124 45 132MB
English Pages 841 [839] Year 2022
Lecture Notes in Electrical Engineering 894
Saad Mekhilef Rabindra Nath Shaw Pierluigi Siano Editors
Innovations in Electrical and Electronic Engineering Proceedings of ICEEE 2022, Volume 2
Lecture Notes in Electrical Engineering Volume 894
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 Yong Li, Hunan University, Changsha, Hunan, China 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 Luca Oneto, Dept. of Informatics, Bioengg., Robotics, University of Genova, Genova, Genova, Italy 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 Walter Zamboni, DIEM - Università degli studi di Salerno, Fisciano, Salerno, Italy 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, Editor ([email protected]) India, Japan, Rest of Asia Swati Meherishi, Editorial Director ([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]) ** This series is indexed by EI Compendex and Scopus databases. **
More information about this series at https://link.springer.com/bookseries/7818
Saad Mekhilef Rabindra Nath Shaw Pierluigi Siano •
•
Editors
Innovations in Electrical and Electronic Engineering Proceedings of ICEEE 2022, Volume 2
123
Editors Saad Mekhilef School of Science, Computing and Engineering Technologies Swinburne University of Technology Hawthorn, VIC, Australia
Rabindra Nath Shaw Office of the International Relations Bharath Institute of Higher Education and Research Chennai, India
Pierluigi Siano Department of Management and Innovation Systems University of Salerno Fisciano, Salerno, Italy
ISSN 1876-1100 ISSN 1876-1119 (electronic) Lecture Notes in Electrical Engineering ISBN 978-981-19-1676-2 ISBN 978-981-19-1677-9 (eBook) https://doi.org/10.1007/978-981-19-1677-9 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 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 Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore
Preface
The book features selected high-quality papers presented at the International Conference on Electrical and Electronics Engineering (ICEEE 2022), jointly organized by University of Malaya and Bharath Institute of Higher Education and Research, India, during January 8–9, 2022, through Webex. The book focuses on current development in the fields of electrical and electronics engineering. The volume one covers electrical engineering topics—power and energy including renewable energy, power electronics and applications, control, and automation and instrumentation, and volume two covers the areas of robotics, artificial intelligence and IoT, electronics devices, circuits and systems, wireless and optical communication, RF and microwaves, VLSI, and signal processing. The book is beneficial for readers from both academia and industry. We are thankful to all the authors that have submitted papers for keeping the quality of ICEEE 2022 at high levels. The editors of this book would like to acknowledge all the authors for their contributions and the reviewers. We have received invaluable help from the members of the International Program Committee and the chairs responsible for different aspects of the workshop. We also appreciate the role of Special Sessions Organizers. Thanks to all of them, we had been able to collect many papers on interesting topics, and during the conference, we had very interesting presentations and stimulating discussions. We hope that the volume will provide useful information to professors, researchers, and graduated students in the area of electrical and electronics engineering and technologies along with AI and IoT applications, and all will find this collection of papers inspiring, informative, and useful. We also hope to see you at a future ICEEE event. Saad Mekhilef Rabindra Nath Shaw Pierluigi Siano
v
ICEEE 2022 Organization
Patron Bhim Singh R. Venkatesh Babu (Pro Vice Chancellor)
Indian Institute of Technology Delhi, India BIHER, India
General Chair Saad Mekhilef
Swinburne University of Technology, School of Science, Computing and Engineering Technologies, Australia
Conference Chair and Chairman, Oversight Committee Rabindra Nath Shaw
Director International Relations, BIHER, India
Conference Secretary Saravanan D.
Galgotias University, India
Technical Chairs Nishad Mendis Ankush Ghosh
Det Norske Veritas, Australia The Neotia University, India
Publication Chairs Valentina E. Balas Sanjoy Das
Aurel Vlaicu University of Arad, Romania IGNTU, Manipur, India
vii
viii
ICEEE 2022 Organization
Springer/ICEEE Liaison Aninda Bose (Senior Editor)
Springer Nature, India
International Advisory Board Vincenzo Piuri Georges Zissis (President) Lakhmi C. Jain Tamas Ruzsanyi Valentina Balas N. R. Pal (President) George T. Yen-Wei Chen Milan Simic M. Paprzycki Maria Virvou D. P. Kothari (Ex Director) B. K. Panigrahi C. Boccaletti Akshay Kumar
University of Milan, Italy IEEE IAS University of Technology, Sydney Ganz-Skoda Electric Ltd., Hungary University of Arad, Romania IEEE CIS University of Piraeus, Greece Ritsumeikan University, Japan RMIT University, Australia Polish Academy of Sciences University of Piraeus, Greece IIT Delhi, India IIT Delhi, India Sapienza University, Italy Concordia University, Canada
Contents
Simulation of Single-Phase Cascaded H-Bridge Multilevel Inverter with Non-isolated Converter for Solar Photovoltaic System . . . . . . . . . . C. Dinakaran and T. Padmavathi Intrusion Detection Based on PCA with Improved K-Means . . . . . . . . . Pralhad Chapagain, Arun Timalsina, Mohan Bhandari, and Roshan Chitrakar Small Signal Stability Improvement of Phillips Heffron Model with Fuzzy Based Control and Its Comparison with PSS . . . . . . . . . . . . Venu Yarlagadda, G. Sasi Kumar, R. Geshma Kumari, and Giriprasad Ambati Computer Vision Approach for Visibility Enhancement of Dull Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chandrika Acharjee and Suman Deb Tomato Leaf Diseases Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vishal Seth, Rajeev Paulus, Mayur Kumar, and Anil Kumar
1 13
28
41 53
Comparing Switching Losses for Automatic and Program-Based Switched Capacitor Model for Active Cell Balancing . . . . . . . . . . . . . . . Vipin Valsan and Gairik Das
62
Control System Design for DC-Link Voltage Control of Quasi-Z Source Inverter Using Bode Diagram . . . . . . . . . . . . . . . . . . Piyush B. Miyani and Amit V. Sant
72
Design and Simulation of Single Phase and Three Phase Wireless Power Transfer in Electric Vehicle Using MATLAB/Simulink . . . . . . . . Tareq Anwar Shikdar, Shornalee Dey, Sadia Mumtahina, Md Moontasir Rashid, and Gulam Mahfuz Chowdhury
83
ix
x
Contents
Automatic Cut Detection: A Probability Based Approach . . . . . . . . . . . 105 Tejaswini Kar, P. Kanungo, and Vinod Jha A Novel Design of T Shaped Patch Antenna Integrated with AMC for WLAN Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 Priyanka Das Neural Network: A Way to Know Consumer Satisfaction During Voice Call . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Divya Mishra and Suyash Mishra Automatic Access Application of ICG and ECG Signals . . . . . . . . . . . . 132 Hadjer Benabdallah and Salim Kerai OntoSpammer: A Two-Source Ontology-Based Spam Detection Using Bagging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 Dev Agrawal and Gerard Deepak SAODFT: Socially Aware Ontology Driven Approach for Query Facet Generation in Text Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 Rituraj Ojha and Gerard Deepak SDPO: An Approach Towards Software Defect Prediction Using Ontology Driven Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 N. Manoj and Gerard Deepak Ubiquitous Control Structure and It’s Comprehensive Application Subject to Controller Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 Ravishankar P. Desai and Narayan S. Manjarekar PCA-Based Random Forest Classifier for Speech Emotion Recognition Using FFTF Features, Jitter, and Shimmer . . . . . . . . . . . . 194 Sheetal Patil and G. K. Kharate An Adaptive and Coordinated Traffic Signal Scheme for Greener Transport 4.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 W. A. G. Weerasundara, D. P. D. Udugahapattuwa, T. D. Munasingha, W. D. K. Gunathilake, and Eng.S. U. Dampage Transient Stability Analysis of IEEE 9 Bus System Using Modified Fourier Filter Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 Dhaval N. Tailor and Vijay H. Makwana District Level Analytical Study of Infant Malnutrition in Madhya Pradesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 Supriya Vanahalli, Sarmista Biswas, Jossy P. George, and Samiksha Shukla Modelling of Microwave Absorption in Pervoskite Based Compounds Using Machine Learning Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248 Aayushi Arya and G. V. V. Sharma
Contents
xi
A Transparent Broadband Millimeter-Wave Absorber for Ka to V Bands Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254 Anupriya Choudhary, Sneha Tiwari, Srikanta Pal, and Gautam Sarkhel AC Analysis on the Modified Johnson Converter . . . . . . . . . . . . . . . . . . 261 Shubham Agrawal, L. Umanand, and B. Subba Reddy Heating and Cooling of Thermoelectric Module Using Modified Johnson Converter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 Shubham Agrawal, L. Umanand, B. Subba Reddy, and Atmanandmaya Hate Detection for Social Media Text with User Alert System . . . . . . . . 279 Jose Ashley, Nefi Nisen, Riyona Lasrado, and Mukta Nivelkar Dual-Band Antenna Design at Sub-6 GHz and Millimeter Wave Band for 5G Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290 Manish Kumar Dabhade and Krishna K. Warhade Bond Graph Modelling and Simulation of Active Heating/Cooling Using Thermoelectric Energy Conversion Systems . . . . . . . . . . . . . . . . . 299 Gautam A. Raiker, B. Subba Reddy, and L. Umanand Deep Learning Based Fruit Defect Detection System . . . . . . . . . . . . . . . 306 Anshul Bhardwaj, Nitasha Hasteer, Yogesh Kumar, and Yogesh Design of 1 kVA Shunt Compensator Test Bench . . . . . . . . . . . . . . . . . 314 Anchal Singh Thakur, L. Umanand, and B. Subba Reddy Real-Time Face Mask Detection Using CNN in Covid-19 Aspect . . . . . . 327 D. L. Kavya Reddy, Kanishka Negi, D. R. Soumya, Gaddam Prathik Kumar, Subrata Sahana, and Anil Kumar Sagar A Smart IoT Based Agro-Management System Using Distributed Ledger and Circular Supply Chain Methodology . . . . . . . . . . . . . . . . . . 345 M. Darshan, Sundeep V. V. S. Akella, S. R. Raswanth, and Priyanka Kumar A Sub-multilevel Topology of 7-Level Cascaded H-Bridge Inverter . . . . 355 Sushree Smrutimayee Barah, Swaraj Pati, and Sasmita Behera A Comparative Harmonic Suppression Analysis of Single Phase Half Bridge Grid Connected Inverter in Rotatory Frame of Reference with Integral Control Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365 Hassan Khalid, Saad Mekhilef, and Marizan Binti Mubin Use of Openflow to Manage Network Devices . . . . . . . . . . . . . . . . . . . . 376 Yogesh Kumar Chekoory and Avinash Utam Mungur
xii
Contents
Intelligent Routing Protocol for Energy Efficient Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387 Nandoori Srikanth, Battula Ashok, Battula Chandini, Mohan Kumar Chandole, and Naga Jyothi Voltage-Lift-Type Z-Source Inverter . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 Vadthya Jagan and Udutha Prashanth Smart Energy Management for a Campus Network . . . . . . . . . . . . . . . 411 B. Subba Reddy, Febin Francis, and L. Umanand Design and Development of a New Cascaded Asymmetric 15-Level Multilevel Inverter Topology with Reduced Number of Switches . . . . . . 419 M. Revathi and K. Rama Sudha A Novel Deep Learning Based Nepali Speech Recognition . . . . . . . . . . . 433 Basanta Joshi, Bharat Bhatta, Sanjeeb Prasad Panday, and Ram Krishna Maharjan Breast Cancer Detection and Prediction Based on Conflicts in Fractal Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444 Krishna Kumar Singh, Deepmala Jasuja, and M. P. Singh A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter (2ph-IBDB2C) for Battery Storage Applications . . . . . . . . . . 456 K. L. Lijin, S. Sheik Mohammed, and P. P. Muhammed Shanir A Comparative Study of Concentrated Photovoltaic ThermalThermoelectric Generator Hybrid Systems . . . . . . . . . . . . . . . . . . . . . . 469 Abhishek Tiwari and Shruti Aggarwal Real Time Exhaustion Detection by Image Classification Using Deep Convolution Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482 Anjani gupta, Prashant Singh, Dhyanendra Jain, Amit Kumar Pandey, Ashu Jain, and Gaurav Sharma A Secured BlockChain Based Facial Recognition System for Two Factor Authentication Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492 M. Darshan, S. R. Raswanth, S. Skandan, S. Shakthi Saravanan, Ranjit Chandramohanan, and Priyanka Kumar Design of an Optimised, Low Cost, Contactless Thermometer with Distance Compensation for Rapid Body Temperature Scanning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503 Pratik Bhattacharjee, Suparna Biswas, and Sandip Roy Sensitivity and Performance Analysis of 10 MW Solar Power Plant Using MPPT Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512 Md Fahim Ansari and Anis Afzal
Contents
xiii
SAR Reduction of Wearable Antenna Using Multiple Slots and AMC Combination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519 Sonali Kerketta, Azharuddin Khan, and Amit Kumar Singh Secure Asset Tracking Using GPS, MQTT, RFID, GSM and Cloud Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528 Jai Garg, Aditya Agrawal, Vansh Gupta, and Surender Dhiman Statistical Oversampling Classification Based Glass Type Identification Through Oxide Content . . . . . . . . . . . . . . . . . . . . . . . . . . 537 M. Shyamala Devi, R. Aruna, S. Vinoth Kumar, G. Vamsi Chowdary, B. V. S. S. Kanaka Raju, and M. Siva Prasad Coordinated Planning of DG and D-STATCOM in Distribution System Considering Polynomial Load Models . . . . . . . . . . . . . . . . . . . . 551 P. Siva Prasad and M. Sushama CRaaS: Cyber Range as a Service . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565 K. S. Srinivas, M. Suhas, P. Srinath, K. C. Sneha, D. G. Narayan, and P. Somashekhar ECG Based Biometric Recognition Using Similarity Measure and Feature Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577 Aditya Vikram Singh, Ishita A. Kumar, and Malaya Kumar Hota Improved Radial Basis Function (RBF) Classifier for Power Transformer Winding Fault Classification . . . . . . . . . . . . . . . . . . . . . . . 587 Sobhana Obulareddy and Surya Kalavathi Munagala Distribution Network Reconfiguration and Capacitor Allocation in Distribution System Using Discrete Improved Grey Wolf Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602 P. Siva Prasad and M. Sushama Machine Learning Algorithms for Predicting the Graduation Admission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618 Krishna Mridha, Suman Jha, Bikash Shah, P. Damodharan, Ankush Ghosh, and Rabindra Nath Shaw Network Load Balancing for Edge-Cloud Continuum Ecosystems . . . . . 638 Andrzej Paszkiewicz, Marek Bolanowski, Cezary Ćwikła, Maria Ganzha, Marcin Paprzycki, Carlos E. Palau, and Ignacio Lacalle Úbeda CNN Based Image Forgery Segmentation and Classification for Forensic Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652 B. Hemalatha, B. Karthik, S. Balaji, K. K. Senthilkumar, and Ankush Ghosh
xiv
Contents
A Novel Approach for Blind - Image to Audio Conversion in Regional Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662 B. Hemalatha, B. Karthik, S. Balaji, G. Vijayalakshmi, and Rabindra Nath Shaw Genetic Fuzzy Based VANET Routing Algorithm for Better Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669 M. Sangeetha, T. Vijayan, N. Aashna Unnikrishnan, G. Kalanandhini, and Rabindra Nath Shaw Design of Reconfigurable Ku-Band Band-Pass Filter for High Performance RF Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 678 H. Umma Habiba, Molimoli Hajamohideen Masood, and Rabindra Nath Shaw Design and Analysis of X-Band Radar Antenna for Self-powered Sensor Application in Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687 H. Umma Habiba, S. Revathy, Rabindra Nath Shaw, and Ankush Ghosh Multiple EBG Cells Based Filtering Antenna for High Frequecny Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691 H. Umma Habiba, M. H. Masood, Rabindra Nath Shaw, and Ankush Ghosh Design of Low Dimensional SIW Bandpass Filter Based on DGS and DMS Technologies for Radar and Satellite Communications . . . . . 702 H. Umma Habiba, A. Banu Priya, Rabindra Nath Shaw, and Ankush Ghosh Studies on Steganography Images and Videos Using Deep Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707 P. Sathish Kumar, K. Fathima, B. Karthik, S. Siva Kumar, B. Sowmya, and Ankush Ghosh Design and Implementation of a Defect Identification Using Image Processing Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734 J. Dhanasekar, A. P. Sharan, M. A. Natarajan, A. Nizamudeen, A. H. Methil Krishnan, and S. R. Senthil Kumar Design and Simulation of a Robotic Manipulator for Ladle with PLC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 752 V. Priya, V. J. Aravinth Raj, K. V. Chethanasai, J. P. Manoj Kumar, V. Manikandan, and K. K. Senthilkumar Design and Implementation of Automatic Goggle Detector for Safety Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 766 V. Balambica, T. R. Vijayaram, M. Achudhan, Vishwa Deepak, and Manikandan Ganesan
Contents
xv
Machine Learning Based Approaches in the Detection of Parkinson’s Disease – A Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774 Arpan Adhikary, Koushik Majumder, Santanu Chatterjee, Rabindra Nath Shaw, and Ankush Ghosh Tactile Internet in Internet of Things Ecosystems . . . . . . . . . . . . . . . . . 794 Ignacio Lacalle, César López, Rafael Vaño, Carlos E. Palau, Manuel Esteve, Maria Ganzha, Marcin Paprzycki, and Paweł Szmeja Artificial Neural Network Based Biometric Palm Print Recognition System for Security Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808 M. Sowmiya Manoj and S. Arulselvi Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 821
About the Editors
Dr. Saad Mekhilef is IET Fellow and IEEE Senior Member. He is Associate Editor of IEEE Transaction on Power Electronics and Journal of Power Electronics. He is Professor in the Department of Electrical Engineering, University of Malaya, since June 1999. He is currently Dean of the Faculty of Engineering and Director of Power Electronics and Renewable Energy Research Laboratory-PEARL. He is Author and Co-author of more than 400 publications in international journals and proceedings (253 ISI journal papers) and five books with more than 21000 citations and 73 H-index, and 116 Ph.D. and master students have graduated under his supervision. He has six patents. He frequently invited to give keynote lectures at international conferences. He was listed by Thomson Reuters (Clarivate Analytics) as one of the Highly Cited (Hi Ci) engineering researchers in the world and included in the Thomson Reuters’ The World’s Most Influential Scientific Minds: 2018. He is actively involved in industrial consultancy for major corporations in the power electronics projects. His research interests include power conversion techniques, control of power converters, renewable energy, and energy efficiency. Dr. Rabindra Nath Shaw is currently working as Director, International Relations, Bharath Institute of Higher Education & Research (Deemed to be University), Chennai, India. Before joining BIHER, he has served also Galgotias University as Director, IR&C. He is an alumnus of the Applied Physics Department, University of Calcutta, India. He is Senior Member of IEEE Industry Application Society, USA, and Fellow of Nikhil Bharat Shiksha Parishad, India. He is a global leader in organizing international conferences. His brand of world leading conference series includes IEEE International Conference on Computing, Power and Communication Technologies (GUCON), IEEE International Conference on Computing, Communication and Automation (ICCCA), IEEE IAS Global Conference on Emerging Technologies (GlobConET), International Conference on Electronics & Electrical Engineering (ICEEE), International Conference on Advances in Computing and Information Technology (ICACIT), etc. He holds the
xvii
xviii
About the Editors
position of Conference Chair, Publication Chair, and Editor for these conferences. These conferences are held in collaboration with various international universities like Aurel Vlaicu University of Arad, University of Malaya, University of Siena. Many world leaders are working with Dr. Shaw in these conferences. Most of these conferences are fully sponsored by IEEE Industry Applications Society, USA. He is also an expert in organizing international seminars/webinars/faculty development program in collaboration with leading institutes across the world. His research interests include optimization algorithms and machine learning techniques for power systems, IoT applications, renewable energy, and power electronics converters. He has published more than hundred twenty-five papers in Scopus/Web of Science/SCI indexed international journal/proceedings including high-impact factor journals in the field of renewable energy. He has also published several edited books from Springer and Elsevier publishing houses. He has worked on many national/international patents. He has successfully executed his duty in various positions like Governing Body Member, Centre in Charge, NBA Coordinator, University Examination Coordinator, University MOOC’s Coordinator, University Conference Coordinator, and Faculty-in-Charge, Centre of Excellence for Power Engineering and Clean Energy Integration. He has more than eleven years of teaching, research, and administrative experience in leading institutes like Motilal Nehru National Institute of Technology Allahabad, Jadavpur University, Gargi Memorial Institute of Technology, Brainware Group of Institution, Dream Institute of Technology. Prof. Pierluigi Siano received the M.Sc. degree in electronic engineering and the Ph.D. degree in information and electrical engineering from the University of Salerno, Salerno, Italy, in 2001 and 2006, respectively. He received the award as 2019 Highly Cited Researcher by ISI Web of Science Group. According to the study conducted by some Stanford University researchers, it has been included in the ranking of 100 thousand international scientists, belonging to different disciplinary fields and with greater scientific impact. The investigation was conducted by some Stanford University researchers who, starting from the SCOPUS database containing the list of scholars of the world classified by scientific production, developed a composite indicator that takes into account all possible bibliometric parameters (H-index, number publications, citations, journal relevance, sector relevance, etc.), in order to generate a wider ranking, focused on the scientific relevance of scientists, in relation to the activity of the last 23 years. He is Professor and Scientific Director of the Smart Grids and Smart Cities Laboratory with the Department of Management & Innovation Systems, University of Salerno. His research activities are centered on demand response, on the integration of distributed energy resources in smart grids and on planning and management of power systems. He has co-authored more than 450 papers including more than 200 international journal papers that received more than 7600 citations with an H-index equal to 44. He has been Chair of the IES TC on Smart Grids. He is Editor for the Power & Energy Society Section of IEEE Access, IEEE Transactions on Industrial Informatics, IEEE Transactions on Industrial Electronics, Open Journal of the IEEE IES and of IET Renewable Power Generation, Editor of the journal Smart Cities, MDPI Publisher since 2018, Editor of Intelligent Industrial Systems, Springer, 2014–2017.
Simulation of Single-Phase Cascaded H-Bridge Multilevel Inverter with Non-isolated Converter for Solar Photovoltaic System C. Dinakaran1(B) and T. Padmavathi2 1 Department of Electrical and Electronics Engineering, GITAM (Deemed to be University),
GITAM Institute of Technology, Visakhapatnam, Andhra Pradesh, India [email protected] 2 Department of Electrical, Electronics and Communication Engineering, GITAM (Deemed to be University), GITAM Institute of Technology, Visakhapatnam, Andhra Pradesh, India [email protected]
Abstract. This manuscript introduces an innovative single-phase cascaded HBridge multilevel inverter (MLI) through a minimum quantity of switches used on behalf of a Solar Photovoltaic (PV) application. Multilevel inverters are measured because of the power conversion system’s ability to support high power and power quality demanding applications. This manuscript intends to intend a modified cascaded Five-Level Inverter topology designed for a practically sinusoidal output voltage, fewer numbers of power switches, decreased complexity and reduced total harmonic distortion (THD). The composition is simple and easy to enlarge higher levels, but its gate driver circuits are also interpreted considering the number of operative switches is decreased. The pulse width modulation technique is employed through the multicarrier system to produce the gating signals supporting the power switches. This ability is attained by analyzing the proposed topology through the conventional topologies as the aforementioned observation point. The performance with efficient accuracy of the proposed topology by novel approach in generating voltage levels for a five-level inverter was proved over simulation using MATLAB/SIMULINK, PROTEUS results. Keywords: Cascaded H-bridge (CHB) · Modified Multilevel Inverter (MLI) · PWM Inverter · Solar Photovoltaic System · Total Harmonic Distortion
1 Introduction At present renewable power consumption obtains more beneficial together economically as well as environmentally [1]. The Solar PV arrangement maintains protected, clean, reliable renewable power sources through the added benefit of zero fuel charge, no moving parts, low operative control, most minor conservation, and extensive lifetime [2]. These points of importance build the consumption of solar photovoltaic in several positions further in off-grid installations [3]. MLI topologies are verified while consequential global awareness through the investigator as well as front-end corporation in a © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 1–12, 2022. https://doi.org/10.1007/978-981-19-1677-9_1
2
C. Dinakaran and T. Padmavathi
different medium as well as high power applications [4] considering of their competence to create an immense quality of output waveforms [5], condensing switching stress over the switches along with concentrating switching frequency [6]. In the current lifespan, several topologies have emerged in the field of grid-associated Multilevel Inverter in non-conventional energy applications [7]. The important categorization of MLI is Diode Clamped Multilevel Inverter, Flying Capacitor Multilevel Inverter, and Cascaded H-Bridge Multilevel Inverter [8]. When essential Multilevel Inverters are observable in high power applications, they contain some drawbacks related to components counts and voltage balancing difficulty except Cascaded H-Bridge MLI [9]. To conquer this difficulty, condensed switch Multilevel Inverter topologies are developed in extinct decades [10]. Several topologies in Multilevel Inverter are concluded by DC source, although only less topology is integrated through Photovoltaic solar applications. The present manuscript develops a novel single-phase grid-connected Multilevel Inverter with a Photovoltaic panel through a boost converter exploit as an alternative to DC sources in Multilevel Inverter [11]. The intended topology desires fewer switches as correlated to the essential form of Multilevel Inverter [12]. A cascaded MLI is various input voltage sources with output voltage levels are situated on the references [13]. The current inverter category is output voltage derivative commencing different input voltage sources [8] and the non-conventional energy source applications [14]. The investigation of cascaded MLI through switching states of a five-level inverter is comprehensive in Sect. 2. The modes of operation for modified topology are detailed in Sect. 3. The results, as well as discussions, are discussed in Sect. 4. The conclusion is formulated in Sect. 5.
2 Cascaded H-Bridge Inverters 2.1 Cascaded Five-Level Inverter Cascaded H-bridge (CHB) Inverters be MLIs constitute a series network of two or more single-phase CHB inverters. Every CHB correlates to two voltage source phase legs, wherever the line-line voltage persists the converter output. As a result, a single CHB converter is competent through achieve three distinct voltage levels. Every portion acquires individual two achievable switching states to deflect DC link capacitor shortcircuit. Moreover, to facilitate available are two legs, four specific switching states remain desirable, despite two of the system contain superfluous output voltage as represented in Fig. 1. Although two or more CHB persist connect in series, its output voltages can survive mutual towards appearance too many variant output levels, expanding the overall inverter output voltage with appeal measured energy as represented in Fig. 2. Figure 3 indicates two H-bridges allied with its series, with an approximate form of its potential being three-level output voltages. The whole converter output voltage obtains too decorated, produce five differential voltage levels. In universal circumstances, as of n CHB in sequence, 2n + 1 various voltage levels persist attained with an ultimate output voltage of nVdc is achievable. Cascaded H-Bridge presents additional redundancies than
Simulation of Single-Phase Cascaded H-Bridge Multilevel Inverter
3
Fig. 1. Power circuit with DC supply
the preceding topologies, in view that every CHB or power cell obtain specific superfluous switching state, along with means of the series association effectively introduces additional redundancies.
Fig. 2. Five-level CHB output voltage waveform (1 - F).
2.2 Proposed Five Level Cascaded H-Bridge Inverter (CHBI) This work, proposed CHB five-level inverters coupled in series, does necessitate comprising identical DC input voltages. In actuality, an appropriate collection of voltage asymmetry within cells can include production at various voltage levels arrangements with reduces repetition. As exposed in Fig. 4, an asymmetry as an alternative voltage ratio of 1:3 meant that a 2-cell Cascaded H-Bridge has to identify five levels accomplished during a 2-cell Cascaded H-Bridge in Fig. 3.
4
C. Dinakaran and T. Padmavathi
Fig. 3. Power circuit by solar photovoltaic module
Fig. 4. Proposed CHB multilevel inverter
Simulation of Single-Phase Cascaded H-Bridge Multilevel Inverter
5
The considered system tells us that the asymmetric fed topology produces the equivalent output voltage for fewer switches. Though, in the case appertaining to adopting the similar power semiconductors technology, the absolute output voltage is decreased as the ultimate blocking voltage of the application resolve limits the consequence of adequate DC source.
3 Modes of Operation of the Proposed Methodology This modified five-level inverter required just six switches correlated to 8 switches of the traditional inverter. The detriment suitable to switching devices persists too condensed. This inverter subsists of an adaptation system in the time of a switching device combined during a voltage source with a diode. The statistic of modification units is corresponding to the computation of producing levels. The switching progression of a five-level inverter is precise in Table 1. The output levels persist unidirectional with CHB, and the output is transformed within bidirectional. The output waveforms of the CHB have five levels, as demonstrated in Fig. 4. Switches Sa and Sd are ON with all through the negative half cycle, Sb as well as Sc are ON in the multilevel CHB inverter. Suitable switching concerning the inverter can yield five output voltage levels (Vdc /2, Vdc , 0, −Vdc /2, −Vdc ) commencing the DC supply voltage. The necessary five levels of the output voltage are achieved as follows. 1) Half of the Positive Output Voltage (Vdc /2): Sa act ON, relating to the positive load terminal through Sd is ON, concerning the negative load terminal through switches S2 by diodes D1 . The entire alternative controlled switches are OFF. The voltage utilized to the load terminal is Vdc /2. 2) Maximum Positive Output Voltage (Vdc ): Sa act ON, relating to the positive load terminal by Sd is ON, relating to negative load terminal through switches S1 and S2 . The real alternative controlled switches are OFF. The voltage utilized to the load terminal is Vdc . 3) Zero Output Voltage (0 & 0*): This level can obtain performed through two switching combination, switches Sa as well as Sd (or) Sc as well as Sb are ON by diodes D1 . The load terminal ab is short-circuited with an applied voltage to the load terminal is zero. 4) Half of the Negative Output Voltage (-Vdc /2): Sc act ON, relating to the negative load terminal by Sb is ON, and connecting to the positive load terminal through switches S2 by diodes D1 . The entire alternative restrained switches are OFF. The voltage related to the load terminal is -Vdc /2. 5) Maximum Negative Output Voltage (-Vdc ): Sc act ON, relating to the negative load terminal by Sb is ON, relating to positive load terminal through switches S2 and S1 . The entire alternative controlled switches are OFF. The voltage connected to the load terminal is -Vdc . Where ‘0’ is OFF & ‘1’ is ON. Table 1 projects the switching sequence to develop five output voltage levels (Vdc /2, Vdc , 0, -Vdc /2, -Vdc ).
6
C. Dinakaran and T. Padmavathi Table 1. Output voltage confer to the switches ON-OFF condition
V0
S1
S2
Sa
Sb
Sc
Sd
D1
Vdc /2
0
1
1
0
0
1
1
Vdc
1
1
1
0
0
1
0
0
0
0
1
0
0
1
1
0*
0
0
0
1
1
0
1
-Vdc /2
0
1
0
1
1
0
1
-Vdc
1
1
0
1
1
0
0
4 Results 4.1 Existing Simulation Results The existing system has been experienced along with simulation results are exposed into this segment as demonstrated in Fig. 5. This demonstration has been implemented through MATLAB/SIMULINK conditions through the SIMPOWER system toolbox as illustrated in Figs. 6 and 7, and total harmonic distortion as represented in Fig. 8.
Fig. 5. Simulation circuit designed used for CHB five-level inverter
Simulation of Single-Phase Cascaded H-Bridge Multilevel Inverter
Fig. 6. PWM based on CHB five-level inverter
Fig. 7. Output Voltage intended for CHB five-level inverter
7
8
C. Dinakaran and T. Padmavathi
Fig. 8. THD meant for CHB five-level inverter
4.2 Proposed System Simulation Results The proposed system possesses implemented in Fig. 9 through MATLAB/SIMULINK conditions through SIMPOWER system toolbox and output waveforms as shown in Figs. 10 and 11, and total harmonic distortion as represented in Fig. 12.
Fig. 9. Simulation circuit designed for proposed CHB five-level inverter
Simulation of Single-Phase Cascaded H-Bridge Multilevel Inverter
9
Fig. 10. Pulse width modulation signal for proposed CHB five-level inverter
Fig. 11. Output voltage meant for proposed cascaded H-bridge five-level inverter
4.3 Proposed System PROTEUS Model Proteus is one of the prominent mainstream simulators. It preserves used to create about every circuit on stimulating fields. It is graceful to facilitate the Graphical User Interface to make it possible compared to the actual prototype board, as illustrated in Fig. 13. Moreover, it conserves to make printed circuit boards and the waveform as shown in Figs. 14 and 15.
10
C. Dinakaran and T. Padmavathi
Fig. 12. THD meant for proposed cascaded h-bridge five-level inverter
Fig. 13. Development of proposed CHB five-level inverter in PROTEUS software
Simulation of Single-Phase Cascaded H-Bridge Multilevel Inverter
11
Fig. 14. Output pulse meant for proposed CHB five-level inverter
Fig. 15. Output voltage for proposed CHB five-level inverter
5 Conclusion This paper has projected the proposed CHBMLI through fewer power switches in MATLAB/SIMULINK and PROTEUS Software representation. The single Pulse width modulation system diminishes the complication in the pulse generation circumference also lowers the total harmonic distortion (THD) with enhanced power factor. Due to the unique appearance, the proposed topology is executive to the traditional system. Therefore single Pulse width modulation is suitable used for the proposed multilevel inverters. Multilevel inverters contain developed commencing, a growing technology toward a well-recognized and attractive medium-voltage resolution with high power applications. The device’s continuous improvement with industrial application development resolves innovative challenges and opportunities by stimulating further improvements
12
C. Dinakaran and T. Padmavathi
to CHBMLI technology. The simulation outcome in competition well during the consequent PROTEUS. This arrangement is predictable to be extremely helpful in high-power applications.
References 1. Majareh, L., et al.: Design, analysis and implementation of a generalised topology for multilevel inverters with reduced circuit devices. IET Power Electron. 12(14), 3724–3731 (2019) 2. Sarwer, Z., et al.: An improved asymmetrical multilevel inverter topology with reduced semiconductor device count. Int. Trans. Electr. Energ. Syst. 30(11), e12587 (2020) 3. Majumdar, S., Mahato, B., Jana, K.C.: Analysis of most optimal multi-unit multi-level inverter having minimum components and lower standing voltage. IETE Tech. Rev. 38(5), 520–536 (2020). https://doi.org/10.1080/02564602.2020.1799874 4. Seifi, A., Hosseinpour, M., Dejamkhooy, A.: A switch-source cell-based cascaded multilevel inverter topology with minimum number of power electronics components. Trans. Inst. Meas. Control. 43(5), 1212–1225 (2021) 5. Wang, X., et al.: A half-bridge IGBT drive and protection circuit in dielectric barrier discharge power supply. Circuit World (2021) 6. Muralikumar, K., Ponnambalam, P.: Analysis of cascaded multilevel inverter with a reduced number of switches for reduction of total harmonic distortion. IETE J. Res. 1–14 (2020) 7. Shaw, R.N., et al.: Review and analysis of photovoltaic arrays with different configuration system in partial shadowing condition. Int. J. Adv. Sci. Technol. 29(9s), 2945–2956 (2020) 8. Siddique, M.D., Iqbal, A., Memon, M.A., Mekhilef, S.: A new configurable topology for multilevel inverter with reduced switching components. IEEE Access 8, 188726–188741 (2020). https://doi.org/10.1109/ACCESS.2020.3030951 9. Siddique, M.D., Mekhilef, S., Rawa, M., Wahyudie, A., Chokaev, B., Salamov, I.: Extended multilevel inverter topology with reduced switch count and voltage stress. IEEE Access 8, 201835–201846 (2020). https://doi.org/10.1109/ACCESS.2020.3026616 10. Arif, MSB, et al.: Asymmetrical multilevel inverter topology with low total standing voltage and reduced switches count. Int. J. Circuit Theory Appl. 49(6), 1757-1775 (2021) 11. Kakar, S., et al.: New asymmetrical modular multilevel inverter topology with reduced number of switches. IEEE Access. 9, 27627-27637 (2021) 12. Siddique, M.D., Iqbal, A., Sarwar, A., Mekhilef, S.: Analysis and implementation of a new asymmetric double H-bridge multilevel inverter. Int. J. Circuit Theory Appl. 49(12), 4012– 4026 (2021). https://doi.org/10.1002/cta.3091 13. Mustafa, U., et al.: Single phase seven-level inverter topology with single DC source and reduce device count for medium and high power applications. In: 2019 Innovations in Power and Advanced Computing Technologies (i-PACT), vol. 1. IEEE (2019) 14. Kumawat, R.K., Palwalia, D.K.: A comprehensive analysis of reduced switch count multilevel inverter. Aust. J. Electr. Electron. Eng. 17(1), 13–27 (2020)
Intrusion Detection Based on PCA with Improved K-Means Pralhad Chapagain1(B) , Arun Timalsina2 , Mohan Bhandari1 , and Roshan Chitrakar1 1 Nepal College of Information Technology, Balkumari Lalitpur, Nepal
[email protected]
2 Pulchowk Engineering Campus, Pulchowk Lalitpur, Nepal
[email protected]
Abstract. An intrusion detection system (IDS) is a network traffic monitoring system that detects malicious activities and issues alerts. Because of the growth of cloud computing, social networks, and mobile cloud computing, intrusion detection systems have become more relevant. Based on the cluster information, we can decide if a behavior is normal or not. In this study, the clustering approach based on PCA with improved K-Means is proposed. Firstly, pre-processing is done to digitize the string and then dataset is normalized using Min-Max normalization so as to improve the cluster efficiency. Secondly, the Principal Component Analysis is used to extract the features in reduced dimensions. Based on cumulative explained variance ratio, the number of required features/components is calculated for the better accuracy. And, for the clustering, improved K-Means is used. Number of clusters for improved K-Means are obtained by Detection Rate (DR) and False Positive Rate (FPR) analysis at different number of cluster and centroid initialization is based on weighted average. The experiment is carried out on NSL-KDD dataset and DR, FPR, Accuracy and clustering time are computed to analyses the proposed IDS. The experiment of proposed method shows the improvements on detection rate, false positive rate, clustering accuracy and clustering time over different methods like PCA Kmeans, Minibatch K-means and ICA-Kmeans. Keywords: Intrusion Detection System · Cumulative Explained Variance · PCA · Improved K-means · NSL-KDD · Detection Rate · False Positive Rate
The swift widening of computer networks, use of computers, and the arrival of electronic commerce, the computer network security has become highly important and international priority. Because of the advancement of social network and mobile cloud computing, the use of computer network is vigorously increased so, intrusions and malicious accesses to computer networks have become commonplace. So, the internet attack has grown considerably [1]. These security problems make it difficult to differentiate between an intrusion and a legitimate user’s usual actions [2]. Intruders take advantage of computer system vulnerabilities, despite the fact that designing a system without any vulnerabilities is nearly impossible. Furthermore, some attackers may employ unknown intrusion patterns, complicating the situation. As a result, intrusion detection systems have become a crucial component of computer hosts and networks [3]. In general, intrusion detection methodologies are of two types: one is © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 13–27, 2022. https://doi.org/10.1007/978-981-19-1677-9_2
14
P. Chapagain et al.
abuse detection and next one is anomaly detection. The first method, misuse detection, detects intrusions based on known patterns, while the second method, anomaly detection, focuses on unknown patterns. The anomaly detection assumes that the intrusion will deviate from normal behaviors, so it establishes the system’s normal behavior pattern [4]. Since an intrusion detection system gathers data from various sources, data analysis becomes more complicated and difficult. Data mining techniques are used to work with huge amounts of data. As a result, the intrusion detection model will produce a normal behavior model using datamining techniques. As a consequence, attack and regular data patterns can be effectively differentiated. In intrusion detection systems, various data mining techniques are used [5]. In general, data mining-based IDS research focuses on two aspects: classification and clustering. Since classification is a supervised learning system, it cannot handle data that has not been classified. When it comes to identifying unknown intrusion patterns, classification techniques are often unsuccessful. Clustering, on the other hand, is an unsupervised learning technique that can cluster unlabeled data in order to detect intrusion [6]. There is different intrusion detection system which are based on clustering. To improve the clustering approach different variations are proposed in literature. But some of the approach has high memory requirement, and some has less accuracy. This thesis work provides a new method for intrusion detection system to fix these flaws. Firstly, pre-processing is done to digitize string in dataset and then d it is normalized using Min-Max normalization so that cluster efficiency can increase. Secondly, Principal Component Analysis is used to extract the features in reduced dimensions. Based on cumulative explained variance ratio, the no of required features/components is calculated for the better accuracy. And, for the clustering, improved K-Means is used. To improve the K-Means with the problem, like: stuck into local optima and proper selection of number of clusters, optimal cluster number are obtained by Detection Rate (DR), accuracy and False Positive Rate (FPR) analysis at different number of cluster and centroid initialization is based on weighted average. The experiment is done by taking NSL-KDD dataset and accuracy, DR, FPR are computed to analyses the proposed IDS.
1 Related Work Intrusion Detection Systems (IDS) can detect malicious activities and anomalies in a network, providing a vital foundation for network protection. Extreme work is done on intrusion detection system. In literature, different feature extraction technique and clustering method are proposed. Xingang Wang and Linlin Wang, developed an intrusion detector is based on autoencoder feature extraction. Autoencoders are used to learn how to code data effectively without the need for monitoring Feature retrieval is a method for discovering essential features in data for coding by learning from the original data set’s coding to construct new ones. This research work realizes improved K-Means clustering for clustering the reduced dimension dataset produced by autoencoder [7]. Anomaly detection using PLS feature obtaining and a core vector machine was proposed by Gan xu-sheng, Duanmu Jing-shun et all. To boost the ability to detect anomaly
Intrusion Detection Based on PCA with Improved K-Means
15
intrusions, partial least square (PLS) feature removal is used. PLS is a radical integration of multiple linear regression, canonical correlation analysis, and principal component analysis. It is a statistical analysis technique of second-generation multivariable. PLS can not only accurately summarize the information generated by independent variables, but it can also effectively describe dependent variables. The speediness superiority of the CVM algorithm in processing large-scale sample data is used to develop an anomaly intrusion detection model for the feature set in this work [8]. Yasir Hamid and M. Sugumaran developed a non-linear dimension cut-off method based on t-SNE for network intrusion detection. They proposed nonlinear dimension reduction technique t-SNE for feature extraction which can enhance the discrimination of data. t-SNE is a refined version of SNE (an unsupervised dimensionality reduction method) that aims to reduce the difference between a pairwise similarities distribution of input objects and the pairwise similarity distribution in the related low-dimensional embedding [9]. Fangei Weng, Qingshan Jiang et al. uses the clustering ensemble for system for detecting intrusion. However, it is difficult to get the great effective detection by using single clustering algorithm. As a result, they proposed a clustering ensemble-based unsupervised anomaly detection scheme. The scheme is based on multiple K-Means runs to gather evidence and avoid false classification of anomalistic results, then using single-link to create a hierarchical clustering tree to obtain the final clustering result [10]. Liu Tao, Hou Yuan-bit et al. conducted a ‘study of a fast-clustering algorithm in intrusion detection based on foregone samples’ This algorithm solved the problem of a higher false positive rate and lower detection rate in network anomaly detection induced by conventional clustering approaches such as choosing original clustering centers at random and computing a single attribute (continual or discrete) only [11]. Cheng Zhong and Na Li proposed a clonal selection-based incremental clustering algorithm for intrusion detection. The problems of computing differences for data, mixed attributes, unknown cluster number, so can fall into local optimization are all solved by using the clonal selection algorithm to optimize the clustering results. Intrusion detection can obtain a high detection rate and low false positive rate by using this form [12]. Md. Sohrab Mahmud, Md. Mostafizer Rahman and Md. Nasim Akhtar, in their work for clustering proposed an improvement of the k-means clustering algorithm with weighted average initial centroids. Cluster analysis is a popular data analysis method, and k-means is one of the most widely used partitioning algorithms. Cluster analysis is computationally intensive, and the resulting collection of clusters is highly dependent on the initial centroids chosen. Weighted average initial centroids have greater cluster precision with less computational time [13]. Mohsen Eslamnezhad and Ali Yazdian Uarjani proposed Minmax K-Means clustering-based intrusion detection. This is a new intrusion detection approach that employs the Minmax clustering algorithm to resolve the k-means algorithm’s lack of sensitivity to initial centers and improve clustering efficiency [14]. Kai Peng worked on intrusion detection, and for clustering, Mini-Batch K means was used. Principal Component Analysis is used to extract features, Mini-Batch K-Mean’s
16
P. Chapagain et al.
clustering is used to cluster the dataset, and k-means++ is used to configure the cluster centroids in the proposed intrusion detection method [15]. Zuherman Rustam and Aini Suri Talita proposed ‘fuzzy kernel robust clustering for anomaly-based intrusion detection.’ In this research work KDD cup’99 dataset is used and classified into five different cluster, one for normal and four for attack type. The experiment is carried on 10% and 90% of the dataset and the accuracy is evaluated [16]. Guo Pu, Lijuan Wang et al., on their research in title ‘A Hybrid Unsupervised Clustering based anomaly detection method’, to detect attacks without any prior knowledge, sub-space clustering and a single class support vector machine are used. They are using NSL KDD dataset for the experiment and compared the obtain result with some of the existing techniques [17].
2 Methodology 2.1 The Proposed Model The proposed model for the intrusion detection system includes different steps. These steps are organized to construct the experiment and process flow diagram (Fig. 1).
Fig. 1. Model of PCA with improved K-means for IDS
2.2 Data Preprocessing Data preprocessing is indeed a data mining method which includes converting raw data into an understandable format. In preprocessing, two tasks have done i.e., converting the string into numerical data using the label encoder in python and normalizing the data using Min-Max normalization, given by Eq. 1. MinMax =
value − min. value value − min. value
(1)
Intrusion Detection Based on PCA with Improved K-Means
17
Where, MinMax = Min-Max Normalization, Max.value = Maximum value and Min.value = Minimum Value. 2.3 Feature Extraction Using PCA Extraction of features is a form of dimensionality reduction that divides a large amount of raw data into smaller units that can be processed. [18] (Fig. 2).
Fig. 2. Flow chart for PCA algorithm
PCA is a technique for reducing the dimensionality of data with a large number of variables. This is accomplished by converting the original variables into a new limited collection of variables while preserving the most relevant details from the original data set. 2.4 Improved K-Means Clustering Clustering may be described as “the process of grouping objects into groups whose members are similar in some way” in a broad sense. Each observation is allocated to the cluster which has the nearest mean in this procedure, which divides data objects (n) into (k) clusters. the following objective function, Eq. (2), is attempted to be minimized by this algorithm (Fig. 3): J =
k n j=1
i=1
(j)
||Xi
− Cj ||2
(2)
18
P. Chapagain et al.
Fig. 3. K-means algorithm
Algorithm 1: Standard K-Means Inputs: D= d1, d2, ……., dm // a list of m data values k // total clusters number C // Cluster Center Output: k cluster set Process: 1: Group the data values into ‘k' clusters, fixed number where ‘k' is known b efore starting the cluster. 2: At random, 'k' points are chosen to serve as cluster centers (C). 3: Using the Euclidean distance equation, assign each object to its nearest cluster center. 4: Determine the mean or centroid of all objects in each cluster. 5: Repeat above steps 2, 3, and 4 until each cluster receives the same number of points in each round.
For the improved K-Means, determination of no of cluster ‘k’ is carried out by observing the better detection rate and accuracy with less false positive rate for different values of ‘k’. And initial centroid is calculated by using weighted average algorithm. This algorithm is given below: Algorithm 2: Weighted Average Algorithm Inputs: D= d1, d2, ……., dm // a list of m data values k // the optimal number of clusters Output: An initial set of total k centroids Process: 1: Calculate the average of each data point's value. di=x1, x2, x3 …….,xn di(avg)=(w1*x1+w2*x2+ +wn*xn)/n where, x= the value of attribute, n=attributes number and w=weight which is used to multiply to ensure the clusters are distributed evenly 2: Order the details by average score. 3: Subdivide the data items into a total of ‘k' subsets. 4: Calculate each subset's mean value. 5: For each data subset, use the nearest possible data point of the mean as the initial centroid.
Intrusion Detection Based on PCA with Improved K-Means
19
Finally, the improved K-Means algorithm is: Algorithm 3: Improved K-Means Input: D= d 1 , d 2 , ……., dm // set of m data items k // the optimal number of clusters calculated by using ‘determination of k’ technique discussed earlier Outputs: a set of k cluster Process: 1: Initialize the initial centers for k cluster by using ‘initial centroid calculation’ technique discussed earlier 2: Repeat this step until no change in cluster means a. Based on the mean value of the data items in the cluster, assign each data items from m to one of the k clusters to which it is most similar. b. Take the mean value of the artifacts for each of the k clusters to change the cluster means.
2.5 Evaluation Criteria Several metrics are examined in this thesis work to analyze the proposed intrusion detection system. Attack detection rate (DR), cluster accuracy, False Positive rate (FPR), and clustering time are the four metrics. total detected abnormal data total abnormal data
(3)
no. of normal data detected as abnormal total normal data
(4)
DR = FPR =
Accuracy =
total correctly detected data total data
(5)
3 Experiment and Result 3.1 Dataset Overview The experiment is conducted using the NSL-KDD dataset. This dataset is improved version of KDD’99 datasets. For this research work, the training and testing dataset with normal and various attack type data is taken for the experiment [19] (Table 1). The dataset contains 41 different features [20]. and different attack types for training and testing dataset. These attack type can be categorized into four different types [21] (Table 2).
20
P. Chapagain et al. Table 1. Dataset overview
Data set
Training dataset
Testing dataset
Normal
67343
9712
Dos
45927
7457
Probe
11656
2421
R2L
995
2754
U2R
52
200
Total
125973
22544
Table 2. Attack category Attack category
Attacks in training set
Attacks in testing set
Dos (Denial of service)
back, neptune, smurf, teardrop, land, pod
apache2, mailbomb, processtable, udpstorm
Probe
satan, portsweep, ipsweep, nmap
mscan, saint
R2L (remote-to-local)
warezmaster, warezclient, ftpwrite, guesspassword, imap, multihop, phf, spy
sendmail, named, snmpgetattack, snmpguess, xlock, xsnoop, worm
U2R (User-to-Root)
rootkit, bufferoverflow, loadmodule, perl
httptunnel, ps, sqlattack, xterm
3.2 Experimental Setup Python is used to run the experiment on a Windows 10 device with an Intel Core i5, 8th Gen, 1.6GHz CPU, 8GB of RAM, and Anaconda3.0 SCIkit-learn as the main software platform. 3.3 Performance Analysis Number of Features Selecting The ability to predict how many components are required to explain the data is an important part of using PCA in practice. The cumulative explained variance ratio can be plotted as a function of the number of features/components we find from the Fig. 4 that value of required component as 10 which can describe the complete dataset more than 95%. Cluster Center Determining The improved k-means uses the weighted average score for the center initialization. To compute initial center, average of all the data point with respective weight is calculated.
Intrusion Detection Based on PCA with Improved K-Means
21
Fig. 4. Cumulative explained variance vs number of components
Weight is determined by correlation of each data point to the cluster label. After calculating average, sorted the data and divided into ‘k’ subgroup. Average of each subgroup is calculated and based on that average; initial center is computed. See Table 3. Table 3. Correlation value with cluster label Principal component
Correlation value
Principal component 1
0.789331
Principal component 2
0.232615
Principal component 3
0.222269
Principal component 4
0.038611
Principal component 5
0.091035
Principal component 6
0.162758
Principal component 7
0.168998
Principal component 8
0.048995
Principal component 9
0.002099
Principal component 10
0.062230
Number of Cluster Finding One of the most essential factors in the clustering method is the cluster number, k, which has a direct effect on the system’s accuracy. The choice of k in this study is based on the high detection, true alarm rate, and clustering accuracy (Table 4). By, observing these data, the detection rate and accuracy has the high value and FPR has lowest for K = 90, so for further computation no of cluster is taken as 90. Result and Discussion After finalizing the number of features for the PCA and number of clusters for the K-Means by using training set of data, now implementing this method to test data for the evaluation of detection rate, false positive rate, accuracy and clustering time by using Eq. (3), (4) and (5) (Table 5).
22
P. Chapagain et al. Table 4. DR, FPR and Accuracy based on different value of k
No of cluster, k
Detection rate (DR)
False positive rate (FPR)
Accuracy
5
0.800
0.199
0.884
10
0.872
0.127
0.930
20
0.902
0.097
0.949
30
0.912
0.080
0.949
40
0.929
0.070
0.955
50
0.938
0.061
0.958
60
0.939
0.060
0.962
70
0.943
0.056
0.965
80
0.944
0.055
0.965
90
0.959
0.0403
0.968
100
0.951
0.048
0.967
110
0.948
0.051
0.968
120
0.946
0.053
0.966
Table 5. DR and FPR of proposed method Category of data
Detection rate
False positive rate
Normal
0.9297
0.0835
Dos
0.9161
0.00795
Probe
0.9339
0.00944
R2L
0.7919
0.034
U2R
0.48
0.00366
This describes that, detection rate of Probe attack is higher than all other data and U2R has the less detection rate. Where, U2R has the less FPR and Normal data has high FPR. These results are represented in bar graph as shown in Figs. 5 and 6. Comparing Results To verify the proposed model, the DR, FPR, Accuracy and clustering time of different clustering methods are compared. For the comparison, PCA with k-means (which has only 2-D features i.e., features are reduced to 2-d), PCA-Kmeans, ICA-Kmeans, Minibatch clustering result are taken (Table 6). The given table shows the Detection Rate (DR) of various approach. Among these different methods, the proposed method has the highest rate of detection so the proposed model is better than other mentioned models. The table data can be presented in bar chart as follows (Figs. 7, 8, 9, 10 and 11):
Intrusion Detection Based on PCA with Improved K-Means
Fig. 5. DR of proposed method
Fig. 6. FPR of proposed method
Table 6. DR comparison of different methods Clustering method
Normal and attack type Normal
Dos
Probe
R2L
U2R
PCA = 2withKmeans
0.8136
0.9102
0.800
0.6597
0.38
PCA-Kmeans
0.9150
0.912
0.8946
0.7846
0.48
ICA-Kmeans
0.9262
0.9031
0.895
0.7476
0.475
Minibatch-Kmeans
0.907
0.904
0.8909
0.6702
0.35
Proposed method
0.9297
0.9161
0.9339
0.7919
0.48
Fig. 7. Normal data DR comparison
Fig. 8. DOS attack DR comparison
Fig. 9. Probe attack DR comparison
Fig. 10. R2L attack DR comparison
23
24
P. Chapagain et al.
Fig. 11. U2R attack DR comparison
The table below shows the False Positive Rate (FPR) of different approach. Among these different methods, the proposed method has the overall lowest false positive rate so the proposed method is better than other mentioned methods (Table 7). Table 7. FPR comparison of different methods Clustering method
Normal and attack type Normal
Dos
Probe
R2L
U2R
PCA = 2 with Kmeans
0.0911
0.0509
0.0206
0.0808
0.0304
PCA-Kmeans
0.067
0.0241
0.0105
0.0458
0.0037
ICA-Kmeans
0.067
0.0239
0.0115
0.0474
0.0039
Minibatch-Kmeans
0.0868
0.0408
0.0185
0.038
0.0021
Proposed method
0.0835
0.0079
0.0094
0.034
0.0036
The observation can also be visualized from bar chart, which are given below (Figs. 12, 13, 14, 15, and 16):
Fig. 12. Normal data FPR comparison
Fig. 13. DOS attack FPR comparison
Now, the accuracy and time for clustering the test data using different clustering approach is given in the table below (Table 8). These data are represented in bar chart as follows (Figs. 17 and 18):
Intrusion Detection Based on PCA with Improved K-Means
Fig. 14. Probe attack FPR comparison
Fig. 15. R2L attack FPR comparison
Fig. 16. U2R attack FPR comparison
Table 8. Accuracy and clustering time comparison Clustering method
Accuracy
Clustering time (second)
PCA = 2 with Kmeans
0.8215
18.02
PCA-Kmeans
0.892
11.01
ICA-Kmeans
0.8894
9.77
Minibatch-Kmeans
0.8707
0.565
Proposed method
0.9048
0.049
Fig. 17. Accuracy comparison
Fig. 18. Clustering time for testing comparison
25
26
P. Chapagain et al.
4 Conclusion and Recommendation The aim of this research work is to develop the method which has better accuracy, detection rate and false positive rate. For this, number of components for PCA are computed by using cumulative variance ration and Custer center is computed using weighted average center initialization is used. Number of clusters is finalized after observing the DR, Accuracy and FPR at different number of clusters. And finally, K-Means with those improvements are implemented, and the results are observed. This illustrates the better response. The experimental results shows that the detection rate of proposed method is maximum for Normal (92.97%), Dos (91.61%), Probe (93.39%), R2L (79.19%) and U2R (48%) as compared to other mentioned methods. The false positive rate of proposed method for DOS (0.79%), Probe (0.94%) and R2L (3.4%) are minimum compared to other methods but for Normal (6.7%), PCA Kmeans has better result and for U2R (0.21%), Minibatch Kmeans is best. The accuracy (90.48%) and clustering time (0.049 s) of proposed method is superior than other mentioned approach. It is possible to develop a system that improves the DR and FPR of R2L and U2R attacks in the future. Furthermore, the research may concentrate on applying applicable domain information from the security sector to the proposed framework in order to enhance its detection capability for new types of attacks.
References 1. Papalexakis, E., Beutel, A., Steenkiste, P.: Network anomaly detection using co-clustering. In: 2012 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, Istanbul (2012) 2. Tang, P., Jiang, R.-A., Zhao, M.: Feature Selection and design of instrusion detection system based on K-Means and triangle are support vector machine. In: 2010 Second International Conference on Future Networks, Sanya, China (2010) 3. Wang, S.: Reasearch of intrusion detection based on an improved k-means algorithm. 2011 Second International Conference on Innovations in Bio-inspired Computing and Applications, Shenzhen, China (2010) 4. Wankhade, K., Patka, S., Thool, R.: An overview of intrusion detection based on data mining techniques. In: 2013 International Conference on Communication Systems and Network Technologies, Gwalior, India (2013) 5. Liao, H.-J., Lin, R.C.-H., Lin, Y.-C., Tung, K.-Y.: Intrusion detection system: a comprehensive review. J. Netw. Comput. Appl. 36(1), 16–24 (January 2013) 6. John, A., Denatious, D.K.: Survey on data mining techniques to enhance intrusion detection. In: 2012 International Conference on Computer Communication and Informatics, Coimbatore, India (2012) 7. Wang, X., Wang, L.: Research on Intrusion Detection Based on feature extraction of Autoencoder and the improved k-means algorithm. In: 2017 10th International Symposium on Computational Intelligence and Design (ISCID), Hangzhou, China (2017) 8. Xu-sheng, G., Jing-Shun, D., Jia-Fu, W., Wei, C.: Anomaly intrusion detection based on PLS feature extraction and core vector machine. Knowl. Based Syst. 40, 1–6 (March 2013) 9. Hamid, Y., Sugumaran, M.: A t-SNE based nonlinear dimension reduction for network intrusion detection. Int. J. Inf. Technol. (2019) https://doi.org/10.1007/s41870-019-00323-9
Intrusion Detection Based on PCA with Improved K-Means
27
10. Weng, F., Jiang, Q., Shi, L., Wu, N.: An intrusion detection system based on the clustering ensemble. In: 2007 International Workshop on Anti-Counterfeiting, Security and Identification (ASID), Xizmen, China (2007) 11. Tao, L., Yuan-Bin, H., Ai-Ling, Q., Xin-Tan, C.: Study of fast clustering algorithm based on foregone samples in intrusion detection. In: 2009 International Conference on Networks Security, Wireless Communications and Trusted Computing, Wuhan, China (2009) 12. Zhong, C., Li, N.: Incremental clustering algorithm for intrusion detection using clonal selection. In: 2008 IEEE Pacific-Asia Workshop on Computational Intelligence and Industrial Application, Wuhan, China (2008) 13. Palimkar, P., Bajaj, V., Mal, A.K., Shaw, R.N., Ghosh, A.: Unique action identifier by using magnetometer, accelerometer and gyroscope: KNN approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 607–631. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_48 14. Eslamnexhad, M., Varjani, Y.A.: Intrusion detection based on minmax k-means clustering. In: 7’th International Symposium on Telecommunications (IST’2014), Tehran, Iran (2014) 15. Peng, K., Leung, C.V., Huang, Q.: Clustering Approach Based on Mini Batch Kmeans for Intrusion detection system over big data. IEEE Access 6,11897–11906 (2017) 16. Palimkar, P., Shaw, R.N., Ghosh, A.: Machine learning technique to prognosis diabetes disease: random forest classifier approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 219–244. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_19 17. Rustam, Z., Talita, A.S.: Fuzzy kernel robust clustering for anomaly based intrusion detection. In: 2018 Third International Conference on Informatics and Computing (ICIC), Palembang, Indonesia (2018) 18. Pu, G., Wang, L., Shen, J. , Dong, F.: A hybrid unsupervised clustering based anomaly detection method. Tsinghua Sci. Technol. 26(2), 146–153 (April 2021) 19. https://www.unb.ca/cic/datasets/nsl.html. [Online]. Accessed 27 Jan 2021 20. http://kdd.ics.uci.edu/databases/kddcup99/kddcup.names. [Online]. Accessed 27 Jan 2021 21. Natesan, P., Balasubramanie, P.: Multi stage filter using enhanced adaboost for network intrusion detection. Int. J. Netw. Secur. Its Appl. (IJNSA) 4(3), 121–135 (May 2012)
Small Signal Stability Improvement of Phillips Heffron Model with Fuzzy Based Control and Its Comparison with PSS Venu Yarlagadda(B) , G. Sasi Kumar, R. Geshma Kumari, and Giriprasad Ambati VNRVJIET, Hyderabad, Telangana, India {venu_y,sasikumar_g,geshmakumari_r,giriprasad_a}@vnrvjiet.in
Abstract. Power system stability playing a major role in the contribution towards the system security and reliability. The loads on the power system due to increased urbanisation and increased usage of power leads to low frequency rotor oscillations. The Single Machine Infinite Bus (SMIB) system with Phillips Heffron constants (K1 to K6) has been developed to carry out the simulation study with three cases, Case 1 is without any controller, Case 2 is with Power System Stabilizer (PSS) and Case 3 is with Fuzzy Logic Controller (FLC) with Filter. The Phillips Heffron model has been implemented and presented the results using Simulink model and carried out the simulation. The comparison of PSS and FLC is made in terms of dynamic responses of speed deviation and swing curves of the said model. The simulation results demonstrate the potency of the FLC in improving the small signal stability of the system compared to PSS. Keywords: Small SignalStability · SMIB · Power System Stabilizer · Fuzzy Logic Controller · Stability Improvement · Phillips Heffron Model · FLC with Filter
1 Introduction In the present scenario the rise of power demand leads to interconnection of power system for various reasons like reducing economical cost and to enhance reliability of the system. It leads to growing complexity which will cause further system collapse due to major outages.Power system stability is defined as a system ability to regain its initial equilibrium state after being subjected to a disturbance. This article focused on the assessment of the small signal stability of the Single Machine and Infinite Bus (SMIB) system realized with Phillips Heffron constants (K1 to K6) has been developed to carry out the simulation study with three cases, Case 1 is without any controller, Case 2 is with Power System Stabilizer (PSS) and Case 3 is with Fuzzy Logic Controller (FLC) with Filter. The load perturbations are causing the oscillatory instability of the SMIB system due to insufficient damping of the system, in which case system has to rely upon the external controller, which can provide the sufficient damping to ensure oscillatory stability. This article focused upon the design of © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 28–40, 2022. https://doi.org/10.1007/978-981-19-1677-9_3
Small Signal Stability Improvement of Phillips Heffron Model
29
the Power System Stabilizer (PSS) and Fuzzy Logic Controller (FLC) with filter, further used to mitigate small oscillations of rotor angle. These two types of compensators being used to enhance the small signal stability or oscillatory stability of the system by providing adequate damping demanded by the poser system. The case study is presented with the Phillips Heffron model with PSS and FLC and results have proven the superior performance of the FLC with Filter [1–7].
2 Power System Stability Power system stability is defined as a system ability to regain its initial equilibrium state after being subjected to a disturbance and its classificationare given in Fig. 1.
Fig. 1. Power system stability stratification
2.1 Rotor Angle Stability The system should remain in synchronism even after being subjected to the disturbance, which involves output power oscillates reflected in rotor oscillations. 2.2 Relation Between Power and Angle Relation between the power and angular position of a rotor in synchronous machine is nonlinear relation, when the synchronous generator is feeding a synchronous motor
30
V. Yarlagadda et al.
through transmission line. The power transferred to the motor from the generator is depends on the function of angular displacement between rotors of the generator and motor this is because of Motor internal angle, generator internal angle, and the angular displacement between motor and generator terminal voltage and power flow is given by the Eq. 1. P=
V1 V2 sin δ X12
(1)
This equation say that the power transferred to a motor from generator is maximum when the angle is 90°, if the angle is further increased beyond 90°, power transferred starts decreasing is shown in Fig. 2. The maximum power transferred is directly proportional to machine internal voltage [8–10]. Pmax =
V1 V2 X12
(2)
Fig. 2. Power angle curve
2.3 Small Signal Stability Whenever the synchronous machine is subjected to small and medium load perturbations then the machine is subjected to rotor oscillations. These oscillations will die out very soon due to damping provided in the machine or external damping provided by the controller or compensator. The swing curves of the machine will illustrates the dynamic instability mechanism of the machine for such disturbances as illustrated in the below Fig. 3 (a) and (b). The fig (a) depicts that the system is stable for small and medium disturbances with quiet effective damping and is unstable for medium disturbances with insufficient damping [4–9].
Small Signal Stability Improvement of Phillips Heffron Model
31
Fig. 3. Power system stability swing curves (a) Stable system (b) Oscillatory unstable system or dynamic instability and (c) Non-oscillatory instability or transient instability
3 Fuzzy Logic Controller Fuzzy control is used when vaguness in the decission making is present or when non linearities are involved in the system dynamics. Figure 4 shows Mamdani Rule based Fuzzy Logic Controller input and output functional relationship in simulink model and Fig. 5 3-d surface of the Fuzzy rules and the Table 1 shows the Fuzzy rules table for SMIB system. There are two input membership functions of Fuzzy logic controller, one is error i.e. the reference voltage minus actual voltage of seven inpit triangular membership functions. The Fuzzy rules table shows the 49 rules framed with input1 variables viz. NB1, NM1,NS1,ZO1,PS1,PM1 and PB1, similarly input2 variables viz. NB2, NM2,NS2,ZO2,PS2, PM2 and PB2 and output variables viz. NB, NM,NS,ZO,PS, PM and PB [1–13].
Fig. 4. Triangular input and output membership functions of Fuzzy controller
32
V. Yarlagadda et al.
Fig. 5. Fuzzy 3-D rule surface
Table 1. The fuzzy rules table for SMIB system de/dt
e NB1
NM1
NS1
ZO1
PS1
PM1
PB1
NB2
N1
N1
N1
N1
N2
N3
ZO
NM2
N1
N1
N1
N2
N2
ZO
P3
NS2
N1
N1
N2
N3
ZO
P3
P2
ZO2
N1
N2
N3
ZO
P3
P2
P2
PS2
N2
NS
ZO
P3
P2
P1
P1
PM2
N3
ZO
P3
P2
P1
P1
P1
PB2
ZO
P3
P1
P1
P1
P1
P1
4 Phillips Heffron Model The single machine infinite bus which is shown in below Fig. 6, The Phillips Heffron model constants and transfer functions of various blocks such as power amplifier, voltage regulator, field system dynamic model and generator load models. 4.1 Mathematical Modelling of SMIB Model The mathematical models of the various components said earlier have been obtained with the following mathematical expressions from Eqs. (3) to (29) and obtained the transfer function models as illustrated in the Fig. 7 below. This model is used to develop the small signal stability analysis with three cases said earlier [4–10]. Xe = External reactance
Small Signal Stability Improvement of Phillips Heffron Model
33
Fig. 6. Single machine infinite bus system
Xd = Direct axis reactance In case of synchronous machine power delivered to infinite bus | | E V |V |2 1 1 ( | − | ) sin 2δ sin δ + Pe = Xt 2 Xq X
(3)
d
|
|
Xt = xd + xe = xd
(4)
x|q + xe = x|q
(5)
|
xd = x|q
(6)
Substitute in power equation Pe =
| | E V Xt
sin δ = Pmax sinδ
(7)
The excitation system transfer function equation is given by Efd =
KE Vref − Vt 1 + sTE
(8)
The electrical torque Te is given by | Te = Eq| iq − xq − xd id iq
(9)
The stator equations while neglecting armature resistance are |
Eq| + xd id = vq
(10)
34
V. Yarlagadda et al.
−xq iq = vd
(11)
The network and load equation are Vq = −xe id + Re iq + vqL
(12)
Vd = Re id + xe iq + vdL
(13)
PL + j QL − VL Bc2 VqL + jVdL = iq − jid 2 + V2 VL = VdL qL PL = PLo QL = QLo
VL VLo VL VLo
(14) (15)
mp (16) mp (17)
Linearizing above equation from (1)–(9) and simplifying can be shown as iq = Yq Eq|
(18)
id = Yd Eq|
(19)
Te = K2 Eq|
(20) |
The torque deviation Te depends only on Eq Te = K1 δ + K2 Eq|
(21)
From Eq. (15), it is identified that this system may not show the angle instability [8–13, 14–18]. Te =
E EB ∂Te δ = cos δo (δ) ∂δ XT
(22)
1 (Tm − Te − KD ωr ) 2H
(23)
pωr =
pδ = ω0 ωr pωr =
1 (Tm − KS δ − KD ωr ) 2H KS =
E EB cos δ0 XT
(24) (25) (26)
Small Signal Stability Improvement of Phillips Heffron Model
35
KD KS
1 d ωr ωr − 2H − 2H = + 2H Tm δ 0 ω0 0 dt δ K3 [Efd − K4 δ] ψfd = 1 + pT3 K2 K3 K4 Te duetoψfd = δ 1 + sT3
(28)
Et = K5 δ + K6
(30)
(27)
(29)
Fig.7. Phillips Heffron model
5 Case Study and Results The Simulink model of Phillips Heffron model has been implemented for simulation study and analysis and results have been presented in this article. Figure 8 depicts the Simulink model of Phillips Heffron model without controller, Fig. 9 shows the Simulink model of Phillips Heffron model with PSS, Fig. 10 illustrates the Simulink model of Phillips Heffron model with FLC. The transient responses of the Phillips Heffron model have been obtained without any controller, which shows system is completely unstable with insufficient damping. Figure 11 shows the Speed deviation of Phillips Heffron model without Controller and Fig. 12 shows the Swing curves of Phillips Heffron model without Controller and both these curves proves the system is completely unstable. Figure 13 depicts the Speed deviation of Phillips Heffron model with PSS and FLC and Fig. 14 illustrates the Swing curves of Phillips Heffron model with PSS and FLC both these plots are proving that the system is predominantly showing superior performance with FLC than PSS.
36
V. Yarlagadda et al.
Fig. 8. Simulink model of Phillips Heffron model without controller
Fig. 9. Simulink model of Phillips Heffron model with PSS
Small Signal Stability Improvement of Phillips Heffron Model
Fig. 10. Simulink model of Phillips Heffron model with FLC
Fig. 11. Speed deviation of Phillips Heffron model without Controller
37
38
V. Yarlagadda et al.
Fig. 12. Swing curves of Phillips Heffron model without Controller
Fig. 13. Speed deviation of Phillips Heffron model with PSS and FLC
Small Signal Stability Improvement of Phillips Heffron Model
39
Fig. 14. Swing curves of Phillips Heffron model with PSS and FLC
6 Conclusions The Single Machine Infinite Bus system with Phillips Heffron constants has been developed to carry out the simulation study with three cases, Case 1 is without any controller, Case 2 is with Power System Stabilizer (PSS) and Case 3 is with Fuzzy Logic Controller (FLC) with Filter. The Phillips Heffron model has been implemented using Simulink model and carried out the simulation and presented the results. The comparison of PSS and FLC is made in terms of dynamic responses of speed deviation and swing curves of the said model. The transient responses of the Phillips Heffron model have been obtained without any controller, which shows system is completely unstable with insufficient damping. The simulation plots of all three cases have proven the fact that the system is completely unstable without any controller. Swing curves of Phillips Heffron model with PSS and FLC both these plots are proving that the system is predominantly showing superior performance with FLC than PSS, further these results demonstrate the potency of the FLC in improving the small signal stability of the system compared to PSS.
References 1. MajidDejamkhooy, A., Banejad, M., Talebi, N.: Fuzzy logic based UPFC controller for damping low frequency oscillations of power systems. In: 2008 IEEE 2nd International Power and Energy Conference, pp. 85–88 (2008). https://doi.org/10.1109/PECON.2008.4762450 2. Tambey, N., Kothari, M.L.: Damping of power system oscillations with unified power flow controller (UPFC). IEE Proc., Gener. Transm. Distrib. 150(2), 129 (2003). https://doi.org/10. 1049/ip-gtd:20030114 3. Arizadayana, Z., Irwanto, M., Fazliana, F., Syafawati, A.N.: Improvement of dynamic power system stability by installing UPFC based on fuzzy logic power system stabilizer (FLPSS). In:
40
4.
5. 6.
7.
8.
9.
10.
11.
12. 13.
V. Yarlagadda et al. 2014 IEEE 8th International Power Engineering and Optimization Conference (PEOCO2014), pp. 188–193 (2014). https://doi.org/10.1109/PEOCO.2014.6814423 He, T., Li, S., Shuijun, W., Li, K.: Small-signal stability analysis for power system frequency regulation with renewable energy participation. Math. Proble. Eng. 2021, 1–13 (2021). https:// doi.org/10.1155/2021/5556062 Liu, J., et al.: Impact of power grid strength and PLL parameters on stability of grid-connected DFIG wind farm. IEEE Trans. Sustain. Energy 11(1), 545–557 (2020) Goyal, S.B., Bedi, P., Rajawat, A.S., Shaw, R.N., Ghosh, A.: Multi-objective fuzzy-swarm optimizer for data partitioning. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 307–318. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_25 Tu, G., Li, Y., Xiang, J.: Analysis and control of energy storage systems for power system stability enhancement. Chin. Control Conf. (CCC) 2019, 560–565 (2019). https://doi.org/10. 23919/ChiCC.2019.8865814 Ghosh, M., et al.: Robustface recognition by fusing fuzzy type 2 induced multiple facial fused image. In: 2021 IEEE 4th International Conference on Computing, Power and Communication Technologies (GUCON), pp. 1–6 (2021). https://doi.org/10.1109/GUCON50781.2021.957 3871 Baadji, B., Bentarzi, H., Bouaoud, A.: SMIB power system model with PSS for transient stability studies. In: 2017 5th International Conference on Electrical Engineering - Boumerdes (ICEE-B), pp. 1–5 (2017). https://doi.org/10.1109/ICEE-B.2017.8191996 Al-Awami, A.T., Abdel Magid, Y.L. , Abido, M.A.: Simultaneous stabilization of power flow controller using particle swarm. In: 15th Power Systems Computation Conference, pp.1–7 (August 2005) Tambey, N., Kothari, M.L.: Damping of power system oscillations with unified power flow controller (UPFC). IEE Proc., Gene. Transm. Distrib. 150(2), 129 (2003). https://doi.org/10. 1049/ip-gtd:20030114 Gupta, N., Jain, S.K.: Comparative analysis of fuzzy power system stabilizer using different membership functions. Int. J. Comput. Electr. Eng. 2(2), 1793-8163 (2010) KarbalayeZadeh, M., Khatami, V., Lesani, H., Ravaghi, H.A.: A fuzzy control strategy to damp multimode oscillations of power system considering UPFC. In: 8th International Conference on Advances in Power System Control, Operation and Management (APSCOM 2009), pp. 1–5 (2009). https://doi.org/10.1049/cp.2009.1780
Computer Vision Approach for Visibility Enhancement of Dull Images Chandrika Acharjee(B) and Suman Deb Computer Science and Engineering Department, National Institute of Technology Agartala, Agartala, India [email protected]
Abstract. Most of the Computer Vision tasks emphasizes images with a standard level of illumination. Adversely, on the pragmatic ground, the applications are often challenged with low light input images. Howbeit, a low vision image captured under poor light conditions is prone to suffer from a considerable redundancy of valuable information. This often makes the image unsuitable for performing any computational task, and hence requires to be radiated and restored. To deal with this problem, in this paper, we discuss Computer Vision-based low vision image enhancement methods and analyse their accuracy. This paper stresses the approach of Histogram Equalization for visibility enhancement. However, this method was found to be inconsistent with actual relics within an image. Hence, furthermore we discuss another approach based on Dual Channel Prior for enhanced outcomes. A comprehensive study of these methods along with their performance and efficacy has been demonstrated in this paper and further research orientations in this work area have been proposed. Keywords: Low vision enhancement · Computer vision evaluation · Histogram equalization · Dual channel prior
1
· Efficacy
Introduction
Recovering information from a poorly lit image and developing a pragmatic low vision image enhancer is quite exacting as alongside mere lightning of the image, it also demands restoring any additional degradation, and effectively amplifying the image. Using a poor camera sensor or recording an image under low light often arouse noises in the image affecting its quality, making the image lack its clarity, and appear shady and indistinct. Unsatisfactory lightning condition infers restricted perception of relics from within the captured image. Hence, enhancing the visibility of such images is recommendable so as to restore the concealed details that result owing to the visible derogation caused by low light conditions. Alternatively stated, amplifying the illumination of the dark regions is required to intensify the image and make the disguised relics discernible. One approach discussed in this paper is the Histogram Equalization Algorithm that utilizes a cumulative distribution function to modify the resulting c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 41–52, 2022. https://doi.org/10.1007/978-981-19-1677-9_4
42
C. Acharjee and S. Deb
pixels to have an even pixel density distribution. It mainly increases the contrast of an image by stretching the intensity range of the most frequent pixels to make the disguised information on an image reappear. Howbeit, the output has been found to have irregular radiance and noise, and was conflicting the original information. A better approach towards solving this issue was to eliminate the problem of noise increment and over saturation. In this regard, another method discussed here is one based on dual dark and bright channel prior [1]. This technique computes transference mapping relation from the bright channel prior and further uses the dark channel prior to rectify any inaccurate transference approximations for enhanced visual outcome. In the next section, the implementation of the Computer Vision-based algorithms along with their preliminary outcomes has been demonstrated, and a comprehensible conclusion on their efficacy has been drawn.
2
Related Works
Numerous approaches to solving the issue of low light image enhancement have been proposed by various researchers till now. Detailed deliberation on these various methods can be found in[2]. Yu, Shun-Yuan et al. [3] proposed a substantial lightning model premised upon whose computed estimations, appearance of low light regions were enhanced. As per this model, illustration of any degradation varies with changing origins of light. Using knowledge about the preliminary surrounding light, the information loss was restored using weighted guide filter. This method was found to produce result quite consistent with the original scenes. Guo, Xiaojie and Li et al. [4] developed a method that considers the maximum pixel values of each of the R, G and B channel, and modifies the initial illumination map by rendering a structure prior on it, for restoring good texture of the image. This technique showed performance superior to any traditional approach with regard to quality and efficacy. Zhou, Zhigang and Sang, Nong and Hu et al. [5] suggested a technique of enhancing a low light image in terms of brightness and contrast in three specific steps - pliable reconciliation of global brightness, enhancement of localized contrast and colour restoration. This method was simple but still effective in its performance. Fu, Xueyang and Liao, Yinghao et al. [6] proposed a probabilistic method based on concurrent computation of illumination and reflectance. This technique was found to give visually satisfying results. Xu, Yadong and Yang, Cheng and Sun et al. [7] proposed a method that fixes the problem of complex physical models. It considered an enhancement technique employed on a remapping function that further considers global and local contrast features to extract information about the real scenes, thus restoring naturalness to the output.
Computer Vision Approach for Visibility Enhancement
43
Cai, Limei and Qian et al. [8] introduced the viability of fuzzy set representation of an image to enhance the low light regions. The algorithm here focuses on impeding the saturated bright regions and uniformly illuminating the dark regions. This technique was found to be efficient and could preserve the tone of the image well. Ren, Wenqi and Liu, Sifei and Ma et al. [9] proposed the approach of Convolutional Neural Network to feed the model learn concise mapping between a blazing and a dusky image. This technique has operational difference with other vision based methods which employ artificial frameworks for setting the enhanced visuals. In our study, we have focused primarily on vision-based methods for enhancing visibility of low vision images, as these methods were found to be of less computational complexity, and yet gave accurate outcomes along with efficient performance.
3 3.1
Image Visibility Enhancement Approach Low Light Image Characteristics
Generally, images captured under low light environment are prone to low contrast and low brightness, along with noise and colour contortion. Figure 1 shows two sample input low vision images along with their colour histogram representations. The X-axis represents the pixel intensities corresponding to red, blue and green color spaces in RGB, and the Y-axis represents the pixel number. From the two histogram representations, we can infer that all the channels have lower range and in the second histogram corresponding to the yet darker picture, it can be observed that all the channels have culminated at the bottom part of the X- axis, owing to the cause that black constitutes the dominant factor in the image in terms of tone. Besides, the edge details are mostly hidden, and the entire colour surfaces get prone to divergence leading to structure detail loss. As a result, such images become unsuitable for operating with any computer vision system. Images we mostly deal with on practical ground are RGB images. So, in order to enhance a low light image premised upon its corresponding RGB colour space, the three color channels R, G, and B are needed to be separately enhanced using the method of gray-scale image enhancement and merged accordingly. Figure 2 shows the schematic representation of this principle of enhancing the visibility of a low light RGB image. Howbeit, this principle neglects the correspondence between the different color components, and thus may lead to divergence in color restored in the final image.
44
C. Acharjee and S. Deb
Fig. 1. Sample low light images and their corresponding color histograms
Fig. 2. A schematic of enhancing a low light RGB image
3.2 3.2.1
Techniques Used Histogram Equalization Approach
Principle The basic proposition behind the Histogram Equalization approach is to take an image’s histogram into consideration and alter it accordingly in order to tune the contrast of the image [10]. This technique can increment the global contrast of a low vision image by expanding the peak over the entire range. The large peak in the histogram correlate with low contrast image with both the forefront and the background settings have similar tone [11]. Figure 3 shows some sample input low light images and their histogram representations.
Computer Vision Approach for Visibility Enhancement
45
Fig. 3. Left: Input low light gray scale images Right: Corresponding histogram representation
Implementation and Results The input in Histogram Equalization is always a gray-scale image. The steps involved in this method begins with the computation of a histogram representing image pixel intensities. This process is next followed by a uniform laying out and distribution of the most frequent pixel estimates. A linear trend is then given to the cumulative distribution function that forms the basis of the Histogram Equalization method. Figure 4 illustrates the basic concept behind the Histogram Equalization method. If the original input image is represented by M and the produced enhanced output image is characterized by N, the process can be defined by the relation N = f(M), where corresponding to a specific characteristic Q(m), M1 , M2 , .........., Mn represent n number of sub images having similar pixels as that in the original image which fulfill certain states. The measure of the image gray estimate is denoted by the parameter m, and N1 , N2 , ......, Nn represent the restored images analogous to the sub images that are later fused in line with the pixel locations after the restoration in order to acquire the final image N. The output results in images with increased global contrast. Figure 4 shows the corresponding output enhanced images obtained by Histogram Equalization.
46
C. Acharjee and S. Deb
Fig. 4. The basic concept of histogram equalization approach
Fig. 5. Left: Enhanced output images corresponding to input images in Fig. 3 produced by the histogram equalization method Right: Corresponding histogram representation
Performance Evaluation and Drawback Analysis The method of Histogram Equalization was found to be computationally simple. The efficiency has been found to be quite satisfactory, and does not require much of complex estimations. This method performs enhancement of the image as a whole. But, at times, some intrinsic details may get lost to the cause of graylevel blending. As this technique disregarding the colour space of the input image,
Computer Vision Approach for Visibility Enhancement
47
takes into account just the gray values for its functioning over the whole image, it cannot retrieve the local brightness and contrast traits as the original scenes. 3.2.2
Dual Channel Prior Based Approach
Principle The instinct behind this technique has been to compute the maximum and minimum pixel value information separately in order to obtain an initial transmission estimation, and to amend any inaccurate transmission, utilizing the bright and dark channel priors. The purpose behind using a dual channel prior roots to the fact that the bright channel alone would comprise most of the missed out details [12]. Howbeit, considering the dark channel furthermore would assist in eliminating block effects in several patches and hence can improve the visibility and precision of the outcome obtained using the bright channel prior [13]. Initially, dark and bright channel images which are nothing but the lowest and highest intensity pixel values in the local image [14] are obtained, following which a parameter called global atmosphere light is estimated. It imparts details about the degree of brightness in different regions of the image [15]. The corresponding transmission maps are created using the channels and atmospheric light value. To create an elevated image with filtered out noise, the transmission maps are further applied with guided filter. Implementation As part of the first step for illumination rectification of an input image, we define a function to obtain the radiance channel, where we estimate the dark and bright channel priors for all channels. Considering count of the channels and measure of each element, the respective channels are derived using matrix set with values from the original image array. In order to set values to the bright and dark channels, the maximum and minimum pixel estimates are obtained by applying a kernel to iterate over the entire image matrix. Numpy’s np.min function is used to compute the minimum pixel estimate in a patch and to access the dark channel. Likewise, the bright channel is accessed by using numpy’s np.max function that computes the maximum pixel estimate in that patch. The image dimensions are kept track of and padding by half the dimension of the kernel is applied to the image for ensuring the size equivalent to that of the original. Following this step, the global minimum and maximum array are further computed using cv::minmaxLoc function. Next, considering the top ten percent pixel intensity from the bright channel, the image array is levelled, remodeled, and arranged in order of the highest intensity. Mean of the top ten percent pixels from the levelled array is taken so as to inhibit even a little aberration from impacting it to any huge extent. The bright channel thus obtained is used to determine the Global Atmosphere lightning.
48
C. Acharjee and S. Deb
From the bright channel prior, the initial transmission map is computed. Illumination of the bright channel is obtained from a max operation done separately on the three color channels. Furthermore, characterizing the transmission as t(x), the illumination as I, Atmosphere light value as A, transmission in a small local pixel as y, denoting all the three r,g and b color channel by c, and assuming the radiance of a well lit patch to be 255, the following equation defines the process of initial Transmission estimation from the bright channel: c
Ibright (x) = max(max(I (y))) I(x)bright − Ac 255 − Ac Transmission map depicts the section of the radiance that is not strewed and get to the camera. On operating with a bright channel alone, for cases like patches lit with explicit radiance sources, the computed transmission disparity is likely to be erroneous, leading to inconsonant transmission. To deal with this problem, it is hence required to consider the lowest value in at the minimum one channel. In this attempt, we utilize the dark channel as an approving state to rectify any erroneous transmission computation of the bright channel. The difference between the bright and dark channel Ichanneldif f erence is obtained and is used to find if there is any irregular depth pixel. A predefined threshold value of 0.4 is taken into account for this particular estimation purpose. Ichanneldif f erence value less then 0.4 denotes a dark patch and irregular depth pixel. In this case, transmission from the bright channel is needed to be rectified using the transmission from the dark channel prior. Basic dark channel prior operation often causes even the rectified transmission to have structure mislaying during computation. Hence, to preserve the edges and to get a more refined image, the transmission map is further smoothen using a guided filter. The filter applied to get corrected transmission map T(x) also eliminates any halo error production. The following equation shows the production of enhanced image I* from the input image with illumination I(x): t(x)bright =
I∗ =
I(x) − A +A T (x)
where A is the estimated Atmosphere lightning value. The refined transmission map is transformed to a grayscale image to insure that the channel count in the original image and in the transformation map are equivalent. The output image obtained was then max-min normalized and returned (Fig. 6).
Computer Vision Approach for Visibility Enhancement
49
Fig. 6. Process flow for low light input image vision enhancement using Dual Channel Prior based method
Experimental Results To evaluate the effectiveness and overall performance of this technique, tests were conducted on a number of diverse images with varying degree of contrast and brightness. This method has been found effective in terms of radiance rectification, quality restoration and accuracy. It could quite efficiently address the problem of general brightness amplification caused by gray level blending in the previous technique of Histogram Equalization. Most of the intrinsic information and local contrast traits got restored. As part of its limitation, this method was found to work only on images with constant illumination property. Figure 5 shows the output obtained by this method. It can be seen that in an image with inconsistent illumination as in the second test image in Fig. 7, the output lacks in precision and cannot restore sharp edges with unambiguity. Hence, even if the output is quite picturesque in this case, a few regions do appear blurry. Also, on testing with an image comprising of some definite light source as in the fourth test image in Fig. 7, it has been observed that this method fell flat to restore and enhance the quality of the image. As this method is premised upon computing the atmosphere light value for constructing the transmission map in the process, such unequivocal light source in the image cause certain areas to be excessively manifested and consequently it retards the performance of the method. Figure 7 shows the enhanced output resulted from the Dual Channel Prior based approach applied on a number of sample test images with different degree of illumination and contrast.
50
C. Acharjee and S. Deb
Fig. 7. Left: Original low light images Right: Corresponding enhanced output images obtained by Dual Channel Prior based approach
Computer Vision Approach for Visibility Enhancement
4
51
Conclusion and Future Work
The Histogram Equalization approach was found to show quite satisfactory performance in terms of amplifying the brightness of the input low light images as a whole. Howbeit, being an algorithm of global functionality, this approach could not recover and optimize dim characteristic of a local region exclusively. As a result, the restored output image may be less detailed and may have inaccurate contrast. Hence, as part of future work into improving this particular approach, it is recommended to split the image into several divisions and apply the Histogram Equalization technique separately to each of the required local areas within the image in order to achieve an optimized output. The Dual Channel prior based approach on the other hand, has shown comparatively much better performance in terms of region-oriented visibility enhancement and could also restore the fine details quite accurately as expected. However, for cases such as the image illumination being inconsonant and the image having an explicit blaze or fluorescence, this method failed to show an optimized performance. Therefore, rectifying these limitations of this method as part of future work would require to modify the algorithm to deal with inconsistent illumination and to limit the over exposure of certain areas caused by the explicit blaze so as to further optimize the performance of this approach.
References 1. Shi, Z., Zhu, M., Guo, B., Zhao, M., Zhang, C.: Nighttime low illumination image enhancement with single image using bright/dark channel prior. EURASIP J. Image Video Process. 2018(1), 1–15 (2018). https://doi.org/10.1186/s13640-0180251-4 2. Wang, W., Wu, X., Yuan, X., Gao, Z.: An experiment-based review of low-light image enhancement methods. IEEE Access 8, 87884–87917 (2020) 3. Yu, S.-Y., Zhu, H.: Low-illumination image enhancement algorithm based on a physical lighting model. IEEE Trans. Circ. Syst. Video Technol. 29(1), 28–37 (2017) 4. Guo, X., Li, Y., Ling, H.: LIME: low-light image enhancement via illumination map estimation. IEEE Trans. Image Process. 26(2), 982–993 (2016) 5. Zhou, Z., Sang, N., Hu, X.: Global brightness and local contrast adaptive enhancement for low illumination color image. Optik 125(6), 1795–1799 (2014) 6. Fu, X., Liao, Y., Zeng, D., Huang, Y., Zhang, X.-P., Ding, X.: A probabilistic method for image enhancement with simultaneous illumination and reflectance estimation. IEEE Trans. Image Process. 24(12), 4965–4977 (2015) 7. Xu, Y., Yang, C., Sun, B., Yan, X., Chen, M.: A novel multi-scale fusion framework for detail-preserving low-light image enhancement. Inf. Sci. 548, 378–397 (2021) 8. Cai, L., Qian, J.: Night color image enhancement using fuzzy set. In: 2009 2nd International Congress on Image and Signal Processing, pp. 1–4. IEEE (2009) 9. Ren, W., et al.: Low-light image enhancement via a deep hybrid network. IEEE Trans. Image Process. 28(9), 4364–4375 (2019) 10. Cheng, H.D., Shi, X.J.: A simple and effective histogram equalization approach to image enhancement. Digit. Signal Process. 14(2), 158–170 (2004)
52
C. Acharjee and S. Deb
11. Senthilkumaran, N., Thimmiaraja, J.: Histogram equalization for image enhancement using MRI brain images. In: 2014 World Congress on Computing and Communication Technologies, pp. 80–83. IEEE (2014) 12. Lee, H., Sohn, K., Min, D.: Unsupervised low-light image enhancement using bright channel prior. IEEE Signal Process. Lett. 27, 251–255 (2020) 13. Lee, S., Yun, S., Nam, J.-H., Won, C.S., Jung, S.-W.: A review on dark channel prior based image dehazing algorithms. EURASIP J. Image Video Process. 2016(1), 1–23 (2016) 14. Park, S., Yu, S., Moon, B., Ko, S., Paik, J.: Low-light image enhancement using variational optimization-based retinex model. IEEE Trans. Consum. Electron. 63(2), 178–184 (2017) 15. Sandoub, G., Atta, R., Ali, H.A., Abdel-Kader, R.F.: A low-light image enhancement method based on bright channel prior and maximum colour channel. IET Image Process. 15, 1759–1772 (2021)
Tomato Leaf Diseases Detection Vishal Seth(B)
, Rajeev Paulus , Mayur Kumar , and Anil Kumar
Vaugh Institute of Agricultural Engineering and Technology (VIAET), Sam Higginbottom University of Agriculture Technology and Sciences, Prayagraj, Uttar Pradesh, India [email protected], {rajeev.paulus,mayur.kumar, anil.kumar}@shiats.edu.in
Abstract. The major risk to food safety is Plant diseases and now its has become important obstacle to identity plant diseases in the starting phase itself to minimize the financial loss. To overcome this challenge a state of the art Convolution Neural Network is proposed to classify tomato leaf diseases with the help of computer vision and provide superb result. The transfer learning method is also used to make the model cost effective. In this paper a transfer learning based CNN architecture is proposed to classify tomato leaf diseases. The dataset is combination of all the main public data available online like PlantVillage dataset, Plant Doc dataset and Mandely dataset which in total consist of 41863 images divided in 10 classes ranging from healthy to different types of tomato leaf diseases. The main focus is on ResNet50 network, a recognized CNN architecture which return the best accuracy of 99.2% and is extensively compared with existing work. The high accuracy makes the model practically useful that can be further extended to integrate with plant diseases classification system to operate in real world scenarios. Keywords: Tomato leaf disease · Deep learning · CNN · Feature extraction
1 Introduction The increasing population around the world has taken a big toll on farming with the demand to grow more on same amount of land or even less in most of the cases. This has created a sense of food insecurity around the world which is increasing even more because of the climate change, degeneration in pollinators and increase plant diseases etc. This leads us to topic of concern with is tomato plants. Tomato is one of a kind important economic crop, which contains vitamins and even can be used as a fruit. According to a survey of Food and Agriculture Organization of the United Nation, tomato diseases is the principal cause for the decrease in global tomato production, with a yearly loss rate up to 8% to 10% [1]. Anyhow, most of diseases in tomato plant begin from leaves and then escalate to the entire plant [2]. In many cases, there is no single treatment for a tomato plant disease, which means it’s up to you to prevent the spread of the disease. Tomatoes are susceptible to fungal infections like early blight, Septoria leaf spot as well as bacterial diseases like late blight, Mosaic virus and many more, so it’s important to keep up with the latest research © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 53–61, 2022. https://doi.org/10.1007/978-981-19-1677-9_5
54
V. Seth et al.
on tomato diseases. The conventional technique used for diseases detection and classification entirely depends upon the observer experience and knowledge or counsel of specialist. These methods are very slow and also require a lot of human labour, accuracy and time. The output is also not that good as it is costly and the efficiency depends on human knowledge. But with rapidly improving technologies of identification and features extraction the cost can be reduced and there is high growth in efficiency and accuracy. The application of deep learning in crop diseases and insect detection can highly reduce the identification time and work load resulting in time response. The paper proposes a deep learning model with the use of transfer learning, in which the pre-trained weights has been used to improve the feature extraction without increasing the cost. One of the main features of this paper is the combination of all publicly available tomato dataset do as to achieve a state-of-the-art model which can be used in field in real world scenario to give more stable and high accuracy in case of deep leaning methodologies [3]. The dataset consists of 10 different classes of diseases and healthy tomato images. The next section presents the problem of the related work followed by approach in the next section with include dataset details, method followed. Then we move to the result of the proposed model and comparison with related works in next section. At last conclusion is provided consisting summary of the proposed model.
2 Problem of Related Work The evolution of existing research come up with a practical and recommendation for the application of CNN in observation of plant disease and pest, can overcome the fault of regular machine visions methods used for feature extraction process. However, the computerized detection and identification of plant insect pests and diseases go through the following troubles [4–8]: • No clear partition exists between the healthy area and the infected area in case of plants. • Diseases have different shape, size and features in different stages of evolution during their life cycle, even at different positions. • Some diseases and insects may show similar or a bit different feature and there can be multiple diseases and insects at a particular position during same time. • It is hard to classify dead plant tissue from infected area. • The complex nature of image background makes the matter worst, as in addition of an infected leaves, it consists of multiple parts such as stem, crop, soil, etc. • During the life span of insects, they grow differently in different stages and the morphological features of insects vary greatly like small target pests.
3 Approach 3.1 Dataset Description The presented model was implemented on a combination of all the tomato dataset available online which also include Plant Village dataset and some other the personal dataset
Tomato Leaf Diseases Detection
55
used in different research papers such as Mendeley data [9–11]. The other dataset consist of 41863 leaves images of tomato plant which are further divided into 10 classes of data varying from healthy. Early blight, Bacterial spot, Leaf mold, and Spider mites etc. are some of the classes. However, the number of images in each class is not uniformly distributed ranging from 2500 to 4500 images in each classes. This entire dataset is divided into three main categories which are- train set, validation set and test set in the ratio of 80/19.8/0.2 where 80% is the train sets for creating the model, and 20% is to test the model accuracy which consist of test and validation sets. The standard augmentation techniques were already available inside these datasets such as taking mirror reflection, rotating image and varying the image intensity over part of images in the range of 20 to 30. So, no further augmentation is required (Table 1). Table 1. Dataset description S. no
Classes (Diseases)
Images
1
Tomato Bacterial spot
4366
2
Tomato Early blight
3493
3
Tomato Healthy
4066
4
Tomato Late blight
4335
5
Tomato Leaf Mold
3397
6
Tomato Mosaic virus
2667
7
Tomato Septoria leaf spot
4104
8
Tomato Spider mites
3856
9
Tomato Target Spot
3690
10
Tomato Yellow Leaf Curl Virus
7889
3.2 Method The growth of technologies has provided multiple ways to solve the problem of image classification in which the use of CNN (Convolution Neural Network) remains the predominant way. CNN’s [12] are the advancements of traditional neural networks with the main goal of extraction the similar patterns pattern is performed on the dataset. The use of layering greatly reduces the number of neurons required as compared to traditional neural network. For the purpose of image classification there are multiple variation of CNN are available which has been effectively to complex image dataset. The basic CNN architectures are like VGG [13], AlexNet [14], LeNet [15], GoogleNet [16], and ResNet [17]. The model proposed in this paper is also based one of this CNN architecture which is Resnet50 and is a 50-layer CNN model. ResNet50 is also known as Residual Neural Network which was developed by Kaiming He et al. in the year 2015. LeNet and VGGNet
56
V. Seth et al.
architectures have created multiple models for image classification. Still when creating models for more complex features to increase classification accuracy we require more deeper neural network which is more difficult as weights reach their saturation limit resulting in no change in accuracy or model degradation. ResNet was designed to overcome these problems which is also known as vanishing gradient problem shown in Fig. 1.
Fig. 1. Residual learning: a building block
In a stacked layer architecture if the weights become zero than the following stacked layers only make identity map resulting in a stagnant model. Whereas if the weights are not zero stacked layer are enabled to learn new features based on the input characteristics, resulting in better performance. 3.3 Performed Test Multiple testing in different experimental setups is executed to analyse the performance of proposed model. Several variations have been done to get the best possible result. The dataset is divided into three portion such as training, testing and validation set in the ratio of 80/19.8/0.02 respectively. Validation set consist of 100 images to test the model performance on unseen data to mimic real life conditions. Resulting in experiment with your proposed model. 3.4 Fine-Tuning Fine tuning is performed to make efficiency reach its optimal level. It makes tiny adjustment to improve the accuracy. It is very important for any neural network to decrease the computational resource consumption and increases the speed of the process. This process was repeated multiple time to improve the accuracy of proposed model. These changes are listed in Table 2 below.
Tomato Leaf Diseases Detection
57
Table 2. Parameters setup for the proposed model Parameter
Value
Batch Size
128
Epoch
12
Learning Rate
Default
Momentum
Default
Optimizer
Adam
Decay
Default
GPU
Tesla P4
RAM
12 GB
4 Result and Comparison This section provides the details of the model created on tomato dataset. The proposed model consists of 50 layers of Resnet50 with pre-trained weight from ImageNet. The layers after faltering have been customized accordingly to accomplice the task in which detection of tomato plant diseases has been done. The proposed model is able to detect 10 classes consisting of diseases and healthy ones with an accuracy of 99.2%. The source code of was written in TensorFlow environment of the deep learning framework and compilation is done on NVIDIA Tesla P4 GPU having 12 GB of ram provided by google Collaboratory. The higher amount of data ensures stable model and higher accuracy. The result of the proposed model is shown below in Fig. 2, 3, 4 and 5 respectively. Where Fig. 2 shows the confusion matrix which is used for calculation of accuracy over the validation set of databases. The accuracy of each class can be calculate with the use of Eq. 1 where numerator will be addition of all value in the primary diagonal and denominator will be addition of all the values in confusion matrix including primary diagonal whereas Fig. 3 shows the epochs and the variation of accuracy during testing and training phase of each epochs from which it is clear that epoch 12 give the maximum accuracy and Fig. 4 shows the variation of loss w.r.t epoch with early stopping implemented which ends the process after 12 epoch because of no significant change in losses. A sample of actual output from the proposed model has been shown in Fig. 5 below. Accuracy =
TP + TN Total no of image
(1)
The proposed model is then compared to not only other ResNet model but also all other important model algorithm of CNN under deep learning framework such as VGG16, AlexNet, GoogleNet and SqueezeNet. Table 3 clearly shows that the proposed model outperform all other algorithm and even other ResNet model.
58
V. Seth et al.
Fig. 2. Confusion Matrix of Proposed Model
Fig. 3. Final model performance with original images
Fig. 4. Variation of loss with Epoch
Tomato Leaf Diseases Detection
Fig. 5. Sample Output of proposed model Table 3. Performance comparison of proposed work with other existing works. S. no
Source
Type of features Method
Size of dataset
Accuracy (in %)
1
Chit Hlaing et al. [18]
Hand-Crafted features
Quadratic SVM
3535
83.5
2
Melike Sardogan et al. [19]
Automatic
CNN with LVQ
500
86
3
J. Shijie et al. [20]
Automatic
VGG16
7040
89
4
P. Tm et al. [21]
Automatic
LeNet
54306
95
5
Suryawati et al.[22]
Automatic
VGGNet
18160
95.24
6
Jayme Garcia et al.[23]
Automatic
GoogleNet
40409
96
7
Sladojevic et al. [24]
Automatic
CNN
30880
96.3
8
HalilDurmus et al. [25]
Automatic
SqueezeNe
54309
97.22
9
Keke Zhang et al. [26]
Automatic
ResNet
5550
97.28
10
Proposed approach
Automatic
ResNet50
41863
99.2
59
60
V. Seth et al.
5 Conclusion In this paper, we have successfully constructed a ResNet50 architecture to identify tomato plant diseases from leaf images. Our model can classify between 10 classes of healthy and diseases affected leaves. The model performance has been compared with other CNN architecture. The complex feature extraction has been done with help of 50-layer model and a combination of public datasets is used to make it more reliable and achieve higher accuracy. This will solve our purpose of early detection of tomato leaf diseases so that it can be cured and save user from financial and degradation losses. In future the model can be extended to other crops also and cured can also be suggested automatically by classification of diseases accordingly.
References 1. Zhu, X.K.: Research on Tomato Disease Identification Based on Convolutional Neural Network. Beijing University of Technology, Beijing, China (2020) 2. Sardogan, M., Tuncer, A., Ozen, Y.: Plant leaf disease detection and classification based on CNN with LVQ algorithm. In: Proceedings of the 2018 3rd International Conference on Computer Science and Engineering (UBMK), Sarajevo, Herzegovina, pp. 382–385 (2018) 3. Kamilaris, A., Prenafeta-Boldú, F.X.: Deep learning in agriculture: a survey. Comput. Electron. Agric. 147, 70–90 (2018) 4. Barbedo, J.G.A.: A novel algorithm for semi-automatic segmentation of plant leaf disease symptoms using digital image processing. Tropical Plant Pathol. 41(4), 210–224 (2016). https://doi.org/10.1007/s40858-016-0090-8 5. Barbedo, J.G.A.: A review on the main challenges in automatic plant disease identification based on visible range images. Biosyst. Eng. 144, 52–60 (2016). https://doi.org/10.1016/j.bio systemseng.2016.01.017 6. Barbedo, J.G.A.: A new automatic method for disease symptom segmentation in digital photographs of plant leaves. Eur. J. Plant Pathol. 147(2), 349–364 (2016). https://doi.org/10.1007/ s10658-016-1007-6 7. Barbedo, J.G.A.: Factors influencing the use of deep learning for plant disease recognition. Biosyst. Eng. 172, 84–91 (2018) 8. Barbedo, J.G.A.: Impact of dataset size and variety on the effectiveness of deep learning and transfer learning for plant disease classification. Comput. Electron. Agric. 153, 46–53 (2018). https://doi.org/10.1016/j.compag.2018.08.013 9. Mohanty, S.P., Hughes, D.P., Salathé, M.: Using deep learning for image-based plant disease detection. Front. Plant Sci. 7, 14–19 (2016) 10. Singh, D.P., et al.: PlantDoc: a dataset for visual plant disease detection. In: Proceedings of the 7th ACM IKDD CoDS and 25th COMAD (2020) 11. Huang, M.-L., Chang, Y.-H.: Dataset of tomato leaves. Mendeley Data 1 (2020) 12. Albawi, S., Mohammed, T.A., Al-Zawi, S.: Understanding of a convolutional neural network. In: Proceedings of 2017 International Conference on Engineering and Technology (ICET), pp. 1–6 (2017) 13. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: Proceeding of International Conference on Learning Representations (2015) 14. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems, vol. 1, pp. 1097–1105 (2012)
Tomato Leaf Diseases Detection
61
15. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proc. IEEE 86(11), 2278–2324 (1998) 16. Szegedy, C., et al.: Going deeper with convolutions. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9 (2015) 17. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9908, pp. 630–645. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46493-0_38 18. Hlaing, C.S., Maung Zaw, S.M.: Tomato plant diseases classification using statistical texture feature and color feature. In: Proceedings of 2018 IEEE/ACIS 17th International Conference on Computer and Information Science (ICIS), Singapore, pp. 439–444 (2018) 19. Sardogan, M., Tuncer, A., Ozen, Y.: Plant leaf disease detection and classification based on CNN with LVQ algorithm. In: Proceedings of 3rd International Conference on Computer Science and Engineering (UBMK), pp. 382–385 (2018) 20. Shijie, J., Peiyi, J., Siping, H., Haibo, S.: Automatic detection of tomato diseases and pests based on leaf images. In: 2017 Chinese Automation Congress (CAC), pp. 2537–2510 (2017) 21. Tm, P., Pranathi, A., SaiAshritha, K., Chittaragi, N.B., Koolagudi, S.G.: Tomato leaf disease detection using convolutional neural networks. In: Proceedings of 2018 Eleventh International Conference on Contemporary Computing (IC3), pp. 1–5 (2018) 22. Palimkar, P., Shaw, R.N., Ghosh, A.: Machine learning technique to prognosis diabetes disease: random forest classifier approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 219–244. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_19 23. Barbedo, J.G.A.: Plant disease identification from individual lesions and spots using deep learning. Biosyst. Eng. 180, 96–107 (2019) 24. Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., Stefanovic, D.: Deep neural networks based recognition of plant diseases by leaf image classification. Computat. Intell. Neurosci. 2016, 1–11 (2016). https://doi.org/10.1155/2016/3289801. Article ID 3289801 25. Chakraborty, A., Chatterjee, S., Majumder, K., Shaw, R.N., Ghosh, A.: A comparative study of myocardial infarction detection from ECG data using machine learning. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 257–267. Springer, Singapore (2022). https://doi.org/10.1007/978-98116-2164-2_21 26. Zhang, K., Wu, Q., Liu, A., Meng, X.: Can deep learning identify tomato leaf disease? Adv. Multimed. 2018, 10 (2018). Article ID 6710865
Comparing Switching Losses for Automatic and Program-Based Switched Capacitor Model for Active Cell Balancing Vipin Valsan(B) and Gairik Das Department of Electrical-Electronics Engineering, Manipal Institute of Technology, Manipal, India {vipin.valsan,gairik.das1}@manipal.edu
Abstract. With the increasing demand and constant comparison between electric vehicles and combustion vehicles. It is pivotal for electric vehicle companies to make the efficiency of the whole system as high as possible and therefore cell balancing circuits are important to extend the life cycle of batteries and extract maximum power from the batteries. In this paper we develop an existing automatic capacitive based cell balancing and compute the MOSFETs losses. Further we improve the existing model with a program-based model, and we improved the switching losses due to the MOSFET by controlling the gate pulse and also balancing time for the overall model. The proposed models are simulated in MATLAB and a comparison is carried out between the two. Keywords: Active cell balancing · Voltage balancing · Switched-capacitor · Battery · Energy storage system
1 Introduction Li-ion is one the most used portable energy storage device. It is highly preferred because of its high-power density and its ability to be recharged again and again. But when lithium-ion cells are used as a battery, each cell might not have the same potential. The reason for this anomalous behavior could have two reasons. One being the production of batteries is a very delicate process and it is very difficult, rather impossible to achieve the same voltage for all the cells. Secondly, each cell does not degrade according to the accepted ideal curve, some degrade faster some slower, causing a voltage difference between the two i.e., the self-discharge rates of the cells are different. Therefore, for maintaining the batteries peak performance we use various balancing techniques to balance out the batteries. Li-ion batteries are used very where from automation industry to the battery in our watches. They are one of the most pivotal energy storage devices, this is because the energy density of the batteries is quite high compared to its counterparts and also lithium lies at top of the periodic table; meaning it has one of the least atomic mass which greatly reduces the weight of the batteries. Even with meticulous planning the manufacturing © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 62–71, 2022. https://doi.org/10.1007/978-981-19-1677-9_6
Comparing Switching Losses for Automatic and Program-Based
63
of the batteries is not perfect. A cell has it goes through numerous cycles of discharge and charging they balance out [1–3]. It might happen so that when charging a cell may reach 100% charge while the other still might be at the 70% in such a case charging the whole system further will cause the former cell to overcharge and we might damage the cell or even worse it might catch fire, this is where balancing is required. Battery packs are stacked as serially as well as parallelly, so that when in parallel their voltage can add up, and when serially their current can add up and give us the desired power to drive that system connected to the batteries. A parallel-connected cell will automatically balance out and is not of concern to us. But in the series connection to maintain the flow of current from the battery to the connected equipment’s its crucial for us to maintain all cells at the same potential. A potential mismatch might cause the current to flow from one cell another, reducing the current for the equipment’s. There are two types of balancing techniques: passive balancing and active balancing. Passive balancing is where we bring the potential of all the cells to the potential of the lowest most cell, dissipating the excess energy through resistors connected. Whereas in active balancing we transfer energy from the cell with the highest potential to the cell with the least, thereby balancing the overall system (Fig. 1).
Fig. 1. Different types of cell balancing
The active cell balancing model can be further divided into capacitive, inductor or converter-based circuits [4–7]. Also, there are numerous balancing techniques cell-tocell, cell-to-pack, and pack-to-cell. In this paper we will be using cell-to-cell based capacitive based cell balancing model.
64
V. Valsan and G. Das
2 Single-Tired Switched Capacitor For this model, for N number of cells, we require 2N switches to operate and N-1 capacitors. It is an automatic cell balancing technique reducing the need for an intelligent control of the MOSFETs.
