Emerging Technologies in Data Mining and Information Security: Proceedings of IEMIS 2020, Volume 1 [1] 9811599262, 9789811599262

Citation preview

Advances in Intelligent Systems and Computing 1286

Aboul Ella Hassanien  Siddhartha Bhattacharyya  Satyajit Chakrabati  Abhishek Bhattacharya  Soumi Dutta   Editors

Emerging Technologies in Data Mining and Information Security Proceedings of IEMIS 2020, Volume 1

Advances in Intelligent Systems and Computing Volume 1286

Series Editor Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Advisory Editors Nikhil R. Pal, Indian Statistical Institute, Kolkata, India Rafael Bello Perez, Faculty of Mathematics, Physics and Computing, Universidad Central de Las Villas, Santa Clara, Cuba Emilio S. Corchado, University of Salamanca, Salamanca, Spain Hani Hagras, School of Computer Science and Electronic Engineering, University of Essex, Colchester, UK László T. Kóczy, Department of Automation, Széchenyi István University, Gyor, Hungary Vladik Kreinovich, Department of Computer Science, University of Texas at El Paso, El Paso, TX, USA Chin-Teng Lin, Department of Electrical Engineering, National Chiao Tung University, Hsinchu, Taiwan Jie Lu, Faculty of Engineering and Information Technology, University of Technology Sydney, Sydney, NSW, Australia Patricia Melin, Graduate Program of Computer Science, Tijuana Institute of Technology, Tijuana, Mexico Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro, Rio de Janeiro, Brazil Ngoc Thanh Nguyen , Faculty of Computer Science and Management, Wrocław University of Technology, Wrocław, Poland Jun Wang, Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, Hong Kong

The series “Advances in Intelligent Systems and Computing” contains publications on theory, applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually all disciplines such as engineering, natural sciences, computer and information science, ICT, economics, business, e-commerce, environment, healthcare, life science are covered. The list of topics spans all the areas of modern intelligent systems and computing such as: computational intelligence, soft computing including neural networks, fuzzy systems, evolutionary computing and the fusion of these paradigms, social intelligence, ambient intelligence, computational neuroscience, artificial life, virtual worlds and society, cognitive science and systems, Perception and Vision, DNA and immune based systems, self-organizing and adaptive systems, e-Learning and teaching, human-centered and human-centric computing, recommender systems, intelligent control, robotics and mechatronics including human-machine teaming, knowledge-based paradigms, learning paradigms, machine ethics, intelligent data analysis, knowledge management, intelligent agents, intelligent decision making and support, intelligent network security, trust management, interactive entertainment, Web intelligence and multimedia. The publications within “Advances in Intelligent Systems and Computing” are primarily proceedings of important conferences, symposia and congresses. They cover significant recent developments in the field, both of a foundational and applicable character. An important characteristic feature of the series is the short publication time and world-wide distribution. This permits a rapid and broad dissemination of research results. Indexed by DBLP, EI Compendex, INSPEC, WTI Frankfurt eG, zbMATH, Japanese Science and Technology Agency (JST). All books published in the series are submitted for consideration in Web of Science.

More information about this series at http://www.springer.com/series/11156

Aboul Ella Hassanien · Siddhartha Bhattacharyya · Satyajit Chakrabati · Abhishek Bhattacharya · Soumi Dutta Editors

Emerging Technologies in Data Mining and Information Security Proceedings of IEMIS 2020, Volume 1

Editors Aboul Ella Hassanien Faculty of Computers and Information Technology Cairo University Giza, Egypt Satyajit Chakrabati Institute of Engineering and Management Kolkata, West Bengal, India

Siddhartha Bhattacharyya Rajnagar Mahavidyalaya Birbhum, India Abhishek Bhattacharya Institute of Engineering and Management Kolkata, West Bengal, India

Soumi Dutta Institute of Engineering and Management Kolkata, West Bengal, India

ISSN 2194-5357 ISSN 2194-5365 (electronic) Advances in Intelligent Systems and Computing ISBN 978-981-15-9926-2 ISBN 978-981-15-9927-9 (eBook) https://doi.org/10.1007/978-981-15-9927-9 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 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

This volume presents proceedings of the 2nd International Conference on Emerging Technologies in Data Mining and Information Security IEMIS2020, which took place in the Institute of Engineering and Management in Kolkata, India, from 2nd to 4th July 2020. The volume appears in the series “Advances in Intelligent Systems and Computing” (AISC) published by Springer Nature, one of the largest and most prestigious scientific publishers, in the series which is one of the fastest growing book series in their programme. AISC is meant to include various high-quality and timely publications, primarily conference proceedings of relevant conference, congresses and symposia but also monographs, on the theory, applications and implementations of broadly perceived modern intelligent systems and intelligent computing, in their modern understanding, i.e. including tools and techniques of artificial intelligence (AI), computational intelligence (CI)—which includes Data Mining, Information Security, neural networks, fuzzy systems, evolutionary computing, as well as hybrid approaches that synergistically combine these areas—but also topics such as network security, cyber intelligence, multiagent systems, social intelligence, ambient intelligence, Web intelligence, computational neuroscience, artificial life, virtual worlds and societies, cognitive science and systems, perception and vision, DNA and immune-based systems, self-organizing and adaptive systems, e-learning and teaching, human-centred and human-centric computing, autonomous robotics, knowledge-based paradigms, learning paradigms, machine ethics, intelligent data analysis, various issues related to network security, big data, security and trust management, to just mention a few. These areas are at the forefront of science and technology and have been found useful and powerful in a wide variety of disciplines such as engineering; natural sciences; computer, computation and information sciences; ICT; economics; business; e-commerce; environment; health care; life science and social sciences. The AISC book series is submitted for indexing in ISI Conference Proceedings Citation Index (now run by Clarivate), EI Compendex, DBLP, SCOPUS, Google Scholar and SpringerLink, and many other indexing services around the world. IEMIS2020 is an annual conference series organized at the School of Information Technology, under the aegis of the Institute of Engineering and Management. Its idea came from the heritage of the other two cycles of events: v

vi

Preface

IEMCON and UEMCON, which were organized by the Institute of Engineering and Management under the leadership of Prof. (Dr.) Satyajit Chakraborty. In this volume of “Advances in Intelligent Systems and Computing”, we would like to present the results of studies on selected problems of Data Mining and Information Security. Security implementation is the contemporary answer to new challenges in the threat evaluation of complex systems. The security approach in theory and engineering of complex systems (not only computer systems and networks) is based on a multidisciplinary attitude to Information theory, technology and maintenance of the systems working in real (and very often unfriendly) environments. Such a transformation has shaped natural evolution in the topical range of subsequent IEMIS conferences, which can be seen over the recent years. Human factors likewise infest the best digital dangers. This book will be of extraordinary incentive to a huge assortment of experts, scientists and understudies concentrating on the human part of the internet and for the compelling assessment of safety efforts, interfaces, clientfocused outline and plan for unique populaces, especially the elderly. We trust this book is instructive yet much more than it is provocative. We trust it moves, driving per user to examine different inquiries, applications and potential arrangements in making sheltered and secure plans for all. The Programme Committee of the IEMIS2020 Conference, its organizers and the editors of these proceedings would like to gratefully acknowledge the participation of all reviewers who helped to refine the contents of this volume and evaluated conference submissions. Our thanks go to Prof Dr. Zdzisław Pólkowski, Dr. Sushmita Mitra, Dr. Pabitra Mitra, Dr. Indrajit Bhattacharya, Dr. Siddhartha Bhattacharyya, Dr. Celia Shahnaz, Mr. Abhijan Bhattacharyya, Dr. Vincenzo Piuri, Dr. SupavadeeAramvith, Dr. Thinagaran Perumal, Dr. Asit Kumar Das, Prof. Tanupriya Choudhury, Dr. Shaikh Fattah and to our all session chairs. Thanking all the authors who have chosen IEMIS2020 as the publication platform for their research, we would like to express our hope that their papers will help in further developments in the design and analysis of engineering aspects of complex systems, being a valuable source material for scientists, researchers, practitioners and students who work in these areas. Giza, Egypt Birbhum, India Kolkata, India Kolkata, India Kolkata, India

