regression calculation when data is held by multiple parties that are not willing or able to share it due to privacy concerns. Multiple SMC protocols have been suggested in the literature for the above problem. We evaluate and compare these algorithms for accuracy and efficiency. As efficiency has been the main factor against the adoption of these algorithms, we will compare semi-honest protocols only, thus, assuming a genuine interest among all parties in learning the outcome of the computation. In terms of security, we consider the protocols that do not share or reveal any raw data or intermediate computation results, thus excluding some of the earlier distributed protocols [5–7]. Moreover, we focus on the protocols that can handle horizontal data partition [3, 8–11]. To our knowledge, this is the first comprehensive comparison of secure linear regression protocols. The paper is organised as follows: Sect. 2 reviews available algorithms and classifies them based on their features. Section 3 describes the experiments performed and discusses the results. The paper is concluded in Sect. 4 with some remarks and limitations.

2 Algorithms 2.1

Evaluation Criteria and Notation

 We introduce the classical setting of a linear regression problem. Let X ¼ xi;j be an N  p matrix of features and Y ¼ ðy1 ; . . .; yN ÞT a corresponding N  1 response vector, where N is the number of samples and p is the number of features. Linear regression consists of modeling the relationship between the set of features (also known as predicting variables) and the response variable. It assumes that the relationship between the response variable and the predicting variables is linear. Fitting a linear regression model consists of feature selection and parameter estimation [12], feature selection is the process of constructing a model that includes all relevant predicting variables. In other words, it is the process of determining the subset of features that best predicts the outcome variable Y. While the parameter estimation consists of finding the linear model parameters b where Y ¼ Xb þ 2 [12]. There are many attempts in the literature at obtaining secure linear regression protocols over m distributed databases (m refers to the number of parties). As privacy is our main concern, we consider the protocols that do not share or reveal raw data or intermediate computation results. The remaining protocols differed in: (i) their security guarantees (IS or CS and the maximum number of corrupt parties they can handle), (ii) the number of third parties/servers employed to help with the computation, (iii) whether model selection is performed (or just parameter estimation), and

168

F. Dankar and N. Madathil

(iv) which partition of raw data they can handle. Essentially, data owned by the different sites can be partitioned vertically, horizontally or a using a mixture of both. Vertical partitions imply that each database holds the same subjects but includes different attributes. Horizontal partitions imply common attributes across the sites’ databases but different subjects. Mixture partitions imply some combination of both partitions. We confine the comparison to semi-honest protocols (non-malicious) that handle horizontally partitioned data. In the next subsection, we review the available secure regression algorithms and classify them based on the aspects listed above (Table 1). 2.2

Secure Regression Algorithms

The first linear regression algorithm with cryptographic security is due to Hall et al. [10], the algorithm uses homomorphic encryption to estimate the regression parameters, and provides only the final result of the computation to the participating parties. The paper reported 2 days for solving a linear regression problem of 51k rows and 22 features [7]. Nikolaenko et al. presented a solution using a combination of garbled circuits and homomorphic encryption in [11]. The algorithm (referred to here as Garbled) uses two non-colluding semi-honest third parties to calculate the parameters of the linear regression. The servers are trusted not to collude with each other or with other parties, thus supporting a maximum of one semi-honest party (one server). The usage of garbled circuits imposed many rounds of interactions and is thus heavy on communication. In the same set of experiments cited above [7], the algorithm required 8.75 h for 108 records with 20 features and a 270 MB of communication, thus significantly outperforming the algorithm of Hall et al. De Cock et al. presented in [9], a method to calculate the parameters of the linear regression by computing b ¼ ðX T XÞ1 X T Y. The algorithm (referred to here as DC) uses secret sharing with the help of a trusted initializer. b is calculated by running Beaver’s matrix multiplication protocol many times [13]. The initial and intermediate matrices are always secret shared among the different parties, and the output of the multiplication algorithm is also secretly shared among them. DC is secure even if all parties except one are semi-honest -as the trusted initializer is only involved in the setup phase and does not engage in the computation phase. The algorithm is information theoretic secure. The theoretical complexity of DC is OðNp2 Þ, however, the protocol is heavy on communication. The multiplication protocol used requires each party to send 2 matrices to every other party (of size p2 Þ, and is repeated OðkÞ, where k is the maximal number of bits required to represent the largest integer. However, experiments done by the authors indicated a capacity to handle over 4 million records with 16 features in the range of 3 h [9]. The protocol is presented for two parties, however, we extended it (along the lines specified in the paper) to a multiple parties’ protocol in our implementation. In a recent paper, Mohassel et al. presented a secure linear regression algorithm based on stochastic gradient descent (referred to as SecureML or SML). Similar to Garbled [11], they use two non-colluding servers [8]. The different data holders

Secure Linear Regression Algorithms: A Comparison

169

(or parties) use secret sharing to distribute their input between the two servers. The servers then use secure two-party protocols to calculate the parameters of the linear regression. The algorithm assumes no collusion between the two servers, and can thus allow a maximum of one semi-honest server. In a series of experiments, the authors report the ability to run millions of records with thousands of features. Recently, Dankar et al. presented the first secure linear regression algorithm (SMA) that performs model selection as well as parameter estimation [3]. They use a semi-trusted third party (a role that is performed by one of the sites) to help with the algorithm execution and methods from statistics to conduct computations (model selection and parameter estimation) on each of the sites independently and then combine these computations (securely) to form one estimator for the collective dataset. Thus, limiting communication to the final step in the estimation process and reducing complexity. The authors discuss at length the necessary assumptions on the model parameters to guarantee the accuracy of the result, in general they require that the number of records per site be bigger than the number of features and number of sites (n [ p and n [ mÞ. They present two versions of their algorithm with varying security level [14]. One version allows only one party at most to be semi-honest (referred to here as SMA1), while the other version uses threshold encryption and is robust against any (pre-defined) number, t, of semi-honest parties (SMAt). In experiments done by the authors themselves, the algorithm required 20 min for 108 records with 50 features. The results exhibit a significant reduction in computation time, however one needs to check whether this comes at the cost of accuracy. A major source of inaccuracy in secure computations (aside from the inability of most algorithms to perform model selection) stems from the need for computation in a finite field. Solutions that rely on encryption (SMA, Garbled, SML) tend to work in a very large finite field and treat decimal numbers as integers [10, 11], or use fixed point multiplication of decimal numbers with fixed decimal bits [8]. DC uses secret sharing and thus relies on some uniformly random blinding factors. As it is not possible to sample uniformly from infinite fields, it uses a modified version of Catrina et al. [15] truncation protocol in order to approximate the computations on real numbers. Table 1. Categorization of the compared algorithms. Algorithms name Garbled – [11] DC – [9] SecureML – [8] SMAt – [3]

Data partition Horizontal Any Any Horizontal

Number of servers/third parties 2, non-colluding 1, trusted initializer 2, non-colluding 1, trusted initializer

Security Max num of corrupt parties CS 1 (server) IS n1 CS 1 (server) CS t; t\n

Model selection No No No Yes

170

F. Dankar and N. Madathil

We compared all the algorithms listed above apart from the algorithm of Hall et al. [10], as prior comparison with the Garbled algorithm showed a considerably slower performance [7]. Table 1 above catalogues the algorithms in terms of the data partition they execute, the number of third parties/initializers required, the security guarantees, the maximum number of acceptable corrupt parties, and whether they perform model selection.

3 Experiments 3.1

Experimental Set-Up

We compared the Secure Multiparty Linear Regression algorithms by implementing them and analyzing the results using real and synthetic datasets. The real datasets were used to test the accuracy of the algorithm while the synthetic datasets were used to compare the performance. We used Python3 as our programming language, which we augmented with the Scikit-learn, numpy, pandas, gmpy2 sympy, secrets crypto and phe libraries for functionality such as socket programming, homomorphic encryption, implementation of garbled circuit, and for dealing with negative and real floating-point arithmetic. We built our system on top of 10 Linux machines with Intel(R) Core(TM) i5-4570 CPU, 3.20 GHz processor, and 16 GB RAM, 4 cores each. The accuracy of the algorithms was tested in comparison with the central algorithm (Cent). The central algorithm performs linear regression on the clear (where data is pooled together in one location). Thus, we require the secure algorithms to produce estimates that are accurate compared to what would be produced if the data was pooled to one place. Our implementation of the central algorithm uses lasso for feature selection and linear least squares method for parameter calculation [16]. 3.2

Experiments on Real Datasets

To test the accuracy of the algorithms we needed real datasets that specify the original collection site. Every site is then treated as an independent data owner. We used 4 real datasets, 3 public datasets contained within the UCI repository and one from Cerner clinical database. The datasets are explained in details in Appendix A, and the parameters of the datasets are summarized in Table 2.

Table 2. The parameters of the real datasets used (number of sites, number of features, number of records, mean and standard deviation of the outcome variable Y). Data set m p N 1. StudentsPerf-Portuguese 2 30 649 2. StudentsPerf-Maths 2 30 395 3. AutoMPG 3 7 392 4. Diabetes (selected features) 6 8 6414 5. Diabetes (with ‘weight’ feature) 6 39 456 6. Diabetes (excluding ‘weight’) 12 38 21205

Y Mean/Std dev 11.906/3.228 10.415/4.576 23.446/7.795 4.361/3.0245 4.414/3.023 4.389/2.972

Secure Linear Regression Algorithms: A Comparison

171

We randomly divided the datasets into 70% training set and 30% testing set. A regression model was trained based on the training set and used to predict the outcome variable in the testing set. The mean of the square prediction error was used to evaluate the model (MSE). The experiments were repeated 50 times each. The results are presented in Table 3. Table 3. MSE of the different algorithms per dataset, and the ratio of the Cent-MSE with the 4 different algorithms tested. ME refers to memory error. Data set MSE Cent 1 3.364 2 7.554 3 13.285 4 7.071 5 7.443 6 5.705

MSE ratio (MSE-Cent/MSE) SMA DC SML Garbled SMA DC SML Garbled 3.417 4.132 4.789 4.531 0.984 0.814 0.702 0.742 7.719 9.971 11.375 10.95 0.979 0.758 0.664 0.690 17.56 14.924 28.21 25.34 0.772 0.909 0.481 0.535 7.283 15.564 17.277 16.785 0.971 0.454 0.409 0.421 7.733 ME 11.84 10.01 0.962 ME 0.629 0.743 5.875 ME 7.95 8.22 0.971 ME 0.718 0.694

As evident from the results, the SMA algorithm has the highest accuracy, this is mainly due to the fact that it is able to perform model selection. For dataset 3, the seven features were all selected by the central algorithm (in almost all 50 rounds, rendering the model selection process redundant), which lead to DC having the highest accuracy. 3.3

Experiments on Synthetic Datasets

For efficiency comparison, we performed a scalability analysis to evaluate the performance of the algorithms as the data size, the number of features and the number of parties’ increase. The synthetic datasets was generated in Python using sklearn.datasets.make_regression (which generates a random regression problem). The total number of records, N, was varied between 100,000 and 100 million, number of features, p, between 2 and 50 and number of sites m, between 2 and 10. The records are always divided equally between the sites. The results are displayed in Figs. 1, 2 and 3.

Fig. 1. Time taken by all algorithms as a function of the number of sites when n ¼ 104 (left) or n ¼ 105 (right) and p ¼ 10. Note that the left graph displays all the results while the right one excludes the DC algorithm.

172

F. Dankar and N. Madathil

When the number of records per site and the number of features are fixed (n and p fixed), the time taken by all algorithms increases with the increase in the number of sites, m (Fig. 1). The SMA algorithms show the least variation in time while the remaining algorithms vary noticeably. Figure 2 displays time performance when the total number of records, N and the number of features p are fixed (Fig. 2). Note that, when N is fixed, n decreases with the increase in m, thus (1) lowering the local per site complexity, and (2) increasing the overall inter-party communication. The SML and the SMA algorithms show small time increase (thus a domination of (2)) while Garbled and DC shows a small decrease in time.

Fig. 2. Time taken by the different algorithms as a function of the number of sites when N ¼ 105 (left graph), N ¼ 106 (right graph), and p ¼ 3. Note that the DC algorithm is excluded from the left graph for displaying memory error at m ¼ 2.

Figure 3 (right), displays the time taken by the SMA algorithm at different levels of security. It shows that the time taken by the algorithm increases when higher security is required. Figure 3 (left) also provides a comparison between SMAm−1 and DC as they are both secure against m  1 corrupt parties (they have the highest level of security among the compared algorithms).

Fig. 3. The graph on the left shows a comparison between the DC algorithm and the SMAm−1 as the number of sites varies, with n ¼ 104 and p ¼ 5. The right graph shows the time taken by different SMAs with varying the security level (number of corrupt parties t varies between 1, m3 ; and m  1) when n ¼ 107 and p ¼ 50.

Secure Linear Regression Algorithms: A Comparison

173

Overall, the SMA algorithm shows consistent superior performance over all other algorithms, while also providing high security (Figs. 1, 2 and 3).

4 Conclusion We compared the Secure Multiparty Linear Regression algorithms by implementing them and analyzing the results using real and synthetic datasets. The real datasets were used to test the accuracy of the algorithms and the synthetic data for efficiency. All datasets used are low dimensional and satisfy n [ p and n [ m. In the performance evaluation, we distinguished between two testing strategies: n fixed or N fixed. Fixed n represents a problem of site constraints in terms of processing capacity or data availability while fixed N represents a requirement for a fixed number of records. For accuracy, the algorithms were compared against the central algorithm where data is deployed on the clear. Overall, the SMA algorithm shows consistent superior performance over all other algorithms, while also providing the desired accuracy and a high security level. Future experiments should relax the requirement on data dimensionality and test the accuracy using additional real datasets.

Appendix A To test the accuracy of our algorithm we needed real datasets that include information about the original collection site (in other words, the sites where the data originates from should be known) and every site is treated as an independent data owner. We were able to find 4 real datasets, 3 public datasets contained within the UCI repository and one from Cerner clinical database: Student Performance Data Set (Portuguese Performance and Math Performance): These two datasets are related to student performance in two Portuguese schools for two subjects: Portuguese language and Mathematics. The total number of students is 649 with 30 attributes. The data attributes include first term grades, demographics, social and school related features. The target attribute is the final year grade (out of 20). Both datasets were divided into 2 sites based on the school name. Auto-Mpg Data: This dataset is related to car fuel consumption in miles per gallon. It contains 392 records with 9 features including car origin (manufacturer), cylinders, weight, acceleration, and miles per gallon. The variable to be predicted is mile per gallon (mpg). The dataset is divided into 3 sites according to the ‘origin’. The number of records in the sites were 245, 68 and 79 respectively. (Note that this dataset was normalized using python library Scikit-Learn). Diabetes Data Set: This is a clinical diabetic dataset with 101,767 records and 41 features, including hospital_id, race, gender, age, weight, admission type, glucose measurements (at different times), insulin doses, A1C results, and diagnosis. The response variable is ‘times in hospital’, that is the length of stay in hospital with a range of 1 to 14 days. The dataset was divided into multiple sites based on the ‘hospital_id’.

174

F. Dankar and N. Madathil

We performed several experiments on this dataset by varying the number of sites used (6 or 12 sites) and number of features (8, 38, or 39). In the last two datasets (5 and 6), the same features were selected in both except for weight which we excluded from dataset 6. In fact, the ‘weight’ is an important predictor for diabetes, however it is missing from 90% of the records.

References 1. El Emam, K., Samet, S., Hu, J., Peyton, L., Earle, C., Jayaraman, G.C., Wong, T., Kantarcioglu, M., Dankar, F., Essex, A.: A Protocol for the secure linking of registries for HPV surveillance. PLoS ONE 7, e39915 (2012). https://doi.org/10.1371/journal.pone.0039915 2. El Emam, K., Arbuckle, L., Essex, A., Samet, S., Eze, B., Wang, L., et al.: Secure surveillance of antimicrobial resistant organism colonization or infection in Ontario. PLoS ONE 9(4), e93285 (2014) 3. Dankar, F.K., Madathil, N., Dankar, S.K., Boughorbel, S.: Privacy-preserving analysis of distributed biomedical data: designing efficient and secure multiparty computations using distributed statistical learning theory. JMIR Med. Inform. 7, e12702 (2019) 4. Lindell, Y., Pinkas, B.: Secure multiparty computation for privacy-preserving data mining. J. Priv. Confid. 1, 5 (2009) 5. Karr, A.F., Lin, X., Sanil, A.P., Reiter, J.P.: Secure regression on distributed databases. J. Comput. Graph. Stat. 14, 263–279 (2005) 6. Du, W., Han, Y.S., Chen, S.: Privacy-preserving multivariate statistical analysis: linear regression and classification. In: Proceedings of the 4th SIAM International Conference on Data Mining (2004) 7. Aono, Y., Hayashi, T., Trieu, L., Wang, P.L.: Fast and secure linear regression and biometric authentication with security update. IACR Cryptol. EPrint Arch. 2015, 692 (2015) 8. Mohassel, P., Zhang, Y.: Secureml: a system for scalable privacy-preserving machine learning. In: 2017 IEEE Symposium on Security and Privacy (SP), pp. 19–38. IEEE, May 2017 9. De Cock, M., Dowsley, R., Nascimento, A.C.A., Newman, S.C.: Fast, privacy preserving linear regression over distributed datasets based on pre-distributed data. In: Proceedings of the 8th ACM Workshop on Artificial Intelligence and Security, pp. 3–14. ACM, New York (2015). https://doi.org/10.1145/2808769.2808774 10. Hall, R., Fienberg, S.E., Nardi, Y.: Secure multiple linear regression based on homomorphic encryption. J. Off. Stat. 27, 669 (2011) 11. Nikolaenko, V., Weinsberg, U., Ioannidis, S., Joye, M., Boneh, D., Taft, N.: Privacypreserving ridge regression on hundreds of millions of records. In: 2013 IEEE Symposium on Security and Privacy, pp. 334–348. IEEE, May 2013 12. Montgomery, D.C., Peck, E.A., Vining, G.G.: Introduction to Linear Regression Analysis. Wiley, Hoboken (2012) 13. Beaver, D.: One-time tables for two-party computation. In: International Computing and Combinatorics Conference, pp. 361–370. Springer (1998) 14. Dankar, F.K., Boughorbel, S., Badji, R.: Using robust estimation theory to design efficient secure multiparty linear regression. In: Proceedings of the 2016 Joint EDBT/ICDT Workshops (2016) 15. Catrina, O., Saxena, A.: Secure computation with fixed-point numbers. In: International Conference on Financial Cryptography and Data Security, pp. 35–50. Springer (2010) 16. Gray, J.B.: Applied Regression Analysis, Linear Models, and Related Methods. Taylor & Francis, London (1998)

Multi-agents Intrusion Detection System Using Ontology for Manets Sara Chadli1(B) , Hajar Chadli1 , Mohammed Saber2 , Mohammed Ghaouth Belkasmi2 , Ilhame El Farissi2 , and Mohamed Emharraf2 1

Laboratory LES, Sciences Faculty, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected], [email protected] 2 Laboratory SmartICT, ENSAO, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected], [email protected], [email protected], [email protected]

Abstract. lately, Mobile ad hoc network continues to increase their existence, hence becoming a essential in several fields. However, these networks operate without any dedicated infrastructure; therefore they are vulnerable to a numerous of threats. The traditional securing networks methods are unsufficient and currently requires to couple a reactive security solution, such as an Intrusion Detection System (IDS). However, most existing IDS suffer from a number of drawbacks, e.g., high rates of false positives, low efficiency, etc. Using a semantic resource such as ontology could be an effective way to enrich the data on intrusions. This paper presents hybrid architecture based on multi-agent system, which uses complete ontology by considering the intrusion at a higher level of abstraction, also to improve the detection process. Keywords: Mobile ad hoc network · Intrusion Detection System (IDS) · Multi-agent system · Ontology

1

Introduction

Mobile ad hoc networks (MANETS) are wireless networks that do not benefit from pre-existing infrastructure or centralized administration for the exchange of information and services provided to users. The topology of these networks is formed according to the appearance and movement of the nodes. The latter communicate with their neighbors via point-to-point wireless links and provide themselves the routing function. As a result, there is no hierarchy between the nodes and no network service can claim to be centralized. These networks are inherently more vulnerable and more difficult to protect than wire line networks. In fact, in a wireless network, access to exchanged data is immediate for any node equipped with a suitable network interface, while c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 175–182, 2020. https://doi.org/10.1007/978-3-030-53187-4_21

176

S. Chadli et al.

it is necessary to have a physical connection in a wired network. Additionaly, the implementation of some security mechanisms developed for wired networks is difficult or impossible in MANETs. Due to their spontaneous nature, they cannot benefit from infrastructure-based security mechanisms, such as a firewall or an authentication server. Consequently, each node is a point of vulnerability that can only rely on its own resources and services to protect itself. Also the vulnerabilities are already identified in wire line networks and often accentuated in the ad hoc context, these networks have unique vulnerabilities such as those related to the physical layer as well as those in connection to the network layer. Therefore, mobile ad hoc network security is becoming a major challenge for both users and administrators of these systems. To address the issue of security, a policy must be defined in regard to the security objectives that we seek to achieve. This policy represents the properties of confidentiality, integrity and availability to ensure a prominent and best security. An Intrusion Detection Systems (IDS) has been of use to detect and defend intrusions more proactively in short period. Even that IDSs have become a standard component in security infra-structures, they still have a number of significant drawbacks [1]. Yet, using a semantic resource such as ontology could be an effective way to enrich the data concerning the intrusions in order to answer more precised complecated questions about the nature and the characteristic of the intrusion, especially to collect new information on possible new intrusions to allow a better integration. In this paper, we will use the classifications proposed in [2–5] to establish a more complete ontology in considering the intrusion at a higher level of abstraction. The paper is organized as follows. Section 2 describes our proposed architecture and Sect. 3 presents a detailed ontology proposal. In Sect. 4, a conclusion and future work are presented.

2

The Proposed Architecture of IDS

In [6–8], We propose a hybrid architecture of IDS, which combine between the hierarchical model and the cooperative model based on a multi-agent system. (see Fig. 1). The proposal architecture is consists of a multi-agent detection system that uses five classes of agents (Fig. 1): Sensor Agent (SA), Manager Agent (MA), Ontology Agent (OA), Agent actuator (ACA) Agent, Analyzer (ANA). The main role of agents is: – Sensor Agent (SA): it captures network raw traffic, and sends them to other agents to be analyzed and processed. – Analyzer Agent: Receives data from Sensor Agent, and compare signatures with predefined patterns. – Ontology agent: help analyzer agent to determine the nature of suspicious activity by using a new model of classification of attacks stored in its data base. – Manager agent: The manager agent can ask other agents from other nodes for local information related to suspicious activity, in this case one or more agents

Multi-agents Intrusion Detection System Using Ontology for Manets

177

Fig. 1. Proposed intrusion detection system

analyzer located in different nodes in cluster can provide local information to the initiator. – Actuator agent: generates alarms and logs. The proposal IDS detects rapidly the attacks through the ontology agent which uses new classification [4,5] to enrich data intrusions and attack signatures by semantic relationships. In what follows, we present the proposed ontology used within our system.

3

Proposed Ontology

The concept of ontology has been used for a very long time, especially in linguistics and the processing of natural languages, ontology defines the terms used to describe and represent an experimental field. Ontologies are used by people, databases and applications, that need to share information about a specific area such as medicine, tool making, financial management... In recent years, Raskin [9] had focus on using Ontology within information security. In fact, ontologies can be used as basic components to provides continuous analysis, and a way to explore known and unknown attacks. The major reasons for using ontology in the proposed IDS, has to do with its effenciency in interoper-ability between nodes. Also it improves the process and quality of the engineering systems. We will, therefore, use the classifications proposed in [4,5,10] to establish a more complete ontology by considering the intrusion at a higher level of abstraction, as well as to improve the detection process. Ontology formally defines the terms used to describe and represent a domain of knowledge. The goal is to allow sharing, reuse and reasoning on built knowledge. The proposed ontology is built around three major classifications that represent the three main areas of knowledge:

178

S. Chadli et al.

– A high-level ontology includes abstract and general concepts of the field of intrusions. This ontology subsumes the concepts existing in the various intrusions. – A general ontology of domain is the one that allows to federate the concepts and relations central of the domain of intrusionsis the one that allows to federate the concepts and relations central of the domain of intrusions. – A low level ontology handles the operational aspects of intrusion detection solutions. 3.1

A High-Level Ontology

Each intrusion in MANETS can be described according to a very precise diagram. We associate three distinct sources constituting the highest level of abstraction as shown in Fig. 2.

Fig. 2. A high-level ontology

The first source is the Security services. It presents the objective of the detection. The second source security policy which represents the security policy aspects implemented to monitor the different nodes constituting Mobile ad hoc network The last source is environment. It highlights the impact of the environment that may occur during a transmission or during the reading of any physical quantity. 3.2

A General Ontology of Domain

The general ontology of the domain describes the various aspects of the domain considered. It provides the necessary characteristics to define and identify the elements causing the intrusions in each domain. In the field of security service, Fig. 3, protection and monitoring are the main sources of information that can lead to intrusion detection. In the area of security-policy, Fig. 4 groups all the concepts related to the intrusion. Several classes are described and they show a clear presentation of security policy for mobile ad ho network.

Multi-agents Intrusion Detection System Using Ontology for Manets

179

Fig. 3. A security-service of general ontology

Fig. 4. A security-policy of general ontology

In the field of environment; Fig. 5 shows the two main classes: Manets Components and vulnerabilities. The first represents the components of a Manets and the different relationships existing between them. We can find the batterypower class which represents how the lack of energy in a battery can lead to a

180

S. Chadli et al.

malfunction. The Bandwidth class, it consists on various known attacks on the wireless transmission. Finally the Routing-protocol class. The second class is Vulnerability, it represents the most exhibiting catalysts a network manets to attacks,

Fig. 5. A environment of general ontology

3.3

A Low Level Ontology

It links implementation examples and algorithmic approaches to different topologies used to operate different detection systems. At the operational level for security service, we detail for example the class IDS-components in Fig. 6. It describes the different elements needed to detect intruders in the surveillance zone.

Fig. 6. A IDS-component of low level ontology

At the operational level for security policy, we detail intrusion-detectionsystem-IDS class in Fig. 7.

Multi-agents Intrusion Detection System Using Ontology for Manets

Fig. 7. Intrusion-detection-system IDS of low level ontology

181

182

4

S. Chadli et al.

Conclusion

This paper offers a formal description of different intrusions in the manets networks using an ontology. We attempt to bring elements to identify and characterize the various intrusions in a global way in manets. Our suggestted solution uses a high level of abstraction to improve the detection process. The future work of this research is to develop detailed detection strategies against complicated distributed intrusions or attacks and test the Ontology in real cases.

References 1. Pinzon, C.I., De Paz, J.F., Herrero, A., Corchado, E., Bajo, J., Corchado, J.M.: idMAS-SQL: intrusion detection based on MAS to detect and block SQL injection through data mining. Inf. Sci. 231(10), 15–31 (2013). https://doi.org/10.1016/j. ins.2011.06.020 2. SOBHTS: Wired and wireless intrusion detection system: classifications, good characteristics and state-of-the-art. Comput. Stand. interfaces 28, 670-694 (2006) 3. Gu, L., Jia, D.: Lightweight detection and classification for wireless sensor networks in realistic environments. In: Proceedings of the 3rd International Conference on Embedded Networked Sensor Systems (SenSys 2005), pp. 205–217, November 2005 4. Chadli, S., Saber, M., Emharraf, M., Ziyyat, A.: Generation of test-cases of attacks in MANETs. IEEE Xplore Digital Libr. 10, 855–860 (2014). https://doi.org/10. 3844/jcssp.2016.495.501 5. Chadli, S., Saber M., Ziyyat A.: Defining categories to select representative attack test-cases in MANETs. In: Proceeding The 2014 Fourth International Conference on Communication Systems and Network Technologies (CSNT 2014), 7–9 April 2014, Bhopal, MP, India, pp. 658–663. IEEE (2014). ISBN 978-1-4799-30692/14/$31.00c 6. Chadli, S., Emharraf, M., Saber, M., Ziyyat, A.: The design of an ids architecture for MANET based on multi-agent. In: 2014 Third IEEE International Colloquium in Information Science and Technology (CIST), pp. 122-128, October 2014. https:// doi.org/10.1109/CIST.2014.7016605 7. Chadli, S., Saber, M., Ziyyat, A.: Implementation an intelligent architecture of intrusion detection system for MANETs, pp. 479-487. Springer (2016). https:// doi.org/10.1007/978-3-319-30298-0 49 8. Chadli, S., Emharraf, M., Saber, M., Ziyyat, A.: Combination of hierarchical and cooperative models of an ids for MANETs. In: 2014 Tenth International Conference on Signal-Image Technology and Internet-Based Systems (SITIS), pp. 230-236, November 2014. https://doi.org/10.1109/SITIS.2014.32 9. Raskin, V., Hempelmann, C.F., Triezenberg, K.E., Nirenburg, S.: Ontology in information security: a useful theoretical foundation and methodological tool. In: Proceedings of the 2001 Workshop on New Security Paradigms, NSPW 2001, Cloudcroft, New Mexico, p. 5359 (2001) 10. Ngankam Kenfack, H., Ndi´e, T.D., Nataf, E., Festor, O.: Une ontologie pour la description des intrusions dans les RCSFs. In: CFIP 2011 - Colloque Francophone sur l Ing´enierie des-Protocoles, UTC, May 2011, Sainte Maxime, France (2011)

Analysis of KDD Dataset Categories to Design a Performing Intrusion Detection System Ilhame El Farissi1(B) , Mohammed Saber1 , Sara Chadli2 , Zineb Bougroun1 , Mohamed Emharraf1 , Mohammed Ghaouth Belkasmi1 , and Rachida El Mehdi1 1

Laboratory SmartICT, ENSAO, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected], {m.saber, r.elmehdi}@ump.ac.ma, [email protected], [email protected], [email protected] 2 Laboratory LES, Sciences Faculty, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected]

Abstract. An Intrusion Detection System (IDS) is a mechanism which is intended to spot the abnormal activities in a network traffic. The recent systems are based on intelligent methods such as the Artificial Neural Network, Na¨ıve Bayes, Random Tree... In fact, due to the learning capacity of the intelligent methods especially the Artificial Neural Network, the IDS becomes able to detect the known attacks and also the unknown or the recent ones. For this reason, it is crucial to use an extensive database in learning phase and also to test the system performance. KDD dataset is the most commonly known set and contains a large variety content. The aim of our research consists on exploiting the relevant data of the KDD data set in order to obtain the optimum system based on neural network. Keywords: Intrusion Detection System · Artificial Neural Network KDD dataset · Evaluation metrics · KDD categories

1

·

Introduction

The attackers try to illegally access into a system in order to steal data or to disrupt its performance. So, it is of vital importance to enhance the security system. For this reason, additional security options must be implemented into the system either to prevent attacks or to detect them. Several researches have been realized in this area. The recent ones are generally based on detecting attacks by distinguishing between a normal behavior and an abnormal one by using the performance of an intelligent method [1]. The Artificial Neural Network is one of the most popular intelligent method which have this ability and respond to this need [2,3]. c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 183–191, 2020. https://doi.org/10.1007/978-3-030-53187-4_22

184

I. El Farissi et al.

Furthermore, in order to build an Artificial Neural Network (ANN), it is required to go through two major phases; learning and test. During these phases, the Neural Network must be alimented by pertinent data to get good results. The KDD dataset [4] is exploited in this context because it is a large dataset with solid parameters. Each record in the KDD dataset depends on four categories; the Basic attributes (B), the attributes which are related to content (C), the attributes which are based on the time using windows of two-second time (T), and the time-based attributes using windows of 100 connections time (H). However, it is essential to select the optimum categories to increase the ANN performance [5,6]. Therefore, we have generated all possible combinations between categories and we have designed several scenarios with different input data. Moreover, we have realized a comparative study between the obtained results from each scenario. This study is based on some evaluation metrics; True Positive (TP), True Negative (TN), False Positive (FP), False Negative (FN), Accuracy, Precision, Sensitivity, Specificity, False Negative Rate (FNR), False Positive Rate (FPR) and F-Score. The rest of the paper is organized as follows: Sect. 1 contains the main concepts of this work. Section 2 presents the related works. Section 3 describes the proposed method. Section 4 presents the obtained results. Finally, the conclusion is presented in Section 5.

2 2.1

Literature Review KDD Dataset

The KDD [4] is the most widely used dataset for the building of an Intrusion Detection System. This is due to its extensive content. KDD dataset contains several records which are divided into two types; bad connections (attacks) and normal connections. Each record depends on 42 features [7]. The first 41 ones represent the record parameters and the last one indicates the status connection which means “bad” or “normal” connection. Furthermore, the 41 parameters are divided into four categories: The Basic attributes (B), the attributes which are related to content (C), the attributes which are based on the time using windows of two-second time (T), and the time-based attributes using windows of 100 connections time (H). However, it is essential to select the optimum categories to increase the ANN performance. 2.2

Intrusion Detection System

An intrusion is an anomaly which has occurred in a system [8]. The Intrusion Detection System is a mechanism of monitoring the system activity in order to detect the generated anomaly and to properly react.

Analysis of KDD Dataset Categories to Design a Performing IDS

2.3

185

Artificial Neural Network

The Artificial Neural Network (ANN) functioning is inspired from the Biological Neural Network. In fact, an ANN contains several neuron layers interconnected between them. Its strength lies in its ability to detect, prevent, classify and learn [9–12]. In fact, the ANN especially the Multi Layer Percpetron (MLP) goes through the training and the test phases. During the first phase the neural network learns and then it moves into the second phase to validate its performance. 2.4

Matlab Tool

We have opted for the use of Matlab tool because it provides an easy-to learn and accessible environment and it contains also a neural network toolbox which presents a framework for designing and implementing neural networks [13]. 2.5

Metrics Evaluation

In order to evaluate a system performance, it is crucial to put into practice some indicators also called the evaluation metrics. The frequently used metrics and with high degree of performance are: – True Positive (TP): the number of cases correctly identified as attack. – True Negative (TN): the number of cases correctly identified as a normal behavior of the network. – False Positive (FP): the number of cases incorrectly identified as attack. – False Negative (FN): the number of cases incorrectly identified as a normal behavior of the network. P +T N – Accuracy: ( T P +TTN +F P +F N ) This measure indicates the ability of system to differentiate the normal and abnormal behaviors correctly. P – Precision: ( T PT+F P ) Called also Detection Rate determines how many of the positively classified were relevant. P – Sensitivity: ( T PT+F N ) Called also recall, it indicates the ability to determine the attack cases correctly. N – Specificity: ( T NT+F P ) The specificity of a test is its ability to determine the normal cases correctly. P – False Positive Rate (FPR): ( T NF+F P ) Called also False Alarm Rate, it indicates how many of positively classified should be negative results. N – False Negative Rate (FNR): ( T PF+F N ) It determines how many of negative cases should be classified as attacks. recision×Recall – F-score: ( 2×P P recision+Recall ) This measure is the balance between precision and recall.

186

3

I. El Farissi et al.

Related Works

Several researchers have realized IDSs based on an intelligent method and alimented by the KDD data. But, the main problem of KDD dataset stems from the lowest interest of some features. In order to overcome this problem, it is of a vital importance to examine the existing studies. Furthermore, some researchers have opted for the classification of input data by determining the attack class and eliminating the features with no impact and some others have concentrated their efforts on detection by distinguishing the normal and bad connection without specifying the attack category. By using random forest, J48, bayes net, na¨ıve bayes, support vector machine, classification via regression, decision table, random table, rules-OneR and Hoeffding Tree, [14] have realized a comparison study based on some evaluation metrics. The summary of the obtained results is indicated in the Table 1. Moreover, the main purpose of [15] was to implement a classifier based on the random tree method which should be able to achieve the maximum detection rate and the minimum false alarm rate. So, knowing that the KDD features belong to one of the four categories; Basic, Content, Traffic or Host, [16] has set up fifteen scenarios. The results are summarized in the Table 1. Table 1. Preview of the obtained results in related works Classifier

Attribute class combinations False alarm rate Detection rate

Random Forest

BCTH

0.12

0.96

J48

0.13

0.95

Bayes Net

0.12

0.94

Na¨ıve Bayes

0.32

0.88

SMO

0.33

0.87

Classification via Regression

0.13

0.95

Decision Table

0.13

0.94

Random Tree

0.36

0.90

Rules-OneR

0.15

0.95

Hoeffding Tree

0.34

0.90

Random Tree

BCTH

0.08

0.77

BCT

0.08

0.79

BCH

0.03

0.73

BTH

0.03

0.69

CTH

0.07

0.61

BC

0.09

0.76

BT

0.08

0.72

BH

0.06

0.71

CT

0.07

0.61

CH

0.07

0.64

TH

0.08

0.62

B

0.07

0.81

C

0.24

0.79

T

0.06

0.53

H

0.08

0.61

Analysis of KDD Dataset Categories to Design a Performing IDS

187

The aim of our study consists on designing and developing a performing IDS by using the Artificial Neural Network and selecting uniquely the crucial categories. Moreover, we have implemented a number of scenarios. According to the established scenario, the network architecture changes which allowed us to compare the obtained results and select the crucial parameters category. In addition, we have used the evaluation metrics; True Positive, True Negative, False Positive, False Negative, Accuracy, Precision (Detection Rate), Sensitivity, Specificity and F-score to measure the IDS performance.