Fig. 2. Single-tired switched capacitor
As per Fig. 2, there is a continuous switching between state A and state B, for we use a pulse generator, each pulse generator is connected to the State A and State B MOSFET, respectively. Each pulse generator operates for one half of the time-period of the switching frequency. Therefore, each pulse generator operates for 0.0002 s with a 50% duty cycle. Since cells operate at around 3.2 V we used voltages ranging from 2.8 V – 3 V. For the calculation of the circuit parameters given by [8] we use the following formulas Ic (t) =
−t νb (t) − vC (t0− ) 1 × e Rc so, Vc (t) = ∫ Ic + Vc (t0 ) R c
(1)
Comparing Switching Losses for Automatic and Program-Based
65
And, VB_max Vc_max = < VB_min 2 2CRSC fs
(2)
Where, VC_max = Maximum ripple voltage across the capacitor. VB_max = Maximum voltage difference between two cells. VB_min = Minimum cell voltage. Based on the above we choose a capacitance C = 1 × 10−5 C. This model continuously switches between two adjacent cells. So, assuming 1st cell is at higher potential and the last cell at the least. The excess voltage will be serially transferred from the 1st to the last and balance out the system [9, 10]. This model has advantages like balance time for the cells is good, which still can further be reduced by connecting another tier of a capacitor between the 1st and the last but is superfluous and is not much of a concern for us and, it is a simple configuration with relatively minimal cost for setting up. The main disadvantage of the set is switching losses are high, it is high because this system is ON continuously connecting all the cells to the capacitor. So as the number of cells increases the switching loss also increases proportionally. The voltages of the cell are taken as V = [2.904, 2.984, 2.956, 2.865.2.811]. In the below graph we see that the cells balance out at 7 s at the average voltage of all the 5 cells i.e., 2.904 V (Fig. 3 and Fig. 4).
Fig. 3. Cell voltages
66
V. Valsan and G. Das
Fig. 4. Graph of the single-Tired balancing model
3 Losses in Mosfet To operate the MOSFET as a switch we either use it in the cut-off region (OFF) or the saturation region (ON). The losses in MOSFET are of two types of conduction loss and the switching loss. In conduction loss every MOSFET has some resistance, it dissipates power when current is conducted through the device, this is known as on-state resistance. Conduction losses are inversely proportional to the size of the MOSFET, the larger is the MOSFET the lower is the in-state resistance. The second type of loss is the switching loss, here as the MOSFET repeatedly switches between the two states, its intrinsic capacitance stores and then dissipates energy during the switching transition [11–14]. This loss is directly proportional to the stitching frequency and the parasitic capacitance of the MOSFET. Increasing the size of the MOSFET, the capacitance also increases and switching loss also increases (Fig. 5, Fig. 6 and Fig. 7). Taking values from the above table we have calculated the losses for each MOSFET. And multiplying it with the number of MOSFETs we get the total loss incurred due to the MOSFET. The important thing to note here is that the losses are completely variable, increasing the number of cells will increase the loss, and the loss has been calculated for one time period. Here we have obtained the losses as 0.238 W which is variable i.e., its directly proportional to the number of batteries in the circuit. Also, this loss is only for one time period, so if this system is ON for 1 min, the switching losses will be 198 W, which won’t be very efficient if the cells required frequent balancing, or the balancing time is large. As we keep on increasing the number of cells the losses will also increase proportional. And for electric vehicle the contain large number of cells, in which cases the losses due to the switching will also be very high, this is one of the biggest disadvantages of any automatic controlled active cell balancing models. Therefore, in the next part we will try to reduce this switching loss to as low as possible.
Comparing Switching Losses for Automatic and Program-Based
Fig. 5. Conduction and switching losses in MOSFET
Fig. 6. Calculation of losses in MOSFET
Fig. 7. Total losses for the single-tired model
67
68
V. Valsan and G. Das
4 Program Based Model In this model we control the MOSFET and turn it ON based on our requirement, we turn it ON only when two cells need to be balanced otherwise, it switched off, therefore reducing the losses. Similar to the previous model we used a cell-to-cell balancing method, and instead of a single-tired capacitive model we have used a single switched capacitive model. The gate pulses for the MOSFET are given based on our requirements. The MATLAB function block is used to control the MOSFET. It takes the input of the voltages of the cells. If the cell difference between the two cells is greater than the offset targets the function block will declare the battery pack as unbalanced. It will further arrange all the cells in an array in ascending or descending order and find the cells with the highestand lowest cell voltages. We find the average of all the cells and turn on the MOSFETs for the least and highest potential cell so that effective voltage transfer can take place. We again check for the voltage of the cells and arrange them in ascending order. This loop continues till the cell voltage of the least cell reaches the average value (Fig. 8 and Fig. 9).
Fig. 8. Schematics for the model
As shown in the above diagram all the MOSFETs with red tick relate to the first pulse generator, while the blue with the other. Assuming the number of cells used is N, the number of MOSFETs used are 4N-2. The two MOSFETs are used to obtain bidirectional control. For better control of all the MOSFETs, having two pulse generators is an added advantage, whenever we are controlling any red tick MOSFET is always
Comparing Switching Losses for Automatic and Program-Based
69
Fig. 9. Algorithm for the proposed model
used for charging the capacitor and when discharging for the blue tick, this perfectly isolates the two different processes. All the cells are provided with the same potential as in the previous model and we can see that the system balances a little over 2.9 V (average of all the cells = 2.904 V) with a balancing time of 5 s (as shown in Fig. 10 and Fig. 11). Coming to the advantage of this model, for anyone time-period, two MOSFETs will operate at any half of the time-period (one half for charging, one for discharging). So, no matter how many cells are out of potential, only 4 MOSFETs will be turned ON in one time-period, making our losses a constant for N number of cells. From the above calculations the switching losses comes out to be 0.09 W which is of a single time period, the losses for 1 min will be 79.33 W which is a constant. No matter how many cells we add up the switching losses for a particular time period will always be constant at that
70
V. Valsan and G. Das
Fig. 10. Losses in MOSFET
Fig. 11. Graph of program-based cell balancing model
particular frequency of operation. Therefore, we have reduced the switching losses by around 50% and also improved the balancing time of the overall system.
5 Conclusion With the increasing competition in the market, everybody wants their system to be as fast and as highly efficient as possible. Technology is very important today as it is used for almost everything and like everything, technology also has its advantages and disadvantages. The program-based model that has been designed also has its own advantages and disadvantages. The advantages, as discussed in the previous section include losses of the MOSFET being constant, having an improved balancing time. But coming to the disadvantage this model requires more MOSFET compared to the former (if for N number of cells, the program-based model required 4N-2 MOSFETs whereas the single-tiered capacitive model requires 2N MOSFETs). This disadvantage might increase the initial cost of production, but I believe we need to find the sweet spot on where each model might fit. For smaller applications, the single-tiered capacitive model would be a much better choice, but for larger models like cars, the program-based model would be an advantage (giving the extra mile of power back-up). As discussed, the disadvantage of a single-tired MOSFET i.e., the high Switching loss can be reduced to an extent by operating the MOSFET at the lowest possible switching frequency of the MOSFET.
Comparing Switching Losses for Automatic and Program-Based
71
References 1. Daowd, M., Omar, N., Van Den Bossche, P., Van Mierlo, J.: Passive and active battery balancing comparison based on MATLAB simulation. In: 2011 IEEE Vehicle Power and Propulsion Conference, pp. 1–7 (2011). https://doi.org/10.1109/VPPC.2011.6043010 2. Cheng, K.W.E., Divakar, B.P., Wu, H., Ding, K., Ho, H.F.: Battery-management system (BMS) and SOC development for electrical vehicles. IEEE Trans. Veh. Technol. 60(1), 76–88 (2011). https://doi.org/10.1109/TVT.2010.2089647 3. Pascual, C., Krein, P.T.: Switched capacitor system for automatic series battery equalization. In: Proceedings of APEC 97 - Applied Power Electronics Conference, vol. 2, pp. 848–854 (1997). https://doi.org/10.1109/APEC.1997.575744 4. Ye, Y., Cheng, K.W.E.: An automatic switched-capacitor cell balancing circuit for seriesconnected battery strings. Energies 9, 138 (2016). https://doi.org/10.3390/en9030138 5. Hemavathi, S.: Overview of cell balancing methods for li-ion battery technology. Energy Storage 3 (2020). https://doi.org/10.1002/est2.203 6. Omar, N., Daowd, M., Antoine, M., Lataire, P., Mierlo Van, J., Bossche, P., Van den: Battery management system—balancing modularization based on a single switched capacitor and bi-directional DC/DC converter with the auxiliary battery. Energies 7, 41 (2014). https://doi. org/10.3390/en7052897 7. Baughman, A.C., Ferdowsi, M.: Double-tiered switched-capacitor battery charge equalization technique. IEEE Trans. Industr. Electron. 55(6), 2277–2285 (2008). https://doi.org/10.1109/ TIE.2008.918401 8. Ye, Y., Cheng, K.W.E.: Analysis and design of zero-current switching switched-capacitor cell balancing circuit for series-connected battery/supercapacitor. IEEE Trans. Veh. Technol. 67(2), 948–955 (2018). https://doi.org/10.1109/TVT.2017.2749238 9. Ye, Y., Cheng, K.W.E., Fong, Y.C., Xue, X., Lin, J.: Topology, modeling, and design of switched-capacitor-based cell balancing systems and their balancing exploration. IEEE Trans. Power Electron. 32(6), 4444–4454 (2017). https://doi.org/10.1109/TPEL.2016.2584925 10. Oeder, C., Barwig, M., Duerbaum, T.: Estimation of switching losses in resonant converters based on datasheet information. In: 2016 18th European Conference on Power Electronics and Applications (EPE’16 ECCE Europe), pp. 1–9 (2016). https://doi.org/10.1109/EPE.2016. 7695275 11. Evzelman, M., Ben-Yaakov, S.: The effect of switching transitions on switched capacitor converters losses. In: 2012 IEEE 27th Convention of Electrical and Electronics Engineers in Israel, pp. 1–5 (2012). https://doi.org/10.1109/EEEI.2012.6377071 12. Kumar, M., Shenbagaraman, V.M., Shaw, R.N., Ghosh, A.: Digital transformation in smart manufacturing with industrial robot through predictive data analysis. In: Bianchini, M., Simic, M., Ghosh, A., Shaw, R.N. (eds.) Machine Learning for Robotics Applications. SCI, vol. 960, pp. 85–105. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0598-7_8 13. Shen, Z.J., Xiong, Y., Cheng, X., Fu, Y., Kumar, P.: Power MOSFET switching loss analysis: a new insight. In: Conference Record of the 2006 IEEE Industry Applications Conference Forty-First IAS Annual Meeting, pp. 1438–1442 (2006). https://doi.org/10.1109/IAS.2006. 256719 14. Guo, J., Ge, H., Ye, J., Emadi, A.: Improved method for MOSFET voltage rise-time and fall-time estimation in inverter switching loss calculation. In: 2015 IEEE Transportation Electrification Conference and Expo (ITEC), pp. 1–6 (2015). https://doi.org/10.1109/ITEC.2015. 7165790
Control System Design for DC-Link Voltage Control of Quasi-Z Source Inverter Using Bode Diagram Piyush B. Miyani(B) and Amit V. Sant Pandit Deendayal Energy University, Gandhinagar, India [email protected], [email protected]
Abstract. The quasi-Z source inverter (q-ZSI) enables single-stage power conversion and voltage boosting with the inclusion of the shoot-through state. The dual-loop control system is mostly adopted for the dc-link voltage control of quasiZ source inverters. In dual loop dc-link voltage control, the proportional integral and proportional controller is used for dc-link voltage control and inductor current control, respectively. However, the procedure of obtaining the controller parameter is not presented yet. This paper presents the procedure of control system design for the quasi-Z source inverter using a bode diagram. Bode diagram is obtained from a small signal model of quasi-Z source inverter after applying a small perturbation around the steady-state operating point. The presented procedure of control system design is further verified with the simulation studies of voltage-fed quasi-Z source inverter for set-point tracking and disturbance rejection using MATLAB Simulink environment. Keywords: Bode diagram · Controller Design · DC-link voltage control · QuasiZ source inverter
1 Introduction The demand for electricity is growing every day due to rising living standards of people and expansion of industry. Due to the clean nature of energy production, power electronics-based renewable energy sources are gaining more attention to fulfill rising electricity demand. Furthermore, due to rising fuel prices and climate change, electric vehicles are gaining popularity. For power conversion and grid integration, a voltage source inverter (VSI) is utilized. However, VSI has a few limitations, including the inability to enhance dc-link voltage and the inability to gate two power electronics switches at the same time [1]. Z-Source inverter (ZSI) has resolved the limitation of traditional VSI [2]. ZSI provides the single stage power conversion with buck-boost capability. However, ZSI has few drawbacks such as higher stress on passive components, discontinuous input current, and high inrush current [3]. The continuous input current with low stress on the passive components is provided by a quasi Z-source inverter (q-ZSI) [3, 4]. The q-ZSI has been © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 72–82, 2022. https://doi.org/10.1007/978-981-19-1677-9_7
Control System Design for DC-Link Voltage Control
73
investigated in renewable energy sources due to its unique advantages over VSI and traditional ZSI [5–10]. The shoot-through duty ratio and modulation index are regulated in order to control the dc-link voltage and output ac voltage, respectively [11]. To control shoot through, proportional-integral (PI) controller based single loop control, PI and P based dual loop control, sliding mode control, Fuzzy logic controller, and discrete-time model predictive control have been reported in [11–14]. The dc-link voltage control with dual loop control for ZSI has been reported in [15]. Dual-loop shoot through duty ratio control based on PI and P controller for q-ZSI have been reported in [16, 17]. However systematic approach for the derivation of controller parameters for dual loop controller using bode diagram is not presented yet. This paper presents a systematic approach for the derivation of controller parameter of PI in outer dclink voltage control and P in inner current control loop using bode diagram. In derivation of controller parameter, bode diagram is obtained from the small-signal model of q-ZSI. This Paper is structured as follows: quasi-Z source inverter is described in Sect. 2. Section 3 presents design of dual loop dc-link voltage controller using bode diagram approach. Section 4 presents simulation results for set-point tracking and disturbance rejection. The conclusion is derived in Sect. 5.
2 Quasi Z-Source Inverter The schematic of q-ZSI is depicted in Fig. 1. The q-ZSI mainly consists of the two inductances, L a , L b , two capacitors, C a , C b , and a diode, D. The voltage across both the inductors and capacitors are V La , V Lb , and V ca , V cb respectively. The voltage across the diode is V diode . The q-ZSI mainly operates in two modes namely shoot-through and non-shoot through mode.
Fig. 1. Schematic of voltage fed quasi-ZSI.
Figure 2(a) and Fig. 2(b) shows the equivalent circuit diagrams of the q-ZSI in shoot through and non-shoot through modes, respectively. Diode, D gets reverse biased in the shoot-through mode. As the voltage across both inductors are positive, both inductors are charged due to the input voltage source, V s , and capacitors. Whereas, Diode, D gets forward biased and the voltage across both inductors becomes negative in the non-shoot
74
P. B. Miyani and A. V. Sant
through mode. As a result, both inductors are discharged and both capacitors are charged in this mode.
Fig. 2. Equivalent circuit of quasi-ZSI in (a) Shoot-through mode. (b) Non-shoot through mode
3 Controller Design The small-signal model is used to design a control system for quasi-ZSI. In both the shoot-through and non-shoot through modes, the average model for quasi-ZSI can be obtained using volt-second balance in the inductors and ampere-second balance across the capacitors. Following the derivation of the average model, a small signal model can be obtained by applying a small perturbation around the steady-state operating condition. The detailed derivation of the small-signal model for q-ZSI has been reported in [16]. Based on a small-signal model, after taking Laplace transform, the two transfer functions (i) Gid (s) for inductor current, ˆiL (s), to shoot-through duty ratio, dˆ (s), and (ii) Gvd (s) for capacitor voltage, vˆ C (s), to dˆ (s) can be derived as shown in Eq. (1) and Eq. (2), respectively [16], where Dsh is the shoot-through duty ratio, iLa and iLb are the inductor currents, RC and rL are the capacitor resistance and inductor resistance respectively, iLoad is current supplied to the load during active states, V ac is phase to ground voltage, and L, C are the inductance and capacitance value of quasi–ZSI impedance network, respectively. Gvd (s) =
LI1 s + (RC + rL )I1 + (1 − 2Dsh )V1 vˆ C (s) = ˆd (s) LCs2 + (RC + rL )Cs + (1 − 2Dsh )2
(1)
Control System Design for DC-Link Voltage Control
Gid (s) =
ˆiL (s) S2 (1 − 2Dsh ) + S1 V1 = ˆd (S) S2 S1 + (1 − 2Dsh )2
75
(2)
In Eq. (1) and Eq. (2), V1 = VCa + VCb − iLoad RC
(3)
I1 = ILoad − iLa − iLb
(4)
S1 = LCs2 + (RC + rL )Cs
(5)
S2 = Ls + RC + rL
(6)
The inner current and outer voltage control loops are the key components of the dualloop control block diagram. The inner current loop is controlled by a proportional (P) controller, while the outside voltage control loop is controlled by a proportional-integral (PI) controller. To achieve the stable close loop control at the desired gain cross-over frequency corresponding to the selected bandwidth, the system should satisfy the following conditions[18]: 76◦ > Phase Margin > 45◦
(7)
|Gloop (s)|Wgc = |Gcomp (s)|wgc · |Gsystem (s)|wgc = 1
(8)
Condition−1 : Condition−2 :
Where Gloop (s) is the overall loop transfer function, Gcomp (s) is the compensator transfer function, and Gsystem (s) is the system transfer of the close path considering Gcomp (s) to 1. The detailed derivation of the controller parameter is described in the subsection (Fig. 3).
Fig. 3. Block diagram of dc-link voltage control scheme.
76
P. B. Miyani and A. V. Sant
3.1 Inner Current Loop Design A low pass filter is included in the inner current control loop configuration. The low pass filter lessens the shoot-through duty ratio’s ripple. As the first order low pass filter (LPF) is taken in this work, it changes the order of the system. Bode diagram is obtained considering the LPF and Gid (s) with the proportional controller, GiK keeping as unity. It can be seen from a bode diagram that the uncompensated inner current loop has a very high crossover frequency. Thus, in this work proportional controller is chosen to 0.0051 in order to obtain gain crossover frequency at 6280 rad/s which is 1/20th of the switching frequency, 20 kHz with the 62.7° phase margin (Fig. 4).
Fig. 4. Bode diagram of inner loop current control with proportional controller.
3.2 Design of Outer Voltage Control The bode diagram of the uncompensated and compensated loop are illustrated in Fig. 5. The uncompensated loop bode diagram is obtained from Gsystem (s), the cascade connection of the Gvd (s), and inner current loop close loop transfer function, GiCL (s) and PI compensator equal to unity. As compared to the inner current loop, the outside voltage control loop’s dynamics is designed to be slower. Thus, the outer voltage control loop is designed at the crossover frequency of 150 rad/s and the inner current loop crossover frequency, 6280 rad/s. The PI controller parameter can be obtained with the following steps: It can be observed from Fig. 5 that the gain and phase of GSystem (s) at wgc are 12.5 dB (4.2331) and 335.8°, respectively. Thus, to achieve stable closed-loop control at the desired gain crossover frequency, wgc , the loop transfer function must offer a gain of 0 dB (unity gain) and the phase margin, 70.8°. Thus, the PI controller should be designed in a manner such that it offers −12.5 dB and −85° phase at the desired gain crossover frequency, wgc .
Control System Design for DC-Link Voltage Control
77
The transfer function of PI compensator can be given as follows: GPI (s) = Gp +
Gi S
(9)
Where Gp and Gi represents the proportional and integral gain of PI compensator respectively. The angle and the amplitude of the PI compensator, GPI (s) at frequency, w, can be found as follows: −1 −Gi (10) Angle: GPI = tan wGp 2 Gi Amplitude: |GPI | = Gp2 + (11) w In order to obtain a phase angle of −85° through the PI controller based compensation, at wgc = 150 rad/s, desired phase angle (−85°) is substituted in Eq. (10) −Gi GPI = tan−1 = −85◦ , Gi = 1714.5Gp (12) wGp The value of gain offered by the system transfer function, GSystem , is 4.2331 (12.5 dB) as shown in Fig. 5(a). Thus PI compensator, GPI (s) should offer gain, |Gpi | = 1/4.2331 = 0.2362 to obtain 0 dB loop gain at wgc = 150 rad/s. Thus, from the relation between Gp and Gi in Eq. (11) and Eq. (12), |GPI | can be determined as 2 1714.5Gp 2 G i 2 2 |GPI | = Gp + = Gp + = 0.2362 (13) 150 150 After solving the above equation, the PI compensator gains Gp and Gi are obtained as 0.0206 and 35.3002, respectively. Hence, the transfer function for the PI compensator for the outer dc-link voltage control can be derived using bode diagram as follows: GPI (s) = 0.0206 +
35.3002 S
(14)
4 Simulation Results To verify the procedure of controller design, the simulation was performed in MATLAB Simulink. The designed controller’s performance was measured in terms of set-point tracking and disturbance rejection.
78
P. B. Miyani and A. V. Sant
Fig. 5. Bode diagram of outer voltage control loop.
4.1 Set-Point Tracking To verify effectiveness of the designed controller for set-point tracking, the source voltage waschanged during simulation as follows: 550 V from 0 to 0.4 s; 580 V from 0.4 to 0.7 s; and 565 V from 0.7 to 1 s. Here, Fig. 6 (a) and (b) illustrate V in and Dsh respectively. Figure 7 (a) and (b) illustrate indirectly estimated and directly measured dc-link voltage respectively. Figure 8 (a) and (b) illustrate I Load and V ac respectively. With the source voltage variation, Dsh was controlled from 0.12 to 0.08 at 0.4 s and 0.08 to 0.98 at 0.7 s, as shown in Fig. 6 and Fig. 7. The ac phase to ground voltage supplied to the resistive load, 32.5 was maintained constant even with the input voltage variationsby the designed controller by maintaining V dc to 750 V as shown in Fig. 8.
Fig. 6. (a) Input voltage. (b) Shoot-through duty ratio.
Control System Design for DC-Link Voltage Control
79
Fig. 7. (a) Indirectly estimated dc-link voltage. (b) Actual dc-link voltage.
Fig. 8. (a) Load current. (b) Phase to ground voltage.
4.2 Disturbance Rejection To verify disturbance rejection by the designed controller, the load was varied during the simulation. The load profile was changed as follows: 65 for 0 < t < 0.4 s; 32.5 for 0.4 < t ≤ 0.7 s; 65 for 0.7 < t < 1.0 s. Figure 9 (a) and (b) depicts I Load and Dsh respectively. Figure 10 (a) and (b) depicts indirectly estimated and directly measured dclink voltage respectively. Figure 11 shows V ac supplied to the load. From the simulation results illustrated in Fig. 9 and Fig. 10, it is revealed that Dsh was regulated from 0.095 to 0.12 at 0.4 s and 0.12 to 0.095 at 0.7 s to maintain dc-link voltage constant to 700 V. It can be observed from Fig. 11 that the load voltage remains constant despite changes in the load condition.
80
P. B. Miyani and A. V. Sant
Fig. 9. (a) Load current (b) Shoot-through duty ratio.
Fig. 10. (a) Indirectly estimated dc-link voltage. (b) DC-link voltage.
Fig. 11. Phase to ground voltage.
Control System Design for DC-Link Voltage Control
81
5 Conclusions The controller parameter of dc-link voltage controller have been derived with bode diagram approach using small signal model of q-ZSI. The performance of the designed dual loop voltage controller have been verified with the simulation in MATLAB Simulink environment. Simulation results verified the effectiveness of the designed controller in terms of set-point tracking and disturbance rejection with good steady state and dynamic response.
References 1. Raja Nayak, M., Tulasi, V.V.K., Divya Teja, K., Koushic, K., Suresh Naik, B.: Implementation of quasi Z-source inverter for renewable energy applications. Materials Today: Proceedings (2021) 2. Peng, F.Z.: Z-source inverter. IEEE Trans. Ind. Appl. 39, 504–510 (2003) 3. Anderson, J., Peng, F.Z.: Four quasi-Z-Source inverters. In: PESC Record - IEEE Annual Power Electronics Specialists Conference, pp. 2743–2749 (2008). https://doi.org/10.1109/ PESC.2008.4592360 4. Anderson, J., Peng, F.Z.: A class of Quasi-Z-source inverters. In: 2008 IEEE Industry Applications Society Annual Meeting. IEEE, pp. 1–7 (2008) 5. Li, Y., Anderson, J., Peng, F.Z., Dichen, L.: Quasi-z-source inverter for photovoltaic power generation systems. In: Conference Proceedings - IEEE Applied Power Electronics Conference and Exposition – APEC, pp. 918–924 (2009) 6. Abu-Rub, H., Iqbal, A., Ahmed, S.M., Peng, F.Z., Li, Y., Baoming, G.: Quasi-Z-source inverter-based photovoltaic generation system with maximum power tracking control using ANFIS. IEEE Trans. Sustain. Energy 4, 11–20 (2013) 7. Law, K.H., Loh, W.N., Wong, K.I.: Mathematics derivation and simulation modelling of maximum power point tracking based controller for Quasi Z source inverter. In: 2019 7th International Conference on Smart Computing and Communications, ICSCC 2019, pp.1–5 (2019) 8. Singh, N., Gupta, K.K., Jain, S.K., Dewangan, N.K., Bhatnagar, P.: A flying squirrel search optimization for MPPT under partial shaded photovoltaic system. IEEE J. Emerg. Sel. Top. Power Electron. 9, 4963–4978 (2020) 9. Santhoshi, B.K., Sundaram, K.M., Padmanaban, S., Holm-Nielsen, J.B., Prabhakaran, K.K.: Critical review of PV grid-tied inverters. Energies 12(10), 1921 (2019) 10. Bajestan, M.M., Madadi, H., Shamsinejad, M.A.: Control of a new stand-alone wind turbinebased variable speed permanent magnet synchronous generator using quasi-Z-source inverter. Electric Power Syst. Res. 177, 106010 (2019). https://doi.org/10.1016/j.epsr.2019.106010 11. Liu, Y., Abu-Rub, H., Xue, Y., Tao, F.: A discrete-time average model-based predictive control for a Quasi-Z-source inverter. IEEE Trans. Industr. Electron. 65, 6044–6054 (2018). https:// doi.org/10.1109/TIE.2017.2787050 12. Liu, Y., Abu-Rub, H., Ge, B.: Z-source/quasi-Z-source inverters: Derived networks, modulations, controls, and emerging applications to photovoltaic conversion. IEEE Ind. Electron. Mag. 8, 32–44 (2014) 13. Huneria, H.K., Yadav, P., Shaw, R.N., Saravanan, D., Ghosh, A.: AI and IOT-based model for photovoltaic power generation. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 697–706. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0749-3_55
82
P. B. Miyani and A. V. Sant
14. Rajaei, A.H., Kaboli, S., Emadi, A.: Sliding-mode control of z-source inverter. In: IECON Proceedings (Industrial Electronics Conference), pp.947–952 (2008) 15. Ellabban, O., Van Mierlo, J., Lataire, P.: A DSP-based dual-loop peak DC-link voltage control strategy of the Z-source inverter. IEEE Trans. Power Electro. 27(9), 4088–4097 (2012). https:// doi.org/10.1109/TPEL.2012.2189588 16. Liu, Y., Ge, B., Ferreira, F.J.T.E., De Almeida, A.T., Abu-Rub, H.: Modeling and SVPWM control of quasi-Z-source inverter. In: Proceeding of the International Conference on Electrical Power Quality and Utilisation, EPQU, pp. 95–101 (2011) 17. Li, Y., Jiang, S., Cintron-Rivera, J.G., Peng, F.Z.: Modeling and control of quasi-z-source inverter for distributed generation applications. IEEE Trans. Industr. Electron. 60, 1532–1541 (2013) 18. Mohan, N.: Power Electronics: A First Course. Wiley Global Education (2011)
Design and Simulation of Single Phase and Three Phase Wireless Power Transfer in Electric Vehicle Using MATLAB/Simulink Tareq Anwar Shikdar , Shornalee Dey , Sadia Mumtahina , Md Moontasir Rashid(B) , and Gulam Mahfuz Chowdhury Department of Electrical and Electronic Engineering, Leading University, Sylhet 3112, Bangladesh [email protected]
Abstract. Electric Vehicles (EV) are the fastest technology in the current world. For the reduction of environmental pollution, carbon dioxide emission and shortage of fossil fuels, the electric vehicle (EV) is the next-generation vehicle for transportation. This high-tech innovation is becoming more reliable and substituting combustion engine vehicles. The main issue with electric vehicles (EV) is the charging methods for the battery which are the main energy sources. Wireless power transmission (WPT) in electric vehicles (EV) is the most popular and advanced method for the charging of electric vehicles (EV). In this paper, a model is designed based on an inductively coupled power transfer methodology. The series-series (SS) wireless power transfer topology is considered for the proposed system, as well as for the system parameters which are discussed in this paper. Theoretical analysis with three simulated models using MATLAB/SIMULINK platform is shown and the performance of the proposed system is briefly discussed. Moreover, simulated results of inductively coupled wireless power transmission are validated for different modes of charging stations. Keywords: Electric vehicles (EV) · Inductive coupling · Simulink · Single-phase WPT · Three-phase WPT · Series-Series topology
1 Introduction Nowadays, air pollution by fossil fuels has become a major problem in our living hood. The most common modes of transportation for people are combustion engine vehicles which are run by fossil fuels and create environmental smog and global warming. For reducing the pollution and promoting green energy electric vehicles (EV) is the next-generation vehicle. The main concern of electric vehicles (EV) is the charging methodology which can be more facilitated by wireless charging technology [1]. The uses of wireless power transmission are not only for electric devices, smartphones, and laptops but also for electric vehicles (EV) [2] Wireless power transfer mode is the greatest innovation since when Nikola Tesla invented [3]. It has been rapidly growing technology for the consumers to get energy without wires. For achieving no power cable © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 83–104, 2022. https://doi.org/10.1007/978-981-19-1677-9_8
84
T. A. Shikdar et al.
lines for power transmission, power transmission faults, no arousing hazard, high efficiency, wireless power transmission has been gaining the trust by industrial application [4–6]. Wireless power transmission methodology is used without conductive wires [2, 7–9]. An inductive wireless charging system (WCS) for electric vehicles (EV) is more expedient, impermeable, safer, high efficiency, dynamic charging, and easier automation than conductive charging [10, 11]. The wireless power transfer as charging system has two parts: the primary side and secondary side. The Primary side (installed in the garage/road/charging station) gets power from the grid and serves to the coil for transferring the power to the secondary side (installed in EV) [12]. The secondary side should be light in weight for the reduction of the extra load [1]. In the designing section, another important factor is defining compensation topology [13]. There are mainly four types of compensation topology [14]: series-series (SP), series-parallel (SP), parallel-series (PS), parallel-parallel (PP) and these stand for how capacitors are connected to the transmitting and receiving coils. A Series-series (SS) compensation topology and parallel-series (PS) compensation topology are the convenience for power transfer efficiency [15]. For wireless power transmission as a charging system, series-series (SS) compensation topology is used in this paper. There are several types of power supply and different kinds of moods that can be implemented for wireless power transfer (WPT) such as single-phase and three-phase power supply can be applied in static, stationary and dynamic moods of WPT. The main aim of this paper is to design wireless power transfer for charging systems of electric vehicles (EV) for different power sources to consumers.
2 Literature Review Wireless power transfer for electric vehicles has become an integral part of modern science and technology innovation. There is so much research going on electric vehicles (EV) charging methodology. Several types of innovation come from the researchers such as: air gap reducing, efficiency increasing, charging moods innovating. In air gap reduction, a research team from MIT has been experimenting with 60 W power transfer for 2 m and it was a very early initiative for wireless power transmission by magnetic coupling [16]. Inspiring from this more experiments were conducted for increasing the transmission distance between two coils [17]. Besides that, Auckland university’s researchers designed a stationary wireless charging station for electric vehicles [18]. In another research, power transfer has been developed at 70% and 96% for 20 cm and 60 cm via magnetic coupling [19, 20]. In [21, 22] 60 KW for bus and 20 KW for SUV with the efficiency of 70% and 83% and 160 mm and up to 200 mm were designed respectively. University of Michigan-Dearborn designed a model of 8 KW WPT with 95.7% efficiency for 200 mm distance [23]. In wireless power transmission, for reducing power leakage and increasing power efficiency compensation topology is playing a major role. Some research reports say that, for the average power and high-efficiency power transmission series-series (SS) compensation technology is more convenient than others [24]. In other hand, inductor-capacitor-inductor (LCL) and inductor-capacitor-capacitor (LCC) is more efficient [25, 26] but series-series (SS) is suitable for large mutual inductance and inductor-capacitor-inductor (LCL) and inductor-capacitor-capacitor (LCC) is fitted for small mutual inductance [27]. The use of ferrite under lower frequency to 100 kHz
Design and Simulation of Single Phase and Three Phase Wireless Power
85
is no different from inductive wireless power transfer (IWPT). For charging the electric vehicles (EV) inductive wireless power transfer (IWPT) with 95% power transfer efficiency has been established in the EV market (reported by Bombardier and Halo) [28]. In the book of Wireless Power Transfer for Electric: Foundations and approach [12] introduce the several types of charging operational modes. Static WPT replaces the EV charging method when it is in the parking section for the charging [29] and it is similar to the conductive charging as many researchers have reported this fact [30, 31]. In Stationary charging, electric vehicles are in standby mode and recharge the battery such as public passenger stoppage and this was proposed in the Victoria project in Spain [32]. Also, there are several types of researches proposed for dynamic wireless power transmission for electric vehicles with high efficiency [33, 34]. In real life implementation Mercedes-Benz, Hyundai Kia, Continental, BMW and Audi companies take over the market [35–38].
3 Methodology 3.1 System Design The system design of the proposed has been shown in Fig. 1. Power supply, AC/DC converter, DC/AC converter, Primary compensation, secondary compensation and DCDC converter were respectively design and simulated as system design described.
Fig. 1. System design of the proposed system
3.2 System Process Figure 2 shows the design process/flowchart of the proposed system. Firstly, a power supply is required to operate the WPT system. There are two modes of power supply which are single-phase and three-phase. Secondly, a rectifier is designed for different power supplies. In the next step, inverter, coil and compensation topology is designed with resonance frequency for maximum power transfer. After power transmission, the rectifier is used for converting AC signal to DC signal. Lastly, a DC-DC buck converter is designed for transferring smooth and desired power to the battery.
86
T. A. Shikdar et al. Start
Power Supply
Designing Design Rectifier Rectifierfor forVarious VariousPower Powersupplies supplies
Modeling Inverter
NO
Coil Design with Compensation Topology
Operated in resonance frequency and maximum power transfer?
YES
Rectification of Receiving Power
Originating Buck Converter NO
Desired Smooth Output for EV Battery?
End YES
Fig. 2. Flow chart of the proposed system
4 System Modeling of Wireless Power Transfer in Electric Vehicle Wireless power transmission has two parts: primary side and secondary side shown in Fig. 3. The Primary side consists of power supply, power electronics, compensation topology, Tx coil. On the other hand, the secondary side consists of Rx coil, compensation topology, and power electronics. The Primary side has a rectifier and inverter but the secondary side has only a rectifier. The primary side gets excited from the power source and fed to the secondary side. The system is operated in resonance frequency for maximum power transfer from the primary side to the secondary side. Compensation topology helps to achieve the resonance frequency as well as the topology along with the coil combined create a resonant circuit [15]. Resonance frequency determined as: 1 ω0 = √ LC ω0 So, f0 = 2 1 = √ 2 LC
(1)
(2)
Co-efficient of magnetic coupling between Tx coil and Rx coil can be determined as: M k=√ L1 L2
(3)
Design and Simulation of Single Phase and Three Phase Wireless Power
87
Fig. 3. Block diagram of the Wireless Power Transmission
Where M is the mutual inductance between L1 and L2 is the Tx coil and Rx coil. By applying the Eq. (2) mutual inductance can be calculated [39–43]. Mutual induction formula can be expressed as M = k L1 L2 (4) For series-series compensation, L1 =L2 is the self-inductance of transmitting and receiving coil, C1 = C2 compensation capacitor of transmitting coil and receiving coil. So, f0 =
1 ω0 1 = = √ √ 2 2 L1 C1 2 L2 C2
(5)
5 Simulation Models and Design of Wireless Power Transfer in Electric Vehicle This paper is based on three individual models for wireless power transfer as a charging method for electric vehicles. These models are: MODEL I: Single-phase wireless power transfer in EV for a single consumer. MODEL II: Three-phase wireless power transfer in EV for a single consumer. MODEL III: Three-phase wireless power transfer for in EV multiple consumers. Simulation models, design, calculation and configuration of the proposed system are described below: 5.1 Circuit Analysis The system configuration of three individual models are shown in Fig. 4 (MODEL I), Fig. 5 (MODEL II), Fig. 6 (MODEL III).
88
T. A. Shikdar et al.
Fig. 4. System configuration of single-phase wireless power transfer for a single consumer (MODEL I)
Fig. 5. System configuration of three-phase wireless power transfer for a single consumer (MODEL II)
Fig.6. System configuration of three-phase wireless power transfer for multiple consumers (MODEL III)
5.2 Simulation Model Analysis From Fig. 7, Fig. 8 and Fig. 9, it is observed that the systems are simulated by several subsystems. The power supply is fed to the rectifier and then the rectifier transmits the filtered power to the inverter subsystem for transferring the power from the primary side to the secondary side via series-series (SS) compensation topology with the primary coil. On the secondary side, a rectifier connected to the buck converter for transferring dc voltage and a load is connected across the buck converter.
Design and Simulation of Single Phase and Three Phase Wireless Power
89
Fig. 7. MATLAB Simulation model of single-phase wireless power transfer for a single consumer (MODEL I)
Fig. 8. MATLAB Simulation model of three-phase wireless power transfer for a single consumer (MODEL II)
Fig. 9. MATLAB Simulation model of three-phase wireless power transfer for multiple consumers (MODEL III)
90
T. A. Shikdar et al.
Table 1 shows the parameters considered for each model. Table 1. Model parameters considered for each model Parameters
Design value
Input voltage
230 V, 50 Hz
Self-inductance of primary side, L1
20 mH
Self-inductance of secondary side, L2
20 mH
Primary side compensation capacitor, C1
5 mF
Secondary side compensation capacitor, C2
5 mF
Primary Side resistance, R1
0.7
Secondary Side resistance, R2
0.7
Mutual inductance, M
10 mH
Filter 1
1000 µH
Filter 2
1000 µH
Co-efficient, k
0.5
5.3 Primary Side Rectifier Design An AC/DC Rectifier subsystem is used in the proposed design. AC/DC Rectifier subsystem of Fig. 7 (MODEL I) and Fig. 9 (MODEL III) are shown elaborately in Fig. 10 and subsystem of Fig. 8 (MODEL II) is shown in Fig. 11. The primary rectifier is connected to the power supply. For an uninterrupted power supply to the inverter section, a controlled rectifier is used in the primary side instead of the uncontrolled rectifier.
Fig. 10. AC/DC Primary Side Rectifier subsystem of MODEL I and MODEL III.
To trigger the rectifier, four pulse generators were added to the detailed thyristor for MODEL I and MODEL III, also six pulse generators were added to the detailed thyristor
Design and Simulation of Single Phase and Three Phase Wireless Power
91
Fig. 11. AC/DC Primary Side Rectifier subsystem of MODEL II
for MODEL II. Period and phase delay must be calculated [44], so the formula can be expressed as: Period , T = andPhasedelay =
1 f
T ∗ Firingangle 360◦
(6) (7)
Rectifier calculation for Single-phase wireless power transfer in EV for a single consumer (MODEL I): Using Eqs. (6) and (7), period is calculated 0.02 and phase delay for detailed thyristor 1, 4 and 2, 3 is 0 and 0.01 respectively. After rectifying the power from power supply, it is served to the inverter section. Rectifier calculation for three-phase wireless power transfer in EV for a single consumer (MODEL II): Period is calculated 0.2 and phase delay is calculated for detailed thyristor 1, 2, 3, 4, 5 and 6 is 0.0016, 0.005, 0.083, 0.116, 0.15 and 0.183 respectively from Eqs. (6) and (7). Rectifier calculation for three-phase wireless power transfer in EV for multiple consumers (MODEL III): Period for all phases (a, b, c) is 0.2 using Eq. (6). Phase delay for each phase is described below by using Eq. (7). ‘a’ phase: Phase delay for thyristor 1, 4 and 2, 3 is 0 and 0.01 respectively. ‘b’ phase: Phase delay for thyristor 5, 8 and 6, 7 is 0.06 and 0.16 respectively. ‘c’ phase: Phase delay for thyristor 1, 4 and 2, 3 is 0 and 0.01 respectively. 5.4 Inverter Design DC/AC Inverter subsystem for all models are shown in Fig. 12. The inverter subsystem is formed with IGBT/Diode. Two pulse generator is connected to the IGBT/Diode for the functioning of the inverter to fed the rectified power supply to the primary coil.
92
T. A. Shikdar et al.
Fig. 12. Inveter subsystem of MODEL I, MODEL II and MODEL III
5.5 Secondary Side Rectifier Design An AC/DC Uncontrolled Rectifier subsystem is used in the proposed models shown in Fig. 8. The secondary side is installed in electric vehicle. The coil and power electronic devices of secondary side should be small in size and lighter in weight so that uncontrolled rectifier’s weight becomes much lighter than controlled rectifier [1] (Fig. 13).
Fig. 13. AC/DC Secondary Side Rectifier subsystem of MODEL I, MODEL II, MODEL III
5.6 DC-DC Buck Converter Design DC-DC buck converter subsystem is used for stepping down the rectified power on the secondary side and smoothing the output power as desired for the electric vehicle’s battery section is shown in Fig. 14. To design the buck converter, MOSFET is used as a switch and a pulse generator is connected to the gate. Values of buck converter’s parameters needed to be calculatedfrom below equations [45]. The formulas are determined as: v0 = vin × D iR =
v0 RL
(8) (9)
Design and Simulation of Single Phase and Three Phase Wireless Power
93
Fig. 14. DC-DC buck converter subsystem of MODEL I, MODEL II, MODEL III
and , L = C=
(vin − v0 )D ˙iL × f
(1 − D)v0 8Lf 2 v0
(10) (11)
6 Simulation Results of the Proposed System Two types of simulation results are observed. Firstly, voltage output, current output and power output are perceived without a buck converter which is shown separately for individual simulation models (Fig. 7, Fig. 8 and Fig. 9). Lastly, output (voltage, current, power) is observed with a buck converter. All the simulation results and observations are described below: 6.1 Simulation Results of Single-Phase Wireless Power Transfer in EV for a Single Consumer 6.1.1 Simulation Results Without Buck Converter Figure 15 shows the output voltage without buck converter and any load. It is the actual voltage of this proposed system. This output voltage is observed with the Filter 2 capacitor of Fig. 8. The Figure shows that the output voltage crossing 200 V where the input voltage is single-phase 230 V.
94
T. A. Shikdar et al.
Fig. 15. Output voltage across Filter 2 (without buck converter and 100 load)
Figure 16 shows the waveform of output voltage, output current and output power with 100 load and without buck converter of single-phase wireless power transfer for a single consumer (MODEL I).
a) Output Voltage
b) Output Current
c) Power Output
Fig. 16. Output results across 100 load (without buck converter)
In Fig. 16(a), output voltage with load is observed. The output voltage is reduced due to 100 load and it is connected across to the secondary rectifier. The output current and output power are shown in Fig. 16(b) and Fig. 16(c) and it is observed with 100 load respectively. Output current and output power are also reduced after connecting load. Values are observed for output voltage, output current and output power are about 210 V, 2.1 A and 440 W respectively. In Fig. 16, the output result’s waveform is not glossy and smooth. To achieving the desired, a buck converter can be used and provided the smooth waveform of output voltage, output current and output power.
Design and Simulation of Single Phase and Three Phase Wireless Power
a) Output Voltage
95
b) Output Current
c) Power Output
Fig. 17. Output results across 100 load (with buck converter)
6.1.2 Simulation Results with Buck Converter Buck converter is mainly used for transferring the desired power to the battery section as converting high voltage to low voltage. Figure 17 shows the output voltage, output current and output power simulation result respectively with a buck converter. Figure 17 shows the output values are stepped down to about 148 V, 1.4 A and 208 W. Figure 17 shows the waveform of output voltage, output current and output power with 100 load and buck converter of single-phase wireless power transfer for a single consumer (MODEL I). In Fig. 16 the output waveforms are not smoothed but after connecting the buck converter the waveform is smoothed and glossy for transferring the power to the battery shown in Fig. 17 and it is observed that the outputs are fully desired waveforms for transferring the power to the battery section for charging the electric vehicle. 6.2 Simulation Results of Three-Phase Wireless Power Transfer in EV for a Single Consumer 6.2.1 Simulation Results Without Buck Converter The definite output voltage without buck converter and any load but with filter 2 of Fig. 9 is shown in Fig. 18. The figure shows that the output voltage crossing 500 V where the input voltage is three-phase 230 V. Figure 19 output voltage, output current and output power with load are spotted. All Output is reduced due to 100 load which is connected across to the secondary rectifier. Values are observed for output voltage, output current and output power is about 505 V, 5.2 A, 2.6 kW after connecting 100 load respectively.
96
T. A. Shikdar et al.
Fig. 18. Output voltage across Filter 2 (without buck converter and 100 load)
Figure 23 shows the waveform of output voltage, output current and output power with 100 load and without buck converter of three-phase wireless power transfer for a single consumer (MODEL II).
a) Output Voltage
b) Output Current
c) Power Output
Fig. 19. Output results across with 100 load (without buck converter)
In Fig. 19, the output result’s waveform is not smooth and desired for the battery. To attaining the smooth graph buck converter is used. 6.2.2 Simulation Results with Buck Converter In electric vehicles, the battery is mainly used as a load for charging. The battery needs smooth power for transferring to the engine section. Figure 20 to shows the output voltage, output current and output power simulation result respectively with a buck converter. Figure 20 shows the waveform of output voltage, output current and output power with 100 load and without buck converter of three-phase wireless power transfer for a single consumer (MODEL II).
Design and Simulation of Single Phase and Three Phase Wireless Power
a) Output Voltage
97
b) Output Current
c) Power Output
Fig. 20. Output results across 100 load (with buck converter)
Figure 20 shows the output result is stepped down about to 350 V, 3.5 A and 1.3 kW. It is observed that the output is fully smooth waveform and desired for transferring the power to the battery. 6.3 Simulation Results of Three-Phase Wireless Power Transfer in EV for Multiple Consumers 6.3.1 Simulation Results Without Buck Converter In this simulation result, output result is observed for each phase (a, b and c phase) as an individual phase has output voltage, output current and output power. The actual and absolute voltage of per phase is observed without buck converter and any load but with Filter 2 of Fig. 10 shown in Fig. 21 for a, b and c phase.
Fig. 21. Output voltage of ‘a’, ‘b’, and ‘c’ phase across Filter 2 (without buck converter and 100 load)
98
T. A. Shikdar et al.
a) Output Voltage
b) Output Current
C) Output power
Fig. 22. Output results of ‘a’, ‘b’ and ‘c’ phase with 100 load (without buck converter)
In Fig. 22, output voltage, output current and power is observed with load is observed with 100 load respectively. Output waveforms are reduced due to 100 load and it is connected across to the secondary rectifier. Figure 22 shows the waveform of output voltage. Output current and output power of ‘a’, ‘b’ and ‘c’ phase with 100 load andwithout buck converter of three-phase wireless power transfer for multiple consumers (MODEL III). The figures show that, every phase has similar waveforms of output voltage, output current and output power and the values are about to 300 V, 3 A and 900 W for a, b and c phase. The output result is not desired waveforms for the battery so that buck converter is used for converting DC-DC glossy output waveforms. 6.3.2 Simulation Results with Buck Converter Buck converter isfor transferring the desired power to the battery section of the electric vehicle. Figure 23 shows the output voltage, output current and output power simulation result respectively with a buck converter. Figure 23 shows the waveform of output voltage, output current and output power with 100 load and buck converter of three-phase wireless power transfer for multiple consumers (MODEL III). The figures show the output results are stepped down about to 210 V, 2.1 A and 440 W. In Fig. 22, the output waveforms are not smoothed. When connecting the buck converter to the proposed system the waveform is smoothed and glossy for transferring the power to the battery shown in Fig. 31 and the outputs are fully desired waveforms for transferring the power to the battery section.
Design and Simulation of Single Phase and Three Phase Wireless Power
a) Output Voltage
99
b) Output Current
c) Output power
Fig. 23. Output results of ‘a’, ‘b’ and ‘c’ phase with 100 load (with buck converter)
7 Overall Performance Analysis In this paper, three individual models are designed and simulated. Table 2 shows the overall results of simulation analysis of all the models. Table 2. Overall performance analysis Model
Input Voltage
Output results without buck converter
Output result with buck converter
Voltage (V)
Current (A)
Power (W)
Voltage (V)
Curren (A)
Power (W)
I
230v Single-phase
210
2.1
440
148
1.4
208
II
230v Three-phase
505
5.2
2600
350
3.5
1300
III
230v Three -phase
300
3
900
210
2.1
440
The output of secondary rectifier of all the models are not smooth and not glossy waveforms. For battery, a buck converter is used. After connecting the buck converter, output shows losses of some values in all the models. As well as, Output results drop after transmission from the primary side to the secondary side. It is observed that there is loss of some voltage, current and power in every section.
100
T. A. Shikdar et al.
8 Operational Modes of the Proposed System There are mainly three operational modes for charging electric vehicles such that i) Static ii) Stationary and iii) Dynamic [32]. In this paper, there are three different types of models proposed. These models can be used in different operational modes namely: 1) Single-phase wireless power transfer in EV for a single consumer as Static Mode 2) Three-phase wireless power transfer in EV for a single consumer as Stationary Mode 3) Three-phase wireless power transfer for in EV multiple consumers as Stationary and Dynamic Mode 8.1 Single-Phase Wireless Power Transfer in EV for a Single Consumer (MODEL I) as Static Mode Single-phase wireless power transfer in EV for a single consumer (MODEL I) can be used as Static mode. A consumer can use this charging method in parking areas, homes, offices, and also in shopping cities. This method will be user-friendly for consumers and it will take time to be fully charged as the output power is low (Fig. 21). Figure 24 shows the illustrative description and operational application of the static mode for single-phase wireless power transfer in EV for a single consumer.
Fig. 24. Illustrative representation of Single-phase wireless power transfer in EV for a single consumer (MODEL I) as Static Mode.
8.2 Three-Phase Wireless Power Transfer in EV for a Single Consumer (MODEL II) as Stationary Mode In stationary mode, three-phase wireless power transfer in EV for a single consumer can be established. This mode can be used in the traffic signal stoppage areas where if an electric car stops at that point, the car can be charged by this mode. MODEL II has high power (Fig. 28) output than MODEL I and in traffic signal stoppage, the time for charging is less than the scenario occurs at homes or parking areas so that MODEL II will be very efficient for stationary mode but the cost will be very high. Figure 25 shows the operational application of this mode by illustrative form.
Design and Simulation of Single Phase and Three Phase Wireless Power
101
Fig. 25. Illustrative representation of three-phase wireless power transfer in EV for a single consumer (MODEL II) as Stationary Mode
8.3 Three-Phase Wireless Power Transfer in EV for Multiple Consumers (MODEL III) as Stationary and Dynamic Mode MODEL III can be used in two modes. In the road stoppage area or traffic signal stoppage area, a big parking load for multiple consumers, stationary mode can be used by MODEL III. Moreover, in the wireless charging lane, MODEL III can be used as dynamic mode. MODEL III has different phases for using different consumers at a time so that MODEL III will be more efficient for these modes. Figure 26 shows an illustrative demonstration of the MODEL III as dynamic mode. MODEL III as stationary mode is similar to Fig. 25.
Fig. 26. Illustrative representation of three-phase wireless power transfer in EV for a single consumer (MODEL III) as Dynamic Mode.
9 Conclusions In this paper, WPT charging methods for different consumers with different power supplies were simulated in MATLAB/SIMULINK. Total observation shown in simulation results and performance analysis was discussed. Firstly, the models were observed without buck converter and figured that the results should be improved for battery, also it is
102
T. A. Shikdar et al.
depicted that when filter 2 is not connected with the load, the secondary rectifier output voltage is crossing the input voltage. After connecting load, the voltage is reduced. Lastly, simulation models were notifying the buck converter results in smooth waveform and it is desired for the battery. Side by side, the operational modes of the proposed systems were illustrated in this paper. However, wireless power transmission has some disadvantages such as loss of power while delivering output. In future work, renewable energy using as a power source and reducing power loss will be necessary and priority tasks for these proposed models.
References 1. Baroi, S., Islam, M.S., Baroi, S.: Design and simulation of a wireless charging system for electric vehicles. In: 2017 2nd International Conference on Electrical & Electronic Engineering (ICEEE), pp. 1–4 (2017). https://doi.org/10.1109/CEEE.2017.8412915 2. Wu, K., Matsumoto, H., Choudhury, D.: Wireless power transmission, technology, and applications. Proc. IEEE 101(6), 1271–1275 (2013) 3. Tesla, N., Marincic, A.: Nikola Tesla and the wireless transmission of energy. IEEE Trans. Power App. Syst. 101(10), 4064–4068 (1982) 4. Covic, G.A., Boys, J.T.: Inductive power transfer. Proc. IEEE 101, 1276–1289 (2013) 5. Li, Y., Mai, R., Lin, T., Sun, H., He, Z.: A novel WPT system based on dual transmitters and dual receivers for high power applications: analysis, design and implementation. Energies 10, 174 (2017) 6. Jang, Y.J., Jeong, S., Lee, M.S.: Initial energy logistics cost analysis for stationary, quasidynamic, and dynamic wireless charging public transportation systems. Energies 9, 783 (2016) 7. Lu, X., Niyato, D., Wang, P., Kim, D., Han, Z.: Wireless charging technologies: fundamentals, standards, and network applications. IEEE Commun. Surv. Tutor. 18(2), 1413–1452 (2015) 8. Garnica, J., Lin, J., Chinga, R.: Wireless power transmission: from far field to near field. Proc. IEEE 101(6), 1321–1331 (2013) 9. Smith, C.: Wireless charging just got even better. http://bgr.com/2014/02/12/a4wp-pma-all iance/ (2014) 10. Yilmaz, M., Krein, P.T.: Review of battery charger topologies, charging power levels and infrastructure for plug-in electric and hybrid vehicles. IEEE Trans. Power Electron. 28, 2151– 2169 (2013) 11. Chopra, S.: Contactless power tranfer for electric vehicle charging application. Master’s Thesis, Delft University of Technology, Delft, The Netherlands (2011) 12. Triviño-Cabrera, A., González-González, J.M., Aguado, J.A.: Wireless power transfer for electric vehicles: foundations and design approach. Springer Nature Switzerland AG, ISSN 1612-1287 ISSN 1860-4676 (electronic) Power Systems ISBN 978-3-030-26705-6 ISBN 978-3-030-26706-3 (eBook) (2020). https://doi.org/10.1007/978-3-030-26706-3 13. Wang, S., Chen, J., Hu, Z., Lui, M.: Study on series-parallel mixed resonance model of wireless transfer via magnetic resonance coupling. In: Progress in Electromagnetic Research Symposium (PIERS), pp. 2941–2945 (2016) 14. Li, S., Nguyen, T.D., Li, W., Deng, J., Mi, C.: A double-sided LCC compensation network and its tuning method for wireless power transfer. IEEE Trans. Veh. Technol. 64(6), 2261–2273 (2015) 15. Rakhybay, A., Bagheri, M., Lu, M.: A simulation study on four different compensation topologies in EV wireless charging. In: 2017 International Conference on Sustainable Energy Engineering and Application (ICSEEA) (2017)
Design and Simulation of Single Phase and Three Phase Wireless Power
103
16. Kurs, A., Joannopoulos, J., Moffatt, R., Fisher, P., Soljacic, M.: Wireless power transfer via strongly coupled magnetic resonances. Science 317(5834), 83–86 (2007) 17. Biswas, S., Bianchini, M., Shaw, R.N., Ghosh, A.: Prediction of traffic movement for autonomous vehicles. In: Bianchini, M., Simic, M., Ghosh, A., Shaw, R.N. (eds.) Machine Learning for Robotics Applications. SCI, vol. 960, pp. 153–168. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0598-7_12 18. Geng, Y., Sun, H., Yang, Z., Li, B., Lin, F.: A high efficiency charging strategy for a super capacitor using a wireless power transfer system based on inductor/capacitor/capacitor (LCC) compensation topology. Energies 10(135), 1–17 (2017) 19. Hao, H., Covic, G., Boys, J.T.: An approximate dynamic model of LCL-T based Inductive power transfer power supplies. IEEE Trans. Power Electron. 29(10), 5554–5567 (2014) 20. Mandal, S, et al.: Motion prediction for autonomous vehicles from Lyft dataset using deep learning. In: 2020 IEEE 5th International Conference on Computing Communication and Automation (ICCCA), pp. 768–773, 30–31 October 2020. https://doi.org/10.1109/ICCCA4 9541.2020.9250790 21. Sungwoo, L., Jin, H., Changbyung, P., Nam-Sup, C., Gyu-Hyeoung, C., Chun-Taek, R.: Online electric vehicle using inductive power transfer system. In: Proc. IEEE ECCE, pp. 1598– 1601 (2010) 22. Jin, H., Wooyoung, L., Gyu-Hyeong, C., Byunghun, L., Chun-Taek, L.: Characterization of novel inductive power transfer systems for on-line electric vehicles. In: Proc. 26th Annu. IEEE APEC Expo., pp. 1975–1979 (2011) 23. Nguyen, T.-D., Li, S., Li, W., Mi, C.: Feasibility study on bipolar pads for efficient wireless power chargers. In: Proc. APEC Expo., Fort Worth, TX, USA (2014) 24. Ahmad, A., Alam, M.S., Chabaan, R.: A comprehensive review of wireless charging technologies for electric vehicles. IEEE Trans. Transp. Electrific. 4(1), 38–63 (2018) 25. Soni, A., Dharmacharya, D., Pal, A., Srivastava, V.K., Shaw, R.N., Ghosh, A.: Design of a machine learning-based self-driving car. In: Bianchini, M., Simic, M., Ghosh, A., Shaw, R.N. (eds.) Machine Learning for Robotics Applications. SCI, vol. 960, pp. 139–151. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0598-7_11 26. Ching, T., Wong, Y.: Review of wireless charging technologies for electric vehicles. In: 5th International Conference on Power Electronics Systems and Applications (PESA), HongKong, pp. 1–4 (2013) 27. Sasaki, H.: Wireless power transfer of magnetic resonance. In: Agilent EEsoft 3D EM Application Series, Agilant Technologies (2010) 28. Palimkar, P., Bajaj, V., Mal, A.K., Shaw, R.N., Ghosh, A.: Unique action identifier by using magnetometer, accelerometer and gyroscope: KNN approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 607–631. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_48 29. Triviño, A., González-González, J.M., Aguado, J.A.: Wireless power transfer technologies applied to electric vehicles: a review. Energies 14, 1547 (2021). https://doi.org/10.3390/en1 4061547 30. Bludszuweit, H.: Project victoria—the first Spanish showcase for DWPT. In Proceedings of the FABRIC conference (Presentation), Brussels, Belgium, 2 February 2016 31. Mi, C.C., Buja, G., Choi, S.Y., Rim, C.T.: Modern advances in wireless power transfer systems for roadway powered electric vehicles. IEEE Trans. Ind. Electron. 63, 6533–6545 (2016) 32. Mandal, S., Biswas, S., Balas, V.E., Shaw, R.N., Ghosh, A.: Lyft 3D object detection for autonomous vehicles. In: Artificial Intelligence for Future Generation Robotics, pp. 119–136. Elsevier (2021). https://doi.org/10.1016/B978-0-323-85498-6.00003-4 33. Zakerian, A., Vaez-Zadeh, S., Babaki, A.: A dynamic WPT system with high efficiency and high power factor for electric vehicles. IEEE Trans. Power Electron. 35(7), 6732–6740 (2020). https://doi.org/10.1109/TPEL.2019.2957294
104
T. A. Shikdar et al.
34. He, H., et al.: Maximum efficiency tracking for dynamic WPT system based on optimal input voltage matching. Access IEEE 8, 215224–215234 (2020) 35. Abbas Moosavi, S., SaeidallahMortazavi, S., Namadmalan, A., Iqbal, A., Al-Hitmi, M.: Design and sensitivity analysis of dynamic wireless chargers for efficient energy transfer. In: Access IEEE, vol. 9, pp. 16286–16295 (2021) 36. Mi, C.C., Buja, G., Choi, S.Y., Rim, C.T.: Modern advances in wireless power transfer systems for roadway powered electric vehicles. IEEE Trans. Ind. Electron. 63(10), 6533–6545 (2016). https://doi.org/10.1109/TIE.2016.2574993 37. Hyundai Motor Company, Kia Motors Corp, Hyundai America Technical Center Inc: Electromagnetic field controlling system and method for vehicle wireless charging system. https:// patents.google.com/patent/US20170120757A1/en (2015) 38. Hyundai Motor Company, Kia Motors Corp, Hyundai America Technical Center Inc: Ev wireless charging adjustable flux angle charger. https://patents.google.com/patent/ EP3229339B1/en?oq=EP3229339B1 (2016) 39. New Mobility: Hyundai-Kia America Technical Center and Mojo Mobility partner on wireless EV charging system. https://www.newmobility.global/technology/hyundai-kiaamericatechnical-center-mojo-mobility-partner-wireless-ev-charging-system (2018) 40. Electronic Tutorials. LC Oscillator Basics. http://www.electronics-tutorials.ws/oscillator/osc illators.html 41. Eberhard Waffenschmidt. Qi Quality Factor. Wireless Power Consortium. https://www.wir elesspowerconsortium.com/technology/qualityfactor.html 42. Electronic Tutorials. Power Diodes and Rectifiers. http://www.electronics-tutorials.ws/diode/ diode_5.html 43. Electronic Tutorials. Mutual Inductance. http://www.electronics-tutorials.ws/inductor/mut ual-inductance.html 44. “how to set firing angle”. Youtube uploaded by rkarthikkumar ACAMEDY, https://youtu.be/ xz4F6Z_g3Po, 12 Jan 2018 45. “Expt 2 # Buck Converter using Matlab/Simulink| DC-DC Converter | Design | PE lab| Simulink model”. Youtube, uploaded by Shaija PJ, https://youtu.be/iY8Pisim1uM, 4 July 2020
Automatic Cut Detection: A Probability Based Approach Tejaswini Kar1(B) , P. Kanungo2 , and Vinod Jha1 1 2
KIIT Deemed to be University, Bhubaneswar, Odisha, India {tkarfet,vjhafet}@kiit.ac.in C. V. Raman Global University, Bhubaneswar, Odisha, India
Abstract. Several techniques have been developed in recent times for shot boundary detection with reliable performances in the detection. Different features are used for the evaluation of the feature difference between two consecutive frames. In this paper, an effort has been made to generalize the appropriate distribution of any feature similarity used in the literature for the abrupt shot transition detection. The strength of our innovative way of classification is independent of the chosen feature and the chosen metric for modeling of the feature distance. While the performance of the shot detection rely on chosen feature. Exhaustive experimental results are presented in support of our novel classification scheme. Keywords: Feature similarity · Cut · Probability distribution · WLD · CSLBP · Poisson distribution · Rayleigh distribution · Rician distribution · Normal distribution
1
Introduction
The video shot boundary detection is essential in a sense that identification of the high-level semantics in video analysis paves the way towards complete understanding of a scene. Details of the different shot boundary detection schemes can be found in [1–3]. A wide range of features like intensity [4], edge [5], texture [6], Histogram [7], SURF [8,9], LDP [10] as well as multiple features [11,12] have been deployed for abrupt transition detection in the literature. The feature similarity between two consecutive frames are used to detect the abrupt transition positions in a video based on certain threshold. Most of the literature, evaluated the automatic threshold [7,13] as T = μ + λσ, where μ and σ are the mean and the standard deviation of the feature difference/similarity and parameter λ is a constant. Whereas, non of the articles communicate about distribution of any feature difference/similarity to address the automatic detection of abrupt transition using simple technique with acceptable performance. Filling this research gap is the primary objective of this letter. Therefore, in this article an experimental study is carried out to justify the appropriate distribution of the feature similarity used for abrupt transition detection. Towards this goal the contributions of this work are of two fold: c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 105–114, 2022. https://doi.org/10.1007/978-981-19-1677-9_9
106
T. Kar et al.
– First it has been demonstrated that the distribution of a feature similarity value of a large number of frames follows approximately a Poisson probability distribution irrespective of the characteristic of the feature and the technique used to model the feature similarity. – Secondly, considering mean and variance of the Normal probability distribution approximation of the feature similarity distribution an automatic threshold can be generated for abrupt transition detection in a video.
2
Similarity Feature for Abrupt Transition Detection
In the proposed approach different features considered are histogram, intensity, correlation, joint histogram, CSLBP [14] histogram and Weber feature [15] based histogram. For different features mentioned above the standard similarity measures S(k) are absolute sum histogram difference, absolute sum intensity difference (ASID), Joint distribution feature based directed divergence [16], correlation, CSLBP histogram feature difference, absolute sum Weber feature difference respectively. Using histogram based feature, the similarity measure between k th and (k + 1)th frame is the absolute sum of histogram difference (ASHD) [7] which is defined as S(k) =
255
|hk (m) − hk+1 (m)|,
(1)
m=0
where, hk is the histograms of k th frame. Joint distribution based directed divergence (DD) [16] and the corresponding similarity measure is given by S(k) = 1 − (Dk,k+1 + Dk+1,k ),
(2)
where, Dk,k+1 and Dk+1,k are the directed divergence between joint distribution of k th , (k + 1)th frame and (k + 1)th , k th frame respectively. The correlation based similarity measure between successive frames [17] selected for identifying abrupt transition is evaluated as S(k) =
M −1 N −1 m=0
n=0
(fk (m, n) − μk )(fk+1 (m, n) − μk+1 ) , σk σk+1
(3)
where fk and fk+1 are the k th and (k + 1)th frame of a video, μk , σk are the mean and standard deviation of the pixel intensities in k th frame, M × N is the size of the frame. The Weber feature [15] based similarity measure is absolute sum Weber feature difference (ASWD) [18] value between k th & (k + 1)th frame is evaluated as follows
Automatic Cut Detection: A Probability Based Approach
S(k) =
23
|Wk (l) − Wk+1 (l)|,
107
(4)
l=0
where, Wk is the Weber feature vector of k th frame. Similarly the CSLBP [14] histogram based similarity measure is CSLBP feature difference (CSLBPFD) between k th and (k + 1)th frame is defined as S(k) =
15 k (k+1) hcf (i) − hcf (i) , i=0
where, hkcf is the CSLBP feature histogram of k th frame. Table 1. Ground truth information for TRECVid 2007 data set Video name
Number of frames Number of transitions
BG37808
22000
76
BG37968
56200
193
BG38438
77400
244
BG38174
54500
199
BG38655
82400
377
BG38117
52200
227
BG38903
52200
132
BG37795
22050
58
BG37613
22830
123
BG38422
49900
96
BG37998
51200
194
BG37721
42170
112
BG37399
35950
198
BG38183
52800
130
BG38423
53440
134
Total frames 727240
2493
(5)
108
T. Kar et al.
Fig. 1. Comparison of ASHD distribution with different distributions: (a) Poisson, (b) Rayleigh, (c) Rician and (d) Normal Table 2. Average mean square error with different distributions ASHD × 10−5 ASID × 10−5 COR × 10−5 CSLBP × 10−5 W LD × 10−5
3
P ois
6.04
60
120
15.55
Rayl
21.84
80
130
25.65
30
Rici
13.85
80
130
25.65
30
N orm 13.63
70
130
34.01
5140
20
Modeling the Distribution of Feature Similarity Measure
Here the abrupt transition detection problem is posed through feature independent probabilistic modeling of the similarity measure. The distribution of the similarity measure is realized by fitting to standard distributions used in probability theory. In this article, the similarity measure is approximated to Poison distribution, Rayleigh distribution, Rician distribution and Normal distribution. Considering the similarity measure between two consecutive frames as a random variable S, different distributions are represented as follows. The normal distribution is most often used to describe the random variation that occurs in the data from many scientific disciplines. The popular bell-shaped curve of normal distribution f (s) is characterized by two parameters the mean (μ) and the standard deviation (σ) as follows −(x−μ)2 1 e 2σ2 fS (μ, σ) = √ 2πσ 2
(6)
Automatic Cut Detection: A Probability Based Approach
109
Fig. 2. Comparison of frame correlation distribution with different distributions: (a) Poisson, (b) Rayleigh, (c) Rician and (d) Normal Table 3. Comparison of average F1 for optimum λ value for different feature distance and for different distribution F eature λopt F 1P ois
λopt λopt F 1Rayl F 1Rice
λopt F 1N or
λ=3 88.11
λ=4 88.71
(λ = 4) 88.53
Ref. [17] λ = 10 87.73
λ=7 91.53
λ=7 91.53
(λ = 4) (91.57)
Ref. [16] λ = 10 82.29
λ=4 79.90
λ=4 79.90
(λ = 4) 81.41
λ=8 86.64
λ=6 87.23
λ=6 87.23
(λ = 4) 86.76
Ref. [18] λ = 9 90.40
λ=7 90.32
λ=7 90.32
(λ = 4) 90.19
Ref. [7]
Ref. [6]
λ=5 88.38
The similarity measure S is the absolute sum of the feature difference value, therefore S is a positive real value. Therefore, it is appropriate to model S as the Rayleigh or Rician distribution. The Rayleigh distribution is characterized by a single parameter σ as follows fS (σ) =
x −x22 e 2σ σ2
(7)
110
T. Kar et al.
Fig. 3. Comparison of DD similarity distribution with different distributions: (a) Poisson, (b) Rayleigh, (c) Rician and (d) Normal
However, for some measurements when mean is low, variance is large, the distribution is typically a skewed distribution. A Poisson distribution with a low mean is highly skewed, but with a high mean can be approximated to Normal distribution. The Poisson distribution f (s) is characterized by a parameter λ. In our problem the similarity measure is discrete in nature and bounded between 0 to 127. Therefore, the pdf is modeled as P (x = k) = e−λ
λk ; k = 0, 1, 2, ..., 127. k!
(8)
Similarly the Rician distribution f (s) is characterized by parameters σ and a and defined as s s2 + a2 as fS (σ, a) = 2 e− ( )I0 ( 2 ), (9) 2 σ 2σ σ where I0 is the Bessel function of first kind and zeroth order.
4
Abrupt Transition Detection
The distribution of similarity measure S is modeled by four distributions, considered in the previous section. Considering the estimated parameters i.e. mean (μ) and standard deviation (σ) are considered to evaluate the threshold parameter T as follows. T = μ + λσ, (10)
Automatic Cut Detection: A Probability Based Approach
111
where λ is a positive constant. If the similarity measure between k th and (k+1)th frame i.e. S(k) is greater than the threshold T , then an abrupt transition is said to exist at the k th frame. The optimum choice of parameter is considered for performance evaluation.
Fig. 4. Comparison of CSLBPFD distribution with different distributions: (a) Poisson, (b) Rayleigh, (c) Rician and (d) Normal
5
Simulations and Analysis
The performance of different distributions on abrupt transition detection based on the threshold in (10) is evaluated for 15 test videos from TRECVid 2007 database. This database video is most appropriate for abrupt transition detection as it contains very high number of abrupt transitions. The videos with number of frames and respective ground-truth abrupt transitions are presented in Table 1. In total 7, 27, 240 number of frames having 2493 number of abrupt transitions are considered to evaluate the performance of abrupt transition detection with all four pdfs with constant λ value. F1 performance measure is used to evaluate the quantitative performance which is defined as 2 × Rec × P r (11) Rec + P r Where Rec refers to fraction of transitions detected correctly out of actual number of transitions and P r is the fraction of transitions detected correctly out of total number of transitions detected by the algorithm. The parameter λ is F1 =
112
T. Kar et al.
Fig. 5. Comparison of ASWD distribution with different distributions: (a) Poisson, (b) Rayleigh, (c) Rician and (d) Normal
varied to evaluate the threshold and the optimum λ is considered for threshold evaluation at which the F1 has the highest value. Five similarity measures such as ASHD, Correlation, DD, CSLBPFD and ASWD are considered for the validation of appropriate pdf model placed in Fig. 1, 2, 3, 4 and 5 respectively. Based on the visual perception it is observed that all these features are closely fitting with Poisson distribution placed in Fig. 1(a), 2(a), 3(a), 4(a) and 5(a) respectively. Further, the mean square error (MSE) between distribution of different feature similarity measures and the different fitted pdfs are tabulated in Table 2. It is observed that the Poisson distribution has the lowest MSE in comparison to Rayleigh, Rician and Normal distribution for all the similarity measures. Therefore, Poisson distribution may be the most appropriate model for approximation of the similarity measure for abrupt transition detection. In the simulation of abrupt transition detection few baseline features and few recently developed features are considered. The parameter λ is varied from 1 to 12 to evaluate the threshold in (10) and the optimum λ is considered at which the F1 measure attains the highest value. Table 3 presents the comparison of average F1 measure along with the optimum λ for different features in different pdf models. In the Table 3, F 1P ois , F 1Rayl , F 1Rice and F 1N or represents the F1 value for Poisson, Rayleigh, Rician and Normal distribution respectively. It is observed that optimum F 1 value based on the threshold model in (10) using the optimum λ is almost same for all the distributions, for all the five similarity measures. The optimum value of λ for Poisson, Rayleigh and Rician varies from feature to feature. However, the λ is fixed to four for Normal distribution for all the similarity
Automatic Cut Detection: A Probability Based Approach
113
measures. Though, the distribution of different similarity measures are closely fitted to Poisson distribution in terms of lowest MSE, Normal distribution is the best choice based on optimum F1 value and fixed λ for abrupt transition detection irrespective of feature type. However, by varying λ value F1 measures can be improved by considering any of the given distributions.
6
Conclusions
In this article an extensive simulation has been carried out considering large number of videos to model the appropriate distribution of the feature similarity values between consecutive frames for abrupt transition detection. It has been validated that the distribution of any similarity value is closely matched with Poisson distribution that retains the characteristics of the feature space in terms of lowest MSE. However, for threshold evaluation for abrupt transition detection based on μ and σ, the parameter λ varies widely from feature to feature by considering Poisson distribution. It has been validated that the parameter λ is constant irrespective of feature similarity for Normal distribution in terms of optimum F 1 measure. Therefore, for the best performance the feature similarity should be modeled as Normal distribution to evaluate the μ and σ. Further the optimum choice of threshold based on the optimum F 1 performance measure is T = μ + 4σ. As a future scope of the work the proposed distribution based modeling for threshold selection can be made adaptive by automatically selecting the parameter λ to improve the F1 score of the abrupt transition detection process.
References 1. Abdulhussain, S.H., Ramli, M.I., Saripan, A.R., Mahmmod, B.M., Al-Haddad, S.A.R., Jassim, W.A.: Methods and challenges in shot boundary detection: a review. Entropy 20(4), 214 (2018) 2. Smeaton, A.F., Over, P., Doherty, A.R.: Video shot boundary detection: seven years of TRECVid activity. Comput. Vis. Image Underst. 114(4), 411–418 (2010) 3. Fabro, D.M., B¨ osz¨ ormenyi, L.: State-of-the-art and future challenges in video scene detection: a survey. Multimedia Syst. 19(5), 427–454 (2013) 4. Lu, Z., Shi, Y.: Fast video shot boundary detection based on SVD and pattern matching. IEEE Trans. Image Process. 22(12), 5136–5145 (2013) 5. Yoo, H.W., Ryoo, H.J., Jang, D.: Gradual shot boundary detection using localised edge blocks. Multimed. Tools Appl. 28, 283–300 (2006) 6. Kar, T., Kanungo, P.: Cut detection using block based centre symmetric local binary pattern. In: 2015 International Conference on Man and Machine Interfacing (MAMI), pp. 1–5, 17–19 December 2015 7. Zhang, H.J., Kankanhalli, A., Smoliar, S.W.: Automatic partitioning of full motion video. Multimed. Syst. 1(1), 10–28 (1993) 8. Birinci, M., Kiranyaz, S.: A perceptual scheme for fully automatic video shot boundary detection. Signal Process. Image Commun. 29(3), 410–423 (2014)
114
T. Kar et al.
9. Hato, E., Abdulmunem, M.E.: Fast algorithm for video shot boundary detection using SURF features. In: 2019 2nd Scientific Conference of Computer Sciences (SCCS), pp. 81–86 (2019) 10. Kar, T., Kanungo, P.: Abrupt scene change detection using block based local directional pattern. In: Sharma, N., Chakrabarti, A., Balas, V.E. (eds.) Data Management, Analytics and Innovation. AISC, vol. 1016, pp. 191–203. Springer, Singapore (2020). https://doi.org/10.1007/978-981-13-9364-8 15 11. Lakshmi Priya, G.G., Domnic, S.: Walsh-hadamard transform kernel-based feature vector for shot boundary detection. IEEE Trans. Image Process. 23(12), 5187–5197 (2014) 12. Kavitha, J., Rani, J.A., Fathimal, P.M., Paul, A.: An efficient shot boundary detection using data-cube searching technique. Recent Adv. Comput. Sci. Commun. Formerly Recent Patents Comput. Sci. 13(4), 798–807 (2020) 13. Singh, A., Thounaojam, D.M., Chakraborty, S.: A novel automatic shot boundary detection algorithm: robust to illumination and motion effect. Signal Image Video Process. 14(4), 645–653 (2019) 14. Heikkil¨ a, M., Pietik¨ ainen, M., Schmid, C.: Description of interest regions with center-symmetric local binary patterns. In: Kalra, P.K., Peleg, S. (eds.) ICVGIP 2006. LNCS, vol. 4338, pp. 58–69. Springer, Heidelberg (2006). https://doi.org/10. 1007/11949619 6 15. Chen, J., et al.: WLD: a robust local image descriptor. IEEE Trans. Pattern Anal. Mach. Intell. 32(9), 1705–1720 (2010) 16. Kim, S.H., Park, R.H.: An efficient algorithm for video sequence matching using the modified Hausdorff distance and the directed divergence. IEEE Trans. Circ. Syst. Video Technol. 12(7), 592–596 (2002) 17. Warhade, K.K., Merchant, S.N., Desai, U.B.: Performance evaluation of shot boundary detection metrics in the presence of object and camera motion. IETE J. Res. 57(5), 461–466 (2011) 18. Kar, T., Kanungo, P.: Motion and illumination defiant cut detection based on weber features. IET Image Process. 12(10), 1903–1912 (2018). https://doi.org/10. 1049/iet-ipr.2017.1237
A Novel Design of T Shaped Patch Antenna Integrated with AMC for WLAN Applications Priyanka Das(B) University of Engineering and Management, Kolkata, India [email protected]
Abstract. This article demonstrates 3 dBi gain improvement of a T-shaped microstrip patch antenna working at 5.2 GHz for WLAN applications by using a single layer artificial magnetic conductor (AMC) surface below the antenna at an optimum spacing. The AMC provides zero reflection phase at the design frequency of the antenna which causes constructive interference between waves radiated from the antenna and those reflected from the AMC surface. A comparative study is made with an equivalent PEC surface in place of AMC. Higher gain is obtained by AMC due to its in-phase reflection of electromagnetic (EM) waves. The AMC unit cell has been designed such that it is polarization insensitive at normal incidence. All simulations and numerical studies of the antenna and AMC have been carried out using Ansys HFSS 2020. Keywords: AMC · Polarization insensitive · Microstrip antenna · WLAN
1 Introduction PEC reflectors when placed below antennas are located at a spacing of 0.5λ which makes the dimensions of integrated antennas large where λ is the wavelength corresponding to the antenna operating frequency. In the modern era, low profile antennas are required for communication systems to combat space constraints. Artificial magnetic conductors (AMCs), when used as reflectors are located closer to the antenna, thus aiding miniaturization. In the recent years, AMC has been widely used for gain enhancement of antennas [1]. Vias have been used with AMC substrate for making the antennas compact [2]. AMC surfaces are used for manipulation of radiation pattern of dipole antennas [3]. An AMC surface has been employed for obtaining unidirectional radiation [4] in a bow-tie antenna. An AMC surface [5] has been used for isolating a dipole antenna from metallic objects or human bodies so that its radiation characteristics are not degraded. An AMC has been integrated with a broadband circularly polarized antenna [6] to make it low-profile. In this article, an AMC has been designed to give zero reflection phase at 5.2 GHz. This is further integrated with a microstrip patch antenna resonating at the same frequency for gain and directivity enhancement. The proposed antenna is low profile and poses to be a suitable candidate for WLAN applications. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 115–120, 2022. https://doi.org/10.1007/978-981-19-1677-9_10
116
P. Das
2 Antenna Design
Fig. 1. Structure of antenna (la = 30, lb = 30, lc = 24.2, ld = 8.6, le = 8, lf = 8, lg = 13, lh = 3. All dimensions are in mm.)
A microstrip patch antenna (MSA) is designed on a FR4 substrate of thickness 1.6 mm with copper ground. A T-shaped patch is constructed for maintaining a compact structure as shown in Fig. 1. Resonance occurs at 5.2 GHz as evident from the reflection coefficient characteristics of the antenna in Fig. 2. The antenna exhibits a peak gain of 3 dBi at 5.2 GHz in the E plane (F = 0°) There is a slight tilt in the direction of peak gain towards left for E plane far field pattern. In the H plane (F = 90°), the maximum
Fig. 2. S11 characteristics of the antenna
A Novel Design of T Shaped Patch Antenna
117
Fig. 3. Radiation pattern of the antenna (a) E plane (b) H plane
radiation is in the broadside direction with a peak gain of 2.7 dBi. This is evident from Fig. 3. The proposed MSA has a fractional bandwidth of 2.88%.
3 Design of AMC The AMC (artificial magnetic conductor) is a periodic surface which is constituted of multiple unit cells made of FR4 substrate having thickness 1.6 mm, arranged periodically both along the transverse directions. The unit cell (0.35λ × 0.35λ, where λ is the wavelength corresponding to the resonant frequency) comprises a patterned slotted layer as shown in Fig. 4 which resonates at 5.2 GHz. The operating frequency of the AMC is identical with that of the antenna. The bottom layer of the unit cell is made of copper ground. Four square shaped slots in conjunction with a cross shaped slot produce minimum reflection at 5.2 GHz. Zero degree reflection phase is obtained at the resonant frequency as observed in Fig. 5. The −90° to + 90° reflection phase bandwidth is 0.38 GHz which covers a fractional bandwidth of 7.3%. Angular stability is obtained up to 40°, beyond which the S11 characteristics degrade. The reflection characteristics gradually shift towards right with the increase in the angle
Fig. 4. Geometry of AMC unit cell (l i = l j = 20 mm, l k = 7 mm, ll =7 mm, lm = 16.6 mm)
118
P. Das
Fig. 5. Reflection coefficient characteristics of AMC unit cell.
Fig. 6. Angular stability in (a) S11 magnitude (b) S11 phase
Fig. 7. Polarization insensitivity in (a) S11 magnitude (b) S11 phase
of incidence, thus bringing instability in the frequency response, as observed in Fig. 6. Due to its four-fold rotational symmetry, high polarization insensitivity up to 90° is obtained as shown in Fig. 7. The reflection characteristics do not change with variation of the polarization angle F. Hence the AMC unit cell depicts analogous reflection characteristics for both TE (F = 0°) and TM modes (F = 90°).
A Novel Design of T Shaped Patch Antenna
119
4 Design of the Proposed Antenna
Fig. 8. Structure of the MSA integrated with AMC. (ln = lo = 160 mm)
The antenna is placed symmetrically above the metasurface consisting of 64 unit cells as shown in Fig. 8. The vertical spacing between the antenna and the AMC surface is fixed at 6 mm for obtaining the maximum gain of 6 dBi. Due to in phase reflection from the AMC, the radiated waves reflected from the antenna and AMC constructively interfere which results in gain enhancement. The front to back ratio improves and a unidirectional beam is obtained. If the AMC is replaced by a perfect electric conductor (PEC), there is a 180° phase reversal when the EM waves from the antenna are reflected from it. A comparative study is conducted between PEC and AMC integrated antenna in terms of reflection characteristics and far field radiation pattern in Fig. 9 and Fig. 10 respectively. The resonance frequency shifts from 5.4 GHz to 5.3 GHz, when the AMC is replaced by a PEC of the same dimensions at the same location. Higher gain is obtained with AMC as compared to PEC. The cross-polarization level is higher with PEC as
Fig. 9. S11 characteristics of AMC integrated antenna
120
P. Das
compared with AMC. Peak gains of 4.9 dBi and 6 dBi have been obtained for PEC and AMC integrated antenna respectively.
Fig. 10. Radiation pattern of (a) AMC integrated antenna (b) PEC integrated antenna
5 Conclusion A 3 dB gain enhancement for MSA operating at 5.2 GHz is acquired by using an AMC of 64 unit cells. A novel slotted unit cell with copper backing has been designed using HFSS 2020, such that 0° reflection phase is obtained at the operating frequency of the MSA. Identical reflection characteristics for TE and TM mode EM waves are obtained due to polarization insensitive nature of the AMC unit cell. Finally a comparative study is conducted between performance of the MSA integrated with AMC and PEC.
References 1. Prakash, P., Abegaonkar, M.P., Basu, A., Koul, S.K.: Gain enhancement of a CPW-fed monopole antenna using polarization-insensitive AMC structure. IEEE Antennas Wirel. Propag. Lett. 12, 1315–1318 (2013) 2. Kim, D., Yeo, J.: Low-Profile RFID tag antenna using compact AMC substrate for metallic objects. IEEE Antennas Wirel. Propag. Lett. 7, 718–720 (2008) 3. Ford, K.L., Rigelsford, J.M.: Street furniture antenna radiation pattern control using AMC surfaces. IEEE Trans. Antennas Propag. 56(9), 3049–3052 (2008) 4. Zhai, H., Zhang, K., Yang, S., Feng, D.: A low-profile dual-band dual-polarized antenna with an AMC surface for WLAN applications. IEEE Antennas Wirel. Propag. Lett. 16, 2692–2695 (2017) 5. Hadarig, R.C., de Cos, M.E., Las-Heras, F.: UHF dipole-AMC combination for RFID applications. IEEE Antennas Wirel. Propag. Lett. 12, 1041–1044 (2013) 6. Feng, D., Zhai, H., Xi, L., Yang, S., Zhang, K., Yang, D.: A broadband low-profile circularpolarized antenna on an AMC reflector. IEEE Antennas Wirel. Propag. Lett. 16, 2840–2843 (2017)
Neural Network: A Way to Know Consumer Satisfaction During Voice Call Divya Mishra1(B)
and Suyash Mishra2
1 G.L. Bajaj Institute of Technology and Management, Greater Noida, (U.P), India
[email protected]
2 Royal Bank of Scotland, Delhi, India
[email protected]
Abstract. Consumer satisfaction is a major factor to grow a business. And if this thing is said about the telecom industry then it becomes even more important. According to Google research, 88% of people still prefer to receive phone calls because it is easier to speak than to write something. In our country, mobile consumer dissatisfaction is increasing as consumer connections are growing fast, due to lack of voice quality during ongoing calls while we are moving towards 5G and 6G. Many academia, researchers and industries are working on this but till now this issue has not been solved. In this research paper, an approach is given to judge the level of customer satisfaction during voice calls. On the basis of few LTE parameters obtained from the nearest cell phone tower, the proposed model will be able to predict whether or not a voice call was satisfactory. The objective of this paper is to measure consumer satisfaction by applying an artificial neural network which is widely referred to for automatic feature extraction, classification and prediction. Keywords: Self optimization · Artificial neural network · Reference signal received power · Reference signal received quality · Call drop · Deep learning
1 Introduction Nowadays, life without a cell phone cannot be imagined. An unsatisfactory phone call can cause you to lose your focus and interrupt your important conversation anytime. Almost everyone, whether they live in the city or the village, is struggling with this issue. Most users in cities have to move from one spot to another for better phone signal quality when trying to call someone. Unsatisfactory calls can also be caused by call drops or poor voice calls. Call drop is an unexpected phenomenon whenever a call is sudden disconnected without any prior information. Whereas Trai states not more than 2% call drop is allowed. A dropped call is what we commonly call when a cellular phone call unexpectedly ends. Many cellular subscribers are struggling with this problem these days. In rural areas, voice calls are dropped due to a loss of coverage, while in urban areas, this problem is caused by a growing gap between an increase in subscribers and a relative lack of self-optimizing infrastructure to address all the issues automatically. In the © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 121–131, 2022. https://doi.org/10.1007/978-981-19-1677-9_11
122
D. Mishra and S. Mishra
world of cellular networks, self-optimizing networks support automated configuration, optimization, diagnosis, and recovery. Introducing self-optimization is considered to be a first-rate necessity for future cellular networks and operations. This is due to the potential savings in capital expenditures and operational expenditures. With the proliferation of self-optimizing functions arising as a consequence of a technological boom, one of the predominant issues for operators will be to decide which functions to introduce and when to activate them to ensure a cost-efficient and well-behaved network. A study in this paper examined how the use of an artificial neural network can predict whether or not a consumer would be satisfied with a phone call, which is a robust algorithm of deep learning and based on hidden layers. The research uses artificial neural networks for predicting consumer satisfaction once the call has been placed. Artificial neural networks are generally used for solving more complex problems. Here, threelayer feed forward neural network has been used to predict customer satisfaction level during the above-said voice call situation. The research paper focused on some LTE parameters which are responsible for call quality degradation during on-going voice call such as the speed of subscriber in km/hr, rsrp in dBm, rsrq in db, rssnr, lte bar level and last is customer review that means call was satisfactory or not. A total of 315 samples were collected by using a google survey form. Number zero stands for satisfactory, and number one stands for unsatisfactory. Used Lte parameter has been checked and collected using a network signal info application. We collected the data from primary sources, which are live and quantitative data because subscribers have provided the information by using a Google form for their phone. We collected data on Bharti Airtel subscribers and Reliance Jio subscribers who are using 4G service with these networks.
2 Related Work Many researchers, industries, telecom operators have done a great job which has own pros and cons but consumer dissatisfaction for poor voice calls or unexpected termination of call is increasing day by day as shown in Fig. 1. Many users face frequent poor voice calls in comparison to call drop. It depends on different causes which can responsible for consumer dissatisfaction in a different form. For example, trai and few researchers conducted the driving test in different locations to analyze the network performance and they observed most of the telecom operators failed to meet the benchmark of network-related parameters. A telecom regulatory authority [1] has conducted a lot of consumer surveys and drive tests to figure out what causes consumer frustrations for instance poor voice call quality, signal strength, or call drop. The test doesn’t depend on any algorithm. It simply collects data about how people react to live interaction. If we talk about the above-mentioned reason, then it can be due to many reasons such as a changing climate, rain, and electromagnetic fields, non-functioning mobile phones, greater distance between the mobile and base stations, traffic jams, handoffs, weak signal strengths should be indirect reasons for poor customer satisfaction. Yi-Bing Lin and Indika A.M. Balapuwaduge [2] proposed the queuing priority channel assignment in their research paper which could be helpful to improve call quality, in this algorithm priority has been given to ongoing calls rather than new introducing calls. Queues are also introduced to buffer the secondary call to entertain later. The drawback of this scheme is the high
Neural Network: A Way to Know Consumer Satisfaction During Voice Call
123
consumption of bandwidth. The prioritization scheme could be getting a better result by combining another scheme for example measurement-based prioritization scheme, guard channel prioritization scheme, queue handoff calls, call admission control protocol, and use auxiliary station [1]. Although many telecom operators used TDMA based dynamic channel allocation in a situation of overloaded cells [3]. A Research paper [4] proposed the bandwidth degradation scheme where bandwidth allocation to various subscribers is adjusted dynamically as required to optimize the utilization of bandwidth for other calls. During high utilization time especially in highly dense areas when a high number of users are trying to make a call using an underline mobile network in order to accommodate large users at a given time telecom operator reduces cell size in a mobile network. Due to this, a number of handovers are taking place and the probability of call dropping has increased. Also, if sufficient bandwidth is not available for new calls, then blocking probability for newly generated calls also increases. Dropping of calls during handover is less desirable than blocking new calls. Operators are also blaming spectrum shortage and state that the spread of radiation fears is making things harder for them. Despite implicit limitations like spectrum shortage, operators must own the responsibility and ideally devise innovative practices to solve the voice call quality issue. Sagadevan [5] has conducted the driving test by using MyMobileCoverage which enables to the view of current network coverage with the signal strength mapping and drop call information. But the research and drive test areas are limited and they need more men’s power and time. On the basis of consumer dissatisfaction factors, the researcher had searched a different solution to improve call quality like trai has launched mycall App to know the consumer review after the phone call. Research has viewed many journals and white papers and after reviewing the earlier proposed methods the research concludes the most of the methods are implemented with the hardware and impacted the telephone companies as capital and operational expenditure increases. The proposed model is a fully self-optimized neural network and it is fully secure to deploy for end-user and 100% accurate to measure the call quality on the basis of real-time parameters.
Fig. 1. Poor voice quality vs. call drop
124
D. Mishra and S. Mishra
3 Methodology Research work is exploratory and experimental where we tuned up the hyperparameters and controls the values for validating the accuracy of the model. It is based on a quantitative approach because research needed numbers and applied statistics on gathered data. We analysed the dataset using Microsoft excel and google colab Jupiter notebook to understand and work on the proposed model. A sampling of data is longitudinal, where we gathered the data at several points by primary resource with the help of google form and network signal info app which has been specially downloaded by consumers by android play store to complete proposed research work. 3.1 Input and Output Parameters Research has collected a total of 315 samples with the help of google form and a network info app which has been used to train a model. Parameters are described below: • RSRP (Reference signal received power) RSRP strength is a major lte signal strength parameter, which measures the received signal power [6]. It is always a negative decimal value and is measured in decibels (dbm). Signal strengths can vary from approximately −40 dBm to −140 dBm. The nearer that variety is to zero, the stronger the call sign. In general, something higher than equal to − 80 decibels is taken into consideration as a usable signal. If RSRP value lies between − 80 dbm to −95 dbm it means there is excellent signal strength. If it is between −96 dbm to −107 dbm, it gives good signal strength. If it is between −108 dbm to −118 dbm, it gives fair signal strength and if it is between −119 dbm to −125 dbm, it means there is poor signal strength. RSRP less than −125 is unusable so there is no signal that will be found on a mobile phone. The formula of RSRP is given below-
Where N shows the number of resource blocks. It depends on bandwidth. Where RSSI is Received Signal Strength Indicator, shows the indication of signal intensity or quality.
• RSRQ Reference Signal Received Quality is a second major lte parameter, which measures the signal quality [6]. It is directly proportional to Reference Signal Power. The range of it is −20 db to −3 db, where −3 db is best. The formula of RSRQ is given below:
Neural Network: A Way to Know Consumer Satisfaction During Voice Call
125
Where N is number of resource block depends on bandwidth. • LTE bar level Lte bar level is the visualization of signal strength on mobile phones in the form of bars line or dot. If the signal is less than or equal to −115 dbm only one bar shows on our mobile phone and if the signal is greater than or equal to −95 dbm there are 4 bars if signals lie between −105 to −114 dBm, two lines, or dot shows in our mobile phone and if it is between −95 to −104, 4 lines show in our mobile phone. • Speed At what speed mobile user is traveling during an on-going call. It can be 0 km per hour to 60 km per hour. • Result Call Test results are divided into two groups which are satisfactory and unsatisfactory. All samples are stored in excel format. Data set information is given below:
Research has used few steps to design the neural network model for measuring consumer satisfaction during voice calls such as feature engineering, designing the model, fitting the dataset which is in form of an excel file, testing a model with giving the unseen dataset to the trained model and now the model is ready to predict the consumer call was satisfactory or not. Now at this point, no manual submission of consumer review is required. It is an automatic process based on a very robust algorithm known as an artificial neural network its working is just like a human brain. Due to a lack of resources, only two network operators are included in the analysis – Jio and Airtel. Samples were taken longitudinally. Numeric values have been used in the research in order to achieve the objectives. The dataset has been pre-processed after it was collected. The pre-processed data had to go through many steps, such as data scaling, data transforming, handling missing data, etc.
126
D. Mishra and S. Mishra
Fig. 2. Co-variance matrix for collected data samples.
The resulting data has been analyzed, and many hidden facts have been revealed. The heat map graphs in Fig. 2 illustrate the covariance of the irrelevant factors. As can be seen in the attached graph, negative values show no relationship between parameters, such as no relationship between speed and rsrp, and it seldom influenced other parameters.
4 Proposed Solution and Technique Self- Optimization is a hidden concept of recent technologies which needs less human intervention and increases the accuracy of the result. In cellular networks, self-optimization means controlling the cellular parameters automatically according to requirements. There are various types of algorithms such as searching the automatic neighbor relation list, mobility load-balancing optimization, handover optimization or mobility robustness optimization are the major features of the self-optimizing network that are so much important to optimize mobile network and will definitely improve voice call quality and consumer satisfaction if these are implemented in the real-time environment [7]. This research is focused on only predicting consumer satisfaction on the basis of real-time parameters with less human intervention. Artificial Neural Network is a supervised feed-forward neural network, can be used for prediction in the cellular networks to handle the handover situation. As shown in Fig. 3, the proposed neural network model is a sequential model having an input layer for taking the input parameters as research has used 4 input parameters so there are 4 neurons in the input layer, two hidden layers for computing the weighted input signals with sigmoid activation function and for calculating the error. Research has adopted mean squared error loss function and in last output, a layer having two neurons as we have to know the call was satisfactory or unsatisfactory. Each neuron is interconnected to other neurons to adjust the information in a forwarding direction. The neural network learned the information in a supervised approach as like a kid is learning flashcards [8]. During the training process model is trying to fit in a given dataset. Research has used 500 epochs with 2 batch sizes with a 0.1 learning rate. As shown in Fig. 4, the proposed model has to adjust fifty parameters and biased is also included in it.
Neural Network: A Way to Know Consumer Satisfaction During Voice Call
127
Fig. 3. Proposed ANN model.
Fig. 4. Proposed model summary
4.1 Model Training As we have seen the structure of the model is ready but now is the time to train the model by samples just like a kid’s learning in a supervised approach. As shown in Fig. 5, the model is learning from samples. And after learning new data sample is given to the model and the model is predicting for that. This is only the visualization of training. But actual training process is done by the backpropagation algorithm. As shown in Fig. 6, the first step is forward propagation, where information is propagated in the forwarding direction. There is the computation of incoming weighted input signals performed in each hidden layer. In the second step, the error is calculated if the expected output is not equal to the predicted output. Error is propagated in the backward direction in the third step. Now in the fourth step weights and bias are updated till the expected output is not received. There are 500 epochs that have been used to train the proposed neural network model.
128
D. Mishra and S. Mishra
Fig. 5. Visualization of model training and prediction
Fig. 6. Model training using back propagation
4.2 Model Accuracy Model evaluation is done by measure the accuracy of the validation dataset and training dataset. The training accuracy of the model is 99.60% and the validation accuracy of the model is 100%. The evaluated training curve and validation curve can be seen in Fig. 7 for reference purposes. Error or loss during training is 0.0023 and during validation of the model, the error in the model is 0.0003. Model accuracy and loss can be seen by below training and validation accuracy and loss curve. The accuracy curve is in increasing order and the loss curve is in decreasing order. The blue color shows the training curve and the red color shows the validation curve. Now the model is ready for a new unseen dataset for testing. After testing the model by fitting the testing samples, the accuracy of a model is 100%. It means the proposed model is so much accurate and its prediction will be 100% true. Now next step is prediction.
Neural Network: A Way to Know Consumer Satisfaction During Voice Call
129
Fig. 7. Model accuracy and loss
5 Results and Discussion Now the model is ready to predict for unseen samples. Research work has tried to weigh the model on every aspect. As shown in Fig. 8 and Fig. 9, unseen samples are given to the model where the model is 100% sure to predict call was satisfactory or not. If −125 is rsrp, −20 is rsrq, 1 is lte bar and 40 is speed. Now the model is 100% sure to predict call was unsatisfactory and the consumer is dissatisfied.
Fig. 8. Test sample for unsatisfactory call
If −110 is rsrp, −11 is rsrq, 3 is lte bar and 40 is speed. Now the model is 100% sure to predict the call was satisfactory and the consumer is satisfied. As shown in Fig. 9 for satisfactory samples.
Fig. 9. Test sample for satisfactory call
In this research paper, we have tabulated the test sample results with the model’s predictions to measure the consumer satisfaction level for all the sample tests (Table 1).
130
D. Mishra and S. Mishra Table 1. Predicted outcomes
RSRP
RESULT
PREDICTED
RSRQ
BARS
SPEED
-94
-7
4
40
Satisfactory
0
-113
-11
2
40
Satisfactory
0
-121
-14
1
0
unsatisfactory
1
-121
-15
1
0
unsatisfactory
1
-120
-16
1
0
unsatisfactory
1
-128
-17
1
0
unsatisfactory
1
-103
-12
3
40
Satisfactory
0
-110
-11
3
40
Satisfactory
0
-114
-15
2
40
Satisfactory
0
-107
-13
2
40
Satisfactory
0
6 Conclusion While the subscriber base in the country is growing very fast, the mobile telecom infrastructure is not growing at the same pace and immense pressure is being put on the existing facilities, leading to a dip in the quality of services (QoS) provided. Poor voice calls or unexpected termination of the ongoing calls, affecting the quality of experience of the subscribers and can increase consumer dissatisfaction. This research proposed the neural network model which is based on a supervised learning approach to predict consumer satisfaction or dissatisfaction during voice calls on the basis of real-time lte parameters. This model can be used by telecom operators and end-user also because research has been done on RSRP, RSRQ, Lte bar level, and speed of user. Of course, these are not sufficient parameters to predict the consumer satisfaction level but we cannot ignore these parameters easily. The model is 100% accurate to predict real-time situations. The proposed model is also compatible to deploy with any android device to predict real-time without needed any internet connection or data pack. Future with self-optimization which is based on deep learning and machine learning is so interesting because this is the time of technological revolution and competition in the research market. This research has some limitations due to lack of resources such that this model alone is not capable to improve voice calls. It can only predict the consumer satisfaction level on the basis of real-time lte parameters during an ongoing voice calls. But it will be helpful for telecom vendors to know the consumer view at every call if it is deployed in a real-time environment.
References 1. Telecom Regulatory Authority of India: Technical Paper on Call Drop in Cellular Networks. http://www.trai.gov.in/web-123/Call-Drop/Call_Drop.pdf (2016)
Neural Network: A Way to Know Consumer Satisfaction During Voice Call
131
2. Lin, Y.-B., Mohan, S., Noerpel, A.: Queuing priority channel assignment strategies for PCS handoff and initial access. IEEE Trans. Veh. Technol. 43 (1994) 3. Li, X.J., Chong, P.H.J.: A dynamic channel assignment scheme for TDMA-based multihop cellular networks. IEEE Trans. Wirel. Commun. 7 (2008) 4. Monica, C.H., Bhavani, K.V.L.: A bandwidth degradation technique to reduce call dropping probability in mobile network systems. Telkomnika Indonesian J. Electr. Eng. 16, 303–307 (2015) 5. Segeran, S.S., Radzi, N.A.B.M.: Determining the drop call rate, failed call rate and signal strength of Telecom mobile network in the Universiti Tenaga National Putrajaya Campus. In: The 3rd National Graduate Conference (NatGrad2015), Universiti Tenaga Nasional, Putrajaya Campus (April 2015) 6. Simpson, O., Sun, Y.: LTE RSRP, RSRQ, RSSNR and local topography profile data for RF propagation planning and network optimization in an urban propagation environment, vol. 21, pp. 1724–1737 (Dec. 2018) 7. Divya, M., Anuranjan, M.: Self-optimization in LTE: An Approach to Reduce Call Drops in Mobile Network: Springer Singapore, FTNCT 2018: Futuristic Trends in Network and Communication Technologies, pp. 382–395 (Dec. 2018) 8. Mitchell, T.M., Carbonell, J.G., Michalski, R.S.: Machine Learning: A Guide to Current Research: The Springer International Series in Engineering and Computer Science. eBook ISBN (978-1-4613-2279-5) (1986)
Automatic Access Application of ICG and ECG Signals Hadjer Benabdallah(B)
and Salim Kerai
Department Biomedical Engineering, University of Tlemcen, Tlemcen, Algeria [email protected]
Abstract. The study of heart diseases has progressed rapidly over the past two decades. For this purpose, we based on a noninvasive and cheaper technique called impedance cardiography (ICG) to diagnose and monitor cardiac disorder, thanks to the calculation of the hemodynamic parameters. In this sense, it is essential to find a simple and effective solution to estimate the cardiac indices and time intervals. We have developed an application of automatic access to help the clinics to analyze ICG and electrocardiogram (ECG) signals, either locally or remotely. This application aims to make available all necessary information to doctors that help them to establish a fast and reliable diagnosis. This diagnosis is based on the estimation of cardiac indices and time intervals for cardiovascular disease. This application is automatic access to the real-time application used to optimize the quality of care and speed of diagnosis, whatever their geographical location. It is performed according to two criteria: information storage and data manipulation. Keywords: ICG · ECG · Hemodynamic parameters · Real-time · Automatic access
1 Introduction Impedance cardiography (ICG) [1] is a novel technique to measure cardiac function in a non-invasive way. The ICG is simple to use, cheaper and safe enabled the estimation of time’s intervals and cardiac indices. To optimize the quality of care and the speediness of diagnosis, whatever their geographical location, we carried out two criteria: • The storage of information; • The manipulation of data through an application of automatic access in real time. In this paper, three different environments are used: Matlab, MySQL, and Java Netbeans, which allowed the development of a new application, presented in Fig. 1 that is easy, fast, and reliable and helps the clinics to analysed ICG and electrocardiogram (ECG) signals and diagnose if there are any abnormalities in heart functions, either locally or remotely. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 132–144, 2022. https://doi.org/10.1007/978-981-19-1677-9_12
Automatic Access Application of ICG and ECG Signals
133
Fig. 1. The principle of our developed application.
The work is carried out under a 64-bit win 7. Three software are installed on the operating system: Java Netbeans, Matlab, WAMP/EASYPHP (MySQL) for web development. NetBeans IDE 7.3 [2] supports Java, PHP, C, C++, and Ruby application development. Netbeans, created at the initiative of sun Microsystems (Kernel of Forte4J/SunOne), has all the characteristics essential for a quality IDE, whether it serves to develop in Jave, Ruby, C, C++, or even PHP. The OpenSource license, NetBeans makes it possible to develop and deploy Swing graphics applications, Applets, JSP/Servlets, J2EE architectures. It is easy to use in an extremely customizable environment. The NetBeans IDE is based on a robust kernel. Especially the NetBeans Platform can also use to develop your own Java applications. It has a powerful plugin system, which allows you to a tailor-made EDI. Besides the full version of the NetBeans IDE, you have different versions (6 in total, not counting the full version) which focus on a specific platform or language: Java ME, Java (SE + ME + EE), Ruby, C/C ++, PHP. NetBeans also allows you to deploy your web applications. Matrix laboratory [3] is interactive software developed by Math Works Inc., particularly for digital data processing. It integrates the numerical calculation, the visualization of the results, and the programming in an environment open to further developments. Matlab exists in Dos, Windows, and UNIX environments, aimed for digital signal processing. It allows a speed visualization of results, modelling, simulation, and design of complex digital systems. It is also suitable for the development of algorithms for processing data series, images, or multidimensional data fields. In this project, the version used is the Matlab R2014a. MySQL databases are created either under ESAYPHP or under WAMP. EasyPHP 2.0b1 [4] is a complete package, allowing you to use all the power and flexibility that the dynamic PHP language offers. The package includes an Apache server, a MySQL database, PHPMyAdmin as well as easy development tools for websites and applications. WampServer [5] is a Windows web development environment. It allows you to create web applications with Apache2, PHP, and a MySQL database. At the same time, PhpMyAdmin allows you to easily manage your databases.
134
H. Benabdallah and S. Kerai
2 Proposed Methodology 2.1 Local Topology Figure 2 presents architecture with a LAN (Local Area Network). It has two machines, a remote server, an RJ45 cable, and a modem that acts as an access point.
Fig. 2. The local area network topology.
2.2 Extended Topology Figure 3 presents architecture with a WAN (Wide Area Network). For the purpose to have remote access to the server; which is located at the central hospital, for example, that has a specialized line with a high speediness connection; from different clinics located in different places installed our application.
Fig. 3. The wide area network topology.
2.3 Our Developed Application • In a treatment room where machine one is available, installed Matlab there is:
Automatic Access Application of ICG and ECG Signals
135
To calculate the hemodynamic parameters and time intervals such as stroke volume (SV), thoracic fluid content (TFC), cardiac output (CO), heart rate variability (HRV), left ventricle ejection time (LVET), and preejection period (PEP) are defined in the following equations [6]: TFC = SV = Vc LVET ∗
1 Z0 1 Z0
(1)
dZ max dt
(2)
CO = HR ∗ SV
(3)
PEP = TQ − TB
(4)
Where, Vc is the intrathoracic blood volume expressed in mL, LVET is the left ventricle ejection time between B and X points of the ICG signals, Z0 : basic thoracic impedance, (dZ/dt)max is the maximum on the ICG signal curve, HR is the heart rate, and PEP is the time interval (T) between Q point of ECG signal and B point of ICG signal. We must applied a features point extraction algorithm for ICG signal and ECG signal: We applied our algorithm on signals of ten healthy subjects, when we recorded and analyzed ECG and ICG signals with a sampling frequency equal to 1000 Hz. We filtered the ECG signal using the low pass filter and high pass filter, we used an algorithm of detection of PQRS. It is based on the detection of the point R and the point-by-point methods for the others points. To denoise ICG signal from artifacts due to motion or respiration of subjects, we applied butter filter with order three. The Pan-Tompkins algorithm [7] is used to identify point C, whereas B and X are the local minima located before and after C, respectively. All results are displayed in the Matlab console and saved directly to the MySQL server using a simple program. We are based on two ways, where the first is a small application developed under Netbeans to send the figures of the ICG and ECG signals analysis to the MySQL server. The second is a Matlab program aimed to send the saved figures to the server automatically. • MySQL server is a remote server that has our database (DB). • Java Netbeans: An application under Netbeans installed on the user’s workstation, allowing access to information. For a remote server, we used WAMP/EASYPHP to create the database. In our project, we adjust the privilege of our database to open a communication with another machine (customers). Then, we use PHPMyAdmin MySQL to create the fifth table (ecg, icg, icg_signal, results, users). Under Matlab, we have implemented two programs: the first for the digital processing of the ICG signal, and the second for sending the results (features characteristics points, time intervals, and cardiac indices), thanks to the fastinsert function, after their processing, to a table (results) under a specific database (database) in
136
H. Benabdallah and S. Kerai
the MySQL server under WAMP or ESAYPHP, As explained above, a small application is developed to send the ICG/ECG figures to the server under a table (icg_signal). The Java Netbeans application has five interfaces with its function (login, menu, users, signals and results). This part is used to visualize and display the data already processed in the treatment room (which contains the Matlab software for the segmentation of ICGs signals), and also to enter various information of patients and users, the latter is identified by a password and a user ID for confidentiality reasons. The purpose of this application is to provide the user with all the necessary information, which helps him to establish a fast and reliable estimate. This application is a real-time automatic access application used to optimize the quality of care and the speed of diagnosis regardless of their geographical location, and is carried out according to two important criteria: the storage of information and the manipulation of data thanks to the connection between softwares. Connections MySQL connection with Matlab The fundamental objective set behind this part is to establish a connection between two environments: “Matlab and MySQl server” to store the information extracted from digital processing of the ICG/ECG signal. Before proceeding with any treatment, compatibility
Fig. 4. The connection of Matlab and MySQL. (a) connect to database, (b) enter the IP address and port number, (c) test the connection, (d) MySQL database appeared in Matlab.
Automatic Access Application of ICG and ECG Signals
137
between the systems used is mandatory. This compatibility facilitates the achievement of the following tasks (see Fig. 4): 1. Installation of ODBC driver; 2. The incorporation of the “database” that has already been created under MySQL in Matlab; 3. Confiscation of information in the MySQL connector window (name, server IP address), to test the connection; 4. Connection of the database. MySQL connection with Java Netbeans To establish MySQL & Java Netbeans connection, you must install a “MySQL JDBC Driver” compatible with the machine used, and then follow the following steps: 1. Connect Netbeans with MySQL: determine the server address. (A public but fixed address). 2. Connect Netbeans with the “database” database. 3. Creation of graphical interfaces. 4. Connect Netbeans with DB tables (so that the content of the tables will be visible in the interfaces created using java Netbeans).