Aboul Ella Hassanien Siddhartha Bhattacharyya Satyajit Chakrabati Abhishek Bhattacharya Soumi Dutta

Contents

Computational Intelligence Mathematical Modelling and Analysis of Walters Liquid Motion with Brownian Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debasish Dey

3

Modelling of Two-Phase Fluid Flow Over a Stretching Sheet and Its Numerical Investigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debasish Dey and Barbie Chutia

15

Boundary Layer Flow and Its Dual Solutions Over a Stretching Cylinder: Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debasish Dey, Rupjyoti Borah, and Bhagyashree Mahanta

27

Towards a Conceptual Modelling of Ontologies . . . . . . . . . . . . . . . . . . . . . . Chhiteesh Rai, Animesh Sivastava, Sanju Tiwari, and Kumar Abhishek Mathematical Modelling of Power Law Fluid Flow Through a Pipe and Its Rheology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debasish Dey and Bhagyashree Mahanta NeurolncRNA: A Database of LncRNAs Associated with Human Neurodegenerative Diseases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Aniruddha Biswas, Aishee De, Kumaresh Singha, and Angshuman Bagchi Affective State Analysis Through Visual and Thermal Image Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Satyajit Nayak, Vivek Sharma, Sujit Kumar Panda, and Satarupa Uttarkabat An Ensemble Deep Learning Method for Diabetes Mellitus . . . . . . . . . . . N. Komal Kumar, D. Vigneswari, Rahul J. Reynold, Jojo Josy, and Jerin C. Prince

39

47

57

65

75

vii

viii

Contents

Memory Fluid Flow Past a Vertical Circular Cylinder and Its Energy Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debasish Dey, Ashim Jyoti Baruah, and Rupjyoti Borah A Comparative Study on Financial Market Forecasting Using AI: A Case Study on NIFTY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bhaskar Mondal, Om Patra, Ashutosh Satapathy, and Soumya Ranjan Behera Hydromagnetic Oscillatory Couette Flow of a Visco-Elastic Fluid with Dust in a Channel with Radiative Heat . . . . . . . . . . . . . . . . . . . Hridi Ranjan Deb Hard Exudates Detection: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Satya Bhushan Verma and Abhay Kumar Yadav Analysis of Visco-elastic Fluid Flow Over an Inclined Cylinder: A Numerical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ardhendu Sekhar Khound, Debasish Dey, and Rupjyoti Borah Convolution Neural Network-Driven Computer Vision System for Identification of Metanil Yellow Adulteration in Turmeric Powder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dipankar Mandal, Arpitam Chatterjee, and Bipan Tudu Identification of Biologically Relevant Biclusters from Gene Expression Dataset of Duchenne Muscular Dystrophy (DMD) Disease Using Elephant Swarm Water Search Algorithm . . . . . . . . . . . . . Joy Adhikary and Sriyankar Acharyya

85

95

105 117

125

137

147

Shadow Detection Using DenseUNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Satyajeet Singh, Sandeep Yadav, Antoreep Jana, and Seba Susan

159

Map Merging for an Autonomous Multi-robotic System . . . . . . . . . . . . . . Arnab Mandal and Chintan K. Mandal

169

On Some Basic Graph Invariants of Unitary Addition Cayley Graph of Gaussian Integers Modulo n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Joy Roy and Kuntala Patra A Brief Survey of Token Curated Registries . . . . . . . . . . . . . . . . . . . . . . . . . Jaspreet Kaur and B. Visveswaraiah Comparative Study of Effective Augmentation Method for Bangla ASR Using Convolutional Neural Network . . . . . . . . . . . . . . . . Md. Raffael Maruf, Md. Omar Faruque, Md. Golam Muhtasim, Nazmun Nahar Nelima, Salman Mahmood, and Md. Maiun Uddin Riad A Novel Human Activity Recognition Strategy Using Extreme Learning Machine Algorithm for Smart Health . . . . . . . . . . . . . . . . . . . . . Dipanwita Thakur and Suparna Biswas

179 189

203

215

Contents

Ranking Countries by Perception of Indigenous Companies: A Social Media Rank Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Saphal Patro, Tejasvin Goyal, Vikky Anand, and N. Jayanthi AIDE—AI and IoT-Enabled Home Automation for Disabled and Elderly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sunita Sahu, Smruti Kshirsagar, Srushti Sachdev, Navjyot Singh, and Anushka Tiwari Synthesizing Uniform Concentric Circular Array Antenna for Minimized Sidelobes Using Teaching–Learning-Based Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kailash Pati Dutta and Gautam Kumar Mahanti Comparing Performance of Ensemble-Based Machine Learning Algorithms to Identify Potential Obesity Risk Factors from Public Health Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ayan Chatterjee, Martin W. Gerdes, Andreas Prinz, and Santiago G. Martinez Krylov Subspace Method Using Quantum Computing . . . . . . . . . . . . . . . Vidushi Jain and Yogesh Nagor

ix

223

233

243

253

271

An IoT-Based Cryptographic Algorithm: Mapping the Credentials for Micro-application in Single Column . . . . . . . . . . . . . . Kishore Kumar, Subhranil Som, Sarvesh Tanwar, and Shishir Kumar

281

Identification of Disease Critical Genes in Preeclampsia Using Squirrel Search Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohitesh Ch Agarwal, Biswajit Jana, and Sriyankar Acharyya

289

An Approach for Bengali News Headline Classification Using LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Md. Rafiuzzaman Bhuiyan, Mumenunnessa Keya, Abu Kaisar Mohammad Masum, Syed Akhter Hossain, and Sheikh Abujar Artistic Natural Images Generation Using Neural Style Transfer . . . . . . Atiqul Islam Chowdhury, Fairuz Shadmani Shishir, Ashraful Islam, Eshtiak Ahmed, and Mohammad Masudur Rahman Routh-Hurwitz Criterion for Stability: An Overview and Its Implementation on Characteristic Equation Vectors Using MATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Aseem Patil Path Finding Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D. D. V. Sai Ashish, Sakshi Munjal, Mayank Mani, and Sarthak Srivastava

299

309

319 331

x

Contents

Improved Transmission Map for Dehazing of Natural Images . . . . . . . . Richa Singh, Ashwani Kumar Dubey, and Rajiv Kapoor

339

Comprehensive Operation of Personalized Astute Wheelchair Having Portable Rooftop and Integrated Piezoelectric Engendered Tires to Assist Societal Advance . . . . . . . . . . . . . . . . . . . . . . . . Rajdeep Chowdhury, Kaushal Kumar Poddar, Sukhwant kumar, Md Adil Alam, Neyaz Ahmed, and Madhurima Som

349

Robust Watermarking Schemes for Copyright Protection of Digital Data: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Poonam Kadian, Vasudha Arora, and Shaifali M. Arora

363

Non-invasive Vein Detection and Visualization Using Fuzzy . . . . . . . . . . . E. Balaraja and M. K. Mariam Bee

375

Finger Vein Pattern Detection Using Neuro Fuzzy System . . . . . . . . . . . . E. Balaraja and M. K. Mariam Bee

385

An Efficient Rash Driving Identification System Based on Neural Networks Using TensorFlow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ayushi Srivastava, Utkarsh Verma, Deepak Arora, and Puneet Sharma

397

Advance Computing Home Automation Using ESP8266 of IOT Module . . . . . . . . . . . . . . . . . . . Swati Swayamsiddha, Diptabrata Mukherjee, and Srinivas Ramavath