4

Research Methodology

The scope of our work is to design and develop a performing IDS by using the Artificial Neural Network and selecting uniquely the crucial categories from the KDD dataset. Thus, from the four features categories, we have generated fifteen scenarios as shown in Table 2. Consequently, the established network architecture changes according to the scenario taken into account. And we have exploited the evaluation metrics mentioned above to realize a comparative study and measure the IDS performance. Table 2. Preview of the obtained results in related works Scenario num Combinaisons B C T √ S1 B X X √ S2 C X X √ S3 T X X S4

H

S5

BC

S6

BT

S7

BH

S8

CT

S9

CH

S10

TH

S11

BCT

S12

BCH

S13

BTH

S14

CTH

S15

BCTH

X X X √ √ X √ √ X √ X X √ √ X √ X X √ X X √ √ √ √ √ √

X √ X √ √

X √ √ √

H # of train patterns # test patterns X 45747 X

3691

116

126

X 21560 √ 29208

18959

X 45926

36921

X 79965 √ 71925

63942

X 22451 √ 29714 √ 89724

19711

X 80246 √ 71934 √ 116461 √ 90269 √ 116468

63949

20802

54793 21160 56723 54823 77256 57074 77284

188

5

I. El Farissi et al.

Results and Discussions

First of all, by using Matlab tool, a matrix confusion is generated for each scenario. This matrix represents the obtained results in terms of TP, TN, FP and FN values. Based on these values, the other metrics are automatically calculated by using the mathematical formulas mentioned above. For best visibility, we have opted to diagnose the obtained results in three axes. On one hand, the performance of an IDS depends in its ability to differentiate the normal and the abnormal behaviors correctly. So, the accuracy is the appropriate indicator to extract the scenario with the highest ability. The histogram Fig. 1 represents the obtained results in terms of accuracy according to the fifteen scenario in test phase.

Fig. 1. Accuracy values

The histogram Fig. 1 indicates that the basic category is necessary to have a performing IDS. On the other hand, we have considered the scenario which produces the maximum precision and the maximum sensitivity and also the minimum False Positive Rate such as the scenario with the best ability to determine the attacks cases. According to the diagram Fig. 2, the scenario BCT gives best results in terms of precision, sensitivity and FPR. Finally, to accomplish our study, we have decided to add a comparison between scpecifity and False Negative Rate to obtain the scenario with the highest ability to determine the normal cases normally. Referring to the diagram Fig. 3, the best compromise specificity/FNR is produced by the BT scenario.

Analysis of KDD Dataset Categories to Design a Performing IDS

189

Fig. 2. Precision, sensitivity and FPR values

Fig. 3. Specificity and FNR values

In sum, the Basic category is necessary to obtain a performing ID but it is not sufficient. Consequently, it is of a vital importance to add the Content and Traffic categories to increase the Detection Rate and decrease the False Negative Rate.

6

Conclusion

In order to build an IDS which based on neural network, it is essential to exploit a dataset containing pertinent data. The KDD dataset is generally the most exploited dataset in this context. The KDD contains four parameters categories; Basic, Content, Traffic and Host. Some of these categories have no impact in Intrusion Detection. Consequently, we have realized this study which consists on evaluating the impact of each parameters category to build a performing IDS based on neural network. In fact, from the four parameters categories we have generated fifteen scenarios and fifteen different neural network architectures and we have established a

190

I. El Farissi et al.

comparative study by using the following evaluation metrics; True Positive, True Negative, False Positive, False Negative, Accuracy, Precision (Detection Rate), Sensitivity, Specificity and F-score. From the obtained results, we conclude that the Basic category is required to obtain a performing IDS but not sufficient. So, it is crucial to add the Content and Traffic categories to increase the Detection Rate and decrease the False Negative Rate.

References 1. Depren, O., Topallar, M., Anarim, E., Ciliz, M.K.: An Intelligent Intrusion Detection System (IDS) for anomaly and misuse detection in computer networks. Expert Syst. Appl.: Int. J. 29(4), 713–722 (2005). https://doi.org/10.1016/j.eswa. 2005.05.00 2. Saber, M., Chadli, S., Emharraf, M. and El Farissi, I.: Modeling and implementation approach to evaluate the intrusion detection system. In: Networked Systems, pp. 1–5. Springer (2015). https://doi.org/10.1007/978-3-319-26850-7 41 3. Saber, M., Belkasmi, M.G., Chadli, S. and Emharraf, M.: Implementation and performance evaluation of network intrusion detection systems, pp. 484-495. Springer (2017). https://doi.org/10.1007/978-3-319-68179-5 42 4. KDD data set. http://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html 5. Akhlaq, M., Alserhani, F., Awan, I., Mellor, J., Cullen, A.J., Al-Dhelaan, A.: Implementation and evaluation of network intrusion detection systems. In: Network Performance Engineering, pp. 988–1016. Springer, Heidelberg. https:// doi.org/10.1007/978-3-642-02742-0 42 6. Sabhnani, M., Serpen, G.: Why machine learning algorithms fail in misuse D detection on KDD intrusion detection data set. In: Intelligent Data Analysis, vol. 6 (2004) 7. Revathi, S., Malathi, A.: A detailed analysis of KDD cup99 Dataset for IDS. Int. J. Eng. Res. Technol. (IJERT) 2, 12 (2013) 8. Saber, M., Emharref, M., Bouchentouf, T., Benazzi, A.: Platform based on an embedded system to evaluate the intrusion detection system. IEEE Xplore Digital Libr. 894–899 (2012). https://doi.org/10.1109/ICMCS.2012.6320253 9. El Farissi, I., Saber, M., Chadli, S., Emharraf, M., Belkasmi, M.G.: The analysis performance of an intrusion detection systems based on neural network. IEEE Xplore Digital Libr. 145-151 (2017). https://doi.org/10.1109/CIST.2016.7805032 10. El Farissi, I., Chadli, S., Emharraf, M., Saber, M.: The analysis of KDD-parameters to develop an intrusion detection system based on neural network, pp. 491-503. Springer (2017). https://doi.org/10.1007/978-981-10-1627-1 39 11. Saber, M., El Farissi, I., Chadli, S., Emharraf, M., Belkasmi, M.G.: Performance analysis of an intrusion detection systems based of artificial neural network. Springer (2017). https://doi.org/10.1007/978-3-319-46568-5 52 12. Slimani, I., El Farissi, I., Achchab, S.: Application of game theory and neural network to study the behavioral probabilities in supply chain. J. Theor. Appl. Inf. Technol. 82(3) (2015) 13. Matlab. https://fr.mathworks.com/products.html 14. Aggarwal, P., Sharma, S.K.: An empirical comparison of classifiers to analyze intrusion detection. In: 2015 Fifth International Conference on Advanced Computing & Communication Technologies. https://doi.org/10.1109/ACCT.2015.59

Analysis of KDD Dataset Categories to Design a Performing IDS

191

15. Aggarwal, P., Sharma, S.K.: Analysis of KDD dataset attributes - class wise for intrusion detection. Procedia Comput. Sci. 57, 842–851 (2015). https://doi.org/ 10.1016/j.procs.2015.07.490 16. Selvakumar, B., Muneeswaran, K.: Firefly algorithm based feature selection for network intrusion detection. Comput. Secur. 81, 148–155 (2019). https://doi.org/ 10.1016/j.cose.2018.11.005

A Comparative Performance Analysis of the Intrusion Detection Systems Mohammed Saber1(B) , Zineb Bougroun1 , Ilhame El Farissi1 , Sara Chadli2 , Mohamed Emharraf1 , Saida Belouali3 , Mohammed Ghaouth Belkasmi1 , and Ilham Slimani1 1

Laboratory SmartICT, ENSAO, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected], [email protected], [email protected], [email protected], [email protected], [email protected] 2 Laboratory LSE, Sciences Faculty, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected] 3 Laboratory LIDICOM, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected]

Abstract. In a comparative analysis, this paper investigates the performance of two open source intrusion detection systems (IDSs) namely SNORT and SURICATA for accurately detecting the malicious traffic on computer networks, an evaluation approach, based on a series of tests. These experiments consisted of a test bed which compared SNORT and SURICATA’s reaction; consist in injecting various traffic loads, characterized by different transmission times, packet numbers, packet sizes and bandwidths, and then analyzing, for each situation, the processing performed on the packets. The study demonstrates that SURICATA would process a higher speed of network traffic than SNORT with lower packet drop rate but it consumed higher computational resources. Keywords: Intrusion detection comparison · Traffic network

1

· SNORT · SURICATA · Performance

Introduction

Recently, attacks made on computer networks have risen dramatically. These attacks are made at various layers in the TCP/IP protocol suite. The attackers act like normal users, generate data and hide their malicious activities under terabytes of data. The monitoring of the network traffic allows to detect malicious activities and perform analysis to differentiate the malicious and non malicious user activities to protect their networks. Detecting malicious activities require intrusion detection systems (IDS). It is critical that an IDS detection mechanism c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 192–200, 2020. https://doi.org/10.1007/978-3-030-53187-4_23

A Comparative Performance Analysis of the Intrusion Detection Systems

193

is accurate enough to differentiate between legitimate and malicious traffic that enter and leave the network. The possible results of using an IDS are as follows: detected malicious traffic (real alarms), undetected malicious traffic, legitimate traffic that IDS detect as malicious (false alarms) and legitimate traffic that IDS detect as good. Intrusion detection is difficult to be accomplished perfectly. With the volume of network traffic rapidly increasing and the number and complexity of network attacks increasing just as quickly, it becomes increasingly difficult for a signaturebased intrusion-detection system to keep up with the current threats. The evaluation performance of intrusion detection systems is a challenging task; it requires a thorough knowledge of techniques relating to different disciplines, especially intrusion detection, methods of attack, networks and systems, technical testing and evaluation. Lately; more research is done on the evaluation of IDS, we cite some research in this area. In [1–4] the researchers evaluated the performance of three IDSs in an environment consisted of physical and virtual computers. The experiment results showed that SNORT could have a negative impact on network traffic more than the other two tested IDSs. In [5] a study demonstrated the lack of ability of SNORT IDS to process a number of packets at high speed and the packet drop rate was higher. The researchers introduced a parallel IDS technology to reduce the packet drop rate as a solution. The proposed approach significantly improved SNORT performance [6]. In papers [7,8] an evaluation of SNORT performance against DDoS. The experiments results show that SNORT packet handling could be improved by using better hardware configurations, but SNORT detection capability was not improved by using better hardware. In [9], authors have tested and analysed the performance of SNORT and SURICATA. In [10], a comparison analysis of the performance of two open-source intrusion detection systems, SNORT and SURICATA is presented, by evaluating the speed, memory requirements, and accuracy of the detection engines in a variety of experiments. In [11] analysed and implemented the SNORT intrusion detection model in a campus network. In [12] presents a thorough comparison of the performance of SNORT and SURICATA. They examine the performance of both systems as they scale system resources. There are other works that looks at measuring the intrusion detection capability as in tweaking IDS performance as in [13], parallel design of IDS on many-core processors. In [14], an approach for unifying rule based deep packet inspection and in [15], improving the accuracy of network intrusion detection systems. Whereas in [16], boosting throughput of SNORT NIDS under Linux. As in [17], evaluation studies of three IDS under various attacks and rule sets. In [18], evaluation based in classification of networks attacks [19]. The evaluation based in optimizes and analysis performance of an Intrusion Detection Systems, it is primordial to exploit uniquely the most important and crucial parameters of each features category KDD [20–22], etc. This paper is structured as follows: 2nd Section is devoted to the proposed approach for evaluating performance of the IDSs. Then a presentation and a

194

M. Saber et al.

discussion of the obtained results are shared in Sect. 3. Finally, 4th Section which is dedicated to a conclusion and future works.

2

The Proposed Approach for Evaluating Performances IDS

The experiments consisted of a test bed which compared the performance of both IDSs. Experiment scenarios were designed to make observations and to take measurements. This study demonstrates rigorous, repeatable, quantitative performance comparison of both IDSs. The network traffic for the experiments was produced using network traffic generator. The default rule set of SNORT and SURICATA were used for the experiments. 2.1

Experiment Scenarios

– Scenario 1 (Consumed resources): The experiment compared the performance of both IDSs by measuring the percentage of CPU (Central Processing Unit), memory utilisation, with different traffic rates. – Scenario 2 (Normal traffic accuracy measurements): The experiment was planned to determine the accuracy for both IDS ruleset inspected the network traffic to correctly classify the legitimate traffic network. – Scenario 3 (Response to high-speed network traffic): The experiment compared the performance of both IDSs by measuring network packet drop rate, by the transmission of the packets (1460 bytes in size) at different transmission time frames (1, 4, 8 and 16 ms). 2.2

Experiment Network

Fig. 1. Experiment network

To perform those tests, the SURICATA version 4.1.1 and the SNORT version 2.9.8.3 are selected. And for experimentations, the network shown in Fig. 1 is

A Comparative Performance Analysis of the Intrusion Detection Systems

195

created. Four computers were used. Depending on the individual experiment requirements, network packets were produced at varying network speeds with network traffic generator tools. The four computers were connected via a switch CISCO SG350XG-2F10 using 10 Gbps Ethernet links. Each IDS was separately installed on identical computers with default performance parameters and rule set. 2.3

Performance Metrics

The metrics listed below in Table 1 are used to measure the detection accuracy of both IDSs. Table 1. Description of performance metrics Performance metrics

Description

False Positive Rate (FPR)

This is the likelihood that the IDS will trigger an alarm when there is no intrusion

False Negative Rate (FNR) This is the likelihood that the IDS does not trigger an alarm when there is an intrusion True Positive Rate (TPR)

This is the likelihood that IDS trigger an alarm when an intrusion is detected

Packets captured (PCA)

The number and percentage of packets received

Packets analysed (PAN)

The number and percentage of packets analysed from the total packets captured

Packets dropped (PDR)

The number and percentage of the packets dropped from the total packets captured

3 3.1

Experiment Scenarios Results and Evaluation Experiment Scenario One : Consumed Resources

This first experimentation supervises the real-time performance of SNORT and SURICATA while processing at a different normal network speed from a network traffic generator. The rational behind the first experiment is to compare SNORT to SURICATA’s performance. To achieve accurate results, the experiment scenario is tested with packets size of 1460 bytes. These packets were injected to both IDSs with a different network speed. The experiment consisted of a logical network diagram as shown in Fig. 1. Each IDS was separately installed on identical computers with default performance parameters and rule set. A number of tools were used to observe and record the measurements of CPU, memory, network utilisation and the packet drop rate. The following packets were injected as the background traffic ranging from a different network speed (100 Mbps, 250 Mbps, 500 Mbps, 750 Mbps, 1.0 Gpbs, 2.0 Gbps and 4.0 Gbps).

196

M. Saber et al.

The collected performance data shows that SURICATA’s CPU usage is greater than SNORT’s as explained in Table 2. SURICATA’s CPU utilisation is increased with different traffic rates, while SNORT’s CPU utilisation is comparatively less using the same metrics. Table 2. (%) CPU and Memory (GB) utilisation for SNORT and SURICATA for different traffic rates Traffic rate (%) CPU utilisation Memory utilisation (GB) SNORT SURICATA SNORT SURICATA 100 Mbps

7

10

0.5

0.7

200 Mbps

12

17

0.7

1

250 Mbps

14

20

0.8

1.3

500 Mbps

23

30

1

1.7

750 Mbps

32

39

1.5

2.1

1.0 Gbps

39

48

1.8

2.7

2.0 Gbps

47

58

2.2

3.2

4.0 Gbps

55

68

2.4

3.5

The collected performance data shows that SURICATA’s memory usage is greater than SNORT’s as presented in Table 2. SNORT’s memory usage was comparatively less. SURICATA’s memory usage has to do more with the multithreaded architecture. 3.2

Experiment Scenario Tow: Normal Traffic Accuracy Measurements

This experiment is planned to determine how accurately SNORT’s and SURICATA’s rule set in order to inspect the network traffic and correctly classify the non malicious traffic. The metrics listed above in Table 1 are used to measure the detection accuracy of both IDSs. Table 3. Normal traffic accuracy measurements Normal traffic SNORT SURICATA FPR FNR TPR FPR FNR TPR UDP

13% 0%

0%

22%

3%

0%

TCP

9% 0%

0%

33% 10%

0%

ICMP

2% 0%

0%

41% 29%

5%

The second experiment analyses the detection accuracy of SNORT and SURICATA while processing the legitimate network traffic. Both the IDSs were kept

A Comparative Performance Analysis of the Intrusion Detection Systems

197

at the default setting. The accuracy test was performed using the legitimate network traffic generator which injected UDP, TCP and ICMP packets to both IDSs, the results are shown in Table 3. SURICATA’s false positive rate (FPR) was higher when processing UDP, TCP and ICMP packets than SNORT’s FPR. However, SNORT did not trigger true positive rate (TPR 0%) and false negative rate (FNR 0%) alarms. As compared to SURICATA, this triggered a 41% FNR and 5% TPR. Therefore, SNORT triggered less false positive alarms. While false negative alarms are observed in both IDSs, SNORT’s detection accuracy is found to be superior to SURICATA in this scenario. 3.3

Experiment Scenario Three: IDSs Response to High-Speed Network Traffic

For this third scenario, the packets 1460 bytes in size (∼ =100000 TCP, and ∼ =100000 UDP, and ∼ =100000 ICMP) are sent at different transmission time frames (1 ms, 4 ms, 8 ms, and 16 ms) for the both systems. Figure 2 shows both IDSs output and obtained results.

Fig. 2. IDSs response to high-speed network traffic

As demonstrated in the results shown in Fig. 2, all the sent packets reached their destinations. Both IDSs have analysed almost all packets in incoming traffic when packets are transmitted in 16 ms time frame. But when the speed of transmission is decreased, both IDSs start to drop packets. The collected performance data showed that SNORT’s dropped packets is greater than that of SURICATA for the metrics (1 ms, 4 ms and 8 ms) as in Fig. 2.

198

M. Saber et al.

In this experiment, it is noticed that for both IDSs; the analysis performance decreased as the speed of transmission is increased. Therfore, the components ability of analysis becomes weaker as the transmission speed is increased.

4

Conclusion

The main contribution of this study is the comparison of the intrusion detection performance of two open source IDSs, namely SNORT and SURICATA. Both are proved to be efficient and high performing IDS, although each one has its own strengths and weaknesses. The analysis of the experiment results shows that SNORT utilises less computational resources to process network traffic whereas SURICATA’s utilisation was higher. Also, SURICATA processes a higher number of packets per second as compared to SNORT, and both IDSs have a high rate of false positives alarms. The obtained results demonstrate a number of signifficant limitations in the use of both IDS. This work identifies specific and replicable bottlenecks in commonly used implementations IDS in high-speed networks. The obtained results can be taken as a benchmark to improve the performance of these systems in future research work.

References 1. Wang, X., Kordas, A., Hu, L., Gaedke, M., Smith, D.: Administrative evaluation of intrusion detection system. In: Proceedings of the 2nd Annual Conference on Research in Information Technology (RIIT 2013), pp. 47–52. ACM, New York (2013). https://doi.org/10.1145/2512209.2512216 2. Saber, M., Chadli, S., Emharraf, M., El Farissi, I.: Modeling and Implementation Approach to Evaluate the Intrusion Detection System. Springer (2015). https:// doi.org/10.1007/978-3-319-26850-7 41 3. Shahbaz, M.B., Wang, X., Behnad, A., Samarabandu, J.: On efficiency enhancement of the correlation-based feature selection for intrusion detection systems. In: 2016 IEEE 7th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), Vancouver, BC, pp. 1–7 (2016). https://doi.org/ 10.1109/IEMCON.2016.7746286 4. Saber, M., Belkasmi, M.G., Chadli, S., Emharraf, M.: Implementation and performance evaluation of network intrusion detection systems, pp. 484–495. Springer (2017). https://doi.org/10.1007/978-3-319-68179-5 42 5. Bulajoul, W., James, A., Pannu, M.: Network intrusion detection systems in highspeed traffic in computer networks. In: 2013 IEEE 10th International Conference on e-Business Engineering, Coventry, pp. 168–175 (2013). https://doi.org/10.1109/ ICEBE.2013.26 6. Trabelsi, Z., Zeidan, S.: IDS performance enhancement technique based on dynamic traffic awareness histograms. In: 2014 IEEE International Conference on Communications (ICC), Sydney, NSW 2014, pp. 975–980 (2014). https://doi.org/10.1109/ ICC.2014.6883446

A Comparative Performance Analysis of the Intrusion Detection Systems

199

7. Saboor, A., Akhlaq, M., Aslam, B.: Experimental evaluation of SNORT against DDoS attacks under different hardware configurations. In: 2013 2nd National Conference on Information Assurance (NCIA), Rawalpindi, pp. 31–37 (2013). https:// doi.org/10.1109/NCIA.2013.6725321 8. Saber, M., Emharref, M., Bouchentouf, T., Benazzi, A.: Platform based on an embedded system to evaluate the intrusion detection system. IEEE Xplore Digital Libr. 894–899 (2012). https://doi.org/10.1109/ICMCS.2012.6320253 9. Alhomoud, A., Munir, R., Disso, J.P., Awan, I., Al-Dhelaan, A.: Performance evaluation study of intrusion detection systems. Procedia Comput. Sci. 5, 173–180 (2011). https://doi.org/10.1016/j.procs.2011.07.024, ISSN 1877-0509 10. Karim, I., Vien, Q.-T., Le, T.A., Mapp, G.: A comparative experimental design and performance analysis of SNORT-based Intrusion Detection System in practical computer networks. Computers 6(1), 1–15 (2017). https://doi.org/10.3390/ computers6010006, ISSN 2073-431X 11. Huang, C., Xiong, J., Peng, Z.: Applied research on SNORT intrusion detection model in the campus network. In: 2012 IEEE Symposium on Robotics and Applications (ISRA), Kuala Lumpur, pp. 596–599 (2012). https://doi.org/10.1109/ISRA. 2012.6219259 12. White, J.S., Fitzsimmons, T., Matthews, J.N.: Quantitative analysis of intrusion detection systems: SNORT and SURICATA. In: Proceedings of SPIE - (2013) The International Society for Optical Engineering, vol. 8757. https://doi.org/10.1117/ 12.2015616 13. Hafeez, K., Masood, M., Malik, O., Anwar, Z.: LASSP: a logic analyzer for tweaking SNORT security and performance. In: 2010 6th International Conference on Emerging Technologies (ICET), Islamabad, pp. 240–245 (2010). https://doi.org/ 10.1109/ICET.2010.5638483 14. Jiang, H., Zhang, G., Xie, G., Salamatian, K., Mathy, L.: Scalable highperformance parallel design for network intrusion detection systems on many-core processors. In: Proceedings of the Ninth ACM/IEEE Symposium on Architectures for Networking and Communications Systems (ANCS 2013), Piscataway, NJ, USA, pp. 137–146. IEEE Press (2013) 15. Munoz, A., Sezer, S., Burns, D., Douglas, G.: An approach for unifying rule based deep packet inspection. In: 2011 IEEE International Conference on Communications (ICC), Kyoto, pp. 1–5 (2011). https://doi.org/10.1109/icc.2011.5963095 16. Papadogiannakis, A., Polychronakis, M., Markatos, E.P.: Improving the accuracy of network intrusion detection systems under load using selective packet discarding. In: Proceedings of the Third European Workshop on System Security (EUROSEC 2010), pp. 15–21. ACM, New York (2010). https://doi.org/10.1145/ 1752046.1752049 17. Salah, K., Qahtan, A.: Boosting throughput of SNORT NIDS under Linux. In: 2008 International Conference on Innovations in Information Technology, Al Ain, pp. 643–647 (2008). https://doi.org/10.1109/INNOVATIONS.2008.4781733 18. Thongkanchorn, K., Ngamsuriyaroj, S., Visoottiviseth, V.: Evaluation studies of three intrusion detection systems under various attacks and rule sets. In: 2013 IEEE International Conference of IEEE Region 10 (TENCON 2013), Xi’an, pp. 1–4 (2013). https://doi.org/10.1109/TENCON.2013.6718975 19. Saber, M., Bouchentouf, T., Benazzi, A., Azizi, M.: Amelioration of attack classifications for evaluating and testing intrusion detection system. J. Comput. Sci. 6(N7), 716–722 (2010). https://doi.org/10.3844/jcssp.2010.716.722

200

M. Saber et al.

20. El Farissi, I., Saber, M., Chadli, S., Emharraf, M., Belkasmi, M.G.: The analysis performance of an Intrusion detection systems based on neural network. IEEE Xplore Digital Libr. 145–151 (2017). https://doi.org/10.1109/CIST.2016.7805032 21. El Farissi, I., Chadli, S., Emharraf, M., Saber, M.: The analysis of KDD-parameters to develop an intrusion detection system based on neural network, pp. 491–503. Springer (2017). https://doi.org/10.1007/978-981-10-1627-1 39 22. Saber, M., El Farissi, I., Chadli, S., Emharraf, M., Belkasmi, M.G.: Performance analysis of an intrusion detection systems based of artificial neural network. Springer (2017). https://doi.org/10.1007/978-3-319-46568-5 52

New Improvement of Malware-Attack Scenarios Modeling Noureddine Rahmoun(&), Yassine Ayachi, Jamal Berrich, Mohammed Saber , and Toumi Bouchentouf LSE2I, National School of Applied Sciences, Mohammed First University, Oujda, Morocco [email protected], [email protected], [email protected], [email protected], [email protected]

Abstract. It is important to state that, from the intrusion detection point of view, the number of steps that appears in a certain session of an attack process is arbitrary. Indeed, in order to prevent detection, attackers can proceed slowly, following different steps, in many days, or even in many weeks. The main objective of this study is to learn how to shape the progression of an attack process in time. We propose a stochastic model based on Markov chains. The basic assumption of our model consists of evaluating malware valid attack scenarios. The algorithm we have proposed shows that we will be able to identify an attack scenario while the attack process is not yet complete which will help IDS to improve the detection rate of malware attacks. Keywords: Intrusion detection
Markov chain model
Evaluating
Malware
IDS


Attacks process
Stochastic

1 Introduction Nowadays, the area of security evaluation is an important concern in ongoing security research. The guarantee of absolute security level for security-critical systems and networks is a very big challenge. In line with other studies (e.g., [1–4]), we also believe in the impossibility of providing an absolute amount of security; some of the most important reasons for this claim are as follows: • The interaction environment between a system and its users is highly unpredictable and in such conditions, and complicated, the exact prediction of the behavior of attackers will be so difficult. A number of questions must be answered to specify the nature of an attack process: When has the attack been initiated? From where has the attack been controlled and originated? How has the attack been conducted? Who has conducted the attack? What is the target of the attack? The problem is that potential attackers have access to a large number of efficient and simple methods and tools for hiding or altering this type of information. In addition, intrusion prevention or detection techniques may not be sufficient to prevent or detect

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 201–211, 2020. https://doi.org/10.1007/978-3-030-53187-4_24

202

N. Rahmoun et al.

malicious and suspected behaviors. As a result, there is always a high probability that a complex attack process against a system would be successfully performed. • The lessons learned from past current and experience position have confirmed that the presence of inherent vulnerabilities, which are the common path of the intrusions, in developed software and hardware components and communication protocols during their entire life, would be almost unavoidable. As one of the most important causes of this problem, designers do not pay particular attention to security-related issues, in the design phases of systems, due efficiency constraints and to cost. Additionally, many systems in their default configurations are completely vulnerable to attacks. • Nowadays, even novice attackers should be regarded as real dangers. With freely available tools, conducting attacks against systems is a highly automated process, and compared to the past position needs a lower skill level and has a higher probability of success. Attacks can cause greater amounts of damage and have become more automated. Indeed, as stated in [5], computer and network security is essentially a battle of wits between a perpetrator who tries to find holes and the designer or administrator who tries to close them. The great advantage that the attacker has is that he or she need only find a single weakness, while the designer must find and eliminate all weaknesses to achieve perfect security. • Flaws frequently occur holes and human-made errors leading to security. Usually, such errors stem from the lack of sufficient experience and professional competence. Intelligent attackers are ready to exploit these novice users’ weaknesses. On the other hand, there may be malicious objectives behind these errors; for example, insiders (e.g., disgruntled employees) can be considered as potential attackers. They have access to sufficient resources and information so that they can easily conduct a successful attack. Although with education and caution, to some extent, humanmade errors may be reduced, they are unavoidable due to their human nature. In this paper, we propose a state-based stochastic model for the purpose of quantitative evaluation. As stated in [1], the, state-space methods have been explored in the context of mathematical models that are based on transition behavior and specify probabilistic assumptions about time durations. For doing so, we present a statetransition model and transform this abstract model into a state-based stochastic model by assigning time distributions to its transitions. Note that the attacker’s or the defender’s activities are implicitly considered in the transitions of the model. When the model is parameterized, quantitative metrics will be assessed with respect to the nature of the model. An appropriate model must predict when, where and how malicious attacks may take place. The proposed state based stochastic model has the Markov property and general probability distributions are assigned to its transitions. The general organization of our paper is as follows: In Sect. 2, the related works on model of malicious attacks process. In Sect. 3, the modeling of problems in the Markov chains. In Sect. 4, the result and discussion. Finally, conclusions of the paper and further research are given in Sect. 5.

New Improvement of Malware-Attack Scenarios Modeling

203

2 Model of Malicious Attacks Process There are several models of attacks [6–16]. They are generally specific to the runtime environment, and therefore require a precise and detailed knowledge of the architecture, network vulnerabilities and topology and considered system. Moreover, these models are based primarily on known vulnerabilities and ignore the attacks that may exploit still unknown vulnerabilities, which would constitute a serious limitation, since the robustness of IDS depends also on unknown new attacks and vulnerabilities. In this paper, we have adopted the model of the attack process of Saber and al. [17], which is based on a preliminary analysis of malware attacks like the most prevalent viruses and worms. This choice is justified by the fact that this model is the result of the analysis of more than seventy malware from the Mitre’s Common Malware Enumeration list (CME) [18], which are representative of the more widespread attacks and dangerous. Indeed, given that worms are autonomous, they must include all the steps in an attack process. In addition, viruses such as worms can be seen as a class of automated attacks developed by skilled attackers, and this can help to understand how interactive attacks can be conducted. This model is described in Figure (Fig. 1). It distinguishes the following steps: • Recognition (Reconnaissance): it is logical for an attacker to find the necessary information on potential victims before targeting them with the most appropriate attack tools (exploit codes, toolkits). • Gain access: to achieve their objectives, attackers usually need access to victim’s resources; the level of access required will obviously depend on the attack. However, some types of attacks such as denial of service attacks do not need access to the victim machine. • Privilege Escalation: Access originally obtained by the attacker is sometimes insufficient to achieve the attack, in which case, the attacker tries to increase its privileges to have more power (for example, switch from user mode to administrator mode to access the system resources). • Browsing Victim: after having acquired sufficient privileges, the attacker usually tries to explore the machine or the target network (e.g., searching files and directories), to search for a particular account (as a guest account or an anonymous ftp account), to identify the hardware components, to identify installed programs or to search for trusted hosts (typically, those with certificates installed on the victim machine). • Principal Actions: as shown in Figure (Fig. 1), this step may take different forms, for example, an attacker can execute a denial of service attacks, install malicious code, compromising the integrity of data or run a program. • Hiding Traces: the most experienced attackers generally use this last step to erase their tracks, thereby making detection more difficult.

204

N. Rahmoun et al.

Fig. 1. The attack process model of Malware type.

To scenarios attack a simplified model called the state machine, illustrated in Figure (Fig. 2). The steps taken by the malware attacks can be classified into only eight primitives and each identified by a symbol, as indicated below: • • • • • • • •

R: Recognition (Reconnaissance) VB: Exploration of the machine/or the network of the victim (Victim Browsing) EP: Program execution (Execute Program) GA: Gain Access IMC: Implementation of malicious code (Implant Malicious Code) CDI: Compromise of integrity (Compromise Data Integrity) DoS: Denial of Service HT: Erasing traces (Hide Traces)

It is important to note that this iterative approach for generating attack scenarios has overcome the problem of combinatorial explosion, inherent problem to convention a approaches to generating attack scenarios. The problem we want to solve is to find efficient algorithms that can generate valid meaningful attack scenarios. It would be easier to incorporate the state machine model

New Improvement of Malware-Attack Scenarios Modeling

205

in an IDS, to test and evaluate it. The modeling of our problem is presented in the next section. The graph in Figure (Fig. 2) allows the generation of attack scenarios at an abstract level.

Fig. 2. State machine representing the attack process.

3 Malware-Attack Scenarios Generation 3.1

Presentation of the Stochastic Process Model

We propose a state-based model consisting of a set of states and a set of transitions between these states. In the context of the proposed model, a state displays the current position of interactions between the attacker and the cible and a transition displays an action of the attacker. Now, we explain the process of building the state transition model, step by step. As we know, a stated-based model consists of a set of states and a set of transitions connecting the states. Before representing the state-based model, it is necessary to define the role of states and transitions in the underlying context. The transition occurrence transfers the model from a state to another state. The underlying context of interest is on security analysis. Using this model, we want to describe the state of the attacker. These states characterize to what extent the attacker has proceeded in the attacking process, or equivalently, to what extent the system has been intruded. The transitions show the malicious activities of the attacker or the defensive activities of the system. The attacker’s transitions display the progression of the attacker towards the security target. The system transitions show the

206

N. Rahmoun et al.

remedial reactions in order to thwart the effects of the malicious activities of the attacker. Based on the definition of the state and the transition, the graphical representation for a single attack path can be characterized. As mentioned earlier, it is supposed that each attack path consists of some elementary attack phases. Thus, a generic attack path can be visualized as in Fig. 1.

Fig. 3. An attack path consisting of n elementary attack phases.

The attack path of Figure (Fig. 3) has n phases. Each state of the attack path specifies a privilege level currently owned by the attacker and a transition between two sequential states specifies that there are some ways for the attacker to obtain new privileges. Clearly, there is a casual dependency between the elementary phases; that is, all steps in an attack path are to be performed sequentially. Now, we take the role of the system into account. Hence, it is desirable to model the interactions between the attacker and the defender. In each attack phase, based on our assumptions, the attacker and the defender interact with each other. If the existing vulnerabilities are detected and removed by the defender before the attacker can detect and exploit them, the state transition model will return to the secure state; otherwise the attacker will move from the current attack phase to the next attack phase. This process is repeated until one of the following two specific security situations is reached: security failure or secure state. In fact, an attack path is said to be successful, if and only if the attacker successfully exploits all of the elementary attack phases of this attack path. Imagine, for gaining a better understanding of the analysis, the attacker and the defender as two competing agents. They compete for different goals: the attacker wants to compromise the security of the system. So far, a state-based model has been represented for describing the interactions between the attacker and the defender in a single attack path. Now, we consider a general attack process consisting of a number of attack paths. We propose, in a similar way to that of a single attack path, the state-transition model of the attack process as shown in Figure (Fig. 4). This model is simply the extended version of the model of Figure (Fig. 3), where instead of an attack path, some attack paths are performed concurrently and independently.

Fig. 4. An attack process consisting of a number of attack paths

New Improvement of Malware-Attack Scenarios Modeling

3.2

207

Markov Chain

A stochastic process {X(t)}t 2 T is a function of time whose value at every moment depends on the outcome of a random experiment. At each time t 2 T, X(t) is a random variable. If we consider a discrete-time, rating is then {Xn} n 2 N a stochastic process in discrete time. If we finally assume that the random variables Xn can take only discrete set of values, we refer to process discrete time and discrete state space. {Xn}n 2 N is a Markov chain in discrete time if and only if: PðXn ¼ j j Xn1 ¼ in1 ; Xn2 ¼ in2 ; . . .; X0 ¼ i0 Þ ¼ PðXn ¼ j j Xn1 ¼ in1 Þ

ð1Þ

The probability that the chain is in a certain state to the nth of the process depends only on the state of the process in the previous step (the n − 1th) and not states in which it was to earlier stages. Is defined as a homogeneous Markov chain when this probability does not depend on n. We can then define the probability of transition from state i to state j denoted pij: pij ¼ PðXn ¼ jjXn1 ¼ iÞ8n 2 N

ð2Þ

By introducing the set of possible states denoted E, we have: X

pij ¼ 1

ð3Þ

j2E

h i We then define the transition matrix p ¼ pij i; j 2 E 0

p11 B p21 B . P¼B B .. @ .. .

3.3

p12 p22 p32




p23 .. .






..

1 C C C C A

ð4Þ

.

Malware Attack Process Application

Implementing modeling attacks (Fig. 2) by Markov chains by implementing the transition matrix for our model is as follows:

208

N. Rahmoun et al.

0

p12 p22 0 0 0 0 0 0

p11 B 0 B B 0 B B 0 P¼B B 0 B B 0 B @ 0 0

p13 p23 p33 p43 p53 0 0 0

0 p24 0 p44 p54 p64 0 p84

0 p25 0 p45 p55 p65 p75 0

0 p26 0 p46 0 p66 0 p86

0 0 0 0 p57 p67 p77 p87

1 0 p28 C C 0 C C p48 C C p58 C C p68 C C 0 A p88

ð5Þ

The analysis of the transient of a Markov chain is to determine the vector p(n) the probability of being in state j at step n: pðnÞ ¼ ½p1 ðnÞ p2 ðnÞ . . . pCardðEÞ ðnÞ

ð6Þ

For our model card (E) = 8. This vector of probabilities depends: – Transition matrix P – The vector of initial probabilities p(0) To investigate this probability vector p(n) can make the following remarks: pj ðnÞ ¼ P½Xn ¼ j

ð7Þ

Or P½Xn ¼ j ¼

X i2E

P½Xn ¼ j j Xn1 ¼ i:P½Xn1 ¼ i

ð8Þ

Which can be written: pj ðnÞ ¼

X i2E

pij :pi ðn  1Þ

ð9Þ

which can be written in matrix form: pðnÞ ¼ pðn  1Þ
P

ð10Þ

When using n times this expression, we get: pðnÞ ¼ pð0Þ
Pn

ð11Þ

One can also introduce the transition probability from state i to state j in m steps, ð mÞ denoted by pij : ðmÞ

pij ¼ P½Xn þ m ¼ j j Xn ¼ i8n 2 N

ð12Þ

New Improvement of Malware-Attack Scenarios Modeling

3.4

209

Equiprobable Case Study of the Malware Attack Process Model

We studied the following special cases. We consider that the system is equally likely. The transition matrix is: 0

1=3 B 0 B B 0 B B 0 P¼B B 0 B B 0 B @ 0 0

1=3 1=6 0 0 0 0 0 0

1=3 1=6 1 1=5 1=5 0 0 0

0 1=6 0 1=5 1=5 1=5 0 1=4

0 1=6 0 1=5 1=5 1=5 1=2 0

0 1=6 0 1=5 0 1=5 0 1=4

0 0 0 0 1=5 1=5 1=2 1=4

1 0 1=6 C C 0 C C 1=5 C C 1=5 C C 1=5 C C 0 A 1=4

ð13Þ

The vector of initial probabilities p(0):   111 pð 0Þ ¼ 00000 333

ð14Þ

4 Result and Discussion Matlab were used to calculate the probability. We have therefore implemented the matrix transition with the transitions’ probabilities from state i to state j in m steps. The application’s results of Algorithm are summed up in Table 1, which represents the percentage of probability to produce an attack from an initial state to the final states. Table 1. Probability of producing an attack from an initial state to the final states. Initial state Final state 3 (DOS) 1 (R) 38% 2 (GA) 0% 3 (DOS) 0% (100% if 3 is final state)

5 (EP) 54% 0% 100%

7 (HT) 8% 0% 0%

8 (CDI) 0% 0% 0%

In Table 1 we notice that the probability of producing an attack of the initial state 1 (R) is successively 38%, 54% and 8% to the final states 3 (DOS), 5 (EP) and 7 (HT). On the other hand, it is 100% from the initial state 3 (DOS) to the two final states 3 (DOS) and 5 (EP). In the light of the results obtained, we conclude that each time we achieve an end state the probability of building an attack increases. The probabilities of the final nodes (3, 5, 7 and 8) are close to 1.