3 Results and Discussions 3.1 Matlab Results Tables below (Table 1, Table 2 and Table 3) present some results of the ICG and ECG signals after processing in Matlab. 3.2 Netbeans Interfaces Login interface Security is necessary for any application to assure access to its functions. For this purpose, we have created an authentication mechanism (see Fig. 5) for the user with a correct user ID and password.
138
H. Benabdallah and S. Kerai Table 1. Feature points extraction for ECG signal.
Code of Number Times (second (s)) patient of cycle P Q R
Amplitudes S
P
Q
R
S
1
7
0.262 0.339
0.304 0.327 0.0088
0.0010
0.4118
−0.4421
2
1
0.432 0.532
0.471 0.494 0.0043
0.0001
0.4148
−0.4475
3
1
0.276 0.3851 0.318 0.341 0.0069 −0.0021
4
5
0.315 0.452
0.358 0.381 0.0042 −0.0056
0.4140
−0.4395
5
6
0.549 0.638
0.589 0.612 0.0077 −3.6 e−005 0.4130
−0.4451
6
2
0.559 0.663
0.598 0.621 0.0064
0.0010
0.4235
−0.4548
7
5
0.257 0.431
0.299 0.323 0.0068 −0.0064
0.4301
−0.4448
8
6
0.506 0.569
0.547 0.57
0.0045
0.4026
−0.4309
9
5
0.314 0.388
0.353 0.376 0.0073 −0.0004
0.4293
−0.4497
10
1
0.34
0.383 0.407 0.0096
0.4044
−0.4315
0.478
0.0111
0.0005
0.42100 −0.4579
Table 2. Feature points extraction for ICG signal. Code of patient
Number of cycle
Times (s)
Amplitudes
B
C
X
B
C
X
1
7
0.282
0.431
0.63
−0.1484
0.9576
−0.4985
2
1
0.482
0.587
0.795
0.0291
0.9033
−0.4306
3
1
0.317
0.455
0.656
0.0765
1.1106
−0.4501
4
5
0.362
0.486
0.702
−0.1304
1.1794
−0.4881
5
6
0.565
0.731
0.869
−0.0235
1.2501
−0.3251
6
2
0.582
0.735
0.904
−0.1073
1.1986
−0.3129
7
5
0.31
0.45
0.635
−0.1947
1.2849
−0.6462
8
6
0.51
0.69
0.82
−0.1631
1.1887
−0.2146
9
5
0.283
0.495
0.64
−0.2785
1.2444
−0.2432
10
1
0.368
0.532
0.702
−0.1617
1.2004
−0.4751
Menu interface The user can manage the database as he wants according to their choice (see Fig. 6) (User, Signals, and Results).
Automatic Access Application of ICG and ECG Signals
139
Table 3. Time intervals and hemodynamic parameters values. Code of patient
Number of cycle
HRV (ms)
LVET (s)
PEP (s)
SV (ml)
TFC (s)
CO (ml/ms)
1
7
835
0.348
0.0570
83.339
1.3525
69.58
2
1
835
0.313
0.05
102.57
1.3557
85.64
3
1
801
0.339
0.0681
85.144
1.348
68.20
4
5
776
0.34
0.090
87.885
1.323
68.19
5
6
826
0.304
0.073
120.26
1.334
99.33
6
2
769
0.322
0.081
127.91
1.333
98.36
7
5
802
0.325
0.0121
75.427
1.303
60.49
8
6
804
0.31
0.059
114.24
1.326
91.85
9
5
783
0.35
0.105
94.74
1.328
74.18
10
1
789
0.334
0.11
96.76
1.338
76.34
Fig. 5. Login interface.
Fig. 6. Menu interface.
140
H. Benabdallah and S. Kerai
Users interface The bottom users in Fig. 7 is a simple mechanism, grouped into a simple interface that stores the different user IDs and passwords of users who have the right to access medical information and who also have the right to read and write.
Fig. 7. User’s management interface; (a) search according to id, (b) add bottom option from the user interface, (c) add option saved to the server under the database.
The ICG images storage A small application developed under Java Netbeans is used to record images automatically under the MySQL server in a table (icg_signal). This interface presented in Fig. 8 allows us to enter the ID, the patient’s code, and the number of cycles, and then click on the browse button to bring the image. The latter will be stored in the table under MySQL.
Automatic Access Application of ICG and ECG Signals
141
Fig. 8. (a), (b), (c) and (d) The images storage interface from Netbeans to the MySQL database in table ‘icg_signal’.
Signals interface The following interface (see Fig. 9) presents the characteristic points of ICG and ECG signals for several cycles and patients according to ID, patient code, and the number of cycles. As well as the images display different signals after their treatments according to the ID.
142
H. Benabdallah and S. Kerai
Fig. 9. Signals interface. (a) search option with category id, (b) tables in the interface that presents the characteristics points of ICG and ECG signal processing in Matlab, (c) ecg table in the server, (d) icg table in the server.
Results interface The following interface presented in Fig. 10 shows the display of the results after the processing of the ICGs and ECGs signals from different subjects. The doctor can read and print the values presented in the interface. Each value has a normal range for a healthy person without suffering from any cardiac disease. For example: • Left ventricle ejection time (LVET) varying from 0.30 to 0.39 s • Preejection period (PEP) varying from 0.05 to 0.12 s • Stroke volume (SV) varying from 70 to 150 ml. Thanks to this information that can be print, the doctor or any medical person can estimate if the subject has cardiovascular disease or not.
Automatic Access Application of ICG and ECG Signals
143
Fig. 10. The results interface. (a) search with category id, (b) print option, (c) complete impression of the results presented in texfields, (d) results table.
144
H. Benabdallah and S. Kerai
4 Conclusion In this paper, we have presented the algorithm developed to detect the characteristic points BCX of the cardiography impedance signal (ICG) to measure various cardiac indices and calculate the time intervals. The results obtained seem very satisfactory and help the patient’s condition analysis. We have also presented another algorithm that allows us to record the values of the ICG signal obtained after its processing in the MySQL server, in parallel, a simple application that allows the insertion of the ICG images directly into the server. This app helps physicians diagnose each patient by cardiac time interval and hemodynamic parameters either locally or remotely.
References 1. Mansouri, S., Alhadidi, T., Chabchoub, S., Salah, R.B.: Impedance cardiography: recent applications and developments. Biomed. Res. 29, 3542–3552 (2018) 2. Boudreau, T., Glick, J., Greene, S., Spurlin, V., Woehr, J.J.: NetBeans: The Definitive Guide: Developing, Debugging, and Deploying Java code. O’Reilly Media, Inc. (2002) 3. Matlab, S.: Matlab. The MathWorks, Natick, MA (2012) 4. MySQL, A.B.: MySQL (2001) 5. Ipswich, D.: Setting up a WAMP Server on your windows desktop. Technology Now at Smashwords (2011) 6. Benabdallah, H., Kerai, S.: Respiratory and motion artefacts removal from ICG signal using denoising techniques for hemodynamic parameters monitoring. Traitement du Signal 38(4), 919–928 (2021) 7. Chakraborty, A., Chatterjee, S., Majumder, K., Shaw, R.N., Ghosh, A.: A comparative study of myocardial infarction detection from ECG data using machine learning. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 257–267. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-21642_21
OntoSpammer: A Two-Source Ontology-Based Spam Detection Using Bagging Dev Agrawal1 and Gerard Deepak2(B) 1 Department of Computer Science and Engineering, Jaypee Institute of Information
Technology, Noida, India 2 Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, India [email protected]
Abstract. With the advancement in technology that has made our lives far more accessible and better, is making our information insecure. Emails are considered the most reliable source for transferring messages, and this technology is hit hardest by spam. Spammers try to send unsolicited emails. Spam is one of the key to many cybercrime methods that use complex methods to trick particular victims. This chapter proposes a hybrid bagging approach by incorporating a pair of distinguished machine learning strategies namely the Support Vector Machine and Random Forest for spam email detection. We have incorporated the Dual Ontology model, i.e., domain ontology and ontology as a regular expression, to implement this model. The proposed OntoSpammer Model records accuracy and precision of 94.53% and 96.12%, respectively, and is compared and examined with benchmark and baseline approaches and is concluded to be one step up in terms of performance. Keywords: Bagging · Ontology · Semantic approach · Spamming
1 Introduction Electronic mail is an exact vehicle to convey information to others. It offers a way to converse with a person, or many, all in a comparatively private manner, which is not always readily available through the internet community. Due to the growing need for online communication, mail safety remains a weak point for many. Regarding email safety, adequate steps like cyber-attacks, mainly advanced social engineering attacks, cannot be blocked by antivirus software. Today more and more email safety risks are growing at an exponential rate. Because of its wide use, this medium is attacked by spammers to use it to spread avoidable advertisements, phishing, viruses, malware codes or Softwares, etc., which results in the installing of undesirable software on our system. Malware can be used to occupy personal data or to exploit the content. Frequently, spam emails are sent for advertisement purposes. With little to no operating costs beyond managing a mailing list, businesses can send thousands of mail on a daily basis. Spam email carries a path from where spammers can quickly gain access © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 145–153, 2022. https://doi.org/10.1007/978-981-19-1677-9_13
146
D. Agrawal and G. Deepak
to our systems. Spam email can be hard to stop, as a machine can send spam emails without the involvement of any humans. It is one of the unpreventable risks of being on the internet. It is hard to stop because any user with a mail address can "spam" any other justified mail address. If a message you received appears to be spam that doesn’t recognize the sender, you should immediately mark the message as spam in your email application. Don’t touch any links or attached files as spammers sometimes include these links to verify that your email address is valid or not, or the links may trigger unwanted web pages or downloads. With the incapacity to hold senders accountable, the list of spammers is increasing rapidly. The amount of unwanted emails arriving in the inbox is growing at an alarming rate, due to which crucial emails are hiding in the bulk of spam emails. Moreover, having quality filters in place doesn’t always stop spam. It sometimes still makes its way. More often, filters make things worse by filtering out the wrong messages, resulting in the removal of wanted emails. Motivation: Emails are considered the most reliable and fast source for transferring messages, but nowadays, spammers use them as a medium to advertise or hack information. Even important mail sometimes gets untouched because of a huge amount of unwanted mail in our inbox. So we need a way to detect spam messages to avoid uncertain troubles. In the present time, the data over the world wide web is increasing exponentially. There is a lot of traffic and a lot of data which is highly scattered because the web is turning toward the data semantic web, and since the data is present over the world wide web in large quantities, reasoning can be done among the data which is available and will be modeled into the ontologies. The existing models are not semantically driven. Therefore a semantic framework for email spam detection is essential which is proposed in this chapter. Contribution: Classifying an email as spam or non-spam mail is an arduous task. The spam box in our Gmail account is the best example of this. So here we are building a spam filter architecture model that uses a dual ontology model with bagging methodology. Firstly, the dataset is pre-processed with multiple techniques. The obtained data are further subjected to the feature extraction and feature weighing process. For the initial classification of the blog dataset, Bagging is installed into the process. The proposed OntoSpammer archives F-Measure, which is the combined measure that assesses the precision and recall tradeoff. This model records the F- Measure of 94.98%. Organization: The latter content of this chapter is structured as follows. Related Works are contained in Sect. 2. Section 3 briefly explains the Proposed System Architecture. Section 4 consists of the Implementation, and the Results and Performance Evaluation are discussed in Sect. 5. Finally, in Sect. 6, the chapter is concluded.
2 Related Works M. Basavaraju et al. [1] proposed a clustering method and implemented it to detect spam emails efficiently. The technique includes the implementation of open source in C language on the distances measured among email attributes. On large datasets, the combination of BIRCH and K-NNC gave better results because BIRCH is a good clustering algorithm, which traverses the dataset once, hence reducing time.
OntoSpammer: A Two-Source Ontology-Based Spam Detection
147
Asha S Manek et al. [2] presented an engrossed pre-processing approach to pick up the best features for the bag-of-words that come in ham and spam emails. The model is evaluated with W-BIF Reader classifiers. Houshmand Shirani-Mehr et al. [3] proposed naive Bayes with Laplace smoothing and SVM with linear kernel classifiers used for SMS spam detection. They added features like the message size based on the count of characters, added on the basis of thresholds, and logically dedudced the curves which are certain and wrongly classified data, which helped in improving the results. Amandeep S. Rajput et al. [4] In the first iteration, they divided it into two sets: a training set and the other as the classified set. In the following iteration, the classified sets are used as the training set, and the remaining sets are to be organized. This activity is recursively done till every set is used as a training set. The classification of the email is done based on the spam percentage gained by each email. Genuine mail is represented by the zero spam percentage, whereas the hundred percent spam percentage indicates spam mail. W.A. Awad et al. [5] used numerous algorithms (Naive Bayes, SVM, K-NN, Artificial Immune system, Rough sets) on the SpamAssassin spam corpus, and this implementation shows satisfying results with Naive Bayes and Rough sets methods with very good accuracy. Khyati Agarwal et al. [6] used ensemble and non-ensemble approaches to classify emails as spam and not spam. The bagging method, which is a part of the ensemble method, gives the best accuracy. Whereas in non-ensemble techniques, the SVM gave better results. In conclusion, the non-ensemble processes performed well in comparison with ensemble methods. Harjot Kaur et al. [7] illustrated many machine learning and non-machine learning algorithms. Discussed several techniques to detect spam emails. 100% results are impossible by any of the strategies, feature selection techniques like MLP neural networks provide high accuracy for detecting spam emails. But it has a drawback of selecting initial information points using a randomized approach which increases the time of the MLP algorithm.
3 Proposed System Architecture The proposed architecture is planned in a way to determine the avail information and unavailing information. The significant part of this proposed method includes an engrossed data pre-processing stage, as mentioned in Fig.1, followed by feature extraction and classification through bagging to detect spam emails. The considerability and quality of the dataset are upgraded to a better version by performing various operations, i.e., data extraction, transformation, integration tokenization, lemmatization. Initially, we need to extract the data, followed by a selection of attributes or instances. Once the instances have been selected, the uniformity of the data is studied, and redundancy has to be eliminated. Then, further cases will be checked, and any special characters involved are looked over. If punctuations are there, they will be removed. Moreover, if it is a URL, it will be subjected to URL normalization or URL canonicalization. After data transformation (The process of transformation covers the
148
D. Agrawal and G. Deepak
Fig.1. Data Preprocessing structure
data into a better form used for the mining process. The numerical or integer, or nominal datasets are converted to categorical datasets.), data integration (It is the preprocessing method to form a data store by involving the merging of data from diverse sources.), we will tokenize which is extracting tokens or individual words. Then it is lemmatized, using a wordnet lemmatizer. After that, stop word removal happens. That is, words with fewer meanings are removed.
Fig.2. OntoSpammer system architecture
From Fig. 2, pre-processed data is further subjected to feature extraction through domain ontologies. The word or item from the dataset is taken, and ontology is generated
OntoSpammer: A Two-Source Ontology-Based Spam Detection
149
using the Ontology tool. From this generated ontology, we will extract the features by using the lesk algorithm similarity. After feature extraction, features are weighted, and for weighting the features, we will be using the concept of information measure or Shannon’s entropy. Let’s define the self-information of an event X = x to be Eq. (1). I(x) = −logP(x)
(1)
Where I(x) is in the units of nats and the volume of information achieved by perceiving an event of probability 1/e is one nat. The volume of uncertainty in an entire probability distribution can be determined using the Shannon entropy in Eq. (2). H(x) = Ex∼P [I(x)] = −Ex∼P [logP(x)]
(2)
An ensemble learning is using multiple learning algorithms at the same time to obtain prediction with the aim of predicting better than individual learning models. It can make better predictions and achieve better performance than just one contributing model. An ensemble reduces the dispersion of the predictions and model performances. So After weighting, the features are sent into the ensemble technique. One ensemble learning technique: bagging, is applied in combination with the base learners. Here multiple models of the same learning algorithm are trained with random subsets of the training dataset, hence work better in terms of accuracy. Two base classifiers: Support Vector Machine and Decision Tree, are used in this chapter. With these base classifiers, ensemble learning will be implemented. Bootstrap aggregating, also called bagging, is a resampling technique used to reduce the high variance of the algorithms by making multiple subsets of the training dataset, and each subset is used to build the classifier. After using bagging, variance can be reduced, and overfitting can be kept away. A support vector machine (SVM) is a selective classifier that is formally designed by a separative hyperplane. It gives better results in high-dimensional spaces. The role of the kernel is to input data and transform it into the required form. In our model, we have to use the sigmoid kernel, which is defined as in Eq. (3) K(X , Y ) = tanh(γ .X T Y + r)
(3)
Decision tree is applicable to two different data mining techniques, classification and regression. It is used to generate rules that can be visualized using a tree, making it simple to interpret and understand. The tree is drawn from top to bottom, with the root at the top giving it a flowchart-like structure. An inner node represents “Test” on the attribute, the output of the test is shown by each branch, and leaf is used to represent a result of the test. The classification rules are represented by the path from root to leaf. It is a way to show an algorithm having conditional statements. Once the bagging is done, we come into the ontologies of rules to find whether the item is spam. So what happens here is, the spam rules generally form as regular expressions, and these regular expression ontologies of rules are sent to the classified result to find deviations. Rules will predict whether the mail is spam or not, that particular rules are made into regular expressions and transformed into the ontology of rules. We will check the item one by one to predict spam if there is a deviation; otherwise, not. So they should have variations because this is the rule to predict whether it is spam or not.
150
D. Agrawal and G. Deepak
4 Implementation The proposed OntoSpammer was designed and successfully implemented, python 3.9.0 was used to implement OntoSpammer’s architecture. Google Colab Notebook was the IDE that has been utilized to design and implement the architecture. Google Colab is offering two different models of GPU NVIDIA K80/T4 for free usage and 100 GB of free space for data, and 16 GB of RAM for which can provide high-performance computing tasks. The data set used in this architecture to achieve the aim is that the Ling Spam corpus, curated from the Linguist List. These mails centered on semantic interests around job opportunities and software discussion. In this dataset from Kaggle, some of the observations are: header information is removed, it contains legitimate messages (2412), and spam messages (481). The algorithm of the proposed OntoSpammer is depicted in Algorithm 1. Algorithm 1: Proposed OntoSpammer Architecture Input: Ling spam dataset, google based knowledge, Domain ontologies, Rules ontologies Output: Classifying the e-mails as spam or ham Begin-Step 1: Extraction, Transformation, Integration, Tokenization, Lemmatization, removal of stop words take place on the dataset. Step 2: For the pre-processed data While(data.next!=NULL){ - feature extraction takes place through domain ontologies. } Step 3: Features are weighed using the concept of information measure. Step 4: Classification takes place by using ensemble technique with random forest classifiers and support vector machine. Step 5: Ontologies of rules are sent to the classified result to find deviations. Step 6: Rules will predict whether the e-mails are spam or not. --End
5 Results and Performance Evaluation One of the most common natural language processing tasks for emails and chat engines categorizes spam and ham messages. Spam messages from ham messages can be separated with the growth in machine learning and natural language processing techniques and even with a high degree of precision. Any recommendation system’s goal is to encompass the highest and most good suggestions to its customers. To achieve this goal, the outcomes of such recommendation systems must be checked and evaluated since it
OntoSpammer: A Two-Source Ontology-Based Spam Detection
151
is critical to analyze how applicable the search content is for a specific user’s desired category. Therefore Precision, Recall, Accuracy, F-Measure, and False Negative Rate (FNR) are considered performance metrics to evaluate the proposed architecture. Precision, Recall, Accuracy, and the F-Measure computes the relevance of results while FNR measures the False Negative Rate made by the framework. Standard Formulations were used for calculating the Precision, Recall, Accuracy, F-Measure, and FNR. Table 1. Comparison of Performance of the proposed OntoSpammer with baseline models. Model
Average Precision %
Average Recall %
Average Accuracy %
Average F-measure %
FNR
NMSTBC [1]
87.61
82.18
85.29
84.80
0.18
RePID-OK [2]
90.21
86.17
89.27
88.14
0.14
SMSSD-MLA [3]
91.33
86.28
89.94
88.73
0.14
Proposed OntoSpammer
96.12
93.87
94.53
94.98
0.07
The proposed OntoSpammer architecture performance was evaluated and compared with the baseline model and is tabulated in Table 1. Table 1 shows that the proposed approach performance overtakes all the compared baseline models. According to Table 1, the performance increase of the OntoSpammer in terms of average Precision, Recall, Accuracy, F-measure, and FNR compared to the NMSTBC [1] (using clustering approach) is 8.51%, 11.69%, 9.24%, 10.18%, 0.11% respectively. When the same perimeters compared to RePID-OK [2] (Repetitive Preprocessing technique using imbalanced dataset by selecting optimal number of keywords), the observed difference is 5.91%, 7.7%, 5.26%, 6.84%, 0.07% respectively, when compared to SMSSD-MLA [3] (using 10fold cross-validation), the observed difference is 4.79%, 7.59%, 4.59%, 6.25%, 0.07% respectively. The Fig. 3 graph depicts the comparison of accuracy vs. the number of instances for the proposed OntoSpammer w.r.t the NMSTBC [1], RePID-OK [2], and the SMSSDMLA [3], respectively. The advantage of OntoSpammer when compared to the proposed base-line models is that the dual ontology model is used, i.e., domain ontology and ontology as a regular expression, bagging as we are using two classifiers (SVM and Decision tree), feature extraction, and combining all of these methods the proposed OntoSpammer gives the highest precision percentage, recall percentage, F-measure percentage, and the lowest FNR.
152
D. Agrawal and G. Deepak
Fig. 3. Accuracy vs. Number of Instances
6 Conclusions An email has been the most crucial medium of communication nowadays. Through internet connectivity, any message can be delivered all over the world. On the other hand, For many companies and individuals, spam is an annoyance and undesired expense. This chapter presented a comprehensive review of the most effective e-mail spam filtering techniques. OntoSpammer is based on ensemble learning technique, achieved by implementing the two machine learning strategies namely Support Vector Machine and Decision tree using bagging for the spam email detection has achieved tremendous success against other spam detection techniques. The proposed architecture records a precision of 96.12%. We can further enhance this architecture by hybridization and incorporating scalable AI techniques.
References 1. Basavaraju, M., Prabhakar, R.: A novel method of spam mail detection using text-based clustering approach. Int. J. Comput. Appl. 5(4), 15–25 (2010) 2. Manek, A.S., et al.: RePID-OK: spam detection using repetitive pre-processing. In: 2013 International Conference on Cloud & Ubiquitous Computing & Emerging Technologies (2013). https://doi.org/10.1109/cube.2013.34 3. Shirani-Mehir, H.: SMS Spam Detection using Machine Learning Approach 4. Rajput, A.A., Athavale V., Mittal S.: “Intelligent Model for Classification of SPAM and HAM”. International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-8 Issue-6S, (2019) 5. Awad, W.A., ELseuofi, S.M.: Machine learning methods for spam e-mail classification. Int. J. Comput. Sci. Inf. Technol. (IJCSIT). 3(1), 173-184 (2011). https://doi.org/10.5121/ijcsit. 2011.3112
OntoSpammer: A Two-Source Ontology-Based Spam Detection
153
6. Agarwal, K., Uniyal, P., Virendrasingh, S., Krishna, S., Dutt, V.: Spam mail classification using ensemble and non-ensemble machine learning algorithms 7. Kaur, H., Verma, E.P.: Survey on e-mail spam detection using the supervised approach with feature selection. Int. J. Eng. Sci. Res. Technol. (IJESRT) 6(4), (2017) ISSN: 2277-9655. https://doi.org/10.5281/zenodo.496096 8. Bedi, P., Goyal, S.B., Rajawat, A.S., Shaw, R.N., Ghosh, A.: A framework for personalizing atypical web search sessions with concept-based user profiles using selective machine learning techniques. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 279–291. Springer, Singapore (2022). https:// doi.org/10.1007/978-981-16-2164-2_23 9. Almeida, T.A., GÃ3 mez Hidalgo, J.M., Yamakami, A.: Contributions to the study of SMS spam filtering: new collection and results. In: Proceedings of the 2011 ACM Symposium on Document Engineering (DOCENG 2011), Mountain View, CA, USA (2011) 10. Chen, Y., Lu, C., Huang, C.: Anti-spam filter based on Naïve Bayes, SVM, and KNN model
SAODFT: Socially Aware Ontology Driven Approach for Query Facet Generation in Text Classification Rituraj Ojha1(B) and Gerard Deepak2 1 Department of Metallurgical and Materials Engineering, National Institute of Technology,
Tiruchirappalli, India [email protected] 2 Department of Computer Science and Engineering, National Institute of Technology, Tiruchirappalli, India
Abstract. Query facet is a collection of topics or information that provides an overall summary of the user query. Only a few words or phrases from the relevant query facet can describe the user query better than the long sentences and this can be helpful during searches made on public search engines. In this chapter, a socially aware and ontology driven query facet generation framework is proposed. The proposed framework takes user query as input which is preprocessed and the obtained individual query words are integrated with domain ontologies. Furthermore, terms and events are integrated from the Twitter API which makes the model socially aware and relatable to real-world social events. The metadata is generated using OpenCalais and RDF Distiller, and is classified using Random Forest, taking individual query words as input and only the top 20% of the classified instances are taken. Similarly, the RandQ dataset is classified by taking individual query words as input using Random Forest. Finally, SemantoSim measure and Gamma Diversity measure is computed between the classified instances from the RandQ dataset and the instances coming from top 20% of the classified metadata and the entities passing through threshold value is formulated into a facet set by grouping them on decreasing value of the semantic similarity. The proposed approach achieves the highest average accuracy with the Precision of 90.61%, Recall of 93.18%, Accuracy of 91.89%, Purity of 0.95, F-Measure of 91.88%, nDCG value of 0.90, and low False Discovery Rate of 0.094 making it better approach than the baseline models. Keywords: Facet recommendation · Metadata generation · Ontology driven approach · Query facet generation · Text classification
1 Introduction The world wide web is populated with a large amount of information, including texts, images, videos, documents, etc. which keeps increasing every day. There is a large amount of data that a user can search on public search engines through his query. With increase in the amount of information available on the internet, it is getting harder and © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 154–163, 2022. https://doi.org/10.1007/978-981-19-1677-9_14
SAODFT: Socially Aware Ontology Driven Approach
155
harder for users to find the relevant information pertaining to their submitted queries. In such cases, query facets play an important role helping the user to get relevant and most important information from the world wide web. Query Facets are collections of topics that summarize and describe an important and essential feature of the query. Facets can be phrases or a word. Several facets can be available for a query which describes the query from all the different perspectives. Hence, query facets can help to improve search results through its knowledge. For example, facets for query watches can describe the query in several distinct ways such as categories, styles, brands, gender and colors. Users can get important aspects of watches, such as brands like Rolex, Omega, etc. and also discover new information without wasting time by browsing through a lot of websites to get appropriate information. Users can also visit the women’s section to gift a watch to their female friend, as recommended by the facets. The query facets can also help in searches for ambiguous queries like, Apple, where searches can contain results from fruit apple and also from technology brand Apple. For this, different query facets will be available to choose for both, which contains detailed information about fruit in one and technology brand Apple, in the other. Facets can also improve the diversity of the information by re-ranking the search results and eliminating the websites containing duplicate information. Hence, there is a need for a query facet generating tool which can be ontology driven and socially aware to work better than the already existing approaches. In the previous work, Ramya et al., has proposed an approach for the extraction of query facets by computing cosine similarity and using a high quality clustering algorithm [1]. Motivation: Query facets can summarize the entire user query from all the diverse perspectives which can help in getting better search results in the first page of the search engine. Hence, a better query facet recommendation framework is needed. Also, there are very few systems in this domain that are socially aware of the current real-world events and also return results with high accuracy. Therefore, building this system was a major motivating factor. Contribution: The socially aware and ontology driven query facet generation approach is proposed. The framework takes the user query as input which is preprocessed and the domain ontologies are integrated with the obtained individual query words. Furthermore, terms and events are integrated from the Twitter API. The metadata is generated and is classified using Random Forest, taking individual query words as input and only the top 20% of the classified metadata instances are taken. Similarly, the RandQ dataset is classified by taking individual query words as input using Random Forest. Finally, SemantoSim measure and Gamma Diversity measure is computed between the classified instances from the RandQ dataset and the instances coming from top 20% of the classified metadata and the entities passing through the threshold value is formulated into a facet set by grouping them on decreasing value of the semantic similarity. The values of metrics namely, Precision, Accuracy, Purity, Recall, nDCG, and F-Measure are increased. Organization: The flow of the remaining chapter is as follows. A condensed summary of the related works is provided in Sect. 2. Section 3 depicts the architecture of the proposed system. Section 4 describes the architecture implementation. Section 5 presents the evaluation of results and performance. The conclusion of the chapter is presented in Sect. 6.
156
R. Ojha and G. Deepak
2 Related Works Kong and Allan [3] have proposed a methodology for query facet extraction using the empirical utility maximization approach. This approach increases the performance and is a precision-focused approach. The proposed approach only shows query facets for a few queries from all the input queries by calculating extraction performance for each query and giving only a few facets based on the extraction performance value. Jiang et al. [4] have proposed a technique for query facet mining using Knowledge Bases instead of using traditional methods of taking topics from top search results. Their approach mines initial facets from search engines and then entities from the Freebase knowledge base are integrated into it. Chakraborty et al. [5] have proposed a model for faceted search of scientific articles on the internet. Their framework, FeRoSA, is based on the random walk model and it also organizes the scientific articles into facets based on their category. Ramya et al. [6] have proposed a technique for mining the query facets for user query automatically. The grouping of the facets happens by using a high quality clustering algorithm and calculating cosine similarity and the top items are collected as facets and recommended to the user. Hu et al. [7] have proposed a method for diversifying the search results of the different users searching with the same query. The framework takes the help of query facets for the input user query and generates subtopics for each topic present in the facet. Vaishakhi and Regi [8] have proposed an approach for mining query facets with high quality by introducing their framework, namely, Anchor. Radhakrishnan and Madhav [9] have proposed a query facet search engine for improving the search results of user queries. The engine will itself fetch the relevant facets for the search query and facets will be selected priority-wise based on their importance. Siddiqui et al. [10] have proposed a model to categorize the large documents using facets. This is done by extracting facets from these documents using their proposed framework, namely FacetGist which constructs a graph-based heterogeneous network to collect data.
3 Proposed System Architecture The architecture for query expansion is depicted in Fig. 1. The proposed approach takes place in several steps. Initially, the User Query is taken as an input and is preprocessed using Tokenization, Lemmatization, stop word removal, and Named Entity Recognition (NER). Tokenization is the process of breaking the texts into small pieces called tokens. Byte Pair Encoding (BPE) is used for the tokenization process. Lemmatization involves grouping together several inflected kinds of the same word so that they are analyzed as one term. During stop word removal, the common ubiquitous words are removed as they add no value for the analysis and only increase the dimension of the feature set. NLTK python library is used for stop word removal. NER is the process of finding and categorizing the data or entity into predefined categories. After the preprocessing stage, individual query words are obtained.
SAODFT: Socially Aware Ontology Driven Approach
157
These individual query words are integrated with domain ontologies. These domain ontologies are both static and dynamic. Static domain ontologies are ready-made ontologies which are already made using available domain ontologies to which the query as well as the dataset belongs to is used and the dynamic ontology is generated using user blogs, community forums, and contents from world wide web, using Onto Collab as a tool [11]. The next step involves integration of terms and events from the Twitter API. This makes the proposed methodology socially aware and relatable to real-world social events, as well as socially popular tags. Metadata is generated using OpenCalais and RDF Distiller. The metadata is classified using Random Forest, taking individual query words as training input. Only, top 20% of the classified metadata instances are selected. The reason for selecting only the top 20% is because the metadata obtained is extensively large and we only want the most relevant and consolidated entities from the metadata. Random Forest is an ensemble machine learning model which can perform classification as well as regression, and any other tasks which can be done by constructing and training a very large number of decision trees. During the classification of metadata using Random Forest in the proposed methodology, the output class is the class which is selected by most decision trees. Random Forest is based on Bagging algorithm, where for each decision tree the model selects random bootstrap samples and a random subset of features from the training set.
Fig. 1. Proposed system architecture
The dataset selected for this approach is RandQ. This dataset is built by Dou et al. [2] by collecting queries and building the query facets by extracting the top results from
158
R. Ojha and G. Deepak
the search engine using their tool, namely, QDMiner. The dataset is in XML format and 105 queries are taken from the dataset. Each query contains 100 documents and each document contains Document id, Title, description, URL, rank of the document, document text, and HTML list. The dataset is classified by taking individual query words as input using Random Forest. Now, we will compute the SemantoSim measure and Gamma Diversity between classified instances from the dataset and the instances coming from the top 20% of the classified metadata. The threshold for the SemantoSim measure is 0.75 and the threshold for Gamma Diversity is 0.5. The entities passing through the threshold values are formulated into a facet set by grouping them in order of decreasing value of the semantic similarity, such that the higher value and most similar entities are kept first as compared to low value semantic similarity. To compute the semantic similarity and Gamma Diversity, a software agent is designed with the state and behaviour. The state will compute both the SemantoSim measure and Gamma Diversity, and behaviour will take the intersection of terms passed from semantic similarity and Gamma Diversity and group them as facets in decreasing value of semantic similarity which is then recommended to the user. Gamma diversity as represented by Eq. (1), is the diversity of total species in a landscape, an island, or an area. The Alpha Diversity, or α in Eq. (1) represents the mean species diversity in the habitats of a local region. The Beta Diversity, or β in Eq. (1) represents the species diversity between two adjacent ecosystems. γ =α+β
(1)
SemantoSim is a semantic similarity measure proposed by Church & Hanks. Equation (2) represents the formula for computing SemantoSim Measure. p(x, y)log p(x, y) + pmi(x, y) (2) SemantoSim(x, y) = log p(y, x)] + [p(x) ∗ p(y)
4 Implementation The dataset used in this proposed work is the RandQ dataset. It contains 105 queries in XML format. Every query contains 100 documents and each document contains Document id, Title, description, URL, rank of the document, document text, and HTML list. The implementation is done in the Google Collaboratory environment. The computer is equipped with 16 GB RAM and an i7 processor. Static domain ontologies are readymade ontologies which are already made using available domain ontologies to which the query as well as the dataset belongs to is used and the dynamic ontology is generated using user blogs, community forums, and contents from world wide web, using Onto Collab as a tool.
SAODFT: Socially Aware Ontology Driven Approach
159
5 Results and Performance Evaluation The Performance of the proposed SAODFT (Socially Aware Ontology Driven Approach for Query Facet Generation in Text Classification) approach is measured by considering metrics namely Precision, Recall, Accuracy, and Purity. Other metrics including Normalized Discounted Cumulative Gain (nDCG), F-Measure and False Discovery Rate (FDR) are also calculated. Precision = Recall =
Retrieved ∩ Relevant Retrieved
(3)
Retrieved ∩ Relevant Relevant
(4)
Precision + Recall 2
(5)
2 × Precision × Recall Precision + Recall
(6)
Accuracy = F − Measure =
nDCG =
DCG ∝ IDCG ∝
(7)
160
R. Ojha and G. Deepak
DCG =
∝
rel i + 1)
(8)
i=1 log 2 (i
False Discovery Rate = 1 − Positive Predictive Value Purity =
1 maxd ∈D |m ∩ d | × m∈M N
(9) (10)
The Eqs. (3), (4), and (5) represent Precision, Recall and Accuracy, respectively. Furthermore, the Eqs. (6), (7), and (8) represent F-Measure, nDCG and Discounted cumulative gain, respectively. Equations (9) and (10) represent FDR and Purity, with ‘M’ clusters and ‘D’ classes partitioning N number of datapoints. The reason for considering Accuracy, Precision, Recall, Purity and F-Measure metrics for evaluation is because they measure the relevance of the results and FDR computes the number of false discoveries made by the proposed system. The nDCG represents the diversity of the relevant results. Table 1. Comparison of performance of the proposed SAODFT with other approaches for the RandQ dataset. Search Technique
Average Precision %
Average Recall %
Accuracy %
Purity
nDCG
F-Measure
FDR
QDMiner [2] 70.14
74.14
72.14
0.88
0.68
72.08
0.298
AEFTM [1]
71.14
77.18
74.16
0.92
0.72
74.04
0.288
Proposed SAODFT
90.61
93.18
91.89
0.95
0.90
91.88
0.094
Eliminating Metadata Generation from the Proposed SAODFT
81.15
85.12
83.14
0.86
0.81
83.08
0.188
Eliminating SemantoSim Measure from SAODFT
83.44
86.17
84.81
0.84
0.85
84.78
0.165
Eliminating Gamma Diversity Index from SAODFT
82.17
85.78
83.98
0.83
0.86
83.93
0.178
Table 1 represents the performance comparison of the proposed SAODFT with that of base paper along with other approaches considered by using elimination techniques. It is
SAODFT: Socially Aware Ontology Driven Approach
161
evident from the table that the proposed approach achieves the highest average accuracy with the Precision of 90.61%, Recall of 93.18%, Accuracy of 91.89%, Purity of 0.95, F-Measure of 91.88%, nDCG value of 0.90, and low FDR of 0.094. The baseline model for this paper AEFTM [1], achieves much less average accuracy, getting Precision of 71.14%, Recall of 77.18%, Accuracy of 74.16%, Purity of 0.92, F-Measure of 74.04%, nDCG value of 0.72, and FDR of 0.288. The performance of the SAODFT framework is better because of several reasons. Firstly, the metadata generation increases the realworld knowledge. Secondly, domain ontology integration enhances the probability of incidence of concepts which are highly relevant to the base framework. Thirdly, the Twitter API makes the framework socially aware of real-world events. Lastly, SemantoSim measure ensures that the relevance of the topics is high and Gamma Diversity ensures that the diversity of the relevant topics is also high. Eliminating any of the used methodologies in the proposed SAODFT framework will lead to a significant decrease in the average accuracy of the framework. In QDMiner [2], the approach extracts the query facets by gathering the similar topics from the top search results on the search engine. This can lead to getting irrelevant noise and duplicate results. The AEFTM [1] is not as good as the proposed framework because it uses only the text mining approach. Also, they are using only Clustering algorithm with cosine similarity which is highly naïve and traditional. They are using domains from the dataset, but the proposed framework uses domains from dataset along with the domain from social network, Twitter. Hence, the proposed SAODFT framework works much better than the baseline models.
Fig. 2. Precision % vs number of recommendations
162
R. Ojha and G. Deepak
Figure 2 represents Precision % vs number of recommendations line graph for all the approaches. It is evident from the figure that the proposed SAODFT framework achieves the highest Precision % for any number of recommendations. Even after the elimination of one methodology used in the proposed approach, the precision remains greater than the two baseline models, AEFTM [1] and QDMiner [2]. Hence, we can conclude that the proposed model is better and more accurate than the baseline models.
6 Conclusions In this chapter, a socially aware ontology driven approach for query facet generation in text classification is proposed. The increase in the amount of information available on the internet has made it harder for users to find the relevant information pertaining to their submitted queries. Therefore, query facets play an essential role in helping the user to get relevant and most important information from the world wide web. Query Facets are collections of topics that summarize and describe an important and essential feature of the query. The proposed framework takes user query as input which is preprocessed and the obtained individual query words are integrated with domain ontologies. Furthermore, terms and events are integrated from the Twitter API. The metadata is generated using OpenCalais and RDF Distiller, and is classified using Random Forest, taking individual query words as input and only the top 20% of the classified instances are taken. Similarly, the RandQ dataset is classified by taking individual query words as input using Random Forest. Finally, SemantoSim measure and Gamma Diversity measure is computed between the classified instances from the RandQ dataset and the instances coming from top 20% of the classified metadata and the entities passing through threshold value is formulated into a facet set by grouping them on decreasing value of the semantic similarity. The proposed approach achieves the highest average accuracy with the Precision of 90.61%, Recall of 93.18%, Accuracy of 91.89%, Purity of 0.95, F-Measure of 91.88%, nDCG value of 0.90, and FDR of 0.094. The overall accuracy of the proposed technique is much better than the existing approaches.
References 1. Ramya, R.S., Raju, N., Pushpa, C.N., Venugopal, K.R., Iyengar, S.S., Patnaik, L.M.: Automatic extraction of facets for user query in text mining [AEFTM]. Int. J. Emerg. Technol. 11(2), 342–350 (2020) 2. Dou, Z., Jiang, Z., Hu, S., Wen, J.R., Song, R.: Automatically mining facets for queries from their search results. IEEE Trans. Knowl. Data Eng. 28(2), 385–397 (2015) 3. Kong, W., Allan, J.: Precision-oriented query facet extraction. In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. 1433–1442 (2016) 4. Jiang, Z., Dou, Z., Wen, J.R.: Generating query facets using knowledge bases. IEEE Trans. Knowl. Data Eng. 29(2), 315–329 (2016) 5. Chakraborty, T., Krishna, A., Singh, M., Ganguly, N., Goyal, P., Mukherjee, A.: Ferosa: a faceted recommendation system for scientific articles. In: Bailey, J., Khan, L., Washio, T., Dobbie, G., Huang, J.Z., Wang, R. (eds.) Advances in Knowledge Discovery and Data Mining: 20th Pacific-Asia Conference, PAKDD 2016, Auckland, New Zealand, April 19-22, 2016,
SAODFT: Socially Aware Ontology Driven Approach
6.
7. 8.
9.
10.
11.
163
Proceedings, Part II, pp. 528–541. Springer International Publishing, Cham (2016). https:// doi.org/10.1007/978-3-319-31750-2_42 Bedi, P., Goyal, S.B., Rajawat, A.S., Shaw, R.N., Ghosh, A.: A framework for personalizing atypical web search sessions with concept-based user profiles using selective machine learning techniques. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 279–291. Springer, Singapore (2022). https:// doi.org/10.1007/978-981-16-2164-2_23 Hu, S., Dou, Z.C., Wang, X.J., Wen, J.R.: Search result diversification based on query facets. J. Comput. Sci. Technol. 30(4), 888–901 (2015) Kumar, A., Das, S., Tyagi, V., Shaw, R.N., Ghosh, A.: Analysis of classifier algorithms to detect anti-money laundering. In: Bansal, J.C., Paprzycki, M., Bianchini, M., Das, S. (eds.) Computationally Intelligent Systems and their Applications. SCI, vol. 950, pp. 143–152. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0407-2_11 Radhakrishnan, A., Madhav, M.L.: Query facet engine for easier search results. In: 2017 International Conference on Circuit, Power and Computing Technologies (ICCPCT), pp. 1–5 (2017) Siddiqui, T., Ren, X., Parameswaran, A., Han, J.: Facetgist: collective extraction of document facets in large technical corpora. In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management, pp. 871–880 (2016) Pushpa, C.N., Deepak, G., Thriveni, J., Venugopal, K.R.: Onto Collab: strategic review oriented collaborative knowledge modeling using ontologies. In: 2015 Seventh International Conference on Advanced Computing (ICoAC), pp. 1–7 (2015)
SDPO: An Approach Towards Software Defect Prediction Using Ontology Driven Intelligence N. Manoj1 and Gerard Deepak2(B) 1 Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Ramapuram, Chennai, India 2 Department of Computer Science and Engineering, National Institute of Technology,
Tiruchirappalli, India [email protected]
Abstract. Software defects arise from various modules and are of various categorical types. Complex software debugging process use up a lot of indispensable computational resources, hence there is an urgent need to optimize the defect detection process. Software defect Classification has been a widely researched topic since a long time, concentrating on predicting the type of software defect from recorded metrics at the time of defect generation. However, the models never concentrated on the sematic understanding of the defects. This chapter proposes a hybrid classifier, that tracks and collects defect metrics by a bug tracker tool, and uses an ontology of software defects to get a semantic understanding of the features of the defects. These features and metrics are passed into a deep learning model to classify and recommend software defect categories. The proposed model significantly improves accuracy of classification compared to notable software defect classification models. Keywords: Software defects · Ontology · Semantic similarity · GRU model · Software defect classification
1 Introduction A software defect can be defined as any fault in a computer program which makes it deviate from its actual performance. Such defects have always caused problems at the time of production and deployment. In this age of rapid deployment and use of various increasingly complex software products, the issues and errors generating from the different components of the product increase and become a lot diverse. Specific classes of software types, or specific patterns of writing code result in error prone instances. Rectifying errors without documenting or preserving the debugging information and related defect causes and terminologies is a wastage of important resources which could have been reused if a similar anomaly was spotted again. This chapter proposes a solution on essentially preserving and representing the information on detecting and debugging software defects in the form of software defect ontologies, which will further be a source of evidence to detect and classify an unknown software defect in the future. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 164–172, 2022. https://doi.org/10.1007/978-981-19-1677-9_15
SDPO: An Approach Towards Software Defect Prediction
165
An ontology is a representation of information about a specific domain, which essentially represents the entities, concepts and definitions involved in that domain in an interconnected coherent manner. An ontology dedicated to defects occurring in particular software will represent the causes, the error classes, debugging concepts and the coherent links between the occurrence of various kinds of defects which will essentially preserve a shared understanding of the concepts, for reference and reuse in the future. Manifesting shared knowledge on a particular domain makes is very easy and highly informative for the agents using the ontology to understand the software. Motivation: The motivation behind this chapter is to optimize software defect detection and debugging of the same, which tends to be an increasingly tedious and computationally expensive process. Using an ontology represents shared knowledge on a particular concept and can be used for analysis and related problem solving in the future. A defect ontology will interconnect descriptions and occurrence of anomalies and inform the developers about the semantic nature of the bug in the software. This knowledge can be further used to classify and detect different kinds of anomalies in a software. Contribution: The chapter describes a cost-effective solution to software defect detection process, by using defect ontologies to get semantic alignment of different classes of errors occurring in a particular software. This is used to train a classification model, which will be able to detect unknown software anomalies accurately in the future. Organization: The chapter is divided into 5 sections, Sect. 2 provides the related work description, Sect. 3 explains the model architecture, Sect. 4 describes the implementation and evaluation of the model with respect to other baseline models, Sect. 5 concludes the chapter.
2 Related Work Traditional methods of software detection and classification was based upon software metrics, which lacked the semantic relation between the defects it predicted or classified. On the other hand, ontologies are designed to semantically represent data which promotes reasoning, sharing and reuse of knowledge, hence they greatly increase the classification accuracy. Bruno Borlini Duarte et al. [1] have proposed a reference conceptual model (domain ontology) called the Ontology of Software Defects, Errors and Failures (OSDEF), to tackle the lack of rational management of software defect resources. OSDEF is grounded on UFO and built using SABiO. Assessing the severity of software defects is fundamental to efficiently test and release software models, as explained by Martin Lliev et al. [2], hence they have proposed a solution to this by automating prediction of defect severity, with the help of reasoning and design knowledge of ontologies. ODCs (Orthogonal Defect Classification models) have been prevalently used in software industries for defect classification and abstraction, but they lack semantic representation when the tasks get very complicated, hence Xuan Hu & Jie Lu [3] propose a solution to this by generating 4 different software-hardware integrated error pattern a sub-category of SREP (Software Error Requirement Stage), according to an error generation mechanism. A corresponding ontology representation is given, focusing on scenario, solution, and error manifestation of the generated errors. Hongliang Liang et al. [4] have proposed a novel framework to combine word embeddings and deep recurrent neural networks
166
N. Manoj and G. Deepak
(LSTM) to classify and predict software defects. They extract tokens from abstract syntax trees, and use the token to get vectors from a mapping function. These vectors are trained on LSTMs, as a supervised learning algorithm. To deal with the problem of data redundancy due to large number of features, Locally Linear Embedding algorithm was modelled along with SVM classifier by Chun Shun et al. [5]. LLE algorithm uses an embedding dimension and K nearest neighbor algorithm to reduce the dimensionalities of the features. SVM is one of the most traditional software used for classification problems, because of its robustness to nonlinear functions and good performance on small samples. The paper shows improvement in performance by the LLE and SVM hybrid model compared to pure SVM models. Mutual Information with Fuzzy Integral, Lu Liu et al. [6], is an innovative approach to classifying software defects using fuzzy measures and fuzzy integrals. The property of non-additivity and monotonicity of fuzzy measures is used to generate mutual information and correlations between different features of software defects, useful to better understand the problem at a deeper level. Fuzzy measures reflect importance of individual attributes and combination of selected attributes in classifying software defects. These fuzzy measures are used to predict the class of software defect in the experiments. Dyana Rashid Ibrahim et al. [7], implemented Random Forest Classifier in their paper to classify and recommend defects. They used Correlation based Feature Selection (CFS) for feature selection by taking into account, the subsets of the most related classes and also subsets that are not correlated. These subsets are searched using the Bat-Search metaheuristic algorithm. The obtained features are used as an input to random forest classifier model. The proposed approach performed better than most traditional software defect classifiers such as SVM, IRF etc.
3 Proposed System Architecture The proposed model architecture is shown in Fig. 1. The software defect messages or issues will be collected as to form a raw defect dataset. This raw data will be preprocessed with various techniques and then, a software defect ontology and a library will be used for semantic alignment of the error messages. A GRU classification model will be trained on the generated enriched and semantically aligned dataset, to classify and recommend defects to developers. The raw software defect dataset will be preprocessed by various prevalent techniques used in Natural Language Processing. First, the dataset will be lemmatized, to generate root words out of its variations, this ensures uniformity in phrases and makes it easier for the computer to analyze. The lemmatized phrases will be cleaned with stop words and rare words removal. Finally, the cleaned dataset will be used to generate Name Entity Recognizers and tokenizers, which are useful semantic analysis, frequency and statistics for the occurring words in the phrases. A defect ontology will be used on the cleaned dataset get semantic alignments of the occurring errors. This will result in a coherent relation between the anomalies, a description of how and where these errors generally arise from and how they can be rectified. This strategic representation of data is obtained by the software defect ontology and a software defect library. The semantic alignment with the help of ontologies will be done by two techniques, NPMI and Regex.
SDPO: An Approach Towards Software Defect Prediction
167
Fig. 1. Proposed architecture
Pointwise Mutual Information (PMI) of two discreet variables x and y is defined by the probability of coincidence of x and y, given their joint distributions and individual distributions, assuming independence. PMI can be normalized to get the values between [−1, +1]. Equations (1), (2) give the PMI and Normalized PMI equations: P(x, y) (1) PMI(x, y) = log P(x)P(y) NPMI(x, y) =
PMI (x, y) h(x, y)
(2)
where, h (x, y) is the joint self – information, estimated as -log2 P (x, y). For defect classification on the generated semantic dataset on software defects, we have used a modern recurrent neural network adaption, the GRU model. A GRU model differs from a regular RNN by the presence of the support gating of the hidden states.
168
N. Manoj and G. Deepak
The model essentially consists of two gates, one is the update gate, another one is the reset gate. The reset gate determines how much of the previous information is mandatory to preserve for the context of the phrase, and the update gate is responsible to determine if the new state is similar to the older state, if it is similar, then discard the old state and update to the new one. If, t ∈ Rn × d where Xt is the minibatch, n is the number of examples, d is the number of inputs, t − 1 ∈ Rn × h is the hidden state at the previous time step, then, the reset gate equation and the update gate equation is given by Equations (3), (4), Rt = σ (XtWxr + Ht − 1Whr + br),
(3)
Zt = σ (XtWxr + Ht − 1Whz + bz),
(4)
where, Wxr, Whr, Wxz, Whz, are weight parameters, and br and bzare biases. When we incorporate the reset gate with the latency state of a recurrent neural network, we obtain a candidate hidden state. After we add the update gating effect to the candidate hidden state, we obtain the final GRU equation. The candidate hidden state equation is given by Eq. (5). ˜ = tanh(XtWxh + (Rt Ht − 1)Wh + bh), Ht
(5)
where Wxh and Whh are weight parameters, bh is the bias and the operator is the element wise product operator. Tanh activation is done to keep the candidate hidden state values in the interval (−1, 1). The final GRU equation is given by Eq. 6. ˜ Ht = Zt Ht − 1 + (1 − Zt) Ht.
(6)
whenever Zt is close to 1 we retain the old state, but if we have Zt close to 0, we update to the new hidden state. This architecture not only helps to retain early observations in a sequence, but also to update the recent observations if required. This is a very suitable choice for our purpose. After the training of the GRU model on the enriched dataset, we obtain recommendations on the class of errors and, their possible cause and the ways to debug them.
4 Implementation and Performance Evaluation The proposed model architecture was built using python programming language. The defect messages are preprocessed using robust python libraries and tags are generated. The Bug-Code (BuCo) Analyzer et al. [8], was used to collect the datasets for our detection model. An upper ontology of software defects is built using Web Protégé Software. Using the upper ontology, the detailed instances are crawled using the beautiful soup crawler. A dynamic ontology is then built with the help of Onto Collab tool, Pushpa et al. [9]. This dynamic ontology is used as an auxiliary knowledge source to generate semantic similarity between the defect features using Regex and NPMI. The obtained features are then passed as an input to a deep learning recurrent model, the GRU model. Deep learning enables automatic feature selection for a robust dealing with the myriad of
SDPO: An Approach Towards Software Defect Prediction
169
software messages that tend to generate. Test results of the model has shown a significant improvement in performance and recommendations of software defect types. The proposed hybrid classifier is tested on 5 metrics – Precision, Recall, Accuracy, FNR, F- measure and compared with RSDC, LLE, SVM, MIFI, Random Forest as baseline models for defect classification. Table 1. Comparison of Performance of different baseline models to proposed model. Search Technique
Average Precision % Average Accuracy % F-Measure FNR Recall %
LDA topic model, Ping 80.27 clasai et al. [11]
84.71
79.97
82.43
15.29
LLE and SVM, Shan et al. [5]
78.45
81.17
88.72
79.98
18.83
MIFI, Liu Lu et al. [6]
79.82
82.87
81.33
81.03
15.22
Random Forest Ibrahim 82.72 et al. [7]
84.78
83.41
83.96
16.59
Proposed hybrid defect classifier
96.24
95.74
95.04
4.29
93.87
Precision% = Recall% =
True number of Positives True number of Positives + False number of Positives
True number of Positives True number of Positives + False number of Negatives Accuracy% = F − Measure% =
Precision + Recall 2 2(Precision × Recall) (Precision + Recall)
False Negative Rate = 1−Recall
(7) (8) (9) (10) (11)
Equations (7), (8), (9), (10), (11) show the selected metrics for the comparison for the proposed hybrid model with the other baseline approaches. Our metric of choice for comparison is the False Negative Rate metric. A graph comparing the FNR metric for the models is depicted in Fig. 2. Table 1 shows a comparison of different metrics between the proposed hybrid model and the other baseline approaches to software defect classification. As observed from the Table 1, the proposed hybrid model produces scores of 93.8%, 96.24%, 95.74%, 95.04%, 4.26%, as the Precision, Recall, Accuracy, F-measure and False Negative rate metrics respectively. The metrics when recorded for LDA [11], produce 80.27%, 84.71%,
170
N. Manoj and G. Deepak
79.97%, 82.43%, 15.29%, respectively, for LLE and SVM [5], they produce 78.45%, 81.17%, 88.72%, 79.98%, 18.83%, respectively, and similarly for Random Forest [7], we get 82.72%, 84.78%, 83.96%, 16.59%, for MIFI [6] we get 79.82%, 82.27%, 81.33%, 79.78%, 18.83%. Hence, we can infer that the proposed model produces significantly better results when compared by each metric individually. This boost in accuracy can be attributed to the novel usage of ontologies as an auxiliary knowledge source in the field of Software Defect Classification, for generating semantic similarities between the features of every software defect. The GRU model used for classification is an effective solution to retain important information about the software defects and use it for classification. Data redundancy techniques proposed by LLE and SVM models work efficiently if the number of features are limited, but it becomes a problem when the number of features increase, the process becomes very computationally expensive. Fuzzy integrals accomplish a significant task by correlating the features of different defects, but then since they are approximations, hence they become a disadvantage, as seen in the metrics. Random forests along with metaheuristic algorithms works well with highly discreet categories of defects, but if the category labels overlap, they cannot produce accurate results. The various baseline approaches deal with metrics recorded about the defects, their features etc. [NASA dataset], whereas our model proposes a novel idea of integrating the use of ontologies to use the defect message very efficiently, finding semantic similarities from the messages as enriched features used by the GRU model to make predictions. Hence our model greatly improves the classification process.
Fig. 2. Comparison of FNR measure of different baseline models to proposed model.
Figure 3 shows the number of feature recommendations of the software defects verses the percentage precision of the models. Comparing the proposed hybrid model
SDPO: An Approach Towards Software Defect Prediction
171
Fig. 3. Number of feature recommendations verses precision percentage of different models
with the baseline approaches we can infer that our model makes a significant jump in precision irrespective of the number of features generated out of the defect. The hybrid model’s contributing features of semantic analysis of defects have been proven as major advantage over other models for software defect classification.
5 Conclusion Software defects tend to arise in different situations and they belong to a variety of overlapping categories. Predicting the cause and type of defect can save a lot of computational expenses and time. Until now, defect classifications are being done using the raw metrics recorded by various bug trackers and software as an input to a myriad of models to best predict the type. Our model uses a novel idea of incorporating the use of ontologies along with metrics to prepare an auxiliary knowledge source to get semantic similarities between the features of defects, thereby exceeding the previous models by a great margin of accuracy. As the complexity of software increase the diversity of the type pf errors, faults and defects become highly complex and noisy. Instead of dealing with the increasing number of features discreetly which can increase data redundancy, it is useful to correlate features by semantic similarity and do an efficient feature selection to generate the highest accuracy. Our model implements this idea to generate state-ofart results. A detailed ontology is useful in a software system to better communicate the software defects and solutions between experienced developers and new developers, which will contribute in saving time and increase collaboration process.
172
N. Manoj and G. Deepak
References 1. Duarte, B.B., Falbo, R.A., Guizzardi, G., Guizzardi, R.S.S., Souza, V.E.S.: Towards an ontology of software defects, errors and failures. In: Trujillo, J.C., Davis, K.C., Du, X., Li, Z., Ling, T.W., Li, G., Lee, M.L. (eds.) ER 2018. LNCS, vol. 11157, pp. 349–362. Springer, Cham (2018). https://doi.org/10.1007/978-3-030-00847-5_25 2. Iliev, M., Karasneh, B., Chaudron, M.R.V., Essenius, E.: Automated prediction of defect severity based on codifying design knowledge using ontologies. In: 2012 First International Workshop on Realizing AI Synergies in Software Engineering (RAISE), pp. 7–11. IEEE (2012) 3. Hu, X., Liu, J.: Orthogonal Defect Classification-based Ontology Construction and Application of Software-hardware Integrated Error Pattern of Software-intensive Systems. In: 2021 23rd International Conference on Advanced Communication Technology (ICACT), pp. 1286– 1298. IEEE (2021) 4. Liang, H., Yue, Y., Jiang, L., Xie, Z.: Seml: A semantic LSTM model for software defect prediction. IEEE Access 7, 83812–83824 (2019) 5. Shan, C., Chen, B., Hu, C., Xue, J., Li, N.: Software defect prediction model based on LLE and SVM, pp. 24–24 (2014) 6. Bedi, P., Goyal, S.B., Rajawat, A.S., Shaw, R.N., Ghosh, A.: A framework for personalizing atypical web search sessions with concept-based user profiles using selective machine learning techniques. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 279–291. Springer, Singapore (2022). https:// doi.org/10.1007/978-981-16-2164-2_23 7. Ibrahim, D.R., Ghnemat, R., Hudaib, A.: Software defect prediction using feature selection and random forest algorithm. In: 2017 International Conference on New Trends in Computing Sciences (ICTCS), pp. 252–257. IEEE (2017) 8. Mauša, G., Grbac, T.G., DalbeloBaši´c, B.: Software defect prediction with bug-code analyzer-a data collection tool demo. In: 2014 22nd International Conference on Software, Telecommunications and Computer Networks (SoftCOM), pp. 425–426. IEEE (2014) 9. Pushpa, C.N., Deepak, G., Thriveni, J., Venugopal, K.R.: Onto Collab: strategic review oriented collaborative knowledge modeling using ontologies. In: 2015 Seventh International Conference on Advanced Computing (ICoAC), pp. 1–7. IEEE (2015) 10. Gao, J., Zhang, L., Zhao, F., Zhai, Y.: Research on software defect classification. In: 2019 IEEE 3rd Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), pp. 748–754. IEEE (2019) 11. Pingclasai, N., Hata, H., Matsumoto, K.-I.: Classifying bug reports to bugs and other requests using topic modeling. In: 2013 20th Asia-Pacific Software Engineering Conference (APSEC), vol. 2, pp. 13–18. IEEE (2013)
Ubiquitous Control Structure and It’s Comprehensive Application Subject to Controller Performance Ravishankar P. Desai(B) and Narayan S. Manjarekar Department of Electrical and Electronics Engineering, BITS Pilani, K K Birla Goa Campus, NH 17B, Zuarinagar, Sancoale 403726, Goa, India {p20180019,narayan}@goa.bits-pilani.ac.in
Abstract. The basics of the feedback control structure and its applicability subject to modelling, control, disturbance rejection, noise suppression techniques are addressed in this paper through a motion control problem of a linearized decoupled speed subsystem model of an autonomous underwater vehicle (AUV). AUV is chosen because vehicle dynamics are highly coupled, time-varying, nonlinear and in reality, its operating environment is highly uncertain in nature. The design of two control schemes is proposed to achieve the control objectives and controller performance. A superiority of proposed control performance is presented and verified through simulation results.
Keywords: Control structure Modelling · Linear controller
1
· Autonomous underwater vehicle ·
Introduction
Control system play an inevitable role in day to day life of technology, in short, we would say technology without a control system is impractical. A control system may define as “An assembly of devices and components connected in such a way that to command, regulate itself or some other system”. Hence, a technical person must be familiar with control system analysis and design. A basic building block of any electromechanical devices are used in industry, military, navy, banking, home appliance etc. consists of a simple to a complex closedloop structure. Here are few highly recommended reference books in the control community are chosen to summarize the basic building block of an ubiquitous control structure. From [1–7], covered five broad areas of the linear control system such as linear algebra, system modelling, analysis and design, state-space analysis and design, some advanced/modern analog and digital control topics. Do all these references will give a strength about how to learn? Why learn?, and what is the need? of control system structure and its importance in linear control system applications. Answer is big “Yes”. A comprehensive information about the system is essential at a given instant of time in practice. To know c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 173–193, 2022. https://doi.org/10.1007/978-981-19-1677-9_16
174
R. P. Desai and N. S. Manjarekar
the details and insight of the system, the complete information is presented in [8]. A system modelling and analysis may limit the controller performance set through by knowing the fundamental specification and analytical analysis of the system. The details of performance limitation in the linear control system presented in [9]. In [10], the basic learning material of classical and modern control technique for linear controller design with its application presented systematically in it. A control problem is formulated precisely under “robust performance problem” in [11] which gives an insight into how modelling uncertainty will affect the performance of the control system. The details of modelling, analysis and design (MAD) with linear/modern controller design technique covered very nicely through the simulation tool MATLAB in [12]. A comprehensive application considered in this paper is AUV which is operated beneath the ocean. Furthermore, we need to analyse the operating environment of the AUV is necessary. The ocean environment is offensive, communication is difficult, and external disturbances such as ocean waves, currents, and tides can contrary affect the AUV dynamics under the mission. AUV modelling and control synthesis is the most challenging and crucial problems, due to vehicle dynamics are inherently nonlinear, time-varying hydrodynamic parameter, uncertain environment, uncertain external disturbances, inadequate knowledge of hydrodynamic coupling, and payload variation of the vehicle. So, the selection of the control problem is made through the above reason, which fulfils the feedback control structure. The complete details of MAD along with linear and nonlinear control technique of an ocean vehicle is presented in [13] and a lightly interacting subsystem of an AUV with robust controller design is presented in [14]. Obtain a hydrodynamic parameter of an AUV, is the most crucial part in order to control the vehicle during operation. The basic idea of parameter estimation and system identification for underwater vehicle modelling presented through the auto-tuning controller in [15]. Also, vehicle body, size, shape, deployment and operational environment etc. have important significance while designing a control system. The design procedure of the control system is conditionally precise but not accurate considering the statics and dynamics of the vehicle. In this work, a Proportional Integral (PI)-controller and Resonant (R)controller design scheme is proposed to control the linearized decoupled speed subsystem model of an AUV. The intention behind this work is to summarize the details of the ubiquitous control structure with its comprehensive application to the user in classical way. Here, a controller is designed to achieve the control objective precisely in the presence of parametric uncertainties, disturbances and noisy environment for sinusoidal reference input. A summary of the paper as follows: Sect. 2 gives the details of the control system structure. In Sect. 3, the basics of AUV and its modelling presented briefly. The controller design and implementation steps presented in Sect. 4. Section 5 shows the efficacy of the proposed controller via simulation result and, lastly, concludes the paper.
Ubiquitous Control Structure and It’s Comprehensive Application
2
175
Basics of Control Structure
To realize the fundamental structures of control, two control structures are depicted in Fig. 1 called the open-loop (OL) control system and Fig. 2 the closedloop (CL) control system. A CL structure possesses sensitivity and disturbance rejection properties that an OL structure can’t. A CL structure deepens the sensitivity with respect to variations or uncertainties in the element which is located in the system forward path. A disturbance rejection refers to the fact that a CL can reduce or eliminate the effects of undesirable disturbances occurring within the CL [1]. Hence, principally for these properties, the CL structure is ubiquitous. From Fig. 1, the output, an error which is the difference between output and input of the system, and transfer function (T.F.) of the OL system is written in Eq. (1), (2) and (3) respectively. Exogoneous Disturbance D(s) Input I(s)
Controller U(s) + C(s)
∑
+
Output O(s)
Model M(s)
Fig. 1. Open loop structure
Exogoneous Disturbance D(s) Input I(s) +
∑ -
E(s)
Controller U(s) + C(s)
∑
+
Output O(s)
Model M(s) +
∑
Measurement Noise N(s) +
Fig. 2. Close loop structure
Ool (s) = M (s)C(s)I(s) + M (s)D(s)
(1)
E(s) = I(s) − O(s) = [1 − M (s)C(s)]I(s) − M (s)D(s)
(2)
T.F. =
I(s) = M (s)C(s) O(s)
(3)
A CL system has unity feedback with the input, I, called as reference/setpoint, the model exogenous disturbance, D, the measurement noise, N, to ignore by the controller and output, O, to track the reference/setpoint with the help of controller design. The typical goal of this structure is to achieve all the time
176
R. P. Desai and N. S. Manjarekar
stable, precise tracking, the capability to reject disturbance and robustness to lumped uncertainty [2]. The output of a CL system and controller is given by Eq. (4)–(5), and the difference between output and input of the system called error is given by Eq. (6). Ocl (s) = U (s) =
M (s)D(s) M (s)C(s)N (s) M (s)C(s)I(s) + − 1 + M (s)C(s) 1 + M (s)C(s) 1 + M (s)C(s)
(4)
C(s)I(s) M (s)C(s)D(s) M (s)N (s) − − 1 + M (s)C(s) 1 + M (s)C(s) 1 + M (s)C(s)
(5)
E(s) = (I(s) − O(s)) =
I(s) M (s)D(s) M (s)C(s)N (s) − − 1 + M (s)C(s) 1 + M (s)C(s) 1 + M (s)C(s)
(6)
The T.F. of the CL system is given by Eq. (7), T.F. =
M (s)C(s) I(s) = O(s) 1 + M (s)C(s)
(7)
The sensitivity and complementary sensitivity T.F. from Fig. 2 is S(s) =
I(s) 1 = O(s) 1 + M (s)C(s)
(8)
T (s) =
I(s) M (s)C(s) = O(s) 1 + M (s)C(s)
(9)
The S(s) + T (s) = 1 equality holds for all frequency. In reality [1], reference input/set point and exogenous disturbances are lowfrequency signal whereas a high-frequency signal is the measurement noise. Thus, we can sum up S(s) and T(s) by keeping S(s) is small which is suitable for disturbance rejection and tracking in the low-frequency range while by keeping T(s) is small which is suitable for noise suppression and control energy reduction in the high-frequency range. Lastly, by keeping M(s)C(s) is small, which is suitable for control energy constraints and it is given as C(s)S(s) =
T (s) C(s) = 1 + M (s)C(s) M (s)
(10)
A feedback control system, defined as a physical system that keeps the commanded account between the expected output and the predefined input by comparing them and taking difference as control action [3]. A CL system provides more advantages over OL, stated as high accuracy, less sensitivity to disturbance, noise and environmental change. Also, more convenient and flexible for steady-state error and transient response with the help of small adjustment in gain and/or controller redesigning. On the other hand, the design structure of CL systems is complex and expensive. Thus, there is a trade-off between the simplicity and inexpensive of an OL system, over the accuracy and expensive
Ubiquitous Control Structure and It’s Comprehensive Application
177
of a CL system [4]. In the CL system, a fundamental trade-off exists between tracking accuracy and disturbances. In frequency domain the tracking signals are quite different from standard test signals, those will appear in system output that must be rejected. This trade-off can be negotiated by choosing the proper frequency band strategically. 2.1
Model
A set of mathematical equations describes the dynamic characteristics of a system when the external perturbing disturbance study is absent called the mathematical model of the system. The development of the model needs to introduce spatial variables (description of geometry), time, dimensional units, assumptions and conditions, formulation of fundamental law (physical, empirical etc.) and verification of the model. Modelling problem is inevitable for the success of the controller design, and it fully depends on a proper modelling of the physical system. A comprehensive understanding, study of operating range and environment is necessary to build the desirable model of the physical system [5]. The first principle of physics gives insight and direction to summarize the knowledge about the physical system through a mathematical model, which is ubiquitous in science and engineering. The physical system is a hierarchy of model which is ranging from simple to complex structure of the model. For the exploratory purpose, a simplex model is preferred to get features of the system behaviour while a complex model for insight performance of the system. The first principle of physics sometimes faces difficulties to get a model of disturbances acting on the physical system. An unmodelled dynamics that is unable to model by using physical laws that can be obtained through experimentally, called a system identification approach. With the help of the balanced equation, model building is easy, but sometimes it is difficult to interpret the physical parameter hence it consumes time. The design method categorized into an internal model in which disturbance act as white noise to the dynamic system and an external model in which disturbance act as covariance function and spectral densities to the dynamic system [6]. An experimental approach is often preferred to get a disturbance model. In the availability of analysis tool and physical situations, disturbances are modelled. Such as a certain class of all signal bounded with square integral and a disturbance signal such as a step or ramp come through as an input at a specific point of the system; referred to as variation in the real model. Measurement noise can be modelled as a random process with certain statistical characteristics are known to us. A permanent uncertainty always exists even though we are very much sure about the applicability of physical laws and identification parameter because a mathematical model is nothing but an approximation of a practical system. Model inaccuracies such as parametric uncertainties, unmodelled dynamics and external disturbances called system behaviour represented in a simplified manner. The difference between proper models and real/actual model referred to as
178
R. P. Desai and N. S. Manjarekar
uncertainty and its representation referred to as a mechanism used to express errors in the physical system [11]. A few of model uncertainty origins [7] are enlisted as 1. An approximately known parameter that may vary due to changes in operating condition or non-linearities in the physical model 2. A reduced-order model by neglecting time delays and higher-order terms 3. Deficient measurement devices 4. Design implementation issues. The model uncertainty can be modelled [7] by considering the origin as 1. parametric or structured uncertainty - the structure of the model and its order is known to us, but tolerances in a concrete physical model. 2. A neglected or unmodelled dynamic uncertainty - a lack of knowledge to understand the physical model. 3. Lumped or unstructured uncertainty - the exact nature of uncertainty is unknown, except that they are bounded. It’s a combination of parametric and neglected uncertainty into a single term. Often, many system models are built through variation in complexity and fidelity. A simplification in the design and analysis of a simple model might catch with some basic characteristics and features of commands. Such as disturbance, noise at the risk of inaccuracy, whereas a very accurately defined and described complex model can create design and analysis task complicated [9]. A precise model of a complex structure of a physical model requires an experimental touch to build the mathematical model of the physical system. Once, we get a precise model then the next part in the modelling is the analysis i.e. quantitative and/qualitative analysis. A computer is preferred for quantitative analysis, here analysis is done through a command input signal and its exact response to an output of the system model. On another side, system properties (stability, controllability, and observability) in the qualitative analysis will play a vital role because controller design is done through it [5]. 2.2
Controller
Two forms of control structure are defined, OL and CL, based on the acknowledgement used in control. In OL control structure, the system output is independent of input, it’s non-causal hence no corrective action takes place whereas, in CL control structure, the system is dependent on input, its causal hence corrective action take place. There is a trade-off between tracking (to follow the change in reference signal) and regulation (to maintain fixed output in the presence of disturbance). So, the control system designer always has a compromise between tracking and regulation (servo and regulator problem). A controller is defined as the internal or external element of the system which controls the physical system. Basically, the design of the control system based on modelling, analysis and design. The implementation of controller in the feedback structure along with
Ubiquitous Control Structure and It’s Comprehensive Application
179
the physical system which is based on a mathematical model, predefined design goal i.e. development of performance specification and analysis of mathematical model through properties and design of control law. In order to estimate the effectiveness of the controller model with the mathematical model of the physical system, it is necessary to check viability through software tool such as MATLAB, Python, LabVIEW etc. before proceeding to the implementation [12]. The design of control law categories into a linear and nonlinear controller. Moving to the conventional control system (signal based control) uses Laplace transforms/Ztransforms and modern/advanced control system uses ODEs (information-based control) even though the basic idea of a control system is ubiquitous. The ODEs developed through modelling are often nonlinear and dominate significantly than linear equations. Because linear models are always a compromised one that give adequate information about the system for the purposes of control design [1]. In reality, all physical systems are nonlinear in nature (large signals). A mathematical tool to estimate and control linear systems are available, accessible and understood easily. Hence, often nonlinear systems are approximated in such a way that the linear systems (small signals) equal to the nonlinear system to satisfy performance requirement in practice. A Linear system is a valuable tool to deal with the nonlinear system, even though it does not exist in reality. Digital computer popularity enhances applicability in continuous-time and discrete-time for linear and nonlinear system. Because estimation and simulation algorithm will work in discrete-time and physical systems work in continuous time [8]. A controller designer must take into account not only the design and concept of the tool, the availability of controller devices along with its capabilities and limitation too. However, small signals are active in the presence of non-linearities, or it can cause a device to saturate. The system behaviour taken into account for nonlinear effects because every nonlinear system in many ways is unique. So nonlinear control design uses a number of approaches such as adequate smallsignal approximation, a heuristic approach with non-linearity to be a varying gain, non-linearity is non-causal and non-linearity has dynamics or causality. In short, a non-linearity is always exist in any physical system at some level of signal strength or at all level of signal strength [1]. We consider mainly two types of problems: 1. Analysis problems- Given a controller, determine if the controlled signal, whether satisfies the desired properties under model uncertainties, disturbances, and noise. 2. Synthesis problems- Controller design must satisfy desired properties under model uncertainties, disturbances, and noise [11]. A classical control theory applies to simple cases where design specification are less demanding, and it refers mainly to a single-input single-output system and time-invariant. The design method of classical control theory usually graphical (root locus plot, bode diagram, Nyquist diagram, polar diagram etc.) and it does not require an advanced mathematical tool. In advance/modern control theory applies to complex cases where design specification are highly demanding, and it mainly refers to multiple-input multiple-output and time-varying system and
180
R. P. Desai and N. S. Manjarekar
its design method are based on analytical approach with advanced mathematical tool. The complexity of controller implementation depends on the nature of the system under control and design requirement i.e. as design requirement increases the complexity of controller implementation also increases [7]. The actuator response depends on the design of a control algorithm or a signal processor i.e. control law. In the design of the classical controller, need to specify parameter and these to be adjusted empirically in real-time. The classical controller varies widely in terms of effectiveness and complexity. Simple controllers include P, PI, PD, PID, and more sophisticated controllers include LQR, estimated state controller, LQG etc. Controller design goals consist of qualitative and quantitative performance, constraints on control law and robustness requirements. An analytical tool is used to make the design goal precise (a small error signal and adequate actuator signal) which can be described as norms of signals and systems. A CL system design specifications include response towards disturbances that may act on it, a robustness that limits the sensitivity, and noise that limit the complementary sensitivity. A framework for controller design consist, norms of the signal and system, geometry, realizability and stability, design specification, analytic solution method, a controller (analysis and optimization) and solution of the controller design problem. For control system design, we must: 1. Comprehensive study of a physical system (to be controlled) is must as per control target, design success, expected outcome, and acceptable range point of view. 2. Analyse the control structure and select the best one for control system design. 3. The control parameters must be fine-tuned. Then, implement the control system, test it and validate the control solution. 2.3
Disturbance
A disturbance signal creates an adverse effect on control system output. The internal disturbances are generated within the system, while external disturbance treated as input to the system is generated outside the system. Disturbances are signals which enter in the feedback system at the input or output of the physical model and cause an undesirable signal at the system output [11]. The effects of the disturbances on the output are reduced when the feedback is negative and increased when the feedback is positive. Always, there is a trade-off between disturbance rejection and noise reduction because: 1. larger feedback bandwidth—larger noise—smaller disturbances 2. smaller feedback bandwidth—smaller noise—larger disturbances
Ubiquitous Control Structure and It’s Comprehensive Application
181
The disturbances effect on: 1. As increases mean square input will reduce the effect on mean square tracking error 2. As reduces mean square input will be a full effect on mean square tracking error Disturbances enter in to the system with random frequency characteristics ranging from constant to white noise. The most influential higher frequency random disturbances get effects while selecting the sample rate. Disturbances are variables that can be deterministic (component) or stochastic (device) and can not be considered while modelling. The random (analog) nature of disturbance adversely affect the performance of tracking and regulating the control system [8]. In control system, a widely exist parameter is disturbance and to reject before it gets an adverse effect on control system performance is one of the prime objectives of the controller design. One of the promising solution to deal with disturbance is a robust controller design. 2.4
Noise
The sensor is a device whose output inevitably contains sensor noise, which introduces conflicts among the system disturbances and sensor noise in order to reduce the efforts on errors [2]. Its effect of observation noise causes an increase in both mean square input and tracking error. In feedback, one of the inevitable aspect is that the system input is influenced by the measurement on which the feedback acts. It implies that erroneous activity produces by feedback whenever measurement noise has an error. The deviation always exists in measurement because none of any device is perfectly accurate. The system behaviour is undesirable in response to noise because the measured signal is the combination of desired plus noise. In order to keep the necessary feedback information into measurement, we have to have a filter to suppress the noise [8]. In the presence of a noisy environment the filter, estimator which is a different form of the feedback path is used to take out possible information from the control system. A distinction between internal and external signal dynamics can be accomplished using the filter. To know the necessary signal for feedback, sometimes, extracted measured signal with internal component signal gives information about the control system which make signal analysis easier [9].
3
Basics of AUV
An AUV is a robotic device i.e. driven through the water by the propulsion system, controlled and piloted by an onboard computer. Earth covered 70% of the surface with water which is unexplored. These marine habitats need to be explored because they affect human directly or indirectly. The underwater vehicle
182
R. P. Desai and N. S. Manjarekar
serves the purpose of better understanding of ocean environment and protects from pollution under ocean resources. AUV has six degrees of freedom (6-DOF) and its rigid body movement i.e. manoeuvrable in three-dimensional space. The kinematic model of a vehicle is the mathematical adherence between the body-fixed and earth-fixed frame. The vehicle links determine the position (P), velocity (V) and acceleration (A) of various parts of the vehicle w.r.t. reference frame. A dynamic model of the vehicle adhere forces and moments with the position and velocity. A rigid body of an AUV is the transformation between the body-fixed frame (X0 − Y0 − Z0 ) and the earth-fixed frame (XE − YE − ZE ) as shown in Fig. 3. A mathematical OE ϕ XE
Rudder θ
ѱ
Fin YE
ZE
rg
O
Fin Rudder
Xo Surge: u, X Roll : p, K Zo Heave: w, Z Yaw : r, N
Yo Sway: v, Y Pitch : q, M
Fig. 3. AUV schematics
model of an underwater vehicle is usually represented by a set of ODEs that describes the motions in 6-DOF (Surge, Sway, Heave, Roll, Pitch, Yaw). Table 1 show the 6-DOF motion components of AUV dynamics. A coupled 6-DOF nonlinear equation of motion represented as [13]: η˙ = J(η)ν Mv˙ + C(v)v + D(v)v + G(η) = τ + τd
(11) (12)
Where M, C, D, and G are the inertial mass, Coriolis and Centripetal terms, damping, gravitational/buoyancy forces and moments respectively. τ is the control input and τd = τwave + τwind + τcurrent is the ocean disturbances. A 6-DOF model of an AUV involves a highly nonlinear and coupled equation that leads to a high level of complexity in designing control systems. So, a 6-DOF
Ubiquitous Control Structure and It’s Comprehensive Application
183
Table 1. 6-DOF motion components of an AUV dynamics DOF Description
EF-FM BF-LAV BF-PEA
1
Surge, (motion in X-direction)
X
u
x
2
Sway, (motion in Y-direction)
Y
v
y
3
Heave, (motion in Z-direction)
Z
w
z
4
Roll, (rotation about the X-axis) K
p
φ
5
Pitch (rotation about the Y-axis) M
q
θ
6 Yaw, (rotation about the Z-axis) N r ψ *EF-FM = Earth Fixed Forces and Moments; *BF-LAV = Body Fixed Linear and Angular Velocity; *BF-PEA = Body Fixed Position and Euler angle Table 2. AUV subsystems Subsystem
Description
Control inputs
Speed
Surge linear velocity (u)
Rudder rate (η)
Steering
Sway linear velocity (v) Rudder deflection (δr ) Yaw angular velocity (r) Heading euler angle (ψ)
Diving
Heave linear velocity (w) Stern deflection (δs ) Pitch angular rate (q) Pitch euler angle (θ)
model of an AUV can be categorized according to lightly interacting or noninteracting subsystems. This can be commonly categorized according to three subsystems; speed, steering, and depth as shown below [14] (Table 2): 1. Speed subsystem - Using the rudder, controls the heading position 2. Steering subsystem - Using the elevator, controls the depth and pitch i.e. responsible for heading errors 3. Diving subsystem - Using the propeller, controls the vehicle speed i.e. responsible for depth errors. The dynamics of AUV get affected by fluid force because AUV experiences lift and drag forces during motion, which causes friction between the surrounding fluid and vehicle body. Hence, vehicle body and shape have important significance under these forces. A well known computational fluid dynamics (CFD) tool is used to obtain the hydrodynamic parameter of a vehicle which will help to reduce skin friction, lift and drag coefficients. A vehicle propulsion system plays an important role to travel the vehicle in an underwater environment. During motion, AUV experiences hydrostatic and hydrodynamic pressure due to water head and movement, respectively. As we increase the depth corresponding pressure will build into the vehicle body, causes the vehicle body to get deform or damaged. These categories into
184
R. P. Desai and N. S. Manjarekar
coastal and deep ocean AUV due to immersion in depth. A coupled, time-varying, nonlinear AUV dynamics and a highly uncertain ocean environment of an AUV challenge the control system during design. In order to regulate and stabilize the vehicle position and velocity, a control system need to regulate the output of the actuator, along with it has self tuning ability to deal the ocean environment. An example of an AUV CL structure, a linearized decoupled speed subsystem model is chosen as a control problem in this paper. In order to generate the lift force on control surfaces, it is inevitable to maintain the relative speed of the submerged vehicle. A submerged vehicle speed can be controlled using a propeller, which acts as a control signal. By assuming movement of AUV only in one degree of freedom (decoupled motion) at a time, considering negligible cross term except for dependency on the dynamics of pitch and yaw for surge velocity. The navigation model for the AUV speed subsystem is expressed as, (MRB + MA )v˙ + (Dl + Dn )v = T
(13)
Where MRB and MA is the matrix of rigid body inertia and added mass respectively, Dl and Dn is the linear and nonlinear damping matrix respectively and T is the force of the thruster. Then the Eq. (13) is written as, (m − Xu˙ )v˙ − (Xu + X|u|u |u|)v = P rop
(14)
The resulting linearized surge dynamic model presents surge velocity is the output and propeller current is the input forms [13,14], u(s) = Gu (s) = ηref (s)
4 4.1
1 b (m − Xu˙ ) = 2|u|X|u|u + Xu s+a s− (m − Xu˙ )
(15)
Design of Controller PI-Controller
A most popular and widely used composite control mode i.e. proportional mode and an integral mode called PI controller is used to design the surge speed control of an AUV. The composite control action improves the steady-state accuracy. In the presence of non-zero error, P-controller adds the correction magnitude while I-controller starts to increase/decrease the integral term from its initial value depending upon the action we have chosen i.e., direct/reverse. On the other hand, in the presence of zero error, PI-controller output is fixed at that value i.e. bias value. Such a kind of composite action is applicable where system load changes frequently or largely [6]. A Mathematical expression for the ideal form of PI-controller is t CP I (t) = Kp e(t) + Kp Ki e(t)d(t) + b(0) (16) 0
Ubiquitous Control Structure and It’s Comprehensive Application
185
where, Kp is proportional controller gain, Ki is integral controller gain which is inverse of integral time constant (1/τi ), and b(0) is the bias value of output at t = 0. The Laplace transform of Eq. (16) is Kp e(s) k1 s + k0 τ = CP I (s) = Kp e(s) + i s s
(17)
Where, Kp = k1 and τi = k1 /k0 . Figure 4 shows that the speed control structure of an AUV. The closed loop transfer function between the reference and output signal is obtained from Eq. (15) and Eq. (17) as u(s) G(s)CP I (s) b(k1 s + k0 ) = = 2 uref (s) 1 + G(s)CP I (s) s + (a + bk1 )s + bk0
(18)
Exogoneous Disturbance d(s) Input uref (s) +
∑ -
e(s)
Controller C(s)
nref(s) +
∑
+
Output u(s)
Model G(s) +
∑
Measurement Noise n(s) +
Fig. 4. Close loop structure of the speed controller of an AUV
For second order polynomial for Eq. (18), a standard second order characteristics polynomial for closed loop system is written as dcl (s) = s2 + (2ζωn )s + ωn2
(19)
Where ζ is the damping coefficient, ωn is the natural frequency. The ζ and ωn are the design parameter to be selected by designer as control objective for desired system performance specification. Comparing the polynomial of Eq. (18) and Eq. (19), (20) s2 : 1 = 1; s1 : a + bk1 = 2ζωn ; s0 : bk0 = ωn2 solving Eq. (20) gives, 2ζωn − a ω2 ; k0 = n (21) b b then the PI-controller parameter is obtained from Eq. (17). The stability of a closed-loop system depends on the gain of the PI-controller. The Kp should be positive and τi should be ∞ > τi > 0. k1 =
186
4.2
R. P. Desai and N. S. Manjarekar
R-Controller
A sinusoidal reference signal tracking and/or sinusoidal disturbance signal rejection is often required in many control applications such as aerospace, robotics, power electronics, communication etc. For sinusoidal reference signal, the parameters of u(t) = A0 sin(ω0 t) is known whereas for sinusoidal disturbance signal d(t) = d0 sin(ω0 t), the frequency is known, but magnitude is unknown. The resonant controller provides sufficient gain at a certain resonant frequency and other frequency almost gain does not exist [10]. Using Eq. (15), the resonant control structure is chosen and written in Laplace transform form as CR (s) =
k4 s2 + k3 s + k2 s2 + ω02
(22)
Where k4 , k3 , k2 are the controller gain and ω0 is the tracking and/or rejecting frequency. The above Eq. (22) forms proportional control action with three unknown coefficients called controller gain. The closed-loop transfer function between the reference and output signal is obtained from Eq. (15) and Eq. (22) as b(k4 s2 + k3 s + k2 ) u(s) = 3 (23) uref (s) s + (a + bk4 )s2 + (ω02 + bk3 )s + (aω02 + bk2 ) For third order polynomial for Eq. (23), a third order characteristics polynomial for closed loop sytem is written as dcl (s) = (s2 + 2ζωn s + ωn2 )(s + λ)
(24)
Where ζ is the damping coefficient, ωn is the natural frequency and λ > ωn > 0. The ζ, ωn and λ are the design parameter to be selected by the designer as control objective for desired system performance specification. Comparing the polynomial of Eq. (23) and Eq. (24), ⎧ 3 s :1=1 ⎪ ⎪ ⎨ 2 s : a + bk4 = 2ζωn + λ (25) s1 : ω02 + bk3 = ωn2 + 2ζωn λ ⎪ ⎪ ⎩ 0 s : aω02 + bk2 = λωn2 solving Eq. (25) gives, 2ζωn + λ − a ω 2 + 2ζωn λ − ω02 λωn2 − aω02 ; k3 = n ; k2 = (26) b b b The stability of a closed-loop system depends on the gains of the R-controller. The k4 , k3 , k2 , ω0 should be positive. k4 =
5
Simulation Result
In this section, a simulation observation was carried out on the characteristics of input disturbance rejection and the effect of measurement noise for decoupled speed subsystem of an AUV. A computational tool MATLAB Simulink R2019b
Ubiquitous Control Structure and It’s Comprehensive Application
187
(9.7.0.1471314) is used to present the simulation result for the proposed design. From Eq. (15) the transfer function for three different velocities of the speed subsystem model [15] is written as 0.001611 u(s) = uref (s) s + 0.4920 0.001611 u(s) = G(u0 =1.5) (s) = uref (s) s + 0.7380 0.001611 u(s) = G(u0 =2.0) (s) = uref (s) s + 0.9836 G(u0 =1.0) (s) =
(27a) (27b) (27c)
In this paper, Eq. (27b) is chosen and controller is designed. 5.1
Model Analysis: OL
A linearized speed subsystem model represented in transfer function form in Eq. (27) has one stable real pole. Figure 5 shows that the pole-zero plot and openloop sinusoidal response of it. The system poles of Eq. (27) lies on the left half of the s-plane, and it has a large error between the reference signal and output signal.
Fig. 5. Open loop pole-zero and sinusoidal reference tracking response
5.2
Model Analysis: CL
To verify the proposed controller design scheme on Eq. (27b) the design parameters of PI-controller are chosen as ζ = 0.707; ωn = 0.5657 and R-controller are ζ = 0.707; ωn = 0.5657; λ = 1 as control objective of speed subsystem. The controller setting parameters are obtained from Eq. (21) and (26) are tabulated in Table 3.
188
R. P. Desai and N. S. Manjarekar Table 3. Speed subsystem: PI-controller and R-controller values Eq.
u0
k0
k1
ω0 k2
k3
k4
21, 26 1.0 191.2 198.7 0.5 170.00 656.42 659.22
PI-Controller
0 -1 0
50
100
Time [sec]
Reference Signal
0
0 -1 0
50
Time [sec]
100
50
100
Time [sec]
0
-500
0
R-Controller
1
500
500
-500
Surge Velocity [m/s]
Reference Signal
1
Propeller Speed [N]
Propeller Speed [N]
Surge Velocity [m/s]
The simulation response of speed control subsystem of an AUV for close loop system of Eq. (18) and (23) as shown in Fig. 6. It clearly shows that R-controller tracks the sinusoidal reference trajectory, but it took a bit of high speed to achieve it as compare to PI-controller. Figure 7 and 8 shows that sinusoidal reference tracking in the presence of disturbance (small and high magnitude) with random noise. R-controller is best for trajectory tracking, disturbance (small and high magnitude) rejection, and less efficient to suppress the noise with a bit high propeller speed. PI-controller has less tracking capability in the presence of disturbance (small magnitude) while it fail to satisfy the tracking capability in the presence of disturbance (high magnitude), and able to suppress the noise with less propeller speed.
0
50
100
Time [sec]
Fig. 6. Close loop sinusoidal tracking response: PI-Controller (left) and R-Controller (right)
The speed variation occurs during motion because of seawater density, underwater pressure, wave disturbance etc. Hence, need to check the performance of the designed controller at different velocities. The low, medium and high-velocity parameter variations are chosen, and controller performance is tested against the medium velocity subsystem presented in Eq. (27b). A sinusoidal reference tracking response for different velocity settings is shown in Fig 9. In the case of the PI controller, it shows minimum tracking error with minimum propeller speed. Also, it observes that as we increase the vehicle velocity, it will affect the tracking magnitude and able to suppress the noise. In the case of the R-controller,
Ubiquitous Control Structure and It’s Comprehensive Application
189
Fig. 7. Sinusoidal tracking response with disturbance (small magnitude) and noisy environment: PI-Controller (left) and R-Controller (right)
Fig. 8. Sinusoidal tracking response with disturbance (high magnitude) and noisy environment: PI-Controller (left) and R-Controller (right)
it shows that good tracking is achieved with a bit of high propeller speed and there is no effect on magnitude tracking as we increase the vehicle velocity. But it is less efficient to suppress the noise. Figure 10 and 11 shows that sinusoidal reference tracking with the effect of disturbance (low and high magnitude) and random noise. As you can observe that using the R-controller a precise tracking is achieved in the presence of disturbance (low and high magnitude) with a bit of high propeller speed and it is less efficient to suppress the noise. Whereas, PI-controller gives tracking error in Fig. 10 and it fails to achieve the tracking in Fig. 11.
190
R. P. Desai and N. S. Manjarekar
The overall observations From Fig. 6, 7, 8, 9, 10 and 11 concludes that Rcontroller provides good sinusoidal reference tracking in the presence of disturbance (low and high magnitude) and random noise. Regarding the noise issue, can be overcome with good quality of the speed sensor. It also shows that the propeller speed is within range. To summarise the controller performance a qualitative analysis of PI- controller and R-controller tabulated in Table 4 and 5. From the qualitative analysis, it clears that the performance of R-controller is excellent than PI-controller. Table 4. Quantitative analysis of speed subsystem using PI-controller Velocity Input
RMS
ISE
IAE
u0 = 0.5 uref u0 = 1.0 uref u0 = 1.5 uref
0.3332 13.32 35.94 0.4191 21.08 45.32 0.4843 28.15 52.38
u0 = 0.5 uref + d0 + n 0.1947 4.551 19.74 u0 = 1.0 uref + d0 + n 0.2326 6.490 24.00 u0 = 1.5 uref + d0 + n 0.2623 8.258 27.32 u0 = 0.5 uref + d1 + n 0.5099 31.20 54.45 u0 = 1.0 uref + d1 + n 0.6367 48.65 68.48 u0 = 1.5 uref + d1 + n 0.7334 64.55 79.01 uref = sin(0.25t); d0 = 0.5 sin(0.25t); d1 = 2.5 sin(0.25t); n = random noise
Fig. 9. Effect of low to high velocity on sinusoidal tracking response: PI-Controller (left) and R-Controller (right)
Ubiquitous Control Structure and It’s Comprehensive Application
191
Fig. 10. Effect of low to high velocity on sinusoidal tracking response with disturbance (small magnitude) and noisy environment: PI-Controller (left) and R-Controller (right)
Fig. 11. Effect of low to high velocity on sinusoidal tracking response with disturbance (high magnitude) and noisy environment: PI-Controller (left) and R-Controller (right)
5.3
Stability Analysis
In reality, the analysis, implementation and input-output stability is not adequate in CL systems. Since there is unbounded internal signals are present in CL systems are finite with the external input signal. The hardware of the physical system get affected in the presence of unbounded internal signals even though the CL system is input-output stable. The input-output stability is assessed by checking i) there are no zeros on the right-hand side of the s-plane in the transfer function and ii)there are no pole-
192
R. P. Desai and N. S. Manjarekar Table 5. Quantitative analysis of speed subsystem using R-controller Velocity Input
RMS
ISE
u0 = 0.5 uref u0 = 1.0 uref u0 = 1.5 uref
0.0201 0.0486 0.4637 0.0245 0.0749 0.6447 0.0303 0.1103 0.8610
u0 = 0.5 uref + d0 + n 0.1012 1.228 u0 = 1.0 uref + d0 + n 0.1014 1.234 u0 = 1.5 uref + d0 + n 0.1017 1.242
IAE
9.672 9.691 9.721
u0 = 0.5 uref + d1 + n 0.1051 1.325 9.971 u0 = 1.0 uref + d1 + n 0.1073 1.383 10.12 u0 = 1.5 uref + d1 + n 0.1103 1.461 10.31 uref = sin(0.25t); d0 = 0.5 sin(0.25t); d1 = 2.5 sin(0.25t); n = random noise
Fig. 12. Stability analysis: (a) PI-controller and (b) R-controller
zero cancellation will occur in the system model, feedback system and designed controller transfer function. The pole-zero plot of the close loop transfer function for PI-controller and R-controller is shown in Fig. 12. It observed that all poles lie in the left half of the s-plane and no pole-zero cancellation occurred in the system model, feedback system and designed controller. Hence, the speed subsystem at various velocity setting is input-output and internally stable.
Ubiquitous Control Structure and It’s Comprehensive Application
193
Conclusion An inevitable and ubiquitous feedback loop structure for the speed subsystem of an AUV is addressed in this paper. The proposed control schemes validate the control objective for sinusoidal reference tracking in the presence of parameter uncertainty, disturbance and noise. The analysis of the proposed controller verified through simulation result, and it found to be effective and efficient. The appended strengths of the proposed design are 1. Effective for disturbance rejection and noisy environment for sinusoidal trajectory tracking under parameter uncertainty. 2. A simple and cost-effective design structure with ease in implementation. The proposed R-controller confirms its effectiveness through simulation results in terms of steady-state tracking error, disturbance rejection, noisy environment and robustness with qualitative analysis.
References 1. Stefani, R.T., Shahian, B., Savant, C.J., Hostetter, G.H.: Design of Feedback Control Systems, 4th edn. Oxford University Press, Oxford (2002) 2. Franklin, G.F., Powell, J.D., Emami-Naeini, A.: Feedback Control of Dynamics Systems, 7th edn. Pearson Education, Inc., Boston (2015) 3. Ogata, K.: Modern Control Engineering, 5th edn. Pearson Education, Inc., Boston (2010) 4. Nise, N.S.: Control System Engineering, 7th edn. Wiley, Hoboken (2015) 5. Chen, C.T.: Linear System Theory and Design, 3rd edn. Oxford University Press, Oxford (1999) 6. Astrom, K.J., Murray, R.M.: Feedback Systems an Introduction for Scientists and Engineers, 2nd edn. Princeton University Press, Princeton (2019) 7. Paraskevopoulos, P.N.: Modern Control Engineering. Control Engineering Series. Marcel, Dekker, Inc., New York (2002) 8. Simon, D.: Optimal State Estimation Kalman, H∞, and Nonlinear Approaches. Wiley, Hoboken (2006) 9. Boyd, S., Barratt, C.: Linear Controller Design: Limits of Performance. Pearson Education, Inc., Boston (1991) 10. Wang, L.: PID Control System Design and Automatic Tuning Using MATLAB/Simulink. Wiley-IEEE Press (2020) 11. Doyle, J., Francis, B., Tannenbaum, A.: Feedback Control Theory. Macmillan Publishing Co. (1990) 12. Xue, D., Chen, Y.: Modeling, Analysis and Design of Control Systems in Matlab and Simulink. World Scientific Publishing Co. Pte. Ltd., Singapore (2015) 13. Fossen, T.I., Pettersen, K.Y.: Guidance and Control of Ocean Vehicles. Wiley, Hoboken (1994) 14. Healey, A.J., Lienard, D.: Multivariable sliding mode control for autonomous diving and steering of unmanned underwater vehicles. IEEE J. Oceanic Eng. 18(3), 327– 339 (1993) 15. Wilman, A.P.M., Alain, G.S., Spartacus, G.C.: The predictive functional control and the management of constraints in GUANAY II autonomous underwater vehicle actuators. IEEE Access 6, 22353–22367 (2018)
PCA-Based Random Forest Classifier for Speech Emotion Recognition Using FFTF Features, Jitter, and Shimmer Sheetal Patil(B) and G. K. Kharate Matoshri College of Engineering and Research Centre, Nashik, India [email protected]
Abstract. In recent years, it has been observed that emotion recognition is a rapidly growing research area. Unlike humans, machines do not possess aptitudes to recognize emotions. This study aims to recognize five emotions, i.e., fear, anger, happiness, neutral, and sadness, from a speech signal. EmoDB dataset comprises the above mentioned emotions; so, this German dataset is used. Prosodic features like energy, voiced and unvoiced frame duration, and silence are extracted. When an individual expresses his/her emotions strongly, his/her voice amplitude and frequency fluctuations turn out to be high. Such fluctuations can detect an emotion, and it is possible to measure jitter and shimmer prosodic features that are extracted for a speech signal’s voiced frames. Spectrum tilt and zero-crossing rate are obtained, and frames of a speech segment are classified as voiced, unvoiced, and silence using hamming windows. Each frame comprises 960 speech samples. Moreover, in our study, a new feature, i.e., Fast Fourier transform (FFT), is extracted. Instead of using FFT of 960 samples, only mean, minimum, maximum, variance, and standard deviation for each FFT are used. We name these features as FFTF. Thirteen Mel frequency cepstral coefficients features are also extracted. A random forest classifier is used; in the first experiment, the effect of different number of trees is observed. With 100 trees, an emotion recognition accuracy of 78.6885% is attained. In the second experiment, different features of voiced and unvoiced frames are used, and it is observed that for FFT features, an accuracy of 78.6885% is attained. In the third experiment, the number of features are reduced using PCA, and it is observed that the accuracy is improved to 83.60%. Keywords: Emotion recognition · FFTF features · Jitter · Shimmer · RF classifier · PCA
1 Introduction If an entertainment robot recognizes an emotion, then it could respond to his/her owner in a different manner depending on his/her emotional state. Emotion recognition through speech is one of the research areas for emotional human–computer interaction [1]. Performance in emotion recognition literatures is not an easy task to compare because of the scarcity of databases. Albeit it is difficult to determine which system is better in © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 194–205, 2022. https://doi.org/10.1007/978-981-19-1677-9_17
PCA-Based Random Forest Classifier for Speech Emotion Recognition
195
general, a few previous systems are described in this study to offer a general idea on current approaches. Wang et al. [2] used Fourier parameters for emotion recognition as one of the features. They represented a voiced segment in terms of Fourier series and then used its first 120 coefficients and their 1st and 2nd order deviations, which led to a large feature set and heavy computation. For EmoDB dataset with an SVM classifier, they achieved 79.9% emotion recognition accuracy. In our research work, we computed Fast Fourier transform (FFT), and instead of using its coefficients, we used its extracted features as minimum, maximum, mean, and standard deviation, which not only reduced the number of coefficients but also improved the rate of emotion recognition. Iqbaland Barua [3] examined tonal properties for identifying emotions. They extracted 34 features including MelFrequency Cepstral Coefficients (MFCC), energy, and spectral entropy. Using RAVDESS and SAVEE datasets with the implementation of GB, SVM, and k-NN classifiers, they achieved an emotion recognition accuracy of~80%. Ververidis and Kotropoulos [4] used spectral features like MFCC, jitter, and shimmer for classifying human speaking styles and animal vocalization arousal levels. In their study, they used SUSAS dataset for human speaking styles and vocalizations labeled by arousal levels for African elephants and Rhesus monkey species. From their study, it was observed that adding jitter and shimmer significantly improved there cognition rate of arousal to 96%. Ke et al. [5] improved emotion recognition rate with the help of a continuous hidden Markov model. They extracted 33 dimensional features including zero-crossing rate (ZCR), formant frequency, and energy. ZCR and fundamental frequency, its differentiation, and second-order differentiation were computed. The use of PCA improvedthe recognition rate to 67.83%. Ayadi et al. [6] emphasized on different features, classifiers, and feature reduction methods. Emotion recognition performance mainly depended on how relevant features can be extracted without changing the speaker, language, and contents. In order to reduce variability in features, it was crucial to normalize the feature. A backend classifier that works on a fixed length feature vector was used for classification. Moreover, it is essential to convert a variable length feature vector into a fixed length feature vector. Because the resulting feature vectors generally have large dimensions, selecting good features is needed to improve not only the overall accuracy but also reduce feature dimensions. In our study, we extracted prosodic features like energy, duration, jitter, and shimmer. System features like FFT features, FFTF (Fast Fourier Transform Features) and MFCC were also extracted. Random forest was used as the classifier. In order to reduce dimensionality, Principal Component Analysis (PCA) was used. Furthermore, the effect of the number of trees on the performance of emotion recognition was observed. The performance of our system was compared using RAVDESS and EmoDB datasets.
2 Emotion Recognition Emotion recognition consists of the following stages.
196
S. Patil and G. K. Kharate
2.1 Framing and Separation of Voiced and Unvoiced Segments and Silence For smoothing and avoiding drastic segmentation in our study, hamming window was used. The complete voice segment was divided into frames of 20ms with 10ms overlapping. When a speech sentence was recorded and analyzed, there were some silent parts in the utterances where there was no energy present, or the energy was very less and was equivalent to small noise. The unvoiced segment had slightly higher amplitude than the silence part, and the voiced segment had higher energy. Voiced speech comprises constant frequency tones of certain duration, which occurs when vowels are spoken. It is produced when periodic air pulses generated by the vibrating glottis resonate through the vocal tract. Approximately 2/3rd of the speech is voiced, which is most important for unambiguousness. Unvoiced speech is non-periodic, which is caused by air passing through a narrow tightening of the vocal tract when consonants are spoken. Voiced speech can be identified and extracted because of its periodic nature. As per emotions, the duration of voiced and unvoiced segments and silence is changed. In addition, unvoiced segment feature can help in recognizing emotions. As per emotions, silent durations can differ. Therefore, speech segmentation in voiced and unvoiced segments and silence is considered in this study. Moreover, instead of using only one method, two methods, i.e., ZCR and spectrum tilt, were combined to acquire better results. ZCR indicates the frequency at which energy is concentrated in the signal spectrum. Voiced speech exhibits allow zero-crossing count. Unvoiced speech exhibits a high zero-crossing count. Silence’s zero-crossing count is expected to be lower for unvoiced speech but quite comparable to that for voiced speech. Spectrum tilt can be represented by first-order normalized autocorrelation, and it is ~1 for a voiced segment, < 0.65 for an unvoiced segment, and ~0 for a silence segment. Herein, we have applied several algorithms by combining ZCR as well as spectrum tilt. 2.2 Feature Extraction We combined prosodic features and system features in this study. Researchers have used different features like MFCC, FP, LPCC, LFPC, and mean Fourier parameters. We selected MFCC as its provides good results and FFTF features. Prosodic features are those that can be perceived by humans such as intonation and rhythm [7]. Different prosodic features like energy, pitch frequency, ZCR, duration, jitter, and shimmer were used by several researchers. In our study, we use energy, duration, jitter, and shimmer. 2.2.1 MFCC MFCC represents parts of a human speech production and perception. It depicts the logarithmic perception of loudness and pitch of a human auditory system. In addition, it eliminates certain characteristics that are dependent on a speaker by removing the elementary frequency accompanied by their harmonics. For each voiced and unvoiced frame, MFCC coefficient is computed. As MFCC values do not differ much after the 13th coefficient, we considered the values up to the 14th coefficient. Next, for each speaker’s utterance, voiced frame, and unvoiced frame, its minimum, maximum, mean, and standard deviation were computed, which is the total form feature vector. Figure 1 shows MFCC features for all emotions in an EmoDB dataset.
PCA-Based Random Forest Classifier for Speech Emotion Recognition
197
Fig. 1. MFCC features for voiced and unvoiced frames for all emotions in an EmoDB dataset.
2.2.2 FFT Features A speech is a complex signal that is an ensemble of different frequencies and harmonics. As per emotions, both pitch and frequencies in a speech change or its amplitude differ. Thus, analysis of frequencies present in a speech signal will help in recognizing emotions correctly. We considered FFT of each voiced and unvoiced frame, as shown below. X [k] =
N −1 n=0
x[n]e
−j2π kn N
; k = 0, 1 . . . N − 1
(1)
As higher coefficients do not carry much information, we decided to take first 120 FFT coefficients. However, if we consider per frame 120 coefficients, it will make the feature vector lengthy, which will not only increase computation burden but also provide an overfitting issue for classification. Therefore, instead of 120 FFT coefficients, minimum, maximum, mean, and standard deviation for each coefficient were calculated separately for voiced and unvoiced frames. As a result, FFT feature vector comprises 480 feature vectors each for voiced and unvoiced frames.
198
S. Patil and G. K. Kharate
2.2.3 Energy The energy of speech is extracted from its intensity. The energy associated with speech differs continuously; therefore, the energy associated with a short speech range is of utmost importance. The energy level for each frame is calculated by squaring each frame’s amplitude. Herein, the energy for voiced frames has been calculated in addition to its minimum, maximum, mean, and standard deviation.
Fig. 2. Energy for voiced frames for all emotions
2.2.4 Duration As silent part can play a crucial role in emotion recognition, silence duration was considered as one of the features in this study. Note that as per an emotion, the number of voiced and unvoiced frame sizes can change. Therefore, the ratio of voiced to unvoiced duration can be considered as one of the features. 2.2.5 Jitterand Shimmer When there are strong emotions of any person, it gives variation in pitch frequency and energy. When a person is angry or extremely happy, there may be variation in the pitch frequency. When a person is speaking with a sad or a surprise emotion, there may be variation in the energy. Therefore, features measuring these variations in amplitude and frequency can play a crucial role in emotion recognition. To observe its effects, jitter and shimmer are considered. For each voiced frame, jitter and shimmer are calculated. In addition, minimum, maximum, mean, and standard deviation are computed for each voice utterance. Jitter: It is a measure of period-to-period fluctuations in a fundamental frequency [8]. Vocal jitter increases in voice disorder and is responsible for a hoarse, harsh, or rough voice quality. Jitter is a measurement of vocal stability. Normal voices are usually 0. Fundamental frequency amplitude gain: KA1 . Fundamental frequency phasor angle shift: Kθ1 . Signal x(t) after passing through first order low pass filter is expressed as, x(t) = A0 +
N −2 n=1
Cn cos(nωt + n ) + De
−t τ
−t
+ D1 e τ1
(13)
222
D. N. Tailor and V. H. Makwana
By applying MFCDFT, xr(N ) =
N 2 2kπ x(k) cos N N k=1
=C1 cos 1 +
N −kT −kT 2kπ 2 De τ + D1 e τ1 . cos N N
(14)
k=1
xi(N ) = −
2 N
N
x(k) sin
k=1
= C1 sin 1 −
2kπ N
N −kT −kT 2kπ 2 De τ + D1 e τ1 . sin N N
(15)
k=1
−T
−kT
Let consider, Z = e τ (unknown), K1 = e τ (known)
N xr(N +1) − xr(N ) = DZ Z N − 1 + D1 K1 K1N − 1 2π 2 cos N
N xr(N +2) − xr(N +1) = DZ 2 Z N − 1 + D1 K12 K1N − 1 4π 2 cos N
N xr(N +3) − xr(N +2) = DZ 3 Z N − 1 + D1 K13 K1N − 1 6π 2 cos N
Now (17) − (16) × K1 = DZ Z N − 1 (Z − K1 )
(18) − (17) × K1 = DZ 2 Z N − 1 (Z − K1 )
(16) (17) (18) (19) (20)
= (20)/(19) D is obtained with the help of Z and (19) and D1 by using Z, D and (16). Consequently, C1 cos 1 is achieved by using Z, D,D1 and (14) &C1 sin 1 by using Z, D,D1 and (15). So, A1 = KCA11 , θ1 = 1 + Kθ1 . Similarly, A1 and θ1 is obtained by using imaginary part Xi(k).