409

Optimization of Quadratic Curve Fitting from Data Points Using Real Coded Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Banashree Mandal

419

Blockchain Leveraged Incentive Providing Waste Management System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Omi Akter

429

Securing Healthcare Data with Healthcare Cloud and Blockchain . . . . . Rohit Ranjan and Shashi Shekhar Automated Multimodal Biometric System with Ear and Side Profile Face for Human Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lija Jacob, G. Raju, and K. T. Thomas A Recommendation System for Customizable Items . . . . . . . . . . . . . . . . . Dhananjai Sharma, Tanupriya Choudhury, Bhupesh Kumar Dewangan, Abhishek Bhattacharya, and Soumi Dutta Hydromagnetic Visco-elastic Boundary Layer Slip Flow and Heat Transfer Over a Flat Plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kamal Debnath and Sankar Singha

439

457 471

483

Contents

xi

Application of Carbon Nanotube in Targeted Drug Delivery System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sayeed Islam and Khandker Shafaet-Uz-Zaman

493

Indoor Air Pollutant Prediction Using Time Series Forecasting Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Joy Dutta and Sarbani Roy

499

Implementation of Automatic Soil Moisture Dearth Test and Data Exertion Using Internet of Things . . . . . . . . . . . . . . . . . . . . . . . . . P. Nagaraj, V. Muneeswaran, M. Pallikonda Rajasekaran, K. Muthamil Sudar, and M. Sumithra IDPchain: Blockchain-Based International Driving Permit and Traffic Crime Reporting System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Md. Naimur Rahman, Rownak Kabir, Md. Abdul Hamid, and M. F. Mridha

511

519

Hydromagnetic Visco-elastic Boundary Layer Flow Past an Exponentially Stretching Sheet with Suction or Blowing . . . . . . . . . . . Kamal Debnath and Bikash Koli Saha

533

Reimagining the Indian Healthcare Ecosystem with AI for a Healthy Smart City . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M. A. Jabbar, K. M. V. V. Prasad, and Rajanikanth Aluvalu

543

Analysis of Blockchain Induced Cryptocurrency: Regulations and Challenges of Cryptocurrencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gurkirat Singh, Siddharth Gautam, Prachi, Aishwarya Verma, and Tanish Kaushal

553

An Optimization Algorithm for IoT Enabling Technology . . . . . . . . . . . . Ishita Chakraborty and Prodipto Das

565

Emerging Trends in Medical Science Using Biometric . . . . . . . . . . . . . . . . Nitin Tyagi, Bharat Bhushan, Shreyas Vijay, Harshit Yadav, and Siddharth Gautam

577

Plant Disease Detection Using Convolutional Neural Networks and Remedy Recommendation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . P. A. Keerthana, B. Babu Nayana, R. Neeraja Krishna, S. Sheba, and P. S. Deepthi Spatio-Temporal Prediction of Noise Pollution Using Participatory Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Baidik Chandra, Asif Iqbal Middya, and Sarbani Roy Two-Way Nanoscale Automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debayan Ganguly, Kingshuk Chatterjee, and Kumar Sankar Ray

587

597 609

xii

Contents

Multi-head Watson–Crick Quantum Finite Automata . . . . . . . . . . . . . . . . Debayan Ganguly, Kingshuk Chatterjee, and Kumar Sankar Ray

615

Handwritten Signature Verification System Using IoT . . . . . . . . . . . . . . . Santosh Kumar, Shivani Mishra, Siddharth Gautam, and Bharat Bhushan

621

A Framework for Blockchain Technology Including Features . . . . . . . . . Nitin Tyagi, Siddharth Gautam, Abhishek Goel, and Prince Mann

633

Blockchain-Based Network Architecture for Accountable Auctions and to Overcome Resource Problems . . . . . . . . . . . . . . . . . . . . . . Dev Arora, Siddharth Gautam, Prachi Sachdeva, Sushil Sharma, and Vaishali Sharma Proposed Secure and Robust Voting System Using Blockchain Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jayesh Solanki and Divykant Meva Dynamic Functional Bandwidth Kernel-Based SVM: An Efficient Approach for Functional Data Analysis . . . . . . . . . . . . . . . . . Anima Pramanik, Vikram Nande, Arka Shankar Pradhan, Sobhan Sarkar, and J. Maiti Software as a Service for Upcoming Industries . . . . . . . . . . . . . . . . . . . . . . Divyansh Chaudhary, Praveen Kumar, and Seema Rawat A Ubiquitous Indoor–Outdoor Detection and Localization Framework for Smartphone Users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sajan Rajak, Ayan Kumar Panja, Chandreyee Chowdhury, and Sarmistha Neogy Coalescence of Artificial Intelligence with Blockchain: A Survey on Analytics Over Blockchain Data in Different Sectors . . . . . . . . . . . . . . Tushar Singhal, M. S. Bhargavi, and P. Hemavathi

647

661

673

683

693

703

Biometric Door Lock Using Mobile Fingerprint . . . . . . . . . . . . . . . . . . . . . Harsh R. Patel, Parth R. Karkar, Shitul G. Borad, Hitesh N. Baradiya, and Vatsal H. Shah

713

Overview for Creating mHealth Application . . . . . . . . . . . . . . . . . . . . . . . . Smit Pateliya and Vatsal H. Shah

721

Application and Challenges of Blockchain with IoT . . . . . . . . . . . . . . . . . . Kartikey Bhasin, Prerna Gulati, Aditi Sharma, and Achyut Shankar

729

Conservation of Water and Power Using Solar Panel and Raspberry Pi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. Yogesh, M. K. Mariam Bee, and D. Dhanasekaran

743

Contents

xiii

Water Monitoring and Conservation Using IOT with User-Friendly Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. Yogesh, M. K. Mariam Bee, and D. Dhanasekaran

749

Delivering Green Supply Chain Using Blockchain Technology for Sustainable Environment: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Anushka Srivastava, Vandana Dubey, and Bramah Hazela

759

An Intelligent Embedded AC Automation Model with Temperature Prediction and Human Detection . . . . . . . . . . . . . . . . . F. M. Javed Mehedi Shamrat, Pronab Ghosh, Imran Mahmud, Naimul Islam Nobel, and Md. Dipu Sultan ColdBlocks: Quality Assurance in Cold Chain Networks Using Blockchain and IoT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Karthik N. Menon, Kevin Thomas, Jim Thomas, Denil J. Titus, and Divya James A Comparative Study for Analysis and Prediction of Stock Market Movement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nabanita Das and Satyajit Chakrabati Autonomic Cloud Computing: A Survey Using IBM SPSS Tool . . . . . . . Sangeeta Sangani and Sunil F. Rodd An Exploration on the Study of Methodology of Fish Tracking System Using Heat Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tanupriya Choudhury, Srishti Nigam, Sakshi Dhasmana, and Mohammad Shamoon Enhancing Number Sense of Dyscalculia Learners by Pedagogical Agents: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nitika Goenka, Tanupriya Choudhury, and Neelu Jyoti Ahuja

769

781

791 805

815

827

An IoT-Based Traceability Framework for Small-Scale Farms . . . . . . . . Divya James and T. K. S. Lakshmi Priya

841

Smart Water Distribution Control System for a Smart Building . . . . . . . Shital Chaudhari, Sujata Khandaskar, and Samay Ahuja

853

A Review on P2P File System Based on IPFS for Concurrency Control in Hadoop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jasmine Sethi, Shashank Srivastava, and Divya Upadhyay

863

Securing Internet of Things: Attacks, Countermeasures and Open Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mini Sharma, Bharat Bhushan, and Aditya Khamparia

873

xiv

Contents

Internet of Things (IoT) Toward 5G Network: Design Requirements, Integration Trends, and Future Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohd Arsh, Bharat Bhushan, and Mayank Uppal