210

N. Rahmoun et al.

5 Conclusion In this paper, we have proposed a state-based stochastic model of Malware type, which allows the evaluation of network security based on Markov chain. Modeling was done in two phases: the first phase is the presentation of a state-transition model and the second phase, which represents the parameterization of the model. The solution we have proposed shows that we will be able to identify an attack scenario while the attack process is not yet complete which may help IDS to improve the detection rate of malware attacks. In our next work, we will try to apply this modeling of attack scenarios on web-based attacks, which represent a worldwide major problem.

References 1. Kabir, S.: An overview of fault tree analysis and its application in model based dependability analysis. Expert Syst. Appl. 77, 114–135 (2017) 2. Madan, B., Goseva-Popstojanova, K., Vaidyanathan, K., Trivedi, K.S.: A method for modeling and quantifying the security attributes of intrusion tolerant systems. Perform. Eval. J. 56(1–4), 167–186 (2004) 3. Collier, Z.A., Panwar, M., Ganin, A.A., et al.: Security metrics in industrial control systems. In: Cyber-Security of SCADA and Other Industrial Control Systems, pp. 167–185. Springer, Cham (2016) 4. Ramos, A., Lazar, M., Holanda, F., Raimir, et al.: Model-based quantitative network security metrics: a survey. IEEE Commun. Surv. Tutor. 19(4), 2704–2734 (2017) 5. Rahmoun, N., Saber, M., Ettifouri, E., Zeaaraoui, A., Bouchentouf, T.: A new approach to detect WEB attacks senario in intrusion detection system, vol. 381, pp. 569–573. Springer (2016). https://doi.org/10.1007/978-3-319-30298-0_59, ISBN 978-3-319-30296-6 6. Geva, M., Herzberg, A., Gev, Y.: Bandwidth distributed denial of service: attacks and defenses. IEEE Secur. Priv. 12, 54–61 (2014) 7. Ben-Asher, N., Gonzalez, C.: Effects of cyber security knowledge on attack detection. Comput. Hum. Behav. 48, 51–61 (2015) 8. Li, W., Meng, W., Kwok, L.-F., et al.: Enhancing collaborative intrusion detection networks against insider attacks using supervised intrusion sensitivity-based trust management model. J. Netw. Comput. Appl. 77, 135–145 (2017) 9. Karapistoli, E., Economides, A.A.: Modeling the Internet of Things under attack: a Gnetwork approach. IEEE Internet Things J. 4(6), 1964–1977 (2017) 10. Mohammed, S., Toumi, B., Abdelhamid, B.: Generation of Attack Scenarios by Modeling Algorithms for Evaluating IDS, pp. 1–5. IEEE (2011). https://doi.org/10.1109/ICMCS.2011. 5945730, ISBN 978-1-61284-732-0 11. Saber, M., El Farissi, I., Chadli, S., Emharraf, M., Belkasmi, M.G.: Performance analysis of an intrusion detection systems based of artificial neural network. In: Advances in Intelligent Systems and Computing, vol. 520, pp. 511–521. Springer (2017). https://doi.org/10.1007/ 978-3-319-46568-5_52, ISBN 978-3-319-46567-8 12. Sheyner, O., Haines, J., Jha, S., Lippmann, R., Wing, J.M.: Automated generation and analysis of attack graphs. In: Proceeding of 2002 IEEE Symposium on Security and Privacy, Oakland, California, USA, pp. 273–284 (2002)

New Improvement of Malware-Attack Scenarios Modeling

211

13. Farnaaz, N., Jabbar, M.A.: Random forest modeling for network intrusion detection system. Procedia Comput. Sci. 89, 213–217 (2016) 14. Hebert, C., Gomez, L.: Penetration test attack tree generator. U.S. Patent No 9,894,090, 13 févr 2018 (2018) 15. Drago, A., Marrone, S., Mazzocca, N., et al.: A model-driven approach for vulnerability evaluation of modern physical protection systems. Soft. Syst. Model. 18(1), 523–556 (2019) 16. Rahmoun, N., Chadli, S., Ettifouri, E., Saber, M., Bouchentouf, T.: A Stochastic Model of Maliciel Process Attack for the Evaluation of Network Security, pp. 269–273 (2015). IEEE. https://doi.org/10.1109/ICoCS.2014.7060955, ISBN 978-1-4799-4647-1 17. Kaaniche, M., Deswarte, Y., Alata, E., Dacier, M., Nicomette, V.: Empirical analysis and statistical modeling of attack processes based on honeypots. In: Proceeding of Workshop on Empirical Evaluation of Dependability and Security (WEEDS), DSN 2006, Philadelphia, USA, pp. 119–124 (2006) 18. El Farissi, I., Saber, M., Chadli, S., Emharraf, M., Belkasmi, M.G.: The Analysis Performance of an Intrusion Detection Systems Based on Neural Network, pp. 145–151 (2017). https://doi.org/10.1109/CIST.2016.7805032, ISBN 978-1-5090-0751-6 19. Mitres Common Malware Enumeration list. http://cme.mitre.org/

IoT Security Management: Model and Design Issues Ghizlane Benzekri1(&), Omar Moussaoui1, and Ali El Moussati2 1

MATSI Lab, EST, Oujda, Morocco [email protected], [email protected] 2 Department of EIT, ENSA, Oujda, Morocco [email protected]

Abstract. The Internet of Things (IoT) is now destroying the barriers between the real and digital worlds. However, one of the huge problems that can slow down the development of this global wave, or even stop it, is security. So, it’s considered as a crucial problematic for IoT from the fact that the minimal capacity “things” being used, the physical accessibility to sensors, actuators, and objects, and the openness of the systems, including the fact that most devices will communicate wirelessly. However, existing security solutions and techniques are not adapted to these developments, which impact security management efficiency. Therefore, there is a need to automate certain security management tasks mainly the detection of security attacks, the deploying reaction and assisting the security administrators for taking the right decisions. Based on several works in that paradigm, we provide an adequate and appropriate solution to IoT security management problems by proposing a dynamic security management model which will aim to simplify the security management process. Keywords: Internet of Things security management


Security
Security management
Dynamic

1 Introduction Internet of things (IoT) is a collection of many interconnected objects, services, humans, and devices that can communicate, share data, and information to achieve a common goal in different areas and application. Indeed, many researchers consider IoT as one of the main technological revolutions of this century [1] and have moved from being a futuristic vision to an increasing market and research reality. One of the major challenges that must be overcome, in order to push the IoT in the real world, is security. IoT architectures are supposed to deal with an estimated population of billions of objects, which will interact with each other and with other entities, such as human beings or virtual entities. And all these interactions must be secured somehow, protecting the information and number of incidents that will affect the entire IoT. Unfortunately, the OWASP Internet of Things Project has listed the most common IoT attacks and vulnerabilities [2]. According to this project, the risk arises because of © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 212–219, 2020. https://doi.org/10.1007/978-3-030-53187-4_25

IoT Security Management: Model and Design Issues

213

the lack of adoption of well-known security techniques, such as encryption, authentication, access control and role-based access control. A reason for this lack of adoption is that existing security techniques, tools, and products may not be easily applied to IoT devices and systems. Based on that statement, it appears crucial to address these security management problems, and develop model or platform, which can simplify the security management process. In other words, it is necessary to automate certain security management tasks (monitoring, reporting, auditing, etc.), to automate the decision and reaction process, and to assist the security administrators in other tasks for taking the right decisions (deep forensics and investigations, security analytics, etc.). In this context, we focus our work on the security management problems by proposing a dynamic security management model which aims to simplify the security management process. This model will allow the IoT system to better leverage intelligence in its resources, also to facilitate the decision against detected intrusions and according to their magnitudes and their intentions to put the necessary reaction in the right place. Thus, this model will make it possible to simplify and automate the decision phase by enriching it with knowledge in order to protect the IoT resources. The next section of this paper provides an overview of IoT security challenges and summary about security management. The Sect. 3 is a brief state-of-the art around IoT security management works. Then, in Sect. 4, we present discussion and therefore the proposed security management model. Finally, we conclude by some perspectives.

2 Background This section gives an overview of the basic concepts necessary to understand the proposed model. 2.1

IoT Security Challenges

Although academic research on the topic of security in the IoT, is still in its infancy, there is a substantial body of work that analyses the existing IoT security challenges [10, 11]. Given below are the security principles that should be enforced to achieve a secure communication framework for the people, software, processes, and things: Confidentiality: It is very important to ensure that the data is secure and only available to authorized users. For the IoT based devices, it ensures that the sensor nodes of the sensor networks don’t reveal their data to the neighboring nodes; similarly, the tags don’t transmit their data to an unauthorized reader. Integrity: The IoT is based on exchanging data between many different devices, which is why it is very important to ensure the accuracy of the data; that it is coming from the right sender as well as to ensure that the data is not tampered during the process of transmission due to intended or unintended interference. Availability: One of the major goals of IoT security is to make data available to its users, whenever needed. Data Availability ensures the immediate access of authorized

214

G. Benzekri et al.

party to their information resources not only in the normal conditions but also in disastrous conditions. Authentication: Each object in the IoT must be able to clearly identify and authenticate other objects. However, this process can be very challenging because of the nature of the IoT. Heterogeneity: The IoT connects different entities with different capabilities, complexity, and different vendors. The devices even have different dates and release versions, use different technical interfaces and bitrates, and are designed for an altogether different functions, therefore protocols must be designed to work in all different devices as well as in different situations. Policy Enforcement: Policy enforcement refers to the mechanisms used to force the application of a set of defined actions in a system. More in details, policies are operating rules which need to be enforced for the purpose of maintaining order, security, and consistency on data. Access Control: Authentication and access control technologies are known as the main elements to address the security issues in the Internet of Things. Actually, any effective access control system should satisfy the main security properties of the CIA triad: Confidentiality, Integrity and Availability. Note that one should not confuse AC with identification and authentication notions. Figure 1 shows the access control process.

Fig. 1. Access control process

Trust in IoT: Trust is a complex notion about which no definitive consensus exists in the scientific literature, although its importance is widely recognized. A main problem with many approaches towards trust definition is that they do not lend themselves to the establishment of metrics and evaluation methodologies.

IoT Security Management: Model and Design Issues

215

Automation: The automation of the security tasks is of paramount importance in the IoT. Indeed, as the amount of resources and their level of associations increase, which are matched with the potential augmentation in usage and the fact that these resources are exposed on a network, it becomes less and less viable for a human based management, at least at the single device level. Interoperability: The interoperability issue in the security of IoT can be separated in three main domains. The first addresses the semantic of communication, the second checks the grammar of communication and finally the third regards the operational connection. 2.2

Security Management

The main objective of security management is to implement the appropriate controls and activities, needs to have for ensuring protection of its assets against the risks of loss, misuse, disclosure or damage, and eliminating or minimizing the impact that various security related threats and vulnerabilities might have on it. There are a large number of services, as we can see in the Fig. 2, the security management imposes security requirements definition (Intrusion detection, security assessment, risk analysis…), to taking them for forming a comprehensive security strategy.

Fig. 2. The security management lifecycle.

However, this strategy is based on Security Risk Management (SRM), which aim is to understand what assets should be protected, from which risks and how these risks could be allayed while covering discovered vulnerabilities of the system [3]. SRM plays a leading role in developing appropriate solutions based on the situation and existing security countermeasures. Typically, it provides a set of rules to lower the risk level or totally prevent possible attacks on the system along with the hints for successful and qualitative system monitoring. There are many different methodologies

216

G. Benzekri et al.

which apply SRM in various domains, each one has his own analysis concepts and principles are used during an analytical procedure in security risk management: CORAS [4], EBIOS [5] AURUM [6], MEHARI [7], CRAMM [8]. After having defined vulnerabilities, risks and security countermeasures, we can define security policy. So, the Common Criteria defines an organizational security policy as: a set of security rules, procedures, or guidelines imposed (or presumed to be imposed) now and/or in the future by an actual or hypothetical organization in the operational environment [9]. Security management will enable to have desirable qualitative characteristics of offered services (i.e. services availability, data confidentiality and integrity preservation etc.). As we can see, IoT paradigm has still to face hard challenges related to the application of security. The following section will focus on works around IoT security challenges aspects and IoT security management.

3 Related Work The scientific community has started several interesting researcher initiatives to address both security management and security challenges cited in Sect. 2, below is a summary of the most recent and relevant ones: A. Ouaddah et al. [12] present the OM-AM authorization reference model which describe and analyze authorization process in IoT based on main IoT security requirements and how each model meets each requirement. A. Outchakoucht et al. [14] focus on access control in the IoT context by proposing a dynamic and fully distributed security policy. This framework is based on the concept of the blockchain to ensure the distributed aspect strongly recommended in the IoT. H. Ouechati et al. [15] propose an access control middleware for the Internet of Things. The latter is an extension of the ABAC model in order to take into account the subject behavior and the trust value in the decision-making process. Therefore, this work introduces a dynamic adaptation process of access control rules based on the risk value, the policies and rule sets. R. Neisse et al. [16] describe an efficient solution to enforcement security policy rules that addresses IoT security challenges. This enforcement solution is based on a Modelbased Security Toolkit named SecKit, and its integration with the Message Queuing Telemetry Transport (MQTT) protocol layer. S. Sicari et al. [17] introduce and discus a flexible security and data quality enforcement framework, coherently integrated within a distributed IoT middleware platform. This framework supports security and data quality enforcement policies, reusable across different domains and able to detect violation attempts.

4 Discussion After reviewing existing work focusing on IoT security challenges and security management, we can note that most of them are limited to access control and policy enforcement of security. Security management in IoT stills a relatively treated subject.

IoT Security Management: Model and Design Issues

217

In fact, existing solutions do not meet all the identified IoT security challenges cited in background. Therefore, a new solution is needed, that will manage security in a way more suited to the needs of existing and future networks, characterized by their openness, complexity and dynamism and will meet to all IoT security challenges. Although, it is too hard to depend on manual approaches that require deep involvement of security managers to deliver the aimed security level. Thus, any security management approach must be a comprehensive approach which intended to implement a security platform in objective supporting solutions and products working independently. This approach must focus on automating the security management tasks. So, even if it exists some model concerning security management task cited in related work, this models are limited to define access control rules such as ABAC model or to describe how we can enforce this rules for the purpose of maintaining data order, security, and consistency and finally to describe risk management. In this context, we propose a dynamic security management model presented in Fig. 3, which aims to simplify the security management process and include all the security management tasks. This model allows the IoT system to better leverage intelligence in its resources, also to facilitate the decision against detected intrusions and according to their magnitudes and their intentions to put the necessary reaction in the right place.

Fig. 3. Dynamic security management model

So, it allows, on receipt of a critical alert, to take a decision on the reaction to be undertaken. This can be a change in the device configuration in the network, a firewall or a machine. As the action can only be to notify the administrator by registering the event and the decision in databases. The choice depends on the analysis result and correlation activity that must be carrying out by taking into account the knowledge available on the Data Base, which represents the set of data and knowledge that must have an administrator on the resources (resources, their level of criticality, the users and their profiles, the critical data location, the network topology and security policy rules…). The objective of integration of this BD is to be able to gather all the data and knowledge that will be

218

G. Benzekri et al.

used in the process of attacks response and contributes to the automation of the attack’s remediation cycle. Thus, this allows the proposed model to have the following characteristics: • It is reactive and so can act on network equipment’s configurations in order to block an attack detected by IDS and which is qualified critical by the correlation and decision module. • Before acting, correlation and decision module analyze the attack based on several criteria, such as attack severity, resource criticality, source and destination address, protocols, ports, etc. Therefore, it consults the Database for taking the most appropriate actions and for applying it on the right place (switch port, host, etc.) at the responses module level. Also, an important part of developing this model concerns tools choice and reasoning techniques to be used (having shown their ability in intelligent systems), for the decision making. The BD also deserves to be defined for choosing the best way for representation data and knowledge on the overseen infrastructure.

5 Conclusion and Perspectives Today, IoT is surrounding us and its aptitudes of sensing, actuation, communication, and control become ever more sophisticated and ubiquitous; however, these advantageous features are also examples of security threats that are already nowadays slowing down the growth and expansion of the Internet of Things when not fulfilled properly. In this paper, we focused on IoT security management. We proposed a dynamic security management model which aims to simplify the security management process. However, this contribution still has perspectives on which we intend to work in our future paper. Indeed, we must develop dynamic policy framework, which takes in the one hand into consideration the context in which the smart devices are, but also which can be improved over time, this improvement obviously does not, and cannot, be managed by a human being given the enormous and heterogeneous amount of data that the IoT generates, and one the other hand, to respond too security management requirements. We therefore think in this paper to use the power of artificial intelligence algorithms, especially those of machine learning, to ensure this task. As a final point, this model needs also a thorough case study as well as an implementation as a concrete proof of concept. Acknowledgment. This work is supported by the Mohammed First University under the PARA1 Program.

References 1. Lopez, J., Rios, R., Bao, F., Wang, G.: Evolving privacy: from sensors to the internet of things. Future Gener. Comput. Syst. 75, 46–57 (2017)

IoT Security Management: Model and Design Issues

219

2. Alkhalila, A., Ramadan, R.A.: IoT data provenance implementation challenges. In: 3rd International Workshop on Tasks on High Performance Computing (THPC), p. 3 (2017) 3. Matulevicis, R.: Fundamentals of secure system modelling (2017). http://www.springer.com/ gp/book/9783319617169 4. Dimitrakos, T., Bicarregui, J., Stolen, K.: CORAS - a framework for risk analysis of security critical systems (2002). https://www.ercim.eu/publication/Ercim_News/enw49/dimitrakos. html 5. EBIOS method for analysis, evaluation and action on risks (2010). https://www.ssi.gouv.fr/ guide/ebios-2010-expression-des-besoins-et-identification-des-objectifs-de-securite 6. Ekelhart, A., Fenz, S., Neubauer, T.: AURUM, a framework for information security risk management. In: International Conference on System Sciences, pp. 1530–1605 (2009) 7. Mihailescu, V.: Risk analysis and risk management using MEHARI. J. Appl. Bus. Inf. Syst. 3(4), 143–162 (2012) 8. SANS Institute. CRAMM: CRAMM, a qualitative risk analysis and management tool. https://www.sans.org/reading-room/whitepapers/auditing/qualitative-risk-analysismanagement-tool-cramm-83 9. Ouaddah, A., Abou Elkalam, A., Ait Ouahman, A.: FairAccess: a new blockchain-based access control framework for the Internet of Things. Secur. Commun. Netw. 9(18), 1–22 (2017) 10. Roman, R., Rios, R., Zhou, J., Lopez, J.: On the features and challenges of security and privacy in distributed Internet of Things. Comput. Netw. 57(10), 2266–2279 (2013) 11. Sicari, S., Rizzardi, A., Grieco, L.A., Coen-Porisini, A.: Security, privacy and trust in Internet of Things: the road ahead. Comput. Netw. 76, 146–164 (2015) 12. Ouaddah, A., Mousannif, H., Abou Elkalam, A., Ait Ouahman, A.: Access control in the Internet of Things: big challenges and new opportunities. Comput. Netw. 112, 237–262 (2017) 13. Novo, O.: Blockchain meets IoT: an architecture for scalable access management in IoT. J. Internet Things Cl. Files 14(8), 1–12 (2018) 14. Outchakoucht, A., Es-samaali, H., Leroy, J.P.: Dynamic access control policy based on blockchain and machine learning for the internet of things. Int. J. Adv. Comput. Sci. Appl. 8 (7), 1–8 (2017) 15. Ouechtati, H., Azzouna, N.B., Bensaid, L.: Towards a self-adaptive access control middleware for the internet of things. In: The 32nd International Conference on Information Networking, pp. 545–550 (2018) 16. Neisse, R., Steri, G., Baldini, G.: Enforcement of security policy rules for the internet of things. In: Third International Workshop on Internet of Things (IoT) Communications and Technologies, pp. 165–172 (2014) 17. Sicari, S., Rizzardi, A., Miorandi, D., Cappiello, C., Coen-Porisini, A.: Security policy enforcement for networked smart objects. Comput. Netw. 108, 133–147 (2016)

Energy and MultiSource Systems Management

Comparison Between Constant and Variable Switching Frequency Strategies Based Direct Torque Control of Asynchronous Motor Soukaina El Daoudi(&), Loubna Lazrak, Chirine Benzazah, and Mustapha Ait Lafkih Laboratory of Automatic, Energy Conversion and Microelectronics (LACEM), Faculty of Science and Technology, Sultan Moulay Slimane University, Beni Mellal, Morocco [email protected], [email protected], [email protected], [email protected]

Abstract. This paper presents a comparative study between a classical and an improved Direct Torque Control (DTC) strategy for induction motor powered by two level three phase voltage source inverter. The main objective is to make a comparative analysis of two control techniques with a variable and a constant switching frequency to which this latter shows a great improvement of the system performance by reducing stator and rotor flux ripples and improving the current sinusoidal form by optimizing the total harmonic distortion (THD). The hysteresis regulators and voltage vectors selection table of the classical DTC, which directly control the inverter states by reducing the torque and flux errors within prefixed band limits, are replaced by proportional–integral (PI) controllers connected to a sine pulse width modulation (SPWM). This latter is used to generate the quadrature and direct voltages. A fair comparison between the two control strategies has been made and simulated under MATLAB/SIMULINK software. Keywords: Classical DTC
Improved DTC
PI controllers
SPWM
THD
Two level three phase inverter

1 Introduction The direct torque control (DTC) was firstly proposed by Takahashi and Noguchi in 1986. It is known by its simple decoupled control scheme of stator flux and torque. It bases on selecting the appropriate voltage vectors via hysteresis regulators and switching table for the associated inverter. The main advantages are it does not require any transformation or current regulators. It minimizes the use of motor parameters [1], which results in less sensitive to parameter variations. However, due to its structure, the main problem of this method is the high level of torque and flux ripples and the variable switching frequency of the inverter [2, 3]. The constant switching frequency DTC strategy which bases on the sine pulse width modulation has been presented to face these kinds of drawbacks. The SPWM-DTC is a technique that uses two (PI) controllers instead of hysteresis regulators to generate direct and quadrature voltage © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 223–231, 2020. https://doi.org/10.1007/978-3-030-53187-4_26

224

S. El Daoudi et al.

components in which the inverter is indirectly controlled using the SPWM modulation [4]. This paper aims to make a comparative study between two strategies based Direct Torque Control. The first technique is the classical DTC while the second one is an improved DTC to which the control is made by PI regulators. The proper voltage vector is selected using sinusoidal modulation technique applied to two level inverter. The proposed strategy maintains a constant switching frequency and minimizes flux and torque ripples. The results will be examined under simulation using MATLAB/SIMULINK software.

2 Mathematical Model of Asynchronous Motor Before establishing mathematical model of the asynchronous motor, assumptions are often made to simplify the system’s modeling. From this, the voltage, current and flux linkage equations are got in the (d − q) coordinate as follows: 8 > Vds ¼ Rs Ids þ dud ds  xs uqs > > > > du < Vqs ¼ Rs Iqs þ d qs þ xs uds ; > > 0 ¼ Rr Idr þ dud dr  xg uqr > > > : du 0 ¼ Rr Iqr þ d qr þ xg udr

8 1 Ids ¼ rL uds  rLLsmLr udr > > s > > > < Iqs ¼ 1 u  Lm u rLs Lr qr rLs qs L 1 m > I ¼  u þ u > dr rLs Lr ds > rLr dr > > : Lm u þ 1 u Iqr ¼  rL qs rLr qr s Lr

8 u ¼ Ls Ids þ Lm Idr > > < uds ¼ L I þ L I s qs m qr qs u ¼ L I þ L Ids > r dr m > : dr uqr ¼ Lr Iqr þ Lm Iqs

ð1Þ

ð2Þ

With: Vds ; Vqs : Stator voltage components. Ids ; Iqs : Stator current components. Idr ; Iqr : Rotor current components. uds ; uqs : Stator flux components. udr ; uqr : Rotor flux components. Rs ; Rr : Stator and rotor resistance, respectively. Ls ; Lr : Stator and rotor inductance, respectively. m : Leakage coefficient. Lm : Mutual inductance and r ¼ 1  LLL s r The mechanical as well as the electromagnetic torque equations of the motor are given as (4) and (5): dX ¼ Ce  Cr  fX dt

ð3Þ

 3  Ce ¼ P uds Iqs  uqs Ids 4

ð4Þ

J

Comparison Between Constant and Variable Switching Frequency

225

With: X: The mechanical speed. Cr : The load torque. f; J: Coefficient of friction and moment of inertia. Ce : The electromagnetic torque. P : The number of pole pairs.

3 Classical Direct Torque Control (DTC) Using Variable Switching Frequency The Direct Torque Control (DTC), based on the stator flux orientation, is one of the techniques proposed by Takahashi in the middle of 1980s [5]. It bases on the direct selection of voltage vector according to the instantaneous errors of the stator flux and the electromagnetic torque. The DTC uses separated hysteresis regulators to ensure a decoupled control of flux and torque without requiring a complex field orientation or current regulation loop [6]. The outputs of the hysteresis comparators choose the appropriate voltage vector through a look-up switching table which results a variable frequency operation. The figure below shows off the DTC based on lookup switching table (Fig. 1).

Fig. 1. General block diagram of the classical DTC offering a variable switching frequency

3.1

Stator Flux Control

Basing on the asynchronous motor model in the stationary frame, the stator flux equation can be expressed as follows: TZ s

us ðtÞ ¼ ðVs  Rs Is Þdt þ us0 0

Where us0 is the stator flux vector at the instant t = 0 s.

ð5Þ

226

S. El Daoudi et al.

To simplify, we can neglect the stator resistance voltage drop Rs Is compared to Vs which means that when the rotation speed increases during a time Te , the stator flux vector’s extremity moves in a direction given by the voltage vector and making a circular trajectory [7] (Fig. 2).

Fig. 2. Stator flux vector trajectory in the (a – b) frame

A two-level hysteresis comparator is used for flux regulation. The choice of the bandwidth depends on the inverter’s switching frequency. 3.2

Torque Control

The electromagnetic torque equation can be expressed in terms of stator and rotor flux vectors as follows: 3 Lm jCe j ¼ P ju jju jsinðdÞ 4 rLs Lr s s

ð6Þ

Where d is the angle between the stator and rotor flux vectors. It is clear that the electromagnetic torque is controlled by the stator and rotor flux amplitudes. If those quantities are maintaining constant, the torque can be controlled by adjusting the load angle. The regulation can be realized using three-level hysteresis comparator; it allows controlling the motor in both rotation senses [8]. This control scheme is used in the industry in low and medium power applications. The main advantages of DTC are summarized in its fast dynamic, the absence of coordinate transformations and current control loops. In the other hand, the main disadvantages of DTC are the variable switching frequency, high torque/flux ripples and high switching losses.

4 Improved Direct Torque Control (DTC) Using Constant Switching Frequency In this DTC strategy, the lookup switching table and the hysteresis comparators are replaced by the sine pulse width modulation (SPWM) and PI regulators in order to select the voltage vectors. This control scheme keeps a number of classical DTC

Comparison Between Constant and Variable Switching Frequency

227

features and can preserve more advantages such as a constant switching frequency that reduces the high torque and flux ripples which is the main drawback of the classical DTC and a minimized total harmonic distortion (THD). The following figure presents the improved DTC based on a sine pulse width modulation (Fig. 3).

Fig. 3. General block diagram of the improving SPWM-DTC offering a constant switching frequency

4.1

Stator Flux Control

Considering the direct stator voltage expression, the control will be done using Eqs. (1–2). By neglecting the voltage drops in the stator resistance and since the quadrature component of the stator flux is zero (since the component d is coincident with the direction of the stator flux vector), the relationship between the stator flux and the stator voltage direct component becomes [9]: 20 1 3 r rL L s 1 4@ us ðsÞ ¼ Rr þ sAIds ðsÞ þ Iqs ðsÞxg ðsÞ5 1 þ Lr s r Lr Rr

Iqs ðsÞ ¼

ð7Þ

Rr

  r rL 1 Rr xg ðsÞ us ðsÞ  Ids ðsÞ 1 þ r Lr s rLs

ð8Þ

Rr

By expressing the direct component of the current as a function of the current quadratic component and the stator flux, the stator voltages become: Vds ðsÞ ¼

us ðsÞ rTr Rs xg ðsÞIqs ðsÞ þ Ts ð1 þ rTr sÞ 1 þ rTr s 1 þ ðTs þ Tr Þs þ rTr Ts s2 Vqs ðsÞ  xs ðsÞus ðsÞ

ð9Þ

ð10Þ

228

4.2

S. El Daoudi et al.

Torque Control

Taking into account that the quadrature component of the stator flux is zero, this makes it possible to have a linear function between the torque and the stator current in quadrature: 3 Ce ¼ Puds Iqs 4

ð11Þ

Since the torque control proposed consists in extracting the quadrature stator voltage, the torque equation can be rewritten using the current expression as a function of the quadrature voltage (9): 2 r ð1  r Þ T 3 Ls us ðsÞxg Ce ¼ P   4 ð1 þ rTr sÞ2 þ rTr xg 2

ð12Þ

By defining the closed loops of the stator flux and torque, the design of the controllers’ parameters is done by applying the optimum symmetry criterion. According to this criterion, the function which describes the studied system can be put in a general form depending on the system closed loops [1–10].

5 Simulation Results The global control algorithm which is presented theoretically in previous has been simulated by MATLAB/SIMULINK software. The simulation results were obtained for a three-phase 1.5 Kw squirrel-cage asynchronous motor. A comparative study between the proposed DTC and the conventional one will be presented using both of them the conventional PI as a speed controller. The following results illustrate the comparative analysis of the two control strategies. Figures from 4, 5 and 6 show respectively: Rotor and stator flux circular trajectories, stator current waveforms and Electromagnetic torque ripples. The figures are specified by (a) for classical DTC, and (b) for SPWM-DTC.

Fig. 4. Stator and rotor flux space vector trajectories for: (a) Classical DTC; (b) SPWM-DTC

Comparison Between Constant and Variable Switching Frequency

229

Figure 4 presents the stator and rotor flux circular trajectories for the two strategies. Comparing the Fig. 4 (a) and (b), it can be clearly seen that rotor and stator flux space vector trajectories in the SPWM-DTC technique have clean circular shape and low ripples level. The SPWM-DTC in Fig. 5 (b) shows a good sinusoid waveform and lower THD level (total harmonics distortion) about 16.4% for SPWM-DTC compared with 31.44% for the classical DTC (Fig. 5 (a)). We can clearly notice that the proposed algorithm has great responses compared to the classical strategy for which the flux and current shapes have less ripples.

Fig. 5. Stator current waveforms for: (a) Classical DTC; (b) SPWM-DTC

Fig. 6. Electromagnetic torque ripples for: (a) Classical DTC; (b) SPWM-DTC

In Fig. 6, the comparison of electromagnetic torque responses is presented, where both responses show a good reference tracking. It can also be seen that the SPWMDTC has reduced torque ripples compared to the classical DTC. Table 1 summarizes characteristics of the two control methods previously used. It can be seen that the SPWM-DTC keep the main advantages of the classical DTC while adding more features.

230

S. El Daoudi et al. Table 1. Comparison of control methods

Classical DTC advantages SPWM-DTC advantages • Structure independent of rotor parameters • Structure independent of rotor parameters • No coordinate transformation • No current control loops • No current control loops • Constant switching frequency • Low switching losses • PWM modulator

6 Conclusion In this paper, we present a comparative analysis of two strategies based on direct torque control. In order to overcome different drawbacks of the classical DTC and obtain a high performance control with reduced ripples and harmonics, the SPWM-DTC has been applied. Due to the use of the sine pulse width modulation, the torque and flux ripples have been reduced considerably compared with the classical strategy while working with a constant switching frequency which greatly reduces the switching losses. The use of the PI regulators instead of the hysteresis comparators improves significantly the system’s control by reducing the response time and eliminating the static error. The simulation results have proven the effectiveness of the SPWM-DTC compared to the classical strategy.

References 1. Lazrak, L., El Daoudi, S., Benzazah, C., Ait Lafkih, M.: Direct control of the stator flux and torque of the three-phase asynchronous motor using a 2-level inverter with sinusoidal pulse width modulation. J. Theor. Appl. Inf. Technol. 96(18), 6199–6210 (2018) 2. Alsofyani, I.M., Idris, N.R.N.: Simple flux regulation for improving state estimation at very low and zero speed of a speed sensorless direct torque control of an induction motor. IEEE Trans. Power Electron. 31(4), 3027–3035 (2016) 3. Ammar, A., Bourek, A., Benakcha, A.: Robust load angle direct torque control with SVM for sensorless induction motor using sliding mode controller and observer. Int. J. Comput. Aided Eng. Technol. 11(1), 14–34 (2019) 4. El Daoudi, S., Lazrak, L., Benzazah, C., Ait Lafkih, M.: Modified strategy of direct torque control applied to asynchronous motor based on pi regulators. In: Farhaoui, Y., Moussaid, L. (eds.) Big Data and Smart Digital Environment 2018, vol. 53, pp. 20–26. Springer, Switzerland (2019) 5. Takahashi, I., Noguchi, T.: A new quick response and high efficiency control strategy of an induction motor. IEEE Trans. Ind. Appl. 11, 820–827 (1986) 6. Casadei, D., Profumo, F., Serra, G., Tani, A.: FOC and DTC: two viable schemes for induction motors torque control. IEEE Trans. Power Electron. 17(5), 779–787 (2002) 7. Sebti, B.: Contribution à la commande directe du couple de la machine asynchrone. Ph.D. thesis, BEJAIA University, Algeria (2016) 8. Mohd Alsofyani, I., Idris, N.R.N., Lee, K.: The performance of lookup-table-based DTC of induction machines. IEEE Trans. Power Electron. 33(9), 7959–7970 (2017)

Comparison Between Constant and Variable Switching Frequency

231

9. Rodríguez, J., Patricio, C.: Predictive Control of Power Converters and Electrical Drives. Wiley, Chichester (2012) 10. Zelechowski, M., Kazmierkowski, M.P., Blaabjerg, F.: Controller design for direct torque controlled space vector modulated (DTC-SVM) induction motor drives. In: IEEE International Symposium on Industrial Electronics, Dubrovnik, Croatia, 20–23 June (2005)

Simulation and Analysis of Enhanced Perturb and Observe MPPT Algorithm Based on an Adaline Neural Network for Standalone PV System Ihssane Chtouki1(&), Houssam Eddine Chakir2, Patrice Wira3, Malika Zazi1, and Bruno Collicchio3 1

2

ERRER Lab, Mohammed V University, ENSET, Rabat, Morocco PMMAT Lab, Faculty of Science, Hassan II University, Casablanca, Morocco [email protected] 3 IRIMAS Lab, Haute Alsace University, Mulhouse, France

Abstract. This paper investigates the effectiveness of the Maximum Power Point Tracking (MPPT) algorithms of solar photovoltaic (PV) systems. Indeed, two efficient new control MPPT algorithms presented are based on the Perturb and Observe (P&O) method with a fixed step. The two suggested controllers are Artificial Neural Network (ANN). The first one uses Multilayer perceptron (MLP) learned by Levenberg Marquard (LM) learning algorithm as a neural regulator named (POPI-LMNN). The second one uses an Adaptative linear Neuron type (Adaline) learned by least mean square algorithm (LMS). The controllers are applied to a DC-DC boost converter inside a standalone solar photovoltaic conversion system used to feed an isolated area. As a result, the responses achieved through the learning approaches are more convincing. That is to say, they are faster and more efficient in terms of the power conversion. This is used to override the limitations of the traditional P&O technique, which are the fluctuations around the maximum power point with their low response tracking performance in the most severe cases of changing weather conditions. A comparative study between the three algorithms is done using Matlab/ Simulink Simpower system environment. The results of simulation demonstrate that the Adaline presents a very high performance in terms of rapidity and elimination of oscillations with a conserving energy. Keywords: Adaline MLP


Solar PV simulator
P&O
MPPT
Boost converter


1 Introduction The photovoltaic solar energy is currently considered as one of the most promising renewable energy sources due to its high availability anywhere in the world and the absence of polluting effects. The operation of the photovoltaic cells or the set of these components to form a panel is defined by its characteristic curves I-V and P-V, which show how these elements behave under different working conditions. They define the maximum power point (MPP), where the power extracted from the photovoltaic cell is © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 232–243, 2020. https://doi.org/10.1007/978-3-030-53187-4_27

Simulation and Analysis of Enhanced P&O MPPT Algorithm

233

maximum. This specific technique or algorithm is called “Maximum Power Point Tracking” (MPPT) [1]. The MPPT makes it possible to achieve a gain of up to 25% that is why it is so important. Technically, the MPPT uses an interface between the PV panel and the load. In this study, we regard as electronic interface a boost converter. Accordingly, various MPPT algorithms have been generally proposed. So we find classical, evolutionary, neuronal and intelligent ones [2, 3]. Indeed, in this work two types of MPPT algorithms are combined: classical and neural. To this end a conventional P&O [4] is chosen thanks to its simplicity of digital implementation. In fact this algorithm is considered as one of the most commonly used methods today. However, the latter still have some optimization problems such as the presence of oscillations around the PPM and confusion about the direction of tracking during changes in atmospheric conditions. Indeed, these search modes have their performances which depend strongly on an increment variable whose chosen value is the result of a compromise between precision and speed. Thus, a high value increment will improve response time at the expense of accuracy, while a low value increment will have the opposite effect and will be beneficial to achieve high accuracy. In this paper a fixed disturbance step P&O algorithm is chosen, the P&O algorithm is improved and the problem of local maxima in the power curve of the photovoltaic generator (PVG) is addressed by using two kinds of ANN. Two controllers are developed. The first one used MLPs NN and the second one is based on NN type Adaline. Two learning algorithms are introduced: LM [5] which gives existence to a new technique that are named POPI-LMNN controller and LMS algorithm with Adaline [6] applied for the control of PV system. The idea of the proposed controllers is to realize a neuronal control loop based on the design of an “Adaline” or on MLP instead of a conventional PI regulator in order to achieve the converter’s input voltage. Thereafter, the converter will have the ability to reduce the error between the reference voltage and the measured PV voltage by varying their duty cycle. For the Adaline proposed MPPT control technique a delay line (D) is introduced [7] in order to modify the impact that synaptic weights generate on the network response. In the reality, this structure is also used as an adaptive filter to minimize the oscillations of the P&O algorithm. A PI controller with fixed gains for a time-varying irradiation and temperature profile, which is subject to random variations, can give way to poor dynamic performance [8]. However, as soon as the weather conditions change, the controller can no longer track this change. Therefore in order to achieve this and to overcome this disadvantage, the Adaline adaptive linear controller is adopted in this article to regulate the voltage regardless of the variation in climatic conditions and to eliminate the oscillations. The main improvements of the two proposed algorithms are found in the elimination of oscillations around MPP, avoids misinterpretation of the location of MPP during a rapid change in climatic conditions, as well as making the system fast. Thereafter, the entire system composed of a PV panel, the boost converter, the proposed MPPT algorithms and the resistive load was implemented simulated using Qucs [9] to validate the sizing parameters of the boost converter and was confirmed using the MATLAB/Simulink SimPowerSystems. As results, the simulation is validated, besides the steady state and transient characteristics of each control algorithm described above are analyzed in depth and compared. After comparison, is found that the Adaline has a

234

I. Chtouki et al.

best dynamic performance and significantly reduces the oscillations problem that occur when applying the P&O function in response to abrupt changes in climatic conditions. Furthermore is finding that the speed and accuracy in extracting the maximum power generally are enhanced with this technique compared to MLP. After this introduction paragraph (Sect. 1), the paper is organized in the following structure, an emulator PV model, using Qucs simulator connected through a DC/DC boost converter, is presented in Sect. 2. Section 3 introduces the two proposed MPPT algorithms to control the boost converter using MLP with LM and PSO. Simulation results of the control schemes are given in Sect. 4 and concluding remarks are provided in Sect. 5 (Fig. 1). PV panel DC - DC

T

Boost

Vpv

Load Converter

G

Vpv

Ipv

Ipv ANN+LM

POPI-LMNN

Conventional P&O

Duty cycle calculation

α or P&O +Adaline

Fig. 1. The considered solar photovoltaic system with boost converter and resistive loads.