4 Result and Discussions This section describes the performance of Conventional DFT algorithm and Modified Full Cycle DFT algorithm. In this section, various types of faults are simulated, such as, three phase-to-ground fault, line-to-line fault and single line-to-ground fault on standard IEEE 9 bus system using MATLAB/ SIMULINK considering wide variation in fault as well as system parameters. Performance of both the algorithms are tested considering variation in fault inception angles (0°, 110° & −60°), wide variation in fault location
Transient Stability Analysis of IEEE 9 Bus System
223
Fig. 2. MATLAB model of Standard IEEE 9 bus system
(10% −70% in step of 20%) and change in fault resistance (1, 5 & 10). Above mentioned fault conditions are seriously disturbing the fundamental signals with unwanted harmonics and decaying dc. These signals are measured in all the above simulated conditions. The measured values are now processed by the conventional DFT and derived modified full cycle DFT algorithms. Figure 2 represents the MATLAB/SIMULINK model of Standard IEEE 9 bus system [24]. To validate the robustness & preciseness of Modified Full Cycle DFT algorithm, different extreme conditions are simulated using MATLAB model. Table 1 shows different cases that are considered for the validation of the MFCDFT algorithm. Table 1. Different cases for validation of Algorithm Case
Fault inception angle (F)
Fault resistance (Rf)
% Fault location on transmission line
Case 1
At 90°
1
10%
Case 2
At 90°
1
50%
Case 3
At 110°
1
10%
Case 4
At 110°
5
50%
Case 5
At 110°
10
70%
Case 6
At −60°
1
10%
Case 7
At −60°
5
50%
Case 8
At −60°
10
70%
Case 9
At 100°
0.1
10%
Case 10
At −60°
0.1
10%
224
D. N. Tailor and V. H. Makwana
Fault inception angle (90°, 110° & −60°) can be changed by controlling the switching time. Different cases have their unique effect on network as well as on computation method. Conventional DFT algorithm is capable to remove harmonics from input signals but fails to remove dc decaying component. On other hand, Modified Full Cycle DFT algorithm has capability to remove unwanted harmonics and delaying dc from input signals. In Case1, the fault inception angle is 90° with fault resistance of 1 & fault occurred at 10% of transmission line. As per the theoretical analysis of the series RL circuit during sudden short circuit conditions, if fault inception angle is 90° then the exponential decaying component call dc component is absent. Figure 3a shows the actual waveform during fault condition & Fig. 3b shows the comparison of output of the conventional & the MFCDFT algorithms. It has been observed from Fig. 3a that very small amount of decaying dc is presented in waveform.. It is to be noted from Fig. 3b that because of absent of decaying dc component at the output of MFCDFT algorithm, it provides stable and constant output within one and half cycle. So, relay can quickly take decision and isolate faulty section from healthy section. Whereas the output of conventional DFT algorithm has small amount of dc delaying which may remain present up to three to four cycles. So, conventional relay must be wait till the output become stable for isolation of faulty section from healthy section and it may result into the delay operation & ultimately step toward the power system instability.
Fig. 3. (a)Actual waveform _Case1_ F = 90°, Rf = 1 & fault location 10% of transmission line.
Similarly, in Case2, fault is assumed to be present at 50% of transmission line section which ultimately increase the exponentially decaying component in signal that is easily observed in Fig. 4a. The simulation results of conventional DFT & MFCDFT algorithm are shown in Fig. 4b. The output of conventional DFT algorithm is failed to remove dc components. So output is not stable for few cycles, and it may lead to delay in relay operation. Where in case of MFCDFT algorithm, Signal passed through low pass filter. Filter can’t remove the dc components from the signal but produces a new decaying dc whose time constant values are obtained according to characteristic equation of the low
Transient Stability Analysis of IEEE 9 Bus System
225
Fig. 3. (b)Comparison of Conventional & modified DFT_Case1_ F = 90°, Rf = 1 & fault location 10% of transmission line.
pass filter. Now this signal passed through MFCDFT algorithm. Mathematical derivation is shown by Eq. 13 to 15. By using MFCDFT algorithm the output stability is achieved within one and half cycle. So, it is concluded that the relay can take quick decision to isolate the faulty network. Now in Case3, the fault inception angle is 110° with fault resistance of 1 & fault occurred at 10% of transmission line. The presence of decaying dc is mainly depending on fault inception angle. As shown in Fig. 5a the current signal contains decaying dc with unwanted harmonics. When signal passed through both algorithms, the conventional
Fig. 4. (a) Actual waveform _Case2_ F = 90°, Rf = 1 & fault location 50% of transmission line.
226
D. N. Tailor and V. H. Makwana
Fig. 4. (b) Comparison of Conventional & modified DFT_Case2_ F = 90°, Rf = 1 & fault location 50% of transmission line.
algorithm failed to remove decaying dc component so that the output signal needed more than five cycles to become stable which shown in Fig. 5b. Where the modified Full cycle DFT algorithm has capability to remove the decaying dc and unwanted harmonics within one and half cycle and make output current signal stable so that relay can take quick decision of isolation of the faulty section. Nowadays, as the power system network is highly complex and interconnected, the relays are required to take quick decision (within two cycles) to maintain the power system stability.
Fig. 5. (a) Actual waveform _Case3_ F = 110°, Rf = 1 & fault location 10% of transmission line.
Transient Stability Analysis of IEEE 9 Bus System
227
Fig. 5. (b) Comparison of Conventional & modified DFT_Case3_ F = 110°, Rf = 1 & fault location 10% of transmission line.
Fig. 6. (a) Actual waveform _Case4_ F = 110°, Rf = 5 & fault location 50% of transmission line.
In Case4 & 5, the fault inception angle is 110° and fault resistance of 5 & 10 with fault occurred at 50% and 70% of transmission line respectively. As shown in Fig. 6a & Fig. 7a, with increasing line impedance and the fault resistance value the intensity of decaying dc is decreased, and it quickly die out within few cycles. When these distorted signals passed through both the algorithms, Fig. 6b & Fig. 7b shown that Modified Full Cycle DFT easily eliminated the unwanted harmonics and decaying dc within two cycles which is the prior requirement of modern interconnected power system protection scheme for maintaining the stability.
228
D. N. Tailor and V. H. Makwana
Fig. 6. (b) Comparison of Conventional & modified DFT_Case4_ F = 110°, Rf = 5 & fault location 50% of transmission line.
Fig. 7. (a) Actual waveform _Case5_ F = 110°, Rf = 10 & fault location 70% of transmission Line
Transient Stability Analysis of IEEE 9 Bus System
229
Fig. 7. (b) Comparison of Conventional & modified DFT_Case5_ F = 110°, Rf = 10 & fault location 70% of transmission line.
Similarly, in Case6, case7 & Case8 fault inception angle is −60° with varied fault locations and fault resistance. Here also the signal is heavily distorted by decaying dc and unwanted harmonics which is shown in Fig. 8a, Fig. 9a & Fig. 10a. The output signal of conventional DFT algorithm takes long time get stabilized as shown in Fig. 8b, Fig. 9b & Fig. 10b. It may result into the instability in the interconnected power system. However, the Modified Full Cycle DFT algorithm provides robust and stable output during all extremely adverse conditions within the one and half cycle which helps to maintain the power system stability.
Fig. 8. (a) Actual waveform _Case6_ F = −60°, Rf = 1 & fault location 10% of transmission line
Case9 & Case10 covers the extremely adverse conditions of IEEE 9 bus system. In both the cases, the fault resistance is nearly zero and the fault occurred at 10% of
230
D. N. Tailor and V. H. Makwana
Fig. 8. (b) Comparison of Conventional & modified DFT_Case6_ F = −60°, Rf = 1 & fault location 10% of transmission line.
Fig. 9. (a) Actual waveform _Case7_ F = −60°, Rf = 5 & fault location 50% of transmission Line
Transient Stability Analysis of IEEE 9 Bus System
231
Fig. 9. (b) Comparison of Conventional & modified DFT_Case7_ F = −60°, Rf = 5 & fault location 50% of transmission line.
Fig. 10. (a) Actual waveform _Case8_ F = -60°, Rf = 10 & fault location 70% of transmission Line
232
D. N. Tailor and V. H. Makwana
Fig. 10. (b) Comparison of Conventional & modified DFT_Case8_ F = −60°, Rf = 10 & fault location 100% of transmission line.
transmission line. As shown in Fig. 11a and Fig. 12a, the signals are extremely distorted by decaying dc. Both the signals are extremely unbalanced and take very long time to come at origin position. These extreme conditions are shown in Fig. 13 & Fig. 14. In this type of fault condition all the algorithm proposed by different researchers got failed to eliminate decaying dc as its decaying time is extremely long. So, it the new scope of research for researchers.
Fig. 11. (a) Actual waveform _Case9_ F = 110°, Rf = 0.1 & fault location 10% of transmis-sion Line
Transient Stability Analysis of IEEE 9 Bus System
233
Fig. 11. (b) Comparison of Conventional & modified DFT_Case9_ F = 110°, Rf = 0.1 & fault location 10% of transmission line
Fig. 12. (a) Actual waveform _Case8_ F = −60°, Rf = 0.1 & fault location 10% of transmission Line
234
D. N. Tailor and V. H. Makwana
Fig. 12. (b) Comparison of Conventional & modified DFT_Case8_ F = −60°, Rf = 0.1 & fault location 10% of transmission line.
Fig. 13. Case9_ F = −110°, Rf = 0.1 & fault location 10% of transmission line
Fig. 14. Case10_ F = - = 60°, Rf = 0.1 & fault location 10% of transmission line
Transient Stability Analysis of IEEE 9 Bus System
235
5 Conclusion During different faulty conditions, the current and voltage signals contained serious unwanted lower order harmonic & dc decaying components. The value of time constant & magnitude of dc decaying were unknown. It was dependent on the fault location, fault inception time & fault resistance. Conventional DFT algorithm used to separate out fundamental frequency component took too much time or having considerable error. On the other hand, the Modified Full Cycle DFT algorithm has quickly eliminated the decaying dc from the signals. The simulation results proved the effectiveness of the MFCDFT algorithm as with the help of this algorithm it could be possible to extract the fundamental frequency components from the input signals quickly and precisely.
References 1. Altuve, H., Diaz, I.V., Vazquez, E.M.: Fourier and Walsh digital filtering algorithms for distance protection. IEEE Trans. Power Syst. 11(1), 457–462 (1996) 2. Schweitzer, E.O., Hou, D.: Filtering for protective relays. In: 19th Annual Western Protective Relay Conference, Spokane, WA (1992) 3. Gu, J.C., Yu, S.L.: Removal DC in current and voltage using novel filteralgorithm. IEEE Trans. Power Delivery 15(1), 73–79 (2000) 4. Sun-Li, Y., Jyh-Cherng, G.: Removal of decaying DC in current and voltage signals using a modified Fourier filter algorithm. IEEE Trans. Power Del. 16(3), 372–379 (2001) 5. Guo, Y., Kezunovic, M., Chen, D.: Simplified algorithms for removal of the effect of exponentially decaying dc-offset on the Fourier algorithm. IEEE Trans. Power Del. 18(3), 711–717 (2003) 6. Nam, S.-R., Park, J.-Y., Kang, S.-H., Kezunovic, M.: Phasor estimation in the presence of DC offset and CT saturation. IEEE Trans. Power Delivery 24(4), 1842–1849 (2009) 7. Kanieski, J.K., Cardoso, R., Pinheiro, H.: Kalman filter-based control system for power quality conditioning devices. IEEE Trans. Ind. Electron. 60(11), 5214–5227 (2013) 8. Sachdev, M.S., Nagpal, M.: A recursive least error squares algorithm for power system relaying and measurement applications. IEEE Trans. Power Del. 6(3), 1008–1015 (1991) 9. Costa, F.F., Formiga, D.A., Ferreira, R.R., Sousa, T., Costa, F.B.: A recursive least-squares aided by pre-filtering for phasor-estimation in distance protection. In: IEEE Grenoble Conference, pp. 1–6 (2013) 10. Benmouyal, G.: Removal of dc-offset in current waveforms using digital mimic filtering. IEEE Trans. Power Del. 10(2), 621–630 (1995) 11. Yu, C.-S.: A discrete Fourier transform-based adaptive mimic phasor estimator for distance relaying applications. IEEE Trans. Power Delivery 21(4), 1836–1846 (2006) 12. Mai, R.K., Fu, L., Dong, Z.Y., Kirby, B., Bo, Z.Q.: An adaptive dynamic phasor estimator considering DC offset for PMU applications. IEEE Trans. Power Del. 26(3), 1744–1754 (2011) 13. Mai, R.K., Fu, L., Dong, Z.Y., Kirby, B., Bo, Z.Q.: Dynamic phasor and frequency estimators considering decaying DC components. IEEE Trans. Power Syst. 27(2), 671–681 (2012) 14. Byeon, G., Oh, S., Jang, G.: A new DC offset removal algorithm using an iterative method for real-time simulation. IEEE Trans. Power Del. 26(4), 2277–2286 (2011) 15. Ren, J., Kezunovic, M.: Real-time power system frequency and phasors estimation using recursive wavelet. IEEE Trans. Power Del. 26(3), 1392–1402 (2011) 16. Serna, D., Maldonado, R.: Taylor–Kalman–Fourier filters for instantaneous oscillating phasor and harmonic estimates’. IEEE Trans. Instrum. Meas. 61(4), 941–951 (2012)
236
D. N. Tailor and V. H. Makwana
17. Tajdinian, Jahromiand Jal, Z., Kouhsari, M.: Comprehensive and innovative method for dynamic phasor and frequency estimation. In: IEEE 2015 9th Power Systems Protection and Control Conference (PSPC), Tehran, Iran, pp. 42–48 (2015) 18. Belega, D., Fontanelli, D., Petri, D.: Dynamic phasor and frequency measurements byan improved Taylor weighted least squares algorithm. IEEE Trans. Instru. Meas. 64(8), 2165– 2178 (2015) 19. Kapoor, G., Mishra, V.K., Shaw, R.N., Ghosh, A.: Fault detection in power transmission system using reverse Biorthogonal wavelet. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 381–393. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0749-3_28 20. Ashrafian, A., Mirsalim, M., Masoum, M.A.S.: An adaptive recursive wavelet based algorithm for real-time measurement of power system variables during off-nominal frequency conditions. IEEE Trans. Ind. Inform. 14(3), 818–828 (2018) 21. Zhan, L., Liu, Y.: A Clarke transformation-based DFT phasor and frequency algorithm for wide frequency range in IEEE Trans. Smart Grid 9(1), 67–77 (2018) 22. Zaki, M.I., El-Sehiemy, R.A., Amer, G.M., Abo El Enin, F.M.: An investigated reactive power measurements-based fault-identification scheme for teed transmission lines. In: International Measurement Confederation, vol. 136, pp. 185–200 (2019) 23. Ferrero, R., Pegorao, P., Toscani, S.: Synchrophasor estimation for three phase systems based on Taylor extended Kalman filtering. IEEE Trans. Instrum. Meas. 6(9), 6723–6730 (2019) 24. MATLAB Central File Exchange, October 2021. https://www.mathworks.com/matlabcentral/ fileexchange/45936-ieee-9-bus
District Level Analytical Study of Infant Malnutrition in Madhya Pradesh Supriya Vanahalli(B) , Sarmista Biswas, Jossy P. George, and Samiksha Shukla Christ University, Bangalore, India {vanahalli.chanvir,sarmista.biswas}@science.christuniversity.in, {frjossy,samiksha.shukla}@christuniversity.in
Abstract. One of the main causes for India’s high infant mortality rate is malnutrition. It can be addressed using three broad groups of conditions: stunting, wasting, and underweight. Other factors such as sanitation, poverty, breastfeeding also contribute to the prevalence of malnutrition. Understanding the contribution of these factors and thus, eliminating them, to reduce malnutrition, is the purpose of this study. In this chapter, the district-level data obtained through NFHS-4 is used for analytical study for infant malnutrition, in Madhya Pradesh. Hierarchical Agglomerative clustering is used to group the districts based on the factors such as exclusively breastfeeding, inoculation, breastfeeding within one hour, no inoculation. The proposed model presents the effect of each factor, on infant malnutrition. It will help decision-makers and the government to shortlist the most appropriate districts contributing to malnutrition and to take curative action to reduce the rate of infant malnutrition. It is a generic model which can be utilized by other states to study infant malnutrition. Keywords: Infant malnutrition · Hierarchical clustering · Stunting · Wasting · Underweight
1 Introduction Malnutrition is used to refer to the deficiencies in an individual’s intake of nutrients. It covers two broad groups of conditions — undernutrition, which involves being underweight i.e., below the required weight as per age, stunting i.e. being below the required height as per age, wasting i.e. being below the weight necessary for one’s height, and deficiency in micronutrients like minerals and vitamins. Obesity, being overweight, and diseases caused by diet are other such conditions. Infantmalnutrition mostly deals with wasting and leads to infant mortality. The factors contributing to child malnutrition can be short-term and long-term. Short-term factors include improper feeding practices, less/no immunization from chronic diseases, feeding of poor-quality food items, irregular growth monitoring, unclean drinking water, and improper sanitation and hygiene practices. Whereas the long-term factors are more socioeconomic ones like poverty, illiteracy, low income, lack of awareness for child care, improper family planning, and poor sanitary environment. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 237–247, 2022. https://doi.org/10.1007/978-981-19-1677-9_20
238
S. Vanahalli et al.
Infant mortality is the most devastating result of malnutrition. Malnourished children are at a higher risk of having infectious diseases like pneumonia, malaria, tuberculosis. Poor mental development is also an impact. These children also have low levels of productivity in the future. So, it becomes very essential to control child malnutrition by understanding the contribution of individual factors. The Joint Child Malnutrition Study conducted by UNICEF, WHO, and World Bank Group shows that children under five years of age facing stunting amounted to 148 million and 47 million suffered from wasting globally in 2019. Southern Asia was among the top contributors to the highest number of stunting and wasting cases in the world. As per the fact sheet released on April 1, 2020, by WHO, 47 million children globally under five years of age were impacted by wasting. 14.3 million children had severe cases of wasting and 144 million were affected by stunting. Around 45% of deaths worldwide among infants are related to malnutrition. Motivation: Although the rate of child mortality has considerably decreased in India since 2008, it is still a major concern. There is a necessity to reduce child mortality and hence, eradicate child malnutrition, especially in India. With this aim, this study analyses the factors responsible for infant malnutrition in the various districts of Madhya Pradesh by accessing data from the National Family Health Survey (NFHS-4) data. As of 2020, Madhya Pradesh recorded the highest infant mortality rate. Therefore, this chapter focuses on the district-level analysis of infant malnutrition in this state. Contribution: Hierarchical Agglomerative clustering has been used in this chapter to group the districts based on the similarity in factors such as exclusively breastfeeding, inoculation, breastfeeding within one hour, no inoculation. Euclidean distance is the similarity measure used and the dendrogram obtained gives the clusters formed using the districts. Correlation is also studied to understand the most effective factor for infant malnutrition. Organization: The remaining part of the chapter is structured in the following manner; Section two presents an overview of the literature and insights available on infant malnutrition. Section three explains the methodology used to achieve the objective of the study, followed by the experimental results and discussion. Finally, a conclusion is made along with the references used for the study.
2 Related Work Striessnig et al. [1] applied two kinds of analysis on the NFHS 2015–2016 data. They used the Principal Component Analysis (PCA) along with Hierarchical Cluster Analysis to locate the geographical markers in the nutritional value across 640 Indian districts. The use of hygienic cooking fuels, water that is safe to drink, better sanitation, female education, and access to media were found to be crucial factors of a child’s development. Low BMI of the mother, household poverty level, high fertility, low immunization are major factors contributing to the prevalence of infants’ nutritional deficiencies, lower height, and weight. It has been found that the children growing up in central rural districts have low growth and nutritional development, which is better for those in southern districts, especially concerning height. In the case of average weight and the amount of children who are underweight and wasted, the districts in the North-east are seen to provide improved child development facilities. However, the study does not consider
District Level Analytical Study of Infant Malnutrition
239
the relationship between childhood obesity and malnutrition, the climatic factors which affect the food supply, and the heterogeneity of the samples. The data used lack representativeness at the national level. Hence, the study lacks the representation of the regional district clusters where infant malnutrition is severe. Univariate and bivariate data analysis having a cross-sectional study pattern was performed by R. Rahmi et al. [2] to study the relationship between the variation in complimentary food and infant nutrition fulfilment in infants of age group 6–12 months. Low variation in complementary feeding led to malnutrition, as it lacked a balanced diet. However, the study was limited to sample size. Lee et al. [3] were able to predict the body mass index (BMI) of a new born by using machine learning techniques involving maternal and delivery information and ultrasound measures. The ultrasound measures employed were estimated fetal weight (EFW), biparietal diameter (BPD), and abdominal circumference (AC). Based on the samples of 3159 patients, the study found that EFW and AC were most accurate predictors of the BMI of a newborn along with the maternal BMI and gestational age during delivery. The ultrasound measures were found to be effective from week 36 and later. Linear Regression and Random Forest outperformed Artificial Neural Network for this purpose. However, effects among variables were not considered and a BMI guideline for the newborn based on the results can be developed. Also, various medical variables were not considered and no parameter tuning was performed for ANN. In [4], Galler et al. offered an overview of the effects of postnatal malnutrition on neurodevelopment and cognition, as well as evidence gaps present in early childhood malnutrition and neurodevelopment research studies. It was concluded that poor neurological development of the infant in the pre-natal stage was associated with high chances of mortality. But the data used was not extensively described. Kumar et al. [5] studied the prevalence of TBM and other associated factors among mother and child pairs in India. To quantify the results achieved by using data from the NFHS 4 study, they used binary and bivariate logistic regression. Age of mother, education, obesity of mother, type of delivery and birth size of the baby are the most important factors for TBM in India. However, no comparative analysis was done with other classification methods of data analysis. To study and analyze the link between alcohol and tobacco abuse by mothers and infant malnutrition in South Africa, logistic regression was built by Modjadji et al. [6]. To gather data on alcohol and tobacco use among mothers, a validated questionnaire was employed in addition to recording their socio-demographic features and obstetric history. Infants aged less than 12 months and 300 samples were considered. The z-scores depending on WHO child growth standards were used to assess stunting, underweight, and thinness. The current use of tobacco was linked to increased odds of being thin, while the current use of alcohol, after adjusting for confounders, was linked to the possibility of being underweight among infants. Sarkar et al., examined 485 sample households [7] in a cross-sectional study of the risk and prevalence factors in West Bengal, India. Bivariate analyses were performed to uncover differences in nutritional indices, and multinomial logistic regression models were fitted to investigate the net effect of various socioeconomic determinants on the
240
S. Vanahalli et al.
possibility of child malnutrition. According to the findings, stunting was the most common kind of malnutrition among infants, followed by underweight and wasting. The results of multinomial analyses show that the child’s age, religion, caste, and birth order are all important predictors of the child’s nutritional health.
3 Methodology 3.1 Data Collection By gathering statistics from the Government of India’s National Family Health Survey 2015–16 (NFSH-4) [8], this survey offered data on nutrition, population, and health for India and each state or union territory. The NFHS-4 provides the district-level information. The survey had four schedules, namely, household, woman, man and biomarker. Household schedule saw the collection of information regarding the socio-economic characteristics, number of members in the household, health insurance, water and sanitation for each household. Both Woman and Man’s schedules individually included the information on schooling, marriage, reproductive health, nutrition and attitude towards other gender. Biomarker Schedule covered the haemoglobin level, height, weight for children. It had the haemoglobin level, blood pressure and random glucose level, height, weight, for men between the ages of fifteen and fifty four and women between the ages of fifteen and forty nine years. HIV testing was also done in this phase. Questionnaires were pre-designed for each of the four survey stages. Hence, data obtained from each stage is present in the questionnaires. Electronic data collected was received on a regular basis via IFSS at the International Institute of Population Sciences. The necessary processing was done for each district and data was integrated at a state level. The factors for the district-level malnutrition of infants were shortlisted from NFSH4 Madhya Pradesh file and final data was made by merging the information in a CSV file. The final dataset contains seven features for fifty districts. The attributes considered for the study are Stunted, wasted, underweight, exclusively breastfeed, breastfeeding within one hour, No inoculation, and full inoculation. 3.2 Data Description According to WHO [9], the definition of wasting is having a low weight-for-height ratio, stunting is having a low height-for-age ratio, and underweight is having a low weightfor-age ratio. The WHO Growth Standards Population Median is used to determine the appropriate standard height, weight, and age for defining the terms (Table 1). Inoculation is another term for vaccination. The dataset includes the infants who are not vaccinated individually with three doses of Hepatitis B, BCG, three doses of DPT, three doses of polio, Measles. It also contains the percentage of infants who have not got all of the core vaccines, which include BCG, measles, three doses of DPT, and polio vaccine (excluding the one given at birth). The act of feeding the infant with only breast milk and no other liquid or solid, is Exclusive breastfeeding. The status of breastfeeding is a 24-h period.
District Level Analytical Study of Infant Malnutrition
241
Table 1. Attributes of the Dataset Parameter
Definition
Stunted
Percentage of children below five years of age who are under the standard height for age (90%
0.107
Proposed Work
25–73
>90%
0.054
*Note: Here λL is the wavelength at the lowest frequency
References 1. Oh, J.-H., Oh, K.-S., Kim, C.-G., Hong, C.-S.: Design of radar absorbing structures using glass/epoxy composite containing carbon black in X-band frequency ranges. Compos. B Eng. 35, 49–56 (2004). https://doi.org/10.1016/j.compositesb.2003.08.011
260
A. Choudhary et al.
2. Gnidakouong, J.R.N., et al.: Electromagnetic interference shielding of composites consisting of a polyester matrix and carbon nanotube-coated fiber reinforcement. Compos. A Appl. Sci. Manuf. 50, 73–80 (2013). https://doi.org/10.1016/j.compositesa.2013.03.007 3. Hippel, A.R., Morgan von, S.O.: Dielectric Materials and Applications. IOP Publishing, Bristol (1955) 4. Chen, L., Wang, G.P.: Pyramid-shaped hyperlenses for three-dimensional subdiffraction optical imaging. Opt. Express, OE. 17, 3903–3912 (2009). https://doi.org/10.1364/OE.17. 003903 5. Bruschi, A., Bin, W., Cirant, S., Gandini, F., Mellera, V., Muzzini, V.: High-Power MillimeterWave Calorimetric Beam Absorbers. Fusion Sci. Technol. 53, 62–68 (2008). https://doi.org/ 10.13182/FST08-A1653 6. Salisbury, W.W.: Absorbent body for electromagnetic waves, https://patents.google.com/pat ent/US2599944A/en (1952) 7. Sheokand, H., et al.: Transparent broadband metamaterial absorber based on resistive films. J. Appl. Phys. 122, 105105 (2017). https://doi.org/10.1063/1.5001511 8. Lai, S., Wu, Y., Zhu, X., Gu, W., Wu, W.: An Optically transparent ultrabroadband microwave absorber. IEEE Photonics J. 9, 1 (2017). https://doi.org/10.1109/JPHOT.2017.2763640 9. Xiao, H., Qin, R., Lv, M., Wang, C.: Highly transparent broadband and polarization-insensitive absorber based on metasurface. Appl. Sci. 10, 9125 (2020). https://doi.org/10.3390/app102 49125 10. Lai, S., Wu, Y., Wang, J., Wu, W., Gu, W.: Optical-transparent flexible broadband absorbers based on the ITO-PET-ITO structure. Opt. Mater. Express, OME. 8, 1585–1592 (2018). https:// doi.org/10.1364/OME.8.001585
AC Analysis on the Modified Johnson Converter Shubham Agrawal(B) , L. Umanand, and B. Subba Reddy Indian Institute of Science, Bangalore 560012, Karnataka, India [email protected]
Abstract. In the contemporary world, the share of AC loads is significant; however, the requirement of DC loads is rapidly increasing, particularly in household and industrial applications. The systems in which both the AC and DC loads are installed require a dual converter, acting as either a DC-DC converter or DC-AC converter. The usage of the dual converter becomes more valuable if it can reduce the output voltage or amplify the output voltage that is buck or boost operation. The present investigation proposes a dual converter, which provides a positive gain with a duty ratio of 0 to 0.5 and a negative gain with a duty ratio of 0.5 to 1. For either range, the converter acts as a DC-DC converter or an inverse gain DC-DC converter. In case of both duty ratios are combined, the converter acts as a DC-AC converter. The paper describes AC analysis of the extended gain range converter (modified Johnson converter), validated through MATLAB simulation. A comparative study of the modified Johnson converter is done with a current-fed converter, and some important conclusions are presented.
Keywords: Extended gain range modulation · DC and AC loads
1
· Current-fed converter · Duty ratio
Introduction
The world is witnessing rapid growth in the renewable energy sector. According to the Government of India (GoI) [1], India presently has installed renewable energy capacity at 101.5 GW. In addition to this, the GoI has ambitious plans to increase the penetration of renewables in the coming days. Solar photovoltaics (PV) has a good share of 45.5% in the installed renewable energy capacity. Solar PV includes ground-mounted, rooftop-mounted, and off-grid installations. The rooftop and off-grid installations are increasing as the cost of PV installations and payback period per watt are reducing due to mass installations across the globe. After generation through various methods, the power electronic converters [2] are essential for serving the voltage requirements of the various loads. For example, the LED bulb uses a DC voltage for its operation; however, the voltage supplied to it is AC from the AC grids. Another example is an inverter and c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 261–270, 2022. https://doi.org/10.1007/978-981-19-1677-9_23
262
S. Agrawal et al.
battery system, which converts the higher AC voltage to a lower AC voltage level through a transformer, then a rectifier system to charge the battery. The reverse of this process happens in the transfer of power from the battery to the connected loads [3]. At present, the loads used for household and industrial purposes require either AC or DC power. The voltage required by these devices may be lesser or more than the grid voltage. Such requirement calls for the development of dual converters, which can play both DC-DC or DC-AC conversion roles. Z-source converter [4–6] combines both the current source and voltage source converter. Z-source topology can convert DC voltage to both DC and AC depending on the requirement of the user. However, it has a limitation [7] of lower average switching device power in the low boost ratio range. The current-fed converter is described for DC [8,9] and AC [10] loads; hence, it can be a dual converter. The drawback of the current-fed converter is that it cannot provide gain between zero and unity in both positive and negative ranges. The motivation of the present investigation is to elaborate the analysis of the extended gain range converter for the AC loads. A critical comparative study is also done between extended gain range converter and current-fed converter for the same design parameters. The paper’s organization is as follows: Sect. 2 describes the DC-DC operation of the extended gain range converter in non-ideal conditions. In Sect. 3, the AC analysis of the converter is done with its duty ratio modulation. Section 4 presents essential simulation results, followed by Sect. 5, where a comparative study is performed between extended gain range converter and current-fed converter. Lastly, some important conclusions are presented.
2
DC-DC Operation of the Modified Johnson Converter
Figure 1 shows a power electronic converter consisting of a DC source, two MOSFETs, two diodes, an inductor, a capacitor, and a resistive load. There are two modes of operation of this converter, which are described in the sub-section.
S1 ig
Vg
vL D1
L
iL
D2 S2 C
iC
R
i0
Fig. 1. Proposed DC-DC converter with appropriate switches
v0
AC Analysis on the Modified Johnson Converter
2.1
263
Non-ideal Operation of the Converter
There are several non-idealities, such as the inductor’s DC resistance RL , diodes’ forward voltage drop Vdo , MOSFETs’ on-state resistance Rds . The operating modes are defined based on switching states. When switches S1 , S2 are in ON state, it is considered as Mode-I, whereas if diodes D1 , D2 are in conduction, it is considered as Mode-II. Mode-I: Figure 2 shows the conduction diagram of the circuit in which switches S1 , S2 provides the path to connect the source and load through various nonidealities. In this mode, diodes D1 , D2 are reversed biased.
vL i L
S1 ig
Rds
RL
Vg
L
Rds S2
iC
C
R
i0
v0
Fig. 2. Mode-I operation in non-ideal conditions
Equations for inductor voltage vL and capacitor current iC can be written using various circuit parameters considered in Fig. 2. vL = Vg + v0 − (2Rds + RL )iL
(1)
iC = −iL − i0
(2)
Mode-II: When the switches S1 , S2 are turned OFF, then due to the inductor’s residual current, the diodes D1 , D2 come into conducting state. Figure 3 shows this mode of operation along with various non-idealities.
vL
ig Vg
RL Vd0
D1
L
iL Vd0
D2 C
iC
R
i0
Fig. 3. Mode-II operation in non-ideal conditions
v0
264
S. Agrawal et al.
vL = −v0 − 2Vdo − iL RL
(3)
iC = iL − i0
(4)
According to volt-second balance, which implies that in steady-state conditions, the area under inductor voltage during a switching cycle should be zero. Mode-I is operated for DTS time duration, and mode-II is operated for (1−D)TS time duration, where D is the duty ratio. Using Eqs. (1) and (3), the following Eq. (5) is obtained. DTS (Vg + v0 − (2Rds + RL )iL ) + (1 − D)TS (−v0 − 2Vdo − iL RL ) = 0
(5)
With reference to the amp-second balance, which implies that in steady-state conditions, the area under capacitor current during a switching cycle should be zero. Using Eqs. (2) and (4), a relation between the inductor current iL and i0 is obtained, which is mentioned in the following Eq. (6). DTS (−iL − i0 ) + (1 − D)TS (iL − i0 ) = 0
(6)
Equations (5) and (6) are combined to determine the relationship between input and voltages in circuit parameters, mentioned in Eq. (7). ⎤ ⎡ 2(1 − D)Vd0 1− ⎥ V0 D ⎢ DVg ⎥ ⎢ (7) = ⎥ ⎢ Vg 1 − 2D ⎣ 2DRds + RL ⎦ 1+ R(1 − 2D)2
S1 ig
Vg
S1'
vL i L L
S4
S3'
S2
S3
S2'
S4'
C
Fig. 4. Bidirectional converter for AC loads
iC
R
i0
v0
AC Analysis on the Modified Johnson Converter
3
265
AC Analysis of the Converter
From Eq. (7), it can be inferred that the converter can be operated for both positive and negative gain using different ranges of the duty ratios. However, for the negative gain, the voltage blocked by switches is in the reverse direction. Hence, bipolar voltage blocking is required to operate the converter in both halves. A thyristor switch can be used to provide the bidirectional voltage blocking; however, the circuitry becomes very large because of its auxiliaries such as commutation circuit and snubber circuit. For this purpose, two MOSFETs can be connected in an anti-series manner. This combination provides bipolar voltage blocking and bidirectional current conduction, which is characteristic of an AC switch. One more advantage of using these AC switches is that the circuit can be used if the input voltage is reverse polarity. In Fig. 4, the converter with four-quadrant or AC switches is shown. Table 1. Swiching rules for AC loads Forward/Reverse polarity of Vg
Output Capacitor charging Capacitor discharging voltage (V0 ) Conducting Blocking Conducting Blocking
Forward (Vg > 0)
V0 > 0 V0 < −Vg
S3 , S4
S1 , S2 S1 , S2
S1 , S2
S3 , S4 S3 , S4
Reverse (Vg < 0)
V0 > V g V0 < 0
S1 , S2
S3 , S4 S3 , S4
S3 , S4
S1 , S 2 S1 , S2
This circuit can be used when the input voltage is either positive or negative polarity. A Switching scheme is derived considering various output voltage levels and capacitor charging or discharging conditions. Table 1 presents the switching rules for various modes. The input voltage Vg is taken as 100 V. Several nonidealities such as diode forward voltage drop Vdo , switches’ on-state resistance Rds and DC resistance of the inductor RL are considered as 0.7 V, 0.05 Ω and 0.1 Ω respectively for a load R equal to 100 Ω. A curve is drawn between the gain and duty ratio. The curve indicates that the gain started from 0 and reached a maximum of around 6; it comes back to 0 at a duty ratio equal to 0.5. Similarly, it started going negative and reached a negative maximum; it comes back to −1 at a duty ratio equal to 1. It can be inferred from the curve that there is no duty ratio for the 0 to −1 range of gain. Figure 5 shows the family of curves which is obtained by varying the circuit parameters.
266
S. Agrawal et al. Gain
6 4
Gain
2 0 −2 −4 −6 0
0.1
0.2
0.3
0.4
0.5 0.6 Duty Ratio
0.7
0.8
0.9
1
Fig. 5. Gain vs duty ratio
3.1
Duty Ratio Modulation
For various duty ratios, the converter gives an extended range of gain. This modulation of duty ratio is helpful to generate a sinusoidal waveform. Various non-idealities are neglected in Eq. (7) to obtain the gain ratio. v0 D = Vg 1 − 2D
(8)
Equation (8) is rearranged to get the duty ratio in terms of input voltage Vg and the output voltage v0 . D=
v0 2v0 + Vg
(9)
In Eq. (9), if the output voltage v0 is sinusoidal, the duty ratio is varied accordingly. However, there is no duty ratio for the gain range of (0, −Vg ). If the various non-idealities are considered, the duty ratio can be computed using all the variables and constants. D=
4Rv0 − a + 6RVd0 − 2Rds v0 + RV g 4(2Rv0 + 2RVd0 + RVg )
2 a = [R2 (Vg2 + 4Vd0 − 4Vd0 V g) − RRds (24v0 Vd0 + 4v0 Vg ) 2 − 16RL R − 16RRds )v02 ]0.5 −8RL R(2v0 Vd0 + v0 Vg ) + (4Rds
(10)
(11)
Figure 6 is plotted between duty ratio and sinusoidal output voltage. The output voltage v0 varies sinusoidally with an amplitude of 325 V. The duty Ratio also varies between 0 to 1 with the variation in the output voltage. The duty is multiplied by a factor of 100 to show it on the same scale. The variation in duty ratio is not very significant if non-idealities are not considered. The curves for duty ratio are drawn in Fig. 7 for a short period to show the variations if the non-idealities are considered.
AC Analysis on the Modified Johnson Converter
267
V oltage vs Duty Ratio
1 0.5 0 −0.5 Output V oltage/Vmax Duty Ratio
−1 0
0.5
1
1.5
2
2.5
3 3.5 Angle (rad)
4
4.5
5
5.5
6
Fig. 6. Duty ratio modulation
0.44
Ideal Duty Ratio N on − ideal Duty Ratio
Duty Ratios
0.43 0.42 0.41 0.4 0.39
0.6
0.8
1
1.2
1.4 1.6 1.8 Angle (rad)
2
2.2
2.4
2.6
Fig. 7. Ideal and non-ideal duty ratios
4
Simulation Results
A validation of the circuitry is conducted using MATLAB software using the simulation parameters as mentioned in Sect. 3. A maximum voltage of 325 V is taken as the amplitude of the output sinusoidal voltage waveform. The inductance and capacitance in the converter are designed using the design parameters such that the converter does not operate in discontinuous conduction mode. The values of inductance and capacitance are 20 µH and 10 µF.
268
S. Agrawal et al. 400
Voltage (V)
Voltage (V)
0 −200 −400 4.980
Inductor V oltage (V )
400
Output V oltage (V ) 200
200 0 −200
5.000 5.020 1.498 Time (s) ·10−3
1.500 Time (s)
−400 4.980
1.502 ×10−2
5.000 5.020 1.498 Time (s) ·10−3
1.500 Time (s)
1.502 ×10−2
(b) Inductor Voltage
(a) Output Voltage 400 V oltage across Switch 1 Voltage (V)
Voltage (V)
400 200 0 −200 4.980
V oltage across Switch 1 5.000 5.020 1.498 Time (s) ·10−3
1.500 Time (s)
0 −200 −400 4.980
1.502 ×10−2
(c) Voltage across switch group 1
V oltage across Switch 2
200
V oltage across Switch 2 5.000 5.020 1.498 Time (s) ·10−3
1.500 Time (s)
1.502 ×10−2
(d) Voltage across switch group 2 400
V oltage across Switch 3 Voltage (V)
Voltage (V)
400 200 0 −200 4.980
V oltage across Switch 3 5.000 5.020 1.498 Time (s) ·10−3
1.500 Time (s)
1.502 ×10−2
(e) Voltage across switch group 3
V oltage across Switch 4
200 0 −200 −400 4.980
V oltage across Switch 4 5.000 5.020 1.498 Time (s) ·10−3
1.500 Time (s)
1.502 ×10−2
(f) Voltage across switch group 4
Fig. 8. AC load simulation results
Figure 8 presents the simulation results for voltage across the output, the inductor, and all four switch groups. Both the positive and negative simulated waveforms are shown, so that the voltage across each switch can be displayed. The following points can be inferred from the simulation results. – The voltage across the load varies even in sub-cycles. However, the change in voltage is not significant. – The switches are turned on and off depending on the output voltage requirement. For the AC loads, eight numbers of switches are required. – Voltage across the inductor is dependent on the duty ratio and the output voltage requirements.
5
Comparison with Current-Fed Converter
It is evident from the literature survey [10] that the current-fed converter also plays a role of a dual converter. Hence, it can be used either as a high gain DC-DC converter or DC-AC inverter. A comparison can be made between the above-explained converter with current-fed converter.
AC Analysis on the Modified Johnson Converter
269
In the current-fed converter, the range of gain is more than one always. It means that it cannot provide an output voltage lesser than the input voltage. Hence, for DC-AC conversion, it will provide a certain amount of disturbances between (−Vg ) i.e., negative input voltage and (Vg ) i.e., positive input voltage. The output voltage waveforms are shown in Fig. 9(a) and (b), considering all the input and output parameters to be the same for both the converters.
Fig. 9. Comparison of dual converters
Total harmonic distortion (THD) is also calculated for the sinusoidal output voltage with an amplitude of 325 V and 50 Hz. For the current-fed converter, the THD is 8.70%, whereas, for the above-explained converter, the THD is 3.46%. It shows the significance of this converter over the current-fed converter when the inversion operation is requisite. For DC-DC conversion, the current-fed converter provides a higher gain as compared to the above-explained converter. Hence, there is a tradeoff in the performance of the converters, which depends on the operation required.
6
Conclusion
The outcome of the simulation study shows that the modified Johnson converter, can be realized as modified Johnson’s converter, this can be used for AC loads. Hence, it is felt the converter can be used either for DC-DC inversion and DC-AC conversion. The modified Johnson’s converter can have a gain ratio between 0 and 1, enabling buck operation with this converter. The maximum gain is limited by various non-idealities such as inductor’s DC resistance, switches’ on-state resistance, voltage drops, and parasitics. The scope of the paper is extended to the comparison of the modified Johnson’s converter with a current-fed converter. One of the observations, the Fast Fourier Transform (FFT) shows that the THD of the current-fed converter is higher than the modified Johnson’s converter, making it more suitable to feed the AC loads.
270
S. Agrawal et al.
References 1. https://mnre.gov.in/the-ministry/physical-progress. Accessed 18 Oct 2021 2. Rocabert, J., Luna, A., Blaabjerg, F., Rodr´ıguez, P.: Control of power converters in AC microgrids. IEEE Trans. Power Electron. 27, 4734–4749 (2012) 3. Umanand, L.: Power Electronics: Essentials and Applications, 1st edn. Wiley, India (2010) 4. Peng, F.Z.: Z-source inverter. IEEE Trans. Ind. Appl. 39, 504–510 (2003) 5. Peng, F.Z., Shen, M., Qian, Z.: Maximum boost control of the Z-source inverter. IEEE Trans. Power Electron. 20, 833–838 (2005) 6. Qian, W., Peng, F.Z., Cha, H.: Trans-Z-source inverters. IEEE Trans. Power Electron. 26, 3453–3463 (2011) 7. Florescu, A., Stocklosa, O., Teodorescu, M., Radoi, C., Stoichescu, D.A., Rosu, S.: The advantages, limitations and disadvantages of Z-source inverter. In: CAS 2010 Proceedings (International Semiconductor Conference) (2010) 8. Agrawal, S., Atmanandmaya, R.B.S., Umanand, L.: Integration of photovoltaic panels with DC grid using high gain DC-DC converter. In: 2020 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT) (2020) 9. Nag, S.S., Mishra, S.: Current-fed switched inverter. IEEE Trans. Ind. Electron. 61, 4680–4690 (2014) 10. Agrawal, S., Umanand, L., Subba Reddy, B.: Bidirectional current-fed converter for high gain DC–DC and DC–AC applications. In: Mohapatro, S., Kimball, J. (eds.) Proceedings of Symposium on Power Electronic and Renewable Energy Systems Control. LNEE, vol. 616, pp. 101–111. Springer, Singapore (2021). https://doi.org/ 10.1007/978-981-16-1978-6 9
Heating and Cooling of Thermoelectric Module Using Modified Johnson Converter Shubham Agrawal(B) , L. Umanand, B. Subba Reddy , and Atmanandmaya Indian Institute of Science, Bangalore 560012, Karnataka, India [email protected]
Abstract. In the present investigation, the application of the modified Johnson converter on a thermoelectric module is described. The electrical voltage and current are the thermoelectric module’s input, and output is the temperature difference between two surfaces. If the reverse voltage is applied to the thermoelectric module, the temperature of the surfaces is also reversed. The modified Johnson converter acts as a power conditioning block between the power source and thermoelectric module. The required operating voltage and current for the thermoelectric module can be obtained by modulating the duty ratio. A temperature gradient is developed for a duty ratio between 0 and 0.5. However, this temperature gradient is negative for the duty ratio between 0.5 and 1. Modeling the thermoelectric module LIARD ETX4-12-F1-4040-TA-RTW6 is done through mathematical equations. A detailed simulation study is done for heating and cooling applications, and some important results are presented. Keywords: Extended gain range · Modified johnson converter ratio modulation · Thermoelectric device
1
· Duty
Introduction
The application of power electronic converters is increasing in AC and DC loads. For a thermoelectric module, the power supply required is dependent on its internal resistor. Hence, either a voltage source or a power interface with a fixed voltage source is required. DC-DC converters play a vital role in interfacing a voltage source and thermoelectric module. A buck converter [1] Is used for power conditioning for automated cooling in a thermoelectric module; however, the voltage source should have more than the maximum operating voltage of the thermoelectric module. A boost or step-up converter [2] and switched capacitor converter [3] is used for energy harvesting through a thermoelectric module; however, it is limited to charging the battery with a higher nominal voltage. A buck-boost converter can be used as a power conditioner to have a wide range of voltage gain. If the gain is positive, then one surface of the thermoelectric module will be hot, and the other surface will be cold. The surfaces c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 271–278, 2022. https://doi.org/10.1007/978-981-19-1677-9_24
272
S. Agrawal et al.
temperatures can be reversed if the voltage polarity is reversed. However, a buck-boost converter has a limitation on the output voltage polarity. The Current fed converter [4,5] has voltage gain in both positive and negative ranges; however, it can only boost the input voltage. The modified Johnson converter is shown in Fig. 1, which can either reduce or amplify the input voltage. The modified Johnson converter is used as a power interface unit between battery and a thermoelectric module in the present study.
S1 ig
Vg
vL D1
D2
iL
L
S2 C
iC
R
i0
v0
Fig. 1. Modified Johnson converter
The motivation of the present investigation is to find the pertinency of modified Johnson converter space and heating/cooling application where the power supply of both negative and positive voltage is desirable as the surfaces’ temperatures can be switched according to the requirement. In Sect. 2, modeling of a Peltier module [6] using the literature [7–10]. Sections 3 and 4 describe the cooling and heating applications of the Peltier device along with simulation results. Lastly, in Sect. 5, the important conclusions are presented.
2
Modeling of a Peltier Module
R J1
T1
Vg J2
T2
Fig. 2. Junction of two metals and temperature gradient
Heating and Cooling of Thermoelectric Module
273
Whenever a current flows through the junctions of dissimilar materials, then a temperature difference establishes between the junctions. The direction of the current determines whether a junction is positive or negative. Figure 2 shows that there are two metal junctions in the circuit path, and a temperature difference ΔT = T1 − T2 develops between these. If the voltage polarity of the source Vg reverses, ΔT also reverses. A reverse of the above condition is also possible; If there is a temperature gradient between two surfaces, a voltage can be generated through it, called the Seebeck effect. The heat absorbed or generated (Q) is proportional to the current (i). Equation (1) gives the relation between heat and current. Q = Πi
(1)
where, Π is called the Peltier coefficient. The modern thermoelectric modules comprise metal and semiconductors as two substances in which a temperature difference is created. Figure 3 shows a model of a commercial thermoelectric module.
surface - 1 (T1) P
N
N
P
ceramic Cu
surface - 2 (T2) Vg
R Fig. 3. Detailed Peltier mudule
2.1
Peltier Module as Heat Pump
Figure 4 shows the Peltier module working as a heat pump. When a voltage is applied across a Peltier module, heat transfer occurs between the two surfaces. A certain amount of heat (QC ) is extracted from the surface - 1, which becomes cold and transfers heat (QH ) to another surface - 2, which becomes hot. The relation between different energy extracted is mentioned in Eq. 2. QC = QH + E
(2)
274
S. Agrawal et al.
hot surface (T1) QH iP R Peltier R vP E P Module QC cold surface (T2)
Vg
Fig. 4. Peltier device as heat pump
2.2
Interface of Peltier and DC-DC Converter
The Peltier resistance RP and voltage required vP is dependent on the QC and the temperature difference between the two surfaces ΔT . Hence, a variable voltage and variable power are required at the ends of the Peltier module. A DCDC converter with a variable duty ratio can supply such variability in power and voltage. Figure 5 shows the power converter interface with the Peltier module.
hot surface (T1) QH iP Peltier R DC-DC vP P Module Converter E QC cold surface (T2)
Vg
Fig. 5. Interface of Peltier device and DC-DC converter
The equations for heat extracted QC and heat supplied QH are dependent on heat transfer phenomena such as convection and conduction. Equations (3) and (4) show the dependency of various factors on QC and QH [7]. QC = αm TC I −
ΔT I 2 Rm − θm 2
(3)
QH = αm TH I −
ΔT I 2 Rm + θm 2
(4)
Heating and Cooling of Thermoelectric Module
275
Here, αm corresponds to specific material used in the Peltier module, called as Seebeck coefficient, θm is the total thermal resistance, and Rm is the total resistance of the thermoelectric module. The subscript m shows the maximum value of the variables. The maximum value is taken as many submodules are connected in series depending on the power rating of the thermoelectric module. The equations for αm , θ, and Rm are as follows in Eq. (5)–(7). Vmax TH
(5)
2TH ΔTmax Imax Vmax (TH − ΔTmax )
(6)
Vmax (TH − ΔTmax ) Imax TH
(7)
αm = θm =
Rm =
Vmax , Imax are the maximum values of voltage and current corresponding to the maximum temperature difference ΔTmax between hot surface TH and cold surface TC . These values corresponds to the fixed hot surface temperature TH , can be calculated from the datasheet given [6]. A series of curves can be obtained between QC and the current I for different temperature differences ΔT . The values for different variables are calculated with consideration of hot surface temperature TH to be 85 ◦ C.
Fig. 6. Variation of energy with current supplied
Figure 6 shows that a higher amount of energy can be transferred for lower ΔT . As The curves shift towards the higher current, the slope of the curves decreases, and after a certain current, the curves become flat.
276
3
S. Agrawal et al.
Cooling Application of the Peltier Module
If the energy is provided to the thermoelectric module, it heats one surface and cools another surface. For simulation studies, the datasheet of the Liard thermoelectric module [6] is taken for reference. Parameters such as hot surface temperature TH , the difference of two surfaces’ temperature ΔT , and energy that needs to be extracted from the cold surface QC are given. Using Eqs. (3)– (7), effective Peltier resistance RP and voltage at terminals v0 can be calculated. Further, with the input voltage Vin , duty ratio d is calculated. Figure 7 shows the block diagram for the calculation required.
eq
(3 )-( 7)
TH (fixed) ΔT QC
RP v0
v0 2v0+Vin
d
Vin Fig. 7. Modeling of Peltier resistance and duty ratio
The calculation for V0 , RP and d is done using the specifications mentioned in Table 1. Table 1. Simulation specifications Parameters
Specifications
Hot surface temperature TH
85 ◦ C
Maximum temperature difference ΔTmax 95.3 ◦ C Maximum voltage Vmax
19.1 V
Maximum current Imax
3.6 A
Input voltage Vin
12 V
The curves for resistance and duty ratios for various ΔT are mentioned in Fig. 8(a) and (b). It shows that a larger voltage and duty ratio is required if the heat extraction from the cold surface increases; however, the resistance decreases, which implies that the current requirement increases.
Heating and Cooling of Thermoelectric Module
Resistance (Ω)
6 5.5 5 4.5
0.4
Duty Ratio
ΔT = 0◦ C ΔT = 10◦ C ΔT = 20◦ C ΔT = 30◦ C ΔT = 40◦ C ΔT = 50◦ C ΔT = 60◦ C ΔT = 70◦ C
6.5
0.3
ΔT = 0◦ C ΔT = 10◦ C ΔT = 20◦ C ΔT = 30◦ C ΔT = 40◦ C ΔT = 50◦ C ΔT = 60◦ C ΔT = 70◦ C
0.2 0.1
5
10
15
20 QC (W atts)
25
30
277
35
5
10
(a) Peltier resistance
15
20 QC (W atts)
25
30
35
(b) Duty ratio
Fig. 8. Cooling application for Peltier module
4
Heating Application of the Peltier Module
In the heating application, the cold side of the thermoelectric module is kept at a fixed temperature while external energy is given to heat the other surface. The simulation parameters are the same; however, the voltage polarity of the external source is reversed. The thermoelectric module continues to act as a sink only, but the voltage polarity is reversed. The same thermoelectric module is used for heating which was used for cooling applications earlier. Figure 9(a) shows the resistance offered by the Peltier module, and Fig. 9(b) shows the required duty ratio for various ΔT . ΔT = 0◦ C ΔT = 10◦ C ΔT = 20◦ C ΔT = 30◦ C ΔT = 40◦ C ΔT = 50◦ C ΔT = 60◦ C ΔT = 70◦ C
4.4 4.3 4.2
0.6 0.59 0.59 0.58 0.58
4.1 4
ΔT = 0◦ C ΔT = 10◦ C ΔT = 20◦ C ΔT = 30◦ C ΔT = 40◦ C ΔT = 50◦ C ΔT = 60◦ C ΔT = 70◦ C
0.6 Duty Ratio
Resistance (Ω)
4.5
0.57
10
20
30
40 QH (W atts)
50
(a) Peltier resistance
60
70
10
20
30
40 QH (W atts)
50
60
70
(b) Duty ratio
Fig. 9. Heating application for Peltier module
It can be inferred from the curves that the resistance and duty ratio does not vary much with the variation in QH , which implies that the efficiency of the thermoelectric module is higher for higher QH .
5
Conclusion
The presented study shows that a modified Johnson converter is suited well to develop a temperature gradient between two surfaces of a Peltier device. The modified Johnson converter reduces or amplifies the battery’s input voltage depending on the temperature gradient and energy that needs to be transferred.
278
S. Agrawal et al.
Various curves for Peltier resistance and duty ratios are shown to validate the current investigation. Such systems are useful in space or heating/cooling applications where surface temperature changes are frequent.
References 1. Mihalache, S., Flamaropol, I., Dumitru, F.-S., Dobrescu, L., Dobrescu, D.: Automated cooling control system through Peltier effect and high efficiency control using a DC-DC Buck converter. In: 2015 International Semiconductor Conference (CAS) (2015) 2. Jurkans, V., Blums, J., Gornevs, I.: Harvesting electrical power from body heat using low voltage step-up converters with thermoelectric generators. In: 2018 16th Biennial Baltic Electronics Conference (BEC) (2018) 3. Marusa, L., Milanovic, M., Valderrama-Blavi, H.: Evaluating a switched capacitorboost converter (SC-BC) for energy harvesting in a Peltier-cells thermoelectric system. In: 2015 International Conference on Electrical Drives and Power Electronics (EDPE) (2015) 4. Agrawal, S., Atmanandmaya, Reddy, B.S., Umanand, L.: Integration of photovoltaic panels with DC grid using high gain DC-DC converter. In: 2020 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT) (2020) 5. Agrawal, S., Umanand, L., Subba Reddy, B.: Bidirectional current-fed converter for high gain DC–DC and DC–AC applications. In: Mohapatro, S., Kimball, J. (eds.) Proceedings of Symposium on Power Electronic and Renewable Energy Systems Control. LNEE, vol. 616, pp. 101–111. Springer, Singapore (2021). https://doi.org/ 10.1007/978-981-16-1978-6 9 6. https://www.lairdthermal.com/products/thermoelectric-cooler-modules/peltierhitemp-etx-series/ETX4-12-F1-4040-TA-RT-W6/pdf/ . Accessed 28 Oct 2021 7. Lineykin, S., Ben-Yaakov, S.: Modeling and analysis of thermoelectric modules. IEEE Trans. Ind. Applicat. 43, 505–512 (2007) 8. Lineykin, S., Ben-Yaakov, S.: Analysis of thermoelectric coolers by a spicecompatible equivalent-circuit model. IEEE Power Electron. Lett. 3, 63–66 (2005) 9. Lineykin, S.B., Ben-Yaakov, S.: PSPICE-compatible equivalent circuit of thermoelectric cooler. In: IEEE 36th Conference on Power Electronics Specialists (2005) 10. Lineykin, S., Ben-Yaakov, S.: SPICE compatible equivalent circuit of the energy conversion processes in thermoelectric modules. In: 2004 23rd IEEE Convention of Electrical and Electronics Engineers in Israel (2004)
Hate Detection for Social Media Text with User Alert System Jose Ashley(B) , Nefi Nisen, Riyona Lasrado, and Mukta Nivelkar Faculty of Information Technology, Fr. Conceicao Rodrigues Institute of Technology, Navi Mumbai, India {jose.ashley,nefi.mery,riyona.judie,mukta.nivelkar}@fcrit.ac.in
Abstract. Social media is an important factor to build social networks and discover various issues going around the world. However, the increase in the use of social media has led people to engage in spreading and sharing hate speech as well. On social media, people may act more aggressively as they can be anonymous. Hate speech not only causes psychological harm to an individual but may also lead to any physical harm when it incites violence. With the help of advanced technology like Machine learning, many researchers are working on this factor specially to solve the emerging issue of hate speech being used on social media. Hate speech is considered offensive in many countries and there are times when people are arrested for the same. Many users may not be in the right state of mind while interacting on social media so they might make the wrong choice of words. A good way to stay out of such situations is to be cautious and review the content of the comment before posting. The main objective of our model is to build a system using a LTSM-based classification system which is a machine learning approach for identifying and classifying hateful comments and alerting the user from posting any abusive comments on social media. We are also using Bi-LSTM in our model to classify hate which gives about 0.9084 accuracy. Here the user can make his own choice of words depending upon the alert message he gets for his comment. Keywords: Hate speech · Natural Language Processing (NLP) · Bidirectional long short term memory (Bi-LSTM)
1 Introduction Social media over the years have been the most popular communication media around the world where users can connect with a number of people and express their opinions. People can act more aggressively in social media as it can be anonymous by using abusive language and hatred to criticize or win an argument. Hate speech can be defined as the way of expressing violence and discrimination against an individual or a group of people for different reasons and it’s also difficult to control and track such spread of content as an unmanageable number of posts are posted within seconds on such platforms. Tracking down hate speech can also be challenging because of the ambiguity of the language in which a single word or sentence can have multiple meanings. Due to such negativity, © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 279–289, 2022. https://doi.org/10.1007/978-981-19-1677-9_25
280
J. Ashley et al.
many people get stressed out and face issues related to mental health which is a serious problem. This motivated us to develop this model where we would identify and classify the hateful comments posted on Twitter from the dataset will be classified as ‘hate’ and ‘not hate’ which will help in classifying the comments on three classes (−1,0,1) which indicates its type of profanity. Further, alerting the user from posting any abusive comments on any social networks with the aspiration of decreasing the negativity on the online platform.
2 Literature Survey Many researchers working on hate detection noticed that the datasets used were mostly imbalanced which led to misclassification of texts. Recently, Chayan Paul et al. [1] worked on a model which used sampling methods to balance the dataset and implemented two different models. One was the Long Short Term Memory (LSTM) and the other was Bi-directional Long Short Term Memory (Bi-LSTM). The authors found out that LSTM had better accuracy for their dataset with good precision and F1 score while Bi-LSTM showed good recall value. The accuracy for the LSTM model was about 0.9598 while for Bi-LSTM it was about 0.9282. Later many authors started working on many neural network models as they gave better accuracy compared to the previously used models. Raghad Alshalan et al. [2] worked on a system which was used to detect hate speech in Arabic tweets. Here the authors built their own dataset which contained about 9316 annotated tweets. They built four models namely CNN, Gated Recurrent Units (GRU), CNN + GRU and BERT but CNN model gave better results for their dataset compared to the other three models with 0.79 F1 score. Though the neural network models gave better accuracy there were some limitations noticed by the authors which included misclassification of texts as the dataset was constructed by the authors themselves. They needed enormous data for better classification. Juan Carlos Pereira-Kohatus et al. [3] worked on a model which was a combination of LSTM + MLP neural network, it not only detects hate but also monitors and visualizes the hate texts in social media. They obtained a precision of about 0.784 but one of the major drawbacks here was the highly imbalanced dataset along with manual tagging of tweets by the authors which led to errors while classification of tweets. Though BOW based approaches showed better accuracy in detection of hate compared to other models, it did have its own limitations. To overcome this many authors started working on N-gram approaches. Priya Rani et al. [4] developed a detection tool which was able to detect offensive language, developed by KMI_Coling under the OffensEval Shared task. Three different systems were trained which were named as A, B and C and obtained an accuracy of about 79.76%, 87.91% and 44.37% respectively which was within the OfensEval task. This method was simple and scalable with large n values but the increase in the n values led to decrease in the processing speed. Shanita Biere et al. [5] worked on a deep learning model, mainly a Convolutional Neural Network model where the classifier assigns each tweet to a particular category which is hate, offensive or neither. This approach was very flexible and obtained an accuracy of 91% with high accuracy but when the classes of tweets were studied separately
Hate Detection for Social Media Text with User Alert System
281
it was noted that lots of tweets were misclassified. Sebastian Köffer et al. [6] worked on a system using NLP techniques to detect the hate speech on German texts where the dataset was labelled as hate and not hate using crowdsourcing approach with a combination of Word2Vec and Extended 2-g approach. This method gave 70% accuracy with improved classification of hate but one of the biggest shortcomings is that the authors used a very small set of data and there was a need for enormous data. Later studies found out that neural network approaches overperforms the existing models which were used for the classification of hate. Shanita Biere et al. [5] worked on a deep learning model, mainly a Convolutional Neural Network model where the classifier assigns each tweet to a particular category which is hate, offensive or neither. This approach was very flexible and obtained an accuracy of 91% with high accuracy but when the classes of tweets were studied separately it was noted that lots of tweets were misclassified.
3 Existing System Many people have worked on hate detection but not everyone has made any existing system on it. One such existing system was made by GedeManggala Putra and Dade Nurjanah where they created a model, Hate Speech Detection In Indonesian Language Instagram [7] which is used to identify hate in Instagram comments by collecting all comments from Indonesian Instagram from the year of June 2019 to April 2020. They divided the dataset into two different classes i.e. “Hate” and “Not hate” had a ratio of 80% and 20% of data which was used to train and test the model respectively. They used the word2vec with window size 15 method with skip-gram models and a modified TextCNN with random sampling methods which obtained an accuracy of about 93.70%. Later it was found out that the randomness of data was oversampled and the author didn’t tune the hyperparameter leading to changes in the oversampled data when the system is re-run.
4 Proposed System The proposed system identifies the hate sentiments from the tweets and classifies them as per the classes in the dataset that is “hate speech” and “not hate speech”, this approach will help in classifying the comments on a scale of -1 to 1 i.e. negative, neither and positive which indicates its type of profanity. The model will be using an LTSM-based classification system which helps us to categorise between hate and offensive language. This system will employ word embeddings with normal LSTM and neural networks for the identification of hate on Twitter. Due to the drawback of LSTM we are using Bidirectional LSTM which has the ability to remember information from both directions which is important for analysing text data.
282
J. Ashley et al.
Fig. 1. Block diagram of the model
Figure 1 shows the block diagram of the system. The components of this system are as follows Module 1 Classifying Tweets in 3 Classes Here in the first module, we have classified the tweets into three different classes which are negative, neutral, and positive. This is done by taking a note of how many people have judged the tweet to determine the sentiment, any negative comments if used, and the confidence about the judgment. Text Pre-processing Here we convert the input text into the more easy form so that it would be easy to perform different machine learning algorithms. This involves removal of the URL, hashtags, stopwords, whitespaces, punctuations, numbers, along with the process of stemming, etc. Module 2 Detection of Hate Text In this module, we train our model so it can detect hate texts. Hence, it detects hate texts every time when the user posts a comment on social media. Module 3 Alerting Users Before Posting a Tweet This part of the system will alert the user every time he/she posts a comment on social
Hate Detection for Social Media Text with User Alert System
283
media. Depending upon the content of the user’s comment an alert message will be displayed if the user’s comment is considered offensive or not. This alert message will be displayed to the user before he/she posts a comment. Module 4 Labeling Displayed Tweets as Sensitive Data Depending upon the previous module, the user is given a choice to post the comment or not. If he/she decides to post the comment even after the alert message was displayed saying the content of his/her comment is considered offensive, then the comment posted by the user will be posted on social media but it will be labeled as sensitive data.
Fig. 2. Flowchart of the model
Figure 2 shows the flowchart of our system. At the beginning the user first logs into the system. The user then writes a comment to post it on social media. If the content of the comment is considered as hateful content then he/she is alerted about it before it is posted but if the content is not considered as hateful content then he/she isn’t alerted about it and the comment is directly posted in the social media. Now when the user is alerted about the content of the tweet, he/she is given a choice to rewrite the tweet or either post it. If the user wishes to post the tweet instead of rewriting it, then the tweet is labeled as sensitive data.
5 Methodology In this model, we have first classified the tweets into three different classes −1, 0, 1 which indicate negative, neutral, and positive respectively. To avoid the issue of an imbalanced dataset we used the Exploratory Data Analysis (EDA) process. Data Analysis basically
284
J. Ashley et al.
comes into the picture when statistics and probability are used in the dataset to figure out different trends. In EDA the datasets are loaded in the Pandas data frame and later it is read by using read_csv() function provided by the Pandas library. The method gave us a balanced dataset compared to the datasets used by the previous researchers. We then did the part of preprocessing where we cleaned the dataset by removing all stopwords, whitespaces, hashtags, numbers, URLs, etc. This was done by importing different packages like Pandas, NumPy, NLTK, etc. Then we converted them to Word2Vec format for word embedding. The system is built using the LSTM-based classification which helps us to categorise between hate and offensive language. Bidirectional LSTM is used for our model.
6 Implementation 6.1 Dataset Preparation As noticed from the previous studies it is clear that the detection of hate text requires an enormous amount of data to prevent any misclassification of data. But it was not easy to obtain such a large amount of data. From the literature survey, we can see that the authors always faced problems detecting hate text because of the dataset. In our model initially, we faced the same issue. The dataset we used was biased. In order to balance the dataset, we used the Exploratory Data Analysis (EDA) process. By using EDA we have combined different datasets together to obtain a single one. After which we were able to obtain an unbiased dataset compared to the previous ones which we used. In the below Fig. 3 it shows the category −1, 0 and 1 indicates negative, neutral, and positive respectively. Below given Fig. 4 is the final dataset put together. But it requires further preprocessing before feeding to the model. 6.2 Data Preprocessing It is a process of converting the raw data into clean data because every time the enormous amount is collected it is not much feasible as the data collected is in a raw format. In Data Preprocessing we have removed whitespaces, URLs, stopwords, removed the characters that are not letters (regardless of case) and numbers, and applied tokenization, stemming, and converted the tweets into lowercase by importing the required libraries for it in order to convert the raw text into clean data. 6.3 Feature Extraction It is a process that converts the text data into numerical vectors. We have done feature extraction for our model by using the Bag of Words approach and Tf-IDF approach wherein the TF will be increased if a token is observed to occur frequently in a document but its IDF is reduced if it occurs frequently in the majority of the documents. Thus the stopwords are penalized while the words which occur frequently get a boost [8].
Hate Detection for Social Media Text with User Alert System
Fig. 3. Balanced dataset
Fig. 4. Final dataset
285
286
J. Ashley et al.
6.4 Tokenization and Padding Tokenization is performed by splitting the texts into tokens. Padding is done to make all these tokens fit into a standard length. The process of padding is done with the help of Keras. 6.5 Bidirectional LSTM We have used Bidirectional LSTM as it can remember the information from both directions compared to LSTM. Figure 5 shows the summary of the given model.
Fig. 5. Summary of the model
6.6 Model Accuracy and Loss Accuracy of the model determines how accurately the model has made predictions as compared with the true data used by it while loss of the model determines how poorly or well it behaves after each iteration of its optimization [9] (Fig. 6).
7 Result In our system, we have first imported the dataset and balanced the dataset using Exploratory Data Analysis (EDA). Then data cleaning is performed which involves removal of the URL, hashtags, stopwords, whitespaces, punctuations, numbers, along with the process of stemming, etc. followed by the Word2Vec model to get the word embeddings. We have used the Bidirectional LSTM approach that uses NN which differentiates between hate speech and not hate and gives us higher accuracy. This approach
Hate Detection for Social Media Text with User Alert System
287
Fig. 6. Model accuracy and loss graph
gave us an Accuracy, Precision, Recall, and F1 score of about 0.9084, 0.9126, 0.9037, and 0.9081 respectively. Our model was able to classify the tweets depending upon negative, neutral, and positive categories as shown in Fig. 7.
Fig. 7. Classification of tweets
The confusion matrix of the model is shown in Fig. 8 which gives the results of 3 classes that are used. This figure tells the number of texts that are predicted correctly or incorrectly for each of the classes.
288
J. Ashley et al.
Fig. 8. Confusion matrix
8 Conclusions Social platforms over the years have been the most active communication channel for people around the globe. People act more aggressively as their identities are anonymous. The use of abusive language and hatred to criticize or win an argument is often observed and it’s a challenge to control and track the spread of such content. Therefore, this system will be classifying and identifying the amount of hate in those comments and then alert the users from posting them on social networks with the aspiration of decreasing the negativity on the online platform.
References 1. Paul, C., Bora, P.: Detecting hate speech using deep learning techniques. Int. J. Adv. Comput. Sci. Appl. (IJACSA), 12(2) (2021) 2. Al-Khalifa, R.A.H.: A deep learning approach for automatic hate speech detection in the Saudi Twittersphere. Appl. Sci. 2020(10), 8614 (2020) 3. Pereira-Kohatus, J.C., Quijano-Sánchez, L., Liberatore, F., Camacho-Collados, M.: Detecting and monitoring hate speech in Twitter. Sensors, 19, 4654 (2019) 4. Rawat, R., Mahor, V., Chirgaiya, S., Shaw, R.N., Ghosh, A.: Sentiment analysis at online social network for cyber-malicious post reviews using machine learning techniques. In: Bansal, J.C., Paprzycki, M., Bianchini, M., Das, S. (eds.) Computationally Intelligent Systems and their Applications. SCI, vol. 950, pp. 113–130. Springer, Singapore (2021). https://doi.org/10.1007/ 978-981-16-0407-2_9 5. Biere, S.: Hate Speech Detection Using Natural Language Processing Techniques. Vrije Universiteit Amsterdam (2018) 6. Gautam, J., Atrey, M., Malsa, N., Balyan, A., Shaw, R.N., Ghosh, A.: Twitter data sentiment analysis using naive bayes classifier and generation of heat map for analyzing intensity geographically. In: Bansal, J.C., Fung, L.C.C., Simic, M., Ghosh, A. (eds.) Advances in Applications of Data-Driven Computing. AISC, vol. 1319, pp. 129–139. Springer, Singapore (2021). https://doi.org/10.1007/978-981-33-6919-1_10 7. Putra, G., Nurjanah, D.: Hate speech detection in indonesian language Instagram. In: International Conference on Advanced Computer Science and Information Systems (ICACSIS) (2020)
Hate Detection for Social Media Text with User Alert System
289
8. Towards Data science - Document feature extraction and classification 9. Intellipaat Community - How to interpret “loss” and “accuracy” for a machine learning model
Dual-Band Antenna Design at Sub-6 GHz and Millimeter Wave Band for 5G Application Manish Kumar Dabhade(B) and Krishna K. Warhade School of Electronics and Communication Engineering, Dr. Vishwanath Karad MIT World Peace University, Pune, India [email protected], [email protected]
Abstract. Future 5G communication technology usage will focus on the coexistence of microwave (i.e., sub-6 GHz) and millimeter wave antenna design on a single layer. In this paper, a novel dual band antenna design has been proposed supporting sub-6 GHz (resonance at 3.5 GHz) and millimeter wave (resonance at 28 GHz) operation with single feed. Initial design of antenna was adopted using cavity model. Proposed novel antenna design consists of slots at 3.5 GHz and recessed microstrip for 28 GHz band. Parametric study has been done to achieve the compactness and improve the performance parameters of the antenna. Proposed design is modelled and simulated on Rogers RT 5880 substrate with relative permittivity of 2.2. The simulated antenna design offers the efficiency and gain of 87% and 1.8 dBi at 3.5 GHz respectively, while efficiency and gain of around 76% and 7.69 dBi at 28 GHz respectively. In addition to the stable antenna parameters, the proposed antenna design maintains the frequency ratio of 8 with the compact size of around 17.51 × 20 × 0.508 mm3 . The design is kept simple and low profile and will offer ease of fabrication. Keywords: Frequency ratio · Current density · Cavity model · Millimeter wave · CST · 3GPP · e-MBB · ITU-R
1 Introduction Smartphones to be used in future for millimeter wave 5G applications are bound to adopt compact antenna design and will be an essential aspect for improving the transmission power and battery life. The main challenge for the researchers in antenna design domain lies in the fact that the millimeter wave frequency experiences enhanced free space path losses and short transmission distances. Extensive research work has been carried out in recent years towards millimeter wave technology implementation in deploying 5G communication. Apart from high end technological applications like internet of things (IoT) millimeter wave band will have the advantage of higher data rates and very low latency maintaining the size compactness. Third generation partnership project (3GPP) releases the standards for 5G, time to time, which enables the researchers to contribute to the technological improvement. Currently content of release 18, priorities and timeline are moving to discussion and decision phase. The prioritization process on release 18 © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 290–298, 2022. https://doi.org/10.1007/978-981-19-1677-9_26
Dual-Band Antenna Design at Sub-6 GHz and Millimeter Wave Band
291
should be completed by year 2021 end [3]. International Telecommunication Union-R has broadly classified usable 5G spectrum in two categories as FR1 (for the usability in Sub-6 GHz) and FR2 (for the usability above 6 GHz) [1, 2]. The most important key element of 5G communication is enhanced mobile broadband (e-MBB) services which provides massive connectivity among people, among machines and also between people and machines. To achieve data rate requirements, and adapt allocated frequency spectrum, 5G networks are designed to have triple layer architecture [4]. The first layer of ultra-experience will exploit the millimeter wave spectrum for providing enhanced data rates at urban areas in the form of hotspots [4]. The second high-capacity layer will ensure wide coverage using sub-6 GHz spectrum [4]. The third, ubiquitous coverage layer will implement wide and in-depth radio coverage assuring universal coverage in UHF band [4]. Multiband antenna is the inevitable aspect of current trend of technology in order to achieve the intended coverage and user experiences in 5G wireless communication. This paper proposes a novel dual band low profile microstrip antenna designed for sub-6 GHz (3.5 GHz) and millimeter wave band (28 GHz). Dual band antennas are relatively simple in structure, occupy less space with minimum complexity and more importantly fed with single port. In order to achieve dual band operation employing microwave and millimeter wave having large frequency ratio, the proposed antenna has a combined structure of two different antenna designs on Rogers RT 5880 substrate. This arrangement makes the complete antenna design flexible while maintaining independent resonances.
2 Antenna Design The compact dual band antenna design operating at sub-6 GHz and millimeter wave frequency band is simulated in Microwave Studio Environment of simulator software known as Computer Simulation Software (CST). The antenna design is initially implemented using cavity model and further optimized using parametric analysis. Dimensions are calculated as per the desired resonant frequencies. Then parametric analysis is done to achieve the optimum antenna parameters keeping in view the size compactness and design simplicity. Antenna maintains the frequency ratio of 8 and provides stable antenna parameters. Plot of resultant surface current distribution, achieved gain, efficiency, return loss and VSWR has been shown in the proceeding sub section. The achieved antenna parameters are tabulated which reflects the antenna design as a potential candidate for supporting future multiband operation for 5G wireless communication. 2.1 Cavity Model Design The area between patch and ground is considered as cavity and exhibit higher order resonances in the cavity model. Due to very thin substrate, the fields inside cavity are uniform along the thickness of the substrate [5]. The resonant frequencies are examined to determine the dominant mode. For a microstrip antenna h L and h W. If L > W > h, the mode with lowest frequency (called the dominant mode) is TMx 010 whose resonant frequency is given by relation [6] (fr )010 =
ϑ0 1 = √ √ 2L με 2L μr
(1)
292
M. K. Dabhade and K. K. Warhade
Where ϑ0 is the speed of light in free space. Inset feed point can be calculated using the equation [6] (Fig. 1) π y0 Rin (y = y0 ) = Rin (y = 0)cos2 L
(2)
Fig. 1. (a) Field Configuration TMx 010 (b) Rectangular microstrip patch geometry [6]
Fig. 2. Cavity model design on CST with dimensions (a) Top view (b) Bottom view
The conventional antenna design is carried out on Rogers RT 5880 (εr = 2.2, tan δ = 0.0009, h = 0.508) substrate with copper conductor on patch and ground as shown in Fig. 2. Return loss, efficiency, E field distribution and gain are plotted on the simulator as a reference for further design (Fig. 3). Simulated conventional cavity model microstrip antenna extracted parameters were efficiency of around 88%, return loss of −14 dB and gain of around 7.5 dBi. However, further design optimization is carried out in order to achieve better stable results with reduced size.
Dual-Band Antenna Design at Sub-6 GHz and Millimeter Wave Band
293
Fig. 3. Antenna parameters plot of conventional cavity model antenna (a) Efficiency (b) Return Loss (c) E Field (d) Gain.
2.2 Miniaturized Sub-6 GHz Antenna Design and Simulation Results The conventional antenna has been made compact by increasing the effective electrical length of the antenna using slots cut in the antenna structure. The resultant miniaturized antenna is designed and simulated at 3.5 GHz. Figure 2(a) depicts a conventional antenna with a patch size of 27.80 × 33 mm2 , whereas Fig. 4 depicts a miniaturised antenna with a patch size of 6 × 16 mm2 for the 5G sub-6 GHz application. When compared to the conventional antenna simulated at 3.5 GHz frequency, the overall patch area is reduced by 89.53% by employing slots cut (L1 and L2). The parametric investigation of L1 and L2 is shown in Fig. 5(a) and 5(b), while the distribution of surface current is shown in Fig. 6. With the increase in length of slot, the increase in the surface current is observed along with the patch length. Parametric analysis helps to measure the best length of the slots L1 and L2 which comes out to be 4 mm and 15.5 mm respectively. The return loss (S11 ), efficiency, voltage standing wave ratio and gain are illustrated in the respective plots in Fig. 7. At 3.5 GHz, a return loss of −32 dB, efficiency of 86.8%, VSWR of 1.04 and resultant gain (1.8 dBi) plot are shown in this paper.
294
M. K. Dabhade and K. K. Warhade
Fig. 4. Miniaturised antenna design at 3.5 GHz
Fig. 5. Parametric analysis results for (a) L1 Length and (b) L2 Length
Fig. 6. Resultant surface current distribution at 3.5 GHz
2.3 Millimeter Wave Antenna Design and Simulation Results For achieving the desired resonance at millimeter wave band at 28 GHz, by using cavity model, antenna design parameters are calculated using Eq. 1. A radiator, to be efficient, practical width which leads to good radiation efficiency can be calculated by using equation mentioned below [6] W =
c √ 2(fr ) (εr + 1)/2
(3)
Dual-Band Antenna Design at Sub-6 GHz and Millimeter Wave Band
295
Fig. 7. Simulated parameters of miniaturized antenna at 3.5 GHz. (a) Return loss of 32 dB (b) Efficiency of 86.8 (c) VSWR of 1.04 (d) Gain plot of 1.8 dBi
Where, f r is the frequency at dominant mode, Er is the relative permittivity of the substrate. The antenna design at 28 GHz with dimension is shown in Fig. 8. The design was simulated in CST microwave studio simulator and the resultant parameters are depicted in Fig. 9.
Fig. 8. Antenna design at 28 GHz (a) Top View (b) Bottom view
The intended low-profile design shows the promising parameter as return loss of around −21 dB, efficiency of around 77%, VSWR of 1.2 and gain of 7.7 dBi. The distribution of surface current is shown in Fig. 10 below.
296
M. K. Dabhade and K. K. Warhade
Fig. 9. Simulated parameters of antenna design at 28 GHz (a) Return loss (b) Efficiency (c) VSWR (d) Gain
Fig. 10. Surface current distribution at 28 GHz antenna design
2.4 Dual-Band Antenna Design and Simulation Results Finally, both the antenna designs discussed in Sect. 2.2 and 2.3 has been combined on a single substrate layer. Resultant dual band antenna has the co-axial feed to microwave band design and the insert gap feed for the millimeter wave band design. Figure 11 shows the overall design, return loss and frequency bandwidth of the proposed antenna.
Dual-Band Antenna Design at Sub-6 GHz and Millimeter Wave Band
297
Fig. 11. Dual-band antenna design at microwave and millimeter wave frequency (a) Top view (b) bottom view (c) Return Loss and Frequency BW
3 Conclusion Simulated novel antenna design presented in this paper is a dual band at microwave (sub6 GHz) and millimeter wave band is a combination of two different designs on a single substrate. The proposed simulated antenna is designed on Rogers RT 5880 substrate with overall size of 17.51 × 20 × 0.508 mm3 . Parametric analysis and optimization technique aided in reduction in antenna size. With the coaxial feed at microwave band design and insert gap feed at millimeter wave band, the antenna provides the promising stable parameters at intended bands of operation and maintain large frequency ratio. In addition to the mentioned simulated parameters in previous sections, proposed antenna also offers the frequency bandwidth of around 69 MHz and 1.2 GHz at 3.5 GHz and 28 GHz respectively as shown in Fig. 11. Due to its compact size, less complexity and stable parameters, the proposed antenna design is suitable for multiband 5G applications.
References 1. Requirements, Evaluation Criteria and Submission Templates for the Development of IMT2020, document Report ITU-R M.2411-0, pp. 1–32 (2017)
298
M. K. Dabhade and K. K. Warhade
2. Zhao, A., Ren, Z.: Wideband MIMO antenna systems based on coupled-loop antenna for 5G N77/N78/N79 applications in mobile terminals. IEEE Access 7, 93761–93771 (2019) 3. https://www.3gpp.org/release18/ 4. 5G-New-Antenna-White paper-V2, Huawei 5. Kumar, G., Ray, K.P.: Broadband Microstrip Antennas. Artech House (2003) 6. Balanis, C.A.: Antenna Theory: Analysis and Design, 4th edn. Wiley, Hoboken (2016) 7. Guo, Y.Q., Pan, Y.M., Zheng, S.Y., Lu, K.: A singly-fed dual-band microstrip antenna for microwave and millimeter-wave applications in 5G wireless communication. IEEE Trans. Veh. Technol. (2021). https://doi.org/10.1109/TVT.2021.3070807 8. Zada, M., Shah, I.A., Yoo, H.: Integration of sub-6-GHz and mm-wave bands with a large frequency ratio for future 5G MIMO applications. IEEE Access (2021). https://doi.org/10. 1109/ACCESS.2021.3051066 9. Kumar, S., Dixit, A.S., Malekar, R.R., Raut, H.D., Shevada, L.K.: Fifth generation antennas: a comprehensive review of design and performance enhancement techniques. IEEE Access (2020). https://doi.org/10.1109/ACCESS.2020.3020952
Bond Graph Modelling and Simulation of Active Heating/Cooling Using Thermoelectric Energy Conversion Systems Gautam A. Raiker(B) , B. Subba Reddy, and L. Umanand Interdisciplinary Center for Energy Research, Indian Institute of Science Bangalore, Bangalore 560012, India [email protected]
Abstract. Thermoelectric Devices operate as a Thermal Heat Pump using Electricity and involve Electrical and Thermal Domains. The Bond Graph Approach is a Multi Domain Modelling Methodology which can be used to model these devices. The Bond Graph Model for Thermoelectric Devices is developed and explained in this paper along with Validation Experimentally. For improved Coefficient of Performance the Temperature Gradient across the Thermoelectric Device should be minimal. To achieve this, a stacked arrangement of devices can be used to divide the temperature gradient across devices. Temperature Balancing is necessary to divide the temperature equally across the arrangement which is achieved using a novel Power Electronic Topology and Duty Cycle based Algorithm. This technique is Simulated and the results are showcased in this work. Finally, the Active Heating/Cooling can be integrated with a Solar Photovoltaic based interface to make the system Energy Independent. Keywords: Bond graph modelling · Thermoelectric devices · Active cooling/heating · Stacked peltier arrangement · Temperature balancing
1
Bond Graph Model
Bond Graph Models are used to graphically represent and model physical systems and helps to convert the system to its state space representation with ease by visual inspection of the model [1]. Bond Graph models are domain agnostic and can incorporate multiple energy domains. It consists of Bonds and Junctions which can exchange energy and each bond has an effort and flow variable whose product is the power [2]. Bond graph models are convenient for simulations of multi-domain systems and hence this approach is used for modelling of Thermoelectric Devices. Bond Graphs have one port, two port and multi port elements. One Port elements include source elements (effort and flow) and passive elements (Dissipative and Energy Storage), two port elements consist of transformers and c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 299–305, 2022. https://doi.org/10.1007/978-981-19-1677-9_27
300
G. A. Raiker et al.
gyrators and multiport elements have 1-junction where flows are additive and 0-junctions where efforts are additive. In addition there are multiport RS elements which include thermal entropy effect which can have several bonds and one of the bonds which dissipate heat are irreversible [3]. In the electrical domain the effort is Voltage and flow is Current while in thermal domain the Effort is temperature and Flow the Entropy Rate.
Fig. 1. Bond graph model for the thermoelectric system
The bond graph model is developed for a Thermoelectric Device which takes a Voltage and Current and develops a temperature difference to pump heat as shown in Fig. 1. The important element here is the RS element which constitutes the Seeback Coefficient (α). The relation between Current I and Entropy Rate S˙ is given as follows: S˙ = αI (1) Also, the temperature at hot junction is given as Th and at cold junction as Tc which gives difference ΔT at the 1 junction. ΔT = Th − Tc
(2)
At the RS element the entropy flow to the cold junction and the hot junctions are derived as follows. The thermal conductivity k causes heat to flow from hot junction to the cold junction and the resistive loss is irreversible dissipated equally at the hot and cold junction ΔS1 Tc = I 2 R/2 + kΔT
(3)
ΔS2 Th = I 2 R/2 − kΔT
(4)
Bond Graph Model for Thermoelectric Devices
301
The relation between Voltage and current with temperature difference ΔT is given by ΔU − αΔT R
(5)
ΔU = U2 − U1
(6)
I=
The cooling and heating power of the Peltier Module is given by Qc and Qh as in (7) and (8) respectively. The electrical power input is given by W as in (10). Qc = αITc − 0.5I 2 R − kΔT
(7)
2
Qh = αITh + 0.5I R − kΔT W = Qh − Qc
(8) (9)
W = αIΔT + I 2 R
(10)
With this section, the preliminaries are over, and next the model is used to simulate active heating and cooling conditions along with validation of the model experimentally.
2
Validation of the Model
Fig. 2. Experimental setup for validation of the model
For validation of the model, a experimental setup of Thermoelectric/Peltier Device is constructed as shown in Fig. 2. The Cold Junction of the Device is connected to a radiator with fan arrangement to allow cooling of a chamber
302
G. A. Raiker et al.
and the hot junction is connected to a water cooled block. The equation for the radiator with cold junction is given as dTc = Qc − U (Tc − Tr ) (11) dt Here U is the convection heat loss coefficient and Tr is chamber temperature. The Chamber is cooled by the fan cooled convection from the cold junction of the Thermoelectric Module. Similarly for Hot side of the Thermoelectric module which is cooled with a water cooling block with water inlet temperature (Tin ) and heat loss coefficient k2 , the equation is as follows Cc
dTh = Qh − k2 (Th − Tin ) (12) dt For the thermal inertia of air in the chamber the following equation is valid. The chamber is cooled by the convection of air by the air cooling block and the some heat may enter from the ambient (Ta ) given by heat loss coefficient (k1 ) Ch
dTr = U (Tc − Tr ) − k1 (Tr − Ta ) (13) dt The parameters for the simulations are tabulated in the following table (Table 1). Cr
Table 1. Parameters used for Simulation. Parameter U Value
k
k1
k2
Tin
Ta
α
R
10 W/K 0.5 W/K 5 W/K 5 W/K 296.06K 300.75K 0.05 V/K 2.2 Ohms
The simulated and experimental results for cooling using the thermoelectric device are shown in Fig. 3.
Fig. 3. Simulation and experimental results for cooling
Bond Graph Model for Thermoelectric Devices
303
By reversal of current, heat can be pumped in the opposite direction using the Thermoelectric Device. Hence the chamber can be heated and the simulation versus the experimental results are shown in Fig. 4 respectively.
Fig. 4. Simulation and experimental results for heating
3
Stacked Thermoelectric Arrangement
The coefficient of performance of thermoelectric devices is best when the temperature difference between the hot and cold side is low. The concept of stacked thermoelectric modules has been explored previously to improve the ΔT and the Coefficient of Performance [4–6]. As modules are stacked the temperature is divided across the individual stacks. The issue with this arrangement is to have constant temperature across each stack and hence the area of subsequent stacked modules have to be reduced. Instead of this the problem can be solved by controlling the current to the stacks to enable temperature balancing.
Fig. 5. Power electronic topology for stacked peltier arrangement
The voltage of bottom peltier stack has to be lower than the top stack assuming that the bottom stack is the cold side. To enable this a power electronic topology as shown in Fig. 5 can be used which divides the input voltage according to
304
G. A. Raiker et al.
the duty cycle applied as shown in below equations. This helps to modulate the voltage to the two stacks of the peltier modules and at a particular duty cycle will balance the temperature of the stack. VP 1 = d × Vin
(14)
VP 2 = (1 − d) × Vin
(15)
An experiment is conducted to validate the idea and the results are shown in Fig. 6. A simulation is carried out based on the bond graph model and it can be seen that both match.
Fig. 6. Simulation and experimental results for stacked peltier arrangement
4
Conclusions
A Bond Graph Technique based Model was developed for Thermoelectric Devices. The model was validated experimentally for Active Cooling and Heating modes. The Bond Graph model can be used independent of the domain of the system and hence is a useful tool in case of Thermoelectric Devices which have both Thermal and Electrical Domains. For improving COP of the system a novel Temperature Balancing technique for Stacked Peltier modules was discussed in this work. This was validated using the bond graph model and experimental investigations.
References 1. Thoma, J., Mocellin, G.: Frictions and irreversibilities. In: Thoma, J., Mocellin, G. (eds.) Simulation with Entropy in Engineering Thermodynamics, pp. 24–25. Springer, Heidelberg (2006). https://doi.org/10.1007/978-3-540-32851-3 2 2. Umanand, L.: Modelling of systems. In: Power Electronics Essentials and Applications, pp. 453–463. Wiley (2016)
Bond Graph Model for Thermoelectric Devices
305
3. Thoma, J.U.: Entropy and mass flow for energy conversion. J. Franklin Inst. 299, 89–96 (1975) 4. Yang, R., Chen, G., Snyder, G.J., Fleurial, J.P.: Multistage thermoelectric microcoolers. J. Appl. Phys. 95, 8226–8232 (2004) 5. Cheng, Y.-H., Shih, C.: Maximizing the cooling capacity and COP of two-stage thermoelectric coolers through genetic algorithm. Appl. Therm. Eng. 26, 937–947 (2006) 6. Meng, F., Chen, L., Feng, Y., Liu, X., Zhao, T.: Performance analysis and optimization of multi-stage combined thermoelectric generators. Int. J. Refrig 34, 2129–2135 (2017)
Deep Learning Based Fruit Defect Detection System Anshul Bhardwaj1(B)
, Nitasha Hasteer1
, Yogesh Kumar1
, and Yogesh2
1 Amity University Uttar Pradesh, Noida 201313, India
[email protected] 2 Department of Information Technology, Amity University, Noida 201313, India
Abstract. During transportation, many defects appear on the surface of fruits due to multiple reasons such as negligence in packing methods, not maintaining the required temperature, and mixing rotten fruits with fresh fruits. This results in the quality of fruits getting degraded and the suppliers facing losses. In this research, a model is being created to detect defects in fruits. Convolutional Neural Network (CNN) is used because of its ability to learn from images, create patterns, then use it to train itself and predict results for fruit from its image. Our database consisted of 8 fruit images which are self-collated from google images and kaggle. Accuracy for each of the fruit classifier is more than 95%. Keywords: Convolutional Neural Network · Fruit defects · Deep learning · Fruit quality assessment
1 Introduction Agriculture sector is a very promising sector in India and a large population is employed in this sector. It accounted for 20.19% of the GDP of the country during the year 2020–21 [1]. There are multiple steps involved before a fruit reaches the consumer. During transportation and storage of these fruits, many surface defects appear on them due to various reasons such as repeated rubbing with other fruits due to improper packaging, dumping from height, not handling with care, and many more [2]. After such defects fruit becomes less attractive. If someone goes to buy fruits, the first thing they check is how it looks and does the fruit has any scars on the surface. They may not know what all diseases a fruit has, and the appearance is given number one priority. In India, fruit farming is done on 6.66 million hectares of land and this resulted in production of 99.07 million metric tonnes of fruits in the year 2019–20 [3]. In a report it is stated that around 16% of fruits and vegetables are wasted due to lack of infrastructure. They become defected, which affects their beauty and then farmers or food processing industries have to face losses because these fruits are generally not accepted by the consumers [4]. When the fruit is moved from farms to processing units or factories, this fruit detection system can be used to check for the defected fruits so that they can reject defected fruits © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 306–313, 2022. https://doi.org/10.1007/978-981-19-1677-9_28
Deep Learning Based Fruit Defect Detection System
307
and exclude them from the further process. It helps the industry to avoid paying for defected fruits. The detection system is also useful for the distributors to avoid purchasing the defected fruits because they are not fit for selling and affect all the other fruits which are in the same container. The retailers are also benefited from this system as they use this system to reject defected fruits and only keep the fruits that are free from any defects. Deep learning which, is a part of Artificial Intelligence is already being implemented at multiple places in agriculture so we propose a model with deep learning, which helps in the image classification of fruits. This paper consists of 4 sections. Section 1 is Introduction. Section 2 consists of literature review and an accuracy comparison table of other research studies. Section 3 is the methodology followed by discussion on result and conclusion in Sect. 4.
2 Literature Review Kazuhiro Nakano [5] implemented 2 CNN’s for the grading of apples. In Japan, apples are given as a gift so the quality of apples must be top-notch and there should be no defect. For this research a very popular and expensive apple of Japan, San Fuji was used. They were directly sourced from the farmer so that they are fresh and then their images were captured with a CCD video camera. Apples were kept on a turntable that rotated 18 degrees in a single turn. After the image is captured, 1st CNN is implemented to classify the color of that apple so that it can be further analyzed. If it is found that the color of the fruit is satisfactory, then these fruits were tested in 2nd CNN where they are categorized in 5 categories – AA, A, B, C, D. Fruit in AA category was the most superior and fruit in D category had defects on its surface. Siyuan Lu et al. [6] designed a 6-layer convolutional neural network to classify fruits. They feel that automatic fruit classification is necessary, so they have designed this CNN model. Their dataset included 9 fruits, each of the fruit had around 200 images in the dataset. These images were collected from Kaggle, google images, and Baidu images. 900 images were used for training and test purposes respectively. With epoch 30 and minimum batch size 256, this model was able to reach an accuracy level of 91.44%. Asia Kausar et al. [7] created a Pure Convolutional Neural Network (PCNN) for classifying fruit images. It consisted of 7 layers, some of them also with stride. For the research, the authors used a recently presented fruit-360 dataset available on Kaggle which consisted of 55,244 color images of fruits. This PCNN was proposed with a Global Average Pooling (gap) layer and strides. They have used GPU for training instead of CPU because it is observed that CPU takes 6 days whereas GPU only takes 22 min. So, using a GPU can pace up the training process. Accuracy is compared of PCCN + GAP, using CNN and fully connected layer without dropout or with dropout. The accuracies reported are 98.88%, 97.41%, and 97.87%. It shows that PCCN is better for classifying fruits. Rucha Dandavate and Vineet Patodkar [8] created a model to classify fruits into 3 categories – Raw, ripe, and overripe. The model was created with 2 hidden layers and 3 fully connected layers. 12,559 images are included in the dataset out of which 80% were used for training the model and the rest 20% used for testing purposes. Only 4 fruits were selected for the model and all of them had 3 same categories - raw, ripe, and overripe.
308
A. Bhardwaj et al.
This model was able to achieve an accuracy of 97.74% in 8 epochs. Researchers felt this model can be used in food processing industries where labour is required for sorting the fruits. Sovon Chakraborty et al. [9] compared multiple models to analyze their accuracies, two CNN-based architectures and one CNN-based mobilenetV2 were included in the study. The goal was to detect rotten fruits with the help of deep learning approach. Agriculture is a large sector in Bangladesh and currently sorting work is done by manual labour which is time-consuming and the standards can also not be maintained because one worker may accept a fruit whereas same fruit could be rejected by another worker as it is solely dependent on what the worker feels about the quality of that fruit. Due to this, there can be an error that a defective fruit is accepted which can also lead to contamination of all other good fruits because of one rotten fruit. For finding defects on apples, V.Leemans and M.-F Destain [10] designed a system to acquire images of apples and then use them to grade these apples. A light tunnel was created with 4 tube lights and 2 cameras which captures the photograph of the apple and then the fruit would be segmented based on the acquired image. The system checks for defects on the surface of the fruit and this segmented region would then be passed to the next step where this defect was analyzed. After the steps are completed that fruit was graded based on its parameters. The accuracy of this system to grade apples was 73%.
3 Methodology A Convolutional Neural Network (CNN) is built to analyze the images. The proposed model identifies the defects in fruits and then provides the results. CNN is vastly used in identification models due to its high accuracy by learning from multiple images of the fruits. CNN is popularly used for analyzing images. It is an artificial neural network that is capable of detecting patterns in images which are later used to make assumptions and predictions. This is the reason why CNN is vastly used. It has hidden layers called convolutional layers and these layers are precisely what Table 1. Images in dataset Fruit name
Number of images
Apple
153
Banana
200
Mango
63
Orange
104
Grapes
58
Papaya
60
Pineapple
59
Pomegranate
68
Random
119
Deep Learning Based Fruit Defect Detection System Table 2. Fruits and their defects
Fruit Name
Defect Name
Apple
Blotch pit
Banana
Red rust thrips
Mango
Body rot
Orange
Chilling injury
Grapes
Spots
Papaya
Stem end rote
Pineapple
Fruitlet core rot
Pomegranate
Stale
Image
309
310
A. Bhardwaj et al.
makes a CNN. The convolutional layer receives the input and then transforms it which is forwarded to the next layer as its input. Model is trained from multiple inputs which makes it capable to identify objects and categorize them in the given classes. For training and testing of this model, images are downloaded from google images and Kaggle [11]. Table 1 shows the number of images in our dataset. There are a total of 884 images of fruits in the database and 119 images of random objects. In this research, eight fruits are included which are commonly found in the Indian market. These are locally grown and do not require special plantation techniques. Then the images of the healthy and defected fruits are collated. For the selected 8 fruits, we had 36 classes that belonged to their defects and 1 class in each fruit is for their fresh image i.e. free from any defect. In Table 2, images of few defects are shown. The CNN model is created on google colab to overcome the limitations of hardware and GPU. With the help of colab, we are able to access systems with GPU so that all the work can be done faster than the CPU. It is proven that GPU increases the speed by a great factor [7]. The first model is used to classify the image into fruit or random object. If the input image is of a random object or of a fruit that is not in our database, then the model gives an error for invalid input. If the image is of fruit, then it is sent to the classifier of that fruit where it is checked which defect is present in that fruit. The accuracy of this classifier is 97.32%. Figure 1 shows the flow diagram of our model, it depicts the process from inserting an image in the model to showing its result. In the model, ‘relu’ activation is used for the layers, softmax for the final layer, and adam used as the optimizer.
Fig. 1. Flow diagram of the model
Similar models are created for all the fruits, by changing some parameters if required anywhere.
Deep Learning Based Fruit Defect Detection System
311
Architecture of a fruit classifier is shown in Fig. 2.
Fig. 2. Architecture of a fruit classifier
Image augmentation is applied using python code and image data generator is used. These are the augmentation parameters in Table 3. Table 3. Image parameters Parameter
Value
Rescale
1./255
Shear_range
0.2
Zoom_range
0.2
Horizontal_flip
True
Target size
(500,500)
Color mode
RGB
Class mode
Categorical
Batch size
16
312
A. Bhardwaj et al.
4 Result and Conclusion CNN model is created and implemented to identify defected fruits. Our model consisted of 8 fruits and 884 images and is able to achieve great accuracy for each of the fruit. All the classifiers are trained separately, each having accuracy above 95%. Table 4 shows the accuracy of all fruit classifiers with the number of epochs which are applied to them while training the model. Table 4. Accuracy of fruit classifiers Fruit name
Accuracy achieved
Number of epochs
Apple
95.83%
30
Banana
95.31%
20
Mango
98.25%
25
Orange
95.96%
10
Grapes
98.15%
8
Papaya
98.21%
20
Pineapple
98.18%
16
Pomegranate
98.41%
15
Accuracy for the initial classifier is 97.81%, this classifier is used to analyze the input image and decide which fruit that is, or if it belongs to the random class. If the image is of a random object, an error will be given stating that it is a random object or this fruit is currently not in our database. A confusion matrix is used to show the result of a classifier.
Fig. 3. Confusion matrix for pomegranate
Deep Learning Based Fruit Defect Detection System
313
It is designed based on the number of classes. Figure 3 shows a confusion matrix for our pomegranate classifier. Predicted defect and actual defect are on the x-axis and y-axis respectively. Our proposed model is beneficial for many fruit suppliers and industries. These systems can be installed in pre-processing zones in industries where fruits are used which will prevent defective fruits to enter further stages. This will help the companies to maintain the quality of supplied fruits as no defective fruits will be included. They can program the system to work with different fruits or may only focus on single fruit as per their requirement.
References 1. Ministry of Statistics and Programme Implementation, “Sector-wise GDP of India,” Statistics Times (2021). https://statisticstimes.com/economy/country/india-gdp-sectorwise.php 2. Altisent, M.R.: Damage mechanisms in the handling of fruits. Handl. Fruits 1, 231–257 (1991) 3. “Fresh Fruits & Vegetables,” Agricultural & Processed Food Products Export Development Authority. http://apeda.gov.in/apedawebsite/six_head_product/FFV.htm. The area under cultivation of, %2C brinjal%2C Cabbages%2C etc. 4. Sharma, S.: India wastes up to 16% of its agricultural produce; fruits, vegetables squandered the most. Finan. Express-Read to Lead (2019). https://www.financialexpress.com/economy/ india-wastes-up-to-16-of-its-agricultural-produce-fruits-vegetables-squandered-the-most/ 1661671/ 5. Nakano, K.: Application of neural networks to the color grading of apples. Comput. Electron. Agric. 18(2–3), 105–116 (1997). https://doi.org/10.1016/s0168-1699(97)00023-9 6. Lu, S., Lu, Z., Aok, S., Graham, L.: Fruit classification based on six layer convolutional neural network. In: International Conference Digital Signal Processing (2018). https://doi. org/10.1049/joe.2019.0422 7. Kausar, A., Sharif, M., Park, J., Shin, D.R.: Pure-CNN: a framework for fruit images classification. In: Proceedings of the 2018 International Conference on Computational Science and Computational Intelligence, CSCI 2018, pp. 404–408 (2018). https://doi.org/10.1109/CSC I46756.2018.00082 8. Dandavate, R., Patodkar, V.: CNN and data augmentation based fruit classification model. In: Proceedings of the 4th International Conference on IoT Social, Mobile, Analytics and Cloud, ISMAC 2020, pp. 784–787 (2020). https://doi.org/10.1109/I-SMAC49090.2020.9243440 9. Chakraborty, S., Shamrat, F.M.J.M., Billah, M.M., Al Jubair, M., Alauddin, M., Ranjan, R.: Implementation of deep learning methods to identify rotten fruits, pp. 1207–1212 (2021). https://doi.org/10.1109/icoei51242.2021.9453004 10. Leemans, V., Destain, M.F.: A real-time grading method of apples based on features extracted from defects. J. Food Eng. 61(1), 83–89 (2004). https://doi.org/10.1016/S0260-8774(03)001 89-4 11. Potdar, R.R., Shrivastava, A., Sohandani, R., Khatwani, N.: Fresh and Stale Images of Fruits and Vegetables. Kaggle. https://www.kaggle.com/raghavrpotdar/fresh-and-stale-images-offruits-and-vegetables
Design of 1 kVA Shunt Compensator Test Bench Anchal Singh Thakur(B) , L. Umanand, and B. Subba Reddy Indian Institute of Science, Bangalore 560012, India {anchalthakur,lums,sreddy}@iisc.ac.in
Abstract. With the need for Power Quality Improvement in Distribution System, the need to test the possible solutions for this dynamic problem increases. Although simulation tools are available, they do not give practical output and the knowledge gained from the design of a test setup is unparalleled. Hardware design not only improves the understanding, but also educates about the various sub-systems involved in the proper functioning of the system. This approach also helps to find better solutions to overcome the shortcomings of the system and fine tune the tested system into a better system. In this paper, the hardware design procedure for a 1 kVA system is presented. The system contains threephase source, load and a parallelly connected shunt compensator. Threephase shunt compensator is configurable into two topologies four-leg and three-leg Voltage source converter. Further, circuits pertaining to the precharging, gate signal conditioning, followed by the combined schematic design and three-dimensional model of the PCB is also described and presented. Keywords: Four-leg · Voltage source converter compensator · Three-leg
1
· Test setup · Shunt
Introduction
Electrical power quality in distribution system has become an important focus due to the recent trends of increase in products with front-end power electronics, distributed generation from renewable sources of energy, increase in electric vehicles etc. [1,2]. With the increase in need for Power quality improvement [3], the need to develop a simple test setup for testing the solutions to the problem of Power Quality Improvement exists. Although simulation tools are available but the modelling is not accurate enough to give practical results for real life situation and it takes long time to develop a model for the system. As an alternative to simulations, a laboratory prototype to test the control algorithms can be developed. The process of developing the test setup would greatly enhance the skill-set and debugging capabilities of the students. So far in the literature many articles are present about the design process, simulations and results of hardware tests [4] for Power Quality improvement, but very few of them explain c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 314–326, 2022. https://doi.org/10.1007/978-981-19-1677-9_29
CiPOS Inverter
315
the gap between design and hardware setup [5]. In this paper we have bridged that gap by presenting the Printed Circuit Board (PCB) design for the test setup. This design was pre-tested using simulation in [6]. All the steps involved in the development of a prototype system with 1 kVA Shunt compensator are presented.
Fig. 1. System
2
System
The complete system consists of three-phase source, three-phase load and a twolevel, four/three-leg Shunt compensator. Same setup can be used to test three-leg and four-leg compensator topologies. The Connection diagram for the system is shown in Fig. 1. The shunt compensator is connected in parallel to the source load system through interface filters. The compensator is rated at 1 kVA with line to line supply voltage as 415 V, 50 Hz. The part of the system highlighted in green is the only part that is on the PCB, rest components like Source, load, Interface filters (Fa, Fb, Fc, Fn) are connected externally to the PCB. The purple rings corresponds to the part where current sensors are placed and the red dots represent the node points for Voltage sensing. MATLAB Simulation for the system with detailed specifications is presented in [6]. This paper shows the design of the PCB. KiCad [7] is used for the schematic design and PCB layout. The schematic of the system is designed using hierarchical sheets. Each hierarchical sheet contains a part of the system and all these sheets combine to form the complete system when external source, load and interface filters are connected to it. The hierarchical sheet contains:
316
1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
A. S. Thakur et al.
Power Supply Source Load system Connection Relay Intelligent power module (IPM) Sensors Gate signal conditioning circuit DC link Pre-charging circuit Digital Signal Processor(DSP) Protection circuit
Fig. 2. Power block diagram
2.1
Power Supply
The Power input to the PCB is 24 V, 3 A. The DC voltages required by various Analog components on the PCB are 15, 5, 3.3, 2.5 V and the DC voltages required by various Digital components on the PCB are 5, 3.3 V. The block diagram of the Power Supply for the PCB is shown in Fig. 2. To prevent the effect of digital switching on the working of Analog components, the Digital and Analog grounds are isolated. And since the Voltage difference between Input and Output Voltage is high so we used Buck converter Integrated circuits (ICs), since significant losses will be involved if Linear voltage regulators are used instead. For the 2.5 V supply, a voltage reference IC is used as the application needs very low current. The Schematic design for the Power supply sheet is shown in Fig. 3. Individual power stages are designed using the WEBENCH tool [8] from Texas Instruments. 2.2
Source Load System
Terminals are provided on the PCB for the connection of three-phase source, three-phase three-wire or four-wire load and Interface filters. The purpose of
CiPOS Inverter
317
Fig. 3. Power block diagram
Fig. 4. (a) Source, load and Interface filter circuit (b) Relay circuit
providing the provision for Interface filter is to test the PCB with multiple possible Interface filters and then select the one compatible with the algorithm. The source, load terminals are provided so that the Compensator can be tested under all possible conditions like balanced, unbalanced, distorted etc. The Schematic design for the source-load sheet is shown in Fig. 4a.
318
A. S. Thakur et al.
2.3
Connection Relay
Connection relays are provided between the source-load system and the Compensator so that the Compensator can be first tested alone to check for any faults within the converter and then connected to the source- load system for proper functioning of the system. The Schematic design for the Connection Relay sheet is shown in Fig. 4b.
Fig. 5. CiPOS IPM - IM818 circuit
2.4
Intelligent Power Module (IPM)
IPMs contain switches and gate drivers within the same IC which reduces the parasitics. In this PCB CiPOS IPM (Maxi series - IM818-MCC) from Infineon technologies is used. Each module contains three legs, so for the four-leg Compensator we used two such modules with two legs from each module. The Schematic design for one of the module is shown in Fig. 5. The input to these modules are gating signals from the DSP post conditioning and output is connected to the poles of each phase. The design of the auxiliary components of the module are done using the data sheet [9]. The IPM used has following features:
CiPOS Inverter
319
1. 1200 V, 10 A TRENCHSTP IGBT4, 1200 V Emitter controlled anti-parallel diode, 6-channel gate driver IC 2. Integrated bootstrap functionality 3. Matched propagation delay 4. Built-in dead times 5. Cross conduction prevention 6. Under voltage lockout at all channels 7. Over current shutdown (ITRIP pin) 8. Temperature monitoring (VTH pin) 9. Programmable fault clear time, fault output and enable input (RFE pin) 2.5
Sensors
In the PCB as per the application [6], the following parameters are sensed: a. b. c. d.
Point of common coupling (PCC) AC Voltages - 3 AC voltages AC Load Currents - 4 AC currents AC Compensator Currents - 4 AC currents DC link Voltages - 2 DC voltages (mid-point and DC link voltage)
For AC current sensing, ACS 712, hall effect based sensors are used because they are the most economic solution for the target measurement range and desired accuracy. DC voltage measurement is done using Operational Amplifier (Opamp) based differential voltage measurement circuit shown in Fig. 6a. The output of the sensor circuit is calibrated to be in the range of 0 to 3.3 V using voltage divider network. For AC voltage measurement additional level shifting is done so that uni-polar output voltage is obtained. The choice of uni-polar voltage output will reduce the required protection circuitry. All the three types of sensing circuit designs are shown in Fig. 6. 2.6
Gate Signal Conditioning Circuit
This part of the circuit is required to disable the gating signals for the switches in case of any faults in the IPMs which is signaled by RFE pin in IM818 IPM. The Schematic for the gate signal conditioning sheet is shown in Fig. 7a. 2.7
DC Link
Dc link mainly consists of banks of electrolytic capacitors. Snap-in type of capacitors are used instead of screw type to reduce parasitics. Along with the electrolytic capacitors, a metal film capacitor is placed in parallel to surpass high frequency components and bleeder resistors are added to the setup to dissipate the energy stored in the capacitors when the setup is turned off. The bank of electrolytic capacitors is made of 6 capacitors, where two banks of three parallel capacitors are connected in series. Voltage across both these banks are measured. The schematic for the DC link is shown in Fig. 7b.
320
A. S. Thakur et al.
Fig. 6. (a) DC voltage sensing circuit (b) AC current sensing circuit (c) AC voltage sensing circuit
CiPOS Inverter
321
Fig. 7. (a) Gate signal conditioning circuit (b) DC link
Fig. 8. (a) Pre-charging circuit (b) Digital Signal Processor (DSP)
2.8
Pre-charging Circuit
This circuit is used to provide the initial charge to the capacitor, before connecting to the grid. Resistors with positive temperature coefficients are used in this circuit to provide inherent over current protection. The schematic for this part is shown in Fig. 8a. 2.9
Digital Signal Processor (DSP)
In this paper we have used DSP from Texas Instruments (TMS320F28379D). For the simplicity of PCB design, we have used the launch pad for the DSP (LAUNCHXL-F28379D). The peripherals needed for the setup are ADC (Analog to digital converter), ePWM (enhanced-Pulse Width Modulation), GPIO (General purpose Input-Output). The following signals are connected to the DSP:
322
1. 2. 3. 4. 5. 6.