887

A Study-Based Review on Blockchain Technology for IoT . . . . . . . . . . . . Tripti Sharma, Sanjeev Kumar Prasad, and Avadesh Kumar Gupta

901

Secure Cloud Services Using Quantum Neural Network . . . . . . . . . . . . . . Surya Bhushan Kumar, Ranjan Kumar Mandal, Kuntal Mukherjee, and Rajiv Kumar Dwivedi

913

A Review of Secure Cloud Storage-Based on Cloud Computing . . . . . . . Piyush Aneja, Akhil Bhatia, and Achyut Shankar

923

IOT-Based Advanced Smart Patient Monitoring System . . . . . . . . . . . . . . Kamakshi Dikshit and Sanjeev Kumar Prasad

935

Analyzing Effects of Architectural Alternatives on Performance of Blockchain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ayushi Srivastava, Abhinav Srivastava, Deepak Arora, and Puneet Sharma Real-Time Driver’s Eye Closure, Yawning Detection for Drowsiness Analysis and Gazing Detection for Distraction Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ankit Sonje, Atharva Dare, Yashrao Swant, Adeeba Rizwan, Roshani Raut(Ade), and Sachin R. Sakhare

943

957

Fog Computing: Issues, Challenges and Tools . . . . . . . . . . . . . . . . . . . . . . . Avita Katal, Vitesh Sethi, Saksham Lamba, and Tanupriya Choudhury

971

Drone Service Management in Emergency Situations . . . . . . . . . . . . . . . . N. Pardhasaradhi and R. Puviarasi

983

Hybrid Vehicle with Power Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . G. Suryanaga Raju and R. Puviarasi

989

Density-Based Adaptive Traffic Control System with OpenCV . . . . . . . . Vikash Chandra Kaushal, Deepak Arora, and Puneet Sharma

997

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1011

About the Editors

Aboul Ella Hassanien is Founder and Head of the Egyptian Scientific Research Group (SRGE) and Professor of Information Technology at the Faculty of Computer and Information, Cairo University. Professor Hassanien has more than 1000 scientific research papers published in prestigious international journals and over 45 books covering such diverse topics as data mining, medical images, intelligent systems, social networks and smart environment. Dr. Hassanien is Founder and Head of Africa Scholars Association in Information and Communication Technology. His other research areas include computational intelligence, medical image analysis, security, animal identification, space sciences and telemetry mining and multimedia data mining. Siddhartha Bhattacharyya is currently the Principal at Rajnagar Mahavidyalaya, Birbhum, India. Prior to this, he was a Professor in the Department of Computer Science and Engineering at Christ University, Bangalore, India. He was also the Principal of RCC Institute of Information Technology, Kolkata, India during March 2017 to December 2019. He served as Senior Research Scientist at the Faculty of Electrical Engineering and Computer Science of VSB Technical University of Ostrava, Czech Republic, from October 2018 to April 2019. He is a co-author of 6 books and co-editor of 75 books and has more than 300 research publications in international journals and conference proceedings to his credit. His research interests include soft computing, pattern recognition, multimedia data processing, hybrid intelligence and quantum computing. Satyajit Chakrabati is Pro-Vice Chancellor, University of Engineering & Management, Kolkata and Jaipur Campus, India, and Director of the Institute of Engineering & Management, IEM. As the Director of one of the most reputed organizations in Engineering & Management in Eastern India, he launched a PGDM Programme to run AICTE approved Management courses, Toppers Academy to train students for certificate courses and Software Development in the field of ERP solutions. Dr. Chakrabati was Project Manager in TELUS, Vancouver, Canada from February 2006 to September 2009, where he was intensively involved in planning, execution, xv

xvi

About the Editors

monitoring, communicating with stakeholders, negotiating with vendors and crossfunctional teams and motivating members. He managed a team of 50 employees and projects with a combined budget of $3 million. Abhishek Bhattacharya is Assistant Professor of Computer Application Department at the Institute of Engineering & Management, India. He did Masters in Computer Science from the Biju Patnaik University of Technology and completed Master of Technology in Computer science from BIT, Mesra. He remained associated as Visiting/Guest Faculty of several universities/institutes in India. He has 5 edited and 3 authored books to his credit. He has published many technical papers in various peer-reviewed journals and conference proceedings, both international and national. He has teaching and research experience of 13 years. His area of research is data mining, network security, mobile computing and distributed computing. He is the reviewer of couple of Journals of IGI Global, Inderscience, and Journal of Information Science Theory and Practice. He is a member of IACSIT, UACEE, IAENG, ISOC, SDIWC and ICSES; Technical Committee member and Advisory Board member of various international conferences such as CICBA, CSI, FTNCT, ICIoTCT, ICCIDS, ICICC, ISETIST and many more. Dr. Soumi Dutta is Associate Professor at the Institute of Engineering & Management, India. She has completed her Ph.D. from the Department of CST, IIEST, Shibpur. She received her B.Tech. in IT and her M.Tech. in CSE securing 1st position (Gold medalist), both from Techno India Group. Her research interests data mining, information retrieval, online social media analysis, micro-blog summarization, spam filtering sentiment analysis and clustering of micro-blogging data. She was the editor in cipr2019, IEMIS2018 Springer Conference for 3 volumes and special issue 2 volumes in IJWLTT. She is a TPC member in various international conferences such as SEAHF, DSMLA, ARIAM and CIPR. She is the peer reviewer in different international journals such as – Journal of King Saud University – Computer and Information Sciences, Springer and Elsevier. She is a member of several technical functional bodies such as IEEE, ACM, MACUL, SDIWC, ISOC and ICSES. She has published several papers in reputed journals and conferences.

Computational Intelligence

Mathematical Modelling and Analysis of Walters Liquid Motion with Brownian Diffusion Debasish Dey

Abstract Time-independent Walters liquid motion past a vertical stretching plate has been studied numerically using MATLAB built-in bvp4c solver technique. Fluid flow is governed by buoyancy force, thermal diffusion and Brownian diffusion. Buoyancy force leads to natural free convection. Conservation laws are represented by partial differential equations. The paper is contributed to find appropriate similarity transformation using Lie group analysis for converting the partial differential equations to its relevant ordinary differential equations. Numerical technique “MATLAB built-in bvp4c solver” has been used to solve the problem and results are discussed graphically. Emphasis is given on the influences of thermal diffusion and Brownian motion parameters on the flow characteristics. Further, both diffusion rates of heat and mass transfers are portrayed in this paper. Keywords Buoyancy force · Walters liquid · Nanofluid · Chemical reaction · Lie-group transformation

1 Introduction In recent years, researchers have shown their interest to derive the solution of boundary layer flow problems using similarity transformations. A lot of work has to be done to define a suitable similarity transformation for converting partial differential equation into its relevant ordinary differential equations. In this work, we have adopted the Lie group technique (following the works of Rosmila et al. [1], Das [2] and Das et al. [3]). Further, we have considered memory fluid containing nano-sized particles. Das [4] and Das et al. [5] in their works have discussed the flow problems of nanofluids over a stretching surface with slips at the surface. Dey [6] has investigated the visco-inelastic nanofluid flow through a porous channel using MATLAB built-in solver bv4pc method. Goyal and Bhargava [7] have studied the visco-elastic nanofluid flow over horizontal stretching surface. Numerical analysis of nanofluid D. Dey (B) Department of Mathematics, Dibrugarh University, Dibrugarh 786004, Assam, India e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_1

3

4

D. Dey

flow with bioconvection has been done by Dey [8] but in this analysis, the constitutive model was based on linear relationship between stress and strain rate. Dey and Borah [9] have discussed visco-elastic nanofluid flow in a rotating system using perturbation scheme. Application of visco-elastic fluid flow is seen in geophysics, chemical engineering (absorption, filtration), petroleum engineering, polymer engineering, soil-mechanics, mechanics of blood flow, paper and pulp technology etc. Buoyancy driven memory fluid flows with heat and mass transfer past a porous medium and inclined surface have been studied by Choudhury and Dey [10, 11] using perturbation technique. Perturbation technique has been used by Dey [12, 13] and Dey and Baruah [14] to study memory fluid flow problems through various geometrical considerations. Memory fluid flow problems mainly observed in the industrial applications and these fluids often constitute nano-sized particles. These types of multiphase flows constitute the concept of nanofluids. Thus, the research in nanofluids has tremendous applications in science and engineering. So, this motivates us for modelling and analysis of nanofluid flow (Brownian motion) using constitutive model of memory fluid. Further, the objective of the work is to extend the work of Goyal and Bhargava [7] by considering the governing memory fluid flow past a vertical stretching surface using Lie-group similarity transformation. Further, in our work, the inclusion of vertical stretching surface influences buoyancy force and as a consequence-free convection appears.