2 Modeling of the Emulate PV System 2.1

Equivalent PV Cell Electronic Model

A single-diode PV cell circuit has been elaborated to emulate the PV cell operations. This circuit is generally used in the literature [3] to produce the dynamic characteristics of the PV panel in Standard Measurement Conditions (SMC) and under changing weather conditions. It has been then implemented using the QUCS [9] free software simulator as shown in Fig. 3. Compared to Matlab/Simulink, the model implemented with Qucs doesn’t need many blocs to model equations. Consequently, we obtain more precise dynamic characteristics under different climatic change. The Fig. 2 presents the I-V and P-V curves of the PV panel for a given value of temperature and irradiation.

Fig. 2. The dynamics operation of pv cell using Qucs.

Simulation and Analysis of Enhanced P&O MPPT Algorithm

235

The considered PV model is based on a datasheet from a real PV panel given by Table 1. 2.2

PV Connected Through a DC/DC Boost Converter

In our application, a DC-DC boost converter acts as an interface between the solar panel and the supplied resistive load. Figure 3 gives the equivalent circuit model simulated using Qucs. The switch S1, symbolized here as a power MOSFET, is rendered conductive periodically with a duty cycle a. The relationship between the input and the output voltages of the boost is given by [10]: VLoad ¼

1 VPv 1a

ð1Þ

Fig. 3. (a) Simulated PV model with boost converter and resistive load using Qucs, (b) input/output boost converter signals.

To size the boost parameters, we are based on the inductance sizing with the switching frequency and the output capacitor sizing. For this reason, we use the following equations [11]: Iemax 4:f s :DVbus

ð2Þ

2Vpv Vbusmax ¼ 4:F:DIe 4:f s :DIe

ð3Þ

C= L=

The calculated values of the boost parameters are deduced from these equations with the PV characteristics given in Table 1 and are: L = 0.2 H, Ce = 10−6F, and Cs = 10−3F. These parameters are validated twice: (i) using Qucs in the first time without control; (i) using Matlab/Simulink with different proposed controllers.

3 Proposed MPPT PV Controllers’ Strategies 3.1

Problem Reformulation

The operating principle of the conventional P&O [4] approach is based on a fixed or variable step disturbance generated to produce a reference signal for the external

236

I. Chtouki et al.

control loop. The disturbed signal is either the reference voltage or the reference current of the network. The most important element in the disturbance is the high increment step which quickly reaches the optimal operating point and thus reduces the system response time. However, this important step causes steady-state oscillations around them and they persist as long as the climatic conditions are stable. On the other hand, a small step reduces these oscillations while slowing down the system which becomes too slow. In this paper, a fixed disturbance step P&O algorithm is chosen. Two multilayer perceptron MPPT neural controllers are suggested to improve the classical P&O: the first is an ANN based on an automatic learning using LM while the second is an adaline. Based on this, the chosen fixed step is determined according to the system following the experimental tests carried out previously. The MPP measures the voltage and the PV current. Then, the P&O algorithm determines the reference voltage. The P&O’s objective is to decide only the Vpv reference value and this is done at a certain time interval. Two control loops are used for each approach based on the neural network design instead of the conventional PI trying to achieve the converter input voltage. It also reduces the error between the reference and the measurement varying the duty cycle of the converter. 3.2

Artificial Neural Network MPPT Based on LM Learning Algorithm

Solar panel DC- DC

Converter

(LM-NN) feedforward network

Coventional PO

Load

Boost

I

[0,1]

Vpv*

+

PI controller

Vpv

Fig. 4. Proposed MPPT control system based on LM-NNPOPI algorithm.

In this study, we are interested in the multilayer feed-forward neural network (or multilayer perceptron) which has a very simple structure in which the neurons are connected only in one direction. The signals from the input layer pass through the hidden layer to eventually reach the output layer. The Neural Networks (NNs) used in this application are in the form of 1-10-1. The global control PV MPPT system with the neuron network is given by Fig. 4. The parameters of the ANN can be calculated using the following equation [12]:
 IH 3ði,jÞ; i 6¼ j Hi;j ¼ r BH i þ wij :Xj

i [j

ð4Þ

Simulation and Analysis of Enhanced P&O MPPT Algorithm

237

with i [ j and i and j are respectively the number of neurons in the hidden layer and the input layer. H : Hidden layer, IH: Input to hidden layer, r: The sigmoid activation function. The output NNs can be calculated using the following equation:

 HO H IH /i;j ¼ r BO i þ wij :r Bi þ wij :Xj

ð5Þ

with: O: Output layer, HO: Hidden to output layer and Xj is the input of NNs   eDVpv ¼ VPVref  VPV that will be improved to maximize the output power of the PV system. The control of the converter is given by the following equation:   H  HO IH /DVpv ¼ r BO 1 þ w101 :r B10 þ w101 :XDVpv

ð6Þ

To replay the objective function condition, the activation function is given by [10]: r ¼ f ð xÞ ¼   f /DVpv ¼

1 1 þ ex

ð7Þ

1 1þe

ðB1 þ w110 :

1

ð

 B10 þ w101 :XDV pv 1þe

Þ

ð8Þ

where the mean square error is 2 1 1
/  /eDVpv E ¼ e2 ¼ 2 2 0 1@ / E¼ 2

ð9Þ 12

1 1þe

ðB1 þ w110 :

ð

1

 B10 þ w101 :XDV pv 1þe

A Þ

ð10Þ

Learning is the process (of calculations) that makes it possible to update the weights of the neurons from one or more measurements. The basic idea of learning, which is behind all NNs learning algorithms, is to adjust the weights to minimize some measures of the error on the learning sample. LM algorithm [5] is an improvement of the classical Gauss-Newton method in solving the no-linear least squares regression problems. This is the recommended method for the non-linear (regression) problems of least squares applied in this paper in order to adapt the architecture of NNs as well as to avoid the problem of the choice of the initial architecture of the network called “selfconstructive” methods exist: it is a question of adding the neurons during the learning so that the learning can be well done. But these methods often encounter the problem of The objective function to solve the problem depends on the NNs input “over-learning”.  eDVpv , the synaptic weight Wi;j , the bias Bj , and the function of /DVpv which is the optimal control estimated value. In this study, the loss function is considered as an objective function used by LM algorithm which can be expressed as a number of errors squared in the form [5]:

238

I. Chtouki et al.

f=

X

e2i;j ¼

X  e2 ; i = 1, 2 and j = 1,. . .:; 10

0  X @ /  f=  

ð11Þ

12     A 1  ðB1 þ w110 :  ðB10 þ w101 :eDv Þ pv 1þe 1

1þe

ð12Þ

with i, j respectively denote the number of the neurons in the input/output layers and the number of the neurons in the hidden layer. 3.3

Adaline MPPT Based Control Algorithm

Solar panel

Ich Ipv

DC - DC

load

Boost

Vpv

Converter

Vch

[0,1]

-

Conventional (P and O)

PWM

Vpv-ref* Vpv

+

+ W0

z-1

n i=1 xi

z-1 z-1

yest

W1

n W2

wi.xi i=1 W3

E(k)

Learning Rule Algorithm W(k+1) =w(k)+alpha*e*x

Fig. 5. The principle of the proposed MPPT algorithm using Adaline controller.

The second proposed technique to control power converter is an adaptive MPPT neural controller type Adaline. The Adaline intelligent controller consists of a single neuron with linear activation function like is shown in the Fig. 5, whose output is calculated as [13]: z(k) ¼ f ðvðkÞÞ

ð13Þ

with: v(k) ¼ bðkÞ þ

3 X

wi ðkÞxi ðkÞ

ð14Þ

i¼0

zi ¼ yi ¼ wTi x þ wi0 x 2 Rn ; y 2 Rm ; z 2 Rm As for Perceptron, the separator planes of the classes have as equation [7]:

ð15Þ

Simulation and Analysis of Enhanced P&O MPPT Algorithm

wTi x þ wi0 ¼ 0 8i ¼ 1; . . .; m

239

ð16Þ

Assuming the linear activation function and neglecting the input bias, we obtain the following formula: zðtÞ ¼ yðtÞ ¼

3 X

wi ðtÞxi ðtÞ

ð17Þ

i¼0

hence f is the activation function, v: argument of the activation function, b bias value, xi input signals, p(t) represents the network inputs over time, wi is the synaptic weight vector, z(t) is the output of the neural network, y(t) is the weighted sum of the inputs and weights before the activation function and k indicates the delay time compared to the first input data. The objective of the LMS algorithm is to reduce the error between the desired response and the actual response of the neural network. the most common error function is the MSE, by developing the Eq. (15), the network response can be obtained from the first instant of time. Let us take the following input-output torque: p(t), d(t) presented at time t. With pðtÞ 2 p1 ; p2 ; . . .; pq is the representation of the network inputs [7]. d(t) is the desired output vector: dðtÞ ¼ ½d1 ðtÞ; . . .; dm ðtÞT

ð18Þ

2 3 p(t) 6 Wð1Þ 7 6 pðt  1Þ 7 6 7 T 7 W ¼6 7; p ¼ 6 4 pðt  2Þ 5; X ¼ ½W ; Z ¼ ½P; yð1Þ ¼ X Z 4 Wð2Þ 5 pðt  3Þ Wð3Þ

ð19Þ

Knowing that 2

Wð0Þ

3

The output of the Adaline is calculated as a function of p(t) at time t by: yi ¼ wi ðtÞT pðtÞ þ wi0 ðtÞ ¼ XiT ðtÞZ(t)

ð20Þ

Through the adjustment of the weights Xi the error can be minimized as given by: ei ðtÞ ¼ di ðtÞ  yi ðtÞ

ð21Þ

The mathematical expectation of the square of the error that the LMS algorithm consists of minimizing, taken on all the pairs: ½Pk ; dk  for k = [1, 2, …, q] [7] is: h i   E e2i ðtÞ ¼ E ðdi ðtÞ  yi ðtÞÞ2

ð22Þ

240

I. Chtouki et al.

4 Results and Discussion

2 Vpv (volts) Irradiation (W/m )

This section presents the simulation results of the two proposed techniques compared with the P&O technique. Based on POPI-LMNN and POPI-PSONN, we have made an improvement on the P&O control. The simulation was done using the Matlab/SimulinkTM environment. The Fig. 6 illustrates the irradiation profile with large variation between 800 w/m2 and 1000 w/m2. It also compares the voltage variations (Vpv) and the power variations (Ppv) provided by the three algorithms as well as the duty cycles. At the time of sudden changes in the environmental conditions, the Adaline controller presents the best dynamic performance thanks to its almost instantaneous convergence time and a significant reduction in oscillations towards the point of maximum power in comparison with the time and the oscillations presented in the P&O algorithm. Also we find that the proposed PIPO-LMNN algorithm presents an acceptable and even a better result than the one of the P&O algorithm. On the other hand, the P&O showed slower behavior to reach the maximum value. 1000 900 800

2

4

6

8

10

12

PO

20

POPI-LMNN Proposed Adaline

0 Dutycycle d % Ppv (Watts)

0

40

0

2

4

6

8

10

12

0

2

4

6

8

10

12

0 0

2

4

8

10

12

150 100 50 0

1 0.5 6 Time (s)

Fig. 6. PV source state parameters, Vpv , Ppv , duty-cycle during rapid periodic change irradiance profile using proposed Adaline, POPI-LMNN and P&O MPPT controllers.

Fig. 7. Output Ppv source produced power, Ps load consumed power during rapid periodic change irradiance using proposed Adaline, POPI-LMNN and P&O MPPT controllers.

Simulation and Analysis of Enhanced P&O MPPT Algorithm

241

Fig. 8. Output Vpv source produced voltage, Vs load output voltage during rapid periodic change Irradiance, using proposed Adaline, POPI-LMNN and P&O MPPT controllers.

From the Fig. 6 we observe that the behavior of the technique, using Adaline, converge rapidly to the maximum value and does not show an oscillation between values as occurred in the P&O. At the start of the profile, we can notice that the Adaline tracks the PPM before the P&O in duration of about 0.12 s. Moreover, with each variation, the P&O loses almost 7% of its energy in joules compared to Adaline which has no losses while the PIPO-LMNNN loses almost 3% of its energy. Evidently, in the steady state, the P&O oscillates around the PPM between 143.6 W and 145 W.

Fig. 9. Output Ipv source produced current, Is load output current during rapid periodic change irradiance using proposed Adaline, POPI-LMNN and P&O MPPT controllers.

Figures 7 and 8 show the tracking performances of two proposed MPPT algorithms in respect to the conventional P&O MPPT algorithm with a high interest in terms of the electrical variables. As shown in Fig. 9, the voltage is amplified, under the standard climatic conditions (G = 1000 w/m2 and T = 25 °C), from 34.14 V: (i) to 101 V by

242

I. Chtouki et al.

using the Adaline approach, (ii) to 86.2 V with the POPI-LMNN approach, (iii) and to 62 V with the P&O approach. Thus, the current is reduced from 4.2 A, to 1.5 A by using the Adaline approach. Also, it is reduced to 1.7 A with the POPI-LMNNN approach, and to 2.2 A with the P&O approach. Therefore, a poor design would represent an increase in power losses in each of the devices and the system in general. From this analysis, we can deduce that the Adaline approach has a good performance without energy losses from the fluctuations and with a fast and clean energy response. Table 1. Parameter of the PV Solarex SOLEX FSM145W-24 in STC (1 kW/m2, 25 °C) and boost converter. Parameter





Typical peak power Pmpp   Voltage et peak power Vmp   Current at peak power Imp Short-circuit current ðISc Þ Open-circuit Voltage ðVoc Þ Temperature coefficient of Isc ð/Þ Temperature coefficient of Voc ðbÞ Series cell ðNs Þ   Parallel cell Np Inductor in the boost converter circuit (L) Capacitor in the boost converter circuit ðCe Þ Capacitor in the boost converter circuit ðCs Þ Switching frequency

Value 145 W 34.4 V 4.2 A 4.7 A 43.5 V 0.0065%/°C 0.36099%/°C 72 1 0.2H 10e−6F 1000e−6F 1 kHz

5 Conclusion This paper demonstrates that the photovoltaic modules must have a maximum power point tracking algorithm to continuously provide the highest possible power to the system in order to improve the performance and the efficiency of the distributed generation system. In this study, the two proposed MPPT algorithms subjected to improve the performance of the P&O algorithm, called POPI-LMNNN and Adaline have shown a high efficiency thanks to the integration of the NNs using the LM and the LMS learning algorithms. Indeed, this combination makes the system response fast, efficient as well as good at eliminating the fluctuations around the maximum power point. Moreover, it minimizes the power losses. The result of the research of the new MPPT algorithms has led manufacturers to market a wide range of the MPPT controllers to quickly and efficiently track maximum power points. Responding to this challenge, we have chosen to improve one of the most popular thanks to its simplicity of implementation according to the cost. Additionally, we have proposed to enhance the MPPT algorithms to have a better quality/price ratio. In the end, we have been able to integrate a new MPPT algorithm

Simulation and Analysis of Enhanced P&O MPPT Algorithm

243

with a good performance using Adaline. As a perspective, the extended version of this work will be an implementation of the proposed algorithms within a real time solar conversion system using the real time dSpace1104 broad.

References 1. Lasheen, M., Abdel-Salam, M.: Maximum power point tracking using Hill Climbing and ANFIS techniques for PV applications: a review and a novel hybrid approach. Energy Con. Manag. 171, 1002–1019 (2018). https://doi.org/10.1016/j.enconman.2018.06.003 2. Ben Salah, C., Ouali, M.: Comparison of fuzzy logic and neural network in maximum power point tracker for PV systems. Electr. Power Syst. Res. 81, 43–50 (2011). https://doi.org/10. 1016/j.epsr.2010.07.005 3. Pillai, D.S., Rajaseka, N.: Metaheuristic algorithms for PV parameter identification: a comprehensive review with an application to threshold setting for fault detection in PV systems. Renew. Sustain. Energy Rev. 82(3), 3503–3525 (2018). https://doi.org/10.1016/j. rser.2017.10.107 4. Seyed Mahmoudian, M., Mohamadi, A., Kumary, S., Maung, A., Oo, T., Stojcevski, A.: A comparative study on procedure and state of the art of conventional maximum power point tracking techniques for photovoltaic system. Int. J. Comput. Electr. Eng. 6(5), 402–414 (2014). https://doi.org/10.17706/ijcee.2014.v6.859 5. Smith, J.S., Wu, B., Wilamowski, B.M.: Neural network training with Levenberg– Marquardt and adaptable weight compression. IEEE Trans. Neural Netw. Learn. Syst. 30(2), 580–587 (2019). https://doi.org/10.1109/TNNLS.2018.284677 6. Zhou, Y., Ai, Q., Xu, W.: Adaline and its application in power quality disturbances detection. In: Proceedings of the 5th WSEAS International Conference on Signal Processing, Computational Geometry & Artificial Vision, Malta, 15–17 September, pp. 41–46 (2005) 7. Borne, P., Benrejeb, M., Haggège, J.: Les réseaux de neurones: présentation et applications. Editions OPHRYS. Technip, Paris (2007). 152 p. 8. Laib, H., Chagh, A.E., Wira, P.: A neural and fuzzy logic controller to improve the performance of a shunt active power filter. Rev. Sci. Technol. – RST 6(1), 1–11 (2015) 9. Pareja, M.: PV cell simulation with QUCS, a generic model of PV cell 20(07) (2013) 10. Chtouki, I., Wira, P., Zazi, M.: Comparison of several neural network perturb and observe MPPT methods for photovoltaic applications. In: The 19th International Conference on Industrial Technology (ICIT 2018) Lyon, France, pp. 1–6 (2018) 11. Özgür, C., Teke, A.: A hybrid MPPT method for grid connected photovoltaic systems under rapidly changing atmospheric and Technology. Electr. Power Syst. 152, 194–210 (2017) 12. Bishop, C.M.: Neural Networks for Pattern Recognition. Clarendon Press, Oxford (1995) 13. Kaminski, M., Orlowska-Kowalska, T.: FPGA implementation of ADALINE-based speed controller in a two-mass system. IEEE Trans. Ind. Inf. 9(3), 1301–1311 (2013)

Performance Assessment of Solar Dish-Stirling System for Electricity Generation in Eastern Morocco Hanane Ait Lahoussine Ouali, Benyounes Raillani, Samir Amraqui, Mohammed Amine Moussaoui(&), and Ahmed Mezrhab Laboratory of Mechanics and Energetics, Faculty of Sciences, Mohammed First University, 60000 Oujda, Morocco [email protected]

Abstract. This paper presents the simulation results of a Dish-Stirling power plant with a nominal power of 100 MW considered to be installed in northeastern Morocco: Oujda (latitude: 34.68 °N, Longitude: −1.9 °E). The DishStirling System of the Stirling Energy system (SES) Company with 4–95 kinematic type motor, contains 4 cylinders and the hydrogen uses as a working fluid. The number of collectors required is 4000 with an area of 900000 m2. For this study, the System Advisor Model (SAM) software has been used in order to investigate a potential technical and economical installation of a 100 MW concentrating solar thermal power plant. The simulation results for an annual Direct Normal Irradiation of 1990 kWh/m2/yr predicted that the system would produces 159.3 GWh annually, achieving a maximum power in May. The Levelized Cost of Electricity (LCOE) of the plant and the capacity factor would be 0.16 $/kWh and 18.3%, respectively. All these results should encourage the Moroccan government to exploit this technology for electricity production which would lead to reduction of CO2 and a sustainable development of this region of Morocco. Keywords: Capacity factor
Dish-Stirling
LCOE
Morocco
Oujda
Solar thermal power plant
System Advisor Model

1 Introduction In last decades, thermodynamic solar energy has been developed around the world and some technologies are reaching maturity to compete with traditional energy sources. One can cite the parabolic trough or the solar tower plant [1, 2]. The exploitation of solar technologies will enable Morocco, a well-known country with abundant sunshine, to reduce its energy dependence and its electricity costs, as well as to meet the energy and environmental challenges related to the depletion of fossil fuels and at risk of danger to the environment. Therefore, the kingdom launched the Moroccan Solar Plan, which aims to produce 2 GW of electricity from solar power by the year 2020 (Fig. 1). Dish-Stirling technology is one of the CSP technologies currently developed and the subject of recent industrial and research projects. Many studies have been achieved © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 244–252, 2020. https://doi.org/10.1007/978-3-030-53187-4_28

Performance Assessment of Solar Dish-Stirling System

245

Fig. 1. Energy mix in Morocco by 2020 compared to 2010 [6].

in terms of the investigation of the solar Dish-Stirling performances. Abbas et al. [3] have been presented simulation results of a large-scale Dish-Stirling plant under the Algerian climate conditions: ‘Algiers’, ‘In Salah’ and ‘Tamanrasset’ using the software SAM (Solar Advisor Model) to assess the net monthly energy production and the levelized cost of electricity. The authors indicate that ‘Tamanrasset’ is the appropriate site that produces the highest net annual electric power production and the lowest LCOE. Bravo et al. [4] evaluated a 10 kW Dish-Stirling facility and compared with a similar photovoltaic facility. The comparison used the environmental outputs of the inventory as well as the environmental impact valued by damage categories. The different results show that the level of environmental impacts is similar for both systems (PV and Dish-Stirling) technologies. In another study, Monné et al. [5] studied the integration of hybridization and thermal energy storage in a stand-alone Dish-Stirling with the capacity of 10 kWe. The hybridization was analyzed for a natural gas, and a biogas. The analysis of performance of the Dish-Stirling system show that hybridization has an advantage that depends on the nature of fuel used, being about 20% for natural gas and 112% for biogas in relation to an output power of solar-only operation. In this context, to highlight the performance of Dish-Stirling system in the eastern region of Morocco, a techno-economic performance of Dish-Stirling system model is realized using the System Advisor Model (SAM).

2 Description of Dish-Stirling System The Dish-Stirling technology consists of a parabolic shaped solar concentrator system that tracks the sun throughout the day. This apparatus focuses the radiation on the Stirling engine’s heat absorption unit placed at the center of the parabola. The solar thermal energy thus concentrated is transformed into electricity directly injectable on the grid. An example of Dish-Stirling solar system is seen in Fig. 2. The Dish-Stirling allows a saving of the ground because the surface of its installation is very small compared to other solar technologies (2.5 ha/MW). In addition, its structure allows it to adapt to all types of land, which can produce electricity in the most difficult access areas.

246

H. Ait Lahoussine Ouali et al.

Fig. 2. (a) Simplified scheme of a Dish striling, (b) Dish-Stirling from TES society.

2.1

Receiver

The receiver absorbs concentrated solar flux and converts it into thermal energy that heats the working gas of the Stirling engine. The absorption surface is generally placed behind the focal point of a concentrator so that the flux density at the absorption surface is reduced. An opening is placed in the fireplace to reduce radiation and convective heat loss from the receiver. The walls of the cavity between the receiving opening and the absorber surface are refractory surfaces. The performance of a receiver is defined by the thermal efficiency of the receiver which is defined as the useful thermal energy delivered to the engine divided by the incoming solar energy into the receiving aperture [7]: grec ¼ sa 

4 4 UðTrec  Tamb Þ þ rFðTrec  Tamb Þ gconc CRg Ib;n

ð1Þ

Where U, F and Ib,n are respectively the coefficient of heat loss by convectionconduction, the equivalent radiative conductivity and insolation. As can be seen in Eq. (1), the efficiency of the receiver can be improved by increasing the transmission factor of the insulation. This induces the increase of the absorption coefficient of the surface and the reduction of the operating temperature or of the capacity of the cavity to lose heat by conduction, convection and radiation. 2.2

Stirling Engine

The Stirling engine consists of a sealed system filled with a working gas (typically hydrogen or helium), which is alternately heated and cooled. The engine operates due to the movement resulting from compression and expansion of the working gas alternatively to the cold or hot source. We note that the overall efficiency of the DishStirling system depends on the efficiency of each component of the system. Overall performance is written:

Performance Assessment of Solar Dish-Stirling System

g ¼ gc gR gsti gG

247

ð2Þ

Where ηc, ηR, ηsti and ηG are respectively the concentrator efficiency, the receiver efficiency, the stirling engine efficiency and the generator efficiency.

3 Methodology and Simulation Inputs A Dish-Stirling power plant with nominal power 100 MW is considered to be installed in eastern Morocco. It is well known that Morocco is a country with abundant sunshine and eastern Morocco is one of the regions with the highest values of direct normal radiation. The model is implemented in the SAM simulation software, which allowed us to predict changes in the net electrical energy produced and the total efficiency of the system under study. In this particular location, we use meteorological data provided from our meteorological station installed at the rooftop of Oujda University. 3.1

Meteorological Data

Figure 3 shows the monthly Global Horizontal Irradiance (GHI), Diffuse Horizontal Irradiance (DHI) and Direct Normal Irradiance (DNI) measured by our meteorological station. The DNI values are delivered with a pyrheliometer (Kipp & CHP1), while the values of DHI are provided with a shaded pyranometer (Kipp & Zonen CMP21). In addition, the GHI total hemispheric sum of the DNI and DHI components is available on a horizontal surface. GHI data represent the amount of solar radiation incident on horizontal flat plate solar collectors. These basic solar components are related by the following equation: GHI ¼ DHI þ DNI  cosðSZAÞ

ð3Þ

Where SZA is the solar zenith angle. As we can see, during the hours of operation of the plants the DNI values vary between 500 and 800 W/m2 for almost every month except November. The diffuse insolation due to the clouds varies between two average values: from 100 W/m2 (December and January) to 350 W/m2 (July and August). Moreover, the horizontal global insolation which is the sum of the two sunstrokes direct and diffuse varies between two average values: from 450 W/m2 to 900 W/m2. Figure 4 shows the daily variation of ambient temperature in Oujda city. Let us note that the maximum daily average ambient temperature reaches 37 °C in August while the minimum temperature is 1 °C in February. We can notice that the Direct Normal Irradiation is the most important parameter to evaluate, simulate and implement the concentrated solar power technologies.

248

3.2

H. Ait Lahoussine Ouali et al.

Input Data

The dish Stirling power plant using in this present work has an output of 100 MWe and the working fluid is hydrogen, based on Dish-Stirling System of 25 kWe for each module, some parameters of which are indicated in Table 1.

Fig. 3. The monthly variations.

Fig. 4. The daily variation of ambient temperature in Oujda city.

Performance Assessment of Solar Dish-Stirling System

249

Table 1. Design characteristics of the proposed plant. Parameter Value Number of collectors, North-south 50 Number of collectors, East-west 50 Number of collectors 4000 Total solar field area 900000 m2 Total capacity 100 MW Working fluid Hydrogen Absorber absorptance 0.9 Absorber surface area 0.6 m2 Single unit nameplate capacity 25 kW

4 Simulation Results Figure 5 presents the monthly variation of net electric output of the Dish-Stirling plant proposed for Oujda city for a simulation period of one year. The results show that the high power values are reached over the year and the net annual energy is 159.3 GWh. The peak power reaches 18.2 GWh in May and the lowest value of 6 GWh is noted in November. Let us note that the capacity factor of the system is 18.3%.

Fig. 5. Net electric output versus months of a year.

As can be seen in Fig. 6, the annual energy flow is a function of the total field strength of each component of the system and the losses that occur during the energy transfer between each component, from solar energy to the net power output. Each bar indicates the losses associated with the following sequential transfers of solar and thermal energy, such as shading errors in the concentrators, thermal losses in the receivers, thermal and mechanical losses in the Stirling engine.

250

H. Ait Lahoussine Ouali et al.

Fig. 6. Dish Stirling plant waterfall chart.

Figure 7 shows the results of a one-year simulation of the system’s net hourly output from the proposed 100% solar power plant without hybridization and without thermal storage. It can be noted that the efficiency of the system reaches 25% per day during the operating hours of the proposed installation (between 9:00 am and 6:00 pm) for the 5 days selected during the summer week (July 27th to 31st) that can be considered better.

Fig. 7. Hourly net system efficiecy.

5 Economic Analysis The economic project study is important for the choice of the technology used. Thus, if the location is well chosen, solar thermal power plants will be economically reliable for the production of electricity.

Performance Assessment of Solar Dish-Stirling System

251

The Levelized Cost of Electricity (LCOE) is a decisive parameter as to the profitability or otherwise of a CSP project. It can be calculated by Eq. 4 [8]: LCOE ¼

Cinv ðCRF þ fins Þ þ CO&M E

ð4Þ

Where CRF, E, fins, Cinv and CO&M are the capital recovery factor, the annual electricity generation, the annual insurance cost, the total investment costs and the cost of operation and maintenance respectively. In order to determine the dominant cost fraction of the whole plant, Fig. 8 shows the relative contribution of each project cost. As can be seen, the cost of the project is mainly composed of the cost of the collector.

Fig. 8. Costs per watt for each system component.

The cost assumptions and financial parameters are listed in Table 2 for the DishStirling plant.

Table 2. Cost assumptions for the Dish-Stirling plant. Parameter Value Collector cost 400 $/m2 Receiver cost 250 $/m2 Engine cost 500 $/m2 Contingency 7% EPC 11% Estimated total installed cost per capacity 2893 $/m2 Fixed cost by capacity 65.00 Real LCOE 0.16 $/kWh

252

H. Ait Lahoussine Ouali et al.

6 Conclusion In this paper, we are interested in thermodynamic solar technology based on the Stirling engine concentrator. The performance of existing systems is improving and the first commercial plants are emerging. To this end, a technical and economic feasibility study of a dish-Stirling solar power station in eastern Morocco is presented. The proposed plant has a nominal power of 100 MW. The results of the simulation are provided for an annual DNI of 1990 kWh/m2 reaching its maximum power in May. The electricity cost of the LCOE power station is USD 0.16/kWh and the capacity factor is 18.3%. The results obtained should encourage the Moroccan government to use this type of technology to cope with a growing energy need and to avoid any degradation of the environment.

References 1. Ouali, H.A.L., Merrouni, A.A., Moussoaui, M.A, Mezrhab, A.: Electricity yield analysis of a 50 MW solar tower plant under moroccan climate, Morocco. In: 1st International Conference on Electrical and Information Technologies ICEIT (2015) 2. Ouali, H.A.L., Guechchati, R., Moussoaui, M.A., Mezrhab, A.: Performance of parabolic through solar power plant under weather conditions of oujda. Appl. Sol. Energy 53(1), 45–52 (2017) 3. Abbas, M., Boumeddane, B., Said, N., Chicouche, A.: Dish stirling technology: a 100 MW solar power plant using hydrogen for Algeria. Int. J. Hydrogen Energy 36(7), 4305–4314 (2011) 4. Bravo, Y., Carvalho, M., Serra, L.M., Monné, C.: Environmental evaluation of dish-Stirling technology for power generation. Sol. Energy 86(1), 2811–2825 (2012) 5. Monné, C., Bravo, Y., Moreno, F., Muñoz, M.: Analysis of a solar dish–stirling system with hybridization and thermal storage. Int. J. Energy Environ. Eng. 5(2), 80 (2014) 6. Hochberg, M.: Renewable energy growth in Morocco an example for the region, MEI policy focus (2016) 7. William, B., Richard, B.A.: Compendium of solar dish-stirling technology. Technical report (1994) 8. Carolina, M.C., Serrano, D., Hernández, J.G., Delgado, S.S.: Solar multiple optimization of a DSG linear fresnel power plant. Energy Convers. Manag. 184, 571–580 (2019)

Real Time Implementation of SPWM Signal Generation Technique for a New Five Level Inverter Using Microcontroller Hajar Chadli1(B) , Zakariae Jebroni2 , Sara Chadli1 , Mohammed Saber3 , Khalid Salmi1 , Abdechafik Derkaoui1 , and Abdelwahed Tahani1 1

Laboratory LES, Sciences Faculty, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected], [email protected] 2 SEEER, ENSAO, Universit´e Mohammed Premier, Oujda, Morocco 3 Laboratory SmartICT, ENSAO, Universit´e Mohammed Premier Oujda, Oujda, Morocco [email protected]

Abstract. For many years, the evolution of power electronics became very important in a world where energy aspects have become an essential issue. The appearance of multilevel inverter is one of the results of this evolution. This type of inverter provides high power quality with fewer harmonics. So as to improve the performance of multilevel inverters, many modulation techniques have been proposed. Generally, pulse width modulation (PWM) control is the most used. In this paper, we used a sinusoidal Pulse Width Modulation (SPWM) strategy to control our new 5-level inverter. The performance of our proposed five-level inverter with respect to harmonic content and number of switches is simulated using MATLAB/Simulink. A hardware prototype is developed to verify the performance of the developed system using microcontroller ATmega2560. Keywords: Multilevel inverter · Sinusoidal Pulse Width Modulation (SPWM) · 5-level inverter · Matlab-Simulink · Microcontroller ATmega2560

1

Introduction

For so long, power electronics has known a very significant growth. Nowadays, this aspect of electrical engineering touches upon vast and very diverse fields of application for powers, ultimately covering a wide range (some watt to several hundreds of megawatts). The static conversion structures that make up mainly power electronics applications are becoming more and more powerful, thus, the technology has had to adapt to this growth in the power to be converted. c The Editor(s) (if applicable) and The Author(s), under exclusive license  to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 253–262, 2020. https://doi.org/10.1007/978-3-030-53187-4_29

254

H. Chadli et al.

This growth was allowed due to the evolution of semiconductor component technologies. The evolution of the voltage and current calibers as well as the improvement of the performances of these components made it possible to use more preferment power electronics for higher power applications [1]. The appearance of multilevel converters is one of the results of this evolution; they are used for high power Alternative machines [2]. A multilevel inverter (MLI) is a linkage structure of multiple input dc levels (obtained from dc sources and/or capacitors) and power semiconductor devices to synthesize a staircase waveform [3]. Several topologies have been proposed for multilevel inverters: diode-clamped (neutral-clamped) [2]; capacitor-clamped (flying capacitors) [4–6]; and cascaded multi cell with separate dc sources [4,7,8]. In this work, we will study new architecture of H-bridge inverter with five levels structure that increases the power delivered to the load, and improve the shape of the output voltage so as to be closer to the sinusoid. Improving the form of the output voltage of converters is a very active area of research, which continues to grow by taking advantage of semiconductor technology and digital computers. To improve the output voltage of an inverter, we can act on its structure or the method of its control. Generally, pulse width modulation commands are used. Several strategies of this type of control are encountered, such as triangulationsine modulation, hysteresis modulation, and vector modulation [9,10]. In this paper we used Sinusoidal Pulse Width Modulation (SPWM) technique, to control a new five-level inverter based on six MOSFETs realized by [11]. The results confirm the efficiency of the proposed controller. The experimental results are presented to confirm the simulation results. This paper is organized as follows: An introduction, a description about the proposed inverter circuit with its modes of operations, the description of used modulation techniques, simulation and experimental results, and conclusion.