A. S. Thakur et al.
Current sensor output signals Voltage sensor output signals Gating signals Relay control signals Temperature monitoring signals Fault signals from IPMs
The Schematic for the DSP launch pad is shown in Fig. 8b.
Fig. 9. Protection circuit
2.10
Protection Circuit
This part of the circuit disables the gating signals to the IPMs in following cases: 1. 2. 3. 4. 5.
Over current condition for Compensator currents Over voltage condition for DC link voltages Under voltage condition for DC link voltages Manual ON-OFF control signal, controlled using jumper RESET signal from DSP
The schematic design for the Protection circuit sheet is shown in Fig. 9.
CiPOS Inverter
323
Fig. 10. Complete schematic
3
Schematic Design
Based on the snippets of schematics shown so far, the complete Schematic for the system is shown in Fig. 10. Each box represents a hierarchical sheet. Threedimensional representation of the PCB can be shown in Fig. 11.
4
Complete Process for Hardware
In this paper it is assumed that the components are procured separately and PCB is fabricated. The PCB assembly will be done by the individual. Following are the steps to get the desired hardware: 1. Freeze the specifications of the system based on the proof of concept to be tested. This will give the rating of the compensator. 2. Calculate the value of the components of the compensator like ratings of the switches, Capacitor sizing, Interface filter sizing etc. 3. Based on the ratings obtained in 2, finalise the components of the power circuit. 4. Based on the nature of control algorithms that will be implemented on the Compensator, sensor placement and quantity is decided. 5. Then based on the range of the measured quantity and the accuracy desired, freeze the sensor and its topology. 6. Based on the total control signals to be handled by the control device and their nature (ADC or PWM or GPIO), we decide the type of control device (controller or processor or Field programmable gate array) to be used for this application. 7. Step-by-step freeze all the components for the PCB. While freezing the components make sure that they are available in stock and is an active
324
A. S. Thakur et al.
Fig. 11. (a) Front side view of 3-D model of PCB (b) Back side view of 3-D model of PCB
CiPOS Inverter
325
component. The stock information can be obtained from various platforms like Mouser, Digikey, Element 14 etc. 8. As soon as you freeze the components, place the order for the components as the timeline for procurement of components is usually higher than the timeline for fabrication of the PCB. 9. After finalising all components, complete the schematic of the PCB and simultaneously add all the specific footprints to the respective components. Once the schematic is completed, add the general footprints to SMDs (Surface mounted device). 10. Once the schematic is complete with the attached footprints, go for PCB layout. 11. Before starting the layout process, arrange the components so as to reduce the length of connecting wires between them. Start with critical connections first and then move towards rest of the connections. 12. Once done with all the connections, go for DRC (Design rule check). After solving all the errors reported in the DRC check, view the 3-D model of the PCB and check. 13. Send the PCB for fabrication to the vendor. 14. After receiving both components and the fabricated PCB, populate one part at a time, starting from the power supply. 15. check and rectify the errors in each part until the complete board functions as desired.
5
Conclusion
A Compensator, source-load test setup with 1 kVA three-leg/ four-leg shunt compensator is discussed. Circuits of all the sub-systems like Pre-charging circuit, gate signal conditioning circuit etc. are discussed, followed by the combined schematic design and three-dimensional model of the PCB is also shown. Along with the schematic design, complete end-to-end procedure from design to physical hardware is explained. Hardware approach helped in better understanding of the subject along with ideas to further improve the real system. The hardware tests on the setup designed are ongoing and will be published soon.
References 1. Li, D., Wang, T., Pan, W., Ding, X., Gong, J.: A comprehensive review of improving power quality using active power filters. Electr. Power Syst. Res. 199, 107389 (2021). https://doi.org/10.1016/j.epsr.2021.107389 2. Sperstad, I.B., Degefa, M.Z., Kjølle, G.: The impact of flexible resources in distribution systems on the security of electricity supply: a literature review. Electr. Power Syst. Res. 188, 106532 (2020). https://doi.org/10.1016/j.epsr.2020.106532 3. Mikkili, S., Panda, A.K.: Power quality issues and solutions-review. Int. J. Emerg. Electr. Power Syst. 16, 357–384 (2015). https://doi.org/10.1515/ijeeps-2014-0052 4. Upamanyu, K., Narayanan, G.: Simplified grid emulator for testing grid-connected power electronic converters. In: 2020 IEEE International Conference on Power Electronics, Smart Grid and Renewable Energy (PESGRE2020) (2020). https://doi.org/ 10.1109/PESGRE45664.2020.9070471 5. Adapa, A.K., Venkatramanan, D., John, V.: Auxiliary subsystems of a generalpurpose IGBT stack for high-performance laboratory power converters. S¯ adhan¯ a 42(8), 1355–1362 (2017). https://doi.org/10.1007/s12046-017-0661-5
326
A. S. Thakur et al.
6. Thakur, A.S., Umanand, L., Reddy, B.S.: Multi-mode operation of converter for power quality improvement. In: Mohapatro, S., Kimball, J. (eds.) Proceedings of Symposium on Power Electronic and Renewable Energy Systems Control. LNEE, vol. 616, pp. 439–449. Springer, Singapore (2021). https://doi.org/10.1007/978-98116-1978-6 38 7. Link for PCB design tool. https://www.kicad.org/ 8. Link for Power supply design tool. https://webench.ti.com/power-designer/ switching-regulator?powerSupply=0 9. Datasheet for CiPOS Maxi IPM, IM818-MCC. https://www.infineon.com/dgdl/ Infineon-IM818-MCC-DataSheet-v02 02-EN.pdf?fileId=5546d46265f064ff0165f9e9 1e3c2e9e
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect D. L. Kavya Reddy, Kanishka Negi, D. R. Soumya(B) , Gaddam Prathik Kumar, Subrata Sahana, and Anil Kumar Sagar School of Computer Science and Engineering, Sharda University, Greater Noida, Uttar Pradesh, India {2019573564.d,2019600654.kanishka,2019533226.durgam, 2019004968.gaddam}@ug.sharda.ac.in, {subrata.sahana, anil.sagar}@sharda.ac.in
Abstract. The COVID-19 pandemic has reshaped everyone’s life as we know it. Many measures and proposals have been suggested to prevent the virus’s rapid spread since scientists detected it. In the midst of many of these, one major piece of advice was to “wear a mask!”. The mask or face coverings can protect the wearer from contracting the virus or spreading it to others. Hence, to ensure that the public is not wearing a mask, one of the applications we have developed is face mask detection, which uses biometrics to map the features of a human face, analyse the data, and determine whether the detected face is a positive image or a non-face or negative image by placing a box around it. The main ideology of studying and researching face mask detection is to use computer technology to directly mimic the action and manner of human facial features and develop a system that reduces the amount of effort required to identify whether or not the person is wearing a mask and notifying them with the assistance of the proposed model. Keywords: Automatic face mask detection · Machine learning · Open CV · COVID-19 · CNN · MobileNetV2
1 Introduction Face mask detection is a computer technology based on AI which can be thought of as a subset of object-class detection. It is a technology that helps in recognizing human faces in digital images or videos [13]. This, like many others, is an easy replacement for human efforts. It uses CCTV surveillance and detects the people without a mask and notifies them to follow the policy to prevent the virus from spreading [14]. As a result, the suggested project’s implementation is important for the benefit of numerous significant institutions and organizations that are striving to prevent COVID-19 from spreading while sustaining productivity.
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 327–344, 2022. https://doi.org/10.1007/978-981-19-1677-9_30
328
D. L. Kavya Reddy et al.
1.1 Machine Learning Machine learning is an application of artificial intelligence (AI) that allows machines to learn and improve themselves without having to be explicitly programmed. It primarily focuses on computer program development as a whole. It has evolved into a technical infection that is so pervasive that we are likely to utilize it multiple times per day. It has a wide range of uses, including automatic objects [6, 7]. 1.2 Computer Vision This is a branch of artificial intelligence that teaches computers to perceive and comprehend the visual world. Machines can properly recognize and classify objects using digital images from cameras and videos, as well as deep learning models, and then react to what they “see” [17]. 1.3 Deep Learning Deep learning is an artificial intelligence (AI) function that mimics the human brain’s processing of data and pattern creation to make decisions. Deep neural learning or deep neural network are other terms for the same thing. Virtual assistants, vision for selfdriving cars, money laundering, face recognition, and many other applications of deep learning are just a few examples [6, 7]. 1.4 OpenCV OpenCV is a large open-source library for computer vision, machine learning, and image processing, and it currently plays a critical part in real-time operations, which are critical in today’s systems. It may be used to detect items, faces, and even human handwriting in photos and movies [18]. 1.5 Tensorflow Tensorflow is a free and open-source machine learning software library. It can be used for a variety of applications, but it focuses on deep neural network training and inference. Classification, perception, understanding, discovering, prediction, and creation are some of the most common uses for TensorFlow [19]. 1.6 Keras Keras is a free and open-source software library for artificial neural networks that include a Python interface. Keras serves as a user interface for TensorFlow. Keras supported a variety of backends up until version 2.3, including TensorFlow, Microsoft Cognitive Toolkit, Theano, and PlaidML. Keras conforms to best practices for lowering the cognitive load, such as providing uniform and simple APIs, limiting the number of user activities required for typical use cases, and providing clear and actionable error messages [19].
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
329
1.7 Application It’s impossible to manually scan the audience all of the time to see if somebody is wearing a mask. The virus is most likely to spread in hospitals because the deceased are more susceptible to the illness. This assures the safety of patients receiving competent care. This can be used in any congested public or private location, such as industrial organizations, malls, ATMs, and so on [20].
2 Literature Survey FrerkSaxen, Philipp Werner, Sebastian Handrich, Ehsan Othman, LasloDinges, Ayoub Al-Hamadi [1] suggested a methodology known as “Face Attribute Detection with MobileNetV2 and NasNetMobile”. They proposed two simple but productive approaches for estimating facial features in unconstrained images in their work employing simple and rapid pre-processingface-alignment approaches. Face attributes were estimated using two lightweight CNN architectures, MobilNetV2 and NasNet-Mobile. In terms of accuracy and speed, both architectures are comparable. A comparison of their proposed approach to state-of-the-art methods in terms of processing time and accuracy reveals that it is faster than the best state-of-the-art model. Furthermore, this method is simple to use and may well be implemented on mobile devices. For mobile devices, they tackled the difficulty of estimating facial features from RGB photos. AnirudhLodh, UtkarshSaxena, Ajmal khan, AnandMotwani, ShakkeeraL, SharmasthVali Y [2] proposed “Prototype for Integration of Face Mask Detection and Person Identification Model – COVID-19”. In their work, they proposed a face mask identification platform that uses artificial networks to distinguish between those who are wearing masks and those who are not. If a person is not wearing a mask and is in the platform’s database, the suggested platform will send a notification to that person. Their classification methodology is MobileNet V2 neural networks, and the person identification model is based on the face recognition module. Their proposed system creates classification and predictive models that can account for accurate face mask identification, grouping, and prediction. MobileNet V2 is used to accomplish a system that will concentrate on improving prediction accuracy and detection probability as well as sending mail to the persons who aren’t wearing masks. Because of the MobileNet algorithm’s execution, accuracy will be improved and time complexity will be reduced. This proposed system monitors people using existing IP cameras from large institutions, making it cost-effective because no additional investment is necessary. In a short amount of time, the proposed system detects several faces and face masks from all angles. Vinitha. V, Valentina. V in their chapter “COVID-19 facemask detection with deep learning and computer vision” [3] suggested that an effective and cost-effective method of applying AI to establish a safe environment in a manufacturing environment. For face mask detection, a hybrid modelcombining deep and classical machine learning would be shown. They utilized OpenCV to accomplish real-time face detection from a live feed via webcam using a face mask detection dataset that consists of with mask and without mask photos. With the help of computer vision and deep learning, they proposed a facemask detection model. The proposed model can be used in combination with surveillance cameras to prevent COVID-19 transmission by detecting people who aren’t wearing
330
D. L. Kavya Reddy et al.
face masks. With Python, OpenCV, and TensorFlow, and Keras, the model combines deep learning and traditional machine learning approaches. ShashiYadav [4] proposed a model called “Deep Learning-based Safe Social Distancing and Face Mask Detection in Public Areas for COVID-19 Safety Guidelines Adherence”. The model proposed uses a combination of the Single Shot Detector (SSD) and MobileNetV2 [lightweight neural network] Considering limited resources and result accuracy the proposed model uses a transfer learning technique allowing it to be used on real-time video surveillance to detect people wearing face masks while maintaining safe social distance. Used OpenCV and MobileNet V2 to evaluate Real-Time Streaming Protocol (RTSP) video streams. The accuracy of the model, which was developed to operate on a Raspberry Pi 4, is between 85 and 95%. Camera video feeds from the Network Video Recorder (NVR) are transmitted using RTSP, and then converted to grayscale to increase speed and accuracy before being sent to the model for additional processing within the Raspberry Pi4. MobileNetV2 architecture was utilized as the framework since it is significantly less expensive than a traditional 2D CNN model. The SSD MultiBox Detector is also used in the process, which is a neural network architecture for high-quality image classification that has previously been trained on a vast number of pictures such as ImageNet and PascalVOC. Single Shot object Detection with the help of MobileNet V2 and OpenCV is used to recognize people in real-time, and it achieves 91.2% mAP, surpassing the similar state-of-the-art Faster R-CNN model.
3 Issue with Existing Model • Multiple faces cannot be detected in the current system. • The current system is not capable of detecting faces from various angles. • The existing system will not be able to recognize masks in a video or through cameras in a real-time situation. • The Existing system required more mathematical operations, parameters which eventually made the system slow [8].
4 Methodology 4.1 Convolutional Neural Networks (CNNs) In recent years in many of the image classification tasks, CNN’s i.e. Convolutional neural networks did a great job and got a great success (CNNs). But, CNN’s i.e. Convolution neural network’s performance depends highly on their architectures [10]. There are two elements as described in Fig. 1 i.e. pooling layer and convolutional layer which are very simple which compose a Convolutional neural network. • Feature Extraction is a component or a convolutional tool in which it extracts, separates, and determines the different image features for the analysis within the process. • Output is utilized by the fully connected layer which is extracted from the process of convolutional layers and then the model predicts the image classes which is based on extracted image features from the previous stages [5].
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
331
Fig. 1. CNN model [16]
4.2 Convolution Layer (CONV) In the first layer of the model, CONV is responsible for extracting the different and various features of the provided images. Between the MxM general-sized filter and the input image, some convolutional mathematical operations are performed to process in this layer. After that process, the dot product between the different parts of the input image and the filter is obtained concerning the filter size (MxM) through sliding the filter upon the provided images. Now the feature map is provided as the output in which it provides the images information like the edges and the corners. After that, these feature maps are used by the different other layers so that the model can learn the input images with various features. 4.3 Pooling (POOL) This pooling layer is also known as down sampling operation in which it is applied and used to diminish or reduce the feature map’s dimensions and size. It has considerably come into operation after the Convolutional layer. It also helps reduce the number of computations and the number of parameters used and performed in the convolutional network. The features map that contains the different features that are obtained in the region by the layer of convolution is summarized by the Pooling layer. Therefore, remaining functioning and operations are done on the features that are summarized rather than on the features that are positioned precisely which is obtained from the layer of convolutional. By this process, the model gets stronger and gets precise to the position variation of the features with the images that are provided to the model. There are two types of pooling techniques:
332
D. L. Kavya Reddy et al.
1) MAX Pooling In Max Pooling, the largest element is taken from the feature map within the form of fixed grid size. 2) Average Pooling Average Pooling calculates the average of the elements in a predefined size Image section within the fixed grid size. 4.4 Fully Connected (FC) In this regular layer neural network, takes the output results of the previous layer as an input and determines the scores of the classes within the network, and provides the output as the 1-dimensional array whose size is similar to that of the number of classes. Benefits of Convolutional Neural Networks • Convolution with the combination of Pooling helps in giving location invariant detection in our model that where the pictures are present in a frame it will detect it. • It also contains a ReLu operation which makes the model non-linear so that CNN can work well. • It makes the computation faster and speeds up the training due to ReLu operation • One of its components is Pooling which reduces the dimensions and computation within the model. • It also reduces over fitting with the help of connection sparsely and Pooling. • Pooling also helps in making the model tolerant towards the small distortion noise and variation in the frames. • CNN is also able to select the filters on its own by calculating the features with the image and deciding the RGB values within the pixels or grids of the image. • Its implementation is fast and easily understandable. • Among all the algorithms the CNN has the highest accuracy in predicting the images. • It also detects the important features within the image with any supervision of the human and which makes it computationally efficient. 4.5 MobileNet V2 MobileNetV2 is the architecture of a convolutional neural network which beliefs to operate well the mobile handheld devices. MobileNetV2 is commonly based on the residual structure that is inverted where the remaining or residual connecting networks are situated between the layer of the bottleneck. This MobileNet works efficiently in that it takes less power of computation due to which it becomes way better than that of other approaches. This efficiency of the MobilenetV2 enables it to work well in the Systems which are embedded, computer systems that don’t have GPU, and Mobile phones which significantly comprise the result’s accuracy. This Architecture also performs well for web browsers that have hurdles in between them which are comprised of storage, processing of graphics, and computations of algorithms [6, 7].
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
333
Benefits of MobileNetV2 • The MobileNetV2 provides the results with great accuracy while the operations of math and parameters are kept as low as possible so that the deep neural networks can be brought into mobile phones. • It also made an improvement in the state of art model mobile performances on benchmarks and different multiple tasks across various sizes of models of the spectrum. • MobileNetV2 is efficient in extracting the features of the image for the segmentation and the detection of objects. • It delivers a mobile orientation that is efficient and complete training can be applied for claims Google and tasks of visual recognition. • MobileNetV2 acquires the operations which are standard that present in all the neural networks frameworks and also has an inference that is memory efficient.
5 Proposed Methodology Design We aim to develop a unique deep learning model that can be used for the detection of the mask on the face of a human, also it will detect if a face has no mask on it. To start with this data set, we begin with collecting images and sending them to the training script (which is used for transferring the face mask sets that are requested).In continuation of that, before the data gets into the classifier model, the pre-processing of the data is conducted first. For training purposes, the MobileNetV2 library is used from the Keras/TensorFlow field. A Keras/TensorFlow library named MobileNet V2 is utilized for training purposes; For Convolutional neural networks, this classifier that we are using is the improved and better version of it because the procedure of training within the model is considered to be much faster in the field of accuracy along with a marginally improved change. After the completion of the operation of training, it is saved in ‘t5’ format to the disc. For monitoring the model with the training process, a graph is plotted with the help of the Matplotlib library.
6 Architect Overview Figure 2 demonstrates the whole functioning of the proposed model. The first step is importing the dataset after which the OpenCV module is used to start launching video streams and the program then recognizes faces in the video stream, applies the face mask classifier to the face region of interest to decide “mask” or “no mask,” and displays the results in a highlighted box around the face region of interest. The application looks for neighboring faces if a mask is found it is highlighted by a green bounding box otherwise it is highlighted using a red bounding box. CNN is a deep learning technique that uses an image as input. The various characteristics of the image are then mapped to weights and biases, allowing them to be distinguished from one another. The amount of pre-processing required for CNN is substantially smaller than for other classification models, which is why we chose it over other classification methods. The architecture has 4 modules:
334
D. L. Kavya Reddy et al.
• Dataset Collection: We collect a large number of data sets, both with and without facemasks. The greater the number of images the more will be the accuracy. • Extracting Datasets: We can extract the features using Mobile Net V2 on the mask and no mask sets. • Training of the model: With the help of extracted features above and using OpenCV, Keras (a python Library) we will train the model. • Facemask Detection: We can recognize facemasks in both pre-processed images and real-time video. If a person wears a mask it will highlight the person with a green bounding box otherwise it will highlight using a red bounding box [12].
Fig. 2. Complete module of face mask detection workflow [12].
7 Proposed System The proposed system employs the OpenCV, MobileNet V2, Keras, and Python libraries to recognize a person wearing a face mask in an image or video stream using computer vision and deep learning algorithms. Face masks on a person’s face are accurately classified, grouped, and predicted using the categorization and prediction model of the proposed system the developed approach will focus on increasing the accuracy of prediction and the likelihood of detection. This is done with the help of MobileNet V2.Because of the MobileNet algorithm’s implementation. The integrity of the data will be improved, and the time complexity will be decreased. This proposed system monitors people using existing IP cameras from major institutions, making it cost-effective because no additional investment is necessary. In a short amount of time, the suggested system detects
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
335
multiple faces and face masks from all angles. Three steps are followed in the proposed system, the workflow of which is shown in Fig. 3. That includes: 1) Dataset creation and preprocessing 2) Model development and training 3) Implementation of the model. 7.1 Dataset Creation and Preprocessiang Collection of face mask datasets- The proposed system makes use of a custom dataset created using a python script that collects images from sites like Kaggle that are free and open-source, and few open-source image libraries and Google Images, etc.
Fig. 3. The workflow module of face mask detection
336
D. L. Kavya Reddy et al.
Loading of face mask dataset - These images are saved in the “no mask” folder. The mask image was then applied to the images in the dataset to create photographs of people wearing masks, which were saved in a folder labeled with the mask. Data or image preprocessing - It’s necessary to specify the dataset’s directory. A list of categories is constructed by combining two entries, one with and one without a mask. Then, for getting a list of photographs from the dataset registry, introducing the information, and classifying it, two more lists, Data, and labels are constructed. After that, the labels are binarized and saved in NumPy arrays. From the given photos, the ROI for the face component is then retrieved. The next step is Data augmentation that creates many images with a single image by adding various properties like rotating the image, shifting the image, etc. 7.2 Model Development and Training Training of the data - The training data set and the testing data set split into the ratio of 80:20 before the custom face mask image dataset. The Training data contains 80% of images from the dataset to successfully train the algorithm and to ensure prediction exactness, we provide input as well as output for training data (80%), on the other hand, the Testing data set contains 20% of images from the dataset to evaluate the system’s prediction accuracy, the only input is fed to the training data. In the beginning, the learning rate is set to 0.0001 the lower the initial learning rates the better the output. EPOCHS is the total number of passes completed by the entire training algorithm, indicating complete training? The EPOCHS’ value is set to 18 which is the aggregate of passes performed by the entire training algorithm, which indicates the training’s achievement. The value of Batch size is set to 32 to indicate the usage of 32 photos by the training method at once. Updating and saving the model - The trained detector model is then saved to disc as the last step. Many steps from the classifier model’s documentation are included in the training procedure. Binarizing class markings, splitting our dataset, and generating a characterization report are all done with SkLearn. Imutils will help us locate the list of images in our dataset. 7.3 Implementation of the Model • Loading the trained and updated model or classifier. • Detecting faces and masks in video streams through OpenCV - After opening the video stream, using OpenCV, we can detect faces and masks as OpenCV has face detection classifiers. • Image preprocessing (extracting ROI) with OpenCV • Apply face mask detection model with Mobile Net. • Detects Mask and No Mask with percentage. • Display result with Box over the face for detection.
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
337
8 Result, Discussion and Analysis 8.1 Creating Dataset The collection of data is the first step in developing the Face Mask Recognition model. 1915 with a mask and 1918 without mask images are used to train and build the model. The data collected is then divided into two categories: with and without a mask, as shown in Fig. 4 and Fig. 5.
Fig. 4. Shows a folder containing face images with mask
338
D. L. Kavya Reddy et al.
Fig. 5. Shows a folder containing face images without a mask
8.2 Training the Model By retaining an acceptable percentage of various classifications, the dataset is separated into training and testing sets. 80% of the total images in the dataset were being utilized in the training phase and the remaining 20% in the testing phase. To ensure that the model can predict effectively it is trained for 20 Epochs. Table 1 shows the outcome of Epochs for checking the loss and accuracy during model training. Table 1. Iteration of Loss, Accuracy, Val_loss, and Val_accfor 20 Epochs Epoch
Loss
Accuracy
Val_loss
Val_acc
1/20
0.3585
0.8339
0.1076
0.9817
2/20
0.1191
0.9591
0.0593
0.9870
3/20
0.0924
0.9703
0.0472
0.9883
4/20
0.0687
0.9766
0.0388
0.9896
5/20
0.0566
0.9799
0.0369
0.9896
6/20
0.0532
0.9822
0.0321
0.9922
7/20
0.0417
0.9875
0.0304
0.9922
8/20
0.0427
0.9858
0.0302
0.9922 (continued)
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
339
Table 1. (continued) Epoch
Loss
Accuracy
Val_loss
Val_acc
9/20
0.0373
0.9888
0.0278
0.9948
10/20
0.0343
0.9875
0.0291
0.9948
11/20
0.0385
0.9868
0.0278
0.9922
12/20
0.0365
0.9881
0.0262
0.9922
13/20
0.0360
0.9868
0.0248
0.9922
14/20
0.0269
0.9921
0.0250
0.9922
15/20
0.0292
0.9904
0.0249
0.9935
16/20
0.0239
0.9921
0.0234
0.9922
17/20
0.0297
0.9885
0.0244
0.9922
18/20
0.0312
0.9888
0.0248
0.9935
19/20
0.0223
0.9937
0.0217
0.9922
20/20
0.0221
0.9947
0.0231
0.9948
As shown in Table 1 starting with the second epoch, accuracy improves and loss reduces. The graph is meant to illustrate loss and accuracy as the epoch advance. It displays the training algorithm’s accuracy. The training and testing stages’ loss curves
Fig. 6. Graph showing training loss and accuracy
340
D. L. Kavya Reddy et al.
are depicted. The training loss is considerably reduced as the number of epochs increases as shown in Fig. 6. 8.3 Working Prototype Images This script contains the face mask detector model, this detector model is then used to make predictions and calculate the probability of a person wearing or not wearing a mask as shown in Fig. 7 and Fig. 8. This model also detects multiple faces wearing or not wearing a mask at once as shown in Fig. 9 shows the Face-mask detection model detecting a person who is wearing a mask (Fig. 10).
Fig. 7. Result on a real-time image with no mask and mask.
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
Fig. 8. Result on a real-time image with no mask and mask.
Fig. 9. Result on a real-time image with no mask and mask.
341
342
D. L. Kavya Reddy et al.
Fig. 10. Result on Multiple Face Detection with and without the mask.
9 Conclusion As the world is suffering from the COVID-19 pandemic and technologies are blooming with recent evolution, we have developed a novel solution for detecting whether a person has a facemask on or not, which can surely contribute towards public wholesomeness. Our modal can detect or monitor the people through IP cameras which can be considered economically feasible. With the lockdown linked restrictions now being rolled back gradually, the world is trying to return to its normalcy. According to WHO a face mask with other preventive measures can help slow the spread of the virus. Our model can be easily installed for automated monitoring. It can keep a track of people wearing face masks at workplaces, which will help make them safer. Our proposed system detects a face mask with an accuracy of 99.92%. We have used the latest architecture of CNN i.e. MobileNetV2, the only architecture that can work on embedded devices. And also our modal can successfully detect multiple faces at the same time.
10 Future Scope The application developer here can be used in any real-life environment like stations, streets, corporate environments, malls, and so on, also the places where accuracy and precision are worth serving the agenda [11]. Moreover, the plan of action here can be used to innovate smart city, since it will help give the necessary boost required in the development of underdeveloped countries like India. Our analysis of the topical problem at hand presents an opportunity to be more than geared up for the next extremities, or to appraise the scope of social reform [9].
Real-Time Face Mask Detection Using CNN in Covid-19 Aspect
343
References 1. Saxen, F., Werner, P., Handrich, S., Othman, E., Dinges, L., Al-Hamadi, A.: Face attribute detection with MobileNetV2 and NASNET-mobile. In: 2019 11th International Symposium on Image and Signal Processing and Analysis (ISPA), pp. 176–180. IEEE (2019) 2. Lodh, A., Saxena, U., Motwani, A., Shakkeera, L., Sharmasth, V.Y.: Prototype for integration of face mask detection and person identification model–COVID-19. In: 2020 4th International Conference on Electronics, Communication and Aerospace Technology (ICECA), pp. 1361– 1367. IEEE (2020) 3. Vinitha, V., Velantina, V.: COVID-19 facemask detection with deep learning and computer vision. Int. Res. J. Eng. Technol 7(8), 3127–3132 (2020) 4. Yadav, S.: Deep learning-based safe social distancing and face mask detection in public areas for covid-19 safety guidelines adherence. Int. J. Res. Appl. Sci. Eng. Technol. 8(7), 1368–1375 (2020) 5. Bhuiyan, Md.R., Khushbu, S.A., Islam, Md.S.: A deep learning-based assistive system to classify COVID-19 face mask for human safety with YOLOv3. In: 2020 11th International Conference on Computing, Communication and Networking Technologies (ICCCNT), pp. 1– 5. IEEE (2020) 6. Kumar, A., Kaur, A., Kumar, M.: Face detection techniques: a review. Artif. Intell. Rev. 52(2), 927–948 (2019) 7. Lee, D.H., Chen, K.L., Liou, K.H., Liu, C.L., Liu, J.L.: Deep learning and control algorithms of direct perception for autonomous driving. Appl. Intell. 51(1), 237–247 (2021) 8. Memish, Z.A., Zumla, A.I., Al-Hakeem, R.F., Al-Rabeeah, A.A., Stephens, G.M.: Family cluster of Middle East respiratory syndrome coronavirus infections. N. Engl. J. Med. 368(26), 2487–2494 (2013) 9. Lee, J., Lee, S.K., Yang, S.I.: An ensemble method of CNN models for object detection. In: 2018 International Conference on Information and Communication Technology Convergence (ICTC), pp. 898–901. IEEE (2018) 10. Vasantha, K.A., Mohd. Umar, P., Subhralina, N., Bhagya, V., Shobana, P.: Automatic face mask detection using MobileNetV2. Int. Res. J. Eng. Technol. 799–803 (2021) 11. Singh, S., Ahuja, U., Kumar, M., Kumar, K., Sachdeva, M.: Face mask detection using YOLOv3 and faster R-CNN models: COVID-19 environment. Multimed. Tools Appl. 80(13), 19753–19768 (2021). https://doi.org/10.1007/s11042-021-10711-8 12. Suganthalakshmi, R., Hafeeza, A., Abhinaya, P., Ganga Devi, A.: COVID-19 face mask detection with deep learning and computer vision. Int. J. Eng. Res. Technol. 9(5), 73–75 (2021) 13. Rota, P.A., et al.: Characterization of a novel coronavirus associated with the severe acute respiratory syndrome. Science 300(5624), 1394–1399 (2003) 14. Bedi, P., Goyal, S.B., Rajawat, A.S., Shaw, R.N., Ghosh, A.: A framework for personalizing atypical web search sessions with concept-based user profiles using selective machine learning techniques. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 279–291. Springer, Singapore (2022). https:// doi.org/10.1007/978-981-16-2164-2_23 15. Bhadani, A.K., Sinha, A.: A facemask detector using machine learning and image processing techniques. Int. J. Eng. Sci. Technol. 1–8 (2020) 16. Yadav, S., Singh, S., Jana, A., Susan, S.: Shadow detection using GAN with residual CNN discriminator. In: 2020 IEEE 15th International Conference on Industrial and Information Systems (ICIIS), pp. 198–203. IEEE (2020) 17. Bobick, A.F., et al.: The KidsRoom: a perceptually-based interactive and immersive story environment. Presence 8(4), 369–393 (1999)
344
D. L. Kavya Reddy et al.
18. Naveenkumar, M., Vadivel, A.: OpenCV for computer vision applications. In: Proceedings of National Conference on Big Data and Cloud Computing (NCBDC 2015) (2015) 19. Bodapati, S., Bandarupally, H., Shaw, R.N., Ghosh, A.: Comparison and analysis of RNNLSTMs and CNNs for social reviews classification. In: Bansal, J.C., Fung, L.C.C., Simic, M., Ghosh, A. (eds.) Advances in Applications of Data-Driven Computing. AISC, vol. 1319, pp. 49–59. Springer, Singapore (2021). https://doi.org/10.1007/978-981-33-6919-1_4 20. Zhu, N., et al.: A novel coronavirus from patients with pneumonia in China, 2019. N. Engl. J. Med. (2020)
A Smart IoT Based Agro-Management System Using Distributed Ledger and Circular Supply Chain Methodology M. Darshan(B) , Sundeep V. V. S. Akella, S. R. Raswanth, and Priyanka Kumar Department of Computer Science and Engineering, Amrita School of Engineering, Coimbatore, Amrita Vishwa Vidyapeetham, Coimbatore, India {cb.en.u4cse19126,cb.en.u4cse17664, cb.en.u4cse19648}@cb.students.amrita.edu, [email protected]
Abstract. To enhance the entire prospective future of renewing, recycling and reusing we propose a solution. The solution we have been working on consists of an IoT Circular Supply Chain Management and Blockchain based approach to conventional grocery shopping. Our distinctive approach would also enhance hyper-local stores/individual cultivators by helping them get better business and visibility. With this vision forming the base of the project we hope to bring in automation for the future, where smart cities are an industry standard. The IoT device developed during this project has helped us in understanding a lot about how conventional agriculture works and how technology can leverage these principles to raise a better yield using the same resources. To achieve the sustainable circular supply chain methodology. Our innovative approach brings in a drastic change for the packaging sector. The learnings from this proposed solution can be used for a wide range of problems faced in many industries. With various stakeholders involved in this solution, it assures that all parts of this solution work hand-in-hand and will lead to a better and connected tomorrow. Keywords: Internet of things · Circular supply chain · Blockchain · Recycling and reusing
1 Introduction As demand rises for better and quality products, it has become even more difficult to locate these products. Many people are unaware of the very small to massive things that are being sold at their hyper-local store and end up paying more for a product found online. Indian demand for groceries has always been increasing and in recent times online delivery systems of these grocery items have come very handy. For all the working class people online grocery shopping was a blessing in disguise. One major problem with the entire online grocery system is that it has a direct impact on the locals who sell all the similar products and yet they barely are able to make a living out of it. Such impact in the larger perspective has a negative impact on the economics of the state. With many © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 345–354, 2022. https://doi.org/10.1007/978-981-19-1677-9_31
346
M. Darshan et al.
corporate companies entering this industry, basic rights of such local business people are affected for the worse. The proposed solution tries to challenge the existing system by placing such local business people as one of the main stakeholders. With rising demands for grocery services there is a lack of platform for individual cultivators to sell their produce in an effective manner. This has been seen as a major concern and the proposed solution tries to eliminate this problem by using an app powered by Blockchain technology and tries to store both the user’s and the consumer’s data in a ledger based format. This way transparency and security is ensured. IoT plays an important role in maximising the profits in terms of both monetary value and produce quantity, but it is not utilized to its maximum at grass-root level. People still stick to conventional methods of farming and irrigation. In a recent study it was found that drop irrigation produces higher yield and uses minimal resources rather than spraying water on the crops. In order to favour such conditions for individual cultivators and help them analyze and understand their crop better, the proposed solution makes use of IoT. This way the cultivators are able to monitor their produce anytime they wish and it becomes even more convenient for them. Getting accurate measurements about soil humidity and temperature can help them in further data analysis and get better results in terms of which crops work out the best based on the features given. Lastly, the proposed solution mimics the entire sustainable supply chain model by giving a lot of importance to the three R’s that are reusing, recycling and reducing in the same priority order. Reusing is the best way to move forward with increasing demand and to meet the expected growth demand other factors come into play. The proposed solution enhances a mindset change in terms of how packaging materials have been used. The solution helps users to return the packaging material in good condition back to the suppliers in return for credits which can be used for shopping later on the same platform. Steps including recycling and reducing are economically expensive in the present day scenario but this research work could push work in this sector. With the perfect balance of IoT, Blockchain Technology and Sustainable Circular Supply Chain methodologies we hope to bring in automation for smart cities in the future that help to bridge the gap between conventional agriculture and cultivation for the upcoming decade.
2 Literature Survey In the research work presented by, “S.Jaiganesh, K.Gunaseelan and V.Ellappan” on “IoT agriculture to improve food and farming technology”, where they have explored various methodologies of using different forms of sensors in the agricultural sector to improve the yield. Their work pushes the limits of research in these areas of technology which can have a direct impact on ground level cultivators. The distinctive methods followed in their proposal proves that experimental analysis and simulation can produce better yield in terms of the final produce [1]. In the research work presented by, “Mohamed Ben-Daya, Elkafi Hassini and Zied Bahroun” on “Internet of things and supply chain management: a literature review” various aspects of IoT technology in supply chain have been studied in detail. It also plunges into various sectors of everyday industry that can use this change for the better.
A Smart IoT Based Agro-Management System
347
Many techniques have been discussed in their solution like quality checking devices and package tracker in real-time. Their work has provided our proposed solution with strong fundamentals and has opened new avenues for research for the same [2]. After careful study of the research work on “Blockchain Technology and its relationship to sustainable supply chain management” by “Sara Saberi, Mahtab Kouhizadeh, Joseph Sarkis and Lejia Shen” we conclude that blockchain is the perfect missing piece in the puzzle of global supply chain management problems. It also provides greater insights into the entire industry and what sort of problems it faces right from interorganisational, intra-organisational, technical and external barriers. With all the features of blockchain such as transparency, security and traceability and advantages of IoT technologies the proposed solution helps in bridging the gap between agriculture and technological sectors [3]. With insights from the meticulous work by “Juhi Shree, N R Kanimozhi, G A Dhanush, Anand Haridas, Anala Sravani and Priyanka Kumar” on “To Design Smart and Secure Purchasing System integrated with ERP using Block chain technology” the proposed solution draws inspiration from automatic self-billing process and transparency of the accounts can be maintained. IoT systems play a major role in understanding the needs of the user and based on inputs the proposed solution tries to bring the best out of both the technologies. Item tracking done using an integrated RFID (Radio Frequency identification) can be incorporated at later stages of the proposed solution [4]. From the astonishing works of “K. Sreelakshmi, S.Akarsh, R. Vinayakumar and K.P Soman” on “Capsule Neural Networks and Visualization for Segregation of Plastic and NonPlastic Wastes” the proposed solution tries to mimic the findings of the above mentioned research work. With recycling, reusing and renewing being the core structure of the proposed solution it is very important that packaging materials are not wasted and instead need to be reused [5]. From the work “The influence of environmental attitudes and perceived effectiveness on recycling, reducing, and reusing packaging materials in Spain” by “José-JuliánEscarioCarlaRodriguez-SanchezLuis V.Casaló” it is vivid that the three R’s of waste management can have a tremendous positive impact for the entire environment. From the research work astonishing figures of carbon emission reduction can be noticed. The research work delves into the socio-economic aspects of recycling, reusing and reducing [6–8].
3 Proposed System With the introduction of Industry 4.0 and Sustainable Development Goals, the world has been moving towards the greener phase of production and consumption. The linear methodology of supply chain is generating exceedingly large amounts of waste in each level of haulage which leads to an economic deadend and volatility. With the increase in the usage of raw materials, there are numerous drawbacks in the linear supply chain model due to resource scarcity as well as environmental impacts. The ‘3 R’s’ of sustainability has been reframed to ‘4 R’s’ which includes - Reduce, Reuse, Recover, Recycle which gives finer results in sustainable waste management. The above specified methodology could be achieved by incorporating a circular supply chain model which would create a closed loop between the consumer and producer thereby ensuring zero waste environment in the near future.
348
M. Darshan et al.
3.1 Sustainable Supply Chain Towards the Circular Economy The interest in the ideology of circular economy has evolved in recent years due to the crucial reduction of excessive usage of natural resources by continuously transmitting the resources, energy and waste in the form of closed loop which results in recovery and reuse of raw materials once again for the production. This has created a powerful impact in the three dimensions of sustainability (social, economic and environmental) which has set off a new greener society. The agro-organic ecosystem focuses on developing enterprises that are sustainable and polyphonic towards the environment. With implementation of a circular supply chain along with leveraging technologies, the organic ecosystem is able to achieve the sustainable development goals. The key stakeholders in the chain includes Producer, Consumer, Logistics and waste management (Figs. 1 and 2).
Fig. 1. Comparison of traditional linear and circular supply chain
4 Proposed Architecture A web based application would be provided for common interaction of producer and consumer in order to eliminate conventional opaque systems and to provide better mode of communication between two entities. After successful login, the producer’s details will be verified and publicized in the application. The necessary product details will be displayed on the dashboard to keep a track of the progression. With the interactive dashboard, the producer would be comprehended about the classification of the soil which are labelled on the basis fertility. The estimated details regarding the outcome would be displayed for enhancing the production sector. A humidity and temperature checker has also been installed which helps the producer to respond promptly according to the classified outcome.
A Smart IoT Based Agro-Management System
349
Fig. 2. Schematic representation of proposed system
4.1 Implication of IoT in the Proposed System In the journey towards smart city development, the Internet of Things (IoT) plays an important role in transforming “Conventional Technology” to “Smart Computing”. A Smart IoT based agro-product haulage system is a sustainable farming methodology with IoT to maximize the productivity to meet the growing world’s demands. Ideal environmental conditions are a predominant factor in the organic farming sector and physical examination do not yield more accurate results. With the implication of IoT and modern automation techniques, human supervision could be diminished and could notify them to make more accurate decisions to deal with dynamic weather and soil conditions. The IoT system could be divided into two subunits namely - Sensing and Processing unit. The sensing unit consists of Soil humidity, air humidity sensors, DHT11 sensor and a temperature sensor for enabling serial communication to the Microcontroller in Arduino uno. Arduino uno is a microcontroller board based on ATmega328 which includes 6 analog pins, 14 digital input/output pins, USB connection etc. An open source electronics computing platform is provided and utilized to program the microcontroller according to the logical requirement. The purpose of the sensing unit is to capture the physical variables according to the environmental factors of the production area. The DHT11 sensor plays a pivotal role in regulating the environmental conditions. Food quality plays a pivotal role in organic agriculture and maximizing the quality factor yields
350
M. Darshan et al.
additional profits to the producer. The DHT11 is an anti-interference sensor which indicates unwanted temperature changes and humidity levels to avoid privations. Usage of a fast-response DHT11 sensor will allow the producer to handle relative humidity changes in correlation with type of the crop. The soil humidity sensor comprises two conductive elements whose terminals would be embedded into the ground for transmission and obtaining electrical signals. These charges could have higher or lower conductivity which would be determined by the level of humidity in the soil. This measure would be used to control humidity level of the soil which would prevent over-watering of the produce. All the measures from the sensing unit would be passed to the processing unit for exploratory data analysis. The data obtained from the system would be visualized for graphical representation of data which would be displayed on the web portal. Using the insightful information, the producer could maximize his production level in order to meet the market demands. 4.2 Implication of Blockchain in the Proposed System Blockchain technology has the power to bring a radical change in the agro-organic ecosystem by providing constructive results to the key challenges. The two way portal of the proposed system has been built to facilitate the communication between consumer and producer. After exploration of the products, the consumer can directly call the producer or enter the required quantity. All the necessary details regarding the producer would be displayed on the responsive dashboard to keep a track of the progress. With the implication of Blockchain, secured ledgers would be maintained to record the status of the product. The decentralized service has resulted in providing financial incentives to the sustainable farming techniques rather than unequal recompensation for yield. The circular economy of the agro-organic ecosystem primarily focuses on less wastage and maximum produce due to recovery and reuse of the raw materials. With the extension of this methodology, a large amount of data has been generated and existing centralized databases are inefficient in handling them. With blockchain, every segment of the supply chain could be traced and access of information is preserved using private keys. Each asset is traceable and immutable which provides security and visibility in the whole supply chain. Waste management increases potentially through bringing in blockchain based proposed system for the circular sustainable supply chain. The consumer would be provided with slots for waste collection and systematic smart logistics would be carried out in order to enter the waste management phase. The closed loop of the chain guarantees an accurate supply-demand algorithm which improvises the consumer-producer relationship with specific data stored on blocks to match the demands thereby avoiding the problem of overproduction (Fig. 3).
A Smart IoT Based Agro-Management System
351
Fig. 3. Workflow of the proposed solution.
5 Implementation After the installment of the IoT device, the whole system is first powered up. With proper conditions, accurate measurement values are recorded by the soil humidity, air humidity sensors and DHT11 sensors. Firstly, the recorded values are sent to the microcontroller of the Arduino Uno and then server scripted to the web portal of the producer. With help of the processed values, the producer selects the suitable crop according to the present environmental conditions and enters the manufacturing phase of the chain (Fig. 4). A computerized smart contract has been developed to reduce convolution of the circular supply chain by providing a traceability of information to both producer and consumer. All the contracts would be compiled and migrated to the ethereum blockchain using development frameworks such as truffle and ganache-cli [1]. The necessary information regarding the product and supplies would be digitized and stored on the decentralized storage thereby eliminating untraceable and insecure methods over the chain. Once the contract is deployed, new blocks would be generated in the ganache-cli and instances of the contract could be tested using truffle [2] (Fig. 5).
352
M. Darshan et al.
Fig. 4. Digital twin of the Arduino-based circuit.
Fig. 5. Compilation of contracts
6 Result The proposed system is a correlation of IoT and Blockchain which has been utilized to provide a smart agro-product haulage model. The whole system is supervised and computed with the robust sensing and dynamic interpretation of IoT technology. All the necessary details of the supply chain have been recorded and stored as state variables
A Smart IoT Based Agro-Management System
353
in the ethereum network using blockchain to ensure traceability and immutability. The above specified system undisputedly provides a trust factor in the relationship among producer and consumer (Fig. 6).
Fig. 6. Migration of contracts
7 Conclusion The essential task of the proposed system is to provide an accumulative solution to the agro-organic ecosystem through leveraging new technologies such as IoT and Blockchain. Implication of IoT in the smart agro product haulage system facilitates data management and intelligent applications which provides a gateway for future enhancements. With the current consensus algorithm of ethereum, sophisticated business logic has been enacted which provides immutable and auditable data to ensure a traceable circular and sustainable economy. We can affirm that our proposed system is not trying to reinvent the wheel but the methodology would definitely fix the already broken wheel. As future work, we plan to upgrade the constrained hardware to produce finer performance and improvise the contract in order to appraise the appropriateness of the proposed system.
References 1. Kirubanand, V.B., Rohini, V., Laxmankumar, V.: Internet of things in agriculture to revolutionize traditional agricultural industry. In: ITM Web of Conferences, vol. 37, p. 01018 (2021). https://doi.org/10.1051/itmconf/20213701018
354
M. Darshan et al.
2. Ben-Daya, M., Hassini, E., Bahroun, Z.: Internet of things and supply chain management: a literature review. Int. J. Prod. Res. 57(15–16), 4719–4742 (2019) 3. Lin, W., et al.: Blockchain technology in current agricultural systems: from techniques to applications. IEEE Access 8, 143920–143937 (2020). https://doi.org/10.1109/ACCESS.2020. 3014522 4. Shree, J., Kanimozhi, N.R., Dhanush, G.A., Haridas, A., Sravani, A., Kumar, P.: To design smart and secure purchasing system integrated with ERP using block chain technology. In: 2020 IEEE 5th International Conference on Computing Communication and Automation (ICCCA), pp. 146–150 (2020). https://doi.org/10.1109/ICCCA49541.2020.9250767 5. Sreelakshmi, K., Akarsh, S., Vinay Kumar, R., Soman, K.P.: Capsule neural networks and visualization for segregation of plastic and non-plastic wastes. In: 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS), pp. 631–636 (2019). https://doi.org/10.1109/ICACCS.2019.8728405 6. Bin Abdul Hamid, D.S., et al.: Application of deep learning with wearable IoT in healthcare sector. In: 2021 IEEE 6th International Conference on Computing, Communication and Automation (ICCCA), pp. 697–701 (2021). https://doi.org/10.1109/ICCCA52192.2021.966 6240 7. Dhanush, G.A., Raj, K., Kumar, P.: Blockchain aided predictive time series analysis in supply chain system. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 913–925. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0749-3_70 8. Kumar, P., Dhanush, G.A., Srivatsa, D., Nithin, A., Sahisnu, S.: A buyer and seller’s protocol via utilization of smart contracts using blockchain technology. In: Luhach, A.K., Jat, D.S., Hawari, K.B.G., Gao, X.-Z., Lingras, P. (eds.) ICAICR 2019. CCIS, vol. 1075, pp. 464–474. Springer, Singapore (2019). https://doi.org/10.1007/978-981-15-0108-1_43
A Sub-multilevel Topology of 7-Level Cascaded H-Bridge Inverter Sushree Smrutimayee Barah1(B) , Swaraj Pati1 , and Sasmita Behera2 1 Department of Electrical Engineering, Veer Surendra Sai University of Technology, Burla,
Sambalpur, India [email protected] 2 Department of Electrical and Electronics Engineering, Veer Surendra Sai University of Technology, Burla, Sambalpur, India
Abstract. The conventional design of the H-bridge inverter starts with four controlled switches in a module and the cascaded modules raise the level of the inverter voltage output. Multilevel inverter (MLI) 7-Level structure has been presented that has less count of switching devices. The configuration is based on the submultilevel principle. This MLI has the output level is the same as the conventional method. The only difference is that it contains fewer switching devices. Due to having fewer switching devices, there is a reduction in power loss across switches which can cut down the expenses for the inverter design. Both the conventional and modified inverters are simulated for five and seven-level using MATLAB /Simulink for both R and R-L load. Here the conventional and optimal H- bridge inverter configuration is contrasted. The pulses are generated by pulse disposition pulse width modulation (PDPWM). The voltage and current THD has considerably reduced in comparison to existing literature. This is because of the increase of level from 5-level to 7-level keeping the count of switching devices as it is. Keywords: Total Harmonic Distortion (THD) · 5-level CHB inverter · 7-level CHB inverter · Modified 7-level CHB inverter · Multilevel inverter (MLI)
1 Introduction An elementary construction of a bi-level inverter displays a rectangular alternation. However sinusoidal ac is prevalently supplied to the load that has less THD. MLI comes up with a solution to provide quasi-sinusoidal output and reduce THD. Multiple lowrated DC sources are used as input to produce the needed AC voltage level at the output. Therefore, for an MLI, the output voltage is stepped of three and above voltage levels. With the increase in the level the waveform smoothers. MLIs can be classified as CSI or VSI. Multilevel VSI is predominant than CSI due to protection issues with its inherent very high fault current [1]. There is four most commonly used anatomy of MLI; DC, NPC, FC, and cascaded H-bridge (CHB) MLI. CHB MLI has preferably low THD [2]. They differ in the supply of the MLI’s input voltage and the PWM control for gating. MLI is in use for large voltage and large power applications. To meet the challenges such as high dv/dt causing doubling effect, %THD to comply, excess electromagnetic © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 355–364, 2022. https://doi.org/10.1007/978-981-19-1677-9_32
356
S. S. Barah et al.
interference (EMI), common-mode voltage rise, and requirements of synthesizing higher voltages for modern industrial applications have subsequently led to the development of various inverter topologies [3, 4]. The foremost concept of series H-bridge (termed as CHB) inverter was patented by Baker and Banister [5]. The drawbacks of DC and FC structures such as extra capacitors and diodes are surpassed in a hybrid structure [6]. This imparts high output voltage from multiple lower-rated inputs. Still higher voltage can be delivered using the low-rated elements and modules in cascade. A review on MLI structure, control, application, and that too with fewer switching devices is put forth within [6]. Because of its ungrounded dc sources, photovoltaic generation can be connected via CHB MLI to an ac grid respecting the power quality [7]. The cascade inverter is used in many applications like regenerative drives and electric vehicles. MLI configurations, control, applications were surveyed thoroughly by Rodriguez et al. [8]. For photovoltaic (PV) power integration MLI structures were thoroughly surveyed by Pharne et al. [9]. A new MLI setup with fewer controlled switching devices was presented by Ebrahimi et al. [10]. However, their structure is a modification of the FC type MLI, where the capacitor is the weakest element to sacrifice reliability. Different designs of MLI have been developed by Najafi et al. [11], Nedumgatt et al. [1]. A CHB structure with reduced switching devices is proposed by Ebrahimi et al. [12]. A CHB with the regenerative operation and a reduced count of switching devices were presented by Lezana et al. [13]. In an analogous structure, Babaei et al. [14] designed hybrid MLI for a large-voltage system. Observing the merits, CHB MLI is considered in this work as it has a better modular structure and the absence of capacitors improves its reliability. A reduced optimal architecture of CHB MLI is articulated in which the output level is the same using fewer switches. This lessening of switches has a multidimensional impact in minimizing the power loss across switches which can cut down the expenses for the inverter design and improve the efficiency of conversion. The flow of the paper leads to Sect. 2 distinguishing the modification of conventional CHB structure with fewer switches, in Sect. 3 we will talk about the PWM control and in Sect. 4 we will talk about the simulation findings that entail to some conclusion.
2 Multilevel Cascaded H-Bridge Inverter 2.1 Operation of Conventional CHB Multilevel Inverter The resultant ac voltage of a CHB MLI will indicate m levels for k dc inputs. Where the number of steps in the output is m. m = 2K + 1
(1)
Number of inputs/Number of units (K) = (m − 1)/2
(2)
Number of switching devices (n) = 2m − 2
(3)
Or it can be written as
A Sub-multilevel Topology of 7-Level Cascaded H-Bridge Inverter
357
Fig. 1. Structure of CHB MLI
Maximum voltage = KVdc
(4)
Conventionally in MLI, each unit contains a dc source, SDCS, and each H-bridge has 4 switching devices that constitute one unit. From each unit, the output ac voltage is stepped from +Vdc , 0, and −Vdc . On the state of S1 and S4 delivers voltage +Vdc whereas on the state of S3 and S2 delivers a voltage −Vdc . The resultant voltage goes to zero when all of the devices in a unit are turned on. The total voltage output is the sum of individual unit voltages Vok , for K = 1, 2, … as they are connected in series. That is Vo = Vo1 + Vo2 + Vok
(5)
Two such units are in series for the 5-level CHB structure. If the inputs are the same in rating as shown in Fig. 1 the voltage Vmax is 2Vdc and so on. Similarly, the 7-level CHB MLI has 3 H bridge units in cascade and 12 switches as articulated in Fig. 2. The dc inputs have 3 different voltage sources. The outputs of the H-bridge units are cascaded such that the superposed voltage waveform aggregates all of the outputs and it is Vo = Vo1 + Vo2 + Vo3
(6)
For the H-bridge 7-level structure, the device switching is given in Table 1 where 0 and 1 indicate inactive and active devices respectively. 2.2 Operation of Modified CHB MLI A new topology is introduced called a sub-multilevel inverter that has a fewer count of switching devices compared to conventional. Figure 3 depicts the elementary configuration for a sub-multilevel inverter. Its elementary unit includes two switching devices
358
S. S. Barah et al.
instead of four and a voltage source as discussed in the earlier subsection. The restriction put here is that the devices are alternately active in a unit and all the dc inputs are identical if they are to be cascaded. This architecture imparts either zero or +Vdc for a single unit or Vo in cascade. To get the full cycle of ac output the full-bridge topology is added as shown in Fig. 4. The devices of the full-bridge S1 , S2 are active during the positive half between 0 to T/2, where the resultant voltage is Vo . At T/2 S1 , S2 are gated inactive and S3 , S4 are active during the negative half from T/2 to T where the resultant voltage is −Vo , here the time period is T (Table 2).
Fig. 2. Structure of conventional 7-level MLI
3 Modulation Technique The controlled devices of the MLI conduct when they are gated with a pulse. Herein, pulse disposition pulse width modulation (PDPWM) a type of SPWM is used. Multiple carrier signals in MLI create various possibilities of mutual locations of those signals. To create the gate pulses in SPWM, a sine wave is matched to a triangular carrier (Fig. 5). In multiple carrier techniques such as PDPWM m-1 no. of carriers are shifted in the phase where m is defined in Subsect. 2.1. Thereby some of the harmonics are discarded in the output. The gating signal is shown in Fig. 6 for a 7-level inverter, one sine wave is compared with 6 triangular waves by using a relational operator. If the sine wave magnitude exceeds the carrier magnitude the output is 1 otherwise 0. In this way, the
A Sub-multilevel Topology of 7-Level Cascaded H-Bridge Inverter
359
Fig. 3. The architecture of cascaded sub-multilevel inverter
Fig. 4. Modified 7-level cascaded MLI
gate pulse is generated for the 8 switches duplicating the same pulse for S1 , S2, and S3 and S4 constituting a total of 6 pulses.
4 Simulation Result MLI of two levels (five and seven) are taken for the study. Models for the conventional and reduced structure are developed with loads of R-L & R. The harmonic spectrum contained in the voltage output was analyzed using MATLAB/Simulink. The dc source
360
S. S. Barah et al. Table 1. Device active/inactive condition for the 7-level conventional MLI
Modes
S0
S1
S2
S3
S4
S5
S6
S7
S8
S9
S10
S11
−Vdc
0
1
1
0
0
0
1
1
0
0
1
1
−2Vdc
0
1
1
0
0
1
1
0
0
0
1
1
−3Vdc
0
1
1
0
0
1
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
3Vdc
1
0
0
1
1
0
0
1
1
0
0
1
2Vdc
1
0
0
1
1
0
0
1
0
0
1
1
Vdc
1
0
0
1
0
0
1
1
0
0
1
1
Fig. 5. PDPWM technique of 7-level MLI
Table 2. Device active/inactive condition for the modified 7-level structure Modes
S0
S1
S2
S3
S4
S5
S6
S7
−Vdc
0
0
1
1
0
1
0
1
−2Vdc
0
0
1
1
1
0
0
1
−3Vdc
0
0
1
1
1
0
1
0
0
0
0
0
0
0
0
0
0
3Vdc
1
1
0
0
1
0
1
0
2Vdc
1
1
0
0
1
0
0
1
Vdc
1
1
0
0
0
1
0
1
A Sub-multilevel Topology of 7-Level Cascaded H-Bridge Inverter
361
Fig. 6. Individual pulse of PDPWM technique of modified 7-level cascaded MLI
voltage is taken 120 V for 5-level and 80 V for 7-level configurations to keep the output ac at the same value. The rest of the parameters are referred to from [15]. The current and voltage waveforms are plotted in Fig. 7, 8, 9 and 10 for resistive load and Fig. 11 and 12 for R-L load (Tables 3 and 4).
Fig. 7. Voltage output for conventional 7-level MLI
Fig. 8. Current output for conventional 7-level MLI
362
S. S. Barah et al.
Fig. 9. Voltage output for modified 7-level MLI
Fig. 10. Current output for 7-level conventional MLI
Fig. 11. Voltage output for 7-level conventional MLI with load RL
Fig. 12. Modified 7-level MLI Output voltage with RL load
A Sub-multilevel Topology of 7-Level Cascaded H-Bridge Inverter
363
Table 3. Comparison between conventional 5-level and 7-level configurations Parameters
Conventional method 5-level
7-level
No. of voltage sources
2
3
No. of switches
8
12
Voltage THD for R load (%)
23.14 [15]
16.73
Current THD for R load (%)
23.14 [15]
16.73
Voltage THD for RL load (%)
23.51 [15]
16.81
Current THD for RL load (%)
23.51 [15]
16.81
Table 4. Comparison between modified 5-level and 7-level configurations Parameters
Conventional method 5-level
7-level
No. of voltage sources
2
3
No. of switches
6
8
Voltage THD for R load (%)
17.66 [15]
15.31
Voltage THD for R load (%)
28.98 [16]
17.02 [16]
Current THD for R load (%)
17.66 [15]
15.31
Voltage THD for RL load (%)
17.63 [15]
15.24
Voltage THD for RL load (%)
33.79 [16]
18.99 [16]
Current THD for RL load (%)
17.62 [15]
15.21
5 Conclusion When the voltage output of the 5-level and 7-level inverters are analyzed, For the 7-level sub-multilevel inverter it is observed to be closer to a sinusoidal waveform and it has less harmonic component than the 5-level inverter. The improved architecture has the benefit of using 25% fewer switching devices than the standard H bridge model. As a result, costs are reduced, device losses are reduced, and efficiency is increased. The modified structure for still higher levels can reduce the harmonics and a give a significant saving of device cost and improve reliability with reduced devices. The THD obtained could be improved with modification of the modulation or control technique. Closed-loop control with the application with renewable power generation is the target for the future.
References 1. Rodriguez, J., Lai, J.S., Peng, F.Z.: Multilevel inverters: a survey of topologies, controls, and applications. IEEE Trans. Ind. Electron. 49(4), 724–738 (2002)
364
S. S. Barah et al.
2. Babaei, E., Laali, S., Bayat, Z.: A single-phase cascaded multilevel inverter based on a new basic unit with reduced number of power switches. IEEE Trans. Ind. Electron. 62(2), 922–929 (2014) 3. Manjrekar, M.D., Lipo, T.A.: A hybrid multilevel inverter topology for drive applications. In: Proceedings of Thirteenth IEEE APEC 1998, Anaheim, CA, USA, vol. 2, pp. 523–529 (1998) 4. Baker, R.H., Bannister, L.H.: Electric power converter. U.S. Patent 3 867 643 (1975) 5. Derakhshanfar, M.: Analysis of different topologies of multilevel inverters. Master of Science Thesis submitted to Department of Energy and Environment, Chalmers University of Technology, Sweden (2010) 6. Wang, Y., Li, Y., Cao, Y., Tan, Y., He, L., Han, J.: Hybrid AC/DC microgrid architecture with comprehensive control strategy for energy management of smart building. Int. J. Electr. Power Energy Syst. 101, 151–161 (2018) 7. Pharne, I.D., Bhosale, Y.N.: A review on multilevel inverter topology. In: Proceedings of IEEE ICPEC, Dindigul, India, pp. 700–703 (2013) 8. Ebrahimi, J., Babaei, E., Gharehpetian, G.B.: A new topology of cascaded multilevel converters with reduced number of components for high-voltage applications. IEEE Trans. Power Electron. 26(11), 3109–3118 (2011) 9. Najafi, E., Yatim, A.H.: Design and implementation of a new multilevel inverter topology. IEEE Trans. Ind. Electron. 59(11), 4148–4154 (2011) 10. Nedumgatt, J.J., Kumar, D.V., Kirubakaran, A., Umashankar, S.: Multilevel inverter with reduced number of switches. In: Proceedings of IEEE Students’ Conference on Electrical, Electronics and Computer Science, Bhopal, India, pp. 1–4 (2012) 11. Ebrahimi, J., Babaei, E., Gharehpetian, G.B.: A new multilevel converter topology with reduced number of power electronic components. IEEE Trans. Ind. Electron. 59(2), 655–667 (2011) 12. Lezana, P., Rodríguez, J., Oyarzún, D.A.: Cascaded multilevel inverter with regeneration capability and reduced number of switches. IEEE Trans. Ind. Electron. 55(3), 1059–1066 (2008) 13. Babaei, E., Hosseini, S.H.: New cascaded multilevel inverter topology with minimum number of switches. Energy Convers. Manag. 50(11), 2761–2767 (2009) 14. Barah, S.S., Behera, S.: An optimize configuration of H-bridge multilevel inverter. In: Proceedings of 1st IEEE ICPEE, pp. 1–4 (2021) 15. Mallikarjun, C., Shanbog, N.S., Modi, S.: Comparative Analysis of Conventional and Modified H-Bridge Inverter Configuration. arXiv preprint arXiv:1908.10032 (2019) 16. Mallikarjun, C., Shanbog, N.S., Modi, S.: Comparative Analysis of Conventional and Modified H-Bridge Inverter Configuration. arXiv preprint arXiv:1908.10032, August 2019
A Comparative Harmonic Suppression Analysis of Single Phase Half Bridge Grid Connected Inverter in Rotatory Frame of Reference with Integral Control Strategy Hassan Khalid1,2(B) , Saad Mekhilef2,1 , and Marizan Binti Mubin1 1 Power Electronics and Renewable Energy Research Laboratory (PEARL), Department of
Electrical Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia [email protected] 2 School of Software and Electrical Engineering, Swinburne University of Technology, Melbourne, VIC 3122, Australia [email protected]
Abstract. In distributed power generation systems energized by multiple renewable energy systems, the harmonics injected by grid connected inverters are seriously affected by the grid uncertainties at the point of common coupling. As several inverters are tied with a single grid, the grid impedance can cause more harmonics and serious damages. Hence at this stage phase locked loop based on phase estimation method must be considered. This paper therefore proposes Integral Control Phase Estimation (ICPE) method as a better solution than Sine Wave Pulse Width Modulation (SPWM) using single phase half bridge grid tied inverter. Three parameters including voltage, phase and frequency are normally used from the grid side converter (GSC) to synchronize battery powered machine side converter (MSC). Here the framework is implemented by estimating only grid side phase and frequency using dynamic quadrature (DQ) rotating frame of reference and thus forming the phase-locked loop (PLL). The THD thus obtained from our proposed technique is around 0.3%, which dominates conventional SPWM. Keywords: Grid-tied inverter · Half bridge inverter · Harmonics · Integral control and THD
1 Introduction For the last two decades the renewable energy sources like photovoltaic (PV) cells and wind energies have gained a lot of attention owing to their advantages [1]. Since wind energy depends mainly on the wind speed therefore requires complex power electronic circuits to synchronize grid power. On the other hand, solar power stations solve this instant problem by installing batteries. Solar power stations are installed to supply excess power to the grid stations and they used inverter schemes to convert DC power into AC power. Since these inverters are tied with the grid station therefore, known as grid tied © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 365–375, 2022. https://doi.org/10.1007/978-981-19-1677-9_33
366
H. Khalid et al.
inverters. Inverters are categorized as single phase and three phase and deep literature has been already presented in [2, 3] regarding grid interconnection. Much of the research work presented in the past have focused in controlling the grid tied inverters based on high order filters as commonly employed in distributed power generation systems (DPGSs). To ensure the stable operation of these filters, LCL resonating filters damped by using delay dependent loop control [4, 5], notch filters as active damping (AD) [6], capacitor current feedback based AD [7] or multi loop feedback systems [8]. For the high order filter design such as LLCL and LCL with multiple traps, their resonance have been resolved by using feedback systems [9]. However, even in practical cases the stability of DPGSs can be quite challenging due to non-ideal point of coupling (PCC) behavior [10]. Such grids are termed as weak grids having non-negligible grid impedance [11]. Due to the uncertain variation in grid impedance makes system unstable and unreliable. Therefore, conventional current controllers, damped LCL filters and feedforward of PCC voltage will also make it impossible to tie the inverter [12]. Therefore, impedance reshaping have been practiced in [13] to make inverter behave as a passive device. Current controllers connected to photovoltaic (PV) inverters ensures a very precise energy exchange between grid and consumer by preventing fault current to let through [14]. In case of incorporating week or distorted grid, or providing auxiliary active power filtering device, synchronous or static frame references, multi-resonant (MR) regulators are often installed with high gain and wide band gap to ensure the steady state current tracking attenuation factor [15]. However, a unified methodology to derive coefficients of MR regulator in frequency domain is not yet developed. Therefore, applying current controllers and their evaluation is quite complex and challenging. However, various studies showed that due to limited PLL bandwidth it cannot be incorporated to locate disturbance injected by harmonics. Therefore, Jinming Xu et al. in [16] showed PLL to be highly effective in suppressing harmonics. Recently, Subhendu Bikash et al. in [17], was able to present grid tied inverter based on PWM technique was able to minimize leakage current not less than THD = 6.5%. Shiqi Jiang et al. in [18], using conventional LCL-LC filter offered 0.91% harmonic distortion, while the modified LCL-LC offers 0.96%, simple LCL offers 1.52% and modified LLCL to offer 1.09% total harmonic distortion. Similar studies based on other types of multilevel inverters have also been presented in [19, 20]. In this paper a reduced order, single phase half bridge inverters based LC resonating tuning tank circuit is implemented to synchronize grid using PLL using based on phase and frequency estimation technique. In order to make inverter independent of grid abnormalities PLL second order generalized integration technique is used to generate orthogonal signals. Since half bridge inverters has higher magnitude of THD as compared to H-bridge or any other type of configuration. Therefore, we have selected this type for our study. The analysis resented here shows that the as the grid gets align with the consumer side the reactive powers become almost negligible. For analysis MATLAB Simulink is used for detailed analysis.
A Comparative Harmonic Suppression Analysis
367
2 Model Description The model consists of two main parts. The grid side control (GSC) and machine side control (MSC). Inside the GSC model, the orthogonal system is implemented using MATLAB as shown in Fig. 1. It utilizes the basic structure of SOGI as discussed in [21]. The transfer function of SOGI is defined by Eq. 1 and has infinite gain. With the help of SOGI we can split a single phase grid side voltage into two components. One known as Vα and the second is termed as Vβ . Both of these signals are said to have angular displacement of π2 . This controller takes two inputs, one is the estimated frequency from the machine side controller block and the other one is the main grid AC voltage. The dynamic and quadrature signal gain controller is defined by Eq. 2 and Eq. 3. GSOGI =
ωn s s2 + ωn2
(1)
Gd =
kωn s s2 + k 2 ωn s + ωn2
(2)
Gq =
kωn2 s2 + k 2 ωn s + ωn2
(3)
Fig. 1. Grid side controller model
From Eq. (1) to Eq. (3), ωn is termed as unamortized natural frequency of the SOGI and is set equal to estimated frequency. Whereas, k is the gain that affects the SOGI bandwidth [22]. The internal block diagram of implemented GSC is shown in Fig. 1. In extreme left, a voltage signal from distribution grid side is fed to get split into two stationary frame of reference signals displaced by π/2, named as Vα and Vβ . Consider the internal block diagram of machine side converter in Fig. 2. The signals from stationary frame of reference are transformed into rotational frame of reference Vdq model using park transformation [23] Eq. 4 with Vo as a zero sequence voltage component. ⎡ ⎤ ⎡ ⎤⎡ ⎤ Vd cos(ωn t) sin(ωn t) 0 Vα ⎣ Vq ⎦ = ⎣ −sin(ωn t) cos(ωn t) 0 ⎦⎣ Vβ ⎦ (4) V0 V0 0 0 1 Inside the machine side converter as shown in Fig. 2, the stationary frame reference voltage Vαβ are transformed into dynamic frame of reference Vdqo . This blockestimates
368
H. Khalid et al.
the required phase and frequency for the given frame of reference voltage. The value of KI = 8.317. This estimation is done by utilizing the information of only one component i.e. Vd and gives us desired estimated phase and frequency information. The complete controller is shown in Fig. 3; this controller block defines the pulses that need to control the switching states of consumer side inverter switches.
Fig. 2. Internal block diagram of machine side control
Fig. 3. Desired controller model
3 Simulation Results The complete system schematic diagram of single phase half bridge inverter tied to grid is shown in Fig. 4. The left side consist of a DC battery. The two switches are controlled by the controlling signals as gate pulse and inverted gate pulse. The components L1 and C3 is the filtration circuit that connects consumer side with the grid model, with R1 as the grid impedance. The parameters used for simulation is summarized in Table 1. As the simulation is initiated, the load is isolated from either sides and controller fetch the desired information from grid side and then process it to generate desired phase and frequency based on our estimation method. The voltage signals in stationary frame of references is shown in Fig. 5. Figure 5 shows that the phase difference between these two voltage is π/2 and this is why it’s called orthogonal voltages. Similarly, the estimated and the one applied by integral control action frequency waveforms are shown in Fig. 6 and Fig. 7 respectively. Whereas the estimated phase waveform is shown in Fig. 8. These estimated phase and frequency are combined to generate switching pulses for S11 and S12 respectively. To test the system characteristics, an inductive load 1000 W is connected between the grid and the consumer side and is separated by switch controlled isolators. The
A Comparative Harmonic Suppression Analysis
369
Fig. 4. General block diagram of grid tied inverter
complete Matlab based model connected with load is shown in Fig. 9. At first the load is disconnected from both sides and then the power is fed by the grid side at T − 0.2 s. The load current start and it power starts to increase. Since the load is inductive, it consumes both active and reactive component of power as shown in Fig. 10 (a) and Fig. 10 (b). At T − 0.65 s, the inverter side power switch is also turned on and the total power in terms of both reactive and active power elements slightly rises as shown in the Fig. 10 (a). As the inverter and grid gets tied to each other the grid and inverter side voltage become completely aligned as shown in Fig. 10 (c). As the grid synchronizes overall system response becomes underdamped and the oscillations die out in less than 300 ms. The overall system power is maintained around 1500 W. Table 1. System parameters Sr. no
Parameter
Value
1.
VDCp
310 V
2.
Vgridp
310 V
3.
fsource
50 Hz
4.
fsw
2 kHz
5.
Capacitor = C
1.23 mF
6.
Inverter side inductor = L1
8.23 mH
7.
Inductive Power Rating
1000 W
370
H. Khalid et al.
The underdamped oscillation in active and reactive power is because the estimated frequency obtained through integral control action response is sluggish. Therefore, it takes time to settle up both the side. Some portion of power is consumed by the reactive elements due to inductive nature while, the remaining power is exchanged as active power. The total harmonic distortion generated by our proposed scheme is around THD = 0.21% shown in Fig. 11. This test validates that the proposed controller design is capable to have better performance as compared to existing controllers. This slight distortion is due to the ripples present in the current and are generated due to switching frequency of the inverter. Since the prime objective of this research was to reduce the THD % therefore, the system is not equipped with inverter side feedback voltage controller that can monitor grid side voltage. However, it is assumed that the inverter side voltage is constantly maintained at the fixed peak voltage level as that of grid side voltage.
Fig. 5. Stationary frame of reference voltage waveforms
Fig. 6. Estimated angular frequency through integral control action
A Comparative Harmonic Suppression Analysis
Fig. 7. Settling time of estimated frequency using Integral control action
Fig. 8. Estimated phase using integral control
Fig. 9. Grid tied inverter complete layout
371
372
H. Khalid et al.
Fig. 10. (a) active power system response, (b) reactive power system response and (c) grid tied inverter voltage response.
A Comparative Harmonic Suppression Analysis
373
Fig. 11. THD for half bridge grid tied inverter
4 Conclusion In this paper we have successfully implemented our controller that not only helps to stabilize the grid along with inverter. But also helps to manage the power flow and drastically reduces the THD less than 1%. The designed controller uses grid signal to extract its information related to phase and frequency and tunes the inverter accordingly. The analysis is controlled using isolators that connects and cut off the load from either one side or both sides. Since the response of our proposed controller is a bit sluggish therefore, the output power takes a little time to stabilize. But this controller has one drawback, it causes overshoot in grid side voltage upon synchronization.
References 1. Narayanan, A., Alex, S.S.: Single phase reconfigurable grid tie inverter using DQ controller. In: 2018 International Conference on Control, Power, Communication and Computing Technologies (ICCPCCT), pp. 260–265. IEEE (2018) 2. Mohammad Noor, S.Z., Omar, A.M., Mahzan, N.N., Ibrahim, I.R.: A review of single-phase single stage inverter topologies for photovoltaic system. In: 2013 IEEE 4th Control and System Graduate Research Colloquium, pp. 69–74. IEEE (2013) 3. Dube, A., Rizwan, M., Jamil, M.: Analysis of single phase grid connected PV system to identify efficient system configuration. In: 2016 Second International Innovative Applications of Computational Intelligence on Power, Energy and Controls with their Impact on Humanity (CIPECH), pp. 173–177. IEEE (2016) 4. Wang, J., Yan, J.D., Jiang, L., Zou, J.: Delay-dependent stability of single-loop controlled grid-connected inverters with LCL filters. IEEE Trans. Power Electron. 31, 743–757 (2016). https://doi.org/10.1109/TPEL.2015.2401612 5. Fantino, R.A., Busada, C.A., Solsona, J.A.: Optimum PR control applied to LCL filters with low resonance frequency. IEEE Trans. Power Electron. 33, 793–801 (2018). https://doi.org/ 10.1109/TPEL.2017.2667409
374
H. Khalid et al.
6. Yao, W., Yang, Y., Zhang, X., et al.: Design and analysis of robust active damping for LCL filters using digital notch filters. IEEE Trans. Power Electron. 32, 2360–2375 (2017). https:// doi.org/10.1109/TPEL.2016.2565598 7. Zhang, X., Tan, L., Xian, J., et al.: Direct grid-side current model predictive control for gridconnected inverter with LCL filter. IET Power Electron. 11, 2450–2460 (2018). https://doi. org/10.1049/iet-pel.2018.5338 8. Xu, J., Xie, S.: LCL-resonance damping strategies for grid-connected inverters with LCL filters: a comprehensive review. J. Mod. Power Syst. Clean Energy 6(2), 292–305 (2017). https://doi.org/10.1007/s40565-017-0319-7 9. Liu, Z., Wu, H., Liu, Y., et al.: Modelling of the modified-LLCL-filter-based single-phase grid-tied Aalborg inverter. IET Power Electron. 10, 151–155 (2017). https://doi.org/10.1049/ iet-pel.2016.0044 10. Guo, X., Yang, Y., Zhang, X.: Advanced control of grid-connected current source converter under unbalanced grid voltage conditions. IEEE Trans. Ind. Electron. 65, 9225–9233 (2018). https://doi.org/10.1109/TIE.2018.2835367 11. Sun, J.: Impedance-based stability criterion for grid-connected inverters. IEEE Trans. Power Electron. 26, 3075–3078 (2011). https://doi.org/10.1109/TPEL.2011.2136439 12. Ebrahimzadeh, E., Blaabjerg, F., Wang, X., Bak, C.L.: Harmonic stability and resonance analysis in large PMSG-based wind power plants. IEEE Trans. Sustain. Energy 9, 12–23 (2018). https://doi.org/10.1109/TSTE.2017.2712098 13. Yang, D., Ruan, X., Wu, H.: Impedance shaping of the grid-connected inverter with LCL filter to improve its adaptability to the weak grid condition. IEEE Trans. Power Electron. 29, 5795–5805 (2014). https://doi.org/10.1109/TPEL.2014.2300235 14. Merabet, A., Labib, L., Ghias, A.M.Y.M., et al.: Robust feedback linearizing control with sliding mode compensation for a grid-connected photovoltaic inverter system under unbalanced grid voltages. IEEE J. Photovoltaics 7, 828–838 (2017). https://doi.org/10.1109/JPH OTOV.2017.2667724 15. Elkayam, M., Kuperman, A.: Optimized design of multiresonant AC current regulators for single-phase grid-connected photovoltaic inverters. IEEE J. Photovoltaics 9, 1815–1818 (2019). https://doi.org/10.1109/JPHOTOV.2019.2937386 16. Xu, J., Qian, Q., Zhang, B., Xie, S.: Harmonics and stability analysis of single-phase gridconnected inverters in distributed power generation systems considering phase-locked loop impact. IEEE Trans. Sustain. Energy 10, 1470–1480 (2019). https://doi.org/10.1109/TSTE. 2019.2893679 17. Santra, S.B., Acharya, A., Choudhury, T.R., Nayak, B., Panigrahi, C.K.: A modified carrierbased PWM technique for minimization of leakage current in transformer less single-phase grid-tied PV system. Electr. Eng. 103(1), 447–461 (2020). https://doi.org/10.1007/s00202020-01092-6 18. Jiang, S., Liu, Y., Liang, W., et al.: Active EMI filter design with a modified LCL-LC filter for single-phase grid-connected inverter in vehicle-to-grid application. IEEE Trans. Veh. Technol. 68, 10639–10650 (2019). https://doi.org/10.1109/TVT.2019.2944220 19. Guichi, A., Mekhilef, S., Berkouk, E.M., Talha, A.: Optimal control of grid-connected microgrid PV-based source under partially shaded conditions. Energy 230, 120649 (2021). https:// doi.org/10.1016/j.energy.2021.120649 20. Siddique, M.D., Iqbal, A., Sarwar, A., Mekhilef, S.: Analysis and implementation of a new asymmetric double H-bridge multilevel inverter. Int. J. Circuit Theory Appl. (2021). https:// doi.org/10.1002/cta.3091 21. Algorithms, L.L.: A comparative study of single phase grid connected phase looked loop algorithms. Jordan J. Mech. Ind. Eng. 11, 185–194 (2017)
A Comparative Harmonic Suppression Analysis
375
22. Ikken, N., Bouknadel, A., Haddou, A., et al.: PLL synchronization method based on secondorder generalized integrator for single phase grid connected inverters systems during grid abnormalities. In: 2019 International Conference on Wireless Technologies, Embedded and Intelligent Systems (WITS), pp 1–5. IEEE (2019) 23. Cheepati, K.R., Maddala, N.R., Munagala, S.K.: A novel reference current extraction technique with multi-functional capability for shunt active filter. J. Electr. Eng. Technol. 15(2), 657–672 (2020). https://doi.org/10.1007/s42835-020-00371-3
Use of Openflow to Manage Network Devices Yogesh Kumar Chekoory(B) and Avinash Utam Mungur ICT Department, FoICDT University of Mauritius, Reduit, Mauritius [email protected], [email protected]
Abstract. With a growing number of devices communicating over networks, it is difficult to manage traditional network architectures. Due to limitations like mobility of this long-established network, to ensure that the devices in the network are operating smoothly, network administrators have to travel to the location of each device in the network if ever there are cases of failures or updates. Software Defined Networking (SDN) is proving to be an alternative network model to help overcome current limitation of network infrastructures by decoupling the control plane from the data forwarding plane. The purpose of this work is to build an SDN topology which uses OpenFlow as a protocol to communicate between the control plane and data plane in an emulator. The whole topology is deployed in GNS3 emulator using OpenDaylight as the SDN Controller, Open vSwitches and OpenDaylight OpenFlow Manager to provide a Graphical User Interface to manage the SDN controller. Lastly, an evaluation was carried out to test the performance of this SDN topology in simulation. This new technology brings along benefits such as remote manageability, network automation and a central platform. Keywords: SDN · OpenFlow protocol · GNS3 emulator · OpenDaylight controller · OpenDaylight OpenFlow manager · Remote · Central platform
1 Introduction With an outgrowing number of devices connecting to networks, the use of traditional Internet Protocol to manage networking devices over a network has become difficult. In addition to configuration errors that arise in the networking area with the use of the standard Internet protocol, network management problems like poor network performance and configuration management are issues that pop out when it comes to managing those devices in the network [1]. Another major problem network administrators face with the use of standard Internet Protocol networks is that the latter has to go to every device in the network and issue vendor specific commands to route information, set policies and configure many network parameters required for the networking device to function and operate smoothly. For instance, if a network administrator has to perform update operations in a network consisting of multiple routers and switches installed in different regions, the latter has to go to each device’s location and update it manually. In traditional networks, the software aspect of the device is integrated inside the hardware. The bundling of the software and hardware in the same device is called the © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 376–386, 2022. https://doi.org/10.1007/978-981-19-1677-9_34
Use of Openflow to Manage Network Devices
377
integrated approach. This approach is costly and requires an excessive amount of time to solve if any failure occurs within the network. If a switch failure occurs in the network, manual testing and configurations of the switch must be done on site which is time consuming. Considering a scenario as an example where a company operates in 6 branches. If a switch in a branch is down, only five branches would then be operational. By the time the networking professionals arrive on the spot to solve this breakdown, the repercussions are that the company is losing the money it could have gained in that particular lapse of time. In this capitalist world where time is considered as money, the Recovery time objective (RTO) is a substantial aspect during these situations. The principal aim of this work is to build a SDN system which uses the Openflow protocol to manage the network devices in an emulator. Network administrators are provided with a central platform for configuring all devices in the network remotely thus making the RTO relatively small. The paper is structured as follows. In Sect. 2, an overview of related works on SDN and network simulators is provided. Section 3 describes the SDN topology Implementation in the selected Emulator. The focus is finally narrowed down to the results and the evaluation of the system in Sect. 4. Section 5 concludes the paper.
2 Background and Related Works Research works carried by multiple researchers in accordance with Software Defined Network systems (SDN) and OpenFlow protocol are reviewed. Shamim et al. used different OpenFlow controllers to test their performance in an SDN environment deployed virtually in an emulator. The emulator used to implement these different controllers is Mininet. Ping tests between host devices connected to multiple OpenFlow switches are used to test the performance of controllers like OpenFlow Ryu, POX and OpenFlow Pyretic across a tree network. A comparison of round trip time for the different controllers is carried out which concludes that the pyretic controller shows a better performance when compared to the POX and Ryu controller [2]. Kim and Ko designed a cost effective solution to emulate a Software Defined Networking environment [3]. Network devices such as Open vSwitches (OVS) and SDN Controller are built on standard raspberry-pi and Linux kernel based operating systems called Raspbian are used to deploy the proposed SDN test-bed architecture. Floodlight is used as the SDN Controller. The testbed provides a GRE connection between two OVS on PIs and hosts attached to these switches. The system has been evaluated by throughput comparison with net-FPGA which resulted in the test-bed showing similar performance as 1Gbps net-FPGA hardware and highlights the Mininet as showing less accurate results because of virtualized network infrastructure. Zulu et al. emulated a SDN topology in the Mininet emulator [4]. OpenFlow Switch and the OpenDaylight SDN Controller have been deployed as a Cloud service on AWS platform. Hence, the flows sent by the OpenDaylight controller towards the OpenFlow switches are controlled via the Internet. A local company network has equally been emulated in a client-server setup in eve-ng. Iperf was used to test the network performance and the results conclude the throughput in variation with data segment length against time.
378
Y. K. Chekoory and A. U. Mungur
Li and Meng provide a detailed literature review on the Software defined networking architecture as well as the security challenges for openflow-based SDN and some counter measurements [5]. The security challenges are related to the switch, the controller and the channel. The major targets are the OpenFlow tables since they contain all information related to the network. Hackers can insert malicious flows in the table or may even flood the flow table with large amounts of useless flows hence making the data plane unable to store normal flows. As for controllers, they are major targets for attacks such as flooding and Dos. The countermeasures are Confidentiality, Integrity and Availability. Preventing sensitive information from disclosure to unauthorized parties, preventing the controller’s information from being modified by unauthorized parties and allowing only authorized users to have access to the controller are ways to counter the attacks in OpenFlow-based SDN. Wibowo and Gregory provided a brief review on SDN aspects and use of multidomain SDN to OpenFlow controllers [6]. The interconnection of multiple controllers across several domains would be needed to have the multi-domain SDN. As a result, there would be interconnection of global SDN domains as a whole. In addition, features of multiple open-source as well as commercial SDN controllers are compared. Scalability, placement, reliability and security are the highlighted major challenges encountered with the SDN controllers across multiple domains. Assegie focuses on deploying an SDN architecture in an emulated environment. The controller used is POX which is a platform implemented in python. The OpenFlow switches connected to the controller are Open vSwitches. The emulator used is Mininet which is a powerful Linux based SDN emulation tool. In addition, a comparison of round trip delay is also performed for tree, linear and hybrid topology which resulted in hybrid topology being the most performant compared to the other two topologies [1]. Based on these research papers, additional studies are carried out on the components of the SDN architecture as well as some Network simulators and emulators. The basic principles of Software defined Network (SDN) architecture can be summarized into three main steps: 1. Control Plane and Data Plane separated 2. A Centralized Platform Long established networking devices like routers and switches contain a strong mixture of control and data planes. This partnership of the data and control plane has rendered the management operations such as troubleshooting and configurations very troublesome. The separation of the control plane and data plane was taken into consideration to alleviate this problem. Consistently, vendors that sell network equipment have applied the logic of packet forwarding into the hardware itself hence, separating the control plane software. As a result, SDN controllers now appear in the networking world which will be used to handle the management of all isolated control planes also known as the brain of the network architecture within a single platform [7]. In comparison to normal networks, this centralized SDN controller will therefore handle the responsibility of controlling traffic for network devices via programmability in the control plane. In addition, since there will be a uniform control platform provided
Use of Openflow to Manage Network Devices
379
by the SDN controller, scalability, decision making and wide network visibility will be achieved. The SDN controller stores all the logical parts like routing, switching rules and decision taking whereas the network devices which the controller commands rely on the latter’s decision. OpenFlow protocol is used by the controller to communicate with OpenFlow switches [8]. 2.1 Data Plane The data plane consists of the infrastructure layer itself and is also known as the forwarding plane [9]. 2.2 Control Plane Also known as the brain of the network, this plane contains all the activities needed by the data plane provides control over the network elements within its topology. It controls how data packets are forwarded from one place to another [10]. It also creates routing tables which are used to store Ip addresses of devices in order to forward data. 2.3 OpenFlow Switch The OpenFlow switch depends on routing and switching similar to a standard switch. However, the most important aspect of an OpenFlow switch is that it needs to be handled by an OpenFlow controller which can configure the switch by manipulating the flow table in the OpenFlow switch [11]. The OpenFlow switch uses the OpenFlow Protocol as a medium to communicate with the external controller and the latter uses this OpenFlow channel to manage the OpenFlow Switch. 2.4 OpenFlow Protocol The OpenFlow protocol is a term associated with SDN. In simple terms, Openflow can be described as a language that the SDN controller uses to communicate with OpenFlow switches. In a non-Openflow protocol switch, both route determination and packet forwarding occur in the same device unlike a switch using the OpenFlow protocol which separates the data path from the control path. Hence, the protocol that enables communication between the software which is the control plane and the hardware which is the data plane consisting of switches in the network is called OpenFlow. The messages sent by the controller are transmitted over a TLS/SSL secured TCP connection [11]. 2.5 OpenFlow Controller The OpenFlow controller manages all the control planes of OpenFlow switches in order to provide greater flexibility and simplicity while programming flow tables of OpenFlow switches in large networks [12]. The centralized controller has a complete view of the whole network. Network administrators can perform any configuration of any switch from the controller itself without interacting manually with the switch provided that the switch is connected to the controller.
380
Y. K. Chekoory and A. U. Mungur
3 Implementation The SDN topology is deployed in GNS-3 emulator as shown in Fig. 1. The topology contains two Open vSwitches (Ovs) connected to the OpenDaylight Controller installed in an Ubuntu Docker container. The OpenDaylight Openflow manager is also installed in a Docker container. Two PCs are used as end devices which are connected to the Ovs. A NAT cloud is used to provide internet access within the topology. In addition, a web browser is also added in the set up in order to access the web interface of the SDN Manager and ODL Controller.
Fig. 1. Deployed SDN topology in GNS-3
Configurations to install OpenDaylight Controller and OpenFlow Manager in the Ubuntu Docker containers [13]. Connecting the OFM Container using SSH with the ODL Container: – – – –
ssh root@ Enter the password set up earlier cd OpenDaylight-Openflow-App cd OpenDaylight-Openflow-App
To connect the Open vSwitches to the ODL Controller, key in the following in the console of the Open vSwitch: -ovs-vsctl set-controller br0 tcp:6633. To display the web interface of both ODL Controller and Openflow manager, enter in two different tabs in the web browser: http://:8181/index.html (ODL Controller) http://:9000 (OFM Manager) Figure 2 shows the basic view panel from the web interface of the OpenDaylight Openflow Manager accessed in the web browser found in the topology. It displays all the devices connected to the OpenDaylight Controller in the topology.
Use of Openflow to Manage Network Devices
381
Fig. 2. OpenFlow manager web interface
From the OFM Web Interface, a drop action is performed in one OVS. This action will drop all traffics in port eth2 in OVS2 as shown in Fig. 3.
Fig. 3. Drop action performed in OpenFlow manager interface
4 Results and Evaluation In this section, the functionalities of the system are tested and the whole system is evaluated. First and foremost, to check if OpenFlow is used as protocol to allow communication between the controller and the Open vSwitches, the data sent between these two are captured in Wireshark which is supported by GNS3. Figure 4 shows a snapshot of packets captured in Wireshark which proves that the protocol used is OpenFlow 1.3. Figure 5 shows the console of PC1 packets were dropped for a period of 30 s after the flow is sent to the controller.
382
Y. K. Chekoory and A. U. Mungur
Fig. 4. OpenFlow protocol displayed in Wireshark
Fig. 5. Packets dropped for 30 s during execution of ping command
4.1 Testing the Performance of the System Round trip time is the amount of time for a packet to be sent along a line plus the amount of time it takes for an acknowledgment of the same packet to be received [14]. Hence, this delay consists of the propagation time between the two points of a signal. The smaller the RTT, the better the network performance. RTT for this project is obtained by using the ping command to send packets from PC1 to PC2. Figure 6 shows the RTT when PC1 pings PC2. The average RTT value for this path is 1.372 ms. A performance analysis graph for values obtained during the ping command from PC1 to PC2 is drawn as shown in Fig. 7. It demonstrates that the time taken for a packet to move from one host in the SDN network to another host passing by Open vSwitches and returning is less than 2 ms which lies in the ideal range for RTT. However, using ping to test and judge the performance is a bit risky. The amount of data successfully transferred from one place to another within a time period is known as throughput [15]. Hence, this value is measured in bits/s. Figure 8 displays the result of a speed test in Iperf using transmission control protocol (TCP). PC1 will act as the server whereas PC2 will be the client. As displayed in Fig. 9, the
Use of Openflow to Manage Network Devices
383
Fig. 6. Round trip time PC1 PC2
Fig. 7. Performance analysis graph
throughput of the network, data transferring from PC1 to PC2 within a period of 10 s is 22.0 Mbits/s. The number of data transferred within this time period is 26.4 Mbytes/s. However, it is difficult to judge the whole architecture between these two hosts. In addition, since the topology is deployed in an emulator, the functionalities of the devices used may affect the throughput. Therefore, it will be biased to judge the performance of the system based on these results.
Fig. 8. Speed Test-Iperf
384
Y. K. Chekoory and A. U. Mungur
Values of processor utilization displayed in the task manager of the host pc for multiple applications used in the deployment of the SDN OpenFlow architecture were jotted down at different time intervals. Figure 9 shows the total percentage of CPU utilization of the applications used during the implementation and testing cycle. For this topology consisting of 9 nodes, an average of 38% of CPU utilization is good. However, if more nodes like Open vSwitches and Docker containers is added in the system, the average CPU usage would increase thus making the PC slower.
Fig. 9. Processor utilization
5 Conclusion The main objective of this research was to alleviate the problems of network administrators. As SDN is a solution to this problem, deep research on SDN and its architecture was made. The OpenFlow protocol then came into practice which could be used to manage network devices away from the hardware. An emulator was used to provide networking devices needed for this project. OpenDaylight and OpenFlow Manager (OFM) were installed in Docker containers in order to provide an SDN controller and a centralized interface to manage the Open vSwitches in the SDN topology. Flows were added in the OFM which eventually directed the Open vSwitches to take actions based on the configurations of the flows. The functionalities of the system were then tested and at the same time, the performance of the system was also evaluated. Taking all these aspects into consideration, it can be concluded that OpenFlow can be used to manage network devices. In addition, it does alleviate the workload of Network Administrators in terms of configurations which can be performed remotely from a centralized platform. As a result, this deployed SDN network topology complies with the aims and objectives of this project and SDN can be a breakthrough in the networking field in the upcoming years. Future Work Since the ODL Controller is the main device in this topology, it must be kept secure. One way to ensure this is to insert a firewall in the topology in order to restrict access to this device. Figure 10 shows an updated design for the SDN topology consisting of a FortiGate firewall having three zones. The ODL Controller, OFM and browser are inserted in the demilitarized zone (DMZ) to ensure only authorized access is granted.
Use of Openflow to Manage Network Devices
385
6 Future Work Since the ODL Controller is the main device in this topology, it must be kept secure. One way to ensure this is to insert a firewall in the topology in order to restrict access to this device. Figure 10 shows an updated design for the SDN topology consisting of a FortiGate firewall having three zones. The ODL Controller, OFM and browser are inserted in the demilitarized zone (DMZ) to ensure only authorized access is granted.
Fig. 10. Securing the ODL controller using a firewall
References 1. Assegie, T.: Software defined Network emulation with OpenFlow Protocol (2020). https:// www.researchgate.net/publication/339708791. Accessed Nov 2020 2. Shamim, S., et al.: Performance Analysis of Different Openflow based Controller over Software Defined Networking (2018). https://globaljournals.org/GJCST_Volume18/3-Perfor mance-Analysis-ofDifferent.pdf 3. Kim, H., Kim, J., Ko, Y.-B.: Developing a cost-effective OpenFlow testbed for small-scale Software Defined Networking (2014). https://www.researchgate.net/publication/269310978. Accessed Dec 2020 4. Zulu, L.L., Ogudo, K.A., Umenne, P.O.: Emulating Software Defined Network Using Mininet and OpenDaylight Controller Hosted on Amazon Web Services Cloud Platform to Demonstrate a Realistic Programmable Network (2019). https://ieeexplore.ieee.org/document/860 1254. Accessed Jan 2021 5. Li, W., Meng, W.: A survey on OpenFlow-based Software Defined Networks: Security challenges and countermeasures (2016). https://www.researchgate.net/publication/301599364. Accessed Dec 2020 6. Wibowo, F.X.A., Gregory, M.A.: Multi-Domain Software Defined Networking: Research Status and Challenges (2017). https://www.researchgate.net/publication/314290818. Accessed Jan 2021
386
Y. K. Chekoory and A. U. Mungur
7. SDxCentral. Understanding the SDN Architecture and SDNControl Plane (2015). https:// www.sdxcentral.com/networking/sdn/definitions/inside-sdn-architecture. Accessed Dec 2020 8. Göransson, P., Black, C., Culver, T.: Software Defined Networks a Comprehensive Approach, 2nd edn. Elsevier (2016) 9. GeeksforGeeks. Software defined Networking (2019). https://www.geeksforgeeks.org/sof tware-defined-networking/. Accessed Dec 2020 10. Raza, M., Hertvik, J.: What is Software Defined Networking? SDN Explained (2020). https:// www.bmc.com/blogs/software-defined-networking. Accessed Jan 2021 11. OpenFlow Switch Specification (2014). https://opennetworking.org/wpcontent/uploads/ 2013/04/openflowspecv1.3.1.pdf 12. CiscoDevNet. OpenDaylight-Openflow-App (2014). https://developer.cisco.com/codeexcha nge/github/repo/CiscoDevNet/OpenDaylight-Openflow-App 13. Bombal, D.: GNS3Talks/ODL OFM install with GNS3 (2017). https://github.com/dav idbombal/GNS3Talks/blob/master/ODL%20OFM%20%20install%20with%20GNS3.txt. Accessed Mar 2021 14. Banerjee, A., et al.: Construction of effective wireless sensor network for smart communication using modified ant colony optimization technique. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 269–278. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_22 15. Wallen, J.: How to perform network throughput tests with Iperf, October 2019. https://www. techrepublic.com/article/how-to-performnetwork-throughput-tests-with-iperf. Accessed Apr 2021
Intelligent Routing Protocol for Energy Efficient Wireless Sensor Networks Nandoori Srikanth1(B) , Battula Ashok2 , Battula Chandini3 , Mohan Kumar Chandole4 , and Naga Jyothi1 1 Department of Electronics and Communication Engineering, St. Peters Engineering College,
Hyderabad, Telangana, India [email protected] 2 Department of Electronics and Communication Engineering, Balaji Institute of Technology and Science, Laknepally, India 3 Department of Computer Science Engineering, Balaji Institute of Technology and Science, Laknepally, India 4 Department of Computer Science Engineering, Koneru Lakshmiah Educational Foundation, Vaddeswaram, India
Abstract. Wireless Sensor Networks is the Key research area from past few decades, lifetime, Energy consumption, Energy Efficiency, Throughput are the parameters to decide the network quality. Deployment is also one of the key aspects to decide lifetime time of a WSN. Deployment in Uniform terrain structures is not a limitation when it is compared with Non Uniform terrain structures. In this paper An Intelligent Routing protocol is proposed to maintain sensor nodes into sleep state until their turn comes. This Routing algorithm makes sensor nodes into sleep states in an intelligent manner without affecting the throughput and QOS. Along with this technique, energy hole removing mechanism is also incorporated to remove Energy Hole Drainage Problem. Keywords: Lifetime · Energy Efficiency · QOS · Terrain structure · Throughput · Deployment
1 Introduction The collection of group of sensors equipped with sensor nodes to perform a particular task is known as wireless sensor network. These wireless sensor networks can also accessible in the harsh environment like where humans cannot do surveillance and operated under remote monitoring [1]. These sensors network has four major units such as power unit, Sensing unit, Transceiver and a Processor unit. These nodes are energy efficient, and they are distributed in an environment, where those nodes cannot be replaced and recharging of batteriesmultiple times. Therefore, due to difficulty of recharging sensors batteries, energy harvesting techniques are implemented [2] (Power harvesting techniques are nothing but energy scavenging from solar energy used to recharge batteries). Here network failure is occurred due to energy drainage problem and this is the major © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 387–396, 2022. https://doi.org/10.1007/978-981-19-1677-9_35
388
N. Srikanth et al.
issue in the wireless sensor networks. So many authors proposed various routing techniques to make network more energy efficient, but still WSN is facing major issues due to lack of improvement in Network Lifetime. Energy utilization is mainly depending on the types of wireless sensor network applications, they are classified into four types based on the utilization of energy such as query-driven applications, continuous monitoring applications, hybrid applications and event-based applications [3]. In some applications like query-driven and event-based applications, the sensor nodes work only when query/Event is driven in the network. Coming to the continuous monitoring applications, nodes has the high amount of energy consumption due to continuous sensing and data transmission and receptions. To improve the performance of the network and also to make the energy efficient network, some popular proposed algorithms of the hierarchical routings are applied [5, 6]. Here routing algorithms are classified into four types based on their performances in wireless sensing networks. They are network flow and QOS aware algorithm, data centric algorithm, location-based algorithm and hierarchical algorithms. In the hierarchical algorithm, it performs the data aggregation and aggregates all message packets in transmitted message and arranged them in an order. This can be implemented by multi-hop communication; Clusters are formed based on the distance between the nodes, and also based on distance from sink. To reduce the over-headed traffic, the higher-level communication is done between cluster-head and base station. The hierarchical routing algorithms are proposed to overcome all the existing algorithms like DEEC, LEACH, TEEN, SEP Etc. This routing algorithm increases the network lifetime and stability of the network. Section 2 of paper gives a brief survey on existing protocols and it also gives the drawbacks of protocols in improving the lifetime of WSN. Section 3 shows the parameters of the proposed system and methodology of proposed routing algorithm. Section 4 deals with consequences and simulation results and comparison of results with existed protocols and algorithms. Section 5 gives a conclusion report on proposed protocol, and also gives the key improvements observed when compared with previous protocols. This section also discuss the improvements can be possible in future.
2 Literature Survey In Low Energy Adaptive Clustering Hierarchy (LEACH) protocol [6], cluster head is responsible for the data transmission. Each node can be elected as a cluster head in various rounds based upon the probability. Cluster heads are formed in randomized manner, there are two different phases in leach protocol, in which first phase is cluster formation and another is data transmission. To avoid data duplication, we use this LEACH protocol. Parameters like information security, scalability, network lifetime are improved. In each round of LEACH excess overhead is occurred due to cluster head, networktopology is changed due to selection of cluster head in every round and consumption of energy is increased.
Intelligent Routing Protocol for Energy Efficient WSN
389
In Modified Low Energy Adaptive Clustering Hierarchy (MOD-LEACH) algorithm [7], two methods are proposed by author, in which one is efficient cluster head replacement method and other one is dual transmitting power method. If the residual energy of a node is greater than threshold energy, then it will always be elected as cluster head. With this approach network topology remains unchanged so that consumption of energy is reduced. If the energy of a node is less than threshold energy then this method will act same as LEACH. Stable Election Protocol (SEP) is the energy efficient algorithm [10]. If the energy of a node is greater than threshold energy, then it will form a cluster head. Depending upon the energy levels of nodes different probabilities are assigned. In the stable election protocol it will not uses the extra power of the sensor nodes. DEEC algorithm [11] is best suitable for heterogeneous WSN where each node having energy of different levels. Based on the energy level of nodes cluster head formation is done. In this protocol the node which has high energy will probably selected as Cluster head.
3 Proposed Intelligent Routing Protocol Routing Protocol: Energy Consumption is the major issue in WSN, which reduces the lifetime of network. To reduce this energy consumption. In proposed intelligent routing protocol, each node should spend minimum amount of energy during processing of a data packet. In order to save the energy of node, we need to reduce the energy consumption of a network that leads to increase life time of the network. Due to aggregation, the overhead of CH, BS will be reduced drastically. To minimize traffic, clustering technique is used. Instead of homogeneous network, here heterogeneous network is used in proposed protocol. The Cluster Head election can be done by following equation p N Ci(t) = 1 Ti(t) = N −p(rmod ( P ) 0 Ci(t) = 0
(1)
Ti (t) = Node probability i = Cluster heads according to round P = required number of cluster head N = Number of rounds Ci(t) = This function determining that whether this node is elected as cluster head before.
390
N. Srikanth et al.
The major flow of intelligent routing protocol (IRP) for every Participating cluster node and cluster head should have a connection. Due to this connection, we can increase or decrease the nodes in a network the energy hole drainage problem can be removed by incorporating the energy hole removing mechanism in a network multiple paths are created in a cluster by using this routing algorithm. Rather than creating multiple paths select the best one as out. The flowchart of IRP is given in Fig. 1. In this intelligent routing protocol have different phases like set-up phase, phase of cluster formation, active and inactive phase. Flow Chart of IRP
Fig. 1. Flow chart of IRP protocol
Intelligent Routing Protocol for Energy Efficient WSN
391
A. Set Up Phase: In general, nodes are deployed randomly. The deployed nodes can communicate among themselves and also with its neighbors by prolonging x and y coordinates and communicates with sensor nodes. The position of the sink should be at center (sink is nothing but base station), due to this placing of sink at center of the sensing area, all cluster head nodes have possibility to communicate without any data loss. In this network, all nodes energies are non-uniform due to Heterogeneous set up. B. Phase of Cluster Formation: After deployment the base station collects all the positions of the nodes. The sensor nodes which have the highest residual energy they will advertise a message to its neighbour that it will be elected as cluster head. The neighbor nodes will give a reply message to the cluster head whose works are ready to be work has a cluster member. The neighbor discovery methods are k-of-n approach, Beacon messaging and ping, these are commonly used methods in discovery of neighbors. After cluster formation and discovery neighbor node it will decide that which node have the cluster head for the present round. Cluster head is selected from available nodes by using the cluster head selection algorithm. C. Active/Inactive Phase: After creating the cluster we can calculate the energy of each node by using Eq. 2. Each node (E0) should be greater than the threshold energy(Eth)then the node will goes to active state otherwise it goes to sleep state. Based on the threshold value (Eth) the node set the sleeping state. To calculate the threshold value, we can use the below equation (2) Eth = ((ETX + EDA) ∗ D) + Eamp ∗ D ∗ d4 Where D = data packet length D = distance between node and base station If the number of nodes are greater than 10, switching of nodes will be take place from sleep to active.
4 Simulation Results In the wireless sensor network, there are hundred sensor nodes deployed in the sensing area, and those nodes energies are non-uniform. In this paper, the sensor nodes are deployed as 400 * 400 m2 square regions. The base station is located at the centre of the sensor nodes; the energy of each sensor node is maximum up to 1j. The below table describes about the input parameters of proposed protocol, this protocol is simulated using MATLAB software and the results of Proposed IRP Protocol compared with relevant and best protocols like TEEN, MODE-LEACH, DEEC, E-MODE-LEACH, SEP. The main comparison is done with the parameters like lifetime of the network, number of the packets that received and the Dead nodes per round.
392
N. Srikanth et al.
A. Election of cluster heads: In the wireless sensor networks cluster head shows the major impact on the network analysis. When consumption of energy increases the number of cluster heads increases. The lifetime of the network is decreases because of the cluster head decrementation. For the balanced energy consumptions, the stability of the cluster heads are required. The below figure shows the energy consumption in each round. In this proposed system initially cluster heads are greater according to the sensor nodes. Then after this it remains constants same as others. In this proposed system the cluster head election process is done. This election process was done before in the LEACH protocol and MODE-LEACH protocol. B. Network life time: The lifetime of the network is nothing but the ratio of last dead node to first dead node. In the below figure its showing the data about dead nodes and alive nodes in the system of network. This data is considered as per the rounds. According to this first and last nodes are died faster than other mode protocols like LEACH and MODE-LEACH. Coming to intelligent routing protocol only some nodes are died due to the energy hole removing mechanism. C. Base station receiving packets: In this intelligent routing protocol more number of packets is received by the sensor nodes and sent to base station. This is the proposed approach as shown in the below figures (Figs. 2, 3, 4, 5 and Table 1).
Table 1. Simulation parameters Parameter
Definition
Value
X
x-axes Distance at
400 m
Y
y-axes Distance at
400 m
N
Total nodes
100 Nodes
E0
Total energy of network
1j
P
Probability of cluster head
0.5
Erx
Receiving energy dissipation
0.0013/pj/bit/m4
Efs
Free space model energy dissipation
10/pj/bit/m2
Eamp
Power amplifier energy dissipation
0.0013/pj/bit/m4
Eele
Electronics energy dissipation
50nj/bit
Etx
Transmission energy dissipation
50/nj/nit
Eda
Aggregation energy dissipation
5/nj/bit
D0
Reference distance
Sqrt(Efs/Emp)
Sn
Number of sleep nodes
10 Nodes
M
Percentage of advanced nodes
0.1
Rmax
Number of rounds
4500
Intelligent Routing Protocol for Energy Efficient WSN
Fig. 2. Cluster heads vs rounds
Fig. 3. Number of nodes alive vs rounds
393
394
N. Srikanth et al.
Fig. 4. Dead nodes vs rounds
Fig. 5. Packet transmission to BS vs rounds
5 Conclusion An intelligent routing protocol is designed and implemented for Energy Efficient wireless sensor network, by using the heterogenous energy property we can increases network life time. In this approach, life time of WSN increased by modifying cluster head electron threshold with the help of a different cluster head appointment mechanism and also by increased the sleep node threshold level. In WSN stability period is increased by removing energy hole drainage problem. Energy efficient sleep Awake aware technique used in this IRP protocol gives best results compared to all other existing protocols. Compared to existing algorithm such as LEACH, MOD-LEACH, SEP, DEEC, this IRP
Intelligent Routing Protocol for Energy Efficient WSN
395
routing protocol gave best results as shown in simulation results and also gave an idea to increase sleep threshold level to increase lifetime of network.