2 Mathematical Formulation We consider a steady two-dimensional flow of memory fluid consists of nano-sized particles past a vertical stretching surface in the presence of Brownian diffusion and thermal diffusion. The fluid has ambient temperatures and concentration as  of the stretching T∞ andC∞ , respectively. Let x —axis be taken along the length   surface and y -axis be taken perpendicular to the surface. Let u  , v  be the compo  nents of velocity along x  and y  axes, respectively, and T and C be temperature and concentration of fluid. In the governing fluid motion, buoyancy force is characterized by the term −ρg and it leads to free convection. The presence of elasticity influences the governing fluid motion and its effect is exhibited by k0 in Eq. (2). Flow is maintained with density difference that leads to buoyancy force and heat is making due to viscosity. We introduce the non-dimensional variables       ∞ ∞ , φ = CCw−C and following the x = √x ν , y = √y ν , u = √ucν , v = √vcν , θ = TTw−T −T∞ −C∞ c

c

above assumptions, the governing equations of motion are:   ∂u ∂ 2u ∂v ∂ 2 u k0 ∂ 2 u ∂ 3u ∂u ∂ 2 u ∂u +v = 2 − u 2 +v 3 − − + Gr θ + Gmϕ u ∂x ∂y ∂y ρ ∂y ∂y ∂ y ∂ x∂ y ∂ y ∂ y2 (1)

Mathematical Modelling and Analysis …

5

  2 ∂θ ∂θ 1 ∂ 2θ DT ∂θ τ1 ∂ϕ ∂θ +v = + u + D B (Cw − C∞ ) (Tw − T∞ ) 2 ∂x ∂y Pr ∂ y ν ∂y ∂y T∞ ∂y (2) T − T∞ ) 1 ∂ 2 θ ∂ϕ 1 ∂ 2ϕ ∂ϕ Kr T∞ (Tw +v = ϕ + − 2 2 ∂x ∂y Pr Le ∂ y D B (Cw − C∞ ) Pr Le ∂ y a

D

u

(3)

Relevant boundary conditions in the dimensionless form are given as u = x, v = 0, θ = 1, φ = 1; y = 0 & u → 0, θ → 0, φ → 0; y → ∞

(4)

3 Method of Solution Let us define a stream function “ψ” such that u=

∂ψ ∂ψ &v=− . ∂y ∂x

(5)

The Eq. (1) reduces to ψ,y ψ,x y − ψ,x ψ,yy = ψ,yyy − a(ψ,y ψ,yyyx − ψ,x ψ,yyyy − ψ,yy ψ,yyx + ψ,x y ψ,yyy ) + Gr θ + Gmφ

(6)

To solve the Eqs. (6), (2) and (3), we use the classical Lie-group approach; the scaling group transformations for the above equations are given as:

: x = e−εC1 x ∗ , y = e−εC2 y ∗ , ψ = e−εC3 ψ ∗ , θ = e−εC4 θ ∗ , φ = e−εC5 φ ∗

(7)

where ε(= 0) is group parameter, C i ’s are the real numbers. The constants, C i ’s are to be determined, so that the equations along with boundary conditions are invariant. Using (7) in (6), (2) and (3) and comparing the exponential terms, we get − 2C3 + 2C2 + C1 = 3C2 − C3 = −2C3 + 4C2 − C4 = −C5 − C3 + C2 + C1 − C4 = −C3 + C2 + C1 − C4 = −C4 + 2C2 = −C5 − C4 + 2C2 = −2C4 + 2C2 − C3 + C2 + C1 = −C3 + C2 + C1 = −C5 + 2C2 = −C4 + 2C2 Solving these, one set of solution is obtained as: C3 = C1 = C4 = C5 , C2 = 0, and the characteristic equations are

6

D. Dey

dy dψ dθ dφ dx = = = = x 0 ψ θ φ

(8)

Solving (8), the scaled transformations are formed as η = y, ψ = x f (η), θ = xθ (η), φ = xφ(η)

(9)

The set of ordinary differential equations are   f 2 − f f  = f  − a 2 f  f  − f f iv − f 2 + Gr θ + Gmφ 1  θ + N bθ  φ  + N tθ 2 Pr

(11)

1 Nt φ  + θ  − Ceφ Pr Le Pr Pr N b

(12)

f θ − f θ  = f φ − f φ =

(10)

The boundary conditions of the problem are η = 0 : f  = 1, f = 0, θ  = 1, φ  = 1 & η → ∞, f  → 0, θ  → 0, φ  → 0 (13) where,g is the acceleration due to gravity and β&β ∗ are coefficient of volume expansions due to heat and mass transfers respectively, ρ density of fluid, k0 viscoelasticity parameter, υ kinematic viscosity, DB Brownian diffusion parameter, DT ∞) thermal diffusion parameter Gr = gβ(Tc√w −T Grashoff number for heat transfer, cν gβ ∗ (C√w −C∞ ) c cν parameter, Pr = αν

Gm =

Grashoff number for mass transfer, a = α DB

k0 c ρν

visco-elastic

Prandtl number, Le = Lewis number, Ce = Kcr ChemDB ical reaction parameter, N b = τ1 ν (Cw − C∞ ) Brownian diffusion parameter, N t = Tτ∞1 DνT (Tw − T∞ ) thermal diffusion parameter.

4 Results and Discussion The dimensionless rate of heat and mass transfers between two fluid layers are given as N u = ∂θ &Sh = ∂φ . Steady two-dimensional flow of memory fluid with Brow∂η ∂η nian diffusion and thermal diffusion past a stretching surface has been studied using Lie-Group similarity transformation. The coupled differential equations cannot be solved analytically, so we have used MATLAB built-in bv4pc solver method. The results are graphical with a special emphasis is given on visco-elasticity, Brownian diffusion and thermal diffusion. In this study, some of the arbitrary values of flow parameters are kept fixed and are given as: Gm = 2; Gr = 5; Pr = 5; Ce = 0.5; Le = 3. Figures 1 and 2 depict the nature of velocity profiles of primary and secondary flows against the displacement variable for various values of flow parameters involved in

Mathematical Modelling and Analysis …

7

Fig. 1 Primary flow against displacement variable For Gm = 2; Gr = 5; Pr = 5; Ce = 0.5; Le = 3; Nb = 0.02; Nt = 0.01

Fig. 2 Secondary flow against displacement variable for Gm = 2; Gr = 5; Pr = 5; Ce = 0.5; Le = 3; Nb = 0.02; Nt = 0.01

the solution. It is seen from the Figs. 1 and 2 that increasing values of visco-elastic parameter decelerate the primary and secondary flows of fluid. Physically, it can be interpreted that there is a reduction in mechanical energy by storage power of elasticity; as a consequence, fluid slows down by its speed. Also, it is seen that in the neighbourhood of the surface, fluid reaches its maximum velocity and then it decreases gradually (0.2 ≤ η ≤ 0.3) as we move towards free stream. A backflow