2 2.1

The Proposed Five-Level Inverter Description of the Proposed Architecture

The proposed multi-level inverter is five-level inverter (see Fig. 1). It contains two DC sources delivered by two solar panels E1 = 36V and E2 = E1/2 and six switching (S1 − S6) each switch is composed of a MOSFET transistor and a diode; the MOSFET switches are used because of its fast switching capability. The portion that forms (S1–S4) gives the voltage at ±E1 and zero levels. The remaining switches S6 and S5 are responsible for making the output voltage at ±E2 levels. 2.2

Operation Mode

For an N-level converter, we have N possible operating sequences to generate the N voltage levels. Particularly for five levels there are five sequences of operation.

Real Time Implementation of SPWM Signal Generation Technique

255

Fig. 1. SPWM command of 5-level inverter

We link a functioned Fsi with each switch. Fsi translates the open or closed state of the switch Si:  1 if S switch is closed (1) Fsi = 0 if S switch is opened The operation can be easily explained with help of the Table 1, below, which contains the output voltage levels and the corresponding switch states. Table 1. Switching sequence for simplified five-level inverter S1 S2 S3 S4 S5 S6 Va - Vb= Vab

3

0

1

0

1

0

0

0

0

0

0

1

0

1

+E1/2

1

0

0

1

0

0

+E1

0

1

0

0

1

0

−E1/2

0

1

1

0

0

0

−E1

Modulation Technique

The principle of this command is to use the intersection of a reference or modulating wave (which is the image of the output wave we want to obtain) sinusoidal, with a modulation wave or carrier, generally triangular where the term triangulo-sinusoidal (see Fig. 2) [12]. If the reference is greater than the carrier signal, then the active device corresponding to that carrier is turned on, and if the reference is less than the carrier signal, then the active device corresponding to that carrier is turned off [13]. This method uses N − 1 carrier signals to generate the output voltage of the N-level inverter. To generate the switching pulses of our 5-level inverter, four carrier signals were compared to a reference sine wave, the amplitude of the

256

H. Chadli et al.

Fig. 2. Operation of spwm technique

Fig. 3. SPWM command of 5-level inverter

carrier signal having been divided into four regions to fit the amplitude of sine wave of reference. As a result, this comparison gave us four pulses “SPWM1, SPWM2, SPWM3, SPWM4”. Figure 3 shows the SPWM carrier signal compared to the reference sinusoidal signal for a 5-level inverter. The signals of two MOSFETS S4 and S2 are known. The signal S4 is ON for the first half cycle of the fundamental frequency and S2 is ON at the second half of it. The other signals S1, S3, S5 and S6 of the other MOSFETS will be generated by a logical combination of the two signals S2 and S4 and the comparison signals as shown in the Fig. 4 below:

Fig. 4. SPWM generator circuit

4

Simulation Results

Simulation tools were used to test the operation for the inverter circuit, the suggested invert is simulated through using MATLAB SIMULINK The reference signal and the four carrier signals are generated and comparator is used to compare these signals and generate the SPWM signals then, logical combination among this signals is made to generate the required pulses for the six MOSFETS switches As displayed in Fig. 5 the system of our five-level inverter consists of six MOSFET switches. Each switch is controlled by the appropriate pulse sequence generated by SPWM control to produce a five-level and a 50-Hz wave voltage.

Real Time Implementation of SPWM Signal Generation Technique

257

Fig. 5. Schematic diagram of proposed five-level inverter using Simulink

Fig. 6. Gating pulses generated by the model spwm

The input power of each DC power source is respectively 36 V and 18 V. The parameters used in the SPWM simulation are as follow: Output frequency (fm) 50 Hz and Switching frequency (fc) 8 KHz. Figure 6 shows gating pulses generated by the model SPWM and Fig. 7 displays the simulated output voltage of the inverter.

258

H. Chadli et al.

Fig. 7. The simulated output voltage of the inverter

Fig. 8. Total harmonics distortion THD of the proposed inverter

The over-modulation is the way to decrease the total harmonic distortion THD as it is presented in the Fig. 8 The THD of five level voltages without modulation [11] is 26.86% and with SPWM is 2.72%.

5

Experimental Results

To generate the SPWM signals we opted to use a digital controller, we used the microcontroller ATmega2560. This microcontroller is programmed to gen-

Real Time Implementation of SPWM Signal Generation Technique

259

erate four SPWM commands based on the comparison of two digital signals: a reference sine signal of frequency Fr = 50 Hz and a triangular signal of a high frequency, and this by exploiting the SPWM outputs of the microcontroller. The treatment is obtained by two functions. A main function and an interrupt function (see Fig. 9).

Fig. 9. Flow charts: (a) Flowchart of the main function, (b) Flowchart of the interrupt function

In the main function, we initialize microcontroller timers, the SPWM outputs and we fix the sample rate. The second function is the interruption function: the sampling frequency chosen is equal to f ec = 31372 Hz, in fact we activate an interruption on the T IM ER1 each T ec = 1/31372 = 31, 8 µs. Similarly, we have chosen the carrier frequency F C = 31372 Hz which is set by the T IM ER2 and the T IM ER3, the comparison of the reference sinusoidal signal is generated using the sine lookup table with the carrier signal generates the different values of the duty cycle. The lookup table contains the numbers of the pulses calculated and their duty cycle values. The number of PWM pulses required to complete the half cycle time of reference signal is shown in Eq. (2) Np =

Tr 2

T ec

(2)

And to complete half cycle (180◦ ) of reference signal in 314 pulses, each incremental step of sine signal is equal to 0,57◦ . For the positive half cycle, the two pulses SPWM1 and SPWM2 are generated respectively on pin 9 and pin 10. Then Flag gets activated, and a the SPWM3 and SPWM4 outputs are generated on pins 5 and 2 for the negative half cycle.

260

H. Chadli et al.

Fig. 10. The bloc diagram of the inverter

The hardware of the inverter is described by the bloc diagram shown in Fig. 10. The first block is the control circuit which is realized by the microcontroller ATmega2520, and it includes an isolation stage between the high power part MOSFET section and the low power part i-e microcontroller circuit by using optocouplers [11]. The second block is the power circuit which consists of six switches MOSFETS and LC filter, to convert the output signal to a sine wave with desired 50 Hz frequency. The proposed hardware of the inverter is implemented practically and tested as shown in Fig. 11.

Fig. 11. Proposed hardware: (a) inverter output signal before filtering, (b) inverter output signal after filtering

Real Time Implementation of SPWM Signal Generation Technique

261

Figure 12 shows the gating pulses generated by the microcontroller with a carrier frequency.

Fig. 12. a) PWM1 et PWM4. (b) PWM2 et PWM3 signals generated by the microcontroller

6

Conclusion

Multilevel inverters have been used in many industrial applications like HVDC, FACTS, EV, PV systems, UPS and industrial drive applications. In this article, the SPWM control method used to improve the shape of the output voltage of the proposed five-level inverter was done and implemented successfully. The algorithm for the generation of the gating signal for the power switches is implemented by microcontroller ATmega2560. The presented inverter is implemented practically, the experimental results given are much close to the simulation results, and they are very reasonable.

References 1. Kouro, S.: Recent advances and industrial applications of multilevel converters. IEEE Trans. Ind. Electron. 57(8), 2553–2580 (2010) 2. Nabae, A., Takahashi, I., Akagi, H.: A new neutral-point clamped PWM inverter. IEEE Trans. Ind. Appl. IA-17, 518–523 (1981) 3. Buticchi, G., Lorenzani, E., Franceschini, G.: A fivelevel single phase gridconnected converter for renewable distributed systems. IEEE Trans. Ind. Electron. 60(3), 906–918 (2013) 4. Lai, J.S., Peng, F.Z.: Multilevel converters-a new breed of power converters. IEEE Trans. Ind. Appl. 32, 509–517 (1996) 5. Meynard, T.A., Foch, H.: Multi-level choppers for high voltage applications. Eur. Power Electron. Drives J. 2(1), 41 (1992) 6. Hochgraf, C., Lasseter, R., Divan, D., Lipo, T.A.: Comparison of multilevel inverters for static var compensation. In: Conference on Record of IEEE IAS Annual Meeting, pp. 921–928, October 1994 7. Hammond, P.: A new approach to enhance power quality for medium voltage AC drives. IEEE Trans. Ind. Appl. 33, 202–208 (1997)

262

H. Chadli et al.

8. Baker, R.H., Bannister, L.H.: Electric power converter, U.S. 3 867 643, February 1975 9. Holtz, J.: Pulse width modulation a survey. IEEE Trans. Ind. Electron. 39, 410-420 (1992). Appl. 34(2), 374–380 (1998) 10. Choudhury, A., Member, S., Pillay, P.: Performance Comparison study of Two and Three-Level Inverter for Electric Vehicle Application, pp. 2–7 (2014) 11. Chadli, H., Jebroni, Z., Chadli, S., Tahani, A., Aziz, A.: Design and implementation of a novel five-level inverter topology. In: 2017 International Conference on Wireless Technologies, Embedded and Intelligent Systems (WITS), Fez, pp. 1–6 (2017). https://doi.org/10.1109/WITS.2017.7934641 12. Hosseini, S.H., Ahmadi, M. Ghassem Zadeh, S.: Reducing the output harmonics of Cascaded H-bridge Multilevel Inverter for Electric Vehicle Applications. In: 8th International Conference on Electrical Engineering/Electronics, Computer Telecommunications and Information Technology, Thailand, pp. 752–755, May 2011 13. Prathiba, T., Renuga, P.: Multi carrier PWM based multi level inverter for high power applications. Int. J. Comput. Appl. 1(9), 67–71 (2010)

Design of a PWM Sliding Mode Voltage Controller of a DC-DC Boost Converter in CCM at Variable Conditions Weam El Merrassi(&), Abdelouahed Abounada, and Mohamed Ramzi Laboratory of Automatic, Energy Conversion and Microelectronics (LACEM), Department of Electrical Engineering. Faculty of Science and Technology, University of Sultan Moulay Slimane, Beni Mellal, Morocco [email protected], [email protected], [email protected]

Abstract. In this paper a state space averaged modeling and control is proposed for a dc-dc boost convert. The constitutional time variation nature and the non-linearity made the control of the power electronics an arduous task. Therefore, linear control techniques cannot achieve effective control effect. Furthermore, the paper advances a SMC controller to regulate the boost converter. Moreover, to a comparison to a linear controller in variable conditions. Keywords: DC boost converter


Sliding mode control
Linear controller

1 Introduction Power electronics is ushering in a new kind of industrial revolution owing to its versatility in terms of fields of application as industrial automation, DC motor drive, energy storage, hybrid vehicles and renewable energy systems. The high-power conversion technique DC-DC Boost finds increasing necessities and power capability demands in widespread applications such as: modern computer, cellular phone, aircrafts, and maximizing energy harvest from photovoltaic systems [1, 4]. Nearly all this application entail that the converter ought to secure a high efficiency, a better power factor and a less total harmonic distortion, moreover to minimize his size and his cost. In this several applications DC-DC converters are requiring a steady output voltage, Vout in spite of variations in the source voltage, Vin, load current, iL, or in elements value of the converter circuit [5, 6]. Therefore, to afford voltage, current and frequency involved for the load, and to promise the desired dynamics, the converter must be adequately controlled. The conventional used control strategies are PWM voltage mode control, and current mode control with proportional (P), proportional integral (PI), and proportional integral derivative (PID) controller. However, the linear controller, PID family miscarried to satisfactorily perform constrained specifications under large parameters changes and load variations [7, 8]. Therefore, non-linear controller has been the purpose of many researches, for their capability to process instantly to a transient condition [9]. Among this various non-linear controller applied to DC-DC converters to solve © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 263–270, 2020. https://doi.org/10.1007/978-3-030-53187-4_30

264

W. El Merrassi et al.

their constantly changing operating points, are sliding mode controllers. These controllers are qualified as a powerful strategy, that is apt to provide a vigorous closed-loop under plant uncertainties and external perturbations. The target of this project is to design a sliding mode controller to regulate a solar DC-DC boost converter.

2 Mathematical Model of the Boost Converter

Fig. 1. Conventional boost converter

The Fig. 1 shows the circuit of boost converter is composed of main inductor, L, two semiconductors devices (Mosfet and Diode) and output capacitor, C, with the load R. They are two operation mode of boost converter: First mode: when the Mosfet is switched on causing the rising input current to flow through the inductor L, storing energy in its magnetics field. During this mode of operation as shown in Fig. 2(a) the load side is completely isolated from the source side.

Fig. 2. Operation mode of the boost converter, (a) boost converter during on state, (b) boost converter during off state

Design of a PWM Sliding Mode Voltage Controller

265

Second mode: when the Mosfet is switched off the input current flow through the diode causing the charge of the capacitor C and supplying the load. The conduction path is shown in Fig. 2b. The inductor voltage, capacitor current and the main inductor current are represented for each operating mode as follow: When the switch is on: 8 i < L dtL ðtÞ ¼ Vin ðtÞ  Ron iL ðtÞ C Vc ðtÞ ¼ VRout : dt iL ðtÞ ¼ iin ðtÞ

ð1Þ

Therefore, when the switch is off: 8 i < L dtL ðtÞ ¼ Vin ðtÞ  Vout ðtÞ  Vd ðtÞ C Vc ðtÞ ¼ iL ðtÞ  VRout : dt iL ðtÞ ¼ iin ðtÞ

ð2Þ

The physical state variables for the boost converter considered in Fig. 1, are the main inductor current iL(t) and the capacitor voltage Vc(t), accordingly the state vector x(t) is defined as follow:  X ðt Þ

i L ðt Þ Vc ðtÞ

 ð3Þ

In [11] the equivalent circuit of dc-dc converter with parasitic elements the Mosfet is modeled by a resistor Ron, while the forward voltage drop of diode D is modeled by an independent voltage source value Vd. In addition, the input vector V(t) has been defined as follow:  VðtÞ

Vin ðtÞ V d ðt Þ

 ð4Þ

In [11], the converter has to be represented in state equation in the following form: P

dxðtÞ ¼ A:XðtÞ þ B dt

ð5Þ

Where, A, B, C, D are matrix containing constant of proportionality. And P is a matrix representing the value of capacitance and inductance in the converter. Therefore, the Eqs. (1) for the state On, can be represented in the state-space equation as follow: 

L 0

    Ron 0 d iL ¼ 0 C dt Vc

0

1 R



  iL 1 þ Vc 0

0 0



Vin Vd

 ð6Þ

266

W. El Merrassi et al.



     Ron 0 L 0 1 0 Where the matrix are defined as: P ¼ ; A1 ¼ ; ; B1 ¼ 0 1 0 C 0 0 R While the system Eq. (2) for the state off, can be written in the state-space form as follow: 

L 0

       0 1 iL 0 d iL 1 ¼ þ 1 1 C dt Vc V 0 c R

1 0



Vin Vd



    0 1 1 L 0 ¼ ; A2 ¼ ; B 2 1 1 0 0 C R In order to get a linear model that is simple to study, we consider there is a signal variation for the system Eq. (5), in say ^x(t) about X(t). the equation of signal AC model is written as [11]: 

Where the matrix are defined as: P ¼

P

  d^x ¼ A^xðtÞ þ B^vðtÞ þ ðA1  A2 Þ:X ðtÞ þ ðB1  B2 Þ: V ðtÞd^ðtÞ dt

ð7Þ  1 0 small small

ð8Þ

Where the average matrix A and B are: A ¼ D:A1  ð1  DÞ:A2 B ¼ D:B1  ð1  DÞ:B2

ð9Þ

3 Slide Mode Controlling of Boost Converter This strategy of control is defined as a variable structure control. The arch aim of the regulation with the variable structure control throughout the sliding mode is to pressure the system to land up a defined surface, designated as the sliding surface specified in the state space [8, 10]. The sliding surface is built quasi-invariant owing to the smalltime delays and perturbations, and respecting the high frequency of the state trajectories of the controlled switch. The switching occurs through available feedback paths, that engender system motions locally directed regarding the sliding surface [9]. To design a SM controller for a boost converter, the first step is to develop a state space description for the aimed control variables in the system (voltage or current). The state variables are illustrated as, 3 3 2 V  bV ref out x1 6 7 d ðVref bVout Þ X ¼ 4 x2 5 ¼ 4 5 dt  R x3 Vref  bVout dt 2

ð10Þ

Where x1 , x2 , and x3 are defined as the voltage error, voltage error dynamics, and integral of voltage error respectively. Besides, bVout , Vref are the sensed output voltage and the reference voltage.

Design of a PWM Sliding Mode Voltage Controller

267

By substituting the boost converter model under continuous conduction mode, the control variable is expressed as [10], 2

xboost

3 Vref  bVout b R out 5 ¼ 4 bV RC Rþ LC ðVout  V in Þu dt Vref  bVout dt

ð11Þ

In a time differentiation, from (5) the matrix form is: 2

xboost _

0 1 6 ¼ Axboost þ Bu ¼ 4 0 1=RC 0

1

3 2 3 0 0 7 0 5xboost þ 4 bðVout  Vin Þ=LC 5u 0 0

ð12Þ

Where, u ¼ 1  u is the inverse logic of the switching function u. which defined as:  u¼

1 when S [ 0 0 when S \ 0

ð13Þ

Accordingly, S is the instantaneous state variable’s trajectory and it is expressed as: S ¼ a1 x 1 þ a2 x 2 þ a3 x 3 ¼ J T x

ð14Þ

With a1 , a2 , and a3 are the parameter control or the sliding coefficients [10].

4 Simulation Results In order to verify the respond of the boost converter via a PWM sliding mode controller. a Simulation is done using a dynamic model of the boost in MATLAB/Simulink environment commanding by a SMCV controller, compared to the linear PID controller. As shown in the Fig. 3, the parameters of the elements in simulated Boost converter are presented in Table 1.

Table 1. Parameters for the boost converter. Parameters Source voltage Vin Output voltage Vout Output power Main inductance L Capacitor C Switching frequency Load R (Kp1, Kp2) sliding parameter

Rating 200 V 400 V 1.6 kw 300 uH 2300 uF 20 kHz 100 X (0.14, 2.89)

268

W. El Merrassi et al.

Fig. 3. The proposed SMCV controller for the Boost Converter

In the first scenario, within any variations, PID and designed SMCV are used to control the boost converter, the results are shown in the Fig. 4. the desired output voltage was fixed to a constant 200 V. both PID and SMCV are simulated and their results are shown separately as follow:

Fig. 4. Results of the converter controlled by SMCV and PID: (a) Voltage output with SMCV, (b) current output with SMCV, (c) Power output with SMCV, (d) Voltage output with PID controller, (e) current output with PID controller, (f) power output with PID controller.

From the results it is clear that the response of SMCV is better than PID controller. with regards to PID controller, the output voltage is a trifle lesser than expected. besides the settling time for the SMV is reduced from 0.05 to 0.007. Therefore, it can be notice that the dynamic response of SMCV is very fast, and it has a better and effective performance. For the next scenario, Load variation is considered, the load increases at time t = 0.25 s from 100 X to 150 X. the results are given in Fig. 5

Design of a PWM Sliding Mode Voltage Controller

269

Fig. 5. Response to load variation for SMCV and PID controller: (a) Output Voltage for PID controller, (b) output current for PID controller, (c) pulse gate for PID controller, (d) Output Voltage for SMCV, (e) Output current for SMCV, (f) Pulse gate for SMV.

As can be seen from the results, when the load varies both PID and SMCV present oscillations, where the SMCV has a variation of the voltage around the operating point equal to 1 V, while the PID controller reveals an oscillation variation more than 3 V. Yet, the SMCV is clearly advantageous concerning the dynamics response and overshooting. In the upcoming scenario, the input voltage increased from 200 V to 240 V at the time t = 0.25 s. As it is exposed in the Fig. 6 the proposed SMVC controller presents a better performance against the voltage fluctuation with a static error around 1 V. While, PID controller attained 3.2 V. With regards, the actual output voltage of the PID controller is slightly lower than expected, it reached barely 397 V.

Fig. 6. Response to input variation: (a) SMCV controller, (b) PID controller

270

W. El Merrassi et al.

However, It is revealed that the SMCV controller is efficient in the control of the output voltage under the different variations in load or input voltage. The proposed controller had presented a dynamic response and a lower overshoot.

5 Conclusion DC-DC converter is nonlinear system, due to the inherent time variation and the characteristics of the switch. The design and the simulation of the slide mode control has been presented in this paper; we can conclude that the proposed controller is feasible for the converter purposes. To investigate the effectiveness of this non-linearcontroller a linear PID controller is applied to carry out a comparative analysis. Therefore, the simulation results for the SMC control has presented for the boost an excellent robustness control, likewise very fast dynamics respond, either a lower overshoot for the different variations of the input voltage; and load increases.

References 1. Chaibi, Y., Salhi, M., El-jouni, A.: Sliding mode controllers for standalone PV systems: modeling and approach of control. Int. J. Photoenergy (2019) 2. Sai Kumar, J., Tikeshwar, G.: A multi input DC-DC converter for renewable energy applications. Int. Res. J. Eng. Technol. 3(6), 380–384 (2018) 3. Josean, R.H., Jose, M., Oscaar, B., Ekaitz, Z., Unai, F.G.: Novel control algorithm for MPPT with boost converters in photovoltaic systems. Int. J. Hydrogen Energ. 42(28), 17831–17855 (2017) 4. Alzgool, M., Nouri, H.: PID controller design for novel multi-input multi-output boost converter hub. J. Electr. Eng. 2, 10–21 (2019) 5. Rathi, M., Kashmira, M., Ali, M.: Design and simulation of PID controller for power electronics converter circuit. Int. J. Innov. Emerg. Res. Eng. 3(2), 26–31 (2016) 6. Kim, I.H., Son, Y.I.: Regulation of a DC/DC boost converter under parametrics uncertainty and input voltage variation using nested reduced-order PI observers. IEEE Trans. Ind. Electron. 64(1), 552–562 (2017) 7. Ibrahim, O., Yahay, N.Z., Saad, N.: Comparative studies of PID controller turning methods on a DC-DC boost converter. In: The International Conference on Intelligent and Advanced Systems, Kuala Lumpur (2016) 8. Rangrang, W., Lei, M., Guangming, Z.: Disturbance observer-based sliding mode control for DC-DC boost converter. In: IEEE 3rd Advanced Information Technology, Electronic and Automation Control Conference, October 2018 9. Souvik, D., Mohd, S.Q., Pankaj, S.: Design of integral sliding mode control for DC-DC converters. In: International Conference of Materials processing and Characterization (2018) 10. Spiazzi, G., Mattavelli, P., Rossetto, L., Malesani, L.: Application of sliding mode control to switch-mode power supplies. J. Cir. Syst. Comput. 5(3), 337–354 (1995) 11. Mattavelli, P., Rossetto, L., Spiazzi, G.: Small signal analysis of DC-DC converters with sliding mode control. IEEE Trans. Power Electr. 1, 153–159 (1995)

Design and Performance Analysis of Super-Twisting Algorithm Control for Direct-Drive PMSG Wind Turbine Feeding a Water Pumping System Benzaouia Soufyane1,2(&), Zouggar Smail1, Rabhi Abdelhamid2, and Mohammed Larbi Elhafyani1 Laboratory of Electrical Engineering and Maintenance – LEEM, High School of Technology, University Mohammed 1st, Oujda, Morocco [email protected] 2 Laboratory of Modelisation, Information and Systems – MIS, University of Picardie Jules Verne, 33 rue Saint Leu, 80039 Amiens, Cedex, France 1

Abstract. This paper deals with the nonlinear control of a direct-drive PMSG wind turbines using the super-twisting algorithm. The studied system is assumed supplying a water pumping system for the use in isolated sites and areas. The aim of the proposed control strategy is tracking the wind turbine maximum power point. The designed controllers are based on one of the high-order sliding mode controller (HOSM) versions, which is the super-twisting algorithm. This latter possess many attractive features as the chattering-free behavior, finite time convergence, less information demand, simplicity, stability and robustness against external disturbances. The performance of the whole system in closedloop mode is assessed through computer simulations. Keywords: Wind energy
Water pumping system
Super-twisting algorithm
Nonlinear systems
PMSG
MPPT
Sliding mode controller

1 Introduction In recent decades, wind energy has received a lot of attention as one of the clean alternative energy sources [1]. In general, renewable energies aim to reduce the negative impact of conventional electricity sources on the environment and also supplying remote areas and isolated sites where access to the classical energy is difficult. Water pumping is one of the principal renewable energy application; photovoltaic energy is the most used and preferred for water supply. In some cases and for some locations, solar energy cannot be the best solution. Wind energy has recently been adopted as a solution for regions with good wind potential. In literature, many wind electric water pumping configurations and control strategies have been proposed and studied [2–4]. The main goal of the existed studies is ensuring an optimal and a maximum efficiency of the water pumping system operation. The studied configuration in this paper is shown in Fig. 1. It consists of permanent-magnet synchronous generator, a controlled © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 271–281, 2020. https://doi.org/10.1007/978-3-030-53187-4_31

272

B. Soufyane et al.

AC/DC converter connected to a permanent-magnet DC motor driving a centrifugal pump (nonlinear load). The permanent-magnet synchronous generator (PMSG) is one of the most preferred choice for standalone systems due to its high efficiency, self-excitation features, reliability and also for allowing a direct-drive systems avoiding by that the use of a gearbox [5]. The major drawbacks of wind energy conversion systems (WECS)
is the highly nonlinear behavior [1]. Figure 2.b represent the power coefficient Cp as function of the pitch angle ðbÞ and the specific speed (k), and Fig. 2.a shows the mechanical power as a function of rotor speed of the turbine for different values of wind speed. Maximizing the captured wind energy power is the main objective, many control strategies can be found in literature as well as controllers improvements in order to overcome the drawbacks of the conventional controllers. In [1], an adaptive fuzzy-PI control is considered to replace the conventional constant gains PI controller for PMSG vector control. A sliding mode control strategy is proposed in [6] and [7] for a PMSG controlled by vector control. In [8] a hybrid fuzzy sliding mode controller is proposed for controlling the permanent magnet synchronous generator speed. In [9], a general regression neural network (GRNN) controller is proposed for induction generator (IG) speed drive. A fuzzy neural network controller is proposed in [10] for the same generator kind. In [11], a hybrid intelligent PMSG control based on sliding mode controller combined with fuzzy inference mechanism and adaptive algorithm is proposed. This paper presents the control of a direct-drive permanent-magnet synchronous generator wind turbine used in an autonomous water pumping system. The proposed optimal control of the PMSG wind turbine is based on super-twisting algorithm. The overall control functions of the wind electric water pumping system are developed including the maximization of the captured power. The presented control strategy allows an efficient operation of the system in a wide range of winds and aims to make the turbine operating on the curve corresponding to the maximum power point. Computer simulations are presented in order to validate and evaluated the performance of the adopted control strategy on the studied system.

Fig. 1. Studied Wind Electric Water Pumping System.

Design and Performance Analysis of Super-Twisting Algorithm Control

273

Fig. 2. a. Wind generator power curves at various wind speed, b. Characteristics Cp vs k for different values of the pitch angle b.

2 Modeling of Wind Electric Water Pumping System 2.1

Modeling of the Wind Turbine

The model of the turbine is modeled from the following system equations [1, 5, 7]:


q:A:v3w pv ¼ 2



1 pm ¼ Cp ðk; bÞ:q:A:v3w 2 k¼

XG mR vw

ð1Þ ð2Þ

ð3Þ

8   C < Cp ðk; bÞ ¼ C1 C2  C3 :b  C4 e c 5 þ C6 :k c

:

1 c

1 ¼ k þ 0:08b  b0:035 3 þ1

ð4Þ

C1 ¼ 0:5176; C2 ¼ 116; C3 ¼ 0:4; C4 ¼ 5; C5 ¼ 21 TmG ¼ (

pm 1 3 ¼ G Cp ðk; bÞ:q:A:vw XG 2:X m m

G G TmG ¼ J G X_ m þ f :XG m þ Tem G G G J ¼ Jturbine þ Jg

ð5Þ

ð6Þ

Where pv is the wind power, q is the air density, A is the circular area, vw is the wind speed, pm is the mechanical power, Cp is the power coefficient, b is the pitch angle, k is G the tip speed ratio, XG m is the turbine rotor speed, R is the turbine radius, Tm is the G mechanical torque, Tem is the electromagnetic torque produced by the generator, f is the friction coefficient and J G is the total moment of inertia of the rotating parts.

274

2.2

B. Soufyane et al.

Permanent Magnet Synchronous Generator (PMSG) Model

The d-q stator voltage equations of this generator are given by the following expressions [12, 18, 19]: 8 Vds ¼ Rs Ids þ Ld I_ds  xr wqs > > < Vqs ¼ Rs Iqs þ Lq I_qs þ xr wds > wds ¼ Ld Ids þ w0 > : wqs ¼ Lq Iqs

ð7Þ

The differential equations of the PMSG can be obtained as follow: 

Ld I_ds ¼ Vds  Rs Ids þ xr Lq Iqs Lq I_qs ¼ Vqs  Rs Iqs  xr Ld Ids  w0 xr

ð8Þ

The electromagnetic torque is represented by:   3 
G ¼ p Ld  Lq Ids Iqs þ w0 Iqs Tem 2

ð9Þ

The PMSG is assumed to be wound-rotor, then Ld ¼ Lq , and the expression of the electromagnetic torque in the rotor can be described as follow: 3 G Tem ¼ pw0 Iqs 2

ð10Þ

Where Ld , Lq are the inductances of the generator on the q and d axis, Rs is the stator resistance, w0 is the permanent magnetic flux, xr is the electrical rotating speed of the PMSG which is given by xr ¼ p:XG m and p is the number of pole pairs. 2.3

Permanent-Magnet DC Motor (PMDC) and Centrifugal Pump Model

The model of the PMDC motor is represented by the following equations [3]: 8 > > >
m > > : T M  T M ¼ J M X_ M þ B :XM þ T M m m m e L f

ð11Þ

Design and Performance Analysis of Super-Twisting Algorithm Control

275

The load torque of the centrifugal pump is given by the following expression: n

TLM ¼ a:XM m þb

ð12Þ

Where RaM is the armature winding resistance, LaM is the armature self-inductance, IaM is the motor armature current, VaM is the applied voltage, e is the back e.m.f of the PMDC motor, Ke is the voltage constant, XM m is the angular speed, Kt is the torque constant, J M is the moment of inertia, Bm is the viscous torque constant, TfM is the torque constant for rotational losses, TeM , TLM are the electromagnetic torque and load torque respectively, and a, b are the constants of the pump.

3 PMSG Side Converter Control In order to control the generator speed, a vector control strategy is applied to the AC/DC converter (Fig. 3). The generator speed XG m can be controlled by adjusting the G G electromagnetic torque (Tem ) to its reference (Tem Þ. That can be done by acting on the  2 G ¼ 3pw Tem . The d-axis stator current (Ids ) q-axis current (Iqs ) using the equation Iqs 0 component is forced to zero to achieve the maximum torque of the generator [12, 18]. k :vw where k The optimal reference rotational speed is calculated using XG mopt ¼ R represents the optimum tip-speed ratio. • Super Twisting Algorithm (STA) The major known drawback associated with variable structure control implementation is the chattering phenomenon [13]. The most used technique to avoid this problem is the approach known as high-order sliding mode (HOSM). This latter is very well known in their stability and robustness against external disturbances and uncertainty. The increasing information demand is the main problem of the high-order sliding mode controllers; the implementation of the rth-order controller requires the knowledge of _ r €; . . .; rðr1Þ (r is the sliding surface). The super-twisting algorithm is the r; r; exception, it has two main advantages. The first one is that the ST algorithm can be applied to any system having a relative degree equal to 1 with respect to sliding variable, and the second and the important advantage is that the ST algorithm does not require any information on the time derivative of the sliding variable and maintains all the distinctive robust features of the SMC [14, 15]. • Outer STA Generator Speed Controller The sliding surface for the STA speed controller is given as follows: G rXGm ¼ XG mopt  Xm

ð13Þ

276

B. Soufyane et al.

It follows that (

G G G T f :X T r_ XGm ¼ X_ mopt  X_ m ¼ X_ mopt  m J Gm em G €XGm ¼ .1 ðt; xÞ þ  1 ðt; xÞT_ em r G

G

G

ð14Þ

Where .1 ðt; xÞ and  1 ðt; xÞ are uncertain bounded functions that satisfy .1 ðt; xÞ [ 0; j.1 ðt; xÞj [ U1 ; 0\Cm1 \ 1 \CM1

ð15Þ

The proposed second-order sliding mode control has been designed using the super twisting algorithm. The control law contains two parts, one is the continuous function of the sliding surface (rXGm Þ and, the other, is the integral of a discontinuous control action [20]: 

G Tem ¼ t1 þ t2

ð16Þ

Where:   t_ 1 ¼ a1 sign rXGm q   : t2 ¼ b r G sign r G 1 Xm Xm 8
< > :

U1 Cm1 4U1 CM1 ða1 þ U1 Þ C2m1 Cm1 ða1 U1 Þ

a1 [

b21  0\q  0:5

ð19Þ

• Inner STA Current Controllers   In order to regulate currents components Iqs and Ids to their references (Ids and Iqs ), the sliding surfaces were chosen as follow:



  Ids rIds ¼ Ids  rIqs ¼ Iqs  Iqs

ð20Þ

Design and Performance Analysis of Super-Twisting Algorithm Control

277

It follows that  €Ids r

 r_ Ids ¼ I_ds  I_ds  and ¼ .2 ðt; xÞ þ  2 ðt; xÞV_ ds

( €Iqs r

  I_qs r_ Iqs ¼ I_ds  ¼ .3 ðt; xÞ þ  3 ðt; xÞV_ qs

ð21Þ

Where .2 ðt; xÞ, .3 ðt; xÞ,  2 ðt; xÞ and  3 ðt; xÞ are uncertain bounded functions that satisfy 

.2 ðt; xÞ [ 0; j.2 ðt; xÞj [ U2 ; 0\Cm2 \ 2 \CM2 .3 ðt; xÞ [ 0; j.3 ðt; xÞj [ U3 ; 0\Cm3 \ 3 \CM3

ð22Þ

The proposed second-order sliding mode control has been designed using the super twisting algorithm. The control voltages of q and d axis are defined as follow: 

 ¼ l1 þ l2  Femd Vds  Vqs ¼ w1 þ w2 þ Femq

ð23Þ

Where:
 8 rIqs l_ 1 ¼ a2 signðrIds Þ 3 sign < w_ 1 ¼ a
 q l2 ¼ b2 jrIds jq signðrIds Þ and w2 ¼ b3 rIqs sign rIqs : : Femd ¼ pXG Femq ¼ pXG m Lq Iqs m ðLd Ids þ w0 Þ 8
ai [ CUmii < 2 i CMi ðai þ Ui Þ ; ði ¼ 2; 3Þ: ð26Þ bi  4U C2mi Cmi ðai Ui Þ > : 0\q  0:5

278

B. Soufyane et al.

Fig. 3. Applied super-twisting algorithm based MPPT control

4 Results and Discussion The following results were obtained for the studied wind electric water-pumping system described in Fig. 1 using Matlab/Simulink software. The control strategy was performed under variable wind speed profile. The wind speed was modeled as a sum of deterministic several harmonics [17]. vw ðtÞ ¼ 7 þ 0:2 sinð0:1047tÞ þ 2 sinð0:2665tÞ þ sinð1:2930tÞ þ 0:2 sinð3:6645tÞ ð27Þ

Fig. 4. Applied wind speed profile.

Fig. 6. d-q axis stator currents components.

Fig. 5. Variation of the turbine mechanical power and the electrical absorbed power by the motor-pump.

Fig. 7. Zoom of d-q axis stator currents components.

Design and Performance Analysis of Super-Twisting Algorithm Control

279

Fig. 8. Power coefficient and its optimal reference

Fig. 9. abc axis current evolution.

Fig. 10. Actual rotor speed and its reference (optimal).

Fig. 11. Motor-pump performances.

The applied wind speed profile is shown in Fig. 4. The used waveforms contains two regions, the high wind speed range and the low one. Figure 5 illustrate the variation of the turbine mechanical power and the electrical absorbed power by the motorpump group. The obtained results shows a good tracking performance of the maximum mechanical power. The dynamic difference between the mechanical turbine power and the electrical absorbed power is due to system inertia, friction and losses at converter
 and generator level. It is observed also on Fig. 8 that the power coefficient Cp has been kept at its maximum value 0:48 even under random wind speed profile. The PMSG vector control can be verified by observing the d and q current axis.  Þ generated by Figures 6 and 7 show a good pursuit of the reference q current axis ðIqs the outer super twisting algorithm controller, similarly to the d current component ðIds Þ, it can been seen on the same figures that the Ids remains around its reference (zero). Figure 9 shows that generator currents are sinusoidal, no harmonics or perturbations are observed at generator level, the thing that will increase the system efficiency. Figure 10 shows a good tracking performance of the speed rotor to the reference (optimal) speed with a very small error. Figure 11 illustrates the motor-pump performances; it shows the evolution of the permanent-magnet DC motor speed, the electrical torque produced by the dc motor and the load torque opposed by the centrifugal pump.

280

B. Soufyane et al.

5 Conclusion In this paper, a wind electric water pumping system based on direct-drive PMSG wind turbine and PMDC motor connected to a centrifugal pump is studied and presented. A vector control strategy based on super-twisting algorithm controllers has been designed and evaluated under varying wind conditions. The obtained results proved the effectiveness and the robustness of the proposed control strategy on the studied system. The selected super-twisting algorithm has shown a good stability without chattering effects. The main advantage of the presented control strategy is that the ST algorithm requires only knowledge of the sign of the sliding variable, which means an easier implementation and good performances with less information demand. The proposed control strategy can be improved by replacing the mechanical sensors with observers state in order to reduce the overall system cost. Acknowledgment. This research was supported by the National Center for Scientific and Technical Research (CNRST) of Morocco and the Embassy of France in Morocco.