References 1. Hu, Y., et al.: An energy-efficient adaptive overlapping clustering method for dynamic continuous monitoring in WSNs. IEEE Sens. J. 17(3), 834–847 (2017) 2. Shurman, M., Awad, N., Al-Mistarihi, M.F., Darabkh, K.A.: LEACH enhancements for wireless sensor networks based on energy model. In: IEEE 11th International Multi-Conference on Systems, Signals & Devices (SSD14), Barcelona, pp. 1–4 (2014) 3. Varshney, S., Kumar, C., Swaroop, A.: A comparative study of hierarchical routing algorithms in wireless sensor networks. In: 2nd International Conference on Computing for Sustainable Global Development (INDIACom), New Delhi, pp. 1018–1023 (2015) 4. Shen, J., Wang, A., Wang, C., Ren, Y., Wang, J.: Performance comparison of typical and improved LEACH algorithms in wireless sensor network. In: 2015 First International Conference on Computational Intelligence Theory, Systems and Applications (CCITSA), Yilan, pp. 87–192 (2015) 5. Patel, K., Sanjay, T., Pradeep, J.: Energy efficient hierarchical routing algorithm in wireless sensor network. Int. J. Adv. Res. Sci. Eng. Technol. 1(3), 103–109 (2014) 6. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: Energy-efficient communication protocol for wireless microsensor networks. In: Proceedings of the 33rd Hawaii International Conference on System Sciences (HICSS-33), vol. 8, pp. 8020–8030 (2000) 7. Mahmood, D., Javaid, N., Mahmood, S., Qureshi, S., Memon, A.M., Zaman, T.: MODLEACH: a variant of LEACH for WSNs. In: Eighth International Conference on Broadband and Wireless Computing, Communication and Applications, Compiegne, pp. 158–163 (2013) 8. Manjeshwar, A., Agrawal, D.P.: TEEN: a routing algorithm for enhanced efficiency in wireless sensor networks. In: Proceedings 15th International Parallel and Distributed Processing Symposium, IPDPS 2001, San Francisco, CA, USA, pp. 2009–2015 (2001) 9. Banerjee, A., et al.: Construction of effective wireless sensor network for smart communication using modified ant colony optimization technique. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 269–278. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_22 10. Smaragdakis, G., Ibrahim, M., Azer, B.: SEP: a stable election algorithm for clustered heterogeneous wireless sensor networks. Boston University Computer Science Department, Boston, Report 022 (2004) 11. Paul, A., Sinha, S., Shaw, R.N., Ghosh, A.: A neuro-fuzzy based IDS for internet-integrated WSN. In: Bansal, J.C., Paprzycki, M., Bianchini, M., Das, S. (eds.) Computationally Intelligent Systems and their Applications. SCI, vol. 950, pp. 71–86. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0407-2_6 12. Pandya, N.K., Kathiriya, H.J., Kathiriya, N.H., Pandya, A.D.: Design and simulation of enhanced MODLEACH for wireless sensor network. In: IEEE International Conference on Computing, Communication & Automation, Noida, pp. 336–341 (2015) 13. Younis, O., Sonia, F.: HEED: a hybrid, energy-efficient, distributed clustering approach for ad hoc sensor networks. IEEE Trans. Mob. Comput. 3(4), 366–379 (2004) 14. Pandya, N.K., Kathiriya, H.J., Kathiriya, N.H., Pandya, A.D.: Design and simulation of advance MODLEACH for wireless sensor network. In: International Conference on Computer, Communication and Control (IC4), Indore, pp. 1–6 (2015) 15. Srikanth, N., Siva Ganga Prasad, M.: Green comp based energy efficient data aggregation algorithm with malicious node identification (GEED-M) for lifetime improvement in WSN. COMPUSOFT Int. J. Adv. Comput. Technol. 8(4), 3117–3125 (2019)
396
N. Srikanth et al.
16. Srikanth, N., Prasad, M.S.: Energy efficient trust node based routing protocol (EETRP) to maximize the lifetime of wireless sensor networks in plateaus. Int. J. Online Eng. 15(6), 113–130 (2019) 17. Srikanth, N., Prasad, M.S.: Family based efficient Routing Protocol for lifetime improvement in WSN. J. Adv. Res. Dyn. Control Syst. 11(06), 113–130 (2019) 18. Srikanth, N., Siva Ganga Prasad, M.: Efficient energy clustering protocol using genetic algorithm in wireless sensor networks. J. Eng. Sci. Technol. Rev. 11(6), 85–93 (2018) 19. Srikanth, N., Siva Ganga Prasad, M.: Efficient clustering protocol using fuzzy k-means and midpoint algorithm for lifetime improvement in WSNs. Int. J. Intell. Eng. Syst. 11(4), 61–71 (2018) 20. Rajawat, A.S., Barhanpurkar, K., Shaw, R.N., Ghosh, A.: Risk detection in wireless body sensor networks for health monitoring using hybrid deep learning. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 683–696. Springer, Singapore (2021). https://doi.org/10.1007/978-98116-0749-3_54 21. Srikanth, N., Siva Ganga Prasad, M.: A compressive family based efficient trust routing protocol (C-FETRP) for maximizing the lifetime of WSN. In: Jain, L.C., Tsihrintzis, G.A., Balas, V.E., Sharma, D.K. (eds.) Data Communication and Networks. AISC, vol. 1049, pp. 69– 80. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-0132-6_6. ISBN 978981-15-0131-9 22. Srikanth, N., Siva Ganga Prasad, M.: An adaptive genetic co-relation node optimization routing for wireless sensor network. In: Jain, L.C., Tsihrintzis, G.A., Balas, V.E., Sharma, D.K. (eds.) Data Communication and Networks. AISC, vol. 1049, pp. 81–106. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-0132-6_7. ISBN 978-981-15-0131-9 23. Srikanth, N., Siva Ganga Prasad, M.: Energy efficient enhanced tree structured compression model (ET-CM) for data aggregation in wireless sensor networks. Int. J. Eng. Technol. 7(2), 1–4 (2018)
Voltage-Lift-Type Z-Source Inverter Vadthya Jagan(B) and Udutha Prashanth Department of Electrical and Electronics Engineering, Vignana Bharathi Institute of Technology, Hyderabad, Telangana, India [email protected]
Abstract. In this article, voltage-lift type of impedance-source inverter (VL-ZSI) is proposed. In a switched-inductor Z-source inverter (SL-ZSI), the central diode in top switched-inductor (SL) cell is replaced with a voltage-lift capacitor, CVL and the bottom SL cell is replaced with a single-inductor to obtain the high boost factor. Therefore, the sum of active and inactive components is reduced, which diminishes the cost, weight, and space requirement of the system when compared with SLZSI. Moreover, this proposed inverter topology utilizes very small duty ratio to acquire the essential voltage boost at high modulation index which improves the output waveform quality with low total harmonic distortion (THD) for a given input voltage. It also results in less switch stress. The proposed voltage-lift ZSI topology is compared with some existing impedance-network topologies. The proposed VL-ZSI is evaluated in the steady-state condition and their theoretical analysis is validated in MATLAB/Simulink. Keywords: CSI · Extended-boost ZSI · EB-ZSI · EB-qZSI · VSI · VL-ZSI · ZSI · SBI · SL-ZSI
1 Introduction The outdated Voltage Source Inverters (VSIs) and Current Source Inverters (CSIs) are used to decrease and to increase the input voltage respectively to the essential level. The VSI can only decrease the input voltage, whereas the CSI can only increase the input voltage. In order to get both the buck and boost operation of VSI/CSI, the dc-dc boost/buck converters are to be linked in amongst the source and the bridge inverter which increases size, space, complexity, and cost of the structure [1]. Besides, the switch on of both the devices in any phase leg of VSI can destroy the power semiconductor switches. In order to overcome the aforesaid downsides of VSI and CSI, the novel impedancesource inverter (ZSI) was presented in [2]. The impedance network comprises of two capacitors and two inductors which are cascaded in amongst the dc source and bridge inverter. In shoot-through state, the inductors are charged by the capacitors and in the non-shoot-through state, the stored energy in these inductors and input energy boost the dc-link voltage. Since the ZSI utilizes the shoot-through state to boost the input voltage in single-stage without dead band circuit, the reliability of the system increases. But, the traditional ZSI agonizes from discrete input current, does not share common ground with source, and takes huge starting inrush current. Furthermore, the stress across the © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 397–410, 2022. https://doi.org/10.1007/978-981-19-1677-9_36
398
V. Jagan and U. Prashanth
capacitors is high. Therefore, to overcome the downsides of ZSI, the qZSI and improvedZSI were projected in [3] and [4] separately with some modifications in the impedance network. The boost factor of the ZSI [2], qZSI [3] and improved-ZSI [4] is identical and is given as B=
ˆ VPN 1 1 = = VDC (1 − 2(T0 /T )) (1 − 2D0 )
(1)
ˆ , is the peak dc-link voltage, V where, B is the boost factor VPN DC is the input voltage, T 0 is the shoot-through period over a period, T and D0 is the duty ratio.
Fig. 1. Switched-inductor Z-source inverter [10].
The paper presented in [5], inserts the two isolated dc sources in the X-shaped impedance network which draws continuous current from the source and also reduces the voltage/current stresses of active and inactive components with the identical boost factor as that of ZSI. The boost factor can also be increased with the modulation techniques proposed in [6–9] without any changes in the impedance networks with reduced switch stress. The inductors in the conventional ZSI are substituted by switched-inductor (SL) cells to boost the voltage gain. As portrayed in Fig. 1, the amalgamations of L 1 – D1 – L 3 – D3 – D5 acts as top SL cell and the amalgamations of L 2 – D2 – L 4 – D4 – D6 acts as bottom SL cell for SL-ZSI. The downsides of SL-ZSI [10] are same as that of the traditional ZSI [2]. Therefore, to overcome these drawbacks, the SL-qZSI and extended-boost qZSIs were proposed in [11, 12] and [13] correspondingly. Similarly, by varying the turns ratio of tapped-inductor (TL) cells, the TL-ZSI and TL-qZSI were proposed in [14] and [15] respectively to increase the boost factor freely. The coupled inductor/transformer-based topologies are proposed in [16–20] with more or less passive components to get the required boost factor. By adding one extra switch, the boost factor of switched boost inverters (SBIs) planned in [21–24] is more or less same as that of conventional ZSI with a smaller number of passive components. The middle diode in SL-cell of SL-qZSI [11] was replaced with the voltage-lift capacitor C VL which forms a voltage-lift unit to improve the boost factor [25–27]. The enhanced-boost ZSI with
Voltage-Lift-Type Z-Source Inverter
399
two switched-impedance networks proposed in [28] provides high boost factor with more active and inactive elements in the impedance network. But, the drawbacks of this inverter are same as that of ZSI and SL-ZSI. The demerits of EB-ZSI were minimized by EB-qZSIs and was proposed in [29, 30]. More number of components were used in [28–30]. Therefore, novel voltage-lift type impedance-source inverter (VL-ZSI) is proposed in this article using a smaller number of passive and active components with high voltage gain when compared with SL-ZSI and EB-(q)ZSIs. The projected topology provides very high boost ability at low duty ratio for the same input voltage which results in high modulation index at low total harmonic distortion (THD) and low switch stress. Due to this small shoot-through period, the conduction losses are less and hence, it improves the overall system efficiency. Even at D0 = 0, the dc-link voltage of the proposed inverter is twice the input voltage. The operating principle and boost factor derivation of the projected VL-ZSI is discussed in Section-II. Section-III describes the performance comparison of SL-ZSI, extended-boost qZSIs and EB-ZSI with the projected VL-ZSI. Designing of the impedance parameters for the proposed inverter is described in SectionIV. Finally, the discussions on the simulation results are included in Section-V to validate the theoretical analysis.
Fig. 2. Circuit diagram of proposed voltage-lift type-ZSI.
2 Proposed Voltage-Lift ZSI As revealed in Fig. 2, the impedance network of proposed VL-ZSI contains of threediodes, three-inductors, and three-capacitors. The top SL-cell of SL-ZSI is replaced by VL unit and bottom SL-cell is replaced by single-inductor L 2 and the capacitors are placed similarly to that of SL-ZSI [10]. The alteration between SL-ZSI [10] and the projected inverter topology is that only one of the two SL-cells is used as a voltage-lift (VL) unit (i.e., it consists of two-diodes, two-inductors, and one voltage-lift capacitor) and other SL-cell is replaced with a single inductor L 2 . Therefore, when compared with the SL-ZSI [10], EB-ZSI [28] and EB-qZSIs [29, 30], the proposed inverter topology
400
V. Jagan and U. Prashanth
uses a smaller number of passive (i.e., capacitor and inductor) and active (i.e., diodes) components which reduces the space, cost, weight, and complexity of the system. Furthermore, it uses a very small duty ratio to produce the required voltage boost which provides less conduction losses. The middle diode of the top SL-cell is replaced with voltage-lift capacitor C VL to obtain the high boost factor at low duty ratio and high modulation index which provides the improved output waveform with low THD for the similar input and output voltages. At low duty ratio (i.e., D0 ≤ 0.25), the boost factor of the proposed inverter topology is more than that of the SL-ZSI [10], extended-boost qZSIs [13], and EB-ZSI [28] the EB-qZSIs [29, 30] for the same input voltage and duty ratio D0 . Similar to the conventional ZSI, the VL-ZSI has seven shoot-through states in addition to the eight non-shoot-through states which are described next.
Fig. 3. Equivalent circuit of proposed VL-ZSI in (a) shoot-through state and (b) non-shootthrough state.
2.1 Shoot-Through State Corresponding configuration of VL-ZSI in the shoot-through state is revealed in Fig. 3(a). The diodes D1, D3 are turned on and diode Din is off. The input current is zero and peak
Voltage-Lift-Type Z-Source Inverter
401
dc-link voltage is also zero during the period, D0 . The inductors L 1 , L 3 are charged by a parallel capacitor C 1 , whereas capacitor C VL obtains energy from C 1 which significantly magnifies the voltage boost and the inductor L 2 is charged by the parallel capacitor C 2. The inductors store the electromagnetic energy. Applying Kirchhoff’s voltage law (KVL) to Fig. 3(a), the following equations can be obtained as ⎫ VL1 = VL3 = VC1 = VCVL ⎪ ⎪ ⎪ ⎬ VL2 = VC2 (2) ⎪ VDin = VC1 + VC2 ⎪ ⎪ ⎭ VD1 = VD3 = 0
2.2 Non-shoot-Through State The corresponding configuration during non-shoot-through state is revealed in Fig. 3(b). Diodes D1 , D3 are off and input diode Din is turned on. The stored energies in the series inductors L 1 , L 2 , and L 3 and capacitor C VL are moved to main dc-link circuit during the period (1 – D0 ) to boost the voltage gain. The capacitors are charged from the input supply through inductors. Applying KVL to Fig. 3(b), the following equations can be obtained as ⎫ VL1 = VDC + VC1 − VL3 − VC2 ⎪ ⎪ ⎪ ⎬ VL2 = VDC − VC1 (3) VL3 = VDC + VC1 − VL1 − VC2 , and VDin = 0⎪ ⎪ ⎪ ⎭ VD1 = VD3 = VC2 − VDC + VL1(non_shoot) ˆ VPN = VC2 − VDC + VC1
(4)
Since the average voltages across the inductors are always zero, the following equations are obtained VC2 =
(VC1 − VDC )(1 − D0 ) D0
(VL1 = VL3 )non_shoot = (VDC + VC1 − VC2 ) + D0 VC2 =
(5) VC1 (1 − D0 )
1 + D0 VC1 + VDC 1 − D0
(6) (7)
From (5) and (7), the voltage across capacitor C 1 can be obtained as VC1 =
1 − D0 VDC 1 − 3D0
(8)
Similarly, after substituting (8) in (7), the voltage expressions for capacitors C 2 and C 3 can be obtained as VC2 = 2
1 − D0 VDC 1 − 3D0
(9)
402
V. Jagan and U. Prashanth
VC3 = VCVL =
1 − D0 VDC 1 − 3D0
(10)
The peak dc-link signal is derived from Fig. 3(b). Substituting the capacitor voltages V C1 and V C2 in (4), the peak dc-link voltage can be obtained as ˆ VPN =
2 VDC 1 − 3D0
ˆ VPN = BVDC
where, boost factor, B =
(11) (12)
2 1−3D0 .
ˆ ) of the projected inverter can be attained as The peak phase output voltage (Van ˆ =M Van
ˆ VPN 2
(13)
where, M is the modulation index. VDC ⎫ ⎪ ⎬ 2 VDC ⎪ ⎭ =G 2
ˆ Van = MB ˆ Van
(14)
The voltage gain (G = MB) for the simple boost control scheme [2] in terms of the modulation index (assuming ideal case) for the proposed topology can be defined as G = MB =
ˆ 2M Van = 3M − 2 (VDC /2)
(15)
The switch stress V S in terms of the voltage gain G for the proposed VL-ZSI topology can be obtained as follows VS = BVDC =
3G − 2 VDC 2
(16)
As can be seen from Fig. 4 and also from Eq. (12), that the boost factor of the projected network is higher than that of SL-ZSI [10], extended-boost qZSIs [13], EBZSI [28] and EB-qZSIs [29, 30] for the same input voltage V DC and duty ratio D0 with fewer active and inactive elements. The Fig. 5 portrays the assessment of voltage gain G of the projected concept with SL-ZSI, extended-boost qZSI, and EB-ZSI at the same modulation index M and input voltage V DC . It can be detected after Fig. 5 that the projected inverter topology has the higher voltage gain at any modulation index, M
Voltage-Lift-Type Z-Source Inverter
403
Fig. 4. Boost factor assessment of projected VL-ZSI with SL-ZSI [10], extended-boost qZSIs [13], and EB-ZSI [28].
Fig. 5. Voltage gain assessment of proposed VL-ZSI with SL-ZSI [10], extended-boost qZSIs [13], and EB-ZSI [28].
3 Performance Assessment of Proposed VL-ZSI with the Existing Z-Network Topologies Figure 6 depicts the comparison of switch stress, which is defined as the ratio of V S and GV DC [7], for the proposed VL-ZSI with that of other existing inverter topologies. As can be observed from this figure that if voltage gain is below six (i.e., G ≤ 6), the stress across the switch is lower than all the existing inverter topologies [10, 13, 28]. But, after G ≥ 6 the stress across the semiconductor switch is slightly higher than enhanced-boost (q)ZSI [28–30] and less than that of SL-ZSI [10], and extended-boost qZSIs [13] for the same input voltage V DC and duty ratio D0 . Therefore, for a given shoot-through duty ratio and input voltage, at low voltage gain the proposed inverter topology provides less switch stress when compared to the existing topologies which decreases the switch conduction loss and thus, the efficiency of the system can be improved.
404
V. Jagan and U. Prashanth
Fig. 6. Switch stress assessment of proposed VL-ZSI with SL-ZSI [10], extended-boost qZSIs [13], and EB-ZSI [28]. Table 1. Passive, active components and input current nature comparision Topology
L
C
D
Switches
Input current nature
SL-ZSI [10]
4
2
7
6
Discrete
DA-qZSI [13]
3
3
3
6
Continuous
CA-qZSI [13]
3
4
2
6
Continuous
EB-ZSI [28]
4
4
5
6
Discrete
EB-qZSI [28]
4
4
5
6
Continuous
Proposed VL-ZSI
3
3
3
6
Discrete
Where, L – number of inductors, C – number of capacitors, D – number of diodes
Fig. 7. Capacitor stress assessment of projected VL-ZSI with SL-ZSI [10], extended-boost qZSIs [13], and EB-ZSI [28].
Voltage-Lift-Type Z-Source Inverter
405
The capacitor stress, which is defined as V C /GV DC [7], is compared for SL-ZSI, extended-boost qZSIs and EB-ZSI with the projected VL-ZSI and is portrayed in Fig. 7. The stress across the capacitors in the projected network topology is slightly less when compared to the SL-ZSI, extended-boost qZSIs, and EB-ZSI. Consequently, lesser rating capacitors can be used which diminishes the cost and space requirement. Similarly, as can be observed from Fig. 8, that the diode stress on diodes is also less in the proposed VL-ZSI when compared with the other existing topologies [10, 13, 28]. Table 1 provides the input current nature and it is concluded from Section II, that the proposed topology delivers discrete input current due to the input diode Din . The passive and active components comparisons are made in Table 1. From this table, it can be observed that the number of inductors used in the projected network is less when compared to the SL-ZSI [10] and enhanced-boost (q)ZSIs [28–30] and is same as that of the extended-boost qZSIs [13]. The sum of capacitors used in the projected inverter topology is also fewer when compared to the enhanced-boost (q)ZSIs and CA-qZSI and is same as that of DAqZSI, but it requires one extra capacitor when compared to SL-ZSI. The sum of diodes used in the projected topology is very less when compare to the SL-ZSI and enhancedboost (q)ZSIs, but more or less same as extended-boost qZSIs. Therefore, for the same input voltage and duty cycle with a smaller sum of active and inactive components, the projected inverter topology gives high boost factor and less switch stress. Table 2 gives the comparison of the peak dc-link voltage, capacitor voltages, switch stress, and boost factor for the proposed topology and conventional inverter topologies with the similar input voltage and duty ratio. The comparison made in Fig. 4, Fig. 5, Fig. 6 and Fig. 7 is based on Table 2 and shows that the boost factor and voltage gain of the projected inverter topologies is higher than the all-other mentioned topologies. Moreover, the switch stress in lower in the proposed topology.
Fig. 8. Diode stress assessment of projected VL-ZSI with SL-ZSI [10], extended-boost qZSIs [13], and EB-ZSI [28].
406
V. Jagan and U. Prashanth
4 Impedance Parameter Design The input voltage appears across the capacitors when there is no boost operation involved and the voltage across the inductor is zero. Consequently, a pure dc current drifts through the inductors. In order to boost the voltage, the shoot-through duty ratio must be inserted. Therefore, the ripple current exits in the inductor. The purpose of the inductor and capacitor is to limit the current and voltage ripples respectively. During shoot-through, the inductors are charged by the capacitors and the inductor current rises linearly. The voltage athwart the inductor is same as the capacitor voltage as given below diL1 ⎫ ⎪ VL1 = VL3 = VC1 = VC3 = L1 ⎬ dt (17) diL2 ⎪ ⎭ VL2 = VC2 = L2 dt Therefore, the inductors of the projected topology can be calculated by using ⎫ TD0 (1 − D0 ) ⎪ VDC ⎪ L1,3 = ⎬ iL1,3 k0 (1 − 3D0 ) (18) ⎪ TD0 2(1 − D0 ) ⎭ L2 = VDC ⎪ iL2 k0 (1 − 3D0 ) Where, iL1 , iL2 , and iL3 are the inductor current ripples. In order to design the capacitors of the proposed inverter, the inductor current is same as the capacitor current in shoot-through state. Therefore, the capacitors can be designed by using ⎫ TD0 iL1,3 ⎪ ⎪ C1,3 = VC1,3 k0 ⎬ (19) ⎪ TD0 iL2 ⎪ ⎭ C2 = VC2 k0 where, k 0 is the no. of shoot-through states over a period, T and V C1 , V C2 , and V C3 are the capacitor voltage ripples.
5 Discussion on Simulation Results To validate the hypothetical study of the projected VL-ZS inverter, the simulation is carried out in MATLAB/Simulink at V DC = 70V, D0 = 0.1, and M = 0.9. Simple boost control modulation technique [9] is used for this purpose. As can be observed from (11), the peak dc-link voltage of the proposed topology is 200V theoretically and therefore, input dc voltage, dc-link voltage, inductor currents, and capacitor C 3 voltage of the projected VL-ZSI is shown in Fig. 9(a). From top to bottom, the input current I in , diode Din and D1 voltages, and capacitor C 1 , C 2 voltages of the projected VL-ZSI are portrayed in Fig. 9(b). Figure 10 shows the line voltage and phase voltage before (top) and after (bottom) the filters. In order to validate the simulation results and theoretical analysis carried out
Voltage-Lift-Type Z-Source Inverter
407
Table 2. DC-link voltage, stress across active switches, current stress, and voltage boost comparison with the same input voltage and ratio
at dc input voltage of 70V with modulation index as 0.9 and duty cycle of 0.1. The proposed VL-ZSI also suffers from several drawbacks such as irregular input current, it does not part joint ground with the source, and has huge inrush current during the initial state similar to ZSI [2], SL-ZSI [10], and EB-ZSI [28]. Moreover, the stress across the capacitors is high. In addition, for solar PV grid-connected systems, the proposed VLZSI needs an extra LC-low pass filter at the input side due to the discrete input current drawn from the supply, which increases the cost and space requirements.
408
V. Jagan and U. Prashanth
Fig. 9. Simulations of (a) Input voltage, peak dc-link voltage, inductor currents and capacitor C 3 voltage of proposed VL-ZSI; (b) Input current, diode Din voltage, diode D1 voltage, and capacitor voltages of projected VL-ZSI.
Fig. 10. Simulation of line and phase voltages before filter, and line and phase currents after filter of proposed VL-ZSI.
6 Conclusion The novel topology of ZSI has been proposed in this paper to boost the voltage gain using voltage-lift unit. For the same input voltage and duty ratio, this topology gives high voltage boost with smaller number of active and inactive elements which reduces the
Voltage-Lift-Type Z-Source Inverter
409
space, cost, weight, and complexity of the system. Furthermore, it features higher initial voltage gain (i.e., at zero shoot-through duty cycle) than any other existing topologies with the same shoot-through duty ratio. To get the same voltage boost, the projected VL-ZSI utilizes less duty ratio and gives high modulation index which reduce stress across the semiconductor switches and provide a better-quality output waveform.
References 1. Shen, M., Joseph, A., Wang, J., Peng, F.Z., Adams, D.J.: Comparison of traditional inverters and Z-source inverter for fuel cell vehicles. IEEE Trans. Power Electron. 22(4), 1453–1463 (2007) 2. Peng, F.Z.: Z-source inverter. IEEE Trans. Ind. Appl. 39(2), 504–510 (2003) 3. Anderson, J., Peng, F.Z.: Four quasi-Z-Source inverters. In: Proceedings of IEEE Power Electronics Specialists Conference, PESC 2008, pp. 2743–2749, June 2008 4. Tang, Y., Xie, S.J., Zhang, C.H., Xu, Z.G.: Improved Z-source inverter with reduced Z-source capacitor voltage stress and soft-start capability. IEEE Trans. Power Electron. 24(2), 409–415 (2009) 5. Gao, F., Loh, P.C., Li, D., Blaabjerg, F.: Asymmetrical and symmetrical embedded Z-source inverters. IET Power Electron. 4(2), 181–193 (2011) 6. Peng, F.Z., Shen, M., Qian, Z.: Maximum boost control of the Z-source inverter. IEEE Trans. Power Electron. 20(4), 833–838 (2005) 7. Shen, M., Wang, J., Joseph, A., Peng, F.Z., Tolbert, L.M., Adams, D.J.: Constant boost control of the Z-source inverter to minimize current ripple and voltage stress. IEEE Trans. Ind. Appl. 42(3), 770–778 (2006) 8. Ali, U.S., Brindha, G., Priya, A.H., Karthikeyan, M.N.: A novel carrier based pulse width modulation technique for quasi-z-source inverter with improved voltage gain. In: Proceedings of International Conference on Circuits, Power and Computing Technologies (ICCPCT), Nagercoil, pp. 199–202 (2013) 9. Loh, P.C., Vilathgamuwa, D.M., Lai, Y.S., Chua, G.T., Li, Y.W.: Pulse-width modulation of Z-source inverters. IEEE Trans. Power Electron. 20(6), 1346–1355 (2005) 10. Zhu, M., Yu, K., Luo, F.L.: Switched inductor Z-source inverter. IEEE Trans. Power Electron. 25(8), 2150–2158 (2010) 11. Nguyen, M.K., Lim, Y., Cho, G.: Switched-inductor quasi-Z-source inverter. IEEE Trans. Power Electron. 26(11), 3183–3191 (2011) 12. Nguyen, M.K., Lim, Y.C., Choi, J.H.: Two switched-inductor quasi-Z-source inverters. IET Power Electron. 5(7), 1017–1025 (2012) 13. Gajanayake, C.J., Lin, L.F., Hoay, G., Lam, S.P., Kian, S.L.: Extended-boost Z-source inverters. IEEE Trans. Power Electron. 25(10), 2642–2652 (2010) 14. Zhu, M., Li, D., Loh, P.C., Blaabjerg, F.: Tapped-inductor Z-source inverters with enhanced voltage boost inversion abilities. In: Proceedings of the IEEE International Conference on Sustainable Energy Technologies, pp. 1–6, December 2010 15. Jagan, V., Das, S.: Two-tapped inductor quasi impedance source inverter (2TL-qZSI) for PV applications. In: Proceedings of 6th IEEE International Conference on Power Systems, (ICPS 2016), New Delhi, India, pp.1–6 (2016) 16. Loh, P.C., Blaabjerg, F.: Magnetically coupled impedance-source inverters. IEEE Trans. Power Electron. 49(5), 2177–2187 (2013) 17. Jun Soon, J., Low, K.S.: Sigma-Z-source inverters. IET Power Electron. 8(5), 715–723 (2015) 18. Nguyen, M.K., Lim, Y.C., Choi, J.H., Choi, Y.O.: Trans-switched boost inverters. IET Power Electron. 9(5), 1065–1073 (2016)
410
V. Jagan and U. Prashanth
19. Li, D., Loh, P.C., Zhu, M., Gao, F., Blaabjerg, F.: Enhanced-boost Z-source inverters with alternate-cascaded switched-and tapped-inductor cells. IEEE Trans. Ind. Electron. 60(9), 3567–3578 (2013) 20. Ahrabi, R.R., Banaei, M.R.: Improved Y-source DC–AC converter with continuous input current. IET Power Electron. 9(4), 801–808 (2016) 21. Ravindranath, A., Mishra, S.K., Joshi, A.: Analysis and PWM control of switched boost inverter. IEEE Trans. Ind. Electron. 60(12), 5593–5602 (2013) 22. Nguyen, M.K., Le, T.V., Park, S.J., Lim, Y.C., Yoo, J.Y.: Class of high boost inverters based on switched-inductor structure. IET Power Electron. 8(5), 750–759 (2015) 23. Nguyen, M.K., Le, T.V., Park, S.J., Lim, Y.C.: A class of quasi-switched boost inverters. IEEE Trans. Ind. Electron. 62(3), 1526–1536 (2015) 24. Babaei, E., Shokati Asl, E., Hasan Babayi, M., Laali, S.: Developed embedded switched-Zsource inverter. IET Power Electron. 9(9), 1828–1841 (2016) 25. Li, L., Tang, Y.: A high set-up quasi-Z-source inverter based on voltage-lifting unit. In: Proceedings of IEEE Energy Conversion Congress and Exposition (ECCE), Pittsburgh, PA, pp. 1880–1886 (2014) 26. Jagan, V., Das, S.: High boosting type Z-source inverter/ improved Z-source inverter for solar photovoltaic system. In: Proceedings of Annual IEEE India Conference (INDICON), New Delhi, pp. 1–6 (2015) 27. Ellabban, O., Abu-Rub, H.: Z-source inverter: topology improvements review. IEEE Ind. Electron. Mag. 10(1), 6–24 (2016). https://doi.org/10.1109/MIE.2015.2475475 28. Fathi, H., Madadi, H.: Enhanced-boost Z-source inverters with switched Z-impedance. IEEE Trans. Ind. Electron. 63(2), 691–703 (2016) 29. Jagan, V., Kotturu, J., Das, S.: Enhanced-boost Quasi-Z-source inverters with two-switched impedance networks. IEEE Trans. Ind. Electron. 64(9), 6885–6897 (2017) 30. Jagan, V., Reddy, J.P.K., Kollati, L., Naresh, G., Medishetty, N.: Continuous input current configurations of enhanced-boost Quasi-Z-source inverters. In: 2019 IEEE 16th India Council International Conference (INDICON), Rajkot, India, pp. 1–4 (2019)
Smart Energy Management for a Campus Network B. Subba Reddy , Febin Francis(B) , and L. Umanand Indian Institute of Science, Bengaluru, Karnataka 560012, India [email protected]
Abstract. Implementation of smart meters with replacing the traditional energy meters are increasing each day because of the possibility of two-way communication between the customer and the utility. The data from the smart meters provide an opportunity for utilities and the clients in maintaining the accurate database, the data will be useful to understand various parameters like: phase balancing, energy monitoring, billing, planning for future loads, outage management, to find energy losses, harmonic distortions etc. The present work describes the development of smart energy management for the institute campus network. The data obtained from the smart sensors/data logger and developed customized software helps in precious monitoring and control of devices employed across multiple locations in the campus microgrid. The work highlights some important outcomes of the smart energy management system (SEMS) adopted in the campus network like phase imbalance, maximum power consumption, total harmonic distortion etc. Keywords: Energy management · Smart metering · Microgrid · Phase imbalance · Harmonics
1 Introduction The use of Smart metering replacing traditional energy meters are increasing day to day because of the two-way communication possible between the customer and utility. Smart meters have many advantages over conventional energy meters [1]. Smart meters are more reliable than the conventional energy meters as these are more accurate and can be developed to the customized requirement. Smart meter along with the data logger, sensors and customized software can be used to monitor and control various electrical parameters thereby controlling electrical devices. The SEMS [2–4], assists in proper load management and control and possibly increasing the reliability of electrical grid. The data obtained from the installed smart meters provides an opportunity for utilities/organizations to carefully plan for the phase balancing, energy monitoring, billing purpose, planning for future loads, calculating energy loss etc. Presently at our campus we have installed more than 260 smart meters at various departments/utility, hostels etc., the data is being regularly monitored and analysed. The SEMS helps to control devices employed in HVAC and lighting installations at different locations in the campus microgrid [3–6]. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 411–418, 2022. https://doi.org/10.1007/978-981-19-1677-9_37
412
B. Subba Reddy et al.
The campus is powered by one Master Unit Substation (MUSS) of 66/11kV, further the loads from various departments/units etc. are connected to 25 substations (SS1SS25) which is fed from MUSS. The position where smart meters are placed are known as nodes, at each entry point of substations there will be a node. This will be useful to measure the losses occurring between the MUSS and the entry point of substations. Smart energy metering system consists of Current transformer, smart energy meters, Data loggers, sensors, server for data storage, application software etc. Figure 1 presents the schematic of various components used for the smart energy management system. The system typically consists of a CT mounted onto a busbar at the incomer, the smart energy meter connects CT and PT in case of HT/EHT connections. The smart meter is connected to data logger through a mod bus. Data logger stores the data available from the smart meter. SMPS powers data logger, further data logger is connected through Modem/WiFi or LAN to the server. Thedata logger sends the data/information in intervals to the main server where the entire data is stored and used for further processing. Currently we are using SEMS for monitoring and is also used for various applications like phase balancing, planning of future loads etc.
2 Smart Energy Management System The Energy management system is generally adopted by individual and commercial organisations for monitoring, measurement, and for controlling the electrical loads for both HVAC and lighting installations across multiple locations in the distribution network. Presently smart energy management is developed for the institute wide campus network. The primary goal of the development is to accurately monitor the distribution system, take appropriate control decisions, and to optimize the entire electrical distribution network in the campus. The monitoring and control system is designed for observing the real time data for various parameters, status of different devices used for communication, historical information for analyzing the past performance of the electrical network. This will help in framing and adopting preventive measures to avoid faults, phase balances in the distribution system. The typical components required for a smart metering include: CTs and PTs, Energy meter, Data logger, Server, application software etc. The smart metering is mainly adopted to measure apparent and active and reactive power, energy consumption, phase and line voltages/currents, phase imbalance, power factor status, peak demands, power down periods, harmonic distortions along with these weekly/monthly power demand, fault alerts etc. are also embedded in the present developed software. Smart meters are installed across the campus micro-grid at more than 250 nodes as shown in Fig. 2. Current, voltage and power data is monitored using smart meters with SCADA network. A web browser-based GUI is developed, wherein prediction of energy usage to make appropriate control decisions for maintenance and actuation. Django based free open-source web framework interface is developed for viewing, monitoring and analysing the data.
Smart Energy Management for a Campus Network
413
Fig. 1. Schematic of installation of Smart Energy Management System
Fig. 2. Installation of Smart meters at different locations in campus
3 Data Monitoring and Analysis The SEMS is developed to monitor and analyse the obtained data for all the major parameters namely phase imbalance, energy consumption, outage time, apparent power, reactive power, current total harmonic distortion, voltage total harmonic distortion etc. Phase imbalance is one of the important parameter of concern, which is mainly due to single phase loads. Figure 3, 4 and 5 present active, apparent and reactive power profiles, while Fig. 6 shows the total energy consumption for the year 2020 respectively. Figure 7,
414
B. Subba Reddy et al.
8 and 9 present typical profiles of phase imbalance, outage time and harmonic distortion for a particular department, while, Fig. 10 presents a typical powerfactor profile for MUSS. Figure 11 shows the integrated solar energy generated at local grid at one of the department. The consumption of energy and billing of various departments/units in our campus is also done through SEMS, a typical billing module developed is shown in Fig. 12. SEMS helps in proper planning for the design of DG sets or integration of renewable sources like solar PVetc.
Fig. 3. Active power profile for the year 2020 for the entire campus
Fig. 4. Apparent power profile for the year 2020 for the entire campus
Smart Energy Management for a Campus Network
Fig. 5. Reactive power profile for the year 2020 for the entire campus
Fig. 6. Total Energy consumption profile for the year 2020 for the campus
Fig. 7. Typical Profile of Phase imbalance for MUSS
415
416
B. Subba Reddy et al.
Fig. 8. Typical Outage time of one of the Dept (DESE) sensor node
Fig. 9. Typical Total Harmonic Distortion profile of DESE sensor node
Fig. 10. Typical Monthly Power factor profile for Feb 2021 at MUSS
Smart Energy Management for a Campus Network
417
Fig. 11. Solar generated power integrated to grid at one of the Dept.
Fig. 12. Typical developed billing module.
4 Summary and Conclusions Smart metering is implemented at the campus microgrid, helps in load monitoring and future grid reliability. The data collected from the smart meters is analyzed and is used
418
B. Subba Reddy et al.
for preventive measures like phase balancing, planning for future loads, addition of loads and so on. The important objective of the present work is to adopt a multimodal arrangement intended towards saving energy, reduce carbon-footprint both at personal and the community level. Further integration of multiple generation including solar PV, building integrated PV, Diesel generating plants is in progress. Smart energy management system with more than 260 sensors are installed at various locations at the campus, the system is used for monitoring microgrid data and taking, appropriate control decisions. Acknowledgment. The research work is supported by the Department of Science and Technology, project entitled: UK India Clean Energy Research Institute (UKICERI) (Sanction No. DST/RCUK/JVCCE/2015/02).
References 1. Zheng, J., Gao, D.W., Lin, L.: Smart meters in smart grid: an overview. In: IEEE Green Technologies Conference (2013) 2. Kadar, P.: Smart meter based energy management system. In: International Conference on Renewable Energy and Power Quality, Spain 13th to 15th April 2011 (2011) 3. Subba Reddy, B., Umanand, L.: Energy monitoring and control for distribution network using smart metering. In: IEEE International Conference of High Voltage Engineering (ICHVE 2018) Held in Athens, Greece, 10–13 September 2018, Paper code: 71401-14613670-6966 (2018) 4. Subba Reddy, B., Umanand, L.: Development of energy monitoring system for a typical microgrid. Paper ID: 56, APEN-MIT-2019, Applied Energy Symposium; MIT A+B, Boston, 22–24 May (2019) 5. Talei, H., et al.: Smart campus microgrid, advantageous and main architectural components. In: 2015 3rd International Renewable and Sustainable Energy Conference (IRSEC) 10–13 December 2015, Paper No. 07455093 (2015) 6. Kadar, P.: Smart meter based energy management system. In: International Conference on Renewable Energy and Power Quality, Spain, 13th to 15th April 2011, p. 4 (2011)
Design and Development of a New Cascaded Asymmetric 15-Level Multilevel Inverter Topology with Reduced Number of Switches M. Revathi(B) and K. Rama Sudha Department of Electrical Engineering, Andhra University College of Engineering (A), Visakhapatnam, Andhra Pradesh, India [email protected]
Abstract. This paper proposes a single phase asymmetrical fifteen level inverter, a novel topology with reduced switch count for different techniques of pulse width modulation. This topology results in less count of power switches (IGBT’S), diodes, gate driven circuits, number of voltage sources, control circuits, and no passive components like capacitors that even reduces the installation area, circuit complexity and hence the inverter cost. Due to which losses will decrease and so the system efficiency increases. In this topology the voltage output total harmonic distortion (THD) reduces. In this paper different techniques of PWM are used to generate output with 15-level. The outputs of the proposed asymmetric15-level inverter are determined using MATLAB/SIMULINK software. Keywords: Multilevel inverter (MLI) · Pulse width modulation (PWM) techniques · Total harmonic distortion (THD)
1 Introduction In recent years, multilevel inverters because of their high voltage operation, low electromagnetic interference (EMI have more importance in the power industry) [1]. These inverters by using multiple dc voltage sources as the input generate stepped voltage waveform. A multilevel inverter has numerous advantages like the output voltage is generated with low distortion, low dv/dt stress, and high voltage capability. It produces small common mode voltage, with low distortion draws the input current. Since the usage of power switches reduced the switching losses are reduced resulting in increase of efficiency for the system [2–4].
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 419–432, 2022. https://doi.org/10.1007/978-981-19-1677-9_38
420
M. Revathi and K. R. Sudha
In this paper asymmetrical multilevel inverter with 15 level topology has been anticipated with less count of dc sources, switches, driver circuits. The proposed topology uses no capacitors so control circuit for balancing the capacitor voltages is not required and results in reduction of complexity in circuit and the cost (Table 1).
Fig. 1. Generalised structure of proposed inverter
Table 1. Switching scheme of proposed 15level inverter Generation of level
Switching scheme
Output voltages
S1
S2
S3
S4
S5
S6
S7
S8
S9
S10
7
0
1
0
0
1
0
1
0
0
1
V1 + V2 + V3
6
0
0
1
0
1
0
1
0
0
1
5
1
0
0
0
1
0
1
0
0
1
V2 + V3 V1 + V3
4
0
0
0
1
0
0
1
0
0
1
V3
3
0
1
0
0
0
1
1
0
0
1
2
0
0
1
0
0
1
1
0
0
1
V1 + V2 V2
1
1
0
0
0
0
1
1
0
0
1
V1
0
0
0
0
0
0
0
1
1
0
0
0
−1
1
0
0
0
0
1
0
1
1
0
−V1
−2
0
0
1
0
0
1
0
1
1
0
−V2
−3
0
1
0
0
0
1
0
1
1
0
−4
0
0
0
1
0
0
0
1
1
0
−(V1 + V2 ) −V3
−5
1
0
0
0
1
0
0
1
1
0
−6
0
0
1
0
1
0
0
1
1
0
−(V1 + V3 ) −(V2 + V3 )
−7
0
1
0
0
1
0
0
1
1
0
−(V1 + V2 + V3 )
Design and Development of a New Cascaded Asymmetric
421
1.1 Proposed Multi Level Inverter The proposed inverter depicted in Fig. 1 is asymmetrical source configuration that generates (2(N+1) −1 levels at the output for N voltage sources, and 3N+1 switches and 2N−2 diodes. In the Fig. 1 h-bridge across the load is used for polarity generation so it is called polarity generator and the other part of the circuit is used for level generation so it is called level generator. In this asymmetric configuration each dc voltage source have different magnitude and in binary form. In binary, the magnitudes of source voltage remains in geometric progression of 2N . This topology uses three voltage sources V1 , V2 , V3 of values Vdc , 2Vdc , 4Vdc and ten switches (Fig. 2).
Fig. 2. Proposed Topology
1.2 Mode of Operation for Proposed Inverter See Fig. 3. Simulation Results The developed fifteen level inverter is simulated with the help of MATLAB Simulink. The three DC voltage sources of values V1 = 33V, V2 = 66V, V3 = 132 V, and load resistance R = 100 ohms is taken. Here results for different carrier signals (triangular, sawtooth, reverse sawtooth and square) with least THD values are shown below (Figs. 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17 and Tables 2, 3):
2 Comparison with Other Toplogies The topologies referred below are compared with the developed topology with respect to no. of voltage sources, switches, diodes, passive components like capacitors, gate driven circuits, control circuit etc. As compared to other typologies, for the proposed topology the cost decreases and the losses will decreases as the no. of switches used will be less. • Topology [5], uses 4 input dc symmetrical sources and 12 switches for 9-level.out of 12 switches, 4switches are for polarity generation and remaining are for level generation.
422
M. Revathi and K. R. Sudha
(a)+1VDC
(c)+2VDC
(e)+7VDC
(b)-1VDC
(d)-2VDC
(f) -7VDC
Fig. 3. Different switching states of the Proposed Inverter
Design and Development of a New Cascaded Asymmetric
423
Fig. 4. APOD-PWM Technique with carrier as triangular and reference as Sine waveforms
Fig. 5. Output voltage obtained for APOD-PWM technique with triangular as carrier waveform at 400 Hz for the proposed topology fifteen level inverter
Fig. 6. THD obtained for APOD-PWM technique with triangular as carrier waveform at 400 Hz for the proposed topology fifteen level inverter
424
M. Revathi and K. R. Sudha
Fig. 7. POD-PWM technique with carrier as sawtooth and reference as Sine waveforms at 100 Hz for the developed fifteen level inverter
Fig. 8. Output voltage obtained for POD-PWM Technique with sawtooth as carrier wave at 100 Hz for the proposed topology fifteen level inverter
Fig. 9. THD obtained for POD-PWM Technique with sawtooth as carrier waveform at 100 Hz for the developed fifteen level inverter
Design and Development of a New Cascaded Asymmetric
425
Fig. 10. POD-PWM technique with reverse sawtooth as carrier and Sine wave reference waveforms for 100 Hz for the developed fifteen level inverter
Fig. 11. Output voltage obtained for POD-PWM Technique with reverse sawtooth as carrier waveform at 100 Hz for the proposed topology fifteen level inverter
Fig. 12. THD obtained for POD-PWM Technique with reverse sawtooth as carrier waveform at 100 Hz for the proposed topology fifteen level inverter
426
M. Revathi and K. R. Sudha
Fig. 13. POD-PWM technique with square as carrier and Sine wave reference waveforms for 100 Hz for the proposed topology fifteen level inverter
Fig. 14. Output waveform obtained for POD-PWM Technique with square as carrier waveform at 100 Hz for the proposed topology fifteen level inverter
Fig. 15. THD obtained for POD-PWM Technique with square as carrier waveform at 100 Hz for the proposed topology fifteen level inverter
Design and Development of a New Cascaded Asymmetric
427
Table 2. Comparison of THD for different carrier waveforms at 100 Hz for PD, POD, APOD SPWM Techniques for the proposed topology fifteen level inverter. Carrier waveform
PD
POD
APOD
Triangle
7.66%
8.12%
7.80%
Sawtooth
7.46%
6.24%
7.73%
Reverse sawtooth
8.58%
5.72%
7.88%
11.35%
7.93%
11.62%
Square
Total Harmonic Distortion for fifteen level inverter at 100Hz for PD,POD,APOD PWM Techniques for multi carrier waveforms 14.00% 12.00% 10.00% 8.00% 6.00% 4.00% 2.00% 0.00%
TRIANGLE
SAWTOOTH PD
POD
REVERSE SAWTOOTH
SQUARE
APOD
Fig. 16. THD comparison for different PWM techniques at 100 Hz frequency for the proposed topology fifteen level inverter
Table 3. THD comparison for PD, POD, APOD-PWM techniques for different values of modulation index for the developed fifteen level inverter Total harmonic distortion Modulation index
PD
POD
APOD
0.2
33.26%
30.83%
34.83%
0.4
17.24%
15.64%
18.68%
0.6
12.30%
9.54%
13.05%
0.8
9.72%
8.32%
10.64%
1
8.58%
5.72%
7.88%
428
M. Revathi and K. R. Sudha TOTAL HARMONIC DISTORTION FOR FIFTEEN LEVEL INVERTER FOR DIFFERENT LS-PWM TECHNIQUES AT DIFFERENT MODULATION INDEX
TOTAL HARMONIC DISTORTION
PD
POD
APOD
40.00% 35.00% 30.00% 25.00% 20.00% 15.00% 10.00% 5.00% 0.00%
0.2
0.4
0.6 0.8 MODULATION INDEX
1
Fig. 17. THD variation with respect to the modulation index for different modulation techniques for the developed topology fifteen level inverter
• In this new topology [6], Developed cascaded cell (DCC) based MLI uses 4 input dc sources and 11switches for 11 level multilevel inverter. Out of 11switches, 4 switches are for polarity generation and remaining are for level generation. • In this topology [7], Novel multilevel inverter uses 1 DC source, 3 capacitors, and 8 switches and 4 diodes for 7 level. In this topology, passive components are used which adds an extra cost. Even control circuit is needed for balancing capacitor voltages. Diodes, capacitors and control circuit adds the extra cost. • In this topology [8], cross connected source based MLI uses 2 voltage sources, 3 capacitors, 8 switches in which 2 are bidirectional for 15 level. By using passive components, capacitors which adds an additional cost. Even bidirectional switches are used i.e. over all 10 switches are used and the inverter cost increases. • In this topology [9], Staircase cascaded MLI uses 4sources, 8 switches in which 2 are bidirectional for 9 level. The no. of voltage sources and switches used are more. • In this new topology [10], Cascaded unit based multilevel inverter uses 2voltage sources, 6 switches, bidirectional auxiliary circuit (four quadrant switch it has one power switch and four diodes) for 11-level. • In this topology [11], a switched series/parallel sources (SSPS) based MLI uses 4 DC sources, and 13 switches for 15 level.
Design and Development of a New Cascaded Asymmetric
429
For all the topologies mentioned above uses more number of voltage sources, power switches, capacitors, diodes, gate driven circuits and control circuits that increases the inverter cost, area of insulation and complexity of the circuit. Due to which switching and conduction losses will increase, resulting in decrease of efficiency of the system (Table 4). Table 4. Parameters comparison table of developed inverter with existing topologies MLI’s
No. of levels
No. of voltage sources
No. of switches
No. of diodes
No. of capacitors
1
9
4
12
–
–
2
11
4
11
–
–
3
7
1
8
4
3
4
15
2
10
–
3
5
9
4
10
–
–
6
11
2
7
4
–
7
15
4
13
–
–
Proposed MLI
15
3
10
4
–
No. of Levels: See Fig. 18.
No. of Levels 16 14 12 10 8 6 4 2 0
Fig. 18. Comparison of developed circuit with existing MLI configurations with respect to number of levels
430
M. Revathi and K. R. Sudha
No. of Voltage Sources: See Fig. 19.
No. of Voltage Sources 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0
Fig. 19. Comparison of developed circuit with existing MLI configurations with respect to number of voltage sources
No. of Switches: See Fig. 20.
No. of Switches 14 12 10 8 6 4 2 0
Fig. 20. Comparison of developed circuit with existing MLI configurations with respect to number of switches
Design and Development of a New Cascaded Asymmetric
431
No. of Diodes: See Fig. 21.
No. of Diodes 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0
Fig. 21. Comparison of developed circuit with existing MLI configurations with respect to number of diodes
No. of Capacitors: See Fig. 22.
No. oCapacitors 3.5 3 2.5 2 1.5 1 0.5 0
Fig. 22. Comparison of developed circuit with existing MLI configurations with respect to number of capacitors
432
M. Revathi and K. R. Sudha
3 Conclusion A new asymmetrical multi-level inverter with 15-level topology is developed and presented in this paper with the main objective to reduce power electronic components. The working and sequence of operation of the proposed circuit is discussed and simulated. Comparison of THD for different carrier waveforms at 100 Hz and comparison for different modulation index have been carried out It was shown that in proposed inverter the count of power switches and gate driven circuits are decreased as compared to other topologies and hence reduces the cost thus improving the reliability and efficiency.
References 1. Rodriguez, J., Lai, J.-S., Peng, F.Z.: Multilevel inverters: a survey of topologies, controls, and applications. IEEE Trans. Ind. Electron. 49(4), 724–738 (2002). https://doi.org/10.1109/TIE. 2002.801052 2. Baker, R.H., Bannister, L.H.: Electric power converter. U.S. Patent 867 643 (1975) 3. Ebrahimi, J., Babaei, E., Gharehpetian, G.B.: A new multilevel converter topology with reduced number of power electronics components. IEEE Trans. Industr. Electron. 59, 655–667 (2012) 4. Rahim, N.A., Chaniago, K., Selvaraj, J.: Single-phase seven level grid-connected inverter for photovoltaic system. IEEE Trans. Ind. Electron. 58(6), 2435–2444 (2011) 5. Bahravar, S., Babaei, E., Hosseini, S.H.: New cascaded multilevel inverter topology with reduced variety of magnitudes of DC voltage sources. In: Proceedings of IEEE 5th India International Conference on Power Electronics (IICPE) New Delhi, India, pp. 1–6 (2012) 6. Babaei, E., Laali, S., Bayat, Z.: A single-phase cascaded multi-level inverter based on a new basic unit with reduced number of power switches. IEEE Trans. Ind. Electron. 62(2), 922–929 (2015) 7. Hsieh, C.H., Liang, T.J., Chen, S.M., Tsai, S.W.: Design and implementation of a novel multilevel DC-AC inverter. IEEE Trans. Ind. Appl. 52(3), 2436–2443 (2016) 8. Agrawal, R., Jain, S.: Multilevel inverter for interfacing renewable energy sources with low/medium- and high-voltage grids. IET Renew. Power Gener. 11(14), 1822–1831 (2017) 9. Samsami, H., Taheri, A., Samanbakhsh, R.: New bidirectional multi-level inverter topology with staircase cascading for symmetric and asymmetric structures. IET Power Electron. 10(11), 1315–1323 (2017) 10. Odeh, C.I., Obe, E.S., Ojo, O.: Topology for cascaded multilevel inverter. IET Power Electron. 9(5), 921–929 (2016) 11. Hinago, Y., Koizumi, H.: A single phase multilevel inverter using switched series/parallel DC voltage sources. In: Energy Conversion Congress and Exposition, ECCE 2009, pp.1962–1967. IEEE (2009) 12. Hinago, Y., Koizumi, H.: A single-phase multilevel inverter using switched series/parallel DC voltage sources. IEEE Transa. Ind. Electron. 57(8), 2643–2650 (2010) 13. Gupta, K.K., Jain, S.: Multilevel inverter topology based on series connected switched sources. IET Power Electron. 6(1), 164–174 (2013) 14. Gupta, K.K., Jain, S.: A novel multilevel inverter based on switched DC sources. IEEE Trans. Ind. Electron. 61(7), 3269–3278 (2014)
A Novel Deep Learning Based Nepali Speech Recognition Basanta Joshi, Bharat Bhatta, Sanjeeb Prasad Panday(B) , and Ram Krishna Maharjan Pulchowk Campus, Institute of Engineering, Tribhuvan University, Lalitpur, Nepal {basanta,sanjeeb,mrrk}@ioe.edu.np
Abstract. Automatic speech recognition has allowed human beings to use their voices to speak with a computer interface. Nepali speech recognition involves conversion of Nepali language to corresponding text in Devanagari lipi. This work proposes a novel approach for developing Nepali Speech recognition model based using CNN-GRU. The data is collected from the Librispeech. The collected data is pre-processed and MFCC is applied on it for feature extraction. CNN-GRU model is responsible for extraction of the features and development of the acoustic model. CTC is responsible for decoding. The performance of the developed model has been assessed using Word Error Rate of the transcribed text. Keywords: Automatic speech recognition · Nepali speech recognition · Recurrent Neural Network (RNN) · Convolution Neural Network (CNN)
1
Introduction
Speech recognition is the process of conversion of raw audio speech to its textual representation. Nowadays, a machine is able to recognize speech and convert the speech in the form of text. Information technology is an integral part of modern society. Recent advancement in modern communication devices helps to make life easier. Although the majority of the people live in the rural areas. Most of them are not able to read and write well. Speech recognition helps them to be familiar with the technology. To build an automatic speech recognition (ASR) system the prerequisite element is data-set. One of the reasons for not conducting the research work in Nepali ASR is due to insufficient Nepali spoken corpus. Currently available datasets are not enough to build a good Nepali ASR system. ASR systems can be developed by using the traditional or Deep Learning based approach. In traditional approach, Speech Recognition requires separate blocks of phonetic/linguistic constructs, acoustic model and the language model and may involve hidden markov models (HMMs), gaussian mixture models (GMMs), hybrid HMM/GMM systems, dynamic time wrapping (DTW’s), etc. But, in deep c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 433–443, 2022. https://doi.org/10.1007/978-981-19-1677-9_39
434
B. Joshi et al.
learning approach, the performance of the ASR system is significantly improved with using separate phonetic/linguistic constructs [1,2]. Most of the previous works that have been built now could not predict well for the unseen data. This is due to the lack of sufficient Nepali dataset (Nepali spoken corpus). Compared to previous models, the model also needs new architecture that must be able to predict well. And also to increase the performance of this ASR system, the model must be trained with a large amount of speech corpus. This paper proposes a Nepali ASR system based on Deep Learning Approach. The system utilizes the MFCC for feature extraction, CNN to enhance the feature and GRU to create the acoustic model.
2
Review
There have been few research works carried out regarding development of Nepali speech recognition systems. One of traditional methods for Nepali Language based on HMM (Hidden Markov Model) is used for speaker independent isolated word Automatic Speech Recognition (ASR) system [3]. This system is trained and tested by collecting Nepali words from 8 different speakers in a room environment out of which 4 were male and 4 were female. The test cases consist of trained and untrained data. The recorded word was trained 20 times & tested 10 times and the overall accuracy of this ASR system is found to be 75% [3]. One of the first deep learning based approaches for Nepali Speech Recognition model uses RNN (Recurrent Neural Networks) for processing sequential audio data and CTC for maximizing the occurrence probability of the desired labels from RNN output [4]. On implementing a trained model, audio features are processed by RNN and Softmax layer successively. The output from the Softmax layer is the occurrence probabilities of different characters at different time steps. The task of decoding is to find a label with maximum occurrence probability. treating However, the model fails to accurately predict sounds like them as single sounds. These systems need some modification to increase the performance of the model. The size of the neural network and dataset must be significantly increased so that the network can learn accurately. The generalization capabilities of the model can be increased by adding a CNN layer. Some preliminary research regarding LSTM-CTC for Nepali Speech Recognition [5] has already been started by the same research group which is the baseline for this work.
3
Method
Feature extraction and model building (Front end processing) and decoding using lexicon and language model (back end) are the integral parts of the speech recognition system. Features are the attributes that provide uniqueness to objects i.e. they can be easily distinguished from others. An acoustic model contains statistical representations of each of the distinct sounds that makes up a word. The architecture of the model is described in Fig. 1.
A Novel Deep Learning Based Nepali Speech Recognition
435
Fig. 1. Proposed ASR system
3.1
Feature Extraction
Features are the individual measurable property or characteristic of a phenomenon being observed, that can be used as a distinctive attribute or aspect of something. MFCC features are sequences of Acoustic feature vectors where each vector represents information in a small time window of signal [6]. The input analog speech after sampling at 16 kHz is represented as S = [s1 , s2 , s3 , ..., sn ]. Pre-emphasis of the input signal is done by using: S d [nd ] = S[nd ] − αS[nd − 1], where 0.95 < α < 0.99
(1)
Emphasis is carried out by taking α = 0.97. Speech signals are not stationary in nature. The main role of a window is to make signals stationary. The signal S d [td ] is windowed to 25 ms by using a hamming window [7]. The consecutive interval between the two windows is 10 ms. If S[n] is sliced window, w[n] is window used on signal S d [td ] then Sl[n] is given by Eq. 4.2. (2) Sl[n] = w[n]S d [td ] where w[n] is hamming windows [7] given by relation 4.3 2πn w[n] = 0.54−0.46cos N −1
(3)
436
B. Joshi et al.
After performing the windowing operation 512 point FFT has been taken. The output frequencies are modeled to human auditory systems. These FFT are passed into Mel- bank filters. The mel-frequency conversion is given by f M (f ) = 1127 ln 1 + (4) 700 Power spectrum is represented by Mel filterbank. Since humans are less sensitive to small energy changes at high energy than small changes at a low energy level, logarithmic is taken from the output from the Mel filterbank. DCT is taken from the result after applying log energy computation. An orthogonal transformation is carried out by using the DCT. The DCT transformation produces uncorrelated features. The final MFCC-13 coefficients are represented by the X = [x1 , x2 , x3 , , , xT ]. These features need to normalize before training the model. MFCC of hundreds random samples are taken to compute mean and standard deviation. Then all the training features get normalized using the relation ¯ X −X (5) Xn = σ + Δe Δe is a very small number to prevent division from zero.
Fig. 2. Steps for Acoustic modeling
A Novel Deep Learning Based Nepali Speech Recognition
3.2
437
Acoustic Modeling and Decoding
A model or hypothesis is needed to build from the seen data in order to predict the unseen data. The model building is carried out by the neural network. The combination of CNN-GRU network is used to build the model. Here the CNN is used to summarize and collect useful features from the input and RNN in association with CTC is used to build the model. i.e. Deep learning model is used to build the model. Here the objective is to build the model and provide the posterior probability at the output so that the symbol having the greatest posterior probability is provided as output. The whole process can be summarized as shown in Fig. 2. The input layer consists of 13- units. Thirteen features supplied by the MFCC are taken as input to these layers. CNN performs 1d-convolution. Convolution and max pooling are two major tasks of CNN responsible for generating the feature vector. Batch normalization is used as per algorithm shown in Fig. 3 for normalization the output of CNN and then fed to RNN network.
Fig. 3. Algorithm for Batch Normalizing Transform applied to activation x over a mini batch
Now, GRU is used to prepare the model for RNN. The output of GRU is again normalized and transferred to time-distributed layers and output is predicted by the Softmax function which involves calculating the posterior probability of each symbol. The decoding is carried out by the CTC [2]. The input to CTC is supplied from the output of the softmax. The decoded output of CTC is later mapped back to Nepali language characters. This mapped output is the predicted output of the model. The loss is computed by comparing the true labels and predicted labels. Later the model weights get adjusted.
438
3.3
B. Joshi et al.
Evaluation
Word Error Rate (WER) is the common evaluation metric for speech recognition system [8]. It is an average of the sum of substituted words, inserted words and deleted words of predicted output to total number of words in a ground sentence. The value of WER lies within 0 and 1 or 0% and 100%.
4
Experiments
4.1
Dataset
In this work, the dataset is the audio voice file which is provided by Open Speech and Language. Resources [9]. The dataset consists of high quality spoken corpus recorded by eighteen unique female speakers. Every voice clip is mapped with the respective transcriptions. The dataset is divided into parts: train and test set. The ratio of train and test is 80%:20%. 4.2
Experimental Setup
The training is carried out in a local machine with CPU Intel Core i7-8550U and GPU MX150 using python and its library. Table 1 and 2 summarizes the parameters used for Feature extraction using MFCC and CNN Network. In the case of GRU, 200 units with learning rate = 0.015, 100 epochs and momentum of 0.9 is taken during the training to obtain the best result. Table 1. MFCC feature summarization Parameter
Value
Description
Window
Hanning
Use for FIR Filter
FFT
256 point Extract Frequency components
Window duration
20 ms
Max frequecncy
8000 khz Maximum Allowable Frequency
MFCC dimension
13
No of Features provided by MFCC
Mini-batch size
50
Size of each mini batch
Maximum duration of audio clip.
10 s
Audio data having length more than 10 s are not allowed during the training
4.3
Size of each chunk
Result and Analysis
Different experiments were carried out in CPU and GPU by varying parameters like learning rate, mini batch size, momentum, number of epochs. The summary of experiments performed is listed in Table 3.
A Novel Deep Learning Based Nepali Speech Recognition
439
Table 2. CNN parameters summarization Parameters
Value Description
Input dimension
13
Number of neurons in input layers
Filter
200
The dimensionality of the output space (i.e. the number of output filters in the convolution)
Conv stride
2
Specifying the stride length of the convolution.
Dilation rate
1
Specifying the dilation rate to use for dilated convolution
Activation function ReLU Specifying the activation function Table 3. Summary of experiments performed Parameters
Experiment1
Experiment2
Experiment3
Experiment4
Training data (Utterances)
16412
16412
1945
1945 (corpus)
Testing data (Utterances)
9588
9588
570
570 (corpus)
Learning rate
0.03
0.03
0.05
0.015
Momentum
0.9
0.9
0.9
0.9
Batch size
100
300
20
50
Epochs
400
100
100
100
Training machine
CPU Intel CPU Intel CPU Intel GPU NVIDIA Core i7-8550U Core i7-8550U Core i7-8550U GEFORCE MAX 150
Training duration
1.5 days
1.3 days
1.3 days
1.5 h
The first experiment is carried out as per parameters and configuration mentioned in Table 4 and training Loss and validation loss curve is shown in Fig. 4. From the graph it can be concluded that the training loss is much less than validation loss. Hence the system gets over fitted. The system can perfectly classify the seen data (data that involves training) but doesn’t classify the unseen data. The second experiment is carried out as per parameters and configuration mentioned in Table 5 and training Loss and validation loss curve is shown in Fig. 5. The third experiment is carried out as per parameters and configuration mentioned in Table 6 and training Loss and validation loss curve is shown in Fig. 6. During the training after 45 epochs, sudden abrupt discontinuity is observed. The training strikes in local maxima. And with an increase in the number of iterations the model does not learn any things. The training loss and validation loss remain unchanged. To eliminate this problem we can apply two
440
B. Joshi et al. Table 4. Summary of first experimental setup Model
CNN+BN+GRU+BN+Dense layer+softmax+CTC
GRU units
200
Filter size
200
Kernel size
11
MFCC
dimension = 13, window = 25 ms, filter-bank = 26
Mini-batch size 100 Optimizer
SGD (learning rate = 0.03, decay = 1e−6, momentum = 0.9
Epoch
400
Fig. 4. Training loss and validation loss curve of first experiment
Fig. 5. Training loss and validation loss curve of second experiment
techniques. First is to rearrange the network and other to increase the value of dropout. Re-arranging the network may be a tedious task so first the dropout is increased to 0.7 from 0.5. No prediction is made and another experiment is set up. The fourth experiment is carried out as per parameters and configuration mentioned in Table 7 and training Loss and validation loss curve is shown in Fig. 7. Some sample of the output of the model are listed as Table 8. The summary of the different experiments can be visualized as shown in Fig. 8.
A Novel Deep Learning Based Nepali Speech Recognition Table 5. Summary of second experimental setup Model
CNN+BN+GRU+BN+Dense layer+softmax+CTC
GRU Units
200
Filter size
200
Kernel size
11
MFCC
dimension = 13, window = 25 ms, filter-bank = 26
Mini-batch size 300 Optimizer
SGD (learning rate = 0.03, decay = 1e−6, momentum = 0.9
Epoch
100 Table 6. Summary of third experimental setup
Model
CNN+BN+GRU+BN+Dense layer+softmax+CTC
GRU Units
200
Filter size
200
Kernel size
11
MFCC
dimension = 13, window = 25 ms, filter-bank=26
Mini-batch size 20 Optimizer
SGD (learning rate = 0.05, decay = 1e−6, momentum = 0.9
Epoch
100
Fig. 6. Training loss and validation loss curve of third experiment Table 7. Summary of forth experimental setup Model
CNN+BN+GRU+BN+Dense layer+softmax+CTC
GRU Units
200
Filter size
200
Kernel size
11
MFCC
dimension = 13, window = 25 ms, filter-bank = 26
Mini-batch size 50 Optimizer
SGD (learning rate = 0.015, decay = 1e−6, momentum = 0.9
Epoch
100
441
442
B. Joshi et al.
Fig. 7. Training loss and validation loss curve of fourth experiment Table 8. Sample Results of Nepali ASR Ground Truth
Model prediction
WER 0.1667 0.111 0.250
Fig. 8. Training Loss and Validation loss curve of fourth experiment
5
Conclusion
In this work, an approach for nepali speech recognition CNN-GRU has been proposed. The MFCC is used for feature extraction and takes window size of 20 ms, FFT of 256 points and hanning windows as FIR filter. The dimension of MFCC is 13. These MFCC features are applied to CNN. Where CNN performs 1-d convolution and pooling to extract useful features. The features are applied
A Novel Deep Learning Based Nepali Speech Recognition
443
to GRU network which is used for acoustic model building. After hyperparameter tuning, the best model is built by taking the parameters learning rate 0.015, batch size 50, momentum 0.9 and dropout rate of 0.65. The built system effectively converts correctly spoken Speech to Nepali text. The WER is found to be 10.67%. The system performance can be increased by adding the Lexicon and building Language model, which can be done in future. Training the system with a large number of valid corpus will increase the system performance. Acknowledgement. This work has been supported by the University Grants Commission, Nepal under a Faculty Research Grant (UGC Award No. FRG-76/77-Engg-1) for the research project “Preparation of Nepali Speech Corpus: Step towards Efficient Nepali Speech Processing”.
References 1. Karita, S., et al.: A comparative study on transformer vs rnn in speech applications. In: 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp. 449–456. IEEE (2019) 2. Passricha, V., Aggarwal, R.K.: Convolutional neural networks for raw speech recognition. In: From Natural to Artificial Intelligence-Algorithms and Applications. IntechOpen (2018) 3. Ssarma, M.K., Gajurel, A., Pokhrel, A., Joshi, B.: Hmm based isolated word nepali speech recognition. In: 2017 International Conference on Machine Learning and Cybernetics (ICMLC), vol. 1, pp. 71–76. IEEE (2017) 4. Regmi, P., Dahal, A., Joshi, B.: Nepali speech recognition using rnn-ctc model. Int. J. Comput. Appl. 178(31), 1–6 (2019) 5. Bhatta, B., Joshi, B., Maharjhan, R.K.: Nepali speech recognition using CNN, GRU and CTC. In: Proceedings of the 32nd Conference on Computational Linguistics and Speech Processing (ROCLING 2020), pp. 238–246 (2020) 6. Gupta, R., Sivakumar, G.: Speech recognition for Hindi language. IIT BOMBAY (2006) 7. Kopparapu, S.K., Laxminarayana, M.: Choice of mel filter bank in computing MFCC of a resampled speech. In: 10th International Conference on Information Science, Signal Processing and their Applications (ISSPA 2010), pp. 121–124. IEEE (2010) 8. Kong, X., Choi, J.Y., Shattuck-Hufnagel, S.: Evaluating automatic speech recognition systems in comparison with human perception results using distinctive feature measures. In: 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 5810–5814. IEEE (2017) 9. Sodimana, K., et al.: A step-by-step process for building TTS voices using open source data and framework for Bangla, Javanese, Khmer, Nepali, Sinhala, and Sundanese. In: Proceedings of the 6th International Workshop on Spoken Language Technologies for Under-Resourced Languages (SLTU), pp. 66–70, Gurugram, India, August 2018
Breast Cancer Detection and Prediction Based on Conflicts in Fractal Patterns Krishna Kumar Singh1(B) , Deepmala Jasuja2 , and M. P. Singh3 1 Symbiosis Centre for Information Technology, Pune, India
[email protected]
2 Amity Global Business School, Noida, India 3 Graphic Era University, Dehradun, India
[email protected]
Abstract. In the era of big data analytics, prediction of various deceases with the help data science methods is common among researchers. Although applying tools and using eco-system which is the integration of various sciences and technology is tough for many. Researcher belief that human brain is the most advance machine with all king of tools, technologies and reasoning exist. Visual analytics is the most simple and effective means to detect and prevent any deceases. In this research, authors used story telling methods based on visual analytics to explore different conflicts exist in the data set. In this research, researcher found that average age of detection of breast cancer is 45 to 49 in most of the cases and data sets. But by seeing different conflicts in the data as well as patterns of classification, we can able to predict and detect breast cancer much earlier based on visual analytics and storytelling method. Conflicts in the data shows variation in the usual patterns and this leads to the different problems in subject. Detection and prediction of diseases like breast cancer is most effect in this era of medical science. Based on the conflicting pattern in the data, authors build story and tried to give resolution of the problem in terms of early detection and prevention of breast cancer. Keywords: Breast cancer · Predictive analytics · Conflicts in pattern · Storytelling method · Fractal theory
1 Introduction Cancer has been recognized as one of the deadly diseases that causes deaths in human. Cancer deaths are being recognized as a major issue in healthcare domain. A lot of women every year are dying due to different types of cancer. One of the major causes of death amongst the women is breast cancer with quite denser breast tissues due to its physiological features [1]. Breast cancer represents 12% of new cancer cases & 25% of all cancers in women. Earlier, the higher chances of contraction of this disease are prominently seen in urban areas, but it has shown an exponential trend globally is 1 million women every year are diagnosed with this type of cancer. It is recognized as a significant health problem worldwide [3]. Scientists across the world have been very early aware of the consequences of the disease hence extant literature is being © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 444–455, 2022. https://doi.org/10.1007/978-981-19-1677-9_40
Breast Cancer Detection and Prediction Based on Conflicts
445
contributed towards the detection of cancer [8]. According to the statistics of Cancer Research UK, “the five years survival rate for BC is almost 100% if detected at its earliest stage but can be as low as 15% when detected at the latest stage”. Breast Cancer is developed by breast tumors usually due to the outgrowth of cells in the lining of breast ducts. A modification in DNA or RNA could turn normal cells into cancer cells. The probable reasons behind theses mutations could be chemicals in the air, bacteria, fungi, food, entropy, radiations, and viruses. Known causes of deadly breast cancer includes ionizing radiations, exogenous hormones of ovary and alcohol beverage. Cancerous cells are detected by MRI, Mammogram, Ultrasound, and biopsy. During the early stages of disease, the symptoms are not clearly visible, due to which mammogram becomes necessary for screening. Mammogram is an X-ray of breast for screening of some tumors without any side effects and seeming the procedure to be safe [7]. Several studies shows that Mammography is the most valuable technique for detecting breast cancer. Diagnosis of breast cancer is done by classifying tumors. Tumors can be either benign or malignant [6]. Malignant tumors are more harmful than the benign. In benign tumor, cells grow abruptly and form a lump. But this doesn’t get spread throughout the body and hence, cannot be called cancers. Lumps can be surgically removed in the form of diagnosis. On the other side, in Malignant tumor, these cells have capacity to spread beyond the breast, if left untreated. There are no preventive techniques as such, only early diagnosis and relevant treatment may determine the chances of survival. Hence, it has become crucial to detect and cure this type of deadly disease at an early stage. If caught at an early stage, it can be cured with fewer risks and lower mortality [2]. Advancements in science, recent technological progress, increase in the number of medical studies, diagnosis of Cancer is still a problem.Data Mining techniques have found to be effective to predict future data easily, especially when large information is available in hand. Enormous amount of data is generated in healthcare systems, there is a strong need to utilize the hidden potential of data generated. Over the past couple of decades, Machine Learning (ML) techniques and algorithms are widely used and accepted in intelligent healthcare systems. These techniques are also known for its speed and accuracy. Further these days, machine learning algorithms are being applied to diagnose Breast Cancer due to high prediction chances and accuracy. Data driven methodologies and ICT tools have become prominent in the decision making [4]. Data Mining techniques are applied in medical science and clinical trials due to high performance, cost reduction, improving patient’s health and making prompt decisions to save lives [5]. The enormous amount of data generated through measurements, examinations and prescriptions have rendered the traditional methods of analyzing trends and patterns redundant. Here comes the role of Big Data and Data Mining in the prediction and determining survivability of various diseases at an earlier stage. The need for the use of ML was triggered because of the incapability of traditional physical examination system of diagnosing the disease [8]. The medical tests provided by the clinics and hospitals are way expensive and difficult for a common man to afford. Without the aid of computing technologies, it was difficult to analyze the complex data and perform critical interrogations. Lately, ML techniques are using the classification techniques to identify people with BC, hence applying an important role in diagnosis and early detection. Various classification methods includes Support Vector Machines (SVM), Artificial Neural Network (ANN), Decision Tree (DTs) and k-nearest
446
K. K. Singh et al.
(k-NNs) techniques which are specifically used in BC diagnosis and prognosis [9]. In the field of medical sciences, big large hospitals and clinics maintain the databases of their patients regarding the symptoms and diagnosis. This historical data is used by the researchers extensively to apply the classification methods and infer the conclusions. Thorough research has been conducted by the academicians and researchers worldwide on the detection and prediction of Breast Cancer using a number of machine learning techniques [1]. This particular research study is based on prediction of breast cancer through conflicts in fractal patterns. Authors have used story telling methods basis the visual analytics to explore and study different conflicts in the dataset of breast cancer patients. Based on the conflicting pattern in data, researchers have attempted to build a story and give a resolution in terms of early detection and prevention of breast cancer. The article is organized as follows: Sect. 2 covers related research and theoretical framework. It is followed by Sect. 3 covering data set specifications and methodology. Section 4 relates to findings and conclusions. Finally, Sect. 5 talks about implications of research and future research directions.
2 Related Work Due to the amount of data created in the healthcare systems, Machine learning can be said as the most appropriate field of machine learning. This section highlights the research studies focused on application of machine learning in detection and diagnosis of Breast Cancer. These studies have applied different approaches to the given classification problem and have inferred the conclusions. Many of the techniques have shown good classification accuracy. Previously in some studies, authors came across varied features from colon biopsy images to detect the cancer and make them distinguish for Malignant and other types [14]. Six distinct features were analyzed including contrast, correlation, entropy, and inverse difference moment. The accuracy was reported to be 90.2%. [15] in their used probabilistic neural network and extracting multi-scale texturing features of the tissues adjacent to micro-calcification to diagnose the breast cancer. Other documents have also applied biochemical analysis, statistical analysis, and tumor characterization for obtaining the results of obese women with breast cancer [16]. Moreover [17] suggested a colon biopsy image based classification method by extracting hybrid features of oriented gradient and variant of statistical moments to distinguish biopsy images and received 98.85% of classification accuracy. Nevertheless, the most of information hidden in micro-calcification of biopsy images were extracted using sample entropy and wavelet entropy features. [10] in their research study applied decision table predictive models for BC survival. Under sampling C5 technique along with bagging algorithm was applied to deal with the classification problem. Decision table model was reported to enhance the accuracy of results. Similar researches have also focused on the comparison of learning classifiers and to extract the best classifier in datasets of breast cancer [2]. The authors have generally used the popular BC detection methods namely Support Vector Machine (SVM) Classifier, Naïve Bayes Classifier boost techniques-CNN (Convolutional Neural Networks) in the prognosis of BC [11]. In [12], researchers analyzed the prediction of survival rate in BC using data mining techniques. SEER Public-Use data was used in investigating three data mining techniques - Naïve Bayes, the back-propagated neural
Breast Cancer Detection and Prediction Based on Conflicts
447
network, and the C4.5 decision tree algorithms. Out of all these, C4.5 decision tree algorithm was found to be the best performing. Neural Networks, a prevailing AI technique carries the ability to construct weight matrixes and represents learning patterns. In [13], an automatic diagnosis system was designed too detect the BC using Neural Network and Association Rules. The proposed performance of the system was compared with NN model. Apart from all these models, Bayesian Network also known as probabilistic network is a powerful tool to describe the uncertainty of any problem. In [18], automatic breast cancer detection support tool was designed using Bayesian Belief Network. This work provided a relationship between diagnosis, physical findings, and laboratory tests. BBN is found to be a technique for medical prediction for breast cancer prognosis & diagnosis. [19] in their research work, applied Naïve Bayes, SVM, Decision tree to classify Wisconsin Breast Cancer and got accuracy results of 96.99% with Support Vector Machines. [20] focussed on the prediction of breast cancer with a total record of 202,932 patient records. All the patient records were split into two groups – Survived and Not Survived. After splitting, Naïve Bayes, Neural Network and c4.5 decision tree algorithms were applied, and it was seen that c4.5 decision tree algorithm achieved highest performance amongst all. It was pointed out by [21] that it is absolutely important to decide parameters than algorithms in order to get better accurate results. Several studies have also applied data mining techniques for the risk prediction of breast cancer. [22] have used Naïve Bayes and J48 decision trees. BC was also diagnosed using JAVA to develop intelligent prediction system and results were performed without any significant delay [23].
3 Cancer Prediction by Finding Conflicts and Building Stories in the Data of Cancer Patients Data Sources: Researchers collected data from medical agencies in India and considered 705 cases of cancer patients. Patients suffering from the disease has a particular common symptom and based on those common symptom data were collected. Parameters considered for the research as symptoms are radius, perimeter, texture, smoothness, area, compactness, concavity, concave points, symmetry etc. Pre-processing of Data: Data collected was inconsistent. It required pre-processing of data to make it consistent as per requirements to apply analytical tools. Researchers filled some of blank data by taking average and filling approximated values. Total number of entries of these data is less than 2%. For the better visualization and understanding different conflicts in the data, researcher tool means of all data and saved these as radius_mean, texture_mean, perimeter_mean, area_mean etc.
448
K. K. Singh et al.
Data sheet – 01.
Finding Conflicts and Story Building: Researcher tool help of various statistical charts and show different angles of the story in data. Data taken from source and supplied to the tool to draw different charts i.e. Box plot distribution, Bar plot, sieve siagra, mosaic display, line plot, heat map, radviz, linear projection, scattered diagram and many more. Tool adopted is anaconda orange for visualization modelling. Line chart of Fig. 1 shows the deterioration of health is more than rate of growth of in the diseases when maximum worst case area is 4254, parameter is 24, 457 and compactness mean is 30. In Fig. 1, 2, 3, researchers used different line and bar charts to explain about data. Variation in the data is visible in each of the graphs and values are varying accordingly. Worst part of the area (in which cancer is visible) has been deteriorating continuously. Because of that, range of affected area is wide. Fractal dimension of the area is also fluctuating continuously (Fig. 2 and 3). Mean value of affected area is tending towards upper-side of the graph. Finding radius of the area is required exponential function smoothing so that it can cover maximum values as per affected area but variation in the data restrict to find exact radius. Variation in the area and its fractality shows its actual nature. Each
Fig. 1. Line graph
Breast Cancer Detection and Prediction Based on Conflicts
449
data clusters and stacked of the data is being visible in line and stacked column chart in Fig. 3.
Fig. 2. Bar chart
Fig. 3. Line and stacked column chart
Figures 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 and 14, researcher tried to explore different conflicts in the data and its implications on the breast cancer prediction in Indian context. In Fig. 4, independent variable is smoothness mean and dependent variable is concavity of the area. At smoothness mean value 0.1 and 0.12, concavity values are relatively smooth but at value 0.11, concavity values are sharply decreasing. In Fig. 5, independent variable is smoothness mean and dependent variable is concavity mean, worst and diagnostic of the area. At smoothness mean value 0.1 and 0.12, concavity values are relatively smooth but in between these, concavity values, worst and diagram are sharply decreasing. Value of diagnosis value of affected area is fluctuating up and down. From Figs. 6, 7 and 8, Conflicts in the data in visible between 0.10 to 0.12 on x-axis. In Fig. 7, classifications of impact on the basis of gene have been plotted and while classifying data, conflict in the same area where other graphs are showing. When one line is going up, others are show opposite behaviour.
450
K. K. Singh et al.
Fig. 4. Line chart after smoothness
Fig. 5. Smoothness meansvrs concavity
Fig. 6. Smoothness meansVrs count
Breast Cancer Detection and Prediction Based on Conflicts
451
Fig. 7. Gene classifiers and impact
On cellularity 742.37 and replace free status (months) 1109.58 there are high chance of larger tumour size even bigger than 230 and its stage is 13.
Fig. 8. Gene classification and first conflict
From Figs. 9, 10, 11, 12 and 13, researchers found various conflicts in the data. Genes are the most important factors in cancer prediction. Classification of data based on genes are by using different graphs are providing great results. Precedence of age and genes provides great results in finding conflicts in the data. Researcher build several stories based upon combining effects of age and genes. Results of diagnosis based on age is visible in Fig. 12 and 13. Most vulnerable age of cancer is between 36 to 40 years, where in detection of cancer if between 42 to 48 years. Increasing and decreasing pattern of the of detection at various age based upon genes is visible in the Fig. 12 and 13. Effect of menopausal state according to the age is visible in the figure below. Below heat map shows menopausal state in increasing order of age and changes of prediction in percentages. Cellularity of affected area affected based on menopausal state and age of the patient. Prediction chances of heterogeneous in its nature. Conflicts in the data are visible in the red. Various colour sheds ranging from green to red depicts condition of based on age, cellularity and menopausal condition of the patients.
452
K. K. Singh et al.
Fig. 9. Gene classification and second conflict
Fig. 10. Gene classification and third conflict
Fig. 11. Gene classification and Fourth conflict
Breast Cancer Detection and Prediction Based on Conflicts
Fig. 12. Age and its impact
Fig. 13. Pie chart showing next conflict based on age
Fig. 14. Heat map with conflict
453
454
K. K. Singh et al.
4 Conclusion There are various stories can be built based on different conflicts in the data as shown in the above figures. Story telling method is one of the most effective method to find out breast cancer and predict it in its early stages. Cases of breast cancer is more in the age between 35 to 50 and cases are based on many factors e.g. age of patient, gene, menopausal state and many more. Fractality in the affected area can initial signs for understanding gravity of the cancer stage. Cellularity and concavity of the area is vital in the prediction of the breast cancer. In this research, researcher found that average age of detection of breast cancer is 45 to 49 in most of the cases and data sets. But by seeing different conflicts in the data as well as patterns of classification, we can able to predict and detect breast cancer much earlier based on visual analytics and storytelling method. Conflicts in the data shows variation in the usual patterns and this leads to the different problems in subject. Detection and prediction of diseases like breast cancer is most effect in this era of medical science.
References 1. Bazazeh, D., Shubair, R.: Comparative study of machine learning algorithms for breast cancer detection and diagnosis. In: 2016 5th International Conference on Electronic Devices, Systems and Applications (ICEDSA), pp. 1–4. IEEE (2016) 2. Chaurasia, V., Pal, S., Tiwari., B.B.: Prediction of benign and malignant breast cancer using data mining techniques. J. Algorithms Comput. Technol. 12(2), 119–126 (2018) 3. Globocan (2012). http://globocan.iarc.fr/Default.aspx. Accessed 28 Dec 2015 4. Kibis, E.Y.: Data analytics approaches for breast cancer survivability: comparison of data mining methods. In: IIE Annual Conference. Proceedings, pp. 591–596. Institute of Industrial and Systems Engineers (IISE) (2017) 5. Asri, H., Mousannif, H., Moatassime, H., Noel, T.: Using machine learning algorithms for breast cancer risk prediction and diagnosis. Procedia Comput. Sci. 83, 1064–1069 (2016) 6. Bharat, A., Pooja, N., Anishka Reddy, R.: Using machine learning algorithms for breast cancer risk prediction and diagnosis. In: 2018 3rd International Conference on Circuits, Control, Communication and Computing (I4C), pp. 1–4. IEEE (2018) 7. Sharma, S., Aggarwal, A., Choudhury, T.: Breast cancer detection using machine learning algorithms. In: 2018 International Conference on Computational Techniques, Electronics and Mechanical Systems (CTEMS), pp. 114–118. IEEE (2018) 8. Yue, W., Wang, Z., Chen, H., Payne, A., Liu, X.: Machine learning with applications in breast cancer diagnosis and prognosis. Designs 2(2), 13 (2018) 9. Alarabeyyat, A., Alhanahnah, M.: Breast cancer detection using k-nearest neighbor machine learning algorithm. In: 2016 9th International Conference on Developments in eSystems Engineering (DeSE), pp. 35–39. IEEE (2016) 10. Liu, Y.-Q., Wang, C., Zhang, L.: Decision tree based predictive models for breast cancer survivability on imbalanced data. In: 3rd International Conference on Bioinformatics and Biomedical Engineering, Beijing, China, 11–13 June 2009 (2009) 11. Chakraborty, A., Chatterjee, S., Majumder, K., Shaw, R.N., Ghosh, A.: A comparative study of myocardial infarction detection from ECG data using machine learning. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 257–267. Springer, Singapore (2022). https://doi.org/10.1007/978-98116-2164-2_21
Breast Cancer Detection and Prediction Based on Conflicts
455
12. Bellaachia, A., Guven, E.: Predicting Breast cancer survivability using data mining techniques, Department of computer science, George Washington University 13. Karabatak, M., Cevdet Ince, M.: Expert system with applications (2009). Elsevier 14. Esgiar, A.N., Naguib, R.N., Sharif, B.S., Bennett, M.K., Murray, A.: Microscopic image analysis for quantitative measurement and feature identification of normal and cancerous colonic mucosa. IEEE Trans. Inf. Technol. Biomed. 2(3), 197–203 (1998) 15. Karahaliou, A.N., et al.: Breast cancer diagnosis: analyzing texture of tissue surrounding microcalcifications. IEEE Trans. Inf. Technol. Biomed. 12(6), 731–738 (2008) 16. Crisóstomo, J., Matafome, P., Santos-Silva, D., et al.: Hyperresistinemia and metabolic dysregulation: a risky crosstalk in obese breast cancer. Endocrine 53(2), 433–442 (2016). https:// doi.org/10.1007/s12020-016-0893-x 17. Kupinski, M.A., Giger, M.L.: Automated seeded lesion segmentation on digital mammograms. IEEE Trans. Med. Imaging 17(4), 510–517 (1998) 18. Gadewadikar, J., et al.: Exploring Bayesian network for medical decision support in breast cancer detection. AJMCSR 3(10), 225–231 (2010) 19. Palimkar, P., Shaw, R.N., Ghosh, A.: Machine learning technique to prognosis diabetes disease: random forest classifier approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 219–244. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_19 20. Delen, D., Walker, G., Kadam, A.: Predicting breast cancer survivability: a comparison of three data mining methods. Artif. Intell. Med. 34, 113–127 (2005) 21. Bernal, J.L., Cummins, S., Gasparrini, A.: Interrupted time series regression for the evaluation of public health interventions: a tutorial. Int. J. Epidemiol. 46, 348–355 (2017) 22. Williams, T.G.S., Cubiella, J., Griffin, S.J.: Risk prediction models for colorectal cancer in people with symptoms: a systematic review. BMC Gastroenterol. 16, 63 (2016). https://doi. org/10.1186/s12876-016-0475-7 23. Pratiwi, P.S.: Development of intelligent breast cancer prediction using extreme learning machine in Java. Int. J. Comput. Commun. Instrum. Eng. 3 (2016)
A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter (2ph-IBDB2 C) for Battery Storage Applications K. L. Lijin, S. Sheik Mohammed(B) , and P. P. Muhammed Shanir TKM College of Engineering, Kollam, India [email protected], [email protected]
Abstract. The bi-directional converter is one of the important subsystems in a modern grid. A bi-directional converter contributes to system stability by allowing power to flow in both directions between the sources and storage system. This paper presents and discusses a two-phase interleaved bidirectional DC-DC buckboost converter (2ph-IBDB2C). The architecture of the proposed converter is analyzed in detail. The selected converter is simulated in MATLAB/Simulink, and the results are presented. The converter performance in boost and buck modes is then examined based on the control input. Keywords: Interleaving · DC-DC converters · Energy Storage · Bidirectional · Buck mode · Boost mode
1 Introduction The rapid increase in usage of conventional energy resources for fulfilling the energy requirements leads to environmental issues and depletion of resources at a much faster rate. The utilization of alternative energy resources could significantly reduce the environmental problems that arise due to the usage of conventional resources. Presently, electric power generation using solar energy has become one of the best technology due to its availability, clean and silent energy production without any emission. Microgrids are controlled small-scale power generation units designed to provide electric power for localized areas. Microgrids can function in one of two modes: (i) islanded mode or (ii) grid-connected mode. In islanded mode, microgrids operate as self-contained generating units capable of producing enough energy to meet load requirements. When a microgrid is connected to the main utility grid, the control mechanism allows power transfer in either direction between the microgrid and the main grid, which is dependent on load power requirements [1, 2]. Microgrid enables the integration of Distributed Energy Resources (DERs), which improves the regional electric grid’s operation and stability. Microgrids primarily generate electricity using renewable energy sources such as solar and wind [3–5]. The major concern associated with renewable energy resources is their varying nature due to their dependency on environmental conditions. For example, solar © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 456–468, 2022. https://doi.org/10.1007/978-981-19-1677-9_41
A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter
457
PV power generation depends on irradiance and temperature. This demands the integration of Energy Storage Systems (ESS) in microgrids. Due to their cost-effectiveness and technological benefits, lead-acid batteries are commonly employed as ESS [6, 7]. [8] introduced a twin battery storage bank arrangement for renewable energy microgrids. The battery charges when the power generated by intermittent energy systems is more than the requirements of the load. This stored energy is consumed during the absence of generation by renewable energy sources or whenever required. The storage networks are linked to the grid bus through a bidirectional DC-DC converter since the power flow of battery storage systems is in either way. Electric vehicle energy storage systems [9, 10], power production systems using fuel cells and PV panels [11, 12], uninterrupted power supplies, and battery chargers [13] are all frequent uses for the bidirectional DC-DC converter Integrating battery storage system with something like a DC-DC bidirectional converter in RES power systems improves the power supply’s reliability and stability. Bidirectional DC-DC converters are available in two topologies, non - isolated and isolated. Non - isolated topologies are mainly preferred for lightweight, more diminutive size, high power applications such as spacecraft due to their simple fabrication with fewer components and higher efficiency [14]. In isolated topology, a transformer is used for electric isolation between the voltage sources, increasing the converters’ size, cost, and weight [15, 16]. The duty cycle of switching pulses delivered to converter switches determines a bidirectional DC-DC converter to operate in either boost or a buck mode. Figure 1 illustrates a simple DC grid system’s basic configuration consisting of a solar PV array, a battery energy storage system with a bidirectional DC-DC converter, and a DC load. Whenever the energy generated by the solar PV system exceeds the load power needed, the extra energy is stored in the battery and used whenever it is required. When the battery is being charged, the bidirectional DC-DC converter is in buck mode, and when the battery is being discharged, it is in boost mode. DC Grid PV Panel Bidirectional DC - DC Converter Load Battery Fig. 1. Basic configuration of power generation using PV panel
Conventional DC-DC converter introduces high ripple in output voltage and input current. These problems can be overcome by using an interleaved DC-DC converter. The parallel connection of numerous converters allows for a current-sharing mechanism between the units, reducing input current and output voltage ripples. Higher efficiency and improved transient response can be achieved via interleaving [17]. This paper presents a two-phase interleaved bidirectional Buck-Boost converter. MATLAB/Simulink is used to build and simulate a 1 kW, 60/96 V Two-phase Interleaved
458
K. L. Lijin et al.
Buck-Boost bidirectional converter with a switching frequency of 20 kHz. The paper is organized into five sections: Sect. 2 describes the working of 2ph-IBDB2 C, Sect. 3 contains the design of 2ph-IBDB2 C. Section 4 includes the simulation and analysis of converter under different duty cycles, and Sect. 5 briefs the conclusion.
2 Two-Phase Interleaved Bidirectional Buck-Boost Converter The 2ph-IBDB2 Cis shown in Fig. 2. The diodes in an interleaved boost converter are replaced by controlled switches that enable bidirectional operation [18]. Two inductors, L 1 and L 2 , and two capacitors, C buck and C boost , make up the circuit. V buck and V boost , respectively, represent the buck and boost voltages in the circuit. S 1 , S 2 , S 3, and S 4 are the four switches used to enable the bidirectional and interleaved operation of the converter.
L1
S3
S4
L2 Vbuck
Cbuck
Vboost Cboost S1
S2
Fig. 2. Two-phase interleaved bidirectional buck-boost boost converter
Switches S 1 and S 2 are triggered during boost mode operation, whereas S 3 and S 4 remain in the OFF state. On the other hand, Switches S 3 and S 4 are actuated for buck operation. S 1 and S 2 will be in the OFF state during this mode of operation. The following sub-sections explain two modes of operation. 2.1 Boost Mode Operation In this mode, switches S 1 and S 2 are triggered with switching pulses having the desired frequency and 1800 phase, enabling interleaved operation. For the analysis of the circuit, ideal components are used [19–21].
A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter
459
Boost mode operation of the converter has four stages of operation as shown in Fig. 3. In the first stage, switches S 1 and S 2 are triggered, and current through the inductor L 1 and L 2 will increase linearly. During this period, the capacitor provides energy to the load. In this case, the load is V boost = V c . S1 is shut off in the second step while S 2 remains in the ON state, resulting in a linear increase in current through L 2 . The energy stored in inductor L 1 forward biases the diode D3 , allowing one part of the input current to flow to the output stage via S 1 and D3 . The inductor L 1 ’s energy is passed to the capacitor C boost and the load. The third step is similar to the first. In the fourth step, S 2 is shut off while S 1 stays in the ON state, resulting in a linear increase in current through L 1 . The energy stored in inductor L 2 forward biases the diode D4 , and the energy stored in inductor L 2 is also transferred to the capacitor C boost and the load. During this operation input voltage of the converter (V buck ) is stepped up to 96 V (V boost ). The converter switching pulses and variation in inductor during switching are shown in Fig. 4.
L2
L2 Vboost
L1 Vbuck
Cboost
Cbuck
Cboost Vbuck
Cbuck
(b) Stage 2
(a) Stage 1
L2
L2 Vboost
L1 Vbuck
(c) Stage 3
Vboost
L1 Cboost
Cboost Cbuck
Vboost
L1
Vbuck
Cbuck
(d) Stage 4
Fig. 3. Different stages of operation of converter in boost mode
2.2 Buck Mode Operation Switches S3 and S4 are operated in this mode. When the switches are triggered, the inductor current increases linearly; when the switches are turned off, the energy stored in the inductor is transmitted to the load. The converter’s input voltage (V boost ) is stepped down to 60 V (V buck ) during this mode of operation.
460
K. L. Lijin et al.
S
T DT
(1-D)T
S
t
S , S
t
I
t
I I
I
t I I
t
Fig. 4. Switching pulses and inductor current waveforms in boost mode operation
3 Design of Two-Phase Interleaved Bidirectional Buck-Boost Converter A 1 KW, 60/96 V 2ph-IBDB2 C with a switching frequency of 20 kHz is considered for the study. The detailed design procedure of the selected converter is presented in [22, 23].
A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter
461
a) Duty cycle (D): D(Boost mode) = 1 − D(Boost mode) =
V1 60 =1− = 0.375 V2 96 60 V2 = 0.625 = V1 96
b) Inductors: The inductor ripple current is taken as 10% of the output current D(1 − D)V2 fs IL 0.375(1 − 0.375) × 96 = 20 × 103 × 0.1 × 16.67 = 674.86 µH
L1 = L2 =
c) filter capacitor: The capacitor voltage ripple is taken as 2% of the output voltage I0 D fs V0 10.42 × 0.375 = 20 × 103 × 0.02 × 96 = 101 µF
Cboost =
d) Input capacitor: Lowest value of the inductor to maintain converter in continuous mode of operation is calculated as follows, D(1 − D)2 R cos−1 θ 2f 0.375 × (1 − 0.375)2 × 9.21 = 2 × 20 × 103 = 33.73 µH
Lmin =
Cbuck =
Vin (1 − D) 8Lmin f 2 VC
48(1 − 0.375) 8 × 33.73 × 10−6 × (20 × 103 )2 × 0.02 × 96 = 144.76 µF
=
462
K. L. Lijin et al.
4 Simulation and Analysis MATLAB Simulink is used to create and simulate the 2ph-IBDB2 C. The important parameters and components values are listed in Table 1. Table 1. Design parameters and value of 2ph-IBDB Sl No
Parameters
Value
1
Buck Voltage (Vbuck )
60 V
2
Boost Voltage (Vboost)
96 V
3
Switching frequency (fs )
20 kHz
4
Rated Power (P)
5
Duty Cycle (D)
1 kW Boost mode
0.325
Buck mode
0.625
6
ΔIL
10% of Iin
7
ΔVc
2% of V 0
8
Inductors (L1 = L2 = L)
674.86 µH
9
Capacitor (Cboost )
101 µF
10
Capacitor (Cbuck )
144.76 µF
Figure 5 illustrates the simulation model of a two-phase interleaved bidirectional buck-boost converter built-in MATLAB/Simulink.
Fig. 5. Simulation model of a two-phase interleaved bidirectional buck-boost converter
This section evaluates the circuit under various conditions and presents and discusses the simulation findings. The converter’s duty cycle is 0.375 from 0 to 1 s. Then from 1 to
A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter
463
2 s, the converter’s duty cycle is increased to 0.625. Again, from 2 to 3 s, the converter’s duty cycle is decreased to 0.375. The duty cycle applied to the converter in response to the control signal is seen in Fig. 6.
Fig. 6. Duty cycle of the converter
Figure 7(a) demonstrates the switching pulses applied to the converter switches for the selected duty pattern, while Fig. 7(b) shows a segment of the switching pulses. The figure indicates that at a time converter can function in either buck or boost mode. Figure 8(a) and 8(b) show the output current of the converter. From 0 to 1-s positive current in Fig. 8(a) and the negative current in Fig. 8(b) indicate the energy transfer from V buck to V boost . During this time converter operates in boost mode. From 1 s to 2 s converter operation changes to buck mode results in reversal of the current direction. During this time, energy is transferred V boost to V buck . Figure 9(a) and 9(b) show the voltage across the energy storage sources during different modes of operation of the converter. Figure 10(a) and 10 (b) show the State of Charge (SOC) of each battery in a different mode of operations. From the above output current and voltage waveforms, it is clear that voltage across the battery increases, and output current remains negative during charging and hence output power is negative. During discharging of batteries, the voltage across the battery reduces, and the output current becomes positive, so the power output is positive. Figure 11(a) and 11(b) show each battery’s power output during charging and discharging operation. The converter’s duty cycle changes enable buck or boost mode operation, which results in bidirectional energy flow between the batteries. The input current and output voltage ripple is relatively small, which increases the converter’s efficiency.
464
K. L. Lijin et al.
Fig. 7. (a) Switching pulses for the buck and boost switches. (b) Portion of Switching pulses in switches at various duty cycles
A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter
Fig. 8. (a) Current at buck side (Vbuck ) (b) Current at boost side (Vboost )
Fig. 9. (a) Voltage at buck side (Vbuck ) (b) Voltage at boost side (Vboost )
465
466
K. L. Lijin et al.
Fig. 10. (a) SOC% of the source (Vbuck ) (b) SOC% of the source (Vboost )
Fig. 11. (a) Power output of the source (Vbuck ) (b) Power output of the source (Vboost )
A Non-isolated Two-Phase Interleaved Bidirectional Buck-Boost Converter
467
5 Conclusion In this study, a one-kilowatt, 60/96-V, 2ph-IBDB2 C for integrating a battery storage system into a DC grid is described and analyzed in detail. The converter’s design and its several modes of operation are covered in detail in this paper. The developed circuit is simulated using MATLAB/Simulink, and the results of the simulations are used to validate the converter’s performance under different conditions.
References 1. Ren, L., Zhang, P.: Generalized microgrid power flow. IEEE Trans. Smart Grid 9(4), 3911– 3913 (2018). https://doi.org/10.1109/TSG.2018.2813080 2. Xin, H., Zhang, L., Wang, Z., Gan, D., Wong, K.P.: Control of island AC microgrids using a fully distributed approach. IEEE Trans. Smart Grid 6(2), 943–945 (2015). https://doi.org/10. 1109/TSG.2014.2378694 3. Nejabatkhah, F., Li, Y.W.: Overview of power management strategies of hybrid AC/DC microgrid. IEEE Trans. Power Electron. 30(12), 7072–7089 (2015). https://doi.org/10.1109/TPEL. 2014.2384999 4. Nasser, N., Fazeli, M.: Buffered-microgrid structure for future power networks; a seamless microgrid control. IEEE Trans. Smart Grid 12(1), 131–140 (2021). https://doi.org/10.1109/ TSG.2020.3015573 5. Sheik Mohammed, S., Krishnendu, J.M.: Energy management control of DC microgrid with electric vehicle and hybrid energy storage system. In: 2019 2nd International Conference on Intelligent Computing, Instrumentation and Control Technologies, ICICICT 2019, no. September, pp. 1360–1364 (2019). https://doi.org/10.1109/ICICICT46008.2019.8993171 6. Carter, R., Cruden, A., Hall, P.J., Zaher, A.S.: An improved lead-acid battery pack model for use in power simulations of electric vehicles. IEEE Trans. Energy Convers. 27(1), 21–28 (2012). https://doi.org/10.1109/TEC.2011.2170574 7. Vasak, M., Kujundzic, G.: A battery management system for efficient adherence to energy exchange commands under longevity constraints. IEEE Trans. Ind. Appl. 54(4), 3019–3033 (2018). https://doi.org/10.1109/TIA.2018.2812138 8. Neto, P.B.L., Saavedra, O.R., De Souza Ribeiro, L.A.: A dual-battery storage bank configuration for isolated microgrids based on renewable sources. IEEE Trans. Sustain. Energy, 9(4), 1618–1626 (2018) https://doi.org/10.1109/TSTE.2018.2800689 9. Ahmadi, F., et al.: Design and implementation of a new transformer less bidirectional DC-DC converter with wide conversion ratios. IEEE Trans. Power Electron. 3(4), 3493–3503 (2012). https://doi.org/10.1109/TPEL.2020.3045986 10. Jm, K., Sheik Mohammed, S., Ahamed, T.P.I., Shafeeque, M.: Design and simulation of stand-alone DC microgrid with energy storage system. In: IEEE International Conference on Intelligent Techniques in Control, Optimization and Signal Processing, INCOS 2019 (2019). https://doi.org/10.1109/INCOS45849.2019.8951384 11. Hintz, A., Prasanna, U.R., Rajashekara, K.: Novel modular multiple-input bidirectional DCDC power converter (MIPC) for HEV/FCV application. IEEE Trans. Ind. Electron. 62(5), 3163–3172 (2015). https://doi.org/10.1109/TIE.2014.2371778 12. Askarian, I., Pahlevani, M., Knight, A.M.: Three-port bidirectional DC/DC converter for DC nano grids. IEEE Trans. Power Electron. 36(7), 8000–8011 (2021). https://doi.org/10.1109/ TPEL.2020.3046453
468
K. L. Lijin et al.
13. Ahmadi, F., Adib, E., Azari, M.: Soft switching bidirectional converter for reflex charger with minimum switches. IEEE Trans. Ind. Electron. 67(10), 8355–8362 (2020). https://doi.org/10. 1109/TIE.2019.2947813 14. Tomar, P.S., Sharma, A.K., Hada, K.: Energy storage in DC microgrid system using nonisolated bidirectional soft-switching DC/DC converter. In: 2017 6th International Conference on Computer Applications In Electrical Engineering-Recent Advances, CERA 2017, vol. 2018-Janua, pp. 439–444 (2018). https://doi.org/10.1109/CERA.2017.8343370 15. Zeng, J., Qiao, W., Qu, L.: An isolated three-port bidirectional DC-DC converter for photovoltaic systems with energy storage. IEEE Trans. Ind. Appl. 51(4), 3493–3503 (2015). https:// doi.org/10.1109/TIA.2015.2399613 16. Wu, Y.E., Ke, Y.T.: A novel bidirectional isolated DC-DC converter with high voltage gain and wide input voltage. IEEE Trans. Power Electron. 36(7), 7973–7985 (2021). https://doi. org/10.1109/TPEL.2020.3045986 17. Belkhier, Y., Achour, A., Shaw, R.N., Sahraoui, W., Ghosh, A.: Adaptive linear feedback energy-based back stepping and PID control strategy for PMSG driven by a grid-connected wind turbine. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 177–189. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0749-3_13 18. Huang, B.J.: Interleaved voltage-doubler boost converter for power factor correction. In: 2018 International Power Electronics Conference IPEC-Niigata - ECCE Asia 2018, pp. 3528–3532 (2018). https://doi.org/10.23919/IPEC.2018.8507419 19. Wang, Y.F., Xue, L.K., Wang, C.S., Wang, P., Li, W.: Interleaved high-conversion-ratio bidirectional DC-DC converter for distributed energy-storage systems-circuit generation, analysis, and design. IEEE Trans. Power Electron. 31(8), 5547–5561 (2016). https://doi.org/10. 1109/TPEL.2015.2496274 20. Metin, N.A., Boyar, A., Kabalci, E.: Design and analysis of bi-directional DC-DC driver for electric vehicles. In: Proceedings of 2019 IEEE 1st Global Power, Energy and Communication Conference GPECOM 2019, pp. 227–232 (2019). https://doi.org/10.1109/GPECOM.2019. 8778532 21. de Melo, R.R., Tofoli, F.L., Daher, S., Antunes, F.L.M.: Interleaved bidirectional DC–DC converter for electric vehicle applications based on multiple energy storage devices. Electr. Eng. 102(4), 2011–2023 (2020). https://doi.org/10.1007/s00202-020-01009-3 22. Mohammed, S.S., Syji, B.M.: Energy management of a standalone low voltage DC microgrid using FPGA based controller. J. Green Eng. 10(5), 1984–2005 (2020) 23. Thomas, S.M., Mohammed, S.S.: Solar-powered EV charging station with G2V and V2G charging configuration. J. Green Eng. 10(4), 1704–1731 (2020)
A Comparative Study of Concentrated Photovoltaic Thermal-Thermoelectric Generator Hybrid Systems Abhishek Tiwari1,2 and Shruti Aggarwal1,2(B) 1 University School of Basic and Applied Sciences, Guru Gobind Singh Indraprastha
University, Sector-16C, Dwarka, Delhi 110078, India [email protected] 2 School of Electrical, Electronics and Communication, Galgotias University, Greater noida, Uttar Pradesh 203201, India
Abstract. In this study, a performance comparison has been conducted between an opaque photovoltaic thermal module integrated with a concentrator and a thermoelectric generator (COPVT-TEG) system (Case 1) and a semitransparent photovoltaic thermal module integrated with a concentrator and a thermoelectric generator (CSPVT-TEG) system (Case 2).The thermodynamic model for both the hybrid systems have been developed and simulated using MATLAB for clear day in the month of May for New Delhi climate. On comparison, the maximum power output of COPVT-TEG hybrid system exceeded the maximum power output of CSPVT-TEG hybrid system by4.19 W. The results have also illustrated that the maximum electrical efficiency of COPVT-TEG hybrid exceeds the maximum electrical efficiency of CSPVT-TEG hybrid system by 0.32% . Keywords: PV Module · Thermoelectric generator · Hybrid system
1 Introduction In the present communication, a performance based comparative study between concentrated opaque photovoltaic thermal-thermoelectric generator (COPVT-TEG) hybrid system and concentrated semitransparent photovoltaic thermal-thermoelectric generator (CSPVT-TEG) hybrid system has been conducted. The parameters used for the comparative study are the solar cell temperature, solar cell efficiency, PV module power output, TEG temperature difference, TEG power, TEG electrical efficiency, hybrid system power and electrical efficiency. The simulation of the theoretical model was conducted using MATLAB for a clear day condition in May for New Delhi climate. Kossyvakis et al. [1] conducted a theoretical and an experimental study to analyse the performance of a tandem photovoltaic-thermal electric hybrid system. Lamba and Kaushik [2] conducted numerical analysis and optimization of PC-TEG hybrid system. Tiwari and Aggarwal [3] conducted a theoretical study of PV thermal-TEG hybrid system in a concentrator with respect to its performance and energy analysis. Li et al. [4] conducted an analytical © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 469–481, 2022. https://doi.org/10.1007/978-981-19-1677-9_42
470
A. Tiwari and S. Aggarwal
study on solar concentrating thermoelectric generator employing micro-channel heat pipe array. Zhang and Xuan [5] conducted a study on the behaviour of solar thermoelectric hybrid system under fluctuating radiation. Liu et al. [6] conducted a simulation-based study of a photovoltaic thermal-compound thermoelectric ventilator system.
2 Description and Working Principle
Fig. 1. Schematic diagram of COPVT-TEG hybrid system (Case 1) [2]
A schematic diagram of COPVT-TEG hybrid system and CSPVT-TEG hybrid system is shown in Fig. 1 and Fig. 2 respectively. The construction of both the hybrid systems is similar with the only difference being the use of an opaque PV module for COPVT-TEG hybrid system and a semi-transparent PV module for CSPVT-TEG hybrid system. Both the hybrid systems, from top to bottom consist of a PV module, a TEG module attached at the back of the PV module. The PV module and the TEG module are thermally attached but electrically isolated from each other, i.e., there is no electrical connection between PV module and TEG module. The PV module acts as heat source for the TEG. The solar radiation CI(t), where C is the concentration ratio, falls on the PV panel. In case of COPVT-TEG hybrid system,
A Comparative Study of Concentrated Photovoltaic Thermal-Thermoelectric
471
Fig. 2. Schematic diagram of CSPVT-TEG hybrid system (Case 2) [3]
partial conversion of concentrated radiation into electrical energy through photovoltaic effect takes place, and the rest is converted into heat energy. A part of this heat energy is liberated from the PV module’s top to the ambient and rest is absorbed by TEG module at its hot end, attached with Tedlar of the opaque PV module. The TEG module converts a part of the absorbed heat energy into electrical power through See beck effect and the remaining heat energy is rejected as shown in Fig. 1 and Fig. 2. In case of CSPVTTEG hybrid system, a portion of concentrated solar radiation in converted into electrical energy and the remaining radiation incident on the non-packing area of is transmitted through the back glass to the TEG where it is absorbed. So, in case of COPVT-TEG hybrid system, the heat energy received or absorbed by the TEG module is only through conduction but in CSPVT-TEG hybrid system, the heat energy received by the TEG module is through both conduction and absorption of solar radiation. The remaining operation and working principle of both the hybrid systems are same. Th and Tc are the TEG hot end and cold end temperature (K) respectively and Ta is the ambient temperature (K).