8

D. Dey

governed by adverse pressure gradient is experienced for higher values of viscoelasticity in the region η ≥ 0.4. A loss of maximum of 58.82% approximately is noticed in fluid speed during the 300% rising in visco-elasticity (Fig. 1). Figures 3 and 4 depict the nature of temperature field against displacement variable for various values of visco-elastic parameter (Fig. 3) and thermal diffusion parameter (Fig. 4),

Fig. 3 Temperature against displacement variable for Gm = 2; Gr = 5; Pr = 5; Ce = 0.5; Le = 3; Nb = 0.02; Nt = 0.01

Fig. 4 Temperature against displacement variable for Gm = 2; Gr = 5; a = 0.4; Pr = 5; Ce = 0.5; Le = 3; Nb = 0.02

Mathematical Modelling and Analysis …

9

respectively. Thermal diffusion parameter (Nt) characterizes the movement of particles from hotter region to cooler region. The figures reveal that temperature is higher at the surface (viscosity dominated region) than the free-stream region. Increasing values of visco-elasticity reduce the flow speed and because of friction, heat is generated. As a result, temperature rises in the neighbourhood of the surface. There is an enhancement in temperature boundary layer during the rise in visco-elasticity and the layer is seen in the region 0 ≤ η ≤ 0.6 (approximately). We have noticed a thermal stability region at η > 0.6 (Fig. 3) but it is reduced during the growth of thermal diffusion. The model must be framed in such a way that the temperature should not be very high and for that combination of thermal diffusion and viscoelasticity of fluid plays a crucial role. The dimensionless temperature of fluid flow experiences negative values in Fig. 4. This negative value of θ indicates that temperature of free-stream region is higher than temperature at the surface (Tw < T∞ ) and it is seen in the region η > 0.5 (Fig. 4). It may be interpreted that the presence of nanofluid increases the thermal conductivity of the fluid, so there is high conductive heat transfer. Thus fluid free-stream region experiences higher temperature than temperature at the surface. There is a fall in temperature during the enhancement of thermal diffusion parameter. The thermal diffusivity reduces the growth of heat energy and so, temperature falls down. Effects of Brownian diffusion and thermal diffusion on the concentration field are shown by Figs. 5 and 6, respectively. It is seen from the figures that deposition of nanoparticles is maximum at the surface and it gradually decreases with the increasing values of η. Movement of Brownian particles reduces the concentration level of fluid. It is also seen that there is no significant

Fig. 5 Concentration against displacement variable for Gm = 2; Gr = 5; a = 0.4; Pr = 5; Ce = 0.5; Le = 3; Nt = 0.02

10

D. Dey

Fig. 6 Concentration against displacement variable for Gm = 2; Gr = 5; a = 0.4; Pr = 5; Ce = 0.5; Le = 3; Nb = 0.02

difference in the concentration at the neighbourhood of the surface and at free-stream region. Maximum impact of Brownian diffusion is seen [0.4, 1.2]. Thermal diffusion characterizes the movement of particles due to temperature gradient and it can be concluded that the deposition is enhanced by thermal diffusion. Nusselt number characterizes the rate of heat transfer and Fig. 7 represents the rate of heat transfer across the flow for various values of Brownian diffusion. It is observed that there is Nusselt number is varying periodically with the Brownian diffusion across the flow (Fig. 7). Also, it can be concluded that at the surface, rate of heat transfer experiences decreasing trend with increasing values of Nb. Sherwood number signifies the rate of deposition and Fig. 8 characterizes the pattern of Sherwood number across the flow for Brownian diffusion. There is a periodical variation on Sherwood number across the flow. Growth of Brownian diffusion reduces the rate of deposition of nano-sized particles in the vicinity of the surface.

5 Conclusion i.

ii. iii.

Strength: This work is based on using the constitutive model of (Walters liquid) memory fluid flow. Memory fluid flows have tremendous applications in science and engineering purposes. Weakness: This theoretical work is carried out by means of some arbitrary values of parameters used in the flow configuration. Future scope: In the future,

Mathematical Modelling and Analysis …

11

Fig. 7 Nusselt number against displacement variable for Gm = 2; Gr = 5; a = 0.4; Pr = 5; Ce = 0.5; Le = 3; Nt = 0.01

Fig. 8 Sherwood number against displacement variable for Gm = 2; Gr = 5; a = 0.4; Pr = 5; Ce = 0.5; Le = 3; Nt = 0.01

12

D. Dey

• this work may be used for the experimental purpose for better accuracy. • real values of the parameters may be used for the required result in engineering and medical sciences. iv.

Major contribution from the work: Some of the points from the above study are highlighted below: • Visco-elasticity of material helps to reduce the momentum and kinetic energy of the system by reducing the speed of the fluid flow. • In the viscous dominant region (neighbourhood of the surface), temperature is seen maximum. • Reduction in heat transfer rate may be controlled by thermal diffusion. • Deposition of Brownian particles is higher in the neighbourhood of the surface.

References 1. Rosmila, A.B., Kandasamy, R., Muhaimi, I.: Lie symmetry group transformation for mhd natural convection flow of nano-fluid over linearly porous stretching sheet in presence of thermal stratification. Applied Mathemat. Mechan. 33(5), 593–604 (2012) 2. Das, K.: Lie group analysis of stagnation-point flow of a nanofluid. Microfluid. Nanofluid. 15(2), 267–274 (2013) 3. Das, K., Acharya, N., Kundu, P.K.: MHD micropolar fluid flow over a moving plate under slip conditions: an application of lie group analysis. U.P.B. Sci. Bull., Series A. 78(2), 225–234 (2016) 4. Das, K.: Slip flow and convective heat transfer of nano fluids over a permeable stretching surface. Comput. Fluids 64, 34–42 (2012) 5. Das, K., Duari, P.K., Kundu, P.K.: Solar radiation effects on cu–water nanofluid flow over a stretching sheet with surface slip and temperature jump. Arabian J. Sci. Eng. 39(12), 9015–9023 (2014) 6. Dey, D.: Brownian motion effects on fluid flow through a channel: an application of power law model. J. Assam Acad. Mathemat. 7, 28–41 (2017) 7. Goyal, M., Bhargava, R.: Numerical solution of MHD viscoelastic nano fluid flow over a stretching sheet with partial slip and heat source/sink. ISRN Nanotechnology, (2013). DOI: http://dx.doi.org/https://doi.org/10.1155/2013/931021 8. Dey, D.: Modelling and analysis of bio-convective nano fluid flow past a continuous moving vertical cylinder. Emerging Technologies in Data Mining and Information Security, 331–340 (2019) 9. Dey, D., Borah, A.J.: Visco-elastic effects on nanao fluid flow in a rotating system in presence of Hall current effect. Emerging Technologies in Data Mining and Information Security, 575–585 (2019) 10. Choudhury, R., Dey, D.: Free convective visco-elastic flow with heat and mass transfer through a porous medium with periodic permeability. Int. J. Heat Mass Transf. 53(9–10), 1666–1672 (2010) 11. Choudhury, R., Dey, D.: Free convective elastico-viscous fluid flow with heat and mass transfer past an inclined plate in slip flow regime. Latin Am. Applied Res. 42(4), 327–332 (2012) 12. Dey, D.: Non-newtonian effects onhydromagnetic dusty stratified fluid flow through a porous medium with volume fraction. Proceedings of the National Academy of Sciences. India Section A: Phys. Sci. 86(1), 47–56 (2016)