References 1. Aissaoui, A.G., et al.: A fuzzy-PI control to extract an optimal power from wind turbine. Energy Convers. Manag. 65, 688–696 (2013) 2. Ouchbel, T., et al.: Power maximization of an asynchronous wind turbine with a variable speed feeding a centrifugal pump. Energy Convers. Manag. 78, 976–984 (2014) 3. Soufyane, B., et al.: A comparative investigation and evaluation of maximum power point tracking algorithms applied to wind electric water pumping system. In: International Conference on Electronic Engineering and Renewable Energy. Springer, Singapore (2018) 4. Lara, D., Merino, G., Salazar, L.: Power converter with maximum power point tracking MPPT for small wind-electric pumping systems. Energy Convers. Manag. 97, 53–62 (2015) 5. Abdullah, M.A., et al.: A review of maximum power point tracking algorithms for wind energy systems. Renew. Sustain. Energy Rev. 16(5), 3220–3227 (2012) 6. Emna, M.E., Adel, K., Mimouni, M.F.: The wind energy conversion system using PMSG controlled by vector control and SMC strategies. Int. J. Renew. Energy Res. 3(1), 41–50 (2013) 7. Errami, Y., Ouassaid, M., Maaroufi, M.: A performance comparison of a nonlinear and a linear control for grid connected PMSG wind energy conversion system. Int. J. Electr. Power Energy Syst. 68, 180–194 (2015) 8. Chen, C.H., Hong, C.-M., Ou, T.-C.: Hybrid fuzzy control of wind turbine generator by pitch control using RNN. Int. J. Ambient Energy 33(2), 56–64 (2012) 9. Hong, C.-M., Cheng, F.-S., Chen, C.-H.: Optimal control for variable-speed wind generation systems using general regression neural network. Int. J. Electr. Power Energy Syst. 60, 14– 23 (2014) 10. Lin, W.-M., Hong, C.-M., Cheng, F.-S.: Fuzzy neural network output maximization control for sensorless wind energy conversion system. Energy 35(2), 592–601 (2010) 11. Lin, W.-M., et al.: Hybrid intelligent control of PMSG wind generation system using pitch angle control with RBFN. Energy Convers. Manag. 52(2), 1244–1251 (2011) 12. Dahbi, A., et al.: Realization and control of a wind turbine connected to the grid by using PMSG. Energy Convers. Manag. 84, 346–353 (2014)

Design and Performance Analysis of Super-Twisting Algorithm Control

281

13. Valenciaga, F., Puleston, P.F.: High-order sliding control for a wind energy conversion system based on a permanent magnet synchronous generator. IEEE Trans. Energy Convers. 23(3), 860–867 (2008) 14. Matraji, I., Al-Durra, A., Errouissi, R.: Design and experimental validation of enhanced adaptive second-order SMC for PMSG-based wind energy conversion system. Int. J. Electr. Power Energy Syst. 103, 21–30 (2018) 15. Kunusch, C., et al.: Sliding mode strategy for PEM fuel cells stacks breathing control using a super-twisting algorithm. IEEE Trans. Control Syst. Technol. 17(1), 167–174 (2008) 16. Mekri, F., Elghali, S.B., El Hachemi Benbouzid, M.: Fault-tolerant control performance comparison of three-and five-phase PMSG for marine current turbine applications. IEEE Trans. Sustain. Energy 4(2), 425–433 (2012) 17. Tran, D.-H., et al.: Integrated optimal design of a passive wind turbine system: an experimental validation. IEEE Trans. Sustain. Energy 1(1), 48–56 (2010) 18. Marmouh, S., Boutoubat, M., Mokrani, L.: Performance and power quality improvement based on DC-bus voltage regulation of a stand-alone hybrid energy system. Electr. Power Syst. Res. 163, 73–84 (2018) 19. Mousa, H.H.H., Youssef, A.-R., Mohamed, E.E.M.: Variable step size P&O MPPT algorithm for optimal power extraction of multi-phase PMSG based wind generation system. Int. J. Electr. Power Energy Syst. 108, 218–231 (2019) 20. Kunusch, C., Puleston, P.F., Mayosky, M.A., Dávila, A.: Efficiency optimisation of an experimental PEM fuel cell system via super twisting control. In: 2010 11th International Workshop on Variable Structure Systems (VSS), pp. 319–324. IEEE, June 2010

Electric System Cascade Analysis for Optimal Sizing of an Autonomous Photovoltaic Water Pumping System Mohammed Chennaif(&), Mohamed Larbi Elhafyani, Hassan Zahboune, and Smail Zouggar Laboratory of Electrical Engineering and Maintenance – LEEM, High School of Technology, University Mohammed 1st, Oujda, Morocco [email protected]

Abstract. Covering the necessary power for the stand-alone work of water pumping systems for irrigation using photovoltaic energy is the primary objective of this paper. To achieve this goal, we have adopted the Power Pinch Analysis as a guideline, and the method is based on the Electric System Cascade Analysis for the sizing of the water supply system, which necessarily contains photovoltaic panels and a storage tank. The sizing procedure for obtaining an optimal design of the system technically and economically is preceded by a modeling of system components, solar radiation model and optimal tilt angles for each month to obtain the largest amount of solar radiation. Then, the procedure start from a developed algorithm with multiple inputs, hourly solar radiation, hourly water demand for irrigation, total dynamic head, as well as cost data. The results prove the capability and accuracy of the method in optimally sizing stand-alone photovoltaic water pumping systems based on load profil and climate resources of the worst month of the year. Keywords: ESCA
Power Pinch Analysis Optimization system


PV pumping system


1 Introduction There is still a growing interest in solar-based applications, due to being a renewable, clean and abundant source of energy reaches most land areas, making it the ideal solution for powering isolated areas in energy. The applications of solar energy vary according to uses and needs, photovoltaic pumping systems are one of its most important applications, especially for the population of isolated areas, where the high water demand for crop irrigation. The various interesting studies have been reported in the literature, aiming at the techno-economic improvement of PV water pumping systems (PVWPS). Chandel et al. notes in [1] that the increasing fuel prices and the deficit in electricity are the main causes of the appeal to the photovoltaic pumping. Bakelli et al. present in their work a model to optimize the capacity sizes of different components of the photovoltaic water pumping system with water tank storage. Two optimization criteria, the loss of power supply probability (LPSP) concept and the life cycle cost (LCC) are taken into account © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 282–290, 2020. https://doi.org/10.1007/978-3-030-53187-4_32

Electric System Cascade Analysis for Optimal Sizing

283

in the optimization procedure, the location is considered in Ghardaia, Algeria [2]. An hourly simulation model is the basis of the optimization method in the study carried out by Campana et al. [3]. The concept of power pinch saw the light for the first time by Bandyopadhyay [4]. The ESCA method is applied in order to obtain the optimal design of autonomous hybrid systems by Hassan et al. [5, 6], and to optimize an isolated photo-voltaic generating unit with battery storage system by Singh et al. [7]. Through this work, we aim to provide a simple and effective methodology on which we can answer the problem of sizing of photovoltaic systems under different constraints, which satisfies all water demand during the study period.

2 The Standalone Photovoltaic Water Pumping System The standalone PV pumping systems operate on the basis of converting primarily the solar energy into electrical by the photovoltaic panels. The electrical energy is then transformed to mechanical energy by the driving AC motor, the movement of fluid is started by the aid of the pump turbine, and the hydraulic energy is created in order to supply water requirement with the water pumped from a deep. Water pumped using the solar system is used directly to satisfy needs, while the excess is stored in the storage tank for use where there is a deficit or lack of supply. The water storage tank in the PVWPS plays the same role of batteries in the electrical hybrid systems and the electric power load demand is replaced by water demand. 2.1

Modelling of System Components

PV Model: The efficiency of each PV panel varies according to its type, as well as a range of variables such as temperature. PV panels supply DC by converting solar energy to electrical energy. The hourly energy produced by a single PV panel during current time step can be expressed by: EPV ðtÞ ¼ gPV :APV :IT ðTÞ

ð1Þ

Where IT(t) is the global solar radiation (Wh/m2) that reaches the collector surface, APV is the PV panel receiving surface (m2) and gPV is the efficiency of PV panel (%). The total generated power by NPV of photovoltaic panels during the current time step is given by: EPV;T ðtÞ ¼ NPV :EPV ðtÞ

ð2Þ

The total Amount of water pumped by the photovoltaic pumping system during the current time step, it is given by: QPV;T ðtÞ ¼

EPV;T ðtÞ:gDCnAC  gP  gM C*TDH

ð3Þ

284

M. Chennaif et al.

Where EPV, T(t) is the generated power by NPV panels; gDC\AC is the inverter efficiency; TDH is the Total Dynamic Head. Load Profil for Irrigation: Water requirements for irrigation depend on the type of crop, weather factors such as temperature, humidity, wind velocity, soil evapotranspiration, the season of the year and the method of irrigation. However, it is important to rely on local practice and experience. The amount of water demand for irrigation in each time step is expressed in our work as QL(t). The net amount of water is the difference between the amount of water pumped by the PVWPS and the load demand QL(t) during the current time step: Qr ðtÞ ¼ QPV;T ðtÞ  QL ðtÞ

ð4Þ

Water Storage Tank Model: The water pumped by the PVWPS is used to satisfy the load demand while the excess is used to charge the storage tank, until the latter is fully charged, and when the amount of water is insufficient, the stored water can be discharged. And so on, the storage tank allows us to store water when there is an excess and covers deficits periods. Then, the net amount of water Qr(t) calculated from (4), controls the quantity of charging (CST(t)) and discharging (DST(t)) of the storage tank.
CST ðtÞ ¼
DST ðtÞ ¼

Qr ðtÞ:gch ; Qr ðtÞ  0 0; Qr ðtÞ\0

ð5Þ

0; Q r ðt Þ  0 Qr ðtÞ=gdisc ; Qr ðtÞ\0

ð6Þ

The net water surplus/deficit QN(t): QN ðtÞ ¼ CST ðtÞ þ DST ðtÞ

ð7Þ

The accumulation of water quantity in the storage tank in the current time interval in m3 is calculated by: Qacc ðtÞ ¼ Qacc ðt  1Þ þ QN ðtÞ

ð8Þ

Qacc(t − 1) is the net accumulated water quantity in the storage tank in the previous time interval in m3. 2.2

Optimization Constraints

In this paper, the maximum allowable Life Cycle Cost (LCC) and the Loss of Water Supply Probability (LWSP) are implemented together in the Electrical System Cascade Analysis method, to optimize the sizing procedure of the system.

Electric System Cascade Analysis for Optimal Sizing

285

The Life Cycle Cost is an estimated measure of the total cost accumulated of the photovoltaic water pumping system during his life cycle. Four principal parts are considered: PV generator, the motor-pump set, the water tank and the inverter.       LCC ¼NPV : CPV þ CM;PV þ CI;PV þ CINV þ CR;INV þ CST þ CI;ST þ ðCMP Þ þ CM;MP þ CI;MP þ CR;MP Þ

ð9Þ

Where, CPV is the cost of a single PV panel, CM,PV is the maintenance cost of PVP, CI, PV is the cost of installation of PV, CINV is the cost of the inverter, CR,INV is the cost of replacement of the inverter, CST is the cost of the storage tank and CI,ST is the cost of installation of the ST. CMP is the cost of the Motor-Pump set, CI,M−P is the cost of installation of the M − P, CM,M−P is the cost of maintenance of the pumping set and CR,M−P is the cost of replacement of the M − P. Loss of Water Supply Probability (LWSP): Stands for the probability of insufficient water when the PVWPS is unable to meet the load demand. Value 1 of LWSP means that load will always be satisfied and the value 0 means that the load will never be satisfied. The LWSP is defined as follows: PT LWSP ¼ PTt¼0 t¼0

QdðtÞ QLðtÞ

ð10Þ

Where Qd(t) is the deficit of water supply during the time step, when the pumped water is not satisfied to meet the water demands, it is given by: Qd ðtÞ ¼ QL ðtÞ  ðQPV ðtÞ þ Qacc ðt  1Þ

ð11Þ

3 Methodology Generally, the sudden water variations represent the big obstacle during the procedure of sizing, for that, The ESCA method used to find the optimal sizing of the system. 3.1

Storage Cascade Table

The forming of the cascade table is a necessary step in performing the power pinch analysis. The construction of the table is done by following the steps below: • • • • • • •

Column 1. Column 2. Column 3. Column 4. Column 5. Column 6. Column 7. using (5).

The The The The The The The

time period arranged in ascending order (1 h). hourly total solar radiation in Wh/m2, denoted by I(t). hourly energy generated by the PV generating units in Wh (2). hourly pumped water by the PVWPS in m3 denoted by QPV,T(t). hourly water demand in m3 denoted by QL(t). net amount of water in m3, calculated using (4). amount of water charging in the storage tank CST(t). If Qr (t) > 0,

286

M. Chennaif et al.

• Column 8. It represents the discharging status of the storage tank DST(t) if Qr (t) < 0. • Column 9. The net water surplus/deficit QN (t) (7). • Column 10. Net accumulated water or the current status of the tank is calculated using (8). During initial analysis Q0(t = 0) is considered zero. • Column 11. Net accumulated water or the current status of the tank with the initial value Q0(t = 0) changing to the minimum value of Qacc (t) in row 11. 3.2

ESCA Algorithm

The algorithm developed checks the ESCA method for the worst month of the year (T = 744 h) (Fig. 1).

Start Data extracƟon and Ɵme step; Models parametrs; Solar radiaƟon and temperature; Profil of water demand; IniƟal number of PVP and iniƟal amount of water in the storage tank. t=0 t=t+1 Calculate QPV,T(t) Calculate NO

Qr(t) = QPV,T(t) - QL(t) YES

Qr(t) > 0 DST(t)

CST(t) Calculate Qacc(t) NO

t=T

Q0 = -Min Qacc

YES Min Qacc

Calculate NPV,New

> < VrVin EW ðtÞ ¼ Pr > > : 0

Vin \V ðtÞ\Vr Vr \V ðtÞ\Vout V ðtÞ\Vin or V ðtÞ [ Vout

ð1Þ

Techno-Economic Sizing of a Stand-Alone Hybrid Energy and Storage

293

Where Pr is the rated power of the wind turbine (W), Vin is the Cut-in wind speed of the wind turbine (m/s), Vout is the Cut-out of the wind turbine (m/s) and Vr is the rated speed of the wind turbine (m/s). The total generated power by NW wind turbines in the current time step: EW;T ðtÞ ¼ NW :EW ðtÞ

ð2Þ

The total Amount of water pumped by NW wind turbines during the current time step is given by: QW;T ðtÞ ¼

EW;T ðtÞ:gDCnAC  gP  gM C*TDHðtÞ

ð3Þ

Where EW, T(t) is the generated power by the wind turbines; gDCnAC is the inverter efficiency; TDH(t) is the Total Dynamic Head; g P and gM are the pump and motor efficiencies, respectively, and C is the hydraulic constant. The total pumping height is PV Model The hourly energy produced by a single PV panel during current time step can be expressed by: EPV ðtÞ ¼ gPV ðtÞ:APV :RG ðTÞ

ð4Þ

Where g PV is the efficiency of PV panel (%), APV is the PV panel receiving surface (m2) and RG(t) is the global solar radiation (Wh/m2). The total Amount of water pumped QPV,T(t) by NPV panels during the current time step [7] is given by: QPV;T ðtÞ ¼

EPV;T ðtÞ:gDCnAC  gP  gM C*TDHðtÞ

ð5Þ

Where EPV, T (t) is the generated power by NPV panels. The total Amount of water pumped by the hybrid PV/Wind pumping system during the current time step: QT ðtÞ ¼ QPV;T ðtÞ þ QW;T ðtÞ

ð6Þ

The net amount of water is the difference between the amount of water pumped by the hybrid PV/Wind system and the load demand for irrigation during the current time step: Qr ðtÞ ¼ QT ðtÞ  QL ðtÞ QL (t): is the water demand for irrigation in each time step.

ð7Þ

294

M. Chennaif et al.

Storage unit model The water pumped by the PV/Wind hybrid pumping system will be stored in the water storage tank if the load is satisfied. In the case where the pumped water is not sufficient to satisfy the load, the water will be discharged from the tank to the load. The net amount of water Qr (t) decides whether storage tank is in charging CST(t) or discharging DST(t) state.  Qr ðtÞ [ 0 :

CST ðtÞ ¼ Qr ðtÞ:gch DST ðtÞ ¼ 0

ð8Þ

nch is the charge efficiency. When Qr (t) is negative, the storage tank is discharging:  Qr ðtÞ\0 :

CST ðtÞ ¼ 0 DST ðtÞ ¼ Qr ðtÞ=gdisc

ð9Þ

ndisc is the discharge efficiency is neglected because the tank is higher. The accumulation of water in the storage tank in the current time interval in m3 is calculated by: Qacc ðtÞ ¼ Qacc ðt  1Þ þ CST ðtÞ þ DST ðtÞ

ð10Þ

Qacc(t − 1) is the net accumulated water quantity in the storage tank in the previous time interval in m3. 2.3

Techno-Economic Optimization

The maximum allowable Loss of Water Supply and the Life Cycle Cost are implemented together to optimize the sizing procedure of the system. Loss of Water Supply Probability (LWSP) The LWSP represents the percentage of deficit for a period. A value of 1 of LWSP represents 100% of deficit and a value of 0 represents 0% of deficit: PT QdðtÞ LWSP ¼ PTt¼0 ð11Þ t¼0 QLðtÞ Where Qd (t) is the deficit of water supply during the time step, it is given by: Qd ðtÞ ¼ QL ðtÞ  ðQPV ðtÞ þ Qacc ðt  1Þ

ð12Þ

Life Cycle Cost (LCC) The Life Cycle Cost is an estimated measure of the total cost accumulated of the hybrid PV/Wind pumping system during his life cycle.

Techno-Economic Sizing of a Stand-Alone Hybrid Energy and Storage

    LCC ¼ NPV : CPV þ CM;PV þ CI;PV þ NW : CW þ CM;W þ CI;W     þ CINV þ CR;INV þ CST þ CI;ST þ ðCMP Þ þ CM;MP þ CI;MP þ CR;MP Þ

295

ð13Þ

Where, CW is the cost of a single wind turbine, CM,W is the maintenance cost of WT, CI,W is the cost of installation of WT, CPV is the cost of a single PV panel, CM,PV is the maintenance cost of PVP, CI,PV is the cost of installation of PV, CINV is the cost of the inverter, CR,INV is the cost of replacement of the inverter, CST is the cost of the storage tank and CI,ST is the cost of installation of the ST. CM-P is the cost of the Motor-Pump set, CI,M-P is the cost of installation of the M-P, CM,M-P is the cost of maintenance of the pumping set and CR,M-P is the cost of replacement of the motor-pump set.

Start Data extraction and time step; Initial number of PVP, WT and initial amount of water in the storage tank. t=0 t=t+1 Calculate QPV,T(t) and QW,T(t) Calculate NO

Qr(t) = QT(t) - QL(t) YES

Qr(t) > 0

DST(t)

CST(t) Calculate Qacc(t) NO

t=T

Q0 = -Min Qacc

YES if

Qacc < 0 NO

YES

Qacc(T) < 0 YES Calculate all possible solutions (NPV,New & NW,New) Calculate QMax = Max Qacc(t) Optimal solution NPV,New ; NW,New ; VST End

Fig. 2. Flowchart of the developed algorithm.

Economic Analysis (LCC)

296

2.4

M. Chennaif et al.

Methodology for Time Period of 744 h

The algorithm developed of the methodology is shown in the Fig. 2. The steps of the methodology are as follows: • Step 1. Data extraction and initial estimation of number of PVP and WT. • Step 2. Check if there is no negative value of Qacc during all the time period of analysis, if not change the initial amount of water in the Storage Tank (Q0 = −min (Qacc)). • Step 3. Check if constraint (Qacc(T) > 0) is satisfied, If not fix NPV and change NW until Qacc(T) becomes higher a little to 0. • Step 4. Each time increment the fixed value of NPV to get all possible combinations (NPV&NW). • Step 3. Calculate the size of the storage tank after constraint is satisfied. • Step 5. Economic analysis based on the LCC.

3 Case Study The previous method is used in this case study to determine the optimal sizing of a hybrid pumping system PV/Wind. The sizing procedure of the system is based on the climatic data of an irrigated area located in the city of Oujda Morocco (latitude: 34° 41′; 1ongitude 1° 54′). The required water of this site varies considerably with the period of the year and the time of day. Solar irradiation, ambient temperature and demand are assumed to be constant at each time step (1 h). The total head is 50 m, and the storage tank height is 5 m, i.e. the total head until the storage tank is 55. The implementation and the programming of the algorithm of the proposed method is done in the MATLAB/SIMULINK environment. The determination of the sizes of the system components is based on the worst month of the year (October in this case study). All data required for the system are presented in Table 1.

Table 1. Required data of the system PV panel Type APV Efficiency Life time CPV CM-PV CI-PV Peak Power of a single panel

Mono-crystalline 1.64 m2 15% 25 y 250 € 10 € 15 € 250 Wp

Storage tank Wind turbine 1.5 KW PR Vin 2.5 m/s CST 3.59 €/m3 Vout 13 m/s Vr 11 m/s LST 25 y Life time 25 y CWT 1300 CM-PV 80 € ηW 98%

Techno-Economic Sizing of a Stand-Alone Hybrid Energy and Storage

297

4 Results and Discussion Table 2. Résults of the method. NPV 0 3 6 9 12 15 18 21 25 28

NW 9 8 7 6 5 4 3 2 1 0

VST (m3) 623 567 510 450 385 317 246 172 239 266

Qacc (t = 0) (m3) 224 172 126 82 43 26 26 25 26 26

LCC (€) 13936 13260 12580 11890 11182 10463 9733 8992 9033 9284

The Table 2 above shows all the possible combinations that are valid as solutions for the sizing of the hybrid PV/Wind pumping system. Figure 3 shows the curve of the accumulated amount of water in the storage tank Qacc(t) for T = 744 h. From the curve, it is clear that: • There is no negative value of Qacc(t). • The condition that the initial quantity of water in the storage tank is equal to the final quantity at t = T is satisfied.

Qacc(Wh)

The economic analysis of the results of the method is presented in Fig. 4. This analysis shows that the ideal solution consists of 21 PV panels, 2 Wind turbines and a storage tank of 172 m3, with a life cycle cost of 8992 €. From the figure, the value of LCC is high when the number of wind turbines is large, and begins to decrease until it reaches the optimal solution at NPV = 21 & NW = 2, and increases again when all the contribution will be by the PV.

210 180 150 120 90 60 30 0 727 694 661 628 595 562 529 496 463 430 397 364 331 298 265 232 199 166 133 100 67 34 1

Time (h)

Fig. 3. The accumulation of water (m3) in the storage tank for T = 744 h.

298

M. Chennaif et al.

15000 12500 10000 7500 5000 2500 0 0_9

3_9

6_7

9_6

12_5

Npv_Nw

15_4

18_3

21_2

25_1

30_0

Fig. 4. The evolution of the LCC for all possible combinations Npv_Nw.

Fig. 5. Evolution of LCC according to the contribution of each energy source.

In Fig. 5, the optimum sizing to be applied for the hybrid PV/Wind pumping system has a configuration consists of 77.2% contribution of solar energy and 22.8% of wind energy, with a minimum Life Cycle Cost of 8992 €.

5 Conclusion In this work, the ESCA method is used to find the configuration for the optimal sizing of a hybrid PV/Wind hybrid pumping system. The algorithm developed for a period of 744 h takes as input the climatic data of the site, the technical and economic data of the various components of the system. The LWSP and the LCC are used as criteria of the system optimization. An application of the method is carried out in a case study on an irrigated surface located in the city of Oujda, Morocco (latitude: 34° 41′ longitude 1° 54′), in order to find the optimal size of the various components of the pumping system according to this surface. Different configurations for the system are obtained by the proposed method. The optimal configuration includes NPV = 21, NW = 2 and VST = 172 m3 with the minimum LCC = 8992 €.

Techno-Economic Sizing of a Stand-Alone Hybrid Energy and Storage

299

References 1. Zahboune, H., et al.: Modified electric system cascade analysis for optimal sizing of an autonomous hybrid energy system. Energy 116, 1374–1384 (2016) 2. Zahboune, H., Zouggar, S., Krajacic, G., Sabev, P., Elhafyani, M.: Optimal hybrid renewable energy design in autonomous system using modified electric system cascade analysis and homer software loss of power supply probability. Energy Convers. Manag. 126, 909–922 (2016) 3. Bakelli, Y., Hadj Arab, A., Azoui, B.: Optimal sizing of photovoltaic pumping system with water tank storage using LPSP concept. Sol. Energy 85(2), 288–294 (2011) 4. Bouzidi, B.: New sizing method of PV water pumping systems. Sustain. ENERGY Technol. Assess. 4, 1–10 (2013) 5. Singh, R., Bansal, R.C., Singh, A.R.: Optimization of an isolated photo-voltaic generating unit with battery energy storage system using electric system cascade analysis. Electr. Power Syst. Res. 164(July), 188–200 (2018) 6. Ho, W.S., Hashim, H., Hassim, M.H., Muis, Z.A., Shamsuddin, N.L.M.: Design of distributed energy system through Electric System Cascade Analysis (ESCA). Appl. Energy 99, 309–315 (2012) 7. Arab, A.H., Gharbi, M.B.A.: Dimensionnement de Systèmes de Pompage Photovoltaïque. Rev. Energ. Ren. 8, 19–26 (2005)

Rotating Machines Energy Loss Due to Unbalance Ali Elkihel1, Bouchra Abouelanouar2(&), and Hassan Gziri1 1

Laboratoire Ingénierie, Management Industriel et Innovation, Faculté des Sciences et Techniques de Settat, Université Hassan 1er, Casablanca, Morocco [email protected] 2 Laboratoire de Génie Industriel et Génie Sismique, Ecole Nationale des Sciences Appliquée, Université Mohamed Premier, Oujda, Morocco [email protected]

Abstract. Unbalance is one of the most common defects in rotating machinery that causes important vibrations and subsequently an increase in energy consumption. Unbalance is defined as un-equal distribution of weight around the center of rotation. Although vibration, heat and noise are usually the results of mechanical defects, vibrations present the most relevant indicator for the identification of unbalanced shaft. By monitoring time indicators such as the Root Mean Square (RMS), unbalance is detected. This paper aims to experimentally analyze the energy losses due to shaft unbalance. The proposed methodology is meant to emphasise on the relationships between vibration level and energy consumption for different degree unbalance defect. A laboratory test rig was designed to create unbalance by adding weights at various eccentricities. For each case, RMS indicator and the electrical characteristics were measured. The results obtained from this study show that vibration analysis is an effective technique to identify unbalance severity and energy losses. Keywords: Unbalance


Vibrations
Energy consumption

1 Introduction In industries for various reasons like design defects, improper foundation and errors during installation, the shafts of the rotating machines may become unbalanced. Unbalance creates centrifugal forces generating vibration at the bearings, seals, gears and couplings, which can accelerate their degradation. Under these permanent vibrations, serious damages can be detected in all machine components. Therefore an unbalanced machine account for a significant costs in terms of lost production, but costs of replacement parts and energy consumption are also increased [1]. Several researchers have focused on the evaluation of energy consumption in the case of imbalanced shafts. It is reported that losses due to unbalance and misalignment are up to 15% [2]. Although some experimental works contested this result, it has been

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 300–308, 2020. https://doi.org/10.1007/978-3-030-53187-4_34

Rotating Machines Energy Loss Due to Unbalance

301

confirmed by several researchers that are beneficial to seek to balance the machinery in order to reduce vibrations, noise and consequently energy consumption [3]. Researchers have been investigating methods that try to find a correlation between the vibratory level and the energy losses due to the unbalance defect. S. P. Mogal and D. I. Lalwani [4] proposed order analysis (phase and amplitude) technique for unbalance and misalignment fault diagnosis. They concluded that order analysis is an effective technique for fault diagnosis. A. Elkhatib [5] treated the effect of vibration levels the effect of the frequencies on the power consumption. The result obtained showed a strong correlation between vibration levels and the power consumption. A. Mukesh et al., [3] experimentally investigated the fact that the power consumption can be reduced if the vibrations can be kept under control. In this work, we develop a smart experimental setup that uses vibration analysis to detect shaft unbalance. For the sake of comparison, a study of energy consumption for different gradations of the unbalanced shaft has also been made.

2 Theory Centrifugal force is one of the basic excitation forces in rotating machinery. Indeed, whatever the care taken in the construction of the machines, it is not possible to obtain a coincidence of the center of gravity with the axis of rotation. This phenomenon is known under the name of unbalance. In general, we can distinguish there types of unbalance: static unbalance, couple unbalance and dynamic unbalance. The difference between these types exists in the distribution of the mass on the rotor and how it will affect the position of the principal axis of inertia with respect to the axis of rotation (see Fig. 1).

Fig. 1. Static, couple and dynamic unbalance

302

A. Elkihel et al.

The unbalance occurs in a rotating machine when the main axis of rotation and the axis passing through the center of gravity do not coincide. In fact, in a balanced rotor of mass M, the addition of a mass m on a radius r displaced its center of gravity G by a distance e called eccentricity (see Fig. 2). Displacement of center of gravity and centrifugal forces created during rotation are the main effects of unbalance.

m

M

Fig. 2. Unbalance schema

The unbalance U, thus created, is defined by: U ¼ m:r ¼ M:e

ð1Þ

The force created by the unbalance depends on the speed of rotation and the amount of unbalance: F ¼ m:r:x2 ¼ U: x2

ð2Þ

Where: F = force (N), m = mass (kg), r = radius (m), x = speed (rad/sec) and S = center of gravity. Between unbalance and machine vibrations there is no proven relationship. However, researchers have been able to demonstrate experimentally that the vibration due to unbalance is directionally proportional to the amount of unbalance.

3 Experimental Procedure Experiments were conducted to find the correlation between the level of vibrations and the energy consumption and the different cases of unbalance. The obtained data were further used to identify energy losses due to shaft unbalance. Figure 3 depicts the experimental set up designed to study the shaft unbalance. It consists of an asynchronous motor, a shaft of 20 mm diameter supported by two

Rotating Machines Energy Loss Due to Unbalance

303

identical ball bearings and connected to the motor with a flexible coupling, and an over hung circular disc with holes evenly distributed to insert the masse.

Fig. 3. Experimental set up

A: DC Motor, B: Bearing Support, C: Coupling, D: Disk, E: Shaft, F: Base, G: Ball Bearing, P1, P2 and P3: Measurement positions in Motor, Bearing 1 and Bearing 2, respectively. Vibration data is acquired in different measurement positions: motor and bearings, when the shaft is in its balanced and unbalanced states, and under different unbalance conditions. Unbalance is created by adding weights at various positions and angles. The vibration measurements were obtained for each of the following imbalance conditions (see Fig. 4):

Angle θ

Radius r Fig. 4. Unbalance conditions

304

A. Elkihel et al.

Case 1: The radius r and the angle h are fixed and the mass m is varied (m1 = 8.9; m2 = 9.4; m3 = 21.6; m4 = 40 g). Case 2: The mass m and the angle h are fixed and the radius r is varied (4.5; 8.5; 12.5 mm). Case 3: The mass m and the radius r are fixed and the angle h is varied (0°; 60°; 120°). Thus, electrical measurements have been taken in ‘healthy’ and defective cases to illustrate the effect of unbalance on energy consumption.

4 Results and Discussion The results obtained from the vibration measurements for the different cases described previously are presented in the following figures: Figures 5, 6 and 7 show the impact of the mass, radius and angle where we inserted the masses on the vibration measurements (RMS). From these figures, we find that all these parameters influence the evolution of the RMS values in the different measurement positions. However, the positioning angle of the masses remarkably influenced the RMS values.

RMS (mm/s)

Good

Measuring positions

Fig. 5. Effect of mass on the RMS

Rotating Machines Energy Loss Due to Unbalance

305

RMS (mm/s)

RMS Good Satisfactory Unsatisfactory

Measuring positions

Fig. 6. Effect of radius on the RMS RMS (mm/s)

RMS Good Satisfactory Unsatisfactory

Measuring positions

Fig. 7. Effect of angle on the RMS

Now, the results obtained from the electric measurements are shown in the following Tables 1 and 2: Initial State: Balanced shaft

Table 1. Measurements of current and voltage in the initial state (Balanced shaft) Motor phases Phase 1 Phase 2 Phase 3

Current (A) Voltage (V) 0.8 315.4 0.9 325.4 1.1 316.2

306

A. Elkihel et al.

From these results we find that the active power is equal to: Pactive ¼ ð315:4  0:8 þ 325:4  0:9 þ 316:2  1:1Þ  0:28 ¼ 250:712w Unbalanced State: Condition of unbalance m = 21.6 g, r = 8.5 cm et h = 60 Table 2. Measurements of current and voltage in the initial state (Balanced shaft) Motor phases Phase 1 Phase 2 Phase 3

Current (A) Voltage (V) 0.9 318.9 1 332.5 1.3 317.8

From these results we find that the active power is equal to: Pactive ¼ ð318:9  0:9 þ 332:5  1 þ 317:8  1:3Þ  0:28 ¼ 289:142w By comparing the initial state with the state of imbalance it can be deduced that the imbalance has increased the energy consumption (see Fig. 8). Also, we find that the mass of imbalance has an important effect on the vibration level and consequently on the energy consumption. (Cases 1 and 2) Unlike the position (i.e. the radius) where the effect was not remarkable and the difference between the energy consumption in the balanced shaft and the defective one was not significant (Case 3).

Fig. 8. Measurements of electrical power for the balanced and unbalanced shaft under different conditions of imbalance:

Rotating Machines Energy Loss Due to Unbalance

307

Fig. 9. Measurements of RMS and electrical power for various unbalance conditions

Case 1 (Unbalance conditions): m = 21.6g, r = 8.5 mm et h = 60° Case 2 (Unbalance conditions): m = 40g, r = 8.5 mm et h = 60° Case 3 (Unbalance conditions): m = 9.4, r=12.5 mm et h = 60° To establish the relationship between unbalances shaft degree and electrical consumption. Figure 9 shows vibration measurements (RMS) and electrical measurements (Electric power) for different imbalanced shaft conditions. Analyzing these graphs, it can be extracted that the unbalance conditions changed the amplitude of the vibration and the energy consumption changed with a similar tendency as the vibrations changed.

5 Conclusion The first results obtained are encouraging. Indeed, as the vibrations increase the energy consumption is also increased for the same unbalance conditions. This study outlines the energy loss due to unbalanced shaft under various cases of unbalance and compared to the same shaft in balanced conditions. It has been experimentally shown that the use of vibration measurements is a potential indicator for the measurement of energy consumption. However, in order to quantify the percentage of energy losses caused by each unbalance conditions, more tests should be done.

References 1. Saleem, M.A., Diwakar, G., Satyanarayana, M.R.S.: Detection of unbalance in rotating machines using shaft deflection measurement during its operation. J. Mech. Civ. Eng. (IOSRJMCE) 3(3), 8–20 (2012) 2. Gaberson, H.A.: Rotating machinery energy loss due to misalignment. In: Proceedings of the 31st Intersociety Energy Conversion Engineering Conference, vol. 3, pp. 1809–1812 (1996)

308

A. Elkihel et al.

3. Mukesh, P.S., Bulsara, A.: Energy loss due to unbalance in rotor–shaft system. J. Eng. Des. Technol. 14(2), 277–285 (2016) 4. Thomson, S.P., Lalwani, D.I.: J. Meas. Eng. 3(4), 114–122 (2013) 5. Elkhatib, A.: Energy consumption and machinery vibrations. In: International Conference on Sound & Vibrations, ICSV14, Cairns, 9–12 July, pp. 1–6 (2007)

Comparative Study Between PI Speed Control and Sliding Mode Control of BLDC Motor Ahmed Loukmane El Idrissi1(&), Jamal Bouchnaif1, Mohammed Mokhtari1, and Anas Bensliman2 1

School of Technology Oujda Laboratory of Electrical Engineering and Maintenance (LEEM), University Mohammed 1, BP 473, 60000 Oujda, Morocco [email protected] 2 University Mohammed 1, Oujda, Morocco

Abstract. The brushless DC motors (BLDC) are used in several applications such as electrical cars, medical and industrial equipment, where speed control with high efficiency is required to satisfy specifications regarding load and tracking references variations. In this paper we propose a comparative study between the classical PI controller and a Sliding Mode Controller (SMC) for closed loop speed control of a trapezoidal back-EMF BLDC using a DC/DC buck converter under various load conditions. Simulation results of speed response, torque and current behaviors of the studied BLDC are also presented to support the noted improvements. Keywords: BLDC


PI control
Speed control
Sliding mode

1 Introduction The brushless DC motors are becoming more popular due to their high efficiency, their high torque-weight ratio, noiseless operation and low maintenance cost. The BLDC motor has a permanent magnet rotor and the stator windings are wounded in different ways with one or multiple phases. The very common BLDC motor used is the three phases BLDC motor with a trapezoidal back-EMF or a sinusoidal back-EMF [1]. In order to make motion control more reliable, more efficient and less noisy, the recent trend has been to use brushless DC motors, they are also lighter compared to brushed motors with the same rated power, the brushes in conventional DC motors wear out over time and may cause sparking, that’s why we use the BLDC motor in operations that demand long life and reliability, The BLDC motor is used in enormous applications that demand very high speed with low torque such the hard drivers or with big torque with high dynamic speed range such electrical cars. In the industry the conventional PI controller still widely used due to his simple control structure, however, there are other more complex control strategies like Neural Network Control or the Fuzzy Logic Control [4] which need continuous learning, optimization and adjustment. In other hand, there are less complex control strategies like the Sliding Mode control [7, 8, 10]. Usually, on those control strategies, a dual loop of the speed and current is used [5, 6, 9]. In order to simplify the complexity of the © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 309–317, 2020. https://doi.org/10.1007/978-3-030-53187-4_35

310

A. L. El Idrissi et al.

control algorithms we use two control loops; the inner loop controls the voltage of the buck converter and the outer loop controls the speed of the motor. The organization of this paper is as follows. It initially begins by the description of the studied system structure, then the adopted control strategy for the speed control, and finally discuss and analysis the simulated results (Table 1). Table 1. List of symbols and abbreviations. Nomenclature Back EMF constant Ke

vk

F (he) he xr Ls M R

Back EMF force reference as function of rotor position Electrical angle of the rotor Angular speed of the rotor Inductance of each phase Mutual inductance Resistance of each phase

ik

SMC Xmes

Sliding Mode Control Measured angular speed

D% Tr5% Xreq

J Tl b p hm

Phase voltage applied from inverter Phase current Rotor inertia Load torque Damping constant Number of poles Mechanical angle of the rotor Percentage overshoot Response time Required angular speed

2 System Modeling and Description The studied system shown in Fig. 1 is based on a three phases trapezoidal back-EMF BLDC motor equipped with Hall, armature current and speed sensors, a three phases power inverter, a DC/DC buck converter and a torque source for load variation.