472
A. Tiwari and S. Aggarwal
3 Modelling 3.1 COPVT-TEG Hybrid System Energy Balance Equation (Case 1) The energy balance for an opaque PV module is given as [8]: CI (t)Apv τg [αsc βc + αt (1 − βc )] = Uta Tsc,op − Ta Apv + Ubh Tsc,op − Th Apv + ηsc,op τg βc Apv CI (t)
(1)
The expressions for Uta-op , Ubh-op and UL-op are given below [8]: −1 Lg Kg + (1 + ho )
(2)
Ubh−op = [(LSC /KSC ) + (Lt /Kt )]−1
(3)
UL−op = Uta + Ubh
(4)
ho = 5.7 + 3.8v
(5)
Uta−op =
Here, v is the wind velocity in m/s.αsc , αt , βc , τg are the solar cell absorptivity, Tedlar absorptivity, opaque PV module packing factor and glass transmittivity respectively. The LHS of Eq. (1) is the solar energy absorption rate of the opaque PV module. The first term, second term and third term in the RHS of Eq. (1) represent the heat energy transfer rate from opaque PV module to the ambient, TEG hot end heat absorption rate and the electrical power of opaque PV module. Uta-op and Ubh-op represent the overall heat transfer coefficient from PV module to the ambient and from the PV module to the TEG module hot end respectively (in W/m2 K).The mathematical expression for opaque PV module temperature Tsc,op is given as: Tsc,op = (B1 + B2 + B3 + B4 − B5 − B6 )/(UL−op − B7 )
(6)
The expression for B1 , B2 , B3, B4 , B5 , B6 and B7 is part of the appendix. The mathematical expression for Tsc,op dependent PV cell efficiency (ηsc,op ) is given as follows [8]: ηsc,op = ηo [1 − βo Tsc,op − To ] (7) ηo , βo and To are the solar cell efficiency at STC, solar cell temperature coefficient and reference temperature of 298K respectively. The expression for opaque PV Module efficiency (ηpv,op )and its power output (Ppv,op ) is given below [8]: ηpv,op = ηsc,op τg βc
(8)
Ppv,op = ηpv,op CI (t)Apv
(9)
The energy balance equations for the TEG module are as given below [8]:
A Comparative Study of Concentrated Photovoltaic Thermal-Thermoelectric
473
Hot end 2 Q˙h = steg Th Iteg + Kteg Tteg − 0.5 Iteg Rteg − 0.5μteg Iteg Tteg
(10)
Cold end Q˙c = steg Tc Iteg + Kteg Tteg + 0.5(Iteg )2 Rteg + 0.5μteg Iteg Tteg
(11)
In Eq. (10) and Eq. (11), the first term of the RHS is Peltier power, the second term is the Fourier heat conduction, the third term is the Joule heating due to electric current Iteg and the fourth term is the Thomson’s heating due to both temperature difference Tteg and electrical current Iteg . Kteg is the thermal conductance of TEG. steg is the See beck coefficient of TEG. μteg is the Thomson coefficient of TEG (in V/K). The expression for power of TEG, Pteg(COPVT-TEG) and its electrical efficiency, ηteg(COPVT-TEG) is given below [8]: 2 ˙h −Q ˙c Pteg(COPVT −TEG) = [ steg − μteg Tteg ] /4Rteg = Q
(12)
ηteg(COPVT −TEG) = Pteg(COPVT −TEG) /Q˙h
(13)
The energy balance for hybrid system are as follows [8]: Q˙h = Ubh Apv (Tsc,op − Th )
(14)
Q˙c = Uc Apv (Tc − Ta )
(15)
The expression for power of COPVT-TEG hybrid system PCOPVT-TEG and its electrical efficiency, ηCOPVT-TEG is given below [8]: PCOPVT −TEG = Ppv,op + Pteg(COPVT −TEG)
(16)
ηCOPVT −TEG = PCOPVT −TEG /I (t)Apv
(17)
3.2 CSPVT-TEG Hybrid System Energy Balance (Case 2) The energy balance for a semitransparent PV module is given below [9]: CI (t)αc βc Apv τg = Uta Tsc,st − Ta Apv + Ubh Tsc,st − Th Apv + ηsc,st τg βc Apv CI (t) (18) The expressions for Uta-st , Ubh-st and UL-st are given below: Uta−st =
−1 Lg Kg + (1 + ho )
−1 Ubh−st = [(LSC /KSC ) + Lg /Kg ]
(19) (20)
474
A. Tiwari and S. Aggarwal
UL−st = Uta + Ubh
(21)
ho = 5.7 + 3.8v
(22)
The mathematical expression for semitransparent PV module temperature Tsc,st is given below: Tsc,st = (B8 − B9 − B10 + B11 + B12 )/(UL−st − B13 )
(23)
The expression for B8 , B9 , B10, B11 , B12 andB13 is part of the Appendix. The mathematical expression for Tsc,st dependent semitransparent solar cell efficiency (ηsc,st ) is given as follows [9]: ηsc,st = ηo [1 − βo Tsc,st − To ] (24) The expression for semitransparent PV Module electrical efficiency (ηpv,st ) and its power output (Ppv,st ) is given below [9]:ηpv,st = ηsc,st τg βc
(25)
Ppv,st = ηpv,st CI (t)Apv
(26)
The energy balance for the TEG module’s hot and cold end for CSPVT-TEG hybrid system are same as those for COPVT-TEG hybrid system. The energy balance for CSPVT-TEGhybrid system is given below [9]: Q˙h = Ubh Apv Tsc,op − Th + τg 2 αteg (1 − βpv )Apv I (t) (27) Q˙c = Uc Apv (Tc − Ta )
(28)
The expression for power of CSPVT-TEG hybrid system, PCSPVT-TEG and its electrical efficiency ηCSPVT-TEG is given below: PCSPVT −TEG = Ppv,st + Pteg(CSPVT −TEG)
(29)
ηCSPVT −TEG = PCSPVT −TEG /I (t)Apv
(30)
4 Results and Discussion The data for solar radiation I(t) and ambient temperature Ta for clear day in the month of May for New Delhi climate has been taken from the Indian Meteorological Department, Pune [9]. The value of other relevant design parameters is given in Table 1. The hourly variation of Tsc,op , Tsc,st , ηsc,op and ηsc,st is shown in Fig. 3. Since electrical efficiency has an inverse relationship with solar cell temperature, the nature of hourly variation of solar cell temperature and electrical efficiency are opposite of each other. The hourly
A Comparative Study of Concentrated Photovoltaic Thermal-Thermoelectric
475
variation of Tsc for opaque and semi-transparent PV module is similar. It first increases, then reaches to maximum value and then starts decreasing. This occurs because as I(t) increases, the incident energy on PV module increases which increases Tsc and when I(t) decreases, Tsc decreases. In Fig. 3, the values of Tsc,st is always greater than the values attained by Tsc,op throughout the observation period. This is because for an opaque PV module, the thermal conductivity of tedlar is lower than that of back glass of semitransparent PV module which results in slow heating of tedlar compared to back glass of semi-transparent module, thereby leading to solar cells attaining lower temperature in opaque PV module. Table 1. Design Parameters for hybrid system (for both cases) Parameter
Value
Lp
10 (−2) m [7]
Ap
1.14 cm2 /couple [7]
ηo
0.13 [8]
Lpv
0.0003 m [8]
Lt
0.000175 m [8]
βo
0.005K−1 [8]
αt
0.5 [8]
τg
0.95 [8]
Rteg
2.88 × 10(−3) /couple [7]
steg
425 × 10(−6) V/K/couple [7]
Kteg
0.02777 W/K/couple [7]
μteg
−0.093 mV/K[7]
Ln
10 (−2) m [7]
An
1 cm2 /couple [7]
βc
0.85 [8]
To
298 K [8]
αc
0.9 [8]
v
1m/s [8]
Kpv
148 W/mK [8]
Kt
0.2 W/mK [8]
ho
9.5W/m2 [8]
Uc APV
20W/K [8]
Lg
0.003m [8]
Kg
1.1W/mK [8]
Apv
0.6324 m2 [8]
476
A. Tiwari and S. Aggarwal
The maximum and minimum value of Tsc,st are 457.04K at 1 pm and 365.69K at 8 am respectively whereas the maximum and minimum value of Tsc,op are 444 K at 1 pm and 360.43 K at 8 am respectively. The maximum and minimum value of ηsc,st are 8.6% at 8 am and 2.66% at 1 pm respectively whereas the maximum and minimum value of ηsc,op are 8.94% at 8 am and 3.51% at 1 pm respectively. Tsc,op Tsc,st
500
10
ηsc,op ηsc,st
Tsc (Kelvin)
460
6
440 4
420 400
ηsc (%)
8
480
2
380 0
360 8
9
10
11
12
13
14
15
16
17
Time (hours) Fig. 3. Hourly variation of Tsc,op, Tsc,st, ηsc,op and ηsc,st
The hourly variation of Ppv,op and Ppv,st is shown in Fig. 4. The values attained by Ppv,op is greater than the values attained by Ppv,st throughout the observation period. This is because of higher electrical efficiency of opaque PV module over semi transparent PV module. In Fig. 4, it is also observed that power obtained from both the opaque and semi transparent PV module shows a dip in the middle of day. This occurs because of higher value of Tsc and lower ηsc value. The maximum and minimum value of Ppv,op are 65.64 W at 9 am and 51.11 W at 1 pm respectively whereas the maximum and minimum value of Ppv,st are 60.76 W at9 am and 38.77W at 1 pm respectively. The hourly variation of Tteg for COPVT-TEG and CSPVT-TEG hybrid system is shown in Fig. 5. The values of Tteg for CSPVT-TEG hybrid system is always greater than that of COPVT-TEG hybrid system throughout the observation period. This is because of the absorption of solar radiation by TEG module passing through the nonpacking factor area of semi transparent PV module in case of CSPVT-TEG hybrid system. In Fig. 5, Tteg first increases, then reaches to maximum value and then starts decreasing. This occurs because I(t) initially increases which increases the heat input to the TEG module thereby increasing Tteg . After 1 pm, I(t) starts decreasing therefore, heat input to TEG also starts decreasing which in turn causes Tteg to decrease. The maximum and minimum value of Tteg for COPVT-TEG hybrid system are 107.67 K
A Comparative Study of Concentrated Photovoltaic Thermal-Thermoelectric
477
at 12 noon and 40.76 K at 5 pm respectively whereas the maximum and minimum value of Tteg for CSPVT-TEG hybrid system are 107.67K at 12 noon and 40.76K at 5 pm respectively. 70
Ppv,op Ppv,st
65
Ppv (W)
60 55 50 45 40 35
8
9
10
11
12
13
14
15
16
17
Time (hours) Fig. 4. Hourly variation of Ppv,op and Ppv,st
The hourly variation of Pteg(COPVT-TEG) ,Pteg(CSPVT-TEG) and ηteg(COPVT-TEG) , ηteg(CSPVT-TEG) is shown in Fig. 6 and Fig. 7 respectively. Pteg is directly proportional to N × (Tteg )2 [9, 10]. Since in this study N is kept constant at 127, Pteg depends only upon (Tteg )2 . So, the hourly variation of Pteg and the hourly variation of Tteg are similar. The value of Tteg as discussed before was higher for CSPVT-TEG hybrid system and lower for COPVT-TEG hybrid system. As a result, Pteg attains higher values for CSPVT-TEG hybrid system compared to COPVT-TEG hybrid system. The maximum and minimum value of Pteg(COPVT-TEG) are 34.29 W at 12 noon and 4.91 W at 5 pm respectively whereas the maximum and minimum value of Pteg(CSPVT-TEG) are 42.04W at 12 noon and 6 W at 5 pm respectively.
478
A. Tiwari and S. Aggarwal
ηteg is directly proportional to N × (Tteg )2 /Q˙h [9, 10]. Since in this study N is kept constant at 127, ηteg depends only upon (Tteg )2 /Q˙h . The value of ηteg for both the hybrid systems increases first, then attains a maximum value and then decreases.This result is in accordance with the result of Tiwari and Aggarwal [9]. The maximum and minimum value of ηteg(COPVT-TEG) are 5.73% at 12 noon and 2.3% at 5 pm respectively whereas the maximum and minimum value of ηteg(CSPVT-TEG) are 6.71% at 12 noon and 2.68% at 5 pm respectively. 140 COPVT-TEG CSPVT-TEG
ΔTteg (Kelvin)
120
100
80
60
40 8
9
10
11
12
13
14
15
16
17
Time (hours) Fig. 5. Hourly variation of Tteg for COPVT-TEG and CSPVT-TEG hybrid systems
The variation of PCOPVT-TEG , PCSPVT-TEG , ηCOPVT-TEG and ηCSPVT-TEG is shown in Fig. 8. The variation of power output for both the hybrid systems is similar to the variation of power output of their respective PV modules as PV modules generate more power compared to TEG module. It is observed that both power output and electrical efficiency of COPVT-TEG hybrid system are always greater than the power and electrical efficiency of CSPVT-TEG hybrid system throughout the observation period respectively. The highest and lowest value of PCOPVT-TEG obtained are 88.47 W at 11 am and 54.8 W at 5 pm respectively whereas the highest and the lowest value of PCSPVT-TEG obtained are84.28 W at11 am and 54.10W at 5 pm respectively. Similarly, the maximum and minimum value of ηCOPVT-TEG obtained are 23.93% at 8 am and 14.13% at 1 pm respectively whereas the maximum and minimum value of ηCSPVT-TEG obtained are 23.61% at 8 am and 13.35% at 1 pm respectively.
A Comparative Study of Concentrated Photovoltaic Thermal-Thermoelectric 45
COPVT-TEG CSPVT-TEG
40 35
Pteg (W)
30 25 20 15 10 5 8
9
10
11
12
13
14
15
16
17
Time (hours)
Fig. 6. Hourly variation of Pteg for COPVT-TEG and CSPVT-TEG hybrid systems
7
COPVT-TEG CSPVT-TEG
6
ηteg (%)
5
4
3
2
8
9
10
11
12
13
14
15
16
17
Time (hours)
Fig. 7. Hourly variation of ηteg for COPVT-TEG and CSPVT-TEG hybrid systems
479
480
A. Tiwari and S. Aggarwal 25.0
90
22.5 80 PHS(COPVT-TEG) PHS(CSPVT-TEG)
17.5
70
ηHS(COPVT-TEG) ηHS(CSPVT-TEG)
PHS (W)
ηHS (%)
20.0
15.0 60 12.5
10.0
8
9
10
11
12
13
14
15
16
17
50
Time (hours)
Fig. 8. Hourly variation of PHS and ηHS for COPVT-TEG and CSPVT-TEG hybrid system
5 Conclusions Based on present study, following are the conclusions: 1. The solar cell electrical efficiency is higher in COPVT-TEG hybrid system compared to CSPVT-TEG hybrid system, leading to higher power from opaque PV module when compared to power obtained semi-transparent PV module. 2. The temperature difference, electrical efficiency and power of the thermoelectric generator is higher in case of CSPVT-TEG hybrid system in comparison to COPVTTEG hybrid system. 3. The power and electrical efficiency of COPVT-TEG hybrid system is greater than the power and electrical efficiency of CSPVT-TEG hybrid system respectively.
Appendix B1 = αc τg βc CI(t), B2 = αt τg (1−βc )CI(t), B3 = Uta-op Ta, B4 = Ubh-op Th , B5 = ηo τg βc CI(t), B6 = ηo τg βc CI(t)βo To, B7 = ηo τg βc CI(t)βo , B8 = αc τg βc CI(t), B9 = τg βc CI(t)ηo , B10 = τg βc CI(t)ηo βo To , B11 = Ubh-st Th , B12 = Uta-st Ta , B13 = τg βc CI(t)ηo βo
References 1. Kossyvakis, D.N., Voutsinas, G.D., Hristforou, E.V.: Experimental analysis and performance evaluation of tandem photovoltaic-thermoelectric hybrid system. Energy Convers. Manag. 117, 490–500 (2016)
A Comparative Study of Concentrated Photovoltaic Thermal-Thermoelectric
481
2. Lamba, R., Kaushik, S.C.: Solar driven concentrated photovoltaic-thermoelectric hybrid system: numerical analysis and optimization. Energy Convers. Manag. 170, 34–49 (2018) 3. Tiwari, A., Aggarwal, S.: Thermal modelling, performance analysis and exergy study of a concentrated semi-transparent photovoltaic thermal-thermoelectric generator (CSPVT-TEG) hybrid power generation system. Int. J. Sustain. Energy 40(10), 947–976 (2021) 4. Shaw, R.N., et al.: Review and analysis of photovoltaic arrays with different configuration system in partial shadowing condition. Int. J. Adv. Sci. Technol. 29(9s), 2945–2956 (2020) 5. Zhang, J., Xuan, Y.: Performance improvement of a photovoltaic-thermoelectric hybrid system subjecting to fluctuant solar radiation. Renew. Energy 113, 1551–1558 (2017) 6. Liu, Z., Zhang, Y., Zhang, L., Luo, Y., Wu, J., Yin, Y., Hou, G.: Modelling and Simulation of a photovoltaic thermal-compound thermoelectric ventilator system. Appl. Energy 28, 1887– 1900 (2018) 7. Wu, C.: Analysis of waste-heat thermoelectric power generators. Appl. Therm. Eng. 16, 63–69 (1996) 8. Lamba, R., Kaushik, S.C.: Modelling and performance analysis of concentrated photovoltaic– thermoelectric hybrid power generation system. Energy Convers. Manag. 115, 288–298 (2016) 9. Aggarwal, S., Tiwari, G.N.: Energy and exergy analysis of hybrid microchannel photovoltaic thermal module. Solar Energy 85, 356–370 (2011) 10. Tiwari, A., Aggarwal, S.: The effect of concentration ratio and number of thermocouples on photovoltaic-thermoelectric hybrid power generation system. In: Bose, M., Modi, A. (eds.) Proceedings of the 7th International Conference on Advances in Energy Research. Springer Proceedings in Energy, pp. 1521–1532. Springer, Singapore (2021). https://doi.org/10.1007/ 978-981-15-5955-6_144
Real Time Exhaustion Detection by Image Classification Using Deep Convolution Neural Network Anjani gupta, Prashant Singh(B) , Dhyanendra Jain, Amit Kumar Pandey, Ashu Jain, and Gaurav Sharma Department of Information Technology, Dr. Akhilesh Das Gupta Institute of Technology & Management, New Delhi, India [email protected]
Abstract. Advances in data and verbal exchange era via the creation of the Internet of Things (IoT) have had a chief effect on clever car systems. Driver fatigue is an essential issue in automobile injuries. Therefore, automobile collisions concerning motive force fatigue and motive force incompetence have been observed with the aid of using experts. In this studies article, a worn-out software is advanced so that it will have a take a observe a person’s face and eye behavior. This model develops a surveillance gadget so that it will experiment people’s eyes for some seconds and locate fatigue with the aid of using linking more than one photos. This gadget works properly in clear light mild situations and whether or not the motive force’s system is used, together with a mirror, listening to aids, or cap. This fatigue detection gadget will notify you whilst fatigue is detected and bring a beep sound to get you closer. Fatigue can once in a while place you at risk, ensuing in life-threatening injuries. OpenCV become used to acquire photos on camera, which have been then located in our CNN binary bar, which decided whether or not people’s eyes have been open or closed. A neural convolution community is a sort of deep neural community this is very powerful in distinguishing photos. This method is meant to lessen the dangers to diverse industries, together with production sites, highways, and industries, as many injuries arise because of fatigue or fatigue. As a result, this software program will notify the consumer earlier than it’s miles too late. The tool can calculate the precise location of the attention and locate fatigue in 85.forty nine percentages of cases. Keywords: CNN · Fatigue recovery · PERCLOS · AVECLOS · EEG
1 Introduction Fatigue indicates excessive tiredness, low stamina, and a strong need for rest that interrupts the daily routine. People may be more prone to fatigue due to lifestyle improvements, eating habits, rest patterns, and other clinical causes. Tired employees are less prepared and may be involved in wrong thinking. Tired drivers are not in a good driving position, which can lead to accidents. Sleep is a major cause of disasters worldwide. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 482–491, 2022. https://doi.org/10.1007/978-981-19-1677-9_43
Real Time Exhaustion Detection by Image Classification
483
Most of the major disasters around the globe are evolved by exhaustion or lethargy. FDS is important for cars to avoid unknowingly passing traffic. Tired students are often unable to concentrate in class, leading to erroneous thinking. Many groups have been searching for ways to improve FDS. Using computer vision, this suggested program aims to detect the decline of the population and prevent complications from fatigue. The concept of a PC is a branch of software engineering that put importance on mimicking the human frame and assisting PCs to recognize and respond to visual and audio objects. PC vision has a variety of uses, one of which is the recognition of facial features. Often, there are several ways to distinguish fatigue, including behavioral and lifestyle indicators. This social metric includes blindfolds, eyeliner, yawning, and more. Physiological measurements include the use of nerves attached to the body to detect heartbeats, pulses, and so on. When physiological parameters are included, the cost of the framework increases. As a result, we have introduced an inexpensive, web-based camera system for dealing with the detection of weaknesses that can separate the eye and lips from the face and issue a warning when fatigue is detected. We are very interested in image processing and machine learning. The important objective of this framework is to make it available to use in Indian industries. [22]. Recognizing fatigue is a technology for ensuring safety that can help prevent road accidents and their adverse effects today. Many drivers should take a closer look. Be focussed on the road so that you can respond rapidly to unforeseen happenings. Exhaustion of drivers is a direct cause of many road accidents, and there is a need to intensify bad weather identification and warning systems. Neuropsychological conditions can significantly lessen quantity of car mishappenings caused by fatigue [1]. The objective of this study is to create a fatigue recovery app that will give you a few seconds to see people’s eyes closed or open. If you’re tired, this gadget will let you know. Fatigue puts you at a greater risk for serious injuries. [8]. Previously, activities that attempted to combat fatigue were performed using mitigation measures such as harcascade separators. These output-based methods allow faster detection and less computer compression, but are more sensitive to light conditions and image correction. Well-trained machine learning methods, which provide greater accuracy and strength in detecting fatigue, can be used to overcome this problem. The CNN-based simplification algorithm is recommended, with an average accuracy of 85.49 percent achieved in all subjects using the required training data sets [5].
2 Widely Usable System for Exhaustion Indication This software is only one example of a driver fatigue detection system. This system checks the track movement for tight turns, difficult steering and unexpected braking. drivers are attentive and drive well, 1 very tired driver should put an halt to driving and rest instantaneously) [2]. PERCLOS is a measure of irritability which is defined by the reader as the percentage of eyelids closed over time and indicates that closing the eyelids is fatal or slower than blinking. [3].Software related to the inspection and management of PERCLOS is used by some real-time pollution detection systems to detect the onset of fatigue. [3] EEG recording is a previously unknown method for processing electrical activity in the brain. The man was discovered by Berger in 1924 and has been transformed
484
A. gupta et al.
into cutting-edge technology for over 90 years [4]. Substantial reductions in the use, gravity and expenditure of EEG, as well as the competence to connect wireless systems with other digital systems, have created a path for the use of the technology in other areas such as entertainment, feedback, learning and memory aids. The computer vision system uses an invisible dash camera and two red light sources to detect and track the driver’s face. the time with eyes closed averaged over a minute. [5] The disinfection software has recently been modified to work on Android phones. To control the user’s eye movements, this method uses a handheld camera installed in the area of the cable control panel. The designers of the show chose to use the eyelid movement method. [6] Even when the external light source is weak, a powerful system can detect rapid head movements and facial expressions.
3 Proposed Solution 3.1 System Structural Design In this fatigue recuperation program, we’ve proposed a three-channel community version that consists of convolution layers, multi-layered layers, and 3 absolutely incorporated layers which have the very last layer of Softamx. The following is an outline of the community shape proposed for fatigue: 1) Convolutional Layout: This is the primary layer, with many gaining knowledge of filters that produce an activation map. Focusing on facial functions in place of pores and skin tone, we flip RGB pictures into gray scales. We use particular layers due to the fact we need to compromise among accuracy and calculation. 2) Max-Pooling Layer: This is an oblique approach of lowering the pattern via way of means of lowering the scale of the illustration location as a way to lessen the wide variety of parameters, thereby lowering computation time and enhancing fairness control. 3) Stop: This layer may be used to attain photograph discount within side the photograph. It closes a part of the nerves within side the absolutely incorporated components to boom normal overall performance via way of means of permitting the layer to examine itself thru diverse neurons. 4) Fully Linked Layer: These are deep hidden layers that take enter quantity from the manufacturing of the pinnacle layer to shape an N-dimensional vector, in which N is the wide variety of classes. In our example, the version must pick categories (worn-out or controlled) to be categorized as follows (Fig. 1):
Real Time Exhaustion Detection by Image Classification
485
Fig. 1. Steps to identify Exhaustion
3.2 Steps to Identify Exhaustion Step 1 - facade recognition Instead, taking statistics from our digital digicam is the primary step. As a result, we have got hooked up a vast loop of get right of entry to an internet digital digicam on the way to document every frame. Then, for face detection, we rent a learning-primarily based totally technique primarily based totally on Viola and Jones’ Haar-like capabilities and cascade classifiers. Step 2 - Get a facade in an figure and construct a Region of Interest (ROI) Combined picture is put to use to simplify the preliminary picture processing required to become aware of the face and to appropriately and appropriately calculate square features. The fee of the photograph related to the point (x, y) can be calculated with the aid of using exceeding one picture in place of the first, so that you can be identical to the sum of all of the pixels above and to the left. This will repair the detection with hyperlinks x and y, in addition to the most width of the item boundary box. We can now flip the paper the other way up and draw border bins on every face [8]. Step 3 – Collect the images of eyes from the ROI and deliver them to the partition The same method has been used for facial popularity as its miles used for visible acuity. First, we set up CNN divisions for every eye. We best want to extract the info of the complete picture now. This may be performed via way of means of casting off the border of the attention, after which making use of this code, we will cast off the picture of the attention from the frame. Incorporates eye picture records best [9]. This may be published on our CNN separator, as a way to expect whether or not eyes are open or not. Similarly, the proper eye may be eliminated from the attention [7, 8]. Step 4 - The separator will differentiate between the open eyes or closed eyes We when employ CNN binary department to expect eye condition. We need to carry out such chores so that it will enter our photo into the version, because the version calls for the right dimensions to start with. We can now estimate every eye the use of our version lpred = version. [8] Predict (l eye) classes. When the fee of lpred [0] is 1, the eyes are open; while the fee of lpred [0] is 0, the eyes are closed. Step 5 - count up Points to verify for Tiredness Points are similarly beneficial as those we can use to discover how lengthy this individual has been blinded. Therefore, if each eyes are closed, we can retain to attain points, and if each eyes are open, we can lose points. The impact is drawn at the display screen the usage of the cv2.putText () function, which suggests the real-time repute of
486
A. gupta et al.
the individual. If there are greater than 15 dots, for example, a border is formed, because of this that that a individual’s eyes stay closed for a protracted duration. Now we use sound.play () to create a warning [8]. 3.3 Proposed Claim Propose The creation of an powerful answer is illustrated and illustrated in Fig. 2 on this step. [9] The digital digicam captures pics of the person’s face. After the picture graph is taken, we can use the ML version to decide if the person is worn-out or not. The version detects and gets rid of facial capabilities from the picture graph earlier than sending the accumulated information to a skilled version, after which assesses the person’s situation or stage of fatigue. Lastly, if the effects display sleepiness or fatigue at the person’s face, the app will sound an alarm.
Fig. 2. Architecture of the application
Real Time Exhaustion Detection by Image Classification
487
4 Result and Investigation During our education, we suggested 85.forty nine percentage verification accuracy. We additionally evaluated the overall performance of our proposed software program in lots of regions and determined a mean of ninety one percentage accuracy, which may be very near good. The accuracy of the take a look at is completed via way of means of modeling the research into articles with out education facts. If the take a look at curves and the accuracy of the education begin consistently, we have to prevent the education early to keep away from the incorrect balance. We are coping with an exaggeration problem, wherein the version over-limits the education facts set. The graph suggests that 15 instances in our version is sufficient to get a dependable find (Figs. 3 and 4).
Fig. 3. Correctness of our given CNN Classifier
Fig. 4. Correctness of the har-cascade Classifier
488
A. gupta et al.
4.1 Actual Time in Video Acquisition from the Camera When our machine detects fatigue, we use a beep to alert us. We have taken many human pics that display how humans talk and the conduct of the eyes. In this way, we will use clever telephones going for walks Android or any digital digicam that takes a couple of snap shots right now in a real-time area and display their associated position. Image processing may be used to perceive consumer situations. We can inform if someone is unsleeping or asleep primarily based totally at the photo on his face. Because the editor unearths the eyes closed, consumer fatigue can also additionally appear [21]. To display sleep quality, eye moves are used to calculate the blink frequency and the ultimate time of the eyes. PERCLOS (The closed a part of the eyelids), which bills for eighty to a hundred percentage of the time the human eye is closed, is one of the maximum broadly used signs and symptoms of sleep quality. PERCLOS is one of the maximum crucial real-time signs and symptoms of sleep deprivation, in line with a have a look at with the aid of using Walter Wierwille and colleagues [30]. (Eye-ultimate time/(eye-ultimate time + eye opening)) * a hundred (1) (Figs. 5, 6 and Table 1).
Fig. 5. Image during open eyes
Real Time Exhaustion Detection by Image Classification
489
Fig. 6. Image during closed eyes Table 1. Accuracy with respect to different conditions S. Nos
Grouping
Correctness
1
high-quality light with the eyeglasses
87.242
2
Down daylight the with no eyeglasses
84.45
3
Down light with the eyeglasses
78.362
4
high-quality light with no the eye glasses
90.167
Average
85.4876
5 Conclusion This paper suggests how to detect exhaustion based on multilayer perceptions classifiers. This system has been designed for embedded programs such as car PCs, cockpits, security systems, etc. The function of the program is to detect facial expressions from photographs and to provide information obtained from a trained human representation. The objective of this approach is to minimise the size of the representation in light of the fact that existing applications cannot be employed in embedded systems owing to their tiny size and storage area [7]. The size of the model employed is tiny based on the test findings. Simultaneously, with an accuracy rate of 85.4975 percent, it may be linked into sophisticated depletion detection systems, driving delay monitoring systems, and Smartphone apps. There is, however, still potential for development. Ongoing research will be conducted to identify human abnormalities such as yawning.
References 1. Sałapatek, D.: Proceedings of the Institute of Vehicles 3(112) (2017) 2. Czapski, P.: Institute of Aviation, Center of Space Technologies 3. Federal Highway Administration: PERCLOS: A valid psychophysiological measure of alertness as assessed by psychomotor vigilance (PDF). Washington, DC: US Department of Transportation (1998)
490
A. gupta et al.
4. EdanSafe (2015). www.smartcaptech.com. Accessed 30 Apr 2017 5. Lan, Q., Zhu, Z., Lan, P.: Real-time nonintrusive monitoring and predition of driver fatigue. IEEE Trans. Veh. Technol. 55(3), 1052–1068 (2004) 6. Mridha, K., et al.: Phishing URL classification analysis using ANN algorithm. In: 2021 IEEE 4th International Conference on Computing, Power and Communication Technologies (GUCON), pp. 1–7 (2021). https://doi.org/10.1109/GUCON50781.2021.9573797 7. Singh, P., Bhardwaj, S., Dixit, S., Shaw, R.N., Ghosh, A.: Development of prediction models to determine compressive strength and workability of sustainable concrete with ANN. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 753–769. Springer, Singapore (2021). https:// doi.org/10.1007/978-981-16-0749-3_59 8. https://data-flair.training/blogs/python-project-driver-drowsiness-detection-system/ 9. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems, pp. 1097–1105 (2012) 10. Mehta, S., Dadhich, S., Gumber, S., Bhatt, A.J.: Real-time driver drowsiness detection system using eye aspect ratio and eye closure ratio. In: 2019SUSCOM. Hosting by Elsevier SSRN (2019) 11. Kamarudin, N.: Implementation of Haar cascade classifier and eye aspect ratio for driver drowsiness detection using raspberry Pi. Univ. J. Electr. Electron. Eng. 6(5B): 67–75 (2019). https://doi.org/10.13189/ujeee.2019.061609. http://www.hrpub.org 12. https://towardsdatascience.com/drowsiness-detection-system-in-real-time-using-opencvand-flask-in-python-b57f4f1fcb9e 13. https://www.pyimagesearch.com/2017/04/24/eye-blink-detection-opencv-python-dlib/ 14. Bodapati, S., Bandarupally, H., Shaw, R.N., Ghosh, A.: Comparison and analysis of RNNLSTMs and CNNs for social reviews classification. In: Bansal, J.C., Fung, L.C.C., Simic, M., Ghosh, A. (eds.) Advances in Applications of Data-Driven Computing. AISC, vol. 1319, pp. 49–59. Springer, Singapore (2021). https://doi.org/10.1007/978-981-33-6919-1_4 15. https://www.researchgate.net/publication/322250457_Real_Time_Driver_Drowsiness_D etection_Based_on_Driver’s_Face_Image_Behavior_Using_a_System_of_Human_Com puter_Interaction_Implemented_in_a_Smartphone 16. Ajeer, S., Mathew, A.: Driver exhaustion detection systems (2020). https://doi.org/10.2139/ ssrn.3553366. https://ssrn.com/abstract=3553366 17. Li, Z., Chen, L., Peng, J., Wu, Y.: Automatic detection of driver fatigue using driving operation information for transportation safety. Sensors 17, 1212 (2017) 18. Mukhopadhyay, M., et al.: Facial emotion recognition based on textural pattern and convolutional neural network. In: 2021 IEEE 4th International Conference on Computing, Power and Communication Technologies (GUCON), pp. 1–6 (2021). https://doi.org/10.1109/GUC ON50781.2021.9573860 19. Forsman, P.M., Vila, B.J., Short, R.A., Mott, C.G., Van Dongen, H.P.: ,Efficient driver drowsiness detection atmoderate levels of drowsiness. Accid. Anal. Prev. 50, 341–350, ISSN 0001-4575 (2013). https://doi.org/10.1016/j.aap.2012.05.005. http://www.sciencedirect.com/ science/article/pii/S0001457512001571 20. Sharma, P., et al.: Vibration-based diagnosis of defect embedded in inner raceway of ball bearing using 1D convolutional neural network. In: Artificial Intelligence for Future Generation Robotics, pp. 25–36 (2021). https://doi.org/10.1016/B978-0-323-85498-6.00011-3 21. Nagargoje, S., Shilvant, D.S.: Drowsiness detection system for car assisted driver using image processing. Int. J. Electr. Electron. Res. ISSN 3(4), 175–179 (2015)
Real Time Exhaustion Detection by Image Classification
491
22. Fuletra, J.D., Bosamiya, D.: A survey on driver’s drowsiness detection techniques. Int. J. Recent Innov. Trends Comput. Commun. 1(11), 816–819 (2013) 23. Patel, D., Tiwari, R., Pandey, S., Nikam, R.: Real-time fatigue detection system using computer vision. Int. J. Eng. Res. Technol. (IJERT) 09(06) (2020)
A Secured BlockChain Based Facial Recognition System for Two Factor Authentication Process M. Darshan(B) , S. R. Raswanth, S. Skandan, S. Shakthi Saravanan, Ranjit Chandramohanan, and Priyanka Kumar Department of Computer Science and Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India {cb.en.u4cse19126,cb.en.u4cse19648,cb.en.u4cse18374, cb.en.u4cse18355,cb.en.u4cse18369}@cb.students.amrita.edu, [email protected]
Abstract. Verification of a human being on the other side of the screen has always been a problem and to encounter these issues captcha and OTP verification had been introduced. We have seen various methodologies being deployed to actual businesses and government initiatives for identity verification. Some include physical verification while some stick to video conferencing methods. Video conferencing methods have been a huge advantage especially during the pandemic. With such technologies in place we firmly believe that this would lead us to the next phase in identity verification. Another challenge that needs to be taken care of is storage of such huge delicate and confidential data. Cloud computing has come so far today and the future predicts that it still has miles to go. Studies show that centralised systems will not be able to handle such huge piles of cluttered data. The answer to this problem was Blockchain. Working towards this problem of face recognition and data storage on blockchain, this research work pushes the technical limits and tries to lay the foundation for a new and better form of identity verification. Index Terms: Face recognition · Blockchain · Identity verification · Decentralised authentication · Secured biometric
1 Introduction Web technology has come a long way since the invention of the same. From posting a few pictures on the web, connecting with friends and peers over the internet to advanced data based social media applications and e-commerce, it is very clear that technology is ever evolving. If not for the internet, the entire world wouldn’t have survived the pandemic. It is vivid that technology and the internet have become an integral part of our life now. With the boost in the telecom space and data rates cheaper than ever, India is producing a staggering amount of data that is used in pushing advertisements and projecting content of our choice. During the pandemic, this effect skyrocketed since everyone had only one means of staying connected. This meant that people had to show © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 492–502, 2022. https://doi.org/10.1007/978-981-19-1677-9_44
A Secured BlockChain Based Facial Recognition System
493
up on video conferencing softwares to mimic a normal work life. However this did prove effective in some businesses and government initiatives, it couldn’t be used to its full potential as people were just getting accustomed to these softwares. Web applications have become so powerful these days that you can literally do all the formalities of getting a passport online and just show up once at the regional passport office for identity verification which brings us to the crux of this research work. Identity verification has forever been a challenge but engineers have come up with effective solutions such as captcha check and OTP verification. People are well-aware that these methodologies work in an ideal case setup but the reality of these techniques doesn’t prove to be effective enough. Since COVID-19, the US FBI reported a 300% increase in reported cybercrimes. As if a pandemic wasn’t scary enough, hackers leveraged the opportunity to attack vulnerable networks as office work moved to personal homes. As of this summer, they recorded 12,377 Covid-related scams (Source: Cybint Solutions). Many applications for instance Netflix, Prime Video and Hulu face heavy loss due to recurring non-paid users. Email and OTP verification have been an integral part of identity verification but it needs to evolve to be a perfect fit for the future. With the problem in place, the proposed solution solves the same using face detection and recognition using convolutional neural networks and stores it in a blockchain based storage system for security and decentralised service. Conventional email and OTP services wouldn’t guarantee that the intended user is the one who is actually logging into the application. With the proposed approach, it can be guaranteed that only the intended user can log in, making it easier for netflix and other streaming platforms for detecting recurring non-paid users. The proposed solution can potentially solve a lot of problems in the future with the convenience of the same existing technology. This essentially would change how users log in on a platform and use it overtime. Another huge challenge with collecting a lot of user data is the security of it is very poor. Data leaks and data breaches have been occurring more frequently than ever. This is a huge matter of concern especially dealing with a large set of user facial data. Hence to encounter this blockchain is used to store the data which makes it very difficult to tamper with. With the added transparency, security and wide arenas of usability, blockchain becomes the best sandwiching layer [1]. With this system developed extensively possible use cases would be to use this at ATMs to check if the intended person is only withdrawing cash or not. This process can also be applied for KYC verification and the data stored would be completely secure with the features of blockchain. From a socio-economic standpoint this solution can be implemented for general and state elections and therefore bringing the total cost of the election tremendously down. From a report in Economic Times, India spent close to $6-$7 billion (approx Rs.50,000 crores) on Lok Sabha Elections. With this system in place identity verification can be done with a few simple steps and help in nation building. This proposed solution would lead us into the future of better facial recognition and help in all dimensions of day to day activities.
2 Literature Survey In the research work presented by, “Klemen Grm, Vitomir Struc, Anais Artiges, Matthieu Caron and Hazim Kemal Ekenel” on “ Strengths and Weaknesses of Deep Learning
494
M. Darshan et al.
Models for Face Recognition Against Image Degradations” where they have explored various CNN models such as SqueezeNet, VGG-FACE, AlexNet and GoogLeNet. From their work it is clear that all face recognition models work at similar accuracy and when feeded with lower quality input images, the model shows gradual degradation. Comparative analysis based on various operations such as blur, compression, gaussian noise, salt and pepper noise, brightness, contrast and missing data have been analysed. This helps in finalising the CNN model that has been used for the proposed solution. From the performance covariates applied on all the different models it was a no brainer to go with the VGG model for the proposed solution [7]. In the research work presented by, “Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi and Andrew Zisserman” on “VGGFace2: A dataset for recognising faces across pose and age” the biggest challenge with any deep learning problem had been solved. The collection of dataset is a primal requirement for any convolutional neural network to function on and this research work does exactly that. The dataset contains a wide variety of images with different pose, age, illumination, ethnicity and profession. Different models have also been used to test the improved performance. In conclusion their work had put forth state-of-the-art performance on the IJB-A and IJB-B face recognition benchmarks, exceeding the previous benchmark by a large margin. The learnings from this research work have laid the foundations of the dataset used in the proposed solution [6]. After careful study of the research work by, “ Musab Coúkun, Ayúegül Uçar, Özal Yıldırım and Yakup Demir” on “Face Recognition based on “Face Recognition Based on Convolutional Neural Network” conclusion of better performing neural networks can be established. Their research work proposed a modified convolutional neural network architecture by adding two normalization operations to two of the layers. Batch normalization provided accelerated results on the network. Distinctive face features had been identified and softmax classifier had been used to classify faces in the fully connected layer of CNN. On testing the model, improved results and performance had been observed [8]. With meticulous study on the research work presented by, “ Saqib Ali, Guojun Wang, Bebo White, Roger Leslie Cottrell” on “ A Blockchain-based Decentralised Data Storage and Access Framework for PingER” inference can be drawn on various different methodologies used to store data on the blockchain. In the research work a blockchain based storage system had been designed to house the framework for PingER. This way the total dependency on a centralised repository is withdrawn. Metadata of the files had been stored on the blockchain and the actual files were stored off-chain through distributed hash tables at multiple locations using a peer-to-peer network. This way decentralised system and distributed processing provided efficient lookup capabilities [4]. In the research work presented by, “Hye-young Paik, Xiwei Xu, Hmn Dilum Bandara, Sung Une Lee, Sin Kuang Lo” on “Analysis of Data Management in Blockchain-based Systems: From Architecture to Governance” detailed analysis has been carried out to study the pros and cons of blockchain based data storage systems. An extensive review had been studied to fully understand blockchain technology in the field of data storage and different steps were executed to conceptualise the exact functioning of conventional
A Secured BlockChain Based Facial Recognition System
495
data storage systems. Insights from governance i.e. policies, processes and structures that could be applied for comprehensive control of the technology have also been discussed in the later section. This provided foundations for the legal sector of the proposed solution since facial data is stored on the blockchain with user consent [2]. In the research work presented by, “Bing Xu1, Tobechukwu Agbele and Richard Jiang” on “Biometric Blockchain: A Better Solution for the Security and Trust of Food Logistics” study has been conducted for food safety on aircrafts and in case of any casualty the responsible person can be held accountable. The research work proposes a protocol called Biometric Blockchain that explicitly incorporates the biometric cues of individuals to unambiguously identify the creators and users in a blockchain-based system. This system helped in understanding the procedural concepts the proposed solution needs to apply in order to ensure similar functionalities [3].
3 Proposed System With the transition of the world towards web 3.0, data analytics and blockchain has turned into a prime aspect in secure authentication service. Biometric techniques have evolved in correlation with computing technologies to bring distinctive and highly secured identification. Although there are many biometric techniques that are widely used for identification purposes, facial recognition has become the most suitable and coherent technique as it is the faster processing technique among all [5]. Data security has been a primary concern due to the growth of enormous amounts of data generation and cyber breaches using phishing or brute force. To encounter this security issue, there is a need for the inclusion of a new layer which enhances security and distributes storage of data. The above specified system could be developed by incorporating a 2FA or two step verification along with blockchain which ensures immutability and inherent security. 3.1 Authentication Factors for Secure Login System A factor is an instance of authentication which could be described by something a person knows or something a person possesses or something a person is. Two factor authentication uses knowledge, possession or inherence factors as instances for its service. The inherent / biometric factor provides the most accurate and secured access to data due to its distinctive attributes and less verification time. Non-contact technological systems have increased spontaneously over the years with better provision of easy and quick verification, audit logs and data analysis, which leads to better training and development of facial recognition systems [11]. 3.2 Methodology A web-based client service is provided for the users in order to provide robust authentication and obstructive unauthorized access within premises. After entering the user’s credentials, a facial recognition process based on transfer learning would take place for identification and verification. The vector value from the inferred model would be mapped to a hash value which would be stored on a public blockchain network with
496
M. Darshan et al.
the objective of providing assured and distributed data storage. Unauthorized access of the application will be eliminated by the implication of blockchain and analysis using a convolutional network-based trained model (Fig. 1).
Fig. 1. Schematic representation of proposed system
3.3 Implication of Face Recognition in the Proposed System Facial recognition is a combined process of identification and verification of images and confining them by imposing a bounding box around their degree. For building a powerful and progressive system, a set of new radical technologies such as deep learning models should be incorporated in order to make the underlying machine cognitive towards specified methodology [9]. The complexity of data increases impetuously with the addition of new sets of patterns and to encounter this deep learning allows flexibility in the learning process by producing nested hierarchies of multi layers of artificial neural networks with abstract rendition of data computed in relation with less abstracted data. The lower level categories
A Secured BlockChain Based Facial Recognition System
497
are utilized to define initial patterns which are further used effectively for higher level categorization. The feature extraction and classification are taken simultaneously, which narrows down the approach towards more accurate inference of the system. Deep learning approach takes reasonable time in training which makes our proposed architecture parse data quickly and process logical deduction from the input [10]. Convolutional neural network is an important class of deep neural networks which is widely applied in computer vision based applications. The schematic representation is similar to a synaptic connection in the brain where the model processes input data into multiple convolution layers. Each layer can be described as a set of nonlinear functions of input data at different correlation of dimensionally nearby subsets of outputs from the preceding layer. To make our proposed architecture persistent towards large input images, the VGG-Face model is used to produce robust results irrespective of the quality of the image. VGG16 is a convolutional neural network model which exploits a series of 3x3 kernel sized filters thus making improvements over AlexNet which utilizes larger kernel sized filters. The structure of each series comprises convolution layers in addition to ReLU (Rectified linear activation function) which is followed by a max pooling layer, then followed by two fully-connected layers and softmax layer. The outcome of the final layer is the VGG image descriptor due to activation of convolution neural networks. All principal features of image are extracted regardless of noise factors with the inclusion of VGG16 model, thus making the system yield better inference of the input images [12] (Fig. 2).
Fig. 2. Schematic representation of VGG16 model
3.4 Implication of Blockchain in the Proposed System Blockchain technology has assured reliability in various upcoming technological sectors and it is a prime factor to be considered to bring a radical change in biometric based security systems. Conventional architecture leads to complex systems which are not quite adaptive to fast paced changes and inefficient at handling a multitude of processes. Hence, the implication of blockchain technology leads to the conception of upcoming technology namely “Biometric Blockchain”. The security of databases for storing significant data such as facial descriptor values is vulnerable to unauthorized access and manipulation of data which results in adverse effects on the systems. Incorporation of a distributive ledger-based framework in accordance with blockchain ensures security through the immutability property of the chain and accessibility is unambiguous through the public
498
M. Darshan et al.
ethereum network. Each asset is traced and preserved through individual decentralised structures using private keys (Fig. 3).
Fig. 3. Schematic representation of blockchain model
4 Implementation The initial step is the collection of relevant dataset for training the VGG16 model. For this proposed system we will be using VGGFace dataset which comprises 2.6 million images of public celebrities spanning a wide range of race-ethnicity whereas the curated version in which the noise labels are removed by human annotators has 8 lakhs images with approximately 300 images per identity. The next step is to find as many subjects as possible that have an adequate distinct set of identities in order to produce better data engineering. Keras is a deep learning API which is part of a machine learning framework namely tensorflow which is used to implement deep learning models as fast and simple as possible for research. From keras we will be importing and using VGGFace model 16 for developing the face identification and verification system. For finer training of the mode, we use hyper parameters to generate a good tradeoff between precision and recall. The next step is Feature extraction where all the convolution features will be filtered corresponding to the training data. It is a linear operation where the dot product of a set of weights results in a single value. This method of translation invariance is carried out to comprehend whether the feature is present rather than its position. A two dimensional array output is generated by multiplication of filters (set of weights) with the input which is called the feature map. The next step is to perform fine tuning of images which involves tweaking the parameters of a pre-trained model in order to make it adaptive towards a new subset of specified tasks. The goal is to freeze the initial layers and retrain the last layers. The final step is to predict the values by utilizing the preprocessed inputs obtained from preceding steps [1] (Fig. 4). Distributed storage of data has been done using the development of smart contracts. Ethereum virtual machines provide a runtime environment for migration and testing of the contracts. Facial recognition data are stored in the form of state variables in the
A Secured BlockChain Based Facial Recognition System
499
Fig. 4. Descriptor values
contract and public events are created for every instance. Separate functions such as createAccount and checkLogin ensure security and prevent unauthorized access of data in the client service. Compilation and migration of contracts are done using frameworks such as Truffle and ganache-gui and new blocks will be generated after successful deployment of the contract (Figs. 5 and 6).
Fig. 5. Smart contract
500
M. Darshan et al.
Fig. 6. Migration of contracts
5 Result The proposed architecture for the authentication system is an interconnection of deep learning network and blockchain which affirms a radical change in the security sector of various industries. The trained VGG16 model has been tested with different sets of images and weights and performance data of the model has been plotted for better understanding of the inference. With inclusion of unspecified images present in the dataset, it produces an accuracy of value 0.7692 and with widely used images it produces an accuracy of 0.8333 (Figs. 7 and 8).
Fig. 7. Test prediction results
The ethereum virtual network is utilized using Truffle which is connected to a personal blockchain called ganache. The network is established in the local address (//127.0.0.1) and in port 8545. After compilation using the ‘truffle compile’ command, the contracts are ready for deployment.After checking the configuration of the truffle, contracts have been deployed using the command ‘truffle migrate’. For testing the instance, ‘truffle test’ command is used which displays the passing of events and functions of the smart contract (Fig. 9).
A Secured BlockChain Based Facial Recognition System
501
Fig. 8. Visualization of test performance of the model
Fig. 9. Testing of instances
6 Conclusion The primary objective of developing a proposed system by incorporation of leveraging new technologies such as deep learning and blockchain has been performed. Implication of deep learning convolutional neural networks have made the proposed system perform robustly irrespective of image quality and thereby providing better security solutions to existing les.s efficient conventional systems. Inclusion of smart contracts provides secured storage for crucial data elements thereby ensuring immutability and prevention of unauthorized access through the public ethereum chain network. As future work, the proposed architecture could be fine-tuned using analysis of deep fakes through metadata analytics and usage of digital signature through blockchain.
502
M. Darshan et al.
References 1. Srivatsa, D., Aakash, N., Sahishnu, S., Kumar, P.: A product authentication scheme for supply chain system using smart contract and facial recognition. In: Congress on Intelligent Systems CIS (2020). https://doi.org/10.1088/1742-6596/1767/1/012057 2. Paik, H.-Y., Xiwei, X., Dilum Bandara, H.M.N., Lee, S., Lo, S.: Analysis of data management in blockchain-based systems: from architecture to governance. IEEE Access 7, 186091– 186107 (2019). https://doi.org/10.1109/ACCESS.2019.2961404 3. Xu, B., Agbele, T., Jiang, R.: Biometric blockchain: a better solution for the security and trust of food logistics. IOP Conf. Ser.: Mater. Sci. Eng. 646, 012009 (2019). https://doi.org/ 10.1088/1757-899X/646/1/012009 4. Khokhar, S., Wang, G., White, B., Cottrell, R.: A blockchain-based decentralized data storage and access framework for PingER, pp. 1303–1308 (2018). https://doi.org/10.1109/TrustCom/ BigDataSE.2018.00179 5. Neha, R., Nithin,S.: Comparative analysis of image processing algorithms for face recognition. In: 2018 International Conference on Inventive Research in Computing Applications (ICIRCA), pp. 683–688 (2018). https://doi.org/10.1109/ICIRCA.2018.8597309 6. Cao, Q., Shen, L., Xie, W., Parkhi, O.M., Zisserman, A.: VGGFace2: a dataset for recognising faces across pose and age. In: 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pp. 67–74 (2018). https://doi.org/10.1109/FG.2018.00020 7. Grm, K., Štruc, V., Artiges, A., Caron, M., Ekenel, H.: Strengths and weaknesses of deep learning models for face recognition against image degradations. IET Biometrics 7, 81–89 (2017). https://doi.org/10.1049/iet-bmt.2017.0083 8. Co¸skun, M., Uçar, A., Yildirim, Ö., Demir,Y.: Face recognition based on convolutional neural network. In: 2017 International Conference on Modern Electrical and Energy Systems (MEES), pp. 376–379 (2017). https://doi.org/10.1109/MEES.2017.8248937 9. Nehru, M., Padmavathi, S.: Illumination invariant face detection using viola jones algorithm. In: 2017 4th International Conference on Advanced Computing and Communication Systems (ICACCS), pp. 1–4 (2017). https://doi.org/10.1109/ICACCS.2017.8014571 10. Rajawat, A.S., Rawat, R., Barhanpurkar, K., Shaw, R.N., Ghosh, A.: Blockchain-based model for expanding IoT device data security. In: Bansal, J.C., Fung, L.C.C., Simic, M., Ghosh, A. (eds.) Advances in Applications of Data-Driven Computing. AISC, vol. 1319, pp. 61–71. Springer, Singapore (2021). https://doi.org/10.1007/978-981-33-6919-1_5 11. Meena, D., Sharan,R.: An approach to face detection and recognition. In: 2016 International Conference on Recent Advances and Innovations in Engineering (ICRAIE), pp. 1–6 (2016). https://doi.org/10.1109/ICRAIE.2016.7939462 12. Tolba, A.S., El-Baz, A.H., El-Harby, A.A.: Face recognition: a literature review. Int. J. Appl. Inf. Syst. 11, 21–31 (2016). https://doi.org/10.5120/ijais2016451597
Design of an Optimised, Low Cost, Contactless Thermometer with Distance Compensation for Rapid Body Temperature Scanning Pratik Bhattacharjee1(B) , Suparna Biswas1 , and Sandip Roy2 1
2
Department of CSE, MAKAUT, Kolkata, West Bengal, India [email protected] Department of Computational Science, Brainware University, Kolkata, West Bengal, India
Abstract. High temperature can be a symptom of various critical and semi-critical diseases. In the present COVID-19 pandemic, the demand for non contact based body temperature measurement devices are quite high. However, such devices are not only costlier than their contact based counterparts, but also their accuracy greatly depends on the distance and the location of the subject from the sensor (usually within 2–4 cm). To increase the accuracy of the measurement, researchers usually couple an ultrasonic sensor that can measure the distance of the object and instruct the sensor to take the reading only when the object is within the acceptable distance. However, the measurement of distance by the ultrasonic sensor depends on various aspects such as the impurities (such as sweat), location of the measurement area (forehead, hand etc.)and the initial distance. To overcome these limitations, the present work designed a low cost working prototype based on Arduino together with MLX90614ESF infrared temperature sensor and HC-SR04 ultrasonic sensor with an automatic distance correction mechanism for improved accuracy in temperature measurement. Keywords: MLX90614ESF IR sensor · HC-SR04 ultrasonic sensor Arduino UNO · Non-contact body temperature · COVID safe · Distance correction · IoT
1
·
Introduction
Modern digital thermometers have greatly substituted the traditional mercury thermometers due to their low cost and easy to use aspects [1]. The temperature reading varies significantly depending on the measured area (such as oral, arm pit, forehead etc.). The temperature measurement on the body surface such as forehead gives lower reading than internal sites such as an arm pit or oral. In the ongoing pandemic situation, using one contact based thermometer for rapid temperature scanning of multiple personals, is not suggested by clinical c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 503–511, 2022. https://doi.org/10.1007/978-981-19-1677-9_45
504
P. Bhattacharjee et al.
experts. The thermometer can act as a vector of pathogen [2] transfer specially for the contagious diseases like COVID-19. This gap may be addressed by the infrared (IR) based contactless thermometers by detecting the thermal radiation (black body radiation). These thermometers are widely used now for routine temperature detection at public places [3]. However, the reliability of IR thermometers for clinical usage, were always questionable [4]. It is often suggested to take repeated measurements of the temperature as the values are greatly affected by the surroundings (ambient temperature, humidity, presence of foreign materials on objects such as sweat, sebum etc.) [5]. The ambient temperature as well as the humidity may further affect the rate of blood circulation of the skin surface, that may reduce the value of the expected body temperature [6]. The self-assembled IR thermometers can be erroneous as they lack the required compensating sensors for distance and temperature. Although few models are available with radiation emitter and receiver, the accuracy of measurement still depends on the user’s capability in placing the device at accurate distance [7]. In order to address the accurate distance measurement problem and ignore the possible outliers, the present work proposed a compensatory algorithm to improve the accuracy in measurement of distance by ultrasonic sensor and thereby improving both the reliability and stability of the temperature reading. Experiments were conducted to compare the modified IR thermometer reading with the contact based digital thermometer reading with the help of the human volunteers.
2 2.1
Materials and Methodology Components and Specifications
The prototype is built using an Arduino UNO board [8]. We have used a factory calibrated MLX90614 ESF non-contact IR temperature sensor which can measure the ambient (−40 to 185 ◦ F) as well as object temperature (−94 to 1886 ◦ F) with a correction of ±0.9 ◦ F. The prototype used a HC-SR04 ultrasonic sensor that uses an ultrasonic beam of 40 kHz and measures the round trip delay from the reflected object to calculate the object distance. This calculation, if not properly calibrated, may induce a significant amount of error in the value [9]. Additionally, a 0.96 inch 128x64 dot SSH1106 organic LED (OLED) display along with a 3v active buzzer (for audible alerts) are mounted on a MB102 830 pin solderless breadboard. A micro-SD card reader module with 8 GB memory is accommodated in the assembly for storing the temperature and distance values. The entire assembly is powered by a 9V cell through Arduino UNO [10]. 2.2
Physical Assembly
Figure 1 shows the physical design under various testing condition. The front section consists of the IR sensor, ultrasonic sensor and the OLED display. The rear section contains the Arduino UNO board, the electronic buzzer, micro-sd card reader and the 9V cell box (with the cell). The jumper wires are used accordingly to complete the circuit.
Design of an Optimised, Low Cost, Contactless Thermometer
Fig. 1. The prototype under different conditions
Fig. 2. Circuit diagram
505
506
2.3
P. Bhattacharjee et al.
Circuit Diagram and Implementation
The Arduino UNO microcontroller is connected through breadboard and jumper wires with the pins of IR sensor, ultrasonic sensor, OLED, buzzer and microSD card reader as shown in Fig. 2. The 9v power supply is distributed by the Arduino as 5v and 3.3 V to the electronic components as per their requirements. 2.4
Programming Interface
The Arduino IDE 1.8.15 was used with required library support for the corresponding sensors. The measurement is taken within 2 cm to 4 cm as suggested by the sensor manufacturer. The output is displayed both on OLED [11] and the serial monitor of the Arduino interface during test run. However, the prototype is installed as an independent unit during implementation and the output is displayed on the attached OLED screen only. 2.5
Sensor Calculations
MLX90614 ESF Temperature Sensor has a viewing field of 80◦ . This leads to a distance-to-spot ratio (D/S ratio) of 1:1.68. Considering 58.3 mm as the forehead height of an average human, the IR temperature sensor can reliably read the target temperature approximately upto a maximum of 4 cm. However, the suggested distance as per the sensor manual is between 2 cm to 4 cm [12]. The HC-SR04 Ultrasound Sensor works on the principle of sonar reflection (similar to bats) to measure the distance from the round trip delay. The present work adds a distance correction mechanism to record a steady distance using the statistical median. As the distance directly affects the accuracy of the IR thermometer reading, this step was very crucial [13]. 2.6
Statistical Analysis
The temperature readings were taken and the mean and standard deviations were calculated and analyzed using python (ver 3.9) and plotted as a linear regression with trendline showing the R2 value. The distance is considered as independent variable and the temperature as dependent variable. The predicted temperature after applying the distance compensation, was able to compare with the actual oral temperature value with R2 = 0.998.
Design of an Optimised, Low Cost, Contactless Thermometer
Fig. 3. Operational flowchart
507
508
3 3.1
P. Bhattacharjee et al.
Experimental Setup and Calculations Data Collection
Five volunteers (above 21 years) had given their consent to participate anonymously in the experiment. The temperature data was collected from their forehead with and without distance compensation. Each value is compared against the oral temperature at that moment taken from digital thermometer. The data is recorded against a random number to preserve the anonymity. The present work is based on three experiments. In the first experiment, the temperature data was collected varying the distance from the forehead from 2 cm to 4 cm with an increment of 0.5 cm without the distance correction for the ultrasonic sensor. The second experiment is used to collect data (under the same setup) after the addition of the distance compensation algorithm. Finally, the oral temperature for the same volunteers were taken using a digital thermometer for benchmarking purpose. 3.2
Distance Compensation
The distance compensation algorithm has the following features1. It specifies a custom minimum and maximum distance as per the prototype requirements. 2. It doesnot lag for a full second if no ping echo is received. 3. Ping sensors consistently and reliably at up to 30 times per second. 4. Takes at least 5 readings and sorts them to calculate and return the median as final value (As the statistical median is a better way to remove the outliers compared to mean and mode). 3.3
Methodology
The overall operation of the system is shown in Fig. 3. The maximum sensible distance is set to be 8cm and the maximum waiting time for the ping reply is fixed at 40 ms. The distance sensed beyond 8 cm and below 2 cm is not considered for the present prototype due to the temperature sensing limitations of the IR sensor. Only when the object is between 2 cm to 4 cm, the IR sensor triggers up and starts taking reading (otherwise it remains in sleep state to conserve energy). The electronic buzzer starts beeping in short interval to indicate that the object is at correct distance from the sensor and the temperature is being recorded. The mean temperature of 5 readings is calculated per person [14]. If the mean temperature is more than 98 ◦ F then the buzzer sound is changed to long continuous beep. The temperature is recorded in the SD card module irrespective of high or normal status. Any temperature value more than 107 ◦ F or below 90 ◦ F is considered as outliers and are discarded from the readings automatically.
Design of an Optimised, Low Cost, Contactless Thermometer
4
509
Result Analysis
Figure 4 shows the comparison of the measured distances (with and without correction) against a 15 cm measuring scale. The prototype was tested for two cases, the reading without (Fig. 4a) and with distance compensation (Fig. 4b), for the distance between 2 cm to 4 cm, with an interval of 0.5 cm. The trendline of linear regression analysis shows that the coefficient of determination (R2 ) value of the measured distance is only 0.7508 (Fig. 4a) without the compensation algorithm. The R2 value is improved to 0.9947 (Fig. 4b) after applying the distance compensation algorithm (R2 > 0.95 is considered as acceptable).
(a) Observed vs actual distance[without correction]
(b) Observed vs actual distance[after correction]
Fig. 4. The actual distance vs measured distance
4.1
Performance of the Digital Thermometer
The temperature was taken from the forehead of five volunteers applying the distance compensation algorithm and is benchmarked against a Hicks digital mercury contact thermometer. The temperature was measured in two phasesfor distance 2 cm to 3 cm and then 3 cm to 4 cm with an increment of 0.2 cm as shown in Fig. 5. It is observed that for the present prototype, the coefficient of determination (R2 ) for the temperatures measured for distance 2 cm to 3 cm is slightly better (0.998) than the measurement at 3 cm to 4 cm (0.994). However, in both the cases the values are well above the acceptable threshold of 0.95. The linear regression analysis shown in Fig. 5 considered the compensated distance as independent parameter while the measured temperature was taken as dependent parameter. The best fit straight line for distance 2 cm, to 3 cm is found to be y = 0.3107x + 34.152 (1) and for distance 3 cm to 4 cm is y = 0.3573x + 33.612 with the R2 values of 0.998 and 0.994 respectively.
(2)
510
P. Bhattacharjee et al.
Fig. 5. Measured temperature and trendlines 2 cm to 3 cm (Black) and 3 cm to 4 cm (Red)
5
Conclusion
The recent pandemic has forced us to search for non-contact based healthcare solutions. The temperature is one of the major parameters to detect many types of contagious diseases such as COVID-19. The present work successfully improved the accuracy of the popular IR sensor based thermometers by augmenting a distance compensation algorithm. However, this technique may be extended to any type of medical or non-medical sensor device that depends on accurate distance measurement through ultrasonic sensor. The prototype has a small form factor and cheap components compared to most of the commercially available non-contact based thermometer models. The accuracy after the distance correction is almost equivalent to contact based mercury thermometers. Interested researchers can extend the model for additional functionalities by adding any ESP 8266 based WiFi shield on the top of Arduino and port the data to the cloud. The implementation of multiple numbers of this model at the entrance of highly populated areas such as shopping mall or movie theaters, along with the cloud integration facility may establish an automated healthcare network for community temperature monitoring. So, there exists a number of opportunities to extend the work based on the research requirements. Conflict of Interest. The authors declare that they have no conflict of interest.
References 1. MacRae, B.A., et al.: Skin temperature measurement using contact thermometry: a systematic review of setup variables and their effects on measured values. Front. Physiol. 9, 29 (2018) 2. House, R.J., Henderson, R.J.: Disinfecting the clinical thermometer. Br. Med. J. 2(5475), 1414 (1965)
Design of an Optimised, Low Cost, Contactless Thermometer
511
3. Bhattacharjee, P., Biswas, S.: Smart walking assistant (SWA) for elderly care using an intelligent realtime hybrid model. Evol. Syst. 1–15 (2021). https://doi.org/10. 1007/s12530-021-09382-5 4. Hsiao, S.-H., et al.: Measurement of body temperature to prevent pandemic COVID-19 in hospitals in Taiwan: repeated measurement is necessary. J. Hosp. Infect. 105(2), 360–361 (2020) 5. Ning, B., Wu, Y.: Research on non-contact infrared temperature measurement. In: 2010 International Conference on Computational Intelligence and Software Engineering, pp. 1–4. IEEE (2010) 6. Cui, S., Sun, B., Sun, X.: A method for improving temperature measurement accuracy on an infrared thermometer for the ambient temperature field. Rev. Sci. Instrum. 91(5), 054903 (2020) 7. Santoso, D., Setiaji, F.D.: Non-contact portable infrared thermometer for rapid influenza screening. In: 2015 International Conference on Automation, Cognitive Science, Optics, Micro Electro-Mechanical System, and Information Technology (ICACOMIT), pp. 18–23. IEEE (2015) 8. Gan, S.K.-E., Yeo, J.Y.: The promises of microcontroller kits and smartphone apps for psychological research. Sci. Phone Apps Mob. Device 6 (2020) 9. Shajkofci, A.: Correction of human forehead temperature variations measured by non-contact infrared thermometer. IEEE Sens. J., 1–1 (2021). https://doi.org/10. 1109/JSEN.2021.3058958 10. Arduino. Arduino Uno Rev3 Datasheet. https://www.farnell.com/datasheets/ 1682209.pdf. Accessed 14 Nov 2021 11. Advance Information 128 × 64 Dot Matrix OLED/PLED Segment/Common Driver with Controller. Systech. https://cdn-shop.adafruit.com/datasheets/ SSD1306.pdf. Accessed 20 Nov 2021 12. Melexis. Datasheet for MLX90614. https://components101.com/sites/default/ files/component datasheet/MLX90614-Datasheet.pdf. Accessed 22 Nov 2021 13. ElecFreaks.Ultrasonic Ranging Module HCSR04. https://cdn.sparkfun.com/ datasheets/Sensors/Proximity/HCSR04.pdf. Accessed 20 Nov 2021 14. Chernysheva, N.S., Ionov, A.B.: Application of a multichannel radiation thermometer for increase in adequacy of non-contact temperature measurement results. In: 2015 16th International Conference of Young Specialists on Micro/Nanotechnologies and Electron Devices, pp. 334–336. IEEE (2015)
Sensitivity and Performance Analysis of 10 MW Solar Power Plant Using MPPT Technique Md Fahim Ansari1(B) and Anis Afzal2 1 Electrical Engineering Department, Graphic Era Deemed to be University, Dehradun, India
[email protected]
2 Department of Electrical Power and Control Engineering, School of Electrical Engineering
and Computing, Adama Science and Technology University, Adama, Ethiopia [email protected]
Abstract. Renewable energy had drawn a lot of attention for researchers, technocrats and industry for sustainable development of nation and also drawn vide attention globally. Specially solar energy has been getting attention since long time because of its availability. This paper present sensitivity and performance analysis of 10MW solar photovoltaic (PV) system for its viability and benefits to environment. The paper also presents reduction in carbon emission to the environment, cost effectiveness. The results shows that using solar PV system pay back period has been improvement. Keywords: Solar PV system · Renewable energy · Payback period · Maximum Power Point Tracker (MPPT)
1 Introduction At present various types renewable energy sources has got lot of attention, solar energy forms an acceptable choice for the applications mostly due to its easy to implement, less operational and maintenance cost, pollution free, inexhaustible and the capability of direct conversion of sun energy to electrical energy using PV system [1–4, 20]. It can be placed anywhere but its initial cost is higher and transmission losses are also more. The energy by the sun is not same all over the day. It depends on several influence including irradiance and temperature which are alternating parameters from one condition to another [5, 21]. The basic element in a PV system is the PV cell which converts received energy by sun radiation into electrical energy. The efficiency of the solar cell is very less; it converts only 30–40% of the incident energy to the electrical energy [6–9]. To maximize the efficiency, it necessary to use a technique which is known as MPPT. Figure 1 shows the block diagram of MPPT. The main objective of MPPT is to keep the operating voltage or current at the maximum power point on the PV characteristic curve at all times. For best usage of the available energy, it is needed to operate the system at its maximum power point (MPP). MPPT approaches have been developed and implemented in a variety of ways [10– 12]. Generally, these techniques are categories in two parts namely conventional and soft © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 512–518, 2022. https://doi.org/10.1007/978-981-19-1677-9_46
Sensitivity and Performance Analysis of 10 MW Solar Power Plant
513
Fig. 1. Block diagram of MPPT with solar PV system
computing methods. The conventional MPPT methods are Constant Voltage Method, this method is faster, easy, and simple to use, but it has limited accuracy [13]. Constant Current (CC) Method, this method works on the same principle as the Constant Voltage (CV) Method [14, 15], but its implementation cost is higher due to practical issues associated with measurement of short circuit current [22]. In this paper Perturb and Observe (P & O) technique has been presented. Due to the simplicity of present scheme, mostly researcher implements this technique to the problem but the operating point of solar PV system is oscillating around MPP [15], Incremental Conductance Method, At MPP, the slope of the PV array power curve is zero, whereas on the right it is negative and on the left it is positive. The incremental conductance (INC) approach suffers the same challenges as the P&O method, particularly the speed-oscillation trade-off [16]. In the case of significant variation in irradiance, according to the author, the INC approach is prone to failure. To address the aforementioned problems, several solutions based on artificial intelligence (AI)-based techniques, such as neural network (NN) and fuzzy logic controller (FLC) [17, 18], have been created. However, these solutions have limitations, such as the need for a large number of input storage and a lot of computations. The various techniques are explained below in detail. P & O is reliable and efficient scheme in present time, so this technique used mostly to determine the MPP. The classical P & O process is based on perturbing the module’s operating point and observing the change in module output power. This approach is widely used because of its ease of implementation, reasonable accuracy, and high performance [19]. The present result has been obtained using Ret Screen Expert using control technique Maximum power point tracking.
2 Results In Table 1 the climate condition of selected site of Dehradun has been shown. In this diagram it is showing latitude, longitude, heating temperature, earth temperature with elevation. The role of elevation also plays an important role for cost of PV module and emission analysis. Table 2 shows the availability of solar PV energy on monthly basis with air temperature, humidity, atmosphere pressure etc. at the selected location. The solar energy from January to december are shown, it is found that during the month June earth and air temperature are highest but energy output is not highest as we know that for high
514
Md. F. Ansari and A. Afzal Table 1. Climate of selected the place
Table 2. Month-wise availability of solar energy
Fig. 2. Variation of solar radiation in a year
temperature output power of solar is small. Daily solar radiation energy in MJ/m2 /d along with air temperature is given in Fig. 2. In Fig. 3 various types of energy output with their costs are shown and that cost is tabulated in Table 1. Minimum and maximum cost of energy are depicted in the Table 1. In this table it is found that the cost of energy of solar using tracking system is low as compared to solar energy without tracking system. All these analysis are done using RET screen software.
Sensitivity and Performance Analysis of 10 MW Solar Power Plant
515
Fig. 3. Plots of various types of energy
Table 3. Energy calculations Types of resources
Energy cost minimum $/kWh Energy cost maximum $/kWh
Gas turbine natural gas
0.048
0.259
Gas turbine combined natural 0.049 gas
0.202
Hydro turbine
0.064
0.328
Photo voltaic
0.057
0.271
Photo voltaic tracking system 0.046
0.260
Reciprocating engine diesel
0.183
0.432
Reciprocating engine natural gas
0.051
0.322
Reciprocating engine biogas
0.047
0.148
Solar thermal plant
0.275
0.429
Steam turbine-coal
0.063
0.114
Steam turbine-biomass
0.146
0.270
Wind turbine
0.067
0.190
Wind turbine offshore
0.141
0.284
Green house gas emission is very big challenge for sustainable development and environment, Fig. 4 shows the result of saving in carbon dioxide using solar PV system. The emission of carbon dioxide saving is 11801 tonnes. Figure 5 is cash flow analysis of solar PV system, before 10 years the cash flow is negative it means more money is given and less output is obtained in terms of money. The earning starts after 10years and cash flow becomes positive.
516
Md. F. Ansari and A. Afzal
Fig. 4. Carbon emission reduction using solar PV system
Fig. 5. Payback period of solar PV system
3 Conclusions From Fig. 1, 2, 3, 4 and Fig. 5 and Tables 1, 2 and 3 it can be concluded that the most suitable and environmentally friendly source of energy is solar system. When solar system with controller is being used it found that energy output is highest and its cost of energy per unit is also low as given in Table 1; minimum cost of energy production from PV is 0.057$/kWh without tracking and 0.046$/kWh with tracking. Figure 5 shows the reduction in carbon dioxide by 11801 tonnes along with a net cash flow analysis, which shows the payback period is 10 years. After 10 years solar PV energy will give us profit.
Sensitivity and Performance Analysis of 10 MW Solar Power Plant
517
References 1. Ahmed, J., Salam, Z.: An improved perturb and observe (P&O) maximum power point tracking (MPPT) algorithm for higher efficiency. Appl. Energy 150, 97–108 (2015) 2. Liu, F., Duan, S., Liu, F., Liu, B., Kang, Y.: A variable step size INC MPPT method for PV systems. IEEE Trans. Ind. Electron. 55, 2622–2628 (2008) 3. Kjaer, S.B.: Evaluation of the hill climbing and the incremental conductance maximum power point trackers for photovoltaic power systems. IEEE Trans. Energy Convers. 27, 922–929 (2012) 4. Safari, A., Mekhilef, S.: Simulation and hardware implementation of incremental conductance MPPT with direct control method using cuk converter. IEEE Trans. Ind. Electron. 58, 1154– 1161 (2011) 5. Mei, Q., Shan, M., Liuand, L., Guerrero, J.M.: A novel variable step size incremental resistance MPPT method for PV systems. IEEE Trans. Ind. Electron. 58(4), 2427–2434 (2011) 6. Houssamo, I., Locment, F., Sechilariu, M.: Maximum power tracking for photovoltaic power system: development and experimental comparison of two algorithms. Renew. Energy 35, 2381–2387 (2010) 7. Rai, A.K., Kaushika, N.D., Singh, B., Agarwal, N.: Simulation model of ANN based maximum power point tracking controller for solar PV system. Solar Energy Mater. Solar Cells 95, 773–778 (2011) 8. Alajmi, B., Ahmed, K., Finney, S., Williams, B.: Fuzzy logic control approach of a modified hill-climbing method for maximum power point in microgrid standalone photovoltaic system. IEEE Trans. Power Electron. 26, 1022–1030 (2011) 9. Jain, S., Agarwal, V.: A single-stage grid connected inverter topology for solar PV systems with maximum power point tracking. IEEE Trans. Power Electron. 22, 1928–1940 (2007) 10. Ahmed, J., Salam, Z.: A modified P&O maximum power point tracking method with reduced steady-state oscillation and improved tracking efficiency. IEEE Trans. Sustain. Energy 7, 1506–1515 (2016) 11. Ahmed, J., Salam, Z.: An enhanced adaptive P&O MPPT for fast and efficient tracking under varying environmental conditions. IEEE Trans. Sustain. Energy 9, 1487–1496 (2018) 12. Haegel, N.M., Kurtz, S.R.: Global progress toward renewable electricity: tracking the role of solar. IEEE J. Photovoltaics 11(6), 1335–1342 (2021). https://doi.org/10.1109/JPHOTOV. 2021.3104149 13. Belkhier, Y., Achour, A., Shaw, R.N., Sahraoui, W., Ghosh, A.: Adaptive linear feedback energy-based backstepping and PID control strategy for PMSG driven by a grid-connected wind turbine. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 177–189. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0749-3_13 14. Zafoschnig, L.A., Nold, S., Goldschmidt, J.C.: The race for lowest costs of electricity production: techno-economic analysis of silicon, perovskite and tandem solar cells. IEEE J. Photovoltaics 10(6), 1632–1641 (2020). https://doi.org/10.1109/JPHOTOV.2020.3024739 15. Kost, C., Shammugam, S., Jülch, V., Nguyen, H.T., Schlegl, T.: Levelized cost of electricityrenewable energy technologies (2018) 16. Sivaneasan, B., Yu, C.Y., Goh, K.P.: Solar forecasting using ANN with fuzzy logic preprocessing. Energy Procedia 143, 727–732 (2017). https://doi.org/10.1016/j.egypro.2017. 12.753 17. Goyal, S.B., Bedi, P., Rajawat, A.S., Shaw, R.N., Ghosh, A.: Smart luminaires for commercial building by application of daylight harvesting systems. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 293–305. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_24
518
Md. F. Ansari and A. Afzal
18. Chen, Y., et al.: Mass production of industrial tunnel oxide passivated contacts (i-TOPCon) silicon solar cells with average efficiency over 23% and modules over 345 W. Prog. Photovolt. Res. Appl 27(10), 827–834 (2019) 19. Green, M.A., et al.: Solar cell efficiency tables (Version 53). Prog. Photovolt. Res. Appl. 27(1), 3–12 (2019) 20. Ansari, M.F., Chatterji, S., Iqbal, A., Afzal, A.: Control of MPPT for photovoltaic systems using advance algorithm EPP. In: Proceedings of third international conference on power systems, IEEE, I.I.T., Kharagpur, India, 27–29 Dec (2009) 21. Afzal, A.: Performance analysis of integrated wind, photovoltaic and biomass energy systems. In: World Renewable Energy Congress to be held on 8–13 May 2011 in Linköping University, Sweden (2011) 22. Afzal, A., Mohibullah, M., Kumar Sharma, V.: Optimal hybrid renewable energy systems for energy security: a comparative study. Int. J. Sustain. Energy 29(1), 48–58 (2010)
SAR Reduction of Wearable Antenna Using Multiple Slots and AMC Combination Sonali Kerketta(B)
, Azharuddin Khan , and Amit Kumar Singh
Department of Electronics Engineering, Indian Institute of Technology (BHU), Varanasi, India [email protected]
Abstract. The emergence of wearable and implantable devices has escalated the demand for efficient antenna designs. It has raised public concerns about their safety as the proximity of such devices to the human body during their operation results in highly non uniform radiation exposure. Long-term exposure to such radiation can rupture the tissues and can even cause lethal diseases. As a result, concern about the safety aspects and perilous effects of EM waves is needed. This paper presents a design of a rectangular microstrip patch antenna operating at 2.4 GHz. Two different SAR reduction techniques have been employed for SAR reduction. One is introducing multiple slots in the radiating patch, and the other is using Artificial Magnetic Conductor as a shield. Ansys HFSS has been used for design and simulations. Keywords: Artificial Magnetic Conductor (AMC) · Microstrip patch antenna · Specific Absorption Rate (SAR) · Slots
1 Introduction Wireless Body Area Network (WBAN) technology has revolutionized our world and daily lives. A WBAN has a wide variety of applications in remote health monitoring, medicine, home/health care, sports, multimedia, and many others. WBAN technology is capable of monitoring patients’ well-being and can provide real-time updates of medical records through the Internet. Wearable and implantable gadgets have increased the need for effective antenna designs due to their increasing popularity. As the frequency is inversely proportional to the depth of penetration by an RF field. At high frequencies, the penetration depth in the body caused by an RF field is low, resulting in a temperature rise in the body tissues. Similarly, the emission of undesirable radiation towards the human body by wearable antennas causes an increase in body temperature. Above a certain level, this temperature rise can cause serious health effects such as tissue damage or burns. So, this unwanted radiation needs to be monitored, and safety limits must be set to avoid the perilous consequences of EM waves. Specific Absorption Rate (SAR) measures the rate at which energy is absorbed per unit mass by a human body when exposed to a radio frequency (RF) electromagnetic field. It is defined as the power absorbed per tissue mass and can be expressed in terms of Watts per kilogram (W/kg). To protect the human body from the perilous consequences of EM waves, regulatory © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 519–527, 2022. https://doi.org/10.1007/978-981-19-1677-9_47
520
S. Kerketta et al.
restrictions are imposed by International Commission on non-Ionizing Radio Protection (ICNIRP). As per the regulations, the SAR value must not exceed 2.0 W/kg measured over 10 g [1, 2]. However, the Federal Communications Commission (FCC) restricts the SAR value to more than 1.6 W/kg for 1g of tissue.
2 Antenna Geometry 2.1 Mirostrip Patch Antenna Structure A microstrip patch antenna comprises a radiating metallic patch mounted on one side of a non-conducting substrate panel, and a metallic ground plane mounted on the other [3]. Microstrip patch antennas come in various geometric forms, out of which rectangular, square, and circular are the most commonly used. Different feeding techniques like inset feed, co-axial feed, aperture coupled, and proximity coupled feed excites the antenna. The significant benefits of a patch antenna are its lightweight, cheap cost simplicity of assembly. The microstrip antenna and the feed can be included in the same circuit. The design requirements of the microstrip patch antenna are shown in Table 1. Table 1. Design requirements Parameters
Value
Operating frequency (f )
2.4 GHz
Dielectric substrate
FR-4
Dielectric constant of the substrate (r)
4.4
Height of the dielectric substrate (h)
1.6 mm
Return loss, S11
_0 [30]. As it have a linear non saturating point is stated that it induces more concurrence of the assumptive ramp slant, compare to the Sigmoid and Tangent-h. Also, correlated to Tangent-h or Sigmoid neurons that imply upscale mathematical act, Relu function can be enabled by combination of activation at nil. Nonetheless, it carries main disadvantage, as they are open to die during the process when the gradient reaches 0. As outcome, in the Table 1, compared to Sigmoid and tangent- h, including their mathematical expressions. 4.6 Evaluation Metrics The most vividly used Metric to measure the performance of automated Learning model is its accuracy. This measure would also be applied to figure out the overall carrying out of the setup, but also resembles when choosing between different models. By this way the mix of parameters that offers much more accuracy can be found. In this submodule Various metrics and evaluation techniques will be put for two different models. The first one will deal all metrics that allows to choose Most appropriate for the model. It mainly includes cross validation for the Choice of Model hyper Parameters and usage of set of Validation techniques for the identification of situation under or over fitted. On the other side the main techniques for evaluating the final output of a classification is Introduced. These will be the accuracy of the model, it’s confusion matrix or Receiver Operating characteristics (ROC) curve.
744
J. Dhanasekar et al. Table 1. Comparison between activation functions Mathematical expression
Advantages
Disadvantages
Sigmoid
1hθ (x) = −x1 + e
Gives binary output in range (0,1) [17] Suitable for binary classification output layer [11]
Stagnant and kills gradients [15] Interpretation of output is not zero [17]
Tanh
2hθ (x) = 1 + e−x + 1
Output is zero cantered (−1,1) [25]
Saturated and kills gradients [35]
Relu
hθ (x) = max (0, x)
Maximum acceleration to convergence due to its linear non saturating form [27]
When neuron becomes 0 the gradient will never activates [12]
Cross validation is required to assess the performance of our model with a variety of parameters. Each model will be trained with each parameter, and their performance will be validated with the validation set. This set will be required because if we utilise the test subset to confirm the dissimilar models and select the most appropriate one, this test will have already been observed by our final model and will no longer be valid for final evaluations. Cross Validation combines of tearing the drill set into k subsets and, at the pace of training, each k sub brackets taken in considered as the model’s investigated set, on the other hand the rest of the data will be considered as the drill set. This trail will be looped k times, and in each monotony a various investigated set are taken in account, while the left off data will be applied, as quoted, as a drill set. The exploit will be the average of the exploit in each repetitiousness. This approach comes handy when a petite dataset is attainable. Added hook to bring out cross validation will be at the hand of the hold-out method. This technique conflict from the preceding one in that the dataset is unequivocally teared into drill and validation subsets. Considered approach will be much fewer computationally upscale, but will desire a preponderant dataset size [14]. Paralleling the two approaches, the k-fold approach has the leverage that every data are employed to train and validate, so more specimen outcome are obtained a priori. On the adverse, by means of the hold-out knack it is fortuitous to have ill fortune when forging the a priori division between training and validation which may outcome in nonepitome samples of the dataset. For this cover, even if the dataset is not petite, if there is ample computational forte, it will always be cap to bring off a k-fold approach in lieu of a hold-out [28]. K fold validation process is particularly described to select the model hyperparameters, whereas the precision to select and flaw are considered and composed to derive favourable outcomes. Hence, the process is beneficial for having a small dataset, despite/however the drawback wil distinctly be its great computational expense. Likely, over-fitting and underfitting situations are fundamentals to examine the presence of Machine Learning, during
Design and Implementation of a Defect Identification
745
the training of the model. A model is proceeding to be over-fitted while performing well with training data, yet its accuracy is preeminently lower with test data. In consideration, the model has retained the data it has perceived and not able to generalize the rules to predict the data it has not recognized/identified. Besides, under-fitting occurs when there is an excess of generalization of the model, which practically ignores most of the training samples. Perhaps, the validation set is essential while training set and simultaneously through validation set. The evolution can be seen along the training epochs. This evolution is signifying through its learning curve that enact the relationship between training set size and its error rate on the training and validation sets respectively. They can e an indispensable tool while diagnosing the performance, of the models. Similarly, can disclose whether the model is suffering from bias, under-fitting, or variance, over-fitting, problems. On the flip side, on one occasion we have elect the right parameters for our model and trained it, its outcome will be weighed with rand-new input data that will harvest in disparate metrics. As stated on top of, the most casual metric to weigh the model will be its certainty. Certainty artlessly calibrates how recurrently the segregator composes the certain fore-cast. It is the proportion at intervals of the number of certain fore-cast and the sum number of fore-cast. Nevertheless, the prime obstruction of the accuracy metric is that it conjectures matched cost for both kinds of errors. in the vicinity of application of the project it will be apt to classify the various defects on different products in manufacturing industries. For as much as it will be imperative to stage the confusion matrix of the model. A perplexing matrix is an illustration that assorts fore-cast bestow to in case they peer the factual value in the data. One of the tabular column’s ambit pinpoint the cinch league of fore-casted values while the alternate ambit pinpoints the ditto for absolute merits. Therefore, through this confusion matrix, the preciseness of an allocation can be weighed by figure out the amount of precisely identified grade illustrations (true positives), the amount of precisely identified illustrations that outcast to the grade (true negatives), and illustrations that each of two were mistakenly accredit to the grade (false positives) or that were overlook as grade paragon (false negatives) (Fig. 12, 13 Table 2 and 3). Table 2. Trained images Defect type
Teeth which are missing
Surface which are scratched
Gear which are normal
Prime of images
Test
Train
Test
Train
Test
Test
100
50
100
50
70
40
746
J. Dhanasekar et al.
Fig. 12. Training process of the dataset
Table 3. Evaluation metrics of the model constructed by Keras Metric
Training
Testing
Accuracy
97.37%
97.10%
Sensitivity
95.15%
95.01%
Specificity
95.98%
95.91%
Fig. 13. Accuracy graph of trained dataset
Design and Implementation of a Defect Identification
747
4.7 Working First, picture of a gear (subject material for here) should be taken. After that, the picture is subject to many image processing techniques such as edge detection and object identification have been done, then it locates the defect and classifies which type of defect it is and it show the output as a picture (Fig. 14 and 15).
Fig. 14. Final setup
5 Results In this process flow, we have tested using different flaws of the gear in each segment. From the flow we have attained the accuracy 97.37% of the flaws in the gear. The trained model gives the default characteristics from the process which took place during the testing of the gears. The following things are the outcomes of testing process which we intend to perform (Fig. 16, 17 and 18).
748
J. Dhanasekar et al.
Fig. 15. Flow chart of the process
Fig. 16. Missing gear teeth is identified
Design and Implementation of a Defect Identification
749
Fig. 17. Perfect gear is identified
Fig. 18. Scratch over the gear is detected
6 Conclusion In this chapter, a new technique to find various defects in a product. This will dramatically reduce the defect identification in quality department and saves a lot of time. For huge scale production, it can be implemented in several times in different production lines. To make it ease we can also integrate with IOT system for better control and monitoring in huge scale production.
References 1. Caggiano, A., Zhang, J., Alfieri, V., Caiazzo, F., Gao, R., Teti, R.: Machine learning-based image processing for on-line defect recognition in additive manufacturing (2019) 2. Amini, M., Chang, S.I.: MLCPM: a process monitoring framework for 3D metal printing in industrial scale (2018) 3. Nuchitprasitchai, S., Roggemann, M., Pearce, J.M.: Factors effecting real-time optical monitoring of fused filament 3D printing. Progress Addit. Manufact. 2(3), 133–149 (2017). https:// doi.org/10.1007/s40964-017-0027-x 4. Wanga, J., Ma, Y., Zhang, L., Gao, R.X., Wu, D.: Deep learning for smart manufacturing: methods and applications (2018) 5. Ghahramani, M., Qiao, Y., Zhou, M.C., O’Hagan, A., Sweeney, J.: AI-based modeling and data-driven evaluation for smart manufacturing processes (2020) 6. Essien, A., Giannetti, C.: A deep learning model for smart manufacturing using convolutional LSTM neural network autoencoders (2020)
750
J. Dhanasekar et al.
7. Bodapati, S., Bandarupally, H., Shaw, R.N., Ghosh, A.: Comparison and analysis of RNNLSTMs and CNNs for social reviews classification. In: Bansal, J.C., Fung, L.C.C., Simic, M., Ghosh, A. (eds.) Advances in Applications of Data-Driven Computing. AISC, vol. 1319, pp. 49–59. Springer, Singapore (2021). https://doi.org/10.1007/978-981-33-6919-1_4 8. Bo, Chen, S., Chen, J., Zhangfeng: Hyperspectral image classification with convolutional neural network and active learning (2020) 9. Feng, J., Feng, X., Chen, J., Cao, X., Zhang, X., Jiao, L., Yu, T.: Generative adversarial networks based on collaborative learning and attention mechanism for hyperspectral image classification (2020) 10. He, T., Zhang, Z., Zhang, H., Zhang, Z., Xie, J., Li, M.: Bag of tricks for image classification with convolutional neural networks (2019) 11. Rajawat, A.S., Rawat, R., Mahor, V., Shaw, R.N., Ghosh, A.: Suspicious big text data analysis for prediction—on darkweb user activity using computational intelligence model. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 735–751. Springer, Singapore (2021). https:// doi.org/10.1007/978-981-16-0749-3_58 12. Jang, K., Kim, N., An, Y.-K.: Deep learning–based autonomous concrete crack evaluation through hybrid image scanning (2019) 13. Kang, D., Benipal, S.S., Gopal, D.L., Cha, Y.-J.: Hybrid pixel-level concrete crack segmentation and quantification across complex backgrounds using deep learning (2020) 14. Tabiaia, I., Tkachev, G., Diehla, P., Frey, S., Ertl, T., Therriault, D., Lévesque, M.: Hybrid image processing approach for autonomous crack area detection and tracking using local digital image correlation results applied to single-fiber interfacial debonding (2019) 15. Samanta, D., et al.: SET based logic realization of a robust spatial domain image watermarking. In: 2008 International Conference on Electrical and Computer Engineering, pp. 986–993 (2008). https://doi.org/10.1109/ICECE.2008.4769356 16. Vardhana, M., Arunkumar, N., Lasrado, S., Abdulhay, E., Ramírez-González, G.: Convolutional neural network for bio-medical image segmentation with hardware acceleration (2019) 17. Ghosh, S., Das, N., Das, I., Maulik, U.: Understanding deep learning techniques for image segmentation (2019) 18. Bedi, P., Goyal, S.B., Rajawat, A.S., Shaw, R.N., Ghosh, A.: A framework for personalizing atypical web search sessions with concept-based user profiles using selective machine learning techniques. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 279–291. Springer, Singapore (2022). https:// doi.org/10.1007/978-981-16-2164-2_23 19. van Roy, J., Keresztes, J.C., Wouters, N., De Ketelaere, B., Saeys, W.: Measuring colour of vine tomatoes using hyperspectral imaging (2017) 20. Liu, Z., He, Y., Cen, H., Lu, R.: Deep feature representation with stacked sparse auto-encoder and convolutional neural network for hyperspectral imaging-based detection of cucumber defects (2018) 21. Palimkar, P., Bajaj, V., Mal, A.K., Shaw, R.N., Ghosh, A.: Unique action identifier by using magnetometer, accelerometer and gyroscope: KNN approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 607–631. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_48 22. Klimt, B., Yang, Y.: The enron corpus: a new dataset for email classification research. In: Boulicaut, J.-F., Esposito, F., Giannotti, F., Pedreschi, D. (eds.) ECML 2004. LNCS (LNAI), vol. 3201, pp. 217–226. Springer, Heidelberg (2004). https://doi.org/10.1007/978-3-54030115-8_22 23. Spanhol, F.A., Oliveira, L.S., Petitjean, C.: Laurent Heutte: a dataset for breast cancer histopathological image classification (2019)
Design and Implementation of a Defect Identification
751
24. Abiodun, O.I., Jantan, A., Omolara, A.E., Dada, K.V., Mohamed, N.A.E., Arshad, H.: Stateof-the-art in artificial neural network applications: a survey (2018) 25. Albawi, S., Mohammed, T.A., Al-Zawi, S.: Understanding of a convolutional neural network (2017) 26. Gangwar, M., Yadav, R.S., Mishra, R.B.: Semantic web services for medical health planning. In: 2012 1st International Conference on Recent Advances in Information Technology (RAIT), pp. 614–618. IEEE, (2012) 27. Rajawat, A.S., Rawat, R., Barhanpurkar, K., Shaw, R.N., Ghosh, A.: Blockchain-based model for expanding IoT device data security. In: Bansal, J.C., Fung, L.C.C., Simic, M., Ghosh, A. (eds.) Advances in Applications of Data-Driven Computing. AISC, vol. 1319, pp. 61–71. Springer, Singapore (2021). https://doi.org/10.1007/978-981-33-6919-1_5 28. Gangwar, M., Mishra, R.B., Yadav, R.S.: Intelligent computing methods for the interpretation of neuropsychiatric diseases based on Rbr-Cbr-Ann integration. Int. J. Comput. Technol. 11(5), 2490–2511 (2013) 29. Molnar, F.J., Byszewski, A.M., Rapoport, M., Dalziel, W.B.: Practical experience-based approaches to assessing fitness to drive in dementia. Geriatr Aging 12(2), 83–92 (2009) 30. Fillenbaum, G.G., et al.: Consortium to Establish a Registry for Alzheimer’s Disease (CERAD): the first twenty years. Alzheimer’s Dement. 4(2), 96–109 (2008) 31. Chakraborty, A., Chatterjee, S., Majumder, K., Shaw, R.N., Ghosh, A.: A comparative study of myocardial infarction detection from ecg data using machine learning. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 257–267. Springer, Singapore (2022). https://doi.org/10.1007/978-98116-2164-2_21 32. Pei, G., Hu, R., Dai, Y., Manuel, A.M., Zhao, Z., Jia, P.: Predicting regulatory variants using a dense Epigenomic mapped CNN model elucidated the molecular basis of trait-tissue associations. Nucleic Acids Res. (2020) 33. Ramzan, F., Khan, M.U.G., Rehmat, A., Iqbal, S., Saba, T., Rehman, A., Mehmood, Z.: A deep learning approach for automated diagnosis and multi-class classification of Alzheimer’s Disease stages using resting-state fMRI and residual neural networks. J. Med. Syst. 44(2), 1–16 (2019). https://doi.org/10.1007/s10916-019-1475-2 34. Palimkar, P., Shaw, R.N., Ghosh, A.: Machine learning technique to prognosis diabetes disease: random forest classifier approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 219–244. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_19 35. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis (2019) 36. Guo, T., Dong, J., Li, H., Gao, Y.: Simple convolutional neural network on image classification (2017) 37. Goyal, S.B., Bedi, P., Rajawat, A.S., Shaw, R.N., Ghosh, A.: Multi-objective fuzzy-swarm optimizer for data partitioning. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 307–318. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_25 38. Kamilaris, A., Prenafeta-Boldú, F.X.: Deep learning in agriculture: a survey (2018) 39. Esteva, A.: A guide to deep learning in healthcare (2019) 40. Singh, P., Bhardwaj, S., Dixit, S., Shaw, R.N., Ghosh, A.: Development of prediction models to determine compressive strength and workability of sustainable concrete with ANN. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 753–769. Springer, Singapore (2021). https:// doi.org/10.1007/978-981-16-0749-3_59 41. Experimental assimilation of various tuning rules with fractional order controller in inverted pendulum. I. J. Eng. Adv. Technol. (2019)
Design and Simulation of a Robotic Manipulator for Ladle with PLC V. Priya1(B) , V. J. Aravinth Raj1 , K. V. Chethanasai1 , J. P. Manoj Kumar1 , V. Manikandan1 , and K. K. Senthilkumar2 1 Department of Mechatronics, Bharath Institute of Higher Education and Research, Chennai,
India [email protected] 2 Prince Shri Venkateshwara Padmavathy Engineering College, Chennai, India
Abstract. In recent times, Robotics play major vital role in all activities in industrial need including human life. Robots are often utilized for several purposes and far of situations also utilized in dangerous environments (including inspections of radioactive materials, bomb detection, and deactivation), Manufacturing Process, Aerospace, Automobile Industry, Defense sectors, where human cannot survive. The Main Objective of this chapter describes the system for Aluminum Pouring in the Casting Industry or Foundry Industry. In the Modern Industrial Manufacturing process consists of Different Products manufactured by the company in Automobile Industry Sector, Aircraft parts Manufacturing Sector, etc. This chapter is to design the control the Robotics Arm auto ladle automatically and manually by using Programmable Logic Controller-(PLC) used to automatically pouring the molten metal into the Die and the Output will be in Casting or Mold product. Manufacturing Industries keep on manufacturing the same models simultaneously and some models have bit variation in Height, Weight, and Shape. Industrial Automation focused on to increase the accuracy, consistency in the quality of the product, Long durability, Making the system to user friendly using the Human Machine Interface-(HMI) as possible, Also, Automation helps to improve Production, Reducing the Human Errors while pouring the aluminum, etc. Keywords: Robotic Arm for automatic ladle · Auto ladle · Aluminum pouring · Casting · Foundry industrial · Programmable Logic Controller-(PLC) · Mold · Accuracy · Long durability Human Machine Interface-(HMI)
1 Introduction In the Foundry Industry or Die Casting Industry which produces a casting of metal. The manufacturing process of casting metal/Mould product by pouring the hot liquid metal into the Die/Mould. Mould contains a hollow container used to shape the molten into solid part the solid part known as the desired Mould product or casting. In the Foundry Industry, all the work is processed by manually and in manual process takes more timeconsuming for creating a job or casting. This Chapter Discuss about the Pouring the © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 752–765, 2022. https://doi.org/10.1007/978-981-19-1677-9_66
Design and Simulation of a Robotic Manipulator for Ladle with PLC
753
Aluminium Alloy metal into Mould. The metal is converted into the liquid form of the metal is known as Molten metal. The Furnace is used to converted the metal into the molten metal, then ladle is used to transfer the liquid form of metal from furnace to Mould/Die. The temperature required for Liquid form of the metal/Molten metal was 640 °C–700 °C (Melting Point) and the temperature required for the furnace was 800 °C–900 °C to melt the Aluminium Alloy metal into Liquid state/Molten metal. The Capacity and Size of the ladle change depends on the Customer/User in this chapter small ladle, capacity 30–50 kg are used to pouring the molten. Due to the manual system Die Casting Industry/Foundries are facing number of Problems/faulty. The liquid type of metal called molten metal is incredibly high in temperature; it can be very dangerous for employees to work at high temperatures. It can be one of the sources of casting failures and cost increases owing to this high temperature. The Main Objective of the chapter used to pour the molten metal into Mould/Die using Robotic manipulator/Arm in End-Effector ladle (U-shaped) it’s like spoon used to transfer the hot liquid metal into mild.
2 Mechanical Design The frame is the main body of the Robotic ladle. The mechanism was designed for testing purposes, The most important step with respect to the mechanical design of the serial robot mechanisms is the choice for the motor for each joint. The proposed
Fig. 1. Design and simulation in Rokisim
754
V. Priya et al.
serial robotic mechanisms consists of five revolute joints and a ladle cup as an endeffector. Considering the high strength and minimum cost requirement two options were available. Considering iron are the materials used for the robotics manipulator and ladle. The pouring temperature of the metal vary the casting size and design, desired rate of pouring. The End-Effector tool for the Robotic ladle material which is used in the design Cast Iron and its coated with Ceramic Fig. 2 and the general range of the melting point temperature of 1800–2000 °C and the general range for the pouring metal temperature; Steel-2850°F to 2950°F, Aluminium-1250°F to 1400°F. The Pouring capacity for the ladle in general range 0.5 to 15 kg. the cycle time of the 18 s. The Robotic manipulator ladle Designed in Solid Works Software and Simulated in Rokisim Software Fig. 1.
Fig. 2. Ladle cup end-effector tool of the Robotic arm
3 DVP20EX200R/Programmable Logic Controller-(PLC) The Programmable Controller (PC) is also known as the Programmable Logic Controller (PLC), and is a sophisticated device for industrial control. Instead of electromechanical devices using Integrated Circuits - (IC) to implement control functions. Control and Programming for the movement of the robot’s parts is important especially when the applications requires and easy to implement. In the field of robot programming and control show various ways of solving the problem. The PLC provided relay capability, thereby mechanically replacing the original hardwired relay logics that are used
Design and Simulation of a Robotic Manipulator for Ladle with PLC
755
for electrically powered machines in switching electrical circuits. In PLC Programming Ladder logic, the instruction set and the functional block diagram is the main programming methods used for the PLCs and it has continued the way of representing electrical sequences of operations and circuits. The relay logic wiring schematics diagram used to take the decision was strategic one. The ladder logic programming method as the main programming method this eliminated the need of teach the electricians, technicians and engineers for how to programming and the ladder logic programming method is easy to understand the functions and sequence of the program. In this chapterwe use the DVP20EX200R is the PLC 32-bit CPU that combines a microprocessor for High- speed processing, an integrated power supply of 110–240 VAC, it has an 8Digital Input + 6Digital Output and 4 Analog Input + 2 Analog Output. The PLC DVP20EX200R, it has built-in PID – (Proportional Integral Derivative) auto tuning function controllers used to control the loop feedbacks and it has complete Analog control solution. The DVP20EX200R Fig. 3 programming capacity 16 K steps and its memory Data register capacity 10 K words. It has a highly efficient processing capacity of 1K program steps which can be completed within 1ms and a maximum 100 kHz pulse control for complex motion control instructions required for multi-axis applications.
Fig. 3. Programmable Logic Controller-(PLC)/DVP20EX200R.
4 Stepper Motor and Stepper Motor Driver 4.1 Stepper Motor The Stepper motor is a brushless DC motor, which can be considered a digital electromechanical system where the pulse modulation of each step (electrical pulse input) results in a discreet angle of travel of the shaft called a step angle of the motor. Ideally, it
756
V. Priya et al.
is switched to an open loop location and speed control. The stepper motor Fig. 4 comes in a wide range of angular positions and the maximum angle of the turn per step can be 90° per step and it can reach a minimum to 72o per step. The stepper motor angle position and speed controlled by the Stepper motor driver.
Fig. 4. Stepper motor
4.2 Stepper Motor Driver (DM542) The DM542 is micro-step and its fully digital stepper driver developed on the latest motion control technology with the advanced bi-polar constant chopper driver technology based on the algorithm of advanced DSP control it has a special level structure stable operation for smoothness, which has an optimum high torque. Also, it reduces the vibrations and the noise of the operating motor. The DM542 has features that can drive a stepper motor at low heat, low noise and smooth motion of the motor. The DM542 has Multi-Steps, which helps low-resolution feedback to generate higher micro-steps and thus provides smoother motor motions. It has the features of 15 selectable microsteps resolutions. The DM542 is a DC18-50 V power supply and ooften suitable for two-phase and four-phase stepper motors with a current of 4.2 A. The driver peak operating current range from the 1.0 A to 4.2 A. Moreover, it has an automatic over-voltage, semi-flow, under-voltage, and over-current protection functions. It can be used in various kinds of instruments and automation equipment such aspacking machines, cutting machines, engraving machines, etc. And also it also adapts to the application performs with low-noise, high-speed, and with high-precision. 4.3 Selecting Driver Output Current and Micro-step a) Current Selection: The DM542 driver use 8-bitDIP switch to set for a motor operating Current Table 1 and micro-step resolution, as shown below:
Design and Simulation of a Robotic Manipulator for Ladle with PLC
757
Table 1. Current selection
b) Micro-step resolution Selection: The driver use8-bit DIP switch to set for a motor micro-step resolution Table 2, as shown below:
Table 2. Micro-Step resolution selection
c) Contorl Signal Connector Interface: The DM542 driver has two connectors, the P1 connector for control of signal connections; and the P2 connector for power and motor connections. The DM542 will accept single-ended and differential inputs (including an anode and cathode connection output) Fig. 5(a), 5(b). The motor driver has three optically separated logic inputs situated on the P1 connector, except for the control signal of the line driver. In comparison, separate inputs suppress or minimize electrical noise when coupled to the drive control signals.
758
V. Priya et al.
Fig. 5. (a) Common anode connection (open-collector signal) (b) Common cathode connection (PNP-Signal)
Design and Simulation of a Robotic Manipulator for Ladle with PLC
759
5 Human Machine Interface-(HMI)/DOP-B03S211 The field of automation has evolved over the past few decades, with humans playing a significant role in the service, troubleshooting, and repair of complex systems. A Human Machine Interface-(HMI) is a computer that permits the user to provide guidance/directions and collect input from the PLC, the feedback obtained to monitor the manufacturing process. HMI programming is also a development, since much of the time is spent modelling the interface on screens rather than coding. In comparison, the programming that can control the inputs and outputs of the HMI will usually live on the PLC, allowing the PLC programmer most control over the features of the HMI. They will show data such as temperature, pressure, process phases, and material count. They would also demonstrate very precise levels in the storage tank and also the precise location of the robot. Human-machine interface-(HMI) that provides operators with the way to manage machine command panels. The Interaction between the user and machine through a Graphical User Interface-(GUI) that facilitates information exchange the data between the User/Machine operators to know the information and process of the machine. There is a range of I/O options available such as the number of analog and digital inputs or outputs, and communication protocols ranging RS-232 links to advance protocols like Ethernetbased communications, SERCOS, and, CAN Open. We are using DOP-B03S211 is the HMI Display model Fig. 6 and the Delta Real-time Operating system-(OS). It has 32-bit RISC Microcontroller the frequency of the HMI Multi-Tone frequency (2 K–4 K Hz). The Operating Voltage has +24VDC (−10%–+ 15%) and the Operating Temperature of the HMI display (0 °C–50 °C). It has two Serial Communication Port/Serial COM Port, COM1 and COM2. And the COM1 have RS-232/RS-485 Serial port and COM2 have RS-422/RS-485 Serial Port.
Fig. 6. Human Machine Interface-(HMI)/DOP-B03S211
760
V. Priya et al.
6 Analytical Calution of Robotic Ladle The concept of a chapter to determine the orientation/direction and location of the endeffector within a fixed reference frame. In robotic manipulator, there are many ways to render a mathematical approximation to find the location and orientation/direction of the robotic manipulator end-effects in space. Axis of the Robotic Ladle and Rotation Angle are shown below in Table 3. Depending on the number of joints and connections, mathematics equations for the direction/orientation and location of the end-effector get longer and more complicated. There are two approaches used in the robotic manipulator for calculations: forward kinematics and inverse kinematics. Table 3. (Axis and rotation angle of robotic ladle). Axis of robotic ladle
Rotation angle
Base
(−120°)–120o
Shoulder
(−65°)–85o
Elbow
(−180°)–70o
Wrist-Roll
(−300°)–300o
Wrist-Pitch
(−120°)–120o
Ladle-Roll
(−360°)–360o
6.1 Forward Kinematics Forward Kinematics applies to the usage of the kinematics equation for the manipulator mechanism, which defines the orientation/direction and location of the end-effector of the robotic manipulator mechanism from the defined values. And the parameters of the joint according to the base coordinate system. In the analysis (Denavit-Hartenberg)-DH approach used to achieve a forward kinematics equation for the process used. 6.2 Inverse Kinematics Inverse Kinematics refer to the use of a manipulator kinematics equation to evaluate the joint parameter of the robotic arm that provides the desired position and direction/orientation values of the end-effector. This approach is relatively complex than the forward kinematics. 6.3 Denavit – Hartenberg The standard mathematics method used by the Denavit-Hartenberg parameter, also called the DH parameter, in the robotic manipulator to determine the direction and location of the end-effector. The DH parameter helps to define the constraint between the links and
Design and Simulation of a Robotic Manipulator for Ladle with PLC
761
the joints of the coordinate system for each link of the manipulator, a 4 * 4 transformation matrix connected with two consecutive coordinate systems. Using the coordinates systems in the corresponding link DH -parameters are listed below in Table 4, we obtain transformation matrices Parameters equation Fig. 7 Table 4. (D-H transformation)
0
A1 =
…. (6.1)
1
A2 =
…. (6.2)
2
A3 =
…. (6.3)
3
A4 =
4
A5 =
…. (6.4)
…. (6.5) Fig. 7. (Transformation matrices)
Multiplying Eqs. (6.1), (6.2), (6.3), (6.4), (6.5) and in Fig. 8 The Final Transformation Matrix that represents, 0
A5 =0 A1 ∗1 A2 ∗2 A3 ∗3 A4 ∗4 A5
762
V. Priya et al.
1
A4 = Fig. 8. (Multiplying equations)
Where 0 A5 is the Final Transformation Matrix for the End-Effector in the Cartesian Coordinates Fig. 9 and 10 was obtained as follows;
ux = u y= uz = vx = Fig. 9. (Cartesian coordinates)
vy = vz = wx = wy = wz = qx = qy = s qz = Fig. 10. (Cartesian coordinates)
In this chapter we used to Simulate the Robotic Manipulator for ladle in Rokisim Simulation software and the Cartesian Joa/Cartesian Coordinates are: (206.929 mm, −1247.789 mm, 2400.793 mm).
7 Ladder Logic Programming PLC-(Programmable Logic Controller) Ladder logic is one of the programming languages used to program PLC. The programming language represents logical notation operations in the ladder diagram program. Ladder logic software is very similar to the rungs and rails of the regular logic relay circuit. Used to perform logical, linear, timing, calculating, and arithmetic operations functions for the processing of industrial automation applications. Ladder Logic Programming Diagrams Fig. 11 can appear as if electrical schematics diagrams going vertical. However, there are some differences. Here are the most important differences:
Design and Simulation of a Robotic Manipulator for Ladle with PLC
763
• In ladder logic programthe line (rung) and executes then run the next lines of the rung. • In electrical system,multiple lines (current pathways) are sometimes executing (activated) an equivalent run time. Ladder logic program is still used in PLC programming since the basic logical rules for control machines and the mechanism are similar. This programming process is designed to specifically construct PLC programming in control circuits and relay-based logic. Although Ladder Logic Diagram programming remains popular in the automation industry. Ladder Diagram system looks like an electrical diagram/wiring diagram, making a graphical programming language and these diagrams a virtual circuit; “virtual power” passes through “virtual contacts” (when closed) to energize the virtual “relay coils” to execute logical functions. 7.1 Software Requirenments a) PLC Software: The tools used to program the PLC-(Programmable Logic Controller) WPL Soft 2.49 and the operating system-(OS) for Windows XP/Vistal/7(32-bit/64-bit)/8/10 (64-bit). The DVP series is supported by this program. b) HMI Software: The software used to program the HMI-(Human Machine Interface) DOP Soft 2.00.04.09.