Mathematical Modelling and Analysis …

13

13. Dey, D.: HydromagneticOldroyd fluid flow past a flat surface with density and electrical conductivity. Latin Am. Applied Res. 47(2), 41–45 (2017) 14. Dey, D., Baruah, A.J.: Stratified visco-elastic fluid flow past a flat surface with energy dissipation: an analytical approach. Far East J. Applied Mathemat. 96(5), 267–278 (2017)

Modelling of Two-Phase Fluid Flow Over a Stretching Sheet and Its Numerical Investigation Debasish Dey and Barbie Chutia

Abstract Three-dimensional dusty fluid flow with volume fraction has been investigated in this paper. The presence of magnetic field and non-uniform heat source/sink is considered along with the flow over a stretching sheet enclosed in a porous medium. Using similarity transformation, the governing equations are transformed into ordinary differential equations. Matlab built-in solver Bvp4c has been used to solve the resulting non-linear ordinary differential equations. The effects of various parameters on the flow have been discussed and characteristics of heat transfer are determined. The results are shown using graphs and tables. Keywords Dusty fluid · Three dimensional · Stretching sheet · Volume fraction · Porous medium · Heat source/sink

1 Introduction The study of two-phase fluid flow has been started by Saffman [1] in 1962. Many researchers have studied the boundary value problems of dusty fluid flow as it has many applications in various fields like waste-water management, combustion, engineering and many more. The investigation of MHD flow of dusty fluid has been done by several scientists such as Vajravelu and Nayfeh [2], Gireesha et al. [3], Sarangi and Mishra [4], Dey ([5, 6]), Kumar et al. [7], Dey ([8, 9]) etc. The study of three-dimensional flow has importance in different areas like aeronautical engineering, geophysics, insulation engineering, solar collector, grain storage devices and many more. Analysis of boundary layer three-dimensional flow was proposed by Wang [10]. Due to its vast applications, many researchers have D. Dey · B. Chutia (B) Department of Mathematics, Dibrugarh University, Dibrugarh 786004, Assam, India e-mail: [email protected] D. Dey e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_2

15

16

D. Dey and B. Chutia

taken keen interest in these problems. Some of such researchers are Mallikarjuna et al. [11], Mohaghegh and Rahimi [12] etc. The presence of volume fraction of dust particles has been neglected by many scientists. Due to which, the high density fluid is not justified. Recently, the influence of volume fraction has been taken into consideration by various researchers. The study of non-Newtonian dusty flow with volume fraction has been discussed by Ahsan et al. [13], Dey [14]. Effect of volume fraction on two-phase incompressible fluid flow has been studied by Rath et al. [15]. Dey [16] has been illustrated the analytical solutions of dusty Jeffrey fluid flow in presence of volume fraction. The main aim of the present study is to discuss the effect of volume fraction on three-dimensional dusty fluid flow towards a permeable stretching sheet with magnetic field. The ordinary differential equations after similarity transformation are then solved numerically using Matlab built-in Bvp4c solver. For various parameters, velocity and temperature graphs have been depicted and skin friction, heat transfer have been characterized in tabular form.

2 Mathematical Formulation of the Problem A time independent, three-dimensional viscous multi phase fluid flow containing dust particles towards a stretching surface has been taken into consideration. Flow is governed by magnetic force due to the transverse magnetic field and resistive force due to permeability. The sheet is stretched along with the xy plane and flow of the fluid is assumed to be in z direction. The coordinate system and flow configuration are shown in Fig. 1. Some of the assumptions are taken into account for dust phase such as dust particles are assumed small spherical in shape where density of dust particles is taken constant. Magnetic Reynolds number is considered small for weak electrical Fig. 1 Flow configuration of the problem

Modelling of Two-Phase Fluid Flow Over …

17

conductivity, magnetic field as compared to applied one. The governing equations of incompressible dusty fluid in three dimensions are as follows: ∂v ∂w ∂u + + = 0 ∂x ∂y ∂z   2 ∂u σ B02 u ∂u ∂u ∂ u νu ∂ 2u ∂ 2u u − +v +w =ν − + + 2 2 2 ∂x ∂y ∂z ∂x ∂y ∂z ρ k   ρp up − u + (1 − ϕ)ρτv   2 σ B02 v ∂v ∂v ∂ v νv ∂v ∂ 2v ∂ 2v − +v +w =ν − u + + ∂x ∂y ∂z ∂x2 ∂ y2 ∂z 2 ρ k   ρp vp − v + (1 − ϕ)ρτv ∂(ρ p u p ) ∂(ρ p v p ) ∂(ρ p w p ) + + =0 ∂x ∂y ∂z   2 ∂v p ∂v p ∂v p σ B02 v ∂ v ∂ 2v ∂ 2v up − ϕ + vp + wp = νϕ + + ∂x ∂y ∂z ∂x2 ∂ y2 ∂z 2 ρ   1 νv − vp − v −ϕ k τv   2 ∂w p ∂w p ∂w p σ B02 w ∂ w ∂ 2w ∂ 2w − ϕ + vp + wp = νϕ up + + ∂x ∂y ∂z ∂x2 ∂ y2 ∂z 2 ρ  1 νu − wp − w −ϕ k τv  2  ρ p (T p − T ) ∂T ∂T α ∂ T ∂T + v + w = + u 2 ∂x ∂y ∂z ρC p ∂z ρC p τT     2  ρp 1 q  2 u p − u + vp − v + + ρC p τv ρC p up

C p (T p − T ) ∂ Tp ∂ Tp ∂ Tp + vp + wp = ∂x ∂y ∂z Cm τT

wher e q  =



 αUw (z)  ∗ A (Tw − T∞ ) f  (η) + B ∗ (T − T∞ ) zϑ

The boundary conditions of the problem are as follows:

(1)

(2)

(3) (4)

(5)

(6)

(7) (8)

(9)

18

D. Dey and B. Chutia

⎫ u = u w , v = vw (z), w = 0, −k ∂∂Ty = h f (Tw − T ) at z = 0 ⎪ ⎪ ⎪ ⎬ u p = u → 0, w p → w, v p = v → 0, ρ p → ρw, ⎪ T → T∞ , T p → T∞ at z → ∞ ⎪ ⎪ ⎭ wher e u w = cλx, vw = cy and C p = Cm (assumption)

(10)

3 Method of Solution The governing equations are transformed into the corresponding ODEs using following appropriate similarity transformations:  ⎫  √ c ⎪ u = cλx f  (η), v = cy f  (η) + g  (η) , η = v z, w = − cv[g(η) + (λ + 1) f (η)], ⎪ ⎬ √ u p = cλx F(η), v p = cy[F(η) + G(η)], w p = cv[G(η) + (λ + 1) K (η)], ⎪ ⎪ ρp T −T ⎭ ρ = = H (η), θ(η) = T −T∞ , θ = p ∞ r

ρ

Tw −T∞

p

(11)

Tw −T∞

Putting Eq. (11) in Eqs. (2)–(9), we have the following ODEs: f  + (g + (λ + 1) f ) f  − λ f 2 + β∈H (F − f  ) − (M + S) f  = 0 (12)

( f  + g  ) + (g + (λ + 1) f ) ( f  + g  ) − ( f  + g  )2 + β∈H (F + G − f  − g  ) − (M + S)( f  + g  ) = 0

(13)

      (G + (λ + 1)K )H + (λ + 1)F + G + G + (λ + 1)K H = 0

(14)

 e (F + G)2 + (G + (λ + 1)K )(F  + G  ) = f  + g  − (M + S)( f  + g  ) − βe(F + G − f  − g  ) (15)

 e λF 2 + (G + (λ + 1)K )F  = f  − (M + S)( f  ) − βe(F − f  )

(16)

(G + (λ + 1)K )(G  + (λ + 1)K  ) + (g  + (λ + 1) f  ) = (M + S)(g + (λ + 1) f ) − βe(G + g + (λ + 1)(K + f ))