Fig. 1. System model under Matlab/Simulink software

Comparative Study Between PI Speed Control and Sliding Mode Control

311

The armature winding of the BLDC motor can be expressed as follows [2, 3]: va ¼ Ra ia þ La didta þ Lab didtb þ Lca didtc þ ea vb ¼ Rb ib þ Lb didtb þ Lba didta þ Lbc didtc þ eb

ð1Þ

vc ¼ Ra ic þ La didta þ Lab didtb þ Lca didtc þ ea Which can be written in the matrix form as: 2

3 2 Ra va 4 vb 5 ¼ 4 0 0 vc

32 3 2 La 0 ia d 0 54 ib 5 þ 4 Lba dt Rc ic Lca

0 Rb 0

Lab Lb Lcb

32 3 2 3 Lca ia ea Lbc 54 ib 5 þ 4 eb 5 Lc ic ea

ð2Þ

Assuming that the windings are symmetric and the system is balanced, we have: ia þ ib þ ic ¼ 0

ð3Þ

La ¼ Lb ¼ Lc ¼ Ls

ð4Þ

Ra ¼ Rb ¼ Rc ¼ R

ð5Þ

Lab ¼ Lac ¼ Lbc ¼ Lba ¼ Lca ¼ Lab ¼ M

ð6Þ

So, the matrix become: 2

3 2 R va 4 vb 5 ¼ 4 0 0 vc

0 R 0

32 3 2 Ls 0 ia d 0 54 ib 5 þ 4 M dt R M ic

M Ls M

32 3 2 3 ia ea M M 54 ib 5 þ 4 eb 5 Ls ic ec

ð7Þ

0 L 0

32 3 2 3 0 ia ea 0 54 ib 5 þ 4 eb 5 L ic ec

ð8Þ

We put: L ¼ Ls  M, then we have: 2

3 2 R va 4 vb 5 ¼ 4 0 0 vc

0 R 0

2 32 3 L 0 ia d 0 54 ib 5 þ 4 0 dt 0 R ic

8 ea ¼ Ke :xr :F ðhe Þ > > <   The trapezoidal back - EMF is given by : eb ¼ Ke :xr :F he  2p 3 > >   : ec ¼ Ke :xr :F he þ 2p 3

ð9Þ

The torque equations can be deduced as: 

   2p 2p Te ¼ Kx :xr :F ðhe Þ:ia þ Kx :xr :F he  :ib þ Kx :xr :F he þ :ic 3 3

ð10Þ

312

A. L. El Idrissi et al.

Te ¼ J

dxr þ bxr þ Tl dt

ð11Þ

Where the function F is defined as below: 8 2p ¼ 1 if 0  > >  he \2p3 2p < 6 ¼ 1  p he  3 if 3  he \p F ð he Þ ¼ 5p > ¼ 1 if p  >  he \5p3 5p : 6 ¼ 1 þ p he  3 if 3  he \2p

ð12Þ

In the other hand, the DC/DC buck converter can be modeled using the following equation: Vdc ¼ aVsource

ð13Þ

With a is the duty cycle to be controlled.

3 Control Strategy Our study will be focused on the dynamic of BLDC speed according to the load variation. As shown in Fig. 2, two control loops are used, the inner loop controls the voltage of the buck converter and the outer loop controls the speed of the motor.

Fig. 2. Control strategy block diagram.

The BLDC motor parameters tested in simulation are defined in the following Table 2:

Comparative Study Between PI Speed Control and Sliding Mode Control

313

Table 2. The parameters of the BLDC motor J Ls M R p pm PN

3.1

0.01 kg/m2 Rotor inertia Inductance of each phase 0.1 mH Mutual inductance 0.01 mH Resistance of each phase 0.2 X Number of poles 4 Maximum permanent magnet flux linkage 0.0175 Wb Rated power 3 KW

PI Controller

The speed regulation is based on controlling the output voltage of the DC/DC buck converter using a PI controller. The proportional coefficient Kpw and the integral coefficient Kiw have been calculated using Tune function in Matlab and chosen value are respectively 0.1 and 15. The results of the PI controller in Fig. 3 show the overshoots of the speed when the speed reference changes, also we notice that the speed drops significantly when we add a load of 5 N.m at T = 1 s.

Fig. 3. The speed variation using a PI control

The PI controller track the speed reference but the response of the BLDC motor isn’t as fast as desired. 3.2

Sliding Mode Control

In order to improve the speed response of the BLDC motor we propose a sliding mode control, which will be replacing the PI controller of the outer loop.

314

A. L. El Idrissi et al.

The sliding surface can be defined as: SX ¼ Xmes  Xreq

ð14Þ

The sliding surface derivative is then given by: dSX dXmes dXreq ¼  ¼ KX :signðSX Þ dt dt dt

ð15Þ

From (8), (10), (11), (13) and (15), the referred Vdc voltage is given by: Vdc

req

¼ K1 Xmes þ K2 signðSX Þ

ð16Þ

With K1 and K2 constants to be chosen in order to insure system stability. As shown in the Fig. 4, using SMC the speed reference tracking has been improved compared to the PI controller.

Fig. 4. BLDC speed response using Sliding Mode Controller

4 Comparative Analysis of Simulation Results In this section, we compare the performance of the proposed SMC controller under the same conditions as the traditional PI controller. From Fig. 5 the response of the sliding mode control is much better than the PI control, and the drop of the speed when we add a load to the system is significant in the PI control and it’s negligible in the SMC.

Comparative Study Between PI Speed Control and Sliding Mode Control

315

Fig. 5. Comparison between the PI and the sliding mode control

Table 3 shows the performances of the SMC controller in comparison with the conventional PI controller.

Table 3. Comparison between the response of the PI and the sliding mode control to the variation of the speed and load.

Conventional PI controller Sliding mode controller

500 rpm

1500 rpm

D% = 22.8% Tr5 % = 93 ms D% = 5.6% Tr5 % = 31 ms

D% = 5.5% Tr5 % = 98.7 ms D% = 0.5% Tr5 % = 64 ms

Adding load of −5 N.m The speed drop by 4.9% The speed drop by 0.75%

Decreasing to 800 rpm D% = 8.6% Tr5 % = 95 ms D% = 1.7% Tr5 % = 47.5 ms

In this study, the current is not controlled, there are two loops, the inner loop is dedicated to regulate the voltage and the outer loop is dedicated to regulate the speed, the current reach a maximum of 100 A in both controllers, and the average of the current is under 50 A, despite of the non-control of the current, the value of the current in the phases of the motor still tolerable (Fig. 6). In both cases, as can be seen from Fig. 7, the motor generates the same amount of torque, which makes the SMC controller a very good option to have a good result acting only on the buck converter.

316

A. L. El Idrissi et al.

PI

SMC

Fig. 6. The current in the three phases ia, ib and ic using SMC vs PI controller

PI

SMC

Fig. 7. The torque variations when SMC controller vs the conventional PI

5 Conclusion In this paper, a particular attention was paid to speed variations of a BLDC motor. A closed loop speed control is carried out using PI and sliding mode controllers. A model of the studied system has been also developed under MATLAB/SIMULINK environment and simulation results with comparative analysis have been presented.

Comparative Study Between PI Speed Control and Sliding Mode Control

317

References 1. Rao, A.P.C., Obulesh, Y.P., Babu, Ch.S.: Mathematical modeling of BLDC motor with closed loop speed control using PID controller under various loading conditions. ARPN J. Eng. Appl. Sci. 7(10), 1321–1328 (2012) 2. Tashakori, A., Ektesabi, M., Hosseinzadeh, N.: Modeling of BLDC motor with ideal backEMF for automotive applications. In: Proceedings of the World Congress on Engineering, vol. 2 (2011) 3. Tibor, B., Fedak, V., Durovský, F.: Modeling and simulation of the BLDC motor in MATLAB GUI. In: 2011 IEEE International Symposium on Industrial Electronics. IEEE (2011) 4. Xu, W., Jiang, Y., Mu, C.: Novel composite sliding mode control for PMSM drive system based on disturbance observer. IEEE Trans. Appl. Supercond. 26(7), 1–5 (2016) 5. Guo, D., et al.: Sliding mode high speed control of PMSM for electric vehicle based on fluxweakening control strategy. In: 2017 36th Chinese Control Conference (CCC). IEEE (2017) 6. Deenadayalan, A., Saravana Ilango, G.: Modified sliding mode observer for position and speed estimations in brushless DC motor. In: 2011 Annual IEEE India Conference. IEEE (2011) 7. Rath, J.J., Saha, S., Ikkurti, H.P.: Sliding mode scheme for speed and current control of brushless DC (BLDC) motor. In: IEEE-International Conference on Advances in Engineering, Science and Management (ICAESM-2012). IEEE (2012) 8. Xiaojuan, Y., Jinglin, L.: A novel sliding mode control for BLDC motor network control system. In: 2010 3rd International Conference on Advanced Computer Theory and Engineering (ICACTE), vol. 2. IEEE (2010) 9. Hao, L., Toliyat, H.A.: BLDC motor full-speed operation using hybrid sliding mode observer. In: 2003 Eighteenth Annual IEEE Applied Power Electronics Conference and Exposition. APEC 2003, vol. 1. IEEE (2003) 10. Shao, Y., et al.: Sliding mode speed control for brushless DC motor based on sliding mode torque observer. In: 2015 IEEE International Conference on Information and Automation. IEEE (2015)

PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine Using a Fuzzy Logic Controller Mhamed Fannakh(&), Mohamed Larbi Elhafyani, Hassan Zahboune, and Smail Zouggar Laboratory of Electrical Engineering and Maintenance – LEEM, High School of Technology, University Mohammed 1st, Oujda, Morocco [email protected]

Abstract. The Wind System Power varies according to the wind speed, and the wind turbine (WT) operates at an optimal point which depends on its rotation speed. For that, this paper presents a Fuzzy Logic Approach (FLA) to tracking the maximum available power of a WT system based on Permanent Magnet Synchronous Generator (PMSG) to supply the load. Its main advantage is not requiring a mechanical sensor or a prior knowledge of the WT characteristic. The PMSG–PWM rectifier combination used is compared with other topologies types in term of cost and efficiency. By modifying the PWM rectifier modulation index, the reflected voltage at the PMSG is varied and consequently its rotational speed, which allows tracking the maximum power point (MPP) of WT. So as to check the proposed strategy feasibility, a PSIM and Matlab co-simulation is made. The simulation results demonstrated the competitiveness of FLA, from the point of view of oscillations around the maximum point and response time compared to P&O technique. Keywords: Wind turbine controller
Co-simulation


PMSG
PWM rectifier
MPPT
Fuzzy logic

1 Introduction In the context of sustainable development, faced with planetary challenges posed by the depletion of fossil energy resources and climatic changes, strong incentives push to develop renewables energies. Non-renewable forms of energy production generate significant environmental pollution caused by emission of greenhouse gases into the air. Several sources of renewable energies like wind energy are under the process of exploiting and researching, whose goal is to develop power extraction techniques to reliable lower costs (of manufacture and maintenance) and increase energy efficiency. The motivating nature of these energies, the development of the wind turbine industry, the evolution of semiconductor technology and new control techniques are reunited to make variable-speed wind turbine with a Permanent Magnet Synchronous Generator (PMSG) more attractive than other types of wind turbine generator [1]. There are several concepts of PMSG dedicated to wind turbine applications, standard construction machines (radial magnetization), disc-shaped generators (axial fields) or © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 318–329, 2020. https://doi.org/10.1007/978-3-030-53187-4_36

PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine

319

with external rotor [2]. The wind systems of this generator type have a low failure rate because of eliminating the gearbox and the system of brushes/commutators. Maintenance costs are then minimized which is very interesting in wind energy applications [3]. The PMSG wind turbine is often associated with power converters in order to maximize the produced power, to regulate the output voltage or to ensure the grid connection. In the remainder of this paper we study a method based on single input fuzzy logic controller to track the maximum power point.

2 System Modelling and Analysis In the interest of optimizing the extracted power in the desired form, the power converters are used with PMSG wind turbine in many topologies [4]. The optimal configuration is chosen based on electrical load types and economic criteria. In this works, we decided to use a DC load to test the MPPT command. Two essential configurations are possible with this load type. In the direction of reducing costs, the diode bridge seems attractive in low power provided that the energy performance is not too degraded. The diode rectifier is placed in this topology between the DC load and the PMSG generator. The power transited between the generator and the DC load is therefore unidirectional. This greatly limits the generator speed adjustment and consequently the possibility of extracting maximum power [5]. The suitable solution is to add a DC/DC converter between diode rectifier and the load [1] as illustrated in Fig. 1. a. In the Fig. 1.b we present a topology using a PWM rectifier. The rectifier bridge is composed of six IGBT transistors with their antiparallel diodes. The PMSG is then perfectly controlled and it is possible to extract the maximum power from the wind turbine by controlling the current in the generator [6]. This last topology appear appropriate to be chosen in view of its simplicity by reducing the complexity of the configuration and minimizing the number of converters, which will decrease the cost of the system, ensuring a high efficiency and less maintenance.

(a)

(b)

Fig. 1. a. Synoptic of topology with two static converters. b. Synoptic of topology with one static converter

320

2.1

M. Fannakh et al.

Wind Turbine Modelling

A wind turbine is a device that transforms a part of the kinetic energy to mechanical energy. Analytical model of the wind turbine have been studied, analyzed and compared in many works [7]. In this paper we are selected the model used by the PSIM software [8]. The power generated by a wind turbine can be expressed as: 1 P ¼ :A:v3wind :q:Cp 2

ð1Þ

Where A is the area of the rotor blade (in m2), vwind is the wind speed (in m/sec.), q is the air density (it is approximately 1.225 kg/m3), and Cp is the power coefficient. The power coefficient Cp is a function of the tip speed ratio k and the blade pitch angle b. The tip speed ratio (k) is the ratio between the rotor speed and the wind speed, it can be expressed as: k¼

xm :Rblade vwind

ð2Þ

Where xm is the rotor rotational speed (in rad/sec) and Rblade is the radius of the rotor blade (in m). Figure 2 describes the relationship between the power coefficient Cp and the tip speed ratio k when the b is set to zero. The figure shows that the power coefficient Cp reaches the maximum of 0.48 when the tip speed ratio k is 8.13

Fig. 2. Turbine power characteristic

PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine

2.2

321

PMSG and PWM Rectifier Modelling

In the permanent magnet machine the magnetic field produced by permanent magnets has the advantage of eliminating brushes and rotor losses which provide a constant excitation field and high efficiency. The mathematical model of the permanent magnet machine assumes some simplifying hypotheses as [9]: ∙ The absence of saturation in the magnetic circuit ∙ The sinusoidal distribution of the MMF created by the stator winding

∙ Hysteresis is neglected with Foucault currents ∙ The notching effect is negligible ∙ The windings resistance is constant

The PMSG electrical equations in the Park system are described by [1]: disd  xe Lsq isq dt

ð3Þ

disq þ xe ðLsd isd þ wf Þ dt

ð4Þ

Vsd ¼ RS isd þ Lsd Vsq ¼ RS isq þ Lsq

Where Vsd and Vsq are the d-axis and q-axis stator voltages (V), respectively; isd and isq are the d-axis and q-axis stator currents (A), respectively; Lsd and Lsq are the daxis and q-axis inductances (H), respectively; Rs is the stator winding resistance (X); Wf is the permanent magnetic flux (Wb); and xe is the electrical rotating speed (rad/s) of the generator. The electromagnetic torque can also be expressed in dq form as:   3

Tem ¼ p Lsd  Lsq isd isq þ isq wf 2

ð5Þ

Where p is the number of the pole pairs. The following mechanical equations express the dynamics of the machine: J

dxm ¼ Tm  Tem  fxm dt

ð6Þ

Where J is the total equivalent inertia (Kg m2); xm is the mechanical rotating speed (rad/s); Tm and Tem are the shaft torque and electromagnetic torque (Nm), respectively and f is the viscous friction coefficient (N.m.s). The PWM rectifier’s bridge consists of six fully-controlled IGBT transistors with antiparallel diode. Each transistor-diode assembly can be considered as an ideal switch and it’s bidirectional in current, and unidirectional in voltage [10].

322

M. Fannakh et al.

3 MPPT Control The power characteristic of a wind turbine is strongly nonlinear. For each wind speed, the system must find the maximum power which is equivalent to finding the optimal rotation speed [11]. The diagram in Fig. 3 illustrates the wind turbine power characteristic as a function of the rotor speed. Each colored curve corresponds to a given wind speed.

Fig. 3. Characteristics of turbine power as a function of the mechanical speed

In order to determine the optimal operating point of the wind turbine, it is necessary to include a Maximum Power Point Tracker (MPPT) algorithm. Many researchers has been written on the topic of MPPT algorithms, especially for wind energy systems. There are two major categories of MPPT algorithms. In the first category, the PMSG is driven by an active rectifier, followed by the DC bus capacitor. At least a mechanical sensor is needed to know the wind speed and/or that of rotation. This algorithm types as Optimal Torque Control or Tip Speed Ratio Control seem simple, fast and efficient. However, a precise measurement for wind speed and/or rotational speed is necessary which increases the cost of the system [11]. In the second alternative, the PMSG is followed by a diode bridge, a boost converter and the DC bus capacitor. The maximum power point can be found by adjusting the duty ratio of DC/DC converter [1]. With this method, the PMSG efficiency is lower and more harmonics are generated. However, robustness, simplicity of control and conversion efficiency are substantially improved. In the small WT, the system cost reducing is widely recommended which is reached with this method, due to elimination of mechanical sensors. In the second discussed topology, perturbation and observation (P&O) strategy might be used if the characteristics of the system are unknown. The AC power

PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine

323

generated by PMSG is rectified into DC power and boosted by using a chopper to find the MPP. Hence, the DC current (Idc) and voltage (Vdc) are expressed as [13]: p qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Idc ¼ pffiffiffi i2qs þ i2ds 2 3 pffiffiffi 3 3 Vs Vdc ¼ p

ð7Þ ð8Þ

Where, Vs is the magnitude of phase voltage, isd and isq are, respectively, the -axis and -axis stator currents. Starting from that PMSG output currents and voltages are proportional to the torque and rotor speed, thus perturbing the output voltage will cause varying in the generator rotor speed and consequently varies the output power [13]. As discussed in Sect. 2, the topology using PWM rectifier is chosen. With PWM rectifier, it is possible to control the DC output voltage by controlling the input voltages Vd and Vq of the rectifier. In view of this fact, the following two modulation factors are used as control variables [14]: pffiffiffi pffiffiffi 2Vq 2Vd md ¼ ; mq ¼ Vdc Vdc

ð9Þ

x The reference signals needed for controlling PWM rectifier are modulation index m and phase angle U, and are given by: m¼

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 m2d þ m2q ; U ¼ tan1 md =mq

ð10Þ

The controlled signals must be in phase with the PMSG voltage. For that, a Phase Locked Loop (PLL), based on a multi-variable filter (MFV), is added to the structure control [15]. The block diagram of the control system is given in Fig. 4.

Fig. 4. Block diagram of the studied system

324

3.1

M. Fannakh et al.

Perturb & Observe MPPT Algorithm

The P&O (Perturb and Observe) algorithm is one of the simplest algorithms. It does not require any prior knowledge of the mechanical parameters. It consists in introducing a small perturbation of the modulation index. Its effect is subsequently noticed in the electric power DC side. If an increase in electrical power is recorded, then the modulation index will be increased again with the same step size as the previous one. If, on the other hand, a decrease in electric power is recorded, the modulation index direction is then reversed [1, 12]. Basic P&O control has some disadvantages related to [1]: • Output power oscillation is permanent even during fixed wind speed. • The process of tracking the point of the maximum power is slowed because the step of modulation index variation is fixed. • This strategy does not allow the detection of small transient wind variations. This can cause unwanted and erroneous power variations. 3.2

Single-Input Fuzzy Logic MPPT Controller

Due to the disadvantages of basic P&O control and in order to act at different speeds of regulation and with increased precisions, the Single-Input Fuzzy Logic strategy is proposed. The Single-Input Fuzzy Logic controller allows a variable variation of modulation index step, which reduces the negative effects of basic P&O strategy. The Single-Input Fuzzy Logic controller is constructed by choosing the error E(k) defined in Eq. 11 as an input signal and the modulation index step as an output signal. In case of MPP, E(k) should be zero. This input is chosen so that the instant value of E(k) shows if the load operation power point is located on the right or in the left compared to the Pmax actual position [15]. The studied control strategy is summarized in Fig. 5.

Fig. 5. Flow diagram of the fuzzy controller

PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine

EðkÞ ¼

Pdc ðkÞ  Pdc ðk  1Þ Vdc ðkÞ  Vdc ðk  1Þ

325

ð11Þ

With Pdc is the DC side power and Vdc is the DC side voltage. The Mamdani method is most commonly used in industrial applications because it allows describing the human expertise in more intuitive. The structure of Mamdani method is explained in [15]: Triangular and trapezoidal membership functions are the most forms used in Fuzzification. The input membership function used is given by Fig. 6.a. A five-term fuzzy set, positive big (PB), positive small (PS), zero (ZE), negative small (NS), and negative big (NB), is defined to describe input variable.

(a)

(b)

Fig. 6. a. Membership functions of input variable. b. Membership functions of output variable

The output variable is the modulation index variation, which is transmitted to the PWM rectifier to drive the load. The modulation index variation membership is given by Fig. 6.b. The five linguistic variables used are in the output variable is positive big (PB), positive small (PS), zero (ZE), negative small (NS), and negative big (NB). To determine the error variation range, we have carried out simulation tests on the studied system for several wind velocity. The results obtained implies that the error interval is about [−20; 20] and the modulation index variation is about [−0.01; 0.01]. The choice of the membership function form depends mainly on the experience and expertise. Triangular and trapezoidal are the most forms used because they can provide the best results. The fuzzy rules where the input is fuzzy sets of error (E) and the output of these rules is the change of modulation index is designed such as: ∙ If E(k) is NB THEN dm is PB ∙ If E(k) is NS THEN dm is PS ∙ If E(k) is ZE THEN dm is ZE

∙ If E(k) is PS THEN dm is NS ∙ If E(k) is PB THEN dm is PB

326

M. Fannakh et al.

Fig. 7. Power system in PSIM software

4 Simulation Results and Discussion The proposed technique is co-simulated using PSIM and Matlab/Simulink softwares. Matlab/Simulink is a suitable platform for control and regulation of simulation processes. Conversely, PSIM is dedicated to power electronic circuits and machine simulation tasks with fast and robust algorithms. The two softwares combination improves the simulation efficiency and increases its speed. To interface the both softwares, Simulink software has an element target called SimCoupler. The SimCoupler interface consists of two parts: the link nodes in PSIM and the SimCoupler model block in Simulink. With the SimCoupler Module, the power part of a system is implemented and simulated in PSIM, and the control system part in Simulink as shown in Fig. 7 and Fig. 8.

Fig. 8. Control system in MATLAB/Simulink software

The power coefficients curve Cp in Fig. 9 show the efficiency of the Single-Input Fuzzy Logic Controller. When the wind speed changes, the algorithm succeeds in putting the system back to its optimal operating point and captures the maximum possible energy. We can see from the curves that the value of the power coefficient regains its maximum value of 0.48 rapidly when using the fuzzy logic. This is because the modulation index step is not fixed as for the control by the P&O algorithm.

PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine

327

Fig. 9. Power coefficient of P&O strategy and FLC strategy

As indicated in Fig. 10, the DC side power remains constant in the steady state with small regular oscillations of the order of 10 W. At each wind velocity variation, the transitional regime has an exponential form of less than 0.2 s due to the use of the capacitor filter. The Fig. 11 present the waveform of three phases currents obtained using the single input fuzzy logic controller. They are deformed because of the harmonics injected by the PWM rectifier.

Fig. 10. DC power output

Fig. 11. Current waveforms of PMSG output

328

M. Fannakh et al.

5 Conclusion The problematic posed in this paper allowed us to study by co-simulation a wind system constituted by a PMSG and PWM rectifier. Therefore two methods were compared. Based on simulation results, the single-input fuzzy logic is better than P&O controller’s in rapidity and efficiency. The co-simulation of PSIM and MATLAB/Simulink softwares improve the simulation time which opens a new perspective in more complex control systems. Experimental validation will be the subject of future work.

References 1. Soufyane, B., Smail, Z., Abdelhamid, R., et al.: A comparative investigation and evaluation of maximum power point tracking algorithms applied to wind electric water pumping system. In: International Conference on Electronic Engineering and Renewable Energy, pp. 510–523. Springer, Singapore (2018) 2. Goudarzi, N., Zhu, W.D.: A review on the development of wind turbine generators across the world. Int. J. Dyn. Control 1(2), 192–202 (2013) 3. Cheng, K.W.E., Lin, J.K., Bao, Y.J., et al.: Review of the wind energy generating system (2009) 4. Baroudi, J.A., Dinavahi, V., Knight, A.M.: A review of power converter topologies for wind generators. Renew. Energy 32(14), 2369–2385 (2007) 5. Milivojevic, N., Stamenkovic, I., Schofield, N., et al.: Electrical machines and power electronic drives for wind turbine applications. In: 2008 34th Annual Conference of IEEE Industrial Electronics, pp. 2326–2331. IEEE (2008) 6. Errami, Y., Benchagra, M., Hilal, M., et al.: Control strategy for PMSG wind farm based on MPPT and direct power control. In: 2012 International Conference on Multimedia Computing and Systems, pp. 1125–1130. IEEE (2012) 7. Slootweg, J.G., De Haan, S.W.H., Polinder, H., et al.: Modeling wind turbines in power system dynamics simulations. In: 2001 Power Engineering Society Summer Meeting. Conference Proceedings (Cat. No. 01CH37262), pp. 22–26. IEEE (2001) 8. Heier, S., Waddington, R.: Grid Integration of Wind Energy Conversion Systems. Wiley, Hoboken (2006) 9. Melício, R., Mendes, V.M.F., Catalão, J.P.S.: Wind turbines with permanent magnet synchronous generator and full-power converters: modelling, control and simulation. In: Wind Turbines, pp. 465–470 (2011) 10. Knapczyk, M., Pieńkowski, K.: Analysis of pulse width modulation techniques for AC/DC line-side converters. Prace Naukowe Instytutu Maszyn, Napędów i Pomiarów Elektrycznych Politechniki Wrocławskiej. Studia i Materiały, vol. 59, no. 26, pp. 194–209 (2006) 11. Abdullah, M.A., Yatim, A.H.M., Tan, C.W., et al.: A review of maximum power point tracking algorithms for wind energy systems. Renew. Sustain. Energy Rev. 16(5), 3220– 3227 (2012) 12. Lahfaoui, B., Zouggar, S., Elhafyani, M.L., et al.: Experimental study of P&O MPPT control for wind PMSG turbine. In: 2015 3rd International Renewable and Sustainable Energy Conference (IRSEC), pp. 1–6. IEEE (2015)

PSIM and Matlab Co-simulation of a Sensorless MPPT for PMSG Wind Turbine

329

13. Ahmed, R., Naaman, A., M’sirdi, N.K., et al.: Sensorless MPPT technique for PMSG micro wind turbines based on state-flow. In: 2014 International Conference on Renewable Energies for Developing Countries, pp. 161–166. IEEE (2014) 14. Iwaji, Y., Fukuda, S.: A parameter design method of PWM voltage-source rectifier. Electr. Eng. Jpn. 113(8), 116–126 (1993) 15. Fannakh, M., Elhafyani, M.L., Zouggar, S.: Fuzzy logic approach to improve the performances of grid-connected photovoltaic power system. In: 2018 6th International Renewable and Sustainable Energy Conference (IRSEC), pp. 1–6. IEEE (2018)

Contribution to Power Maximization of an Asynchronous Wind Electric Water Pumping System Using Single Input Fuzzy Logic Controller and Modified Enhanced Perturb and Observe Mohammed Mokhtari1(&), Smail Zouggar1, Nacer K. M’sirdi2, and Mohamed Larbi Elhafyani1 1

School of Technology (L.E.E.M), University Mohammed 1st, 60000 Oujda, Morocco [email protected] 2 LSIS, CNRS, UMR 6168, Dom. Univ. St Jrme, Av. Escadrille Normandie - Niemen, 13397 Marseille, France

Abstract. This paper investigates the efficiency of an original approach for Maximum Power Point Tracking (MPPT) algorithm applied to a Wind Electric Water Pumping System (WEWPS). The studied model is developed under Matlab/Simulink software and consists of an asynchronous wind turbine, a Static Var Compensator (SVC) and a centrifugal water pump driven by a three phase Induction Motor (IM). The proposed control technique seeks to improve water flow rate by exploring the maximum amount of electrical power produced by the asynchronous wind turbine in a wide range of wind speed. Theoretical analysis as well as simulation results have shown that the highest electrical power rate depends on the value of the produced voltage which can be controlled by the SVC using single input fuzzy logic regulator. Modified Enhanced Perturb and Observe (MEPO) algorithm is then used in this application to calculate the optimal value of the voltage reference that ensure maximum electric power extraction and hence maximal water flow rate. Moreover, a comparison have been made with the conventional P&O algorithm to prove the superior performance of the proposed approach which does neither require to measure wind speed nor to know the WEWPS parameters. Keywords: Asynchronous wind turbine MEPO
SIFLC
SVC


Centrifugal water pump
MPPT


1 Introduction Wind energy sources are today an attractive alternative for electric power generation, it represent a competitive and promising renewable energy, especially, for energizing relatively small stand-alone systems in developing countries, where many people have no access to an AC power grid [8, 13, 15].

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 330–342, 2020. https://doi.org/10.1007/978-3-030-53187-4_37

Contribution to Power Maximization of an Asynchronous WEWPS

331

In the literature, several system for water pumping application have been studied using different topology, configurations and control strategy, most of them uses SEIG or PMSG or BDFIG etc. as generator and either DC or AC motor to drive the pump [4– 6, 14]. In the different cases, an MPPT control algorithm can be employed in order to capture the maximum power from available wind. A variety of MPPT techniques have been employed for Wind Energy Conversion System (WECS) such as Incremental and Conductance method (INC), Perturb and Observe (P&O) method, Fuzzy Logic Controller (FLC) and many Evolutionary Algorithms [9, 16]. In this paper, a modified enhanced MPPT algorithm is applied to the structure of the proposed wind electric water pumping system in order to maximize the water flow rate. This algorithm is simple, fast, efficient, and more importantly, does not require wind speed measurement. The present work is organized as follows: Sect. 2 presents the mathematical model of the studied system. Section 3 presents the MPPT techniques. Section 4 presents the simulation and its results with a discussion and Sect. 5 concludes this paper (Table 1). Table 1. Nomenclature. Parameters Vas ; Vbs ias ; ibs iar ; ibr iam ; ibm uar ; ubr

Stator terminal voltages in ab reference Stator phase current in ab reference Rotor phase current in ab reference Magnetizing current in ab reference Rotor flux in ab reference

uas ; ubs

Stator flux in ab reference

Rr ; Rs Lm ls ; lr xs V Cp ðkÞ Xturb

Per phase rotor and stator resistances Magnetizing reactance Stator and rotor leakage reactances Synchronous angular velocity Wind Speed Aerodynamic performance of the turbine Turbine shaft speed

2 System Description and Modeling The complete system shown in Fig. 1 can be divided into three major parts. The first one is the renewable power generator source that capture and transform wind power to electric power, it consists of a three blade wind turbine, a multiplier and an Induction Generator (IG) with its self-excitation capacitors. The second part consists of a Static Var Compensator (SVC) with its control strategy designed to maximize the electric power flow by controlling the produced voltage. Finally, the third part is the electric

332

M. Mokhtari et al.

water pumping system with its starter that will be explored to transfer water from a suction point to the storage tank to be used for irrigation or other needs [10, 12].

Fig. 1. Schematic diagram of the studied system

2.1

Wind Turbine Model

Several models of wind turbines have been developed and can be found throughout the bibliography [7, 11]. Since the electrical behavior of the system is our main point of interest the following model is assumed (Fig. 2):

Fig. 2. Schematic diagram of the turbine model.

2.2

Self-excited Induction Generator Model

Because of its high efficiency and less maintenance, the self-excited induction generator offers important advantages in construction and operation for standalone small wind turbines application [1]. The SEIG model developed in this work is represented in ab reference frame. Its equivalent circuit at a steady-state operation is shown in Fig. 3 with Lm been a function of the magnetizing current im [2].

Contribution to Power Maximization of an Asynchronous WEWPS

333

Fig. 3. Equivalent circuit of the SEIG in ab reference frame

The electromagnetic torque produced by the induction generator is defined as: Cem ¼ p:

2.3

 Lm
u :isb  urb isa Lr ra

ð1Þ

Water Pump Model

The chosen hydraulic system for this application is formed by a three phase induction motor to which a centrifugal pump is attached. Induction machines are widely used in water pumping activities because they are robust, inexpensive and easily replaceable in case of failure [3]. Hence, the IM is also modeled in ab reference frame: Vspa ¼ Rsp :ispa þ

duspa dt

Vspb ¼ Rs :ispb þ

duspb dt

Vrpa ¼ 0 ¼ Rrp :irpa þ

durpa  xrp :urpb dt

Vrpb ¼ 0 ¼ Rrp :irpb þ

durpb þ xrp :urpa dt

uspa ¼ Lsp :ispa þ Lmp :irpa

ð2Þ

334

M. Mokhtari et al.

uspb ¼ Lsp :ispb þ Lmp :irpb urpa ¼ Lmp :ispa þ Lrp :irpa urpb ¼ Lmp :ispb þ Lrp :irpb

ð3Þ

The dynamic equation of the mechanical motion can be given as: Jmp

dXmp ¼ Cmp  Crp dt

ð4Þ

Where Cmp is the produced mechanical torque by the IM: Cmp ¼ p:

 Lmp
urpa :ispb  urpb :ispa Lrp

ð5Þ

The pump power and its torque depends on the angular speed and can be expressed as: Pp ¼ kp  X3mp Crp ¼

Pp ¼ kp  X2mp Xmp

ð6Þ ð7Þ

Where kp is the centrifugal pump constant. The peak current of the electric water pump at transient phase when direct start is realized is higher than nominal current provided by the SEIG. Therefore a starter is used during starting phase to limit the starting current and avoid voltage collapse when the pump is connected to the asynchronous wind turbine. 2.4

Static Var Compensator Model

The Static Var Compensator FC-TCR type consists of a fixed capacitance C and an inductance L in series with a bi-directional thyristor valve that are fired symmetrically in an angle control range of 90° to 180°. At fundamental frequency, the total equivalent impedance of the SVC is shown in Fig. 4 and can be represented using the following expression [5]: XSVC ðaÞ ¼ Xc

p=rx

  sinð2aÞ  2a þ p 2  r1x

ð8Þ

Where rx ¼ X c =XL the limits of the compensator given by the firing angle limits and fixed by design.

Contribution to Power Maximization of an Asynchronous WEWPS

335

Fig. 4. Schematic diagram of the equivalent impedance of the FC-TCR

3 Proposed Control Strategy The control technique aims to optimize the generator voltage by means of SVC in order to maximize the water electric pump output power. Many strategies were investigated to achieve the MPPT. Two control methods are presented in this paper: the P&O and MEPO [8]. 3.1

Maximum Power Point Tracking Technique

To maximize the water flow rate, the rotational speed of the motor which drive the centrifugal pump must be maximized. That’s can be done by maximizing the power absorbed by the motor pump group. The MPPT process in the proposed system is based on directly adjusting the total excitation capacitor according to the result of the comparison of successive motor pump group input-power measurements. From the characteristics Fig. 5 and Fig. 6, we can see that our system operate in the points on the left, so the objective is to led the system to work in the yellow star which correspond to the maximum absorbed power and maximum produced AC voltage Vsp.

Fig. 5. Pelec_pump as function of Cext

336

M. Mokhtari et al.

Fig. 6. Vs_pump as function of Cext

The required Maximum Power Point to Track is so function of the voltage Vsp, the excitation capacitor Cext and can be defined by the following objective function: dPelec pump dPelec pump dVsp ¼ ¼0 dCext dVsp dCext

ð9Þ

Perturb and Observe Algorithm It’s one of the simplest MPPT techniques as it involves measurement of the power only. It is based on perturbing the voltage in small step and perceiving the resulting changes in power, as illustrated by Fig. 7, [5]. This algorithm is based on the following procedure: if the operating voltage of the WEWPS is perturbed in a given direction and if the power supplied by the generator increases, it means that the operating point has moved toward the MPP, and therefore the voltage of the IG must still be settled in the same direction. Otherwise, if power operated generator decreases, the operating point is far from the MPP and therefore the direction of the disturbance in the voltage of operation must be reversed. Additionally, selecting an appropriate step size is not a simple task: though larger step-size means a faster response and more oscillations around the peak point, and hence, less efficiency, a smaller step-size improves efficiency but decreases the convergence speed. MEPO: Modified Enhanced P&O Algorithm We propose, as a modified P&O Algorithm which will be more robust, the reference voltage given by: Vsp

ref

¼ Vsp

RMS

þ K:DPelec

pump :sign




DVsp



ð10Þ

In case of no change in the output power after perturbation: DPelec pump ¼ 0; then Vsp ref ¼ Vsp RMS In case of DPelec pump [ 0, the power increase after positive perturbation of DVsp then let us continue in the same direction. In case of DPelec pump \0, the power decreases after positive perturbation of DVsp then let us continue in the reverse direction. This method gives an enhanced variable step size algorithm. The step size is adjusted in proportionally to the power variation produced in the previous step. The

Contribution to Power Maximization of an Asynchronous WEWPS

337

step adjustment gain K is used for weighting this adjustment step. It may be useful for oscillation avoidance and noise sensitivity.