8 Block Diagram The block diagram represents the simplify the principal of project. Here, the block diagram of the Robotic ladle using (Programming Logic Controller)-PLC is as shown in a Fig. 12 which is one of the simple examples of the whole working process of the robotic ladle. The block diagram has following represents parts like Power supply, SMPS, PLC, Input modules, Input sensors and switches, Stepper Motor Driver, Stepper Motor, HMI and SCADA, etc. Here PLC are the Main part of this Block diagram. The PLC requires 110 V or 220 V AC power supply to operate, PLC takes inputs from the switches and sensors. Which signals may be digital or analog. Also, it gives output to the relay module/output module here the PLC receives inputs and outputs as a 24VDC supply. The Relays module is an intermediate medium between the Motor drivers and the PLC. Relays are transfer the PLC output to the motor Driver. Moreover, the Driver is a part which runs the Robotics motor/Stepper Motor for the movement of the robotics joints. Drivers take input from the PLC via Relays as a 24VDC signal and gives the required output to the Driver and motors. Stepper motor forwarding, motor revising, to start and stop the motor, and the speed control the motor, etc. functions plays by the Motor Driver. The movement moves a particular limb of the robotic ladle. Here motors we used for the moving in particular axis and specific purpose. Also, the Stepper motor movements can be controlled by the sensors. Here the Sensors are detecting the position
764
V. Priya et al.
Fig. 11. (Ladder logic program)
of the robotics ladle and gives the sensors output feedback to the PLC. also SCADA(Supervisory Control and Data Acquisition) and (Human Machine Interface)-HMI and we are using for the graphical presentation of the robotic ladle to analysis the position of the robotic ladle and angle of the joints and links of the robotic manipulator for the ladle. Also, SCADA will provide a real-time data monitoring and control of the robotic manipulator for the ladle.
Fig. 12. (Block diagram for robotic manipulator for ladle)
Design and Simulation of a Robotic Manipulator for Ladle with PLC
765
9 Conclusion This chapter has presented the design, modelling, simulation and methods of an 5-DoF robotic arm for ladle. This robotic ladle consists of an industrial 5-Dof interface with PLC module. The definition is done with the aid of the PLC ladder logic system language that processes the 5-DoF movement of the robotic manipulator. The PLC command console is made up of the HMI Monitor support for distance and displacement speed Allows both start and stop of the stepper motor application on the panel and also controls the rotation of the robotic manipulator ladle. Before the development of this control system, any mechanism required by humans to manually move molten metal at an enormous temperature, the automation of this control system process discussed in this article. This presented system is to reduce time, eliminates the hazards to employees and decreases work effect of the labour, maintain the cycle time, pressure, flow of oil of the machine. And this automation control system process the machine accurately and repeatedly fills the molten aluminium metal into the mould/die and this system increase performance, efficiency, cycle time process of the machine and the productivity to produce the high quality of the casting products.
References 1. Momin, N.C., Patil, M.S.: Automatic Ladle Control with Job Database Using PLC-SCADA (2010) 2. Veena, C.D., Sharath, H.K., Rajendra, S., Shivashankara, B.S.: Design and fabrication of Plc and Scada based Robotic Arm for material handling (2018) 3. Myint, K.M., Htun, Z.M.M., Tun, H.M.: Position control method for pick and place robot arm for object sorting system 4. Shah, J., Rattan, S.S., Nakra, B.C.: End-effector position analysis using forward kinematics for 5DOF Pravak robot arm. Int. J. Robot. Autom. (IJRA) 2(3), 112–116 (2013) 5. Amritha Ashok, K., Savy, A., Shijoh, V., Shaw, R.N., Ghosh, A.: Hospital assistance robots control strategy and machine learning technology. In: Bianchini, M., Simic, M., Ghosh, A., Shaw, R.N. (eds.) Machine Learning for Robotics Applications. SCI, vol. 960, pp. 35–46. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0598-7_3 6. Patil, M.D.: Robot manipulator control using PLC with position based and image based (2017) 7. Tsai, L.-W.: Robot analysis: the mechaincs of serial and parallel manipulator (1999) 8. Barz, C., Latinovic, T., Erdei, Z., Domide, G., Balan, A.: Practical Application with Plc in Manipulation of a Robotic Arm 9. Kumar, M., Shenbagaraman, V.M., Shaw, R.N., Ghosh, A.: Digital Transformation in smart manufacturing with industrial robot through predictive data analysis. In: Bianchini, M., Simic, M., Ghosh, A., Shaw, R.N. (eds.) Machine Learning for Robotics Applications. SCI, vol. 960, pp. 85–105. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-0598-7_8 10. Anjankar, S., Jawarkar, P., Heda, L., Dahake, V.: Pick and place automation machine using plc and ladder programming 11. Aphale, D., Kusekar, V.: PLC based pick and place robot with 4 DOF 12. Bao, Y.: Design of manipulator PLC control system (2017) 13. Bhambere, A.: Design of 6-Axis robotic arm 14. Pachaiyappan, S., Balraj, M.M., Sridhar, T.: Design and analysis of an articulated robot arm for various industrial applications (2014) 15. Tayade, A.., Chitre, A., Rahulkar, A.D.: PLC based Hydraulic auto ladle system (2014) 16. Bogdan, L.: Programming and Controlling of RPP robot by using a Plc (2010) 17. Sarhan, H.: PLC-controlled stepper motor drive for NC positioning system (2014)
Design and Implementation of Automatic Goggle Detector for Safety Measure V. Balambica1(B) , T. R. Vijayaram2 , M. Achudhan2 , Vishwa Deepak2 , and Manikandan Ganesan3 1 Department of Mechatronics Engineering, Bharath Institute of Science and Technology,
BIHER, Chennai 600073, TN, India [email protected] 2 Department of Mechanical Engineering, Bharath Institute of Science and Technology, BIHER, Chennai 600073, TN, India [email protected] 3 Faculty of Manufacturing, Institute of Technology, Hawassa University, Hawassa, Ethiopia
Abstract. Automatic Goggle Detector is predominantly used in the Industries to detect whether the humans are wearing the goggles in workstations. There are no such innovations have been developed in this era to detect whether the humans are wearing the goggles or not. In this innovation we are using device which will be mounted on the goggle temples. This device will detect whether humans are wearing the goggles or not by using sensors. Moreover, this device will give a feedback to the machine if goggles are not used by the humans and it will stop the machine abruptly. This device has been developed to prevent accident which happens harming the eye. There are many critical operations in industries which lead to accident. For instance: Chemical companies, Spring winded parts, Pain shops, Gas emitting companies, oil companies, welding companies etc.. The goggles detector uses a temperature sensor MLX90614 to detect human body temperature. From that input the machine is controlled using RF & TX (Receiver and Transmitter) module. Moreover, the controller which is used to communicate is ARDUINO_ATMEGA328. Further, the temperature will be displayed in the LCD screen to cross verify the temperature. Nowadays, in many industries, there are no such devices to detect goggles in contrast there are some companies were they are using cameras to detect but the cost is too high. By implementing this device in the market many accidents can be prevented. Keywords: Temperature sensors · RF & TX module · Arduino Uno · Beeper · LCD screen
1 Introduction Goggles are obligatory in many industries to safeguard the eye. Although employers provide necessary personal protective equipment, many individuals are unaware about the significance of goggles; eventually it affects the eye owing to some inevitable reasons. According to the survey by many researchers, most of the accidents harming the eyes © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 766–773, 2022. https://doi.org/10.1007/978-981-19-1677-9_67
Design and Implementation of Automatic Goggle Detector for Safety Measure
767
mainly in factories. There should be some device to detect whether people are wearing goggles in the workplace. This was an opportunity provided todevelop a device which can be used anywhere and it shields the eyes from reportable or non-reportable accident. It will be helpful to use in industries such as that makes forging component, casting components, winded spring, machining shops and more on. The gadget which was developed is wireless, so it can be used without any issues to the operator. It will stop the machine abruptly if operator does not wear the goggle.
2 Literature Review Although the goggle detector plays an vital role in the society, it has not been invented earlier times. It was cross verified with some types of goggles which people used for disparate methodology. [1] NT le, in 2019 innovated goggles for ski diving and ice skaters which used to identify water particles, dust, snow and spotlight. In the year 2017 BL Jian [2] introduced night vision goggles which can detect the honey comb during night, it is mainly used for inspection purpose and it is used identify the minute particles clearly. In 2019 N Das [3] introduced goggles which draws affinity diagram automatically and it identifies the programs which others used in IT field, these are done by software coding In 2004 GW HO [4] innovated night vision goggles which assists to operate in dark and dim light conditions. RS Allison [5], in the year 2016, invented airborne optical goggle which utilize to detect the wildfire in forests by using image and thermal capturing sensors. In 2004 U Svensson [6] invented goggles which detects the eye blink pattern and inform whether the driver is sleeping or not while he or she is driving a car. Further, Venk Sanj [8], in 2017 introduced electronic goggles which used for blind peoples, by using these devices people can walk independently without any others assistance, the sensors which they have used is UV & IR sensors. Then in 2013 [9] Liu, invented the Non-Infrared goggle system that which detects the lymph in the human body and it’s mostly used in the hospitals to detect the cancer. In 2019 CJ de Nauois [10] had innovated the spectacles which are worked with artificial neural network, it not only detects the driver’s activeness but also check whether they are sleeping or not. Whereas in 2014 [11] D Wu found the spectacles with multi sensor image fusion system, it take numerous picture in a same time and give the desired output in a prompt manner. In 2013 Sanjvenk [12] invented goggles by using UV sensors which gives clear views to the eyes and used to detect issues in the welding process. CJ de Mark [13] in 2019 innovated device which identifies the external surface of the vehicle and it inform the chauffeurs before the accident happens, the sensor used is distance measurement sensor. Also, in 2013 CJ. Dickson [14] created solar welder eye glasses which used in applications for mainly welding. In addition to these entire image tagging device was introduced by Steve [15] in June 2018 which use to show the image of the people like video call process
3 Methodology Brief Working Scenario of the Prototype: The above two pictures are the transmitter and receiver modules. The significant controller unit in this module is Arduino ATmega328 which controls the device to pass the input and output signals. The whole
768
V. Balambica et al.
module consists of electronic systems like temperature sensor, LCD display and battery to power up (Fig. 1).
Fig. 1. Transmitter and a receiver module
Design and Implementation of Automatic Goggle Detector for Safety Measure
769
3.1 Transmitter and Receiver Module The transmitter kit is 433 MHz and the receiver kit for Arduino is ARM MCU and it is totally wireless. The description about the receiver module models is XD-FST and the total distance it covers is upto 200 m. It totally vary upon its operating voltage, the voltage of direct current is 5 V. However, it operate between 3.5–12 V and its frequency is 433.92 MHz. Moreover, the transmitting frequency is 433 M and its pin out from left to right (DATA – VCC – GND) (Fig. 2). 3.2 Temperature Sensor: MLX90614
Fig. 2. MLX90614
3.3 General Description The MLX90614 is a non contact module Infra-Red thermometer sensor user to measure temperature. The sensor is low noise amplifier, it has 17-bit ADC and it has the most powerful DSP unit, with these units the thermometer can show the follow perfectly. The temperature value is varies from −20 to 120 °C, and its output resolution is 0.14 °C (Fig. 3).
Fig. 3. Arduino_ATMEGA328
770
V. Balambica et al.
3.4 Product Overview The Arduino Uno is a microcontroller which has ATmega328 chip. It consists of 14 digital input/output pins, further, in these fourteen pins six can be utilized as PWM output. In addition to that it has 6 analog inputs, and it has a 16 MHz crystal oscillator. It has the USB connection where it can used to charge as a power jack and it can be connected with PC to send and receive programs. The board can also be charged by using adaptor and the voltage should be less than 12 V.
4 Working Principle of Goggle Detector The model version consists of a Goggles, Arduino Uno, LCD display screen, buzzer, Transmitter, Receiver, and Temperature sensor. To initiate the process first the receiver side - Arduino should be plugged with adaptor, the Arduino board will be powered on. In addition to that goggles temperature sensor should be powered with 9 V battery. Now, the temperature senses the human body temperature and gives the analog input to the receiver. The receiver will collect the data from it and it will transmit it to Arduino. Coding was done in UNO board to compare the value of temperature sensor. If the temperature is less than 93 °C the UNO will not give the output signal, it presume that goggles is not wore by humans. However if the temperature is between 93° to 100° the UNO board will give output signal. Moreover, if the temperature is above 100° it sends the message that human body temperature is high. Now if the human wear the goggles the temperature will be sent and it will confirm that human is wearing goggle and the message will be displayed in LCD screen. Nevertheless, if the human was not wearing the goggles, the temperature will be low and then automatically the buzzer will sound (Fig. 4, 5 and 6).
Fig. 4. Fully designed view:
Fig. 5. Prototype module transmitter:
Design and Implementation of Automatic Goggle Detector for Safety Measure
771
Fig. 6. Prototype receiver module
5 Results and Discussion This module can be used in disparate environment. It was tested in many other industries and the device depicts accurate results. This can be implemented in all companies to avoid accidents harming the eyes of an operator. The cost of the device is at an affordable price (Table 1). Table 1. Various levels of temperature testing Ambient temperature in degree
Human temperature in farenheit
Result
Remarks
88
93.5
OK
Human wearing the goggles
85
95.8
OK
Human wearing the goggles
84
89
NOK
Violation – Human not wearing the goggles
89
98.6
OK
Human wearing the goggles
90
101.50
NOK
Caution – Body temperature High
88
95.5
OK
Human wearing the goggles
The module prototype has got some limitations for an user as the primary drawback of the Temperature sensor output varies from one person to another and the consistency of the sensor is poor. More than more than 3 sensors were changed to achieve the desired output. Moreover, earlier the sensor only gives desired output when it touches the human body; however, now the module sensor being changed MLX90614, it any human body can be sensed from a distance of 5 mm and the desired output was achieved. The comparative analysis of this module is elucidated. a) Temperature Sensor detection ranges from 0 to 5 mm. b) This sensor can detect any object which is between 0 to 5 mm and its check the temperature of the object which front of it.
772
V. Balambica et al.
c) The voltage used to power up the temperature sensor ranges from 5 v to 9 v. Further, it can withstand up to 12 V. d) The communication between transmitter and receiver varies while we are testing, then we have fixed it ranges up to 3 m distance. e) While we are testing the goggles in industries which has high heat (furnace industries) the temperature and sensor varies according to the atmospheric temperature. We had taught the sensor according to the room temperature and defined the correct value. f) On real time while we are communicating the goggles with PLC the communication lags more than 2 s, then we have reduced is distance, thereby, reducing the lag time. g) After all the corrections the transmitter and receiver worked efficiently without any malfunction.
6 Conclusion and Future Scope To conclude, from the aforementioned points, it is evident that goggles detector will detect precisely whether mortals are wearing goggles. The device is advent for many industries where they faced reportable and non-reportable accidents. Ergo, I strongly confide that, this innovation can be used in all the industry wherever the goggles are used. To exemplify a few industries, molding, fabricating, welding, and winded spring making factories, cement, brick and many more.
References 1. Le, N.T.: Novel framework based on HOSVD for ski goggles defect detection and classification. MDPI (2019) 2. Jian, B.L., Pen, C.C.: Development of an automatic testing platform for Aviator’s night vision goggle honeycomb defect inspection, pub med. (2017) 3. Das, N., Chaba, S., Wu, R.: Goggles: automatic image labeling with affinity coding. Sigmod 20 (2019) 4. Parush, A., Gauthier, M.S., Arseneau, L., Tang, D.: The human factors of night vision goggles: perceptual, cognitive, and physical factors. Rev. Hum. Factors Ergon. (2004) 5. Allison, R.S.: Airborne optical and thermal remote sensing for wildfire detection and monitoring. MDPI J. (2016) 6. Svensson, U.: Drowsy driver detection system using eye blink patterns. IEEE Explore Digital Library (2004) 7. Das, N.: GOGGLES: automatic image labeling with affinity coding. ArXiv J. (2019) 8. San, V.: Be my third eye - a smart electronic blind stick with Goggles. Int. Res. J. Eng. Technol. (IRJET), 17 June (2017) 9. Liu, Y.: Near-infrared fluorescence goggle system with complementary metal–oxide–semiconductor imaging sensor and see-through display. Pubmed (2013) 10. Rajawat, A.S., Rawat, R., Barhanpurkar, K., Shaw, R.N., Ghosh, A.: Sleep Apnea detection using contact-based and non-contact-based using deep learning methods. In: Bansal, J.C., Paprzycki, M., Bianchini, M., Das, S. (eds.) Computationally Intelligent Systems and their Applications. SCI, vol. 950, pp. 87–103. Springer, Singapore (2021). https://doi.org/10.1007/ 978-981-16-0407-2_7 11. Mandal, S., et al.: Single shot detection for detecting real-time flying objects for unmanned aerial vehicle. Artif. Intell. Future Gener. Robot. 37–53 (2021). https://doi.org/10.1016/B9780-323-85498-6.00005-8
Design and Implementation of Automatic Goggle Detector for Safety Measure
773
12. Sanjvenk: Welding goggle auto darkening with elitevision technology unimig U21001K (2013) 13. de Mark, C.J.: Method and system for detecting objects external to a vehicle (2019) 14. Dickson, C.J.: Welder eyes glasses, solar powered auto darkening welding TIG MIG goggles welder eyes glasses (2013) 15. Rajawat, A.S., Barhanpurkar, K., Shaw, R.N., Ghosh, A.: Risk detection in wireless body sensor networks for health monitoring using hybrid deep learning. In: Mekhilef, S., Favorskaya, M., Pandey, R.K., Shaw, R.N. (eds.) Innovations in Electrical and Electronic Engineering. LNEE, vol. 756, pp. 683–696. Springer, Singapore (2021). https://doi.org/10.1007/978-98116-0749-3_54 16. Balambica, V., Anto, A., Achudhan, M., Deepak, V., Juzer, M., Selvan, T.: PID sensor controlled automatic wheelchair for physically disabled people. Turkish J. Comput. Math. Educ. 12(11) (2021). e-ISSN 1309–4653 5848–5857 17. Ganesan, M., Loganathan, G.B., Dhanasekar, J., Ishwarya, K.R., Balambica, V.: Implementing industrial robotics arms for material holding process in industries. J. Harbin Inst. Technol. 53(9), 17–27 (2021). ISSN: 0367–6234
Machine Learning Based Approaches in the Detection of Parkinson’s Disease – A Comparative Study Arpan Adhikary1 , Koushik Majumder1(B) , Santanu Chatterjee1 , Rabindra Nath Shaw2 , and Ankush Ghosh3 1 Department of Computer Science and Engineering, Maulana Abul Kalam Azad University of
Technology, Kolkata, West Bengal, India [email protected] 2 Bharath University, Chennai, India 3 The Neotia University, Sarisha, West Bengal, India
Abstract. Parkinson’s disease (PD) is one of the enfeeble diseases that is incurable. It occurs when dopamine is not properly produced in substantia nigra of our central nervous system. Dopamine is responsible for our movement and activities like speaking, walking and writing. In the recent papers, researchers have found that speech, Freezing of Gait (FOG), writing problems are the most common symptoms for PD. However, most of the people who suffer from Parkinson have speech and FoG disorder. With the increase of severity, some other symptoms also arise gradually. Some works have been done by the researchers in early prediction and detection of PD using Machine Learning (ML) and Deep Learning (DL) classifiers. In this paper a detailed survey of the different works in this area has been carried out. We identified certain drawbacks of the existing works. Based on this, we tried to explore the future scope of works which can enhance the performance of the existing works in early detection of PD. Keywords: Parkinson’s Disease (PD) · Machine learning · Deep learning
1 Introduction Parkinson becomes a very common nerve and brain disorder in the recent past. People above 60 are mainly being affected by PD. Depending upon some external scenarios, people of different ages are also suffering from this disease. Some basic symptoms of PD are tremoring, slowness of movement, rigidity or stiffness, speech disorder. But, before the symptoms for PD appears, there exists a pre-motor phase that can last from 5 to 20 years before classic clinical motor symptoms arise. In that time, a patient has symptoms like Sleep Behavior Disorder, rapid eye movement, olfactory loss etc. PD is a very slow progressive disease. There is no particular reason found yet, for which PD can happen. It can be genetic or if a person is exposed to pesticides or medicines like reserpine or phenothiazine, then also PD may arise. No medicines or surgery till not found to permanently cure PD. So early detection can prevent a patient from more © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 774–793, 2022. https://doi.org/10.1007/978-981-19-1677-9_68
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
775
harm. Manual detection of PD is more time consuming, and it may suffer from erroneous detection. To overcome this problem, we need some automated system, which is less time consuming and less erroneous. By Machine learning and deep learning technique, we can solve this problem. We can take the symptoms as input parameters to train the model.So thatthe model can learn about PD symptoms and differentiate PD patients from healthy people. It also takes less time for analysis as the parameters are already known. We can also differentiate between the severity of PD by separating the input parameters into different clusters. Both motor and non-motor features and pre-motor stage data were taken by the researchers. Different techniques like RF, different SVM kernels, AdaBoost, CNN architectures like LeNet and AlexNet have been used by the researchers for predicting Parkison’s disease.
2 Related Works In the recent past, many researchers took ML and DL classifiers into consideration to classify PD. Most popular features taken for the test were Freezing of gait and speech disorder data. For this data was collectedfrom patients’ walking test and vowel pronunciation. Some of the related works in this area are discussed here. Juutinen et al. [1] compared three feature selection and nine classification methods to identify PD patients from control subjects based on accelerometer and gyroscope signals measured with a smartphone during a 20-step walking test. kNN gave them the highest accuracy of 84.5%, sensitivity of 88.5%, specificity of 81.3% and an error rate of 15.5%. Prashanth et al. [2], showed that the combination of non-motor symptoms like hyposmia and sleep behavior disorder, cerebrospinal fluid measurement and imaging markers are more useful in PD diagnostics. They used Parkinson’s Progression Makers Initiative (PPMI) database and after the statistical analysis of the features, they took NB, SVM (LIBSVM library), Boosted trees, RF and Logistic Regression classifiers for classification. Rahman et al. [3] took speech sample as speech disorder is one of the main symptoms for PD. It is also helpful in tele-diagnosis and telemonitoring of PD patients. They used signal processing algorithm i.e., PLP and RASTA-PLP for feature extraction. Four different types of SVM kernels were used to classify the extracted features. SVM with RBF kernel gave them 74% accuracy when the feature was extracted using PLP. Ortiz et al. [4] have taken the data from PPMI (Parkinson’s Progression Markers Initiative, RRID: SCR_006431) which is an observational clinical study to verify the progression makers in PD. They used two very popular CNN architecture LeNet and AlexNet with average accuracy of 95.1% and AUC of 97% where LeNet architecture comprises of 7 layers and AlexNet comprises of 12 layers. Jindal et al. [5], in their study used an open database: a digital image tablet dataset from Parkinson Spiral Drawings. They used CNN for PD identification. Here AlexNet CNN architecture was used after the data collection part. Their result shows 96.5% precision. Tiwari et al. [6] have taken their dataset from UCI machine learning repository. Their work indicated that the ensemble technique XGBOOST classification algorithm
776
A. Adhikary et al.
achieved the highest test accuracy rate of 95% and training accuracy rate of 100% while classifying people with PD from the healthy people. For this, they used different types of prediction models i.e., Logistic Regression, DT, SVM, kNN, XGBOOST and Bagging techniques. Arora et al. [7], have aimed to distinguish PD patients accurately using selfadministered tests of postural sway and gait from healthy people by using smartphone accelerometer data. Random Forest classifier gave them the best result of 98.0 (±1.1)% of Balanced Accuracy. Freezing of Gait (FoG) is one of the major causes of PD. Mazilu et al. [8] used smartphones and wearable accelerometers for detecting FoG through their application. The system provides auditory feedback and a vibration alert when FoG is detected. They took Random Trees (RT), RF, DT, Pruned Decision Tree, NB, Bayes Nets, kNN with one neighbour and two neighbours, MLP, bagging and boosting and achieved the best sensitivity, specificity, F1 score and accuracy for AdaBoost of 98–99%. Tripoliti et al. [9] used six accelerometers and two gyroscope sensors for FoG detection of PD patients. Firstly, the missing values were removed which were generated due to signal loss and then low frequency components were removed and entropy was calculated. They have classified the final result using Naïve Bayes, Random Forest, Decision Trees and Random Trees classifier. They have achieved 96.11% accuracy, 81.94% sensitivity and 98.74% specificity using the Random Forest classifier. Elazari et al. [10] compared the motor symptoms of PD in patients and healthy people. They also differentiated mild PD patients from the severe patients by using body-fixed sensors. Their system detected PD patients from healthy people with an accuracy of 92.3%, mild from severe patients with 89.8% accuracy and mild patients from healthy people with 85.9% accuracy. Zhan et al. [11] implemented a smartphone application to track the motor symptoms of PD. They used Random Forest classifier with 500 trees with splitting criteria based on gini index. The patients were tracked before and after their medication dose. They passively monitored the patients for over 46000 h to collect the data and obtained 71.4(±4)% accuracy. Patel et al. [12] took the accelerometer data to estimate the severity of PD symptoms like tremor, bradykinesia and dyskinesia and also the motor complications in PD patients. They compared different SVM kernels to estimate the severity of PD by applying their trained model on the videos of patients’ activity. Polynomial kernel gave them the best accuracy with respect to other kernels. Unified Parkinson’s Disease Rating Scale (UPDRS) is used for comparing severity of PD as it is a widely used Parkinson standardization scale. Nilashi et al. [13] used incremental SVM method to predict motor-UPDRS and total-UPDRS. For data dimensionality reduction they have used iterative non-linear partial Least Square method and for clustering, they have used self-organizing map. The authors predicted the accuracy by Mean Absolute Error for both motor-UPDRS and total-UPDRS. They achieved the accuracy 0.4967 for motor-UPDRS and 0.4656 for total-UPDRS. They have compared the result with neural network, multiple linear regression (MLR), adaptive neuro-fuzzy interface (ANFIS) and SVR on the same dataset.
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
777
Almeida in [14] processed voice data signal to detect PD. They used smartphone and acoustic cardioid microphone to record speech samples, mainly for vowel “a”. They have taken 5 metrics into consideration to analyze the performance of classification, mainly EER, AUC and accuracy. They used kNN, MLP, OPF and SVM for their research and obtained 92.94% accuracy when the data was collected using smartphone and 94.55% accuracy when the data was collected using microphone. Sriram in [15] used voice samples from UCI machine learning repository. For statistical analysis, classification and evaluation they used Orange v2.0b and Weka v3.4.10 software and for classification they have used Bayes Net, NB, Logistic Regression, ADT, RF, kNN and SVM. SVM gave an accuracy of 88.9%, RF gave an accuracy of 90.26% and NB gave an accuracy of 69.23%. Pahuja in their study [16] used voice dataset from UCI machine learning repository and classified the same using machine learning algorithms for early detection of Parkinson. They compared MLP, ANN, SVM and k-NN to see which algorithm is more accurate in classifying PD. ANN with Levenberg-Marquardt algorithm gave the highest accuracy of 95.89% and SVM with RBF kernel gave an accuracy of 88.21% after that. Shi et al. [17] developed a fall detection algorithm with tri-axial accelerometer and magnetometer to detect falls as it is also a motor symptom of Parkinson. In their algorithm, they used Signal Vector Magnitude base length, post-impact velocity and peak value to detect falls using daily activities. They asked the subjects to perform walking, jumping, going upstairs and downstairs to take the data and got better accuracy and specificity than similar previous works. Ali et al. [18] used handwritten drawings to detect Parkinson. Their main goal was to reduce biasness due to imbalanced data and for this they proposed random under sampling method for training process balancing. Another problem was low classification accuracy and to overcome this problem, they proposed cascaded learning system which cascades Adaboost with Chi2 model. Chi2 model is used to select and rank some relevant features whereas Adaboost is used to predict Parkinson using those features. They obtained 76.44% accuracy, 70.94% sensitivity and 81.94% specificity. Abdulhay et al. [19] used gait and tremor as their clinical data to distinguish patients with PD from normal subjects. In their paper, they broke down various phases of gait with different periods which can differentiate normal gait from gait with Parkinson. They developed a pulse duration algorithm that can extract different types of temporal features of gait. They obtained an average accuracy of 92.7% and best accuracy of 94.8% with Medium Gaussian SVM, which was used to determine Parkinson severity. Wang et al. [20] observed the PD pre-motor stage symptoms, i.e., Rapid Eye Movement, Sleep Disorder, Olfactory loss, Dopaminergic Imaging Markers and Cerebrospinal Fluid data and applied different ML and DL techniques to classify Parkinson. They have achieved the best accuracy of 96.60% using deep learning method but algorithms like SVM, kNN, Random Forests also gave an average accuracy of 95.5%. Sample Heading (Forth Level). The contribution should contain no more than four levels of headings. The following Table 1 gives a summary of all heading levels.
778
A. Adhikary et al.
3 Discussion 3.1 Methodology Juutinen et al. [1] collected data by using smartphone inbuilt sensors which was a great idea and a step towards telemedicine and telephonic check-ups for remote patients. They collected the data from Satakunta Hospital District, Pori, Finland during 2018. For data collection, the patients were asked to walk indoors for 20 steps. Though this test was done twice, some of the tests were unsuccessful since the patients had taken an extra step after being asked to stop. At pre-processing step, they selected only the recordings which included both accelerometer and gyroscope data. These data were synchronized by removing the excessive parts of data and sampled at 20 Hz frequency. The main goal of the researchers was to detect each step and to reject the inactive periods of walking from start and end of the test. From the three-axis measurement, the researchers detected the beginning of each step using vertical signal and then inactive periods of walking was distinguished from the acceleration signal. Two segments were measured in this step, the time taken from minimum to maximum stride amplitude and vice versa. After calculating the mean and segmenting individual steps from the walking tests, they calculated statistical features for each step. They compared mRMR, SFS and SBS for feature selection methods. For classification part, they took nine different machine learning algorithms into consideration as mentioned earlier. Using kNN, they received the best accuracy of 84.5% and 15.5% of error rate (Fig. 1).
Fig. 1. Workflow of PD classification using mobile sensors [1]
Before PD occurs, the patients face pre-motor stage symptoms where some nonmotor features like rapid eye movement, sleep disorder and olfactory loss occur. This can last for 5 to 20 years. Prashanth et al. [2] took the non-motor features and cerebrospinal fluid measurement from PPMI database (Reference) to detect early PD. They used UPSIT to evaluate RDB screening questionnaire and olfactory dysfunction tests. For the study the researchers took 183 normal subjects and 401 early-stage PD patients. Then the statistical analysis of the features was done, where the subject smelled a scent and marked the option in smell test. Then for sleep disorder screening the subjects were given some questions, which they had to answer with yes or no, and the high scorer had the most chance of having sleep disorder. Here the researchers compared the results from NV, Logistic Regression, SVM, boosted trees and random forests classifiers. The dataset was divided in 7:3 ratio where 70% data was used for training and the rest was used for testing. They iterated 100 times to compute the accuracy, sensitivity and specificity along with the average. Though all the classifiers performed very well but SVM gave the best accuracy of 94.60% and 98.88% AUC. It was observed from the plots that SVM
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
779
and LR gave best performance in accuracy, RF and SVM gave best performance in AUC and in specificity, and the performance of SVM was best. They used the combination of the three pre-motor features, CIF and dopaminergic imaging markers, which boosted up the accuracy (Fig. 2).
Fig. 2. Flowchart of PD detection from pre-motor stage symptoms [2]
Speech disorder is one of the many symptoms of PD. Rahman et al. [3], took large voice samples with different vowel phonation to detect the PD patients. They collected the data from a medical teaching institution in Pakistan. Different types of vowel phonation were used, and the data collection was done using smartphones by asking the patients to pronounce the vowels “a”, “o” and “u”. They took 29 PD and 30 healthy peoples’ data for their work. It was observed that PD patients face uneasiness while spelling “a” correctly. They used smartphones to collect the data keeping the device 10 m away from the subjects. For pre-processing, PLP and RASTA-PLP algorithms were used. Firstly, they tuned the signals by reducing the signal breaks and smoothing the ends of each signal. Then they used hamming window to tap the signals to zero for each signal at the beginning and the end. Then in critical band spectral resolution, they wrapped the shortterm power spectrum as well as it’s frequency axis and pre-emphasized the samples. Finally, the data was modelled and RASTA-PLP was used to increase the robustness of PLP algorithm. Using this algorithm, they calculated the critical-band colour and estimated the temporary derivative of log-deduction band spectrum. For this they took regression line into consideration. Just like PLP, they pre-processed the data non-linearly, plotted the graph and retrieved the audio spectrum co-related with it. SVM gave them an 74% accuracy, 70% sensitivity and 77% specificity when the feature was extracted using PLP and SVM model was used by RBF kernel. When RASTA-PLP along with MLP kernel was used, they got 68% accuracy, 70% sensitivity and 65% specificity which is quite low in terms of classifying PD (Fig. 3). Automatic recognition of image patterns to characterize and detect PD using CNN is very useful in PD detection. Ortiz et al. [4], used 3D brain images from PPMI database (Parkinson’s Progression Markers Initiative, RRID: SCR_006431) which is an observational clinical study to verify the progression makers in PD. The main focus was to reduce the amount of data from 3D images and to use only the relevant information by extracting the features using isosurfaces. Then using spatial normalization, they removed the
780
A. Adhikary et al.
Fig. 3. Method for PD detection using Signal Processing [3]
differences in shape and size of the brain images and then pre-processed it. To minimize the system complexity, they reduced the image size. Subjects with dopaminergic activity were chosen for this study. Intensity normalization was used to ensure that the patients had taken similar drugs and this could be measured from some indirect measure of neuro physical activity. For feature extraction, they used isosurfaces where interpolation was used instead of thresholding. Classification results using the isosurfaces computed with different thresholds were compared with each other and then the best classification result was chosen. Both forward and backward selection were taken into consideration for this. CNN was used for classification and two architectures were used, one is LeNet and another is AlexNet, where both of these networks were given the pre-processed data. Classification performance was evaluated by accuracy, sensitivity and specificity. From these, the ROC curve was computed and k-fold cross-validation was done on the classification result. They obtained an average accuracy of 95.1% and AUC of 97% with 7-layer LeNet architecture and 12-layer AlexNet architecture (Fig. 4).
Fig. 4. Methodology for PD classification using brain image classification [4]
Image dataset gives researchers more features in finding PD. Jindal et al. [5] took a spiral drawing dataset to detect PD, where they collected the dataset using a tablet and a computerized pen. They showed three injuries on the board where Archimedean spiral was shown. Then some red dot was given on the screen, and the patients were asked to hold the pen on some certain red dot that was appearing on the screen. The main objective was to see the shaking of hand of the patients. They recorded the data in 110 Hz as well as in 140 Hz but to receive consistent details, they checked the signals at specific rate of 110 Hz. They used CNN for PD identification in two parts: extraction and arrangement of layers that were completely linked with each other. CNN was divided into two sections. The first one had two coevolutionary layers which consisted of 16 channels. Here AlexNet was used by the researchers. The cross-approval was done using five subsets of layers where CNN was designed in 4 out of 5 subsets and fifth feature was used to validate the structure. This choice was altered frequently, and the experiments were also modified accordingly. They used 4 folds for preparing the split where 3 folds were used to prepare the loads and one-fold was used for adjustment. They obtained a precision of 96.5%, F1-score of 97.7% and region of 99.2% which is higher than the previous studies done by other researchers. Use of drawing sketches as input parameters gives a less expensive method of PD classification (Fig. 5).
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
781
Fig. 5. Methodology for PD detection using signal processing [5]
Tiwari et al. [6] took different ML and DL technique into consideration to predict PD. UCI machine learning repository dataset was taken for their workings which contains the voice data of PD and healthy people. 75% of the data was of PD patients and 25% was of healthy people. The dataset consisted of biomedical voice measurement in ASCII CSV format. They analysed fundamental frequency, shimmer, and jitter of different vocal samples. Then the samples were categorized into maximum, minimum and average frequency. In feature extraction, the researchers took vocal frequencies, fundamental frequencies, amplitude variations, noise to tonal component ratio, nonlinear measurements of fundamental frequency variations and status of disease to find whether subject had PD or not. Then feature importance analysis was done and corelation between the features was defined. For prediction they used the ML techniques like LR, DT, SVM and k-NN classifier and Deep-Learning techniques like Bagging and xgboost. xgboost performed better than the other algorithms (Fig. 6).
Fig. 6. Methodology for PD classification [6]
Arora et al. [7] differentiated PD from healthy people by some self-directed assessments of gait and postural sway. The goal was mainly to re-produce the clinical assessment accurately using the acceleration time series for the two most common PD symptoms mentioned above. They recorded tri-axial acceleration and then extracted different features from it. The pattern of the PD patients’ movement was taken into consideration to differentiate PD patients accurately. For data collection they took advantage of inbuilt sensors in smartphone: accelerometer and gyroscope. All the participants were asked to walk 20 steps forward, turn back and return to the initial position for gait test. For postural sway test, the patients were asked to stand upright without any support and this test was repeated 4 times per day, i.e., before taking medication in the morning, one hour after the dose, in the mid-afternoon and before sleep. The features were then extracted
782
A. Adhikary et al.
from the accelerometer based on time series and frequency-domain properties. Some of the metrics were not recorded in their study: stride length, minimum foot clearance and trunk flexion. Lomb-Scargle periodogram was used in this work to extract the frequencybased features. 30 features were extracted from accelerometer time series. Some features were extracted from tri-axial spherical transformation of accelerometer. Here, RF was used to avail the output from the signals. To investigate the accuracy, Random classifier and Conditional random classifier was used. For cross-validation, they used 10-fold cross-validation technique with 100 repetitions. 90% of the data was used to train the model and the other 10% of data was used to test it. Using the RF classifier, they obtained an average sensitivity of 98.5% and an average specificity of 97.6% (Fig. 7).
Fig. 7. Methodology for distinguishing PD from healthy people [7]
Mazilu et al. [8] introduced an online FoG detection system using wearable accelerometer sensor and smartphone. They built a system that took the internal accelerometer data from smartphone as well as from external sensors by the help of Bluetooth. Their work was separated into two different classes. First, an offline FoG detection classifier was built and second, a real-time FoG detection app was developed. The researchers used the DAPHNet dataset for this work. The sensors were attached above the ankle, the knee and on the lower back. They asked the subjects to walk forward, turn back and walk backward. Random walk test was done in a hall with 360° full turn and random stop. The test was done in patients’ own space for taking the daily activity result. These recorded data were pushed into queue for the window-based classification. They read the data in a fixed time interval for pre-processing. Mean, Standard Variation, Variance and Entropy were computed for FoG detection. An online FoG detection system was built based on the offline classifications and the mobile application for real-time FoG detection. Offline FoG detection classifier was fed to the collected labelled data and the model was trained using it. Firstly, the researchers took the most invidious features to pick out FoG from the normal gait subjects. Based on this, the model was trained offline. For online detection, the mobile application was given the same classifier that was used to train the model offline. It took the recorded data, pre-processed, and extracted the features that were necessary for FoG detection. If FoG was detected, it provided the feedback to the user. Using AdaBoost and Random Forest classifier they achieved almost same result of 98–99% but Random Forest classifier is more complex compared to AdaBoost (Fig. 8). As gait is one of the main symptoms for PD, Tripoliti et al. [9] took wearable sensors into consideration for detecting the freezing of gait event. 6 accelerometers and 2 gyroscope sensors were used here. The sensors were placed in different body parts of the patients: right wrist, left and right legs, chest and waist. Two gyroscope sensors were also placed on the chest and waist. The data was collected using Bluetooth and a portable
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
783
Fig. 8. Classification process for online FoG detection [8]
storage device. This experiment was done with 16 subjects from which 5 were healthy, 5 were PD patients with gait symptoms and 6 without gait symptoms. They asked the patients to perform some tasks like lying on the bed for 5 min, standing up from the chair, walking 5 m of distance, opening or close doors, walking to the corridor, walking back to the room, drink water and walking back to the chair. The data was verified twice, while recording and during the data collection process. They took linear interpolation to perform the imputation due to data loss and different starting time of the sensors and to remove the low frequency components. In feature extraction, sliding window was taken which can correctly identify all FoG events. FoG identification was done based on two factors. Entropy was used to detect gait event as it is non-linear in nature, and differences between the normal people with the PD patients was observed from the plotted histogram. The output of feature extraction was a vector with 3N features, where N is the number of sensors used for the analysis. To detect the FoG events the researchers used different classification algorithms like NB, DT, RT and RF. For data obtained from waist, NB, RF and DT gave the best result. For data obtained from chest, RF, DT and RT algorithm gave the best result. They got 96.11% accuracy, 81.94% sensitivity and 98.74% specificity using RF classifier (Fig. 9).
Fig. 9. Methodology for FoG detection [9]
784
A. Adhikary et al.
As body-fixed sensors are the most acceptable method for PD detection so Elazari et al. [10] took a single body-fixed sensor into consideration. These sensors were attached to patients’ body for few days to detect the activity and record the motor symptoms like tremor, FoG, postural control, turning ability, fall risk etc. But they excluded the patients with brain surgery, any cerebrovascular accidents, major depression, dementia or who had major head injury. They divided the PD samples into two categories, stage I and II for mild PD and stage III and IV for severe PD. They evaluated gait using tri-axial body sensor which was placed on the subjects’ lower back. Gait speed while daily walking, and speed of performing some dual tasks were checked. These dual tasks were given to test daily walking in challenging conditions. This task was done in lab and after lab assessment, the patients still wore the sensor for 3 more days to detect the daily activity of the patients. Then the transition between walking and sitting was determined after detecting the activities. Then they determined the starting and ending point of each transition by selecting the maximum and the minimum points of the signal. Data was analysed based on the duration of the task, the smoothness of the task and amplitude. Temporal features included range, duration, jerk and standard deviation. SVM was used to differentiate the entire features. In every round, a weak learner was added to the ensemble and the weighting vector was adjusted. Classification was repeated leaving one sample out at any given time. Using multivariate linear regression model, they compared PD patients with healthy adults and mild PD patients with severe PD patients. The outcome that they got using linear regression was applied to SVM to assess the feature sets’ ability to distinguish between the groups. Then they determined the accuracy, specificity and sensitivity to compare each group i.e., mild to severe PD and PD to healthy people. They achieved an accuracy of 92.3% for patients to healthy people, 89.8% for mild to severe patients, 85.9% for mild patients to healthy people. But the accuracy decreased when only the lab-based measurements were used (Fig. 10).
Fig. 10. Methodology for FoG detection using single sensor [10]
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
785
As remote monitoring system is now having a great impact on telemedicine and teletherapy, many researchers are focusing on building remote monitoring systems using ML algorithms. Zhan et al. [11] built a remote-monitoring system using smartphone where they actively and passively collected the data from their subjects. Data was collected actively at the time of clinical check-ups. Passive data was collected in background to train the model more accurately. Here phone usage logs, GPS location, Wi-Fi parameters were taken to measure the movements and social behaviour of the subjects. From the collected data they choose five PD symptoms - voice, balance, gait, dexterity and reaction time. Firstly, they implemented the application in the subjects’ smartphone. The application was a fully automated system to capture and compress the data. Then the data was encrypted and uploaded to the server storage. An option was given to the subjects to take self-reported evaluation. For feature extraction, they took only active test data. For voice data extraction, they divided 20 s voice sample into 40 data frames in 5 s each. Dexterity features were extracted based on the stay duration (i.e., for how long subjects touched the smartphone screen), move duration (i.e., that was the time interval between the one finger release to another finger touching the screen) and the reaction feature mainly focused on the lagging of the finger reactions. Collected active dataset was then used to detect the dopaminergic medication response. Here they used RF classifier to represent whether the tests were performed before medication or after it. In their research, they showed the effect of medication and the effect of it on PD (Fig. 11).
Fig. 11. PD assessment methodology using smartphone [11]
Patel et al. [12] used accelerometer to find the severity of PD symptoms and motor complications. They mainly estimated the severity of tremor, dyskinesia and bradykinesia from the sensor data. Some standardized motor tasks were given to the patients to perform. An inspection was done of the video recorded at the time of patients’ performance. For collecting the data, the researchers asked the patients to perform some standardized motor tasks like finger tapping, hand movements, heel tapping etc. They positioned the sensors on the upper and lower limb of patients’ body. The accelerometer data was then filtered, and 5 different features were extracted. The data was collected every 30 min for seven trials after the patients had their medication. The baseline trial was taken before medication. Based on severity they built different clusters. For this, polynomial, exponential and radial basis SVM kernels were used. They combined different features to reduce the error and showed that three or four feature types from the accelerometer data gave them the minimum error. Using the rms value, data range value and frequency values (i.e., dominant frequency and the ratio between dominant
786
A. Adhikary et al.
frequency component and total energy) gave the average estimation error of 3.4% for tremor, 2.2% for bradykinesia and 3.2% for dyskinesia. But they asked the patients not to take medications before their data collection process. Though this procedure gave them maximum accuracy, but this is one disadvantage in their study (Fig. 12).
Fig. 12. Methodology for Parkinson severity detection [12]
4 Comparative Study In the above discussed papers, we found different classification techniques that were used to detect PD and its severity. The researchers collected data using smartphone sensors, wearable sensors etc. In some works, brain image was used. Whereas, in some other works, subjects’ performance was recorded as image sets. Data was collected through walking test in some papers. In [1, 7, 8, 11] researchers took smartphone sensors to collect the data. In [1], kNN, LR, NB, RF and SVM were used to detect PD using walking test. kNN gave the best accuracy for this test. kNN is used for classification. Based on the similarity of previous data points, it classifies new data points. Here parameter tuning is done to choose the best value for k. To calculate the nearest neighbor, Euclidean distance is used. But kNN is computationally expensive. For high dimensional data and large dataset with missing values, kNN does not work well. SVM is better than kNN as it is easier to interpret. SVM can separate the features using hyperplane. For both classification and regression SVM can be used. It is suitable when there are more features than training examples. In higher dimension also, SVM is effective. But SVM has certain problems. For large data sets, SVM takes more time. SVM’s performance also decreases if the target classes overlap in the dataset. SVM has different kernels. But choosing the right one is difficult. RF algorithms achieve better performance than SVM. SVM is better for regression whereas RF is better for classification. RF is consisted of multiple decision trees. The trees have their own prediction and the tree with maximum votes becomes the model’s prediction. Decision trees are also used for both classification and regression problem. It needs less effort in data pre-processing. DT has no effects of missing values while building tree. But any small changes in data can have a huge impact on the tree structure. Sometimes it takes more time in training models. Overfitting is one of the largest problems for DT. For continuous numerical variables, DT performs less efficiently. As random forest builds large number of trees, the complexity also gradually increases. Training period is more for RF, and it depends upon the number of trees.
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
787
Walking test was also done in [7] for gait and postural sway test. RF gave them highest accuracy. In [8] Random Trees (RT), RF, DT, NB, BN, kNN and AdaBoost were used for FoG detection where RF and AdaBoost gave almost same accuracy. AdaBoost is used as Ensemble method. Unlike RF, it uses decision stumps, which are decision trees having single split. AdaBoost uses multiple decision stumps. But it reduces error from the dataset by reducing bias, whereas RF reduces error by reducing variance. AdaBoost works less efficiently with noisy data and outliers. But using AdaBoost with RF makes better prediction for larger population and they are less prone to overfitting. In [11], the researchers used smartphone to collect data. We have found certain weaknesses while comparing the papers. Dataset taken in [7] was smaller than [1]. It gave lesser variation to PD symptoms. Mean age of the subjects was 60–70 years in [1] and they had several health issues. But in [7] mean age was 57–65 years. For [8] and [11], data collection was done using smartphone. FoG was detected by walking test in [8]. But in [11], relevant PD symptoms were tested using voice, balance, gait, dexterity, reaction time and tremor. RF gave the best accuracy for both works. SVM gave the best result in both [2] and [3]. In [6], xgboost gave highest accuracy. But it may suffer from overfitting due to improperly tuned parameters. Interpretation and visualization are also tough for xgboost. It is also slower than RF. In [4] and [5], the researchers took CNN to classify image dataset. In [4], 3D brain image was used, which had huge information. The authors used LeNet and AlexNet CNN architecture to extract the features. In [5] and [8] also, AlexNet architecture was used to extract the image data. In detecting PD, models are trained faster using AlexNet. Being a deeper architecture, AlexNet is better in extracting features than LeNet. Training time is also less for AlexNet, and it is approximately 4 times faster than other models. We have found certain weaknesses for the work. It takes more time for higher accuracy. Also, it suffers from overfitting problem. LeNet architecture faces problems in scanning all the features from images, for which it faces difficulties in generalizing and creating accurate models. In [9, 10] and [12], the researchers used wearable sensors which were attached to subjects’ body. Here the focus was to track patient’s daily activity. In [9], the main purpose of the researchers was to use the sensors to detect the FoG events. NV, RF, DT and RT algorithms were used for PD classification. Among them RF gave the best result. Main objective in [10] was to find the daily transition. Multi linear regression was used in this work to compare between PD patients with normal subjects. In [12] the researchers took SVM to evaluate the severity of PD. SVM is memory efficient and effective in high dimensional spaces, but certain drawbacks were found in the works. If the dataset is noisy i.e., target classes are overlapping each other, then the performance for SVM is reduced. For wearable sensors, security mechanism was not discussed in the papers, which is the major drawback in using the sensors for data collection and storage.
788
A. Adhikary et al. Table 1. Advantages and disadvantages of the above-mentioned papers.
Paper name
Method used
Result obtained
Advantage
Disadvantage
Juutinen et al. [1]
Classification tree, Gaussian kernel, Linear discriminant analysis, ensemble, kNN, LR, NB, SVM, RF for classification on accelerometer and gyroscope data
Highest accuracy of 84.5% and an error rate of 15.5% using kNN
Availability of smartphone and use of default android sensors made this method easy and acceptable Working with motor symptoms(walking test)madethe work useful
Some extra steps were taken while walking. So, some data are biased The measured data was manually transferred to database Running mRMR technique multiple times made it less efficient
Prashanth et al. [2]
NB, SVM, Boosted trees and RF classifiers on non-motor features (RBD and olfactory loss)
SVM gave them the best accuracy of 96.40% and 98.88% of AUC
Use of non-motor features with dopaminergic imaging features classified PD more accurately Use of the combination of pre-motor features, CSF and dopaminergic imaging markers boosted up the accuracy
Comparison of motor and non-motor features would be more relevant for this work Only early stage of PD patients’ data was taken for classification Any pre-motor subjects who had the symptoms but did not recognize as PD were not taken
Rahman et al. [3]
PLP and RASTA-PLP for feature extraction and SVM to classify the extracted features
SVM RBF kernel with PLP based feature extraction gave 74% accuracy
Remote diagnosis and monitoring made easy As speech disorder is one of the very common motor-symptoms for PD, so it was very useful in classifying PD from healthy people
Accuracy in this paper waslow, and it needs further improvement. Only vowel pronunciation data was recorded
(continued)
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
789
Table 1. (continued) Paper name
Method used
Result obtained
Advantage
Disadvantage
They got an accuracy of 95.1% and AUC of 97%
Use of isosurfaces reduced the complexity of the architecture and got a high accuracy due to CNN architectures
Computation cost was high as brain image was used The architecture faced problems in scanning all the features from the images
Jindal et al. [5]
CNN on Parkinson 96.5% of spiral drawing precision, 97.7% F1-score and 99.2% region
Using tablet-based drawing gave the opportunity to classify PD without taking the motor symptoms in concern Remote supervision of the patients made possible
The dataset was small, so a slight fall in the output was detected when spectrums from the 2D input grill was expelling
Tiwari et al. [6]
LR, DT, SVM, kNN, xgboost, bagging to predict PD
Xgboost gave them 95% test accuracy and 100% training accuracy
Result obtained from different classifiers were compared. It gave us idea about the classifiers that can be taken for PD classification
The dataset was very small, only 31 people’s data was considered, more improvement can be done using large samples
Arora et al. [7]
RF, random classifier and conditional random classifier were used on postural sway and gait
Random Forest gave an average sensitivity of 98.5% and an average specificity of 97.5% and accuracy of 98%
By using smartphone accelerometer data, it was possible to classify PD from healthy people outside the clinic also
Some primary gait and postural sway matrices like stride length, trunk flexion, minimum foot clearance was not computed
Mazilu et al. [8]
RT, RF, DT, Pruned Decision Tree, NB, BN, kNN with one neighbour and two neighbours, multi-layer perceptron, bagging and boosting to detect gait
AdaBoost gave the best sensitivity, specificity, F1 score and accuracy of 98–99%
Online FoG detection system gave opportunity for the doctors to classify PD from healthy people with a less effort Number of sensors were less
Result was slightly biased, as training and testing data was randomly selected A user dependent algorithm can be built in future
Ortiz et al. Used CNN to see [4] the progression of PD
(continued)
790
A. Adhikary et al. Table 1. (continued)
Paper name
Method used
Result obtained
Advantage
Disadvantage
Tripoliti et al. [9]
NB, RF, DT and RT classifier on freezing of gait using sensors data
96.11% of accuracy using the Random Forest classifier
Automated system was built based on patient’s movement types (walking, turning, or standing) for FoG detection Sensor data give brief understanding of the subjects’ movement
Only 16 subjects were involved in this study Some misclassification was there as different duration of the FoG events was taken
Elazari et al. [10]
Compared PD and healthy people’s motor symptoms and differentiate the type of PD
They achieved an accuracy of 92.3% for patients to healthy people, 89.8% for mild to severe patients, 85.9% for mild patients to healthy people using SVM
PD progression and response to the therapy was tracked PD severity was also taken into consideration
Comparison was tough as different features were observed for different group of subjects The correlation between the individual measures for PD severity and sensor-derived matrices were weak
Random forest achieved 71.4(±4) % accuracy
Easily available software was used for remote monitoring Clinical monitoring was done at low frequency Based on clinical data, patients were observed remotely and frequently with high frequently data
Data was collected booth actively and passively, but only active data was chosen for the research Tracking fluctuations in PD symptoms and medication response may help in monitoring disease progression Accuracy was only 71% which can be further improved
Zhan et al. Implemented a [11] smartphone application to track the motor symptoms of PD
(continued)
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
791
Table 1. (continued) Paper name
Method used
Patel et al. Tested severity of [12] PD for symptoms of tremor, bradykinesia, and dyskinesia
Result obtained
Advantage
Disadvantage
Different types of SVM kernel were used and polynomial kernel gives the best result
They monitored and grouped patients into different groups to identify the severity of PD
Data was collected before medication to maximize the accuracy Monitor patients in normal motor fluctuation cycle can be more useful
5 Conclusion and Future Scope From the above papers’ discussion, we find that PD mostly occurs for elderly people. But the symptoms, researchers worked with, are very common at old age. Many people forget things at 70’s, speak and walk slowly. Some has tremors and rigidity also. But in all the cases it cannot be said that the people having the symptoms must have PD. As PD is not fully curable so it is very important to detect it at its early stage. For early detection, premotor symptoms should be taken along with the motor symptoms. Comparison between these two can provide better result. Collecting medical data and storing it securely using wearable sensors is a critical task. In the above-mentioned papers, the security aspect of the data is not mentioned properly. This can be a huge threat for the patients if the data is leaked. So, IOT sensors and devices should be employed following proper security framework. Secured IOT architectures can ensure proper confidentiality, integrity and availability of the critical medical data during the phases of data collection and transfer to servers. Also, it has been observed that, in the existing works, the data in most of the cases were recorded at time of clinical check-up only. This results in a very small dataset which can affect the accuracy of the prediction models. Whereas the use of IoT devices in the home setup, can enable us to measure clinical data for a longer period. Smart home appliances can be used for subject’s walking and speech test. This will result in a larger and non-biased data set which will improve the accuracy of the prediction model. Availability of smartphones made PD detection easier. In our future work an android application can be developed to collect the data. The application once installed in the patient’s smartphone, will help in PD classification to be done remotely. To monitor the motor difficulties, an automated measurement system can be used. But it should be made sure that the application is user-friendly, so that, anyone can use and understand the functionality of the application. This can be an important step towards tele-diagnosis. This low-cost solution will be highly beneficial to many people in the rural areas in a country like India. It has also been observed that optimization techniques can be used to find the best solution from several features. This can be an important area of work in PD classification. Till date it is not clear for which particular reason, PD occurs the most. So, if PD reasons can be optimized, then we can find the best reason for it, and it will pave a path for proper treatment of the disease.
792
A. Adhikary et al.
References 1. Juutinen, M., et al.: Parkinson’s disease detection from 20-step walking tests using inertial sensors of a smartphone: machine learning approach based on an observational case-control study. Plos one 15(7), e0236258 (2020) 2. Prashanth, R., Roy, S.D., Mandal, P.K., Ghosh, S.: High-accuracy detection of early Parkinson’s disease through multimodal features and machine learning. Int. J. Med. Inform. 90, 13–21 (2016) 3. Rahman, A., Khan, A., Raza, A.A.: Parkinson’s disease detection based on signal processing algorithms and machine learning. CRPASE: Trans. Electr. Electron. Comput. Eng. 6, 141–145 (2020) 4. Ortiz, A., Munilla, J., Martínez-Ibañez, M., Górriz, J.M., Ramírez, J., Salas-Gonzalez, D.: Parkinson’s disease detection using isosurfaces-based features and convolutional neural networks. Front. Neuroinform. 13, 48 (2019) 5. Jindal, M., Tripathi, Y.: Parkinson’s disease detection using convolutional neural networks. Eur. J. Mol. Clin. Med. 7(6), 1298–1307 (2020) 6. Tiwari, H., Shridhar, S.K., Patil, P.V., Sinchana, K.R., Aishwarya, G.: Early prediction of Parkinson disease using machine learning and deep learning approaches, no. 4889. EasyChair (2021) 7. Arora, S., Venkataraman, V., Donohue, S., Biglan, K.M., Dorsey, E.R., Little, M.A.: High accuracy discrimination of Parkinson’s disease participants from healthy controls using smartphones. In: 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 3641–3644. IEEE (May 2014) 8. Sinha, T., Chowdhury, T., Shaw, R.N., Ghosh, A.: Analysis and prediction of COVID-19 confirmed cases using deep learning models: a comparative study. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 207–218. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-21642_18 9. Tripoliti, E.E., et al.: Automatic detection of freezing of gait events in patients with Parkinson’s disease. Comput. Methods Programs Biomed. 110(1), 12–26 (2013) 10. Palimkar, P., Shaw, R.N., Ghosh, A.: Machine learning technique to prognosis diabetes disease: random forest classifier approach. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 219–244. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-2164-2_19 11. Zhan, A., et al.: High frequency remote monitoring of Parkinson’s disease via smartphone: platform overview and medication response detection (2016). arXiv preprint arXiv:1601. 00960 12. Chakraborty, A., Chatterjee, S., Majumder, K., Shaw, R.N., Ghosh, A.: A comparative study of myocardial infarction detection from ECG data using machine learning. In: Bianchini, M., Piuri, V., Das, S., Shaw, R.N. (eds.) Advanced Computing and Intelligent Technologies. LNNS, vol. 218, pp. 257–267. Springer, Singapore (2022). https://doi.org/10.1007/978-98116-2164-2_21 13. Nilashi, M., Ibrahim, O., Ahmadi, H., Shahmoradi, L., Farahmand, M.: A hybrid intelligent system for the prediction of Parkinson’s Disease progression using machine learning techniques. Biocybern. Biomed. Eng. 38(1), 1–15 (2018) 14. Almeida, J.S., et al.: Detecting Parkinson’s disease with sustained phonation and speech signals using machine learning techniques. Pattern Recogn. Lett. 125, 55–62 (2019) 15. Sriram, T.V., Rao, M.V., Narayana, G.S., Kaladhar, D.S.V.G.K., Vital, T.P.R.: Intelligent Parkinson disease prediction using machine learning algorithms. Int. J. Eng. Innov. Technol. (IJEIT) 3(3), 1568–1572 (2013)
Machine Learning Based Approaches in the Detection of Parkinson’s Disease
793
16. Pahuja, G., Nagabhushan, T.N.: A comparative study of existing machine learning approaches for Parkinson’s disease detection. IETE J. Res. 67(1), 4–14 (2021) 17. Shi, T., Sun, X., Xia, Z., Chen, L., Liu, J.: Fall detection algorithm based on triaxial accelerometer and magnetometer. Eng. Lett. 24(2) (2016) 18. Ali, L., Zhu, C., Golilarz, N.A., Javeed, A., Zhou, M., Liu, Y.: Reliable Parkinson’s disease detection by analyzing handwritten drawings: construction of an unbiased cascaded learning system based on feature selection and adaptive boosting model. IEEE Access 7, 116480– 116489 (2019) 19. Abdulhay, E., Arunkumar, N., Narasimhan, K., Vellaiappan, E., Venkatraman, V.: Gait and tremor investigation using machine learning techniques for the diagnosis of Parkinson disease. Futur. Gener. Comput. Syst. 83, 366–373 (2018) 20. Wang, W., Lee, J., Harrou, F., Sun, Y.: Early detection of Parkinson’s disease using deep learning and machine learning. IEEE Access 8, 147635–147646 (2020)
Tactile Internet in Internet of Things Ecosystems Ignacio Lacalle1 , C´esar L´opez1 , Rafael Va˜ no1 , Carlos E. Palau1 , 1 2 Manuel Esteve , Maria Ganzha , Marcin Paprzycki2(B) , and Pawel Szmeja2 1
Communications Department, Universitat Polit`ecnica de Val`encia, Valencia, Spain {iglaub,csalpepi,ravagar2}@upv.es, {cpalau,mesteve}@dcom.upv.es 2 Systems Research Institute Polish Academy of Sciences, Warsaw, Poland {maria.ganzha,marcin.paprzycki,pawel.szmeja}@ibspan.waw.pl
Abstract. The aim of this contribution is twofold. First, to summarize the state-of-the-art in the area of Tactile Internet. Second, to outline, based on pilots of the ASSIST-IoT project, an architecture needed to realize Tactile Internet in Internet of Things ecosystems. Keywords: Tactile internet · Distributed systems Internet of Things · Haptic interfaces
1
· 5G networks ·
Introduction
Tactile Internet (TI) is one of recent pivotal technological trends. According to the International Telecommunication Union – Telecommunication Standardization Sector (ITU-T), Tactile Internet is defined by “extremely low latency, in combination with high availability, reliability and security” [1]. It refers to systems leveraging extra fast networks, enabling haptic human-machine interaction. Applications of TI are multiple and diverse, and its relevance, observed in the literature, is growing fast (e.g., since the introduction of the term in 2014, in the first paper [2], more than 135 related articles have been published in 2020, and the trend continues to grow in 2021). All works support the claim that Tactile Internet reaches far beyond data streaming over fixed and/or mobile networks, and provides a new dimension to the Internet, by improving availability, reliability and latency. In addition, the upsurge of related technologies such as social robots, wireless sensing networks, and the Next Generation Internet of Things (NGIoT), contribute to the rise of the Tactile Internet. What follows summarizes key aspects of Tactile Internet of today (Q1, 2021) and outlines key aspects of the NGIoT architecture that can be seen as TI-ready.
2
Why Is Tactile Internet Needed?
For Tactile Internet to be understood, let us outline why it came into being. The advent of TI emanates directly from needs expressed by the industry and the technological community. The following sections contextualise the need of TI. c The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 794–807, 2022. https://doi.org/10.1007/978-981-19-1677-9_69
Tactile Internet in IoT
2.1
795
Foundations of Tactile Internet
Internet communication infrastructure is used mainly to transfer information between “specific points”. As Internet evolved, it was possible to improve the key parameters of data transmission, e.g. latency, or data rate. This brought new possibilities, such as remote operation of heavy machinery, or remote healthcare. As a result, a technological trend has emerged, focused on offering point-to-point communication with extreme low latency and high reliability. Not only does it bring about new communication paradigms (e.g. haptic data1 ), but also provides foundation for the next generation of applications (see also, Sect. 2.2). Standardisation/Research Groups. The ITU-T was one of the first research and standardization groups that mentioned Tactile Internet [1]. The IEEE SAB presented the IEEE 1918.1 as an IEEE Standards Working Group devoted to Tactile Internet [4]. This standard defines the concept, and includes definitions, terminology, and application scenarios. Moreover, it includes the reference model and the architecture, which define architectural entities, interfaces, and mappings. Finally, the European Telecommunications Standard Institute (ETSI) has created a group that considers impact of IPv6 across technologies, including Tactile Internet [5]. There have been other actions targeting TI, but using other names, such as the 5G-NR (New Radio; [6]), released by the 3GPP that aims at providing enhancements related to Ultra-Reliable Low Latency Communication. Communities/Clusters. In 2016, an Alliance for the Internet of Things Innovation (AIOTI) was initiated, to contribute to development of the European IoT ecosystem. It defined TI, in the Standardisation Working Group, as a technology enabler for ultra-high reliability, and ultra-low latency applications, where the network is part of a haptic feedback loop [7]. The NG-IoT Initiative [8] (part of the ICT-56 cluster [9]), introduced TI as one of the enabling technologies for the next generation IoT. Aiming at creating a competitive ecosystem of technologies for IoT, a series of projects seeks to include TI based on human-centric sensing, and new IoT services, e.g. integration of computing capabilities, neuromorphic computing, etc. Here, it is expected that TI will be enabled by, and grow with, IoT, AR/VR/MR and contextual computing. Strategic Technological Trends. Tactile Internet is focused on enabling real-time physical interaction, involving real and/or virtual objects. Technical requirements for the TI will need to be addressed by innovative technologies, which are fast maturing [10]. Among them, key is the deployment of 5G networks, to achieve Ultra Reliable Low Latency Communication, and enable human control of real and virtual objects, in real-time [11]. Furthermore, developments 1
Usually, defined as any data created to provide haptic feedback, which allows users to touch, feel, and manipulate physical, or virtual, objects through remote machines. Here, the data rate is 1 KHz or higher [3]; see, also, Sect. 3.1.
796
I. Lacalle et al.
are accelerated by unexpected needs, such as the health emergency, caused by COVID-19. Here, use of remote communication has “exploded”, pushing development of applications requiring TI. Moreover, studies of 2021 technology trends ([12]), emphasis such “interaction” as a key emerging trend at the macro level. 2.2
How the Industry Expresses the Necessity of Tactile Internet
Recently, the Industry has been facing problems that go beyond visual and/or hearing interactions, and require introduction of the Tactile Internet. One example is use of robots in manufacturing. Managing such machinery remotely, replicating movements of operators, without risking their health, has been reported, among others, in [13]. Moreover, critical applications, where humans are involved, may benefit from the TI. Here, consider remote surgery, which requires ultra reliable, low latency, robust systems [14]. The number of verticals in need of TI is huge. Consider, for instance, addressing the following needs: – To facilitate the remote on-the-field intervention of experts, who would be put at risk if present at a given physical location. Here, availability of haptic movements, actionable through a machine (e.g., robot) is needed. – To facilitate remote operations in inaccessible places (e.g., drones flying over fires, actions within deep seabed locations, rescue missions, etc.). – To create remote experiences, allowing visiting sites without being physically present. Here, large scale deployment could imply enormous benefits in various contexts (e.g., reducing traffic, transfer of people, etc.). – To boost teleworking. Some jobs require physical presence of operators, or other workers that must perform “manual/touchable” tasks. These actions could be replicated by robots, provided that TI becomes fully operational. This has being recognised as paramount in the context of COVID-19 pandemic, and the need for social distancing measures [15]. Furthermore, TI will generate new business opportunities, and academic/research lines. In particular, companies devoted to communication technologies are well positioned to play a major role exploring potential of Tactile Internet.
3
What Does Tactile Internet Entail?
Some works, like [16], have classified the Tactile Internet using an application taxonomy (wireless networked control systems, remote operation systems, immersive entertainment and edutainment systems, intelligent mobility systems, smart energy systems), or as a global field composed of enabling technologies (high-performance wireless, connectivity requirements, and vision). The authors of this work consider, drawing from the experience gained in the ASSIST-IoT project (see Sect. 5), that TI must be understood as any system that (i) successfully carries haptic data by (ii) relying on a communication meeting the requirements of (a) low latency, (b) high reliability, and (c) high data throughput. Following sub-sections aim at describing the meaning of each of these aspects.
Tactile Internet in IoT
3.1
797
Haptic Data and Haptic Codecs
Tactile Internet’s mission is to enable exchange of haptic data, facilitating novel human-machine interaction. Haptic information can be expressed through various parameters, e.g.: force, movement, vibration or texture. Usually, such data is generated by haptic feedback devices that can be divided into: cutaneous (involving skin features, e.g. pressure, touch intensity, temperature or pain) and kinesthetic (relying on relative position of neighbouring body parts, and capturing, e.g., muscle tension) [17]. Two scenarios can be further defined, depending on the communication: (i) closed-loop feedback (i.e. perception) that require being as-close-to-real-time-as-possible (delay intolerant) and (ii) open-loop feedback, which are less restrictive, and accept time-delayed communication. The most relevant characteristic in haptic data transmission is information encoding (IEEE SAB group 1918.1.1). Similarly to audio signals (using e.g., MP3 or HVXC), or image frames (using e.g. PNG or JPEG), encoding schemes are needed for haptic feedback to allow final ends of communication to understand each other (speak the same language, set plug-and-play communication, etc.). The algorithms taking care of encoding are called “codecs”. Currently, the encoding proposals (based on haptic codecs, such as block-based processing, or frequency-domain models [18]), are combined with audiovisual transmission, to provide the feeling of being present in a remote environment, allowing to work in distant or inaccessible situations (see Sect. 2.2). In the most restrictive TI scenario (closed-loop, above), there is the need for the codecs (apart from enabling communication) to reduce amount of data transmitted (number of packets). To do so, current approaches rely on mathematical models of human perception. Overall, encoding schemas – and associated codecs applying data transformation algorithms – are an open research line aimed at reducing the amount of transmitted haptic data, to reach validity of human perception. 3.2
Infrastructure Requirements
While the technology stack, needed to support TI, is still under development, the necessary requirements can be assessed. Tactile Internet communication must be characterised by strict latency and Quality of Experience (QoE) restrictions. Due to human involvement, there is a limitation imposed by the sensation of touch, estimated at a time resolution of 1000 Hz [17]. To offer a realistic experience, TI must offer communication latency of 1 ms, from the moment the action starts (e.g. user moves a finger, or remote operator presses the edge of a glove), till the feedback is received. Assuming that communication paths cannot reach the speed of light, the maximum distance between TI endpoints is limited to 150km [1]. However, TI also requires communication infrastructure delivering both high availability and high reliability. Based on [19], let us summarize the infrastructure requirements (see also Fig. 1): – Low-latency. In a generic case, the time-consuming operations are: running user interface, data transmission, and computing. For TI to deliver seamless,
798
I. Lacalle et al.
Fig. 1. Visual summary of the infrastructure requirements of the Tactile Internet. Copyright note: Flaticon
low-latency, connectivity, it should take full advantage of 5G networks, delivering round-trip latency in the millisecond range. However, 5G may not be available. Hence, robust implementation should be possible also over WLAN, sub-GHz technology, and their combinations [20]. This is why technologies, like Software-Defined Networking (SDN) and Virtualised Network Functions (VNFs), are explored to significantly reduce latency, by optimising data transmission [21]. When fast-enough communication cannot be guaranteed, other techniques may be used. Here, in [17], use of Artificial Intelligence is suggested (based on regression models and linear predictions) leveraging previously received values. However, such complex approaches may be useful if the computation is realized “close to the user”. Therefore, they should be placed at the edge of the network (e.g., Mobile Edge Computing; MEC [22]). – High reliability. In all communications, the goal is to make them speedoptimal and error-less. However, this can be very difficult, due to environmental conditions, or strict requirements of the application. This involves, for instance, need for extremely high level of reliability, due to the consequences that can arise from communication errors. Consider a target of 99.99999% availability (of a remote surgery system), allowing outages of only 3.17 s per year. Separately, to ensure the integrity of data, security mechanisms are needed. Obviously, security is important also when haptic data is processed on a remote server, as data travels between the device and the MEC server. Here, encryption techniques are needed. However, they imply additional delay in communication [17], so a trade-off must be found. – High data throughput. Haptic-enabled VR/AR/MR applications require high data throughput. However, transmission is the most critical point for moving large volume of data in (near) real time, becoming one of the most demanding limitations of TI. This is because of the amount of data that
Tactile Internet in IoT
799
moves, the needed bandwidth, and lack of buffer that manages the reception. A promising solution would be to increase the transmission frequency. However, use of narrow beamwidth directional antennas, required to reduce the high path loss at mmWave, may result in frequent link outages, due to antenna misalignment. To support TI applications, a hybrid radio access architecture was proposed [23]. Here, sub-6 GHz access is used for transmission of haptic information, while mmWave access is used for high data-rate transmission of audiovisual information. Overall, Tactile Internet has strict technological requirements for end-to-end communication, data encoding, and overall network availability, to achieve actual real time interaction. As noted, several open challenges are currently faced, while alternative technologies (e.g. SDN, NFV, narrow beamwidth antennas, MEC at 4G/5G) have been proposed to meet strict communication requirements. However, the state of the art does not currently point to a clear reference product/solution that gathers all previous innovations in a single specification.
4
Reference Initiatives
Lately, some efforts have started to formalise a reference Tactile Internet specification, with tentative applications in real-world scenarios. However, up to this point, there is not a clear standard to stick to in terms of TI deployment. This section briefly outlines the most relevant initiatives with that purpose. 4.1
Theoretical and Application Agnostic Solutions
Since being proposed, Tactile Internet has been associated mainly with a communication structure, focused on the interaction between the end-points (transmitter-receiver, client-server, publisher-subscriber, user-remote machine). In 2014, the ITU-T [1] proposed to frame communication within TI systems using a master-slave schema, reasoning that it is most natural, and drawing from the legacy, pre-TCP/IP era telecommunication systems. Here, the master as the “commander”, issues actuation orders to the slave, which is in charge of “performing” the command. In the context of Tactile Internet communication, the master would be, for instance, a remote controller entity (human or machine) that, through an interface, codes haptic input into command signals and sends them to the slave. The slave could be an object/robot, directly guided by the remote controller (master), capable of providing feedback and closing the loop [24]. The architecture proposed by the IEEE 1918.1 follows the same masterslave concept, but improves the edge connectivity, by setting a gateway node to connect slaves to the network. The architecture included the gateway node at the edge, near to the master and slave actuators, or in the domain network. Contributions from ITU-T and IEEE, along with other architectures found in literature, are mainly theoretical and generic, and do not include real/practical
800
I. Lacalle et al.
aspects (e.g., including a list of technologies and how to “glue them together”) to implement TI. Moreover, they lack a link to the role that IoT (sensors, gateways, actuators, data exchange, access networks) are to play in specifications. Separately, multiple projects introduce alternative vision of Tactile Internet. In FlexNGIA [25], a TI infrastructure is based on network applications. Leveraging computing capacities of network elements, FlexNGIA proposes to deploy services as network functions. FlexNGIA defines a business model where network operators could offer not only data delivery but also service chains, with stringent requirements in terms of performance, reliability and availability. Other projects, like TACTILENet [26], implement TI in a network architecture that adapts to the QoE requirements of specific scenarios, instead of uniformly delivering higher capacity and higher reliability. TACTILENet has roots in cloud densification (spatial overload and spectral aggregation), green energy efficiency, and Cross-layer Machine-type Communications. Next, TACNET 4.0 [27] is focused on 5G integration, through an architecture designed to support remote control of mobile machines/robots, allowing workers to interact through AR devices. TI is achieved using 5G, and a management and control plane that ensures efficient use of resources, to improve reliability. Other proposals offer integration of specific novel technologies to support vertical application requirements. For instance [20] enables flexible and dynamic slicing, by integrating SDN and NFV with fog computing, and creating a hierarchy where the SDN controller manages the network. To improve latency and reliability, by reducing network congestion, a multi-level cloud system, described in [22], provides offloading through MEC. Improvements in network capacity can be based on enhancing LTE-A heterogeneous networks fiber backhaul and WiFi offloading capabilities [28], while in [29] this is achieved by simultaneously transmitting over multiple wavelength channels in Ethernet Passive Optical Networks. 4.2
Specific Solutions
As noted, early realizations of TI can be found in the literature. Here, we briefly overview key publications, dividing them into “application domains”. It should be noted that the described approaches implicitly express urgent need of a valid, robust, scalable reference architecture, focused on practical implementation of the Tactile Internet. Provided examples illustrate variance of significance of TI requirements, from one application to another. – Mechatronics: human-humanoid robots interaction, combines electronics, computing, telecommunications, etc., to create robotic systems. These systems can be managed in real-time by a human-machine communication following the master-slave concept. In mechatronics [13], the human interacts with a remote body to execute various skills, physically linking the virtual and the real worlds. Here, applications face strong real-time requirements, as in remote surgery [30], where an increase of latency, or packet transmission errors, can lead to serious consequences. An example of this vertical can be found in [31], proposing a remote controlled exoskeleton, which is not that
Tactile Internet in IoT
–
–
–
–
–
801
demanding on reliability, but requires a reaction time fast enough to feel realtime movements. Some approaches set multi-robot applications, allocating physical and/or digital human tasks to robots. In [32], an integrated FiberWireless multirobot network architecture is proposed, coordinating local and non-local Human-to-Robot task allocation. Other projects, as Ocean One [33] design an architecture that allows human-robot interaction, not only with video and haptic data, but also with GPS coordinates. Reliability requirements materialize in industrial applications (e.g., safety control or health interventions) where 99.9999% is acceptable, but 99.99999% is desirable. Virtual Reality (VR): video transmission applications do not involve haptic data, either because the master only needs image and sound to control the system, or because a machine-to-machine communication is carried out, and video data is absent. However, high resolution images, and 3-D stereo audio, in VR and AR applications demand massive flow of information, and bring different challenges. For instance, network throughput, or delay performance, has to become less than 10 ms latency, to avoid cybersickness [17]. Some use scenarios, e.g. a free-viewpoint video, allows digital image processing to synthetically render the viewpoint of the viewer to another spot. Helmet-mounted applications: a remote VR phobia treatment architecture is proposed in [34], using MEC networks to reduce computation delay, and mmWave communications to increase network capacity. Note that, thanks to similar innovations, new helmet-mounted VR devices have emerged, e.g. Oculus VR, HTC Vive, or Microsoft Hololens. Ultra-realistic videogames: VR games may use TI to perform real-time interaction. System, described in [35], proposes mmWave Access Points, to reduce latency, and edge computing to minimise communication errors. Unmanned aerial vehicles (UAV): video surveillance application [36], use UAV with cameras, interacting with sensors and actuators, installed on the ground. Due to the long time required on the transmission of video, from the drone to the cloudlet, or MEC ground-servers, a microcontroller is installed on board of the drone, to process incoming data, make decisions, and send activation messages to ground-based-actuators. Here, TI applications that do not require human presence, use data transmission optimisation; e.g., in autonomous driving, vehicles make decisions during driving events. This application requires low-latency and high-reliability, with extremely highavailability, due to potentially tragic consequences of accidents. Vehicle-to-Vehicle (V2V) applications: in [37], a scenario, for broadcasting messages only, covers communication patterns, where either a vehicle, or a pedestrian, can send the information of its location or velocity, so the receiver can calculate the relative position to avoid collision, or to facilitate operation in emergency situations. In [38], a testbed, based on flexible and re-configurable software-defined radio, is proposed, to improve cooperative automated driving. The system includes a re-configurable frame structure with fast-feedback, a novel P-OFDM waveform, low-latency multiple-access scheme, and a robust hybrid synchronization.
802
I. Lacalle et al.
– Smart Cities applications: also use TI and relevant M2M communications. In [39], a quality of experience (QoE)-driven five-layer Tactile Internet architecture is proposed. It includes inter-band spectrum aggregation in transmission, and osmotic computing to handle offloading in cloud edge integration. Network optimisation, adapting resources to the demands, is achieved by intelligent decision-making in the application layer.
5
ASSIST-IoT Approach Towards Tactile Internet
Let us now introduce ASSIST-IoT2 , a H2020-funded project that aims at design of a decentralised NGIoT architecture to support human-centric applications. The solution will integrate 5G, AI-based functions, Edge/Fog computing and SDN/NFV (among other technologies), considering Tactile Internet as one of key objectives. In this context let us note that, as discussed, TI has been positioned mainly as a way to solve problems within vertical application areas. However, it was also shown that infrastructure requirements play a major role in TI systems, and those requirements might vary between application domains. Henceforth, it is crucial to include TI’s strict limitations in the architectural decisions of any forthcoming IoT/distributed systems reference architecture(s). In this context let us note that although the architecture designed in ASSISTIoT will be application-agnostic, it will be bottom-up validated. Specifically, the ASSIST-IoT project is grounded in three application areas (industries): maritime port terminals, construction and automotive. Within these industries, four pilots have been formulated. The scope of these pilots, as related to Tactile Internet, has been summarized in Table 1. Material presented thus far, supported by the content of Table 1 allows us to outline the key aspects of the ASSIST-IoT architecture that have been conceived to support Tactile Internet. Note that this is a preliminary vision of the architecture, and only the TI-focused aspects are reported. Overall, ASSIST-IoT approach to the next generation of IoT architectures is based on three premises: – The architecture has to be domain-agnostic and the implementation has to support all kinds of applications. – Architecture has to be flexible and adapt to different situations, depending on the requirements (note that all scenarios require high capacity network). Recall that applications may require (a) high-availability (such as autonomous driving), (b) bandwidth (such as video surveillance monitoring in drones, or (c) latency (such as remote mechatronics). In this context, ASSIST-IoT goes beyond the current landscape in two areas: (i) proposing and intelligent management of the available network resources to avoid interruptions in communication, or congestion in any node of the network, and (ii) allowing dynamic (automation to be explored) modification of the network, for changing scenarios and requirements within the same deployment, leveraging SDN and NFV techniques. 2
https://assist-iot.eu/.
Tactile Internet in IoT
803
Table 1. Tactile internet in ASSIST-IoT Pilots
Actors
Equipment
Requirements
Pilot 1 – Port automation
Remote operator
Container Handling Equipment (CHE)
Latency Remote control with end-to-end latencies below 50 ms Reliability Remote Operating System expected to work 99% of the time Others Enough communication bandwidth to ensure video streaming (around 30 Mbps)
Pilot 2 – Smart Construction safety of workers worker
Fall arrest equipment
Reliability Precise construction worker localisation service is expected to work 99,99%
Pilot 2 – Smart Construction safety of workers worker
Dangerous zone alarm
Latency Alarm construction worker about entering dangerous area (e.g. construction plants) Reliability Precise geo-positioning of construction worker vs. construction plants
Pilot 2 – Smart Construction safety of workers worker
Evacuation instructions
Reliability Precise geopositioning of construction worker Others Enough communication bandwidth to ensure that in case of jeopardizing events all construction workers will receive their evacuation plans and will be navigated. Evacuation plans will be regularly updated
Pilot 3 – Cohesive vehicle monitoring and diagnostics
AR interfaces
Others Enough communication bandwidth to ensure video streaming of the vehicle, on a mobile device (around 30 Mbps)
Driver
– To achieve this, ASSIST-IoT proposed a novel meaning to the term “enabler”, which can be understood as a grouping of containerized components, that together deliver a specific (micro)service and act towards a single goal (to provide a specific functionality) framed within a specific plane of the architecture (see Fig. 2), providing deep modularity. Taking into account information available within Fig. 2, we can summarize the way that ASSIST-IoT proposes to introduce TI, in its NGIoT architecture. – Introduction of the interaction with the human in the Application and Services plane. AR components will provide new ways of accessing and con-
804
I. Lacalle et al.
Fig. 2. ASSIST-IoT approach towards Tactile Internet.
–
–
–
–
figuring IoT environments for operation and training, supported by edge computing nodes and smart network capabilities, to achieve a low latency environment, and using advanced visual information systems to make better informed decisions by people/workers. Here, the human interface will act as a master role, existing in the three pilots to be deployed in the project. At the Data Management plane, enablers for haptic data encoding mechanisms (softwarised codecs) will be provided. Additional enablers, for aggregating and semantically annotating haptic data, needed to encode the information are envisioned. In the Smart and Network Control plane, virtualised low-latency networks (with embedded DLT for security) will provide tactile support for real-time application. This will be achieved leveraging SDN-enabled switches through customised controllers. Regarding the Device and Edge plane, the IoT gateway will support multiple access networks (5G, 4G, WiFi, Ethernet, Zigbee, LoRa, Bluetooth) and will be able switching between them for redundancy, and to minimise latency. Finally, haptic actuators and sensing networks will support human-centric approach to development of TI-oriented services and applications, thanks to use of decentralised edge approach to deliver low-latency for haptic interfaces.
Tactile Internet in IoT
6
805
Concluding Remarks
The presented work has reviewed the current state of Tactile Internet as a concept, digging also into its technological nature, and outlining how it has emerged both as an industry need and as the next step needed for evolution of Internet of Things ecosystems. In addition, the actual meaning of the Tactile Internet, in term of infrastructure requirements and the type of data that manages was considered. It was suggested that, while the requirements and constraints may vary from one application to the other, there exists the common quest of reducing round-trip latency, in a highly reliable system, looking for providing real feeling of participatory perception to the user. Besides, it has been argued that TI’s wide scope has opened a myriad of different deployment approaches (promoted – or not – by standardisation entities), preventing the state-of-the-art to come up with a clear implementation reference (architecture) to build on. Finally, it has been suggested why ASSIST-IoT may deliver such a reference, drawing from a novel, dynamic, intelligent architecture that will be validated in four pilots, each with Tactile Internet ambitions. Acknowledgments. This work is part of the ASSIST-IoT project that has received funding from the EU’s Horizon 2020 research and innovation programme under grant agreement No 957258.
References 1. ITU-T: The Tactile Internet – Technology Watch Report (2014). http://www.itu. int/ITU-T/techwatch 2. Fettweis, G.P.: The tactile internet: applications and challenges. IEEE Veh. Technol. Mag. 9(1), 64–70 (2014) 3. You, Y., Sung, M.Y.: Haptic data transmission based on the prediction and compression. In: 2008 IEEE International Conference on Communications. IEEE (2008) 4. Holland, O., et al.: The IEEE 1918.1 “tactile Internet; standards working group and its standards. Proc. IEEE 107(2), 256–279 (2019) 5. ETSI-WP-2017-2018.html. https://www.etsi.org/e-brochure/Work-Programme/ 2017-2018/mobile/index.html#p=28 6. 3GPP – Release 15. https://www.3gpp.org/release-15 7. Standardisation AIOTI. https://aioti.eu/wg standardisation/ 8. Research & Innovation - Next Generation IoT. https://www.ngiot.eu/researchinnovation/ 9. ICT-56-2020. https://ec.europa.eu/info/funding-tenders/opportunities/portal/ screen/opportunities/topic-details/ict-56-2020 10. Gartner Top 10 Strategic Technology Trends for 2020 – Smarter With Gartner. https://www.gartner.com/smarterwithgartner/gartner-top-10-strategictechnology-trends-for-2020/ 11. The 5G tactile internet is here. What next? – Ericsson. https://www.ericsson.com/ en/blog/2019/4/5g-tactile-internet 12. Deloitte – Top Tech Trends 2021. https://www2.deloitte.com/content/dam/ insights/articles/6730 TT-Landing-page/DI 2021-Tech-Trends.pdf
806
I. Lacalle et al.
13. Haddadin, S., Johannsmeier, L., Diaz Ledezma, F.: Tactile robots as a central embodiment of the tactile internet. Proc. IEEE 107(2), 471–487 (2019) 14. Gupta, R., Tanwar, S., Tyagi, S., Kumar, N.: Tactile-internet-based telesurgery system for healthcare 4.0: an architecture, research challenges, and future directions. IEEE Netw. 33(6), 22–29 (2019) 15. Milasi, S., Gonzalez-Vazquez, I., Fernandez-Macias, E.: Telework in the EU before and after the COVID-19 : where we were , where we head to. Technical report (2020) 16. Aijaz, A., Sooriyabandara, M.: The tactile internet for industries: a review (2019) 17. Van Den Berg, D., et al.: Challenges in haptic communications over the tactile internet. IEEE Access 5, 23502–23518 (2017) 18. Steinbach, E., et al.: Haptic codecs for the tactile internet. Proc. IEEE 107(2), 447–470 (2019) 19. Accenture: Tactile internet enabled by pervasive networks. Technical report (2018) 20. Bojkovic, Z.S., Bakmaz, B.M., Bakmaz, M.R.: Vision and enabling technologies of tactile internet realization. In: 2017 13th International Conference on Advanced Technologies, Systems and Services in Telecommunications, TELSIKS 2017 – Proceeding. vol. 2017-October, pp. 113–118. Institute of Electrical and Electronics Engineers Inc. (2017) 21. Prados-Garzon, J., Adamuz-Hinojosa, O., Ameigeiras, P., Ramos-Munoz, J.J., Andres-Maldonado, P., Lopez-Soler, J.M.: Handover implementation in a 5G SDNbased mobile network architecture. In: IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, PIMRC. Institute of Electrical and Electronics Engineers Inc. (2016) 22. Ateya, A.A., Vybornova, A., Kirichek, R., Koucheryavy, A.: Multilevel cloud based tactile internet system (2017) 23. Joshi, K.C., Niknam, S., Prasad, R.V., Natarajan, B.: Analyzing the tradeoffs in using millimeter wave directional links for high data-rate tactile internet applications. IEEE Trans. Industr. Inf. 16(3), 1924–1932 (2020) 24. Promwongsa, N., et al.: A comprehensive survey of the tactile internet: state-of-theart and research directions. IEEE Commun. Surv. Tutor. 23(1), 472–523 (2021) 25. Zhani, M.F., ElBakoury, H.: FlexNGIA: a flexible internet architecture for the nextgeneration tactile internet. J. Netw. Syst. Manag. 28(4), 751–795 (2020). https:// doi.org/10.1007/s10922-020-09525-0 26. TACTILENet. https://tactilenet.sabanciuniv.edu/en/Objectives 27. TACNET 4.0. http://www.tacnet40.de/index englisch.html 28. Maier, M., Ebrahimzadeh, A.: Towards immersive tactile internet experiences: lowlatency FiWi enhanced mobile networks with edge intelligence. J. Opt. Commun. Netw. 11(4), B10 (2019) 29. Neaime, J., Dhaini, A.R.: Resource management in cloud and tactile-capable nextgeneration optical access networks. J. Opt. Commun. Netw. 10(11), 902 (2018) 30. Luvisotto, M., Pang, Z., Dzung, D.: Ultra high performance wireless control for critical applications: challenges and directions. IEEE Trans. Industr. Inf. 13(3), 1448–1459 (2017) 31. Luengas, Y.R., Lopez-Gutierrez, R., Salazar, S., Lozano, R.: Robust controls for upper limb exoskeleton, real-time results. Proc. Inst. Mech. Eng. Part I: J. Syst. Control Eng. 232(7), 797–806 (2018) 32. Chowdhury, M., Maier, M.: Collaborative computing for advanced tactile internet human-to-robot (H2R) communications in integrated FiWi multirobot infrastructures. IEEE Internet Things J. 4(6), 2142–2158 (2017)
Tactile Internet in IoT
807
33. Khatib, O., et al.: Ocean one: a robotic avatar for oceanic discovery. IEEE Robot. Autom. Mag. 23(4), 20–29 (2016) 34. Wrzesien, M., et al.: Treating small animal phobias using a projective-augmented reality system: a single-case study. Comput. Hum. Behav. 49, 343–353 (2015) 35. Elbamby, M.S., Perfecto, C., Bennis, M., Doppler, K.: Toward low-latency and ultra-reliable virtual reality. IEEE Netw. 32(2), 78–84 (2018) 36. Grasso, C., Schembra, G.: Design of a UAV-based video surveillance system with tactile internet constraints in a 5G ecosystem. In: 2018 4th IEEE Conference on Network Softwarization and Workshops, NetSoft 2018, pp. 132–140. Institute of Electrical and Electronics Engineers Inc. (2018) 37. Tanwar, S., Tyagi, S., Budhiraja, I., Kumar, N.: Tactile internet for autonomous vehicles: latency and reliability analysis. IEEE Wirel. Commun. 26(4), 66–72 (2019) 38. Cao, H., Gangakhedkar, S., Ali, A.R., Gharba, M., Eichinger, J.: A 5G V2X testbed for cooperative automated driving. In: IEEE Vehicular Networking Conference, VNC. IEEE Computer Society (2016) 39. Wei, X., Duan, Q., Zhou, L.: A QoE-driven tactile internet architecture for smart city. IEEE Netw. 34(1), 130–136 (2020)
Artificial Neural Network Based Biometric Palm Print Recognition System for Security Analysis M. Sowmiya Manoj(B) and S. Arulselvi Department of Electronics and Communication Engineering, Bharath Institute of Higher Education and Research, Chennai, India [email protected]
Abstract. The detection of human features can be achieved using various techniques. There are many demerits in the detection and in the process of maintaining the forensic tasks. So the authentication process is developed to the level of palmprint recognition. The palm print recognition of the people can be done through the left or right that can execute the same as the stamp of the finger. The size of the palm is bigger when it is compared to the fingers print. It is the very easier step to obtain the data and the processing speed is also very high. The palm print authentication is used for the good authentication process. It is very new when it is compared to the fingers print and their organization of face the palm print has many merits when it is compared to other methods and it is also very reliable. In this research work, palmprint is detected with the help of an ANN classifier. It shows a better result than the conventional model. This proposed model is implemented using the MATLAB tool and the performance is assessed based on the accuracy rate. Keywords: Biometric · Authentication · ANN (Artificial Neural Network) · Images · Database · Accuracy
1 Introduction Automatic and reliable personal authentication for effective security control is demanded with the rapid improvement in the utilization of e-commerce based applications. For the matching of feature data the pattern of the line, lines of the branches, and the wrinkles present in the form or utilized. It consists of low-resolution images for the matching data. The parameters that are mentioned above will not be the same for all humans it will vary from one to another. Even the twins from the same mother don’t have the same branches and wrinkles in the palm so it is a critical task. Till now many techniques is proposed for the recognition of palm print that is categorized into two stages. They are structural feature and statistical feature-based. The technique of structural feature-based system can obtain directly the structural information, like principle lines, wrinkles, minutiae points, etc. that can indicate the feature of palm print in a clear form. In this type, the features can be obtained even from the images with low resolution this type of technique has to spend more computation cost © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 808–819, 2022. https://doi.org/10.1007/978-981-19-1677-9_70
Artificial Neural Network Based Biometric Palm Print Recognition System
809
on matching the segment of line with the templated that is stored in the database. In statistical features-based technique, palmprint images are considered as the whole and palmprint features are obtained by the transformation of the image. The features that were obtained are continuously used for the classification. Many feature extraction techniques like Fourier transform, Gabor filter, ICA are explored. The characters of the statistical feature based approaches are dependent on the effectiveness of feature extraction techniques and also scheme of classification. Comparing with the Fourier transformation and Eigen’s palm, ICA is linked with the property of statistic data that can be utilized for multi-dimensional data. Figure 1 shows the major features of the palmprint (Goh Kah Ong Michael et al. 2011).
Fig. 1. Major features of the palmprint (Goh Kah Ong Michael et al. 2011)
Figure 1 palmprint consists of three important lines. From the above Fig. 1 denotes the heart line, 2 indicates the headline and three represents the lifeline of the palmprints. ANN commonly called neural networks, which are the computing systems that are inspired by the biological neural networks which constitute the brain structure.The main intention of this research work is to identify the people using their palmprints with the help of an ANN classifier.
2 Literature Review In the image processing field, palmprint detection is the most important research topic. Palmprint detection faces various issues because of image illumination level, shadow, blur, and obstacles of the images. Istiqamah. I et al., 2016 recommend a new palmprint detection method using the hand attributes, and LVQ (Learning Vector Quantization) ANN. Palmprint detection is exactly to identify the people due to its uniqueness. Compare with the fingerprint, the palmprint contains various information like wrinkles, ridges, and principal lines. The major advantages of the palmprint detection methods are less
810
M. S. Manoj and S. Arulselvi
cost of capturing system and low-resolution picture. In this detection method, various methods like scaling, rotating are used to increase the quality of the captured image. Based on the investigational output this proposed system produces an 88.50% accuracy rate with 0.01 s execution time. The proposed model is executed successfully with dry, soggy condition images also [1]. LinLin Huang et al., 2010 propose a forceful palmprint detection technique. Initially, the author uses salient-point methods are offered to split and find out the ROI region from the input image. Then, ICA is applied to the ROI region to retrieve the major attributes. At last, uses PNN (polynomial neural network) to classify decreased features. The efficiency of the recommended model is illustrated in testing [2]. Mrinal Kanti Dhar et al., 2017 present a successful approach base on texture for palmprint identification for security. This approach contains three important steps. In the first step, ROI is taken out from the given input hand picture. In the processing stage, attributes are retrieved from palmprint with the help of DCT(discrete cosine transform). In this second stage, the redundant features are removed and frequency coefficients with high values are identified. At last, RBPNN (Radial basis probabilistic function network) classifier is applied for classification. It is the novel method with the combination of PNN and RBFN used to classify the images. Preprocessing approaches are used to identify five major points from the hand picture. The proposed work is tested with the CASIA database and compared with various ANN-based classifiers to validate its effectiveness. Based on the experiment results the proposed work shows detection rate is 88.75% [3]. Peng Chen et al., 2018 proposes a personal detection system known as LPVPPIS economically. This system uses the palmprint and palm vein features for detection. It contains LED arrays, infrared, low price sensors, Xilinx-type chips, and various components. The IQA (image quality assessment) approach is offered in this work for merging palmprint and palm vein attributes. Two kinds of LED components are used to extract the palmprint and palm vein. The deep diffusion CNN model is recommended for retrieving attributes from the fused type images and multi-class SVM is applied for the detection and training process. The proposed model was executed with the IQA technique with the fusion method. It also working with fewer requirements than existing fusion techniques, uses low memory, and produces less error value than traditional feature retrieval techniques. The efficiency of the developed system is tested using the PolyU dataset [4]. Huikai Shao et al., 2017 explore a better palmprint identification system with the combination of hashing technology and information refinement. Here the palmprint images are changed into binary type codes using hashing approach and store for quick comparison. Here the author uses 30,000 palmprint pictures gathered from five various smartphones and also they labeled 14 major points on every image of ROI retrieval. Key points are generated using the regression approach. The outcome indicates the achievability of the current dataset and the importance of palmprint detection in authentication systems. With hashing technique, this proposed system is executed with a 97.49% accuracy rate and the EER value is 0.607% [5]. Out of various palmprint detection systems, SRC (sparse representation for classification) is one of the major techniques and it offers the best accuracy rate. Mad Rida et al., 2017 suggest an idea to increase the performance of the palmprint detection system
Artificial Neural Network Based Biometric Palm Print Recognition System
811
using SRC with ensemble type sparse illustration. The major advantage of ensemblebased learning is to decrease the value of sensitivity due to the small size of data used in a training phase. To satisfy the SRC theory, a novel space is created using increasing and decreasing inter and intraclass differences two dimensional LDA. The recommended approach is tested on a multispectral dataset and PolyU dataset. An investigational result demonstrates the above-mentioned two datasets have shown a better result than normal datasets [6]. Palmprint and palm vein can be retrieved easily with the help of a multispectral system. Selma Trabelsi et al., 2019 uses the CNN classifier to retrieve the palmprint attributes and also create an efficient DL-based multispectral-based palmprint detection model. Finally, they concluded the fusion-based model with multimodal devices provides better results than unimodal devices [7].
3 Proposed System Palmprint detection system is one of the major issues in identification systems like mobile phones and other smart devices. Palmprint identification had more benefits than other biometric systems. Typical DL approaches are not fit for compact authorization systems. Comparing with the other biometric, palm prints have many benefits like stable line feature, the text of rich features, imaging with low resolution, capturing of low-cost pieces of equipment. Thus personal authentication based on palm print has become an active research topic. Figure 2 demonstrates a block diagram of the proposed palmprint detection system.
Fig. 2. Block diagram of palm print detection system
The proposed detection system is divided into various stages. Image capturing is the initial phase. Here the palmprint hand images are scanned with the help of a flat model scanner. In the next stage, the unwanted information is removed using preprocessing methods. The scanner scans the entire hand image. During preprocessing stage the palmprint is extracted from the scanned hand image. A proper preprocessed image produces better classification results. The exact preprocessing process retrieves the exact ROI region. It also recovers the constant size ROI region from captured data that has
812
M. S. Manoj and S. Arulselvi
normalized to decrease the rotating modifications. The main intention of split the ROI is to normalize the selected portions automatically. During the feature retrieval phase, the major attributes of the palmprints are extracted using an ANN classifier. ANN is based on the collection of interconnected nodes or units called artificial neurons.ANN utilizes the brain for developing the algorithm which can be utilized for the complex patterns and the detection of issues. According to Huikai Shao et al., 2019 [5] the NN changes the output value of the final layer transfer to the softmax type later to attain the usage of categorization. exp(zi ) qi = zi exp(zi )
(1)
From Eq. 1 zi indicates the output value of the ith neuron in the final layer, similar to ith stage. qi , indicates the probability value of the given input picture fits with the ith group. After extracting the features it is compared with the stored database features. If the features are matched with the stored features it displays the message as the people are identified. If not it displays that the people are not identified.
4 Result and Discussion Now let us see the see how the palm print is been segmented and classified by our proposed ANN. The input original images is been collected from the online datasets. Then this image is fed into the program. At first the image will be Resized and converted into black and white image according to the program and the noise present inside the
Fig. 3. Input palm image
Artificial Neural Network Based Biometric Palm Print Recognition System
813
Fig. 4. Grayscale palm image
image is been analyzed and removed by median filter. Then segmentation is done by Fuzzy C means Clustering algorithm and classification is done through ANN. The following Fig. 3 represents the input palm image and Fig. 4 represents the Greyscale image. Figure 5 represents the contrast image. Then Fig. 6 represents the segmentation process and Fig. 7 represents the classified user image using ANN. The last Fig. 8 represents the comparison graph between ANN and SVM algorithms.
814
M. S. Manoj and S. Arulselvi
Fig. 5. Contrast palm image
Specificity Table No of Images
Random Forest Accuracy %
Support Vector Machine Accuracy %
Artificial Neural Network Accuracy %
100
81.3
85.6
87.54
200
81.52
86
88
300
81.91
86.14
88.5
400
82
86.2
88.7
500
83.15
87.64
88.9
600
82.73
87.82
89
Artificial Neural Network Based Biometric Palm Print Recognition System
Fig. 6. Segmented palm output
Fig. 7. Classified palm output
815
816
M. S. Manoj and S. Arulselvi
Fig. 8. Comparison graph
Artificial Neural Network Based Biometric Palm Print Recognition System
Specificity Graph 90 88 86 84 82 80 78 76 100
200
300
Random Forest Accuracy %
400
500
600
Support Vector Machine Accuracy %
ArƟficial Neural Network Accuracy %
Accuracy Table No of Images
Random Forest Accuracy %
Support Vector Machine Accuracy %
Artificial Neural Network Accuracy %
100
81
87.1
91.21
200
81.5
88.5
92.43
300
82.74
89.22
92.72
400
83.41
89.74
93.44
500
83.53
89.97
93.78
600
84.55
90
93.92
817
818
M. S. Manoj and S. Arulselvi
Accuracy Graph 95 90 85 80 75 70 100
200
300
Random Forest Accuracy %
400
500
600
Support Vector Machine Accuracy %
ArƟficial Neural Network Accuracy %
5 Conclusion In recent days, biometric authentication systems are used in most security devices. Biometric detection means to verify the people depends on physiological or behavioral features. From the various physiological features, palmprint contains various attributes. Computer-based personal authentication is also called a biometric which is considered an effective solution. There are many weaknesses in the palm stamp and the position of the hand differs from one to another and the grip of the hand tends to shrink when the input information is obtained. In this research article, palmprint features are retrieved using an ANN classifier. It is implemented using the MATLAB software tool. The proposed model is tested with the online dataset. The outcome of the proposed model is compared with the other classifiers in terms of accuracy rate.
References 1. Istiqamah, I., Yanuar, F., Wibawa, A.D., Sumpeno, S.: Line hand feature-based palm-print identification system using learning vector quantization. In: IEEE 2016 International Seminar on Application for Technology of Information and Communication, pp. 253–260 (2016) 2. Huang, L., Li, N.: Palmprint recognition using polynomial neural network. In: Zhang, L., Lu, B.-L., Kwok, J. (eds.) ISNN 2010. LNCS, vol. 6064, pp. 208–213. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-13318-3_27 3. Dhar, M.K., Dhar, R.K., Hussain, M.S., Islam, M., Rahman, Y.F.: Palmprint identification using radial basis probabilistic neural network. In: IEEE International Conference on Power, Control, Signals and Instrumentation Engineering (ICPCSI-2017), pp. 49–53 (2017) 4. Chen, P., et al.: Design of low-cost personal identification system that uses combined palm vein and palmprint biometric features. IEEE Access 7, 15922–15931 (2018) 5. Shao, H., Zhong, D., Du, X.: Efficient deep palmprint recognition via distilled hashing coding. In: IEEE Xplore, pp. 714–723 (2019)
Artificial Neural Network Based Biometric Palm Print Recognition System
819
6. Rida, I., Al-Maadeed, S., Mahmood, A., Bouridane, A., Bakshi, S.: Palmprint identification using an ensemble of sparse representations. IEEE Access 6, 3241–3248 (2017) 7. Trabelsi, S., Samai, D., Meraoumia, A., Bensid, K., Taleb-Ahmed, A.: An improved multispectral palmprint system using deep CNN-based palm-features. In: IEEE International Conference on Advanced Electrical Engineering, pp. 1–6 (2019)
Author Index
A Acharjee, Chandrika, 41 Achudhan, M., 766 Adhikary, Arpan, 774 Afzal, Anis, 512 Aggarwal, Shruti, 469 Agrawal, Aditya, 528 Agrawal, Dev, 145 Agrawal, Shubham, 261, 271 Akella, Sundeep V. V. S., 345 Ambati, Giriprasad, 28 Ansari, Md Fahim, 512 Arulselvi, S., 808 Aruna, R., 537 Arya, Aayushi, 248 Ashley, Jose, 279 Ashok, Battula, 387 Atmanandmaya,, 271 B Balaji, S., 652, 662 Balambica, V., 766 Barah, Sushree Smrutimayee, 355 Behera, Sasmita, 355 Benabdallah, Hadjer, 132 Bhandari, Mohan, 13 Bhardwaj, Anshul, 306 Bhatta, Bharat, 433 Bhattacharjee, Pratik, 503 Biswas, Sarmista, 237 Biswas, Suparna, 503 Bolanowski, Marek, 638
C Chandini, Battula, 387 Chandole, Mohan Kumar, 387 Chandramohanan, Ranjit, 492 Chapagain, Pralhad, 13 Chatterjee, Santanu, 774 Chekoory, Yogesh Kumar, 376 Chethanasai, K. V., 752 Chitrakar, Roshan, 13 Choudhary, Anupriya, 254 Chowdary, G. Vamsi, 537 Chowdhury, Gulam Mahfuz, 83 ´ Cwikła, Cezary, 638 D Dabhade, Manish Kumar, 290 Damodharan, P., 618 Dampage, Eng.S. U., 206 Darshan, M., 345, 492 Das, Gairik, 62 Das, Priyanka, 115 Deb, Suman, 41 Deepak, Gerard, 145, 154, 164 Deepak, Vishwa, 766 Desai, Ravishankar P., 173 Devi, M. Shyamala, 537 Dey, Shornalee, 83 Dhanasekar, J., 734 Dhiman, Surender, 528 Dinakaran, C., 1 E Esteve, Manuel, 794
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 S. Mekhilef et al. (Eds.): ICEEE 2022, LNEE 894, pp. 821–823, 2022. https://doi.org/10.1007/978-981-19-1677-9
822 F Fathima, K., 707 Francis, Febin, 411 G Ganesan, Manikandan, 766 Ganzha, Maria, 638, 794 Garg, Jai, 528 George, Jossy P., 237 Ghosh, Ankush, 618, 652, 687, 691, 702, 707, 774 Gunathilake, W. D. K., 206 gupta, Anjani, 482 Gupta, Vansh, 528 H Habiba, H. Umma, 678, 687, 691, 702 Hasteer, Nitasha, 306 Hemalatha, B., 652, 662 Hota, Malaya Kumar, 577 J Jagan, Vadthya, 397 Jain, Ashu, 482 Jain, Dhyanendra, 482 Jasuja, Deepmala, 444 Jha, Suman, 618 Jha, Vinod, 105 Joshi, Basanta, 433 Jyothi, Naga, 387 K Kalanandhini, G., 669 Kanungo, P., 105 Kar, Tejaswini, 105 Karthik, B., 652, 662, 707 Kavya Reddy, D. L., 327 Kerai, Salim, 132 Kerketta, Sonali, 519 Khalid, Hassan, 365 Khan, Azharuddin, 519 Kharate, G. K., 194 Krishnan, A. H. Methil, 734 Kumar, Anil, 53 Kumar, G. Sasi, 28 Kumar, Gaddam Prathik, 327 Kumar, Ishita A., 577 Kumar, J. P. Manoj, 752 Kumar, Mayur, 53 Kumar, P. Sathish, 707 Kumar, Priyanka, 345, 492 Kumar, S. R. Senthil, 734 Kumar, S. Siva, 707 Kumar, S. Vinoth, 537
Author Index Kumar, Yogesh, 306 Kumari, R. Geshma, 28 L Lacalle Úbeda, Ignacio, 638 Lacalle, Ignacio, 794 Lasrado, Riyona, 279 Lijin, K. L., 456 López, César, 794 M Maharjan, Ram Krishna, 433 Majumder, Koushik, 774 Makwana, Vijay H., 218 Manikandan, V., 752 Manjarekar, Narayan S., 173 Manoj, M. Sowmiya, 808 Manoj, N., 164 Masood, M. H., 691 Masood, Molimoli Hajamohideen, 678 Mekhilef, Saad, 365 Mishra, Divya, 121 Mishra, Suyash, 121 Miyani, Piyush B., 72 Mohammed, S. Sheik, 456 Mridha, Krishna, 618 Mubin, Marizan Binti, 365 Mumtahina, Sadia, 83 Munagala, Surya Kalavathi, 587 Munasingha, T. D., 206 Mungur, Avinash Utam, 376 N Narayan, D. G., 565 Natarajan, M. A., 734 Negi, Kanishka, 327 Nisen, Nefi, 279 Nivelkar, Mukta, 279 Nizamudeen, A., 734 O Obulareddy, Sobhana, 587 Ojha, Rituraj, 154 P Padmavathi, T., 1 Pal, Srikanta, 254 Palau, Carlos E., 638, 794 Panday, Sanjeeb Prasad, 433 Pandey, Amit Kumar, 482 Paprzycki, Marcin, 638, 794 Paszkiewicz, Andrzej, 638 Pati, Swaraj, 355 Patil, Sheetal, 194 Paulus, Rajeev, 53
Author Index Prasad, M. Siva, 537 Prasad, P. Siva, 551, 602 Prashanth, Udutha, 397 Priya, A. Banu, 702 Priya, V., 752 R Raiker, Gautam A., 299 Raj, V. J. Aravinth, 752 Raju, B. V. S. S. Kanaka, 537 Rama Sudha, K., 419 Rashid, Md Moontasir, 83 Raswanth, S. R., 345, 492 Reddy, B. Subba, 261, 314 Revathi, M., 419 Revathy, S., 687 Roy, Sandip, 503 S Sagar, Anil Kumar, 327 Sahana, Subrata, 327 Sangeetha, M., 669 Sant, Amit V., 72 Sarkhel, Gautam, 254 Senthilkumar, K. K., 652, 752 Seth, Vishal, 53 Shah, Bikash, 618 Shakthi Saravanan, S., 492 Shanir, P. P. Muhammed, 456 Sharan, A. P., 734 Sharma, G. V. V., 248 Sharma, Gaurav, 482 Shaw, Rabindra Nath, 618, 662, 669, 678, 687, 691, 702, 774 Shikdar, Tareq Anwar, 83 Shukla, Samiksha, 237 Singh, Aditya Vikram, 577 Singh, Amit Kumar, 519 Singh, Krishna Kumar, 444 Singh, M. P., 444 Singh, Prashant, 482
823 Skandan, S., 492 Sneha, K. C., 565 Somashekhar, P., 565 Soumya, D. R., 327 Sowmya, B., 707 Srikanth, Nandoori, 387 Srinath, P., 565 Srinivas, K. S., 565 Subba Reddy, B., 271, 299, 411 Suhas, M., 565 Sushama, M., 551, 602 Szmeja, Paweł, 794 T Tailor, Dhaval N., 218 Thakur, Anchal Singh, 314 Timalsina, Arun, 13 Tiwari, Abhishek, 469 Tiwari, Sneha, 254 U Udugahapattuwa, D. P. D., 206 Umanand, L., 261, 271, 299, 314, 411 Unnikrishnan, N. Aashna, 669 V Valsan, Vipin, 62 Vanahalli, Supriya, 237 Vaño, Rafael, 794 Vijayalakshmi, G., 662 Vijayan, T., 669 Vijayaram, T. R., 766 W Warhade, Krishna K., 290 Weerasundara, W. A. G., 206 Y Yarlagadda, Venu, 28 Yogesh,, 306