Fig. 7. Flowchart P&O algorithm

Fig. 8. MEPO algorithm Simulink implementation

3.2

Voltage Control

To meet the referred voltage Vsp_ref, Static Var Compensator FC-TCR type is used to adjust the value of the total excitation capacitor (Cext + Csvc) of the WT. The control of the produced voltage is done via a single input fuzzy logic regulator as shown in Fig. 9.

338

M. Mokhtari et al.

Fig. 9. Voltage control diagram

The per-phase thyristor firing angle a is calculated by adding the new FLC output change in da to the old one. The operation of this technique is explained as shown in Fig. 10.

Fig. 10. Voltage regulator

The proposed voltage controller is constructed by choosing the error between the reference and the measured RMS value of the generated voltage as an input signal and the thyristor firing angle as an output signal. a is defined by Eq. 11. It depends on a constant value a int ¼ 114:33 used initially to keep the value of Csvc equal to zero until the lunch of the MPPT algorithm. X a ¼ a int þ daðtÞ ð11Þ

• If Vs_RMS is equal to Vs_ref then error is equal to zero and da must also be equal to zero to keep a equal to a_int to avoid injecting or absorbing any reactive power. • If Vs_RMS is greater than Vs_ref (inductive mode) the error will be negative and da must be positive and its value should increase to absorb the excess of reactive power forcing the generated voltage to drop to the rated value.

Contribution to Power Maximization of an Asynchronous WEWPS

339

• If Vs_RMS is less than Vs_ref (capacitive mode) the error will be positive and da must be negative and its value should increase to inject the needed reactive power to reach the rated voltage value. The selected range for the input/output variables must be well chosen. Usually, it’s desirable to select the standard range of (±1). Yet, since the power system was well studied, the range for a variable is already known, and so the normalization step for the input was abandoned (Fig. 11).

Fig. 11. Input and output membership functions of the SFLC

The seven linguistic variables used are Negative Big (NB), Negative Medium (NM), Negative Small (NS), Zero (ZO), Positive Small (PS), Positive Medium (PM) and Positive Big (PB). The 7 rules have been built as represented in Fig. 12.

Fig. 12. The chosen rules of the SIFLC

4 Simulation Results and Analysis To evaluate the MPPT performance, the designed control system is simulated using MATLAB/Simulink software package and the obtained results will be discussed in this section. The wind speed variation profile is considered as depicted in Fig. 13. It can be devised to fourth zone. The first one is between 0 s and 3.3 s representing the average

340

M. Mokhtari et al.

wind speed that ensure the excitation of the IG. Zone 1 and 3 are the same and help to observe the behavior of the system when the wind speed rises above the rated wind speed (8 m/s). Zone 2 in the other side is to test the aero-generator at low wind speed value.

Fig. 13. Wind speed profile

At second 3.3 the MPPT control is lunched synchronously with the wind speed variation. From Fig. 14 we can observe that the absorbed power by the electric water group have been increased using both MPPT algorithm. However, the MEPO is showing better results especially when the wind speed get under the rated value. At the 7th second without MPPT control the Pelec_pump was at 590 W. Once the MEPO is applied the power rises up to 630 W unlike the P&O algorithm which have improved the power to only reach 600 W.

Fig. 14. Variation of the absorbed power by the motor pump with and without MPPT controllers

Contribution to Power Maximization of an Asynchronous WEWPS

341

The observed operating voltage of the wind electric water pumping system in Fig. 15 when P&O algorithm is used is lower than the one generated with MEPO algorithm, this is due to the inability of the MPPT algorithm to generate quickly the optimal value with a fixed step, which represent another drawback for the P&O method.

Fig. 15. Variation of the produced voltage Vs during power maximization

5 Conclusion This work has allowed us to complete our research on an autonomous wind electric water pumping system. The modeling of the studied system have been first presented, then a comparative analysis between P&O and MEPO algorithms has been discussed to support the proposed approach for power maximization that gives more efficient and reliable results at variable wind speed condition.

References 1. Mokhtari, M., Zouggar, S., Elhafyani, M.L., Ouchbel, T., Benzaouia, S., Fannakh, M.: Design, simulation and performance analysis of voltage regulator based on STATCOM for asynchronous wind turbine. In: International Conference on Electronic Engineering and Renewable Energy, pp. 498–509. Springer, Singapore, April 2018 2. Mokhtari, M., Zouggar, S., Elhafyani, M.L., Ouchbel, T., M’sirdi, N.K., Naaman, A.: Voltage stability improvement of an asynchronous wind turbine using static var compensator with single input fuzzy logic controller. In: 2018 6th International Renewable and Sustainable Energy Conference (IRSEC), pp. 1–6. IEEE, December 2018 3. Elhafyani, M.L., Zouggar, S., Benkaddour, M., Zidani, Y.: Permanent and dynamic behaviours of self-excited induction generator in balanced mode. Moroccan J. Condens. Matter 7 (2006)

342

M. Mokhtari et al.

4. Daili, Y., Gaubert, J.P., Rahmani, L.: Implementation of a new maximum power point tracking control strategy for small wind energy conversion systems without mechanical sensors. Energy Convers. Manag. 97, 298–306 (2015) 5. Ouchbel, T., Zouggar, S., Elhafyani, M.L., Seddik, M., Oukili, M., Aziz, A., Kadda, F.Z.: Power maximization of an asynchronous wind turbine with a variable speed feeding a centrifugal pump. Energy Convers. Manag. 78, 976–984 (2014) 6. Zeddini, M.A., Pusca, R., Sakly, A., Mimouni, M.F.: PSO-based MPPT control of winddriven self-excited induction generator for pumping system. Renew. Energy 95(162), 177 (2016) 7. Lara, D., Merino, G., Salazar, L.: Power converter with maximum power point tracking MPPT for small wind-electric pumping systems. Energy Convers. Manag. 97, 53–62 (2015) 8. Abdullah, M.A., et al.: A review of maximum power point tracking algorithms for wind energy systems. Renew. Sustain. Energy Rev. 16(5), 3220–3227 (2012) 9. Aubrée, R., Auger, F., Macé, M., Loron, L.: Design of an efficient small wind-energy conversion system with an adaptive sensorless MPPT strategy. Renew. Energy 86, 280–291 (2016) 10. Atawi, I.E., Kassem, A.M.: Optimal control based on maximum power point tracking (MPPT) of an autonomous hybrid photovoltaic/storage system in micro grid applications. Energies 10(5), 643 (2017) 11. Lahfaoui, B., Zouggar, S., Mohammed, B., Elhafyani, M.L.: Real time study of P&O MPPT control for small wind PMSG turbine systems using Arduino microcontroller. Energy Procedia 111, 1000–1009 (2017) 12. Fathabadi, H.: Novel high efficient speed sensorless controller for maximum power extraction from wind energy conversion systems. Energy Convers. Manag. 123, 392–401 (2016) 13. Tiwari, R., Ramesh Babu, N.: Fuzzy logic based MPPT for permanent magnet synchronous generator in wind energy conversion system. IFAC-PapersOnLine 49(1), 462–467 (2016) 14. Sefidgar, H., Asghar Gholamian, S.: Fuzzy logic control of wind turbine system connection to PM synchronous generator for maximum power point tracking. Int. J. Intell. Syst. Appl. 6 (7), 29 (2014) 15. Farhat, M., Barambones, O., Sbita, L.: Efficiency optimization of a DSP-based standalone PV system using a stable single input fuzzy logic controller. Renew. Sustain. Energy Rev. 49, 907–920 (2015) 16. Eltamaly, A.M., Farh, H.M.: Maximum power extraction from wind energy system based on fuzzy logic control. Electr. Power Syst. Res. 97, 144–150 (2013)

Hybrid System Energy Management in a Low Power Isolated Site Mohammed Larbi El Hafyani1(&), Abdelmalek El Elmehdi2, Smail Zouggar1, and Toufik Ouchbel1 Laboratory of Electrical Engineering and Maintenance – LEEM, ESTO, Oujda, Morocco [email protected], [email protected] Laboratory SmartICT, National School of Applied Sciences, Oujda, Morocco [email protected] 1

2

Abstract. Our study focuses on the problem of multi-source load management in a hybrid energy production system, photovoltaic/wind, associated with a storage system. The connection of these elements to the photovoltaic panel and the wind generator is performed at a DC voltage bus. This continuous bus has the advantage of interconnecting more easily the different elements of the hybrid system. This solution being the one adopted in this work. From the DC bus, the connection to the load is made using DC/DC power converters. The goal is to find a strategy for managing power exchanges between the different elements of the hybrid system. This strategy should optimize the overall performance of the system, properly utilize each source and load, and adapt to the configuration change. The system command should be distributed so that items can be removed without having to change the command policy. It is necessary to study the operation of hybrid power systems in extreme northern climates in order to optimize the efficiency of these systems once in use in these environments. Keywords: Renewable energy energy system (HES)


MPPT controller
Isolated site
Hybrid

1 Introduction Available in quantities greater than the current energy needs of humanity, renewable energy resources also represent an opportunity for more than two billion people, living in remote areas, to access electricity. These advantages, combined with increasingly efficient sectors, favor the development of renewable energies. For Morocco, the challenge of developing renewable energies is important. Indeed, these energies will increasingly cover the necessary and legitimate growth of basic energy services in the areas of rural development, housing, health, education and industry. For its geographical location, Morocco favours the development of the use of solar and wind energy, given its rate of sunshine (300 sunny days per year), while for © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 343–354, 2020. https://doi.org/10.1007/978-3-030-53187-4_38

344

M. L. El Hafyani et al.

the speed of the winds, the Moroccan territory, with Its 3.500 km of coastline, enjoys winds of an average of 9 m/s (meters per second), at 40 m height [1]. However, the production of electricity from only one of the two sources of energy mentioned above is very often limited by the final cost price of the electric KWh produced, because of the irregularity of the speed the wind and the discontinuity in time of the solar radiation which poses the problem of the storage of energy. However, this aspect of these energies is still today one of those which slow the most their development, because this storage, is often a big part of the investment of an installation of production of electrical energy from the wind or radiation solar. Considering their respective seasonal characteristics, these two energies do not compete with each other but on the contrary can be mutually valued. This is why we propose here a hybrid system composed of these two sources of energy, which consists in the optimal exploitation of the complementarily between them [2–4]. Thus, this complementarily of energy is supplemented by a storage system provided by lead batteries. The aim of this work is to develop an optimal energy management algorithm for autonomous hybrid renewable energy networks [3]. To achieve this goal, this work is divided into five paragraphs: After this introduction, where we exposed the problem and showed the need to integrate renewable energy sources, with a management strategy to feed an isolated energy site. The second paragraph is devoted to the presentation of the site and the load, and the solar and wind energy potentials. In the third paragraph, a configuration is chosen and a modelling of the different constituents of our system is given as well as the principle of energy management. Fourth paragraph summarizes the results of the implementation of the algorithms developed in the control systems of the hybrid energy system (SEH) under the Matlab/Simulink environment. Finally, a conclusion and perspectives summarize the work developed and outline the perspectives of actions to be considered in future work.

Fig. 1. Synoptic of topology studied

Hybrid System Energy Management in a Low Power Isolated Site

345

2 Sizing, Configuration and Modeling of HES Depending on the application and available energy sources, it is possible to integrate different energy sources, both renewable and conventional. In addition, the system can be easily expanded by adding electrical components or generators to meet increasing energy needs. By combining several sources of energy, the advantages of each of them are thus added: – – – –

The The The The

photovoltaic system reduces consumption and uptime. storage capacity of the batteries reduces the cost of the system. reliability of the overall system is increased. powers and energies involved are more important and the load to feed higher.

The following problems arise: – Correctly choose the size of each component of the energy system. – Optimize energy management within this system. – Look for the optimal configuration, i.e. the minimum production cost. The configuration of the HES obviously depends on the available energy resources as well as the constraints of the use. This requires a measurement companion and a preliminary analysis of site specificities. 2.1

Presentation of the Site and Characteristics of the Load

This autonomous system must allow supplying an isolated site with a daily average value of 200 W. For our study, we have chosen an isolated site in the outskirts of Oujda-Angad whose characteristics are given by the following meteorological variables Latitude Longitude Altitude

2.2

34.680 −1.910 50

Dimensioning of the Hybrid System

The design of the SEH requires the selection and sizing of the most appropriate combination of energy sources, converters and storage system as well as the implementation of an efficient operating strategy. Sizing software is an indispensable tool for analyzing and comparing the different possible combinations of sources used in the SEH. The main factors of sizing are [5, 6]: • The environmental conditions of the site (wind speed, irradiance, temperature, humidity); • The load profile; • Customer preferences and requests; • Financial resources; • Availability of technology and technical support;

346

M. L. El Hafyani et al.

a. The daily distribution of our load The load profile is shown in Fig. 2. This is a load with a daily average value of 200 W. b. Estimation of solar and wind energy resources In our study, the analysis of local resources (wind speed and radiation) requires a measurement companion and a preliminary analysis of site specificities. So to create a baseline for our study he chose to adopt relevant data of technical measures made during a day. • Wind potential To evaluate the wind potential of the site, measurements of wind speeds were carried out during one day, with an interval of one hour. 2.3

Solar Potential

Like wind speeds, measurements of solar radiation were made. To satisfy the load, the system will consist of two photovoltaic panels, a wind turbine and a battery of 210 Ah (Figs. 3 and 4).

350

Load Power (w)

300 250 200 150 100 50

0

5

10 15 T emps ( H)

20

25

Fig. 2. Load profile 1000 Irradiation ( W/m≤)

14

Wind ( m/s)

12 10 8 6 0

5

10

15

20

25

Temps(H)

Fig. 3. Representation of wind speeds

800 600 400 200 0

0

5

10

15

20

Tem ps (H)

Fig. 4. Solar irradiations

25

Hybrid System Energy Management in a Low Power Isolated Site

347

3 Modeling of the System SEH 3.1

Modeling of the Photovoltaic Array

Figure 5 shows the equivalent circuit of a photovoltaic cell under illumination. It corresponds to a current generator Ipv connected in parallel with a diode. Two parasitic resistances are introduced in this schema. These resistors have some influence on the characteristic I = f (V) of the cell: • The series resistance (Rs) is the internal resistance of the cell. • The shunt resistor (Rp) is due to a leakage current at the junction.

Fig. 5. Equivalent circuit of a solar cell

The mathematical model for the current-voltage characteristic of the ideal PV panel [7] is given by:     V þ Rs :I V þ IRs I ¼ Ipv  I0 exp 1  Ns Vt Rp

ð1Þ

Where Ipv and I0 are the photovoltaic (PV) and saturation currents, respectively, of the panel, Ns is the cells connected in series, Vt ¼ a:k:T =q is the thermal voltage of the panel, q is the electron charge (1.60217646  10−19 C), k is the Boltzmann constant (1.3806503  10−23 J/K), T(in Kelvin) is the temperature of the p-n junction and a is the diode ideality constant [8, 9]. The technical parameters of a photovoltaic panel used in this application are demonstrated in Table 1. Table 1. Parameters of the PV panel Parameters Maximum power rating Open circuit voltage (Voc) Short circuit current (Isc) Maximum power voltage (Vmp) Maximum power current (Imp) Number of Cells a

Value 180 W 30.4 V 8.03A 24.2 V 7.45A 50 1.2

348

M. L. El Hafyani et al.

The mathematical model of the photovoltaic panel with MPPT [10–12] allows obtaining a schematic block of SIMULINK. 3.2

Modeling of the Wind Turbine

A small wind turbine based on PMSG [13]; with the characteristics proposed by the manufacturer mentioned in Table 2; is installed in our Laboratory of Electrical Engineering and maintenance (LEEM), for the objective of modeling, simulating and studying the wind turbine behavior. Table 2. The characteristics of the wind turbine Start up wind speed 5 m/s (9.7 knots) 11 m/s (21.4 knots) 15 m/s(29.2knots) Bladespan Turning radius

5 knot (2.5 m/s) 14 W 108 W 180 910 mm 462

Based on the experimental data taken during the tests, the characteristic of our wind turbine (the power according to the voltage and according to the wind speed) is represented with the help of the Simulink Library tool “Look Table (2D)”. The Fig. 6 (curve model) illustrates the aerodynamic model of our wind turbine. It is advisable to also note that each curve presents a point of maximum power, which is the optimum point for the efficient use of the wind turbine. The optimum power as a function of the voltage follows the formal model represented by the equation below: Popt ¼ 0; 0021  U 3 þ 0; 048  U 2 þ 0; 21  U  0; 05

ð2Þ

Popt : optimum powerðWÞ; U : voltage ðVÞ: MPPT Curve (PV) of Wind turbine

140

150

12m/s

120

MPPT curve

curve model

v=12m/s 100

80 9m/s

60

7m/s

40

v=9m/s 50

v=7m/s

5.4m/s

20 0

v=10m/s

Power(W)

Power(W)

100

v=5.4m/s

0

10

20

30

40 50 Voltage(V)

60

70

80

Fig. 6. Curve of the aerodynamic model of the proposed wind turbine [14]

0

0

10

20

30

40 50 Voltage(V)

60

70

80

Fig. 7. MPPT Curve of the Wind turbine [14].

Hybrid System Energy Management in a Low Power Isolated Site

349

The Fig. 7 (MPPT curve) illustrates the points of maximum power as a function of the output voltage and the wind speed [14, 15]. To determine the electrical characteristics of the proposed wind turbine, we performed experimental tests in the wind tunnel. Several measurements were made in the laboratory by connecting the wind turbine to a load. For each wind speed (generated by the wind tunnel) and for each load value, they measured the voltage and current at the output of the generator. Based on the experimental data taken during the tests, the characteristic of the wind turbine (P (V)) is represented using the tool “Look Table (2D)” of the simulink library. 3.3

Simulation of the HES Under the MATLAB/SIMULINK Environment

The complete hybrid system consists of renewable energy sources (2 photovoltaic panels, and a wind turbine) a battery, a load and a management block (Fig. 1). In order to satisfy the load whatever the illumination and wind speed, a management block is necessary. Where Sl is the switch of the load, Sdel is the load shed switch, Sch is the switch of the battery charge and Sdch is the switch of the battery discharge.

Fig. 8. The energy management of the HES is as follows

Figure 8 shows the management algorithm developed under Simulink. The inputs of the block are the power from renewable energy sources and the state of charge of the battery, its outputs are the switches used to control the state of each device. The energy management strategy consists in optimizing the exploitation of the electrical energy of a hybrid multi-source network consisting of a photovoltaic

350

M. L. El Hafyani et al.

generator, a wind generator, a storage system and a load. This management must satisfy the load whatever the conditions. So the load is the element is the central element of the system [16–18].

4 Simulation of the Hybrid System for a Day Taking into account the sizing of the hybrid system, realized in Sect. 4, a simulation is carried out for one day with the climatic data given above. We give below the results obtained during the simulation: • Power of a photovoltaic panel 200 180 160

Ppv (W)

140 120 100 80 60 40 20 0

0

1

2

3 temps(s)

4

5

6 6

x 10

Fig. 9. Power of a photovoltaic panel

140

120

Pel(W)

100

80

60

40

20

0

0

1

2

3 temps(s)

4

5

6 6

x 10

Fig. 10. Wind power

• Power of renewable energy sources (Per) The power provided by the combination of two photovoltaic panels and a wind turbine takes the following form

Hybrid System Energy Management in a Low Power Isolated Site

351

Figure 11 gives the profile of the power provided by two photovoltaic panels and the turbine. The powers of a photovoltaic panel and the turbine are respectively represented in Figs. 9 and 10.

400 350 300

Per(W)

250 200 150 100 50 0

0

1

2

3 temps(s)

4

5

6 6

x 10

Fig. 11. Power of renewable energy sources (Per)

14

tension de bus dc (V)

12

10

8

6

4

2

0

0

1

2

3 temps(s)

4

5

6 6

x 10

Fig. 12. DC bus voltage

250

200

Pch (W)

150

100

50

0

0

1

2

3 temps(s)

Fig. 13. Load power

4

5

6 6

x 10

352

M. L. El Hafyani et al. 80 80 79.9999

SOC%

79.9999 79.9998 79.9998 79.9997 79.9997 79.9996 79.9996

0

1

2

3 temps(s)

4

5

6 6

x 10

Fig. 14. SOC of battery

250

Pdélestage (W)

200

150

100

50

0

0

1

2

3 temps(s)

4

5

6 6

x 10

Fig. 15. Power of load shedding

Figure 12 shows that the DC bus voltage still reaches the value of 12 V. It is also possible to satisfy the load demand (Fig. 13) thanks to the battery which discharges to t = 1.5 s since the power of the two solar panels and the turbine reaches 200 W requested by load (Fig. 14), and if there is an excess of power, a charge of load shedding will be satisfied after the charge of the battery (Fig. 15). Interpretation and Discussion The results clearly show that the management strategy makes it possible to satisfy the power demanded by the load for different values of radiation and wind speed. Indeed, it has been verified that the DC bus voltage is regulated at 12 V. The load is still satisfied (200 w), using only renewable energy sources (2PV wind turbine). The excess power produced (>200 w) will be used to charge the battery. If the power produced by renewable energy sources is not sufficient, the battery will be used to satisfy the load Once the power of the renewable sources goes to 240 W, the battery starts to charge since its state of charge is less than 80%.

Hybrid System Energy Management in a Low Power Isolated Site

353

5 Conclusion The works presented in this manuscript concern, in a general way, the study, the modelling and the simulation of electricity production systems, integrating mixed renewable resources (photovoltaic and wind), with a battery storage system using the Matlab software. In this context, we have introduced the different constituents of a SEH as well as their characteristics. Each component of the SEH is described with its mathematical model. Our objective was to optimize the energy management of a hybrid system, made up of renewable energy sources and a storage battery, and to size the installation in order to supply the load with its requested power, whatever the power requirements, and weather conditions. Indeed, one always check if the power provided by the HES is sufficient, one feeds the load, and the surplus is used to charge the battery if the SOC is lower than its minimum authorized. In the case where the SOC is superior to the SOCmax, the load is powered and the load shedding activated. When the power coming from HES is not sufficient to supply the load, we use the storage system (battery) if it is possible (SOC > SOCmax), otherwise we activate the delestage The results show that the management strategy achieves the objectives This work offers some perspectives that we present below: • Couplings of the hybrid energy system on the high-power grid. • Improving the efficiency of the hybrid energy system by introducing other renewable energy sources such as hydrogen • Introduce the notion of hybridization in storage systems to improve the efficiency of the hybrid energy system.

References 1. Ouammi, A., et al.: Artificial neural network analysis of Moroccan solar potential. Renew. Sustain. Energy Rev. 16(7), 4876–4889 (2012) 2. Lazarov, V.D., Notton, G., Zarkov, Z., et al.: Hybrid power systems with renewable energy sources types, structures, trends for research and development. In: Proceedings of International Conference on ELMA2005, Sofia, Bulgaria, pp. 515–520 (2005) 3. Kadda, F.Z., Zouggar, S., El Hafyani, M.: Contribution to the optimization of the electrical energy production from a hybrid renewable energy system. In: 5th IEEE, International Renewable Energy Congress (IREC) (2014) 4. Seddik, M., Zouggar, S., Ouchbel, T., Oukili, M., Rabhi, A., Aziz, A., Elhafyani, L.: A stand-alone system energy hybrid combining wind and photovoltaic with voltage control feedback loop voltage. Int. J. Electr. Eng. IJEET 6(2), 9–13 (2010) 5. Kadda, F.Z., Zouggar, S., Elhafyani, M.L.: Optimal sizing of an autonomous hybrid system. In: 1st IEEE, The International Renewable and Sustainable Energy Conference, pp. 269–274 (2013)

354

M. L. El Hafyani et al.

6. Zahboune, H., Zouggar, S., Krajacic, G., Varbanov, P.S., Elhafyani, M., Ziani, E.: Optimal hybrid renewable energy design in autonomous system using Modified Electric System Cascade Analysis and Homer software. Energy Convers. Manag. 126, 909–922 (2016) 7. Villalva, M.G., Gazoli, J.R., Ruppert Filho, E.: Comprehensive approach to modeling and simulation of photovoltaic arrays. IEEE Trans. Power Electron. 24(5), 1198–1208 (2009) 8. Sera, D., Teodorescu, R., Rodriguez, P.: PV panel model based on datasheet values. In: 2007 IEEE International Symposium on Industrial Electronics. ISIE 2007. IEEE (2007) 9. Fannakh, M.E., Elhafyani, M.L., Zouggar, S.: Fuzzy logic approach to improve the performances of grid-connected photovoltaic power system. In: Proceedings of 2018 6th International Renewable and Sustainable Energy Conference. IRSEC 2018, 29 April 2019. https://doi.org/10.1109/irsec.2018.8703028 10. Chang, Y.-H., Chang, C.-Y.: A maximum power point tracking of PV system by scaling fuzzy control. In: 2010 Proceeding of the International MultiConference of Engineers and Computer Scientist. IMECS 2010, Hong Kong, 17–19 March 2010, vol. II (2010) 11. Fannakh, M., Elhafyani, M.L., Zouggar, S.: Hardware implementation of the fuzzy logic MPPT in an Arduino card using a Simulink support package for PV application. IET Renew. Power Gener. 13(3), 510–518 (2019). https://doi.org/10.1049/iet-rpg.2018.5667 12. Seddik, M., Zouggar, S., Lahfaoui, B., Elhafyani, M., Aziz, A.: The new architecture of a BUCK/BOOST SHUNT converter non-inverter dedicated to the wind turbine system with a high efficiency. In: Proceedings of 2014 International Renewable and Sustainable Energy Conference. IRSEC 2014 (2014). https://doi.org/10.1109/irsec.2014.7059898 13. http://www.marlec.co.uk/?s 14. Badreddine, L., Zouggar, S., Elhafyani, M.L., Kadda, F.Z.: Experimental Modeling and Control of a Small Wind PMSG Turbine, pp. 978–984. IEEE (2014) 15. Koutroulis, E., Kolokotsa, D., Potirakis, A., Kalaitzakis, K.: Methodology for optimal sizing of stand-alone photovoltaic/wind-generator systems using genetic algorithms. Solar Energy 80, 1072–1088 (2006) 16. Kadda, F.Z., Zouggar, S., El Hafyani, M.: Optimization of the managed electrical energy within a hybrid renewable energy system. Int. J. Adv. Eng. Manag. Sci. 2(4) (2016)

Machine Learning, Intelligent Systems and Applications

Citation Classification Using Natural Language Processing and Machine Learning Models Syyab Rahi1 , Iqra Safder1 , Sehrish Iqbal1 , Saeed-Ul Hassan1(&) , Iain Reid2 , and Raheel Nawaz2 1

Information Technology University, Ferozepur Road, Lahore, Pakistan 2 Manchester Metropolitan University, Manchester M15 6BH, UK [email protected]

Abstract. In this paper, we address the problem of identifying the quality of citation as important or unimportant to the developments presented in the research papers. We gather features represented by four state-of-the-art machine learning techniques and combined them with newly engineered, natural language-based features. Using a known dataset of 465 citations, manually labeled by experts, our approach out-performed state-of-the-art by using fine-tuned Random Forest Classifier with 90.7% F1 score and 97.7% precision. We also employ Convolutional Neural Networks with AdamW optimizer with focal loss function - that converges quickly on small data to achieve considerably significant results. Keywords: Citation classification
Machine learning
Deep learning
Natural language processing

1 Introduction The publication databases consist of large knowledge-base that helps to mine scientific structures and citation-based impact of research [1–3]. More recently, the scientific community has been focusing on better understanding the citations semantics by tapping the power of full-text publication corpora [4–6]. We believe that citations context play a vital role in indicating the purpose of citations, helping in classifying them into important and unimportant citations to the developments presented in the research papers. Moravcsik and Murugesan [7], discovered that almost 40% of the citations in their corpus of articles was only perfunctory i.e. just general acknowledgment of others work. Since more than half of the citations are insignificant, this enlightens the importance of citation context. Various annotation schemes are devised to judge whether a citation is important or just incidental. Abu-Jbara, et al. [8] has proposed a set of features which are very useful in determining the context of a citation and its classification. They used classic supervised techniques of machine learning, like random forests and support vector machines, to identify important and unimportant citations. The following are two main contributions of this paper. At first, we improve on the classification accuracy of the existing state-of-the-art [9] citation classification model © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 A. El Moussati et al. (Eds.): SmartICT 2019, LNEE 684, pp. 357–365, 2020. https://doi.org/10.1007/978-3-030-53187-4_39

358

S. Rahi et al.

by employing Natural Languages Processing (NLP) techniques - more specifically a novel language dependence model. The second and the foremost important contribution is the implication of deep learning models that solely seek to use the textual information of citation in full-text for the classification task. The rest of the paper has been organized as follows. The Sect. 2 presents a recent review on important studies in the area of automatic citation classification. The Sect. 3 presents data and employed methods following by the results sections. Finally, we present conclusions along with future research directions.

2 Literature Review Teufel et al. [10] were among the pioneers to propose automated techniques for the citation classification task. Later, Abu-Jbara et al. [8] extended their model with an additional work of identifying citation context, classifying citations, and performing sentiment analysis of citations. This required features which were on context level as well as on polarity level. Moreover, they utilized a famous machine learning algorithm, SVM, for classification with 10-fold cross-validation and achieved 81.4% accuracy. Their research highlighted that including citation context can improve results, also their designed classification technique performed better than existing classification techniques. Xu et al. [11] introduced three labels classification: functional, perfunctory and ambiguous. Cue patterns, positional features, structural features, and network-based features were used for measuring the relation of author and article. Ding et al. [12] suggested a technique for identifying significant citation references, those which are mentioned to use or extended the research work. Overall, the citations were categorized into four categories: related work, comparison, using the work, and extending the work. They employed supervised machine learning techniques for classification, using SVM and Random Forest (RF), with 3-fold cross-validation and achieved an accuracy score of 80%. Zhu et al. [13] suggested a model for citation identification with central academic influences. Their approach consisted of supervised machine learning which predicted the academic influence by using features based on: count, similarity, context, and position. However, the features based on count appeared promising among all by giving the best results. More recently, Hassan et al. [14] extended features proposed by Valenzuela et al. [15] that classify citations into important and unimportant classes. The newly designed features appeared as important features for the task of citation classification. They employed SVM, KNN, NB, Decision Tree, and RF as classification models. Note that RF was the best performing model among all which outperformed the Valenzuela’s model with 84% AUCPR.

Citation Classification Using Natural Language Processing

359

3 Data and Methodology 3.1

Dataset

The dataset under consideration is manually labeled and open data set of Association for Computational Linguistics (ACL) [15]. The ACL anthology consists of more than 20,000 research papers. These research papers further include *100000 citations from which, 450 citations were randomly selected and then annotated. The dataset was annotated into binary class, “0” and “1”, labels. Where “0” is the class label for “Incidental/Unimportant” citation and “1” represents the class label for “Important” class. The data annotation was performed by a domain expert, which were then verified by inter-annotator agreement between two domain experts for a subset of the dataset. The agreement between experts was 93.9% [15]. 3.2

Dependency Parsing

The syntactic information was obtained with the help of dependency parsing. A dependency parser was used to extract word dependencies of the citing sentence, along with the citing context, to create the syntactic understanding of the citation. A dependency parser is a function for extracting structure of sentence. In other words, a dependency parser analyses the grammatical structure of a sentence, creating the exact relation of headwords and the words which change, alter, or modify their head. The result of dependency parsing is shown in Fig. 1.

Fig. 1. Dependency parsing of citing sentence

3.3

Employed Approaches

In this paper, we mainly employed two techniques: Classical Machine Learning, and Deep Learning. Furthermore, in the classical machine learning, we experimented with various arrangements of features and their combinations. Similarly, we tried a few deep neural networks for the given task of citation classification. The details of all the experiments are explained later in this section. The results, evaluations, and outcomes of these experiments will be explained in detail in the later sections. Classical Machine Learning Model. In order to improve the classification accuracy, we employed RF model for the task of classification. Note that RF has shown the best classification results in the existing studies [9]. Therefore, we carried out our classical

360

S. Rahi et al.

machine learning experiments only with RF model. We performed various experiments by modifying the arrangements of dependent features. The first arrangement is named as “previous features only” i.e. the features created by Teufel et al. [10], Abu-Jbara et al. [8], Valenzuela et al. [15] and Hassan et al. [14]. In the second arrangement, the previous features were used along with the text of citation context i.e. one sentence before citing sentence and two sentences after citing sentence. The third experiment consists of more improved features. We additionally used the technique of dependency parsing for the third experiment. The concept of dependency parsing is very important concept in the field of NLP and has been extensively used for text classification from last many years. Hence, we incorporated feature of words, from citation context, concatenated with their dependencies along with previous features and citation context. Deep Learning Model. We have utilized the Convolutional Neural Network (CNN) for our citation classification problem. The CNN’s are constructed from neurons that have weights and biases which can be learned during the training phase. Each neuron performs dot product of the inputs and transforms the input into non-linear form. This transformation towards non-linearity is optional. This is followed by maxpooling layers. The CNN has a fully connected layer as an output layer with some loss function. CNN’s have proven to be successful in various text classification problems. A simple convolutional neural network with the tuning of hyperparameters and static vectors can beat the results of existing benchmark techniques – improving the state-ofthe-art on 4 out of 7 tasks [16, 17]. We used the convolutional neural network with slightly different parameters than basic or conventional usage. The used AdamW optimizer for this specific case because it has a great tendency to converge quickly with small training data with self-balancing loss function, called focal loss [18].

4 Results and Discussion For all the experiment arrangements mentioned in the previous section, the data was split, randomly, into two sets; training set and testing set, 70% for training data and 30% for testing data. For the evaluation purpose, we used ROC and PR curves as evaluation metrics. 4.1

Training and Testing Data

In order to design supervised machine learning algorithms, we need annotated data for training and testing. We used manually annotated open data from the work of Valenzuela et al. [15]. This data has a total of 465 citations with important or unimportant labels annotated by experts. The distribution of data is 85.4% unimportant class whereas 14.6% belong to the important class. This data has a significant class imbalance, hence limited algorithms can be used which are not affected by the problem of class imbalance. We used this dataset for the experimentation of employed machine learning and deep learning models.

Citation Classification Using Natural Language Processing

4.2

361

Experiments

We performed three experiments; classical machine learning, an experiment of language model and an experiment of deep learning. Furthermore, we computed ROC curve, PR curve, learning curve, loss curve, and F1-score as evaluation measures for the comparisons among the above-mentioned experimental arrangements. In this study, we mainly focused on identifying important citations using a bundle of features. Generally, in classical machine learning the first jump of evaluation results, from baseline results, is easily achieved. However, further improvements in results occur in small steps that require intense and highly sophisticated feature engineering and algorithmic fine tuning. In our work, we carefully curated features and fine-tuned model to outperform the results of state-of-the-art techniques [14]. Machine Learning Model. We performed machine learning experiments in three arrangements: Previous features only, Previous features with citation context, and previous features with citation context and dependency parsing while using RF as classifier. The reason for using this model is that this model operates on ensemble method techniques, removing the need for performing cross-validation. The classification report for all the machine learning experimental arrangements is shown in Table 1. It is evident that precision and F-score is improved from Hassan’s [14] results. However, the recall is slightly lesser. Table 1. Classification Report Comparison with state-of-the-art Hassan et al. [14] Experiment Precision Hassan et al. [14] 0.89 Known features + Citation Context + Dependency Parsing 0.977 CNN [Citation Context only] 0.731

Recall 0.89 0.861 0.737

F1-score 0.89 0.907 0.732

The PR curves of all the three machine learning experimental setups are demonstrated in Fig. 2. We have observed a significant improvement in AUCPR from 0.91 to 0.97, with different experimental setups. Results clearly indicate that “Known features + Citation Context + Dependency Parsing” has outperformed all the existing techniques in terms of AUCPR. The AUCPR for only 64 previous features [14], was 0.91. However, when citation context was also included with 64 features, the AUCPR jumped to 0.94 which is slightly higher than achieved by state-of-the-art technique [14]. Furthermore, with the addition of dependency parsing along with citation context and 64 features, the AUCPR reached upto 0.97. Figure 3 presented the ROC curve for our designed model (Known features + Citation Context + Dependency Parsing) with RF classifier. We achieved AUROC 0.98, higher the 0.90 AUROC obtained by Hassan et al. [14]. Evidently, our approach has improved the receiver operating characteristic curve from state-of-the-art as well. Deep Learning Model Results. The deep learning experiments could not produce better results than machine learning experiments. However, although the amount of data was not large enough to implement deep learning models with simple

362

S. Rahi et al.

Fig. 2. PR curves for all models using RF classifier.

configurations and hyper-tuning parameters. We implemented a CNN with parameters conditional to our data set and attained promising results. Our model obtained 0.731 precision with 0.737 recall and 0.732 F-score. We believe that provided more amounts of data, this customized CNN can produce pretty attractive results. Figure 4 shows the learning loss curves for our CNN model. The sub-figure on the left is the learning curve of CNN, showing the trend of training and testing accuracy with respect to epochs. The model achieved 0.95 training accuracy on 50th epochs whereas, the test accuracy remains lower than training accuracy with 0.73 accuracy on the almost 33rd epoch. The sub-figure on the right is the trend of training and validation loss. It is observed that the training loss reached the lowest value of less than 1 at around 55th epoch. Whereas the testing loss showed less reduction in loss and could only reduce