(17)

Modelling of Two-Phase Fluid Flow Over …

19

θ  + Pr(g = (λ + 1) f )θ  + Pr βτ (θ p − θ )H  + Pr β Ecx (F − f  )2 + Ec y (F − f  + G − g  )2 H + λ(A∗ f  + B ∗ θ ) = 0 (18)



(G + (λ + 1)K )θ p +

 Cp  βτ θ p − θ = 0 Cm

(19)

Boundary conditions of Eqs. (10) and (11) after transformation become: 





f = 0, f = 1, g = g = 0, θ = −Bi (1 − θ ) at η = 0 



f = F → 0, g = G → 0, K = − f −

g , H → ω, θ p = θ → 0 as η → ∞ λ+1

The surface shear stresses and heat transfer rate are given by: τzx

      ∂w ∂w ∂u ∂v ∂T + + = −μ , τzy = −μ qw = −k ∂z ∂ x (z=0) ∂z ∂ y (z=0) ∂ y z=0

Using the non-dimensional variables, we obtain the following friction factor and local Nusselt number: 1

C f x Rex2 = − f  (0),

1

C f y Re y2 = −g  (0),

 N u x = − Rex θ  (0)

4 Results and Discussion Numerical procedure Matlab built-in solver Bvp4c has been adopted to carry out the solution of the Eqs. (12)–(19) with reference to boundary condition. Influence of flow parameters such as fluid-particles interaction parameter for velocity (β), fluidparticles interaction parameter for temperature (βτ ), velocity ratio (λ) on the flow and heat transfer of dusty fluid has been shown graphically and in tabular form respectively. Figure 2 describes the effect of fluid-particle interaction parameter for temperature (βτ ) on the temperature profiles of fluid phase and dust phase, respectively. We have obtained from Fig. 2 that as there is an increment in the values of βτ , the temperature profile for fluid phase and dust phase boost up. As a result, the increasing values of βτ lead to the intensification of conductive heat transfer.

20

D. Dey and B. Chutia

0.6 0.5 0.4 0.3 0.2

0.025, 0.5, 1

0.1 0 -0.1

0

0.5

1

1.5

2

2.5

3

Fig. 2 Temperature profiles of fluid and dust phase for distinct values of βτ

1 0.9 0.8 0.7

f’, F

0.6

= 0.025, 0.3, 0.5

0.5 0.4 0.3 0.2 0.1 0

0

0.5

1

1.5

2

2.5

Fig. 3 Velocity profiles (u, u p ) of fluid and dust phase for distinct values of β

3

Modelling of Two-Phase Fluid Flow Over …

21

4 3.5 3 2.5

'

2 1.5 1 0.5 0

0

0.5

1

1.5

2

2.5

3

Fig. 4 Velocity profiles (v, v p ) of fluid and dust phase for distinct values of λ

Figure 3 is plotted to obtain the effect of fluid-particle interaction parameter (β) on velocity components u, u p for both fluid and dust phase. From Fig. 3, we get to know that as the fluid-particle interaction parameter raises, the velocity components for both the phases decline. Since whenever there is an enhancement in the values of β, the contribution of dust particles of the fluid velocity increases. As a result, change in the value of β affects more in the dust phase than the fluid phase. Therefore, from Fig. 3, we can see that the velocity profile of dust phase is higher than that of the fluid phase. Figures 4 and 5 are plotted to see the influence of velocity ratio (λ) on the velocity components v, w for both the phases. From Fig. 4, we observed that the velocity profile of the dust phase is greater than the fluid phase. Also, we depicted the values of the velocity ratio λ rises with the declination in the velocity profiles for both fluid and dust phases. Figure 5 describes the effect of λ for the velocity component w of both the phases. It can be seen that with enhancement in the values of λ, the velocity component w for fluid phase falls off while velocity component w p of the dust phase rises. Table 1 shows the variation of various parameters with the magnitude of the coefficient of skin friction and heat transfer rate. We have obtained the results from which we can say that the increasing values of permeability parameter, fluid-particle interaction parameter, volume fraction and velocity ratio leads to the increment in the magnitude of the coefficient of skin friction along x direction, while the magnitude of the coefficient of skin friction along y direction declines for higher values of volume fraction parameter and velocity ratio. On the other hand, the rate of heat transfer rises for increasing values of e, βT , Ecx and Ec y .

22

D. Dey and B. Chutia

)K

4 2

+(

0 -2

= 0.025, 0.5, 1

-4

,

-6 -8 -10

0

1

0.5

1.5

2.5

2

3

Fig. 5 Velocity profiles (w, w p ) of fluid and dust phase for distinct values of λ Table 1 The value of βT , S, e, λ, Ecx and Ec y βT

f  (0), −g  (0) and

− f (0) 1.5590

2.4585

−0.3336

0.5

1.2

1

0.01

0.01

1.5590

2.4585

−0.3225

1.5590

2.4585

−0.3121

1.7151

3.2948

−0.3229

2.0092

4.7454

−0.3034

2

1.2

1

0.01

0.01

0.5

1.4

1

0.01

0.01

1.6 1 0.5

1.2

2

0.01

0.01

3 0.2 0.5

1.2

1

0.3

0.01

0.4 0.5



Ec y

1.2

0.5



0.5

1.2

1

0.01

−g (0)

values

Ecx

3

0.5

various

λ

1

0.5



for

e

0.5 0.5

θ  (0)

S

0.025 0.3

−

−θ (0)

2.8421

6.6399

−0.0064

1.5590

2.4585

−0.3121

1.5532

2.3622

−0.3227

1.5496

2.2894

−0.3308

1.5590

2.4585

−0.3121

1.8345

0.6848

−0.3645

2.0879

0.3547

−0.4188

1.5590

2.4585

−0.3119

1.5590

2.4585

−0.3118 −0.3117

1.5590

2.4585

0.2

1.5590

2.4585

−0.3082

0.3

1.5590

2.4585

−0.3061

0.4

1.5590

2.4585

−0.3040

of

Modelling of Two-Phase Fluid Flow Over …

23

5 Conclusions In this paper, steady three-dimensional dusty fluid flow past a stretching sheet in the presence of volume fraction and non-uniform source/sink embedded in a porous medium has been investigated numerically. Some of the conclusions are as follows: • A rise in the values of fluid-particle interaction parameter leads to the declination in the velocity profile along x axis for both the fluid and dust phases. • Temperature profile of dust and fluid phases increases as the fluid-particle interaction parameter for temperature boosts up. • The magnitude of the skin friction coefficient (C f x ) in x direction increases for the increasing values of permeability parameter, fluid-particle interaction parameter and relaxation parameter, volume fraction and velocity ratio. On the other hand,(C f y ) along y direction decreases for increasing values of volume fraction parameter and velocity ratio. • A rise in the values of permeability parameter, Eckert number in x and y directions and thermal particle-interaction parameter enhance the thermal diffusion rate. Limitations: Some of the limitations are following: • The flow is confined to laminar and incompressible where density is taken constantly. • Constitutive equations are based on linear relation. • The study has been done theoretically taking arbitrary values of the flow parameter. Future Scope: The investigations could be done further in different geometric configurations with the influence of various parameters like thermophoresis, micropolar material parameter, Grashof number, radiation parameter and non-Newtonian models. There is a possibility that a collaborative work considering turbulent flow could be done in the laboratory. Also, if we have the proper industrial values of parameters then it could be useful in the experimental process.

6 Nomenclature (u, v.w) → Velocity components of fluid phase, (up , vp , wp ) → Velocity components of dust phase, ρ → Density of the fluid, ρ p → Density of dust phase, K(= 6) → Stoke’s constant, β = τvl c → Fluid-particle interaction parameter, βT = τTl c → Fluidparticle interaction parameter for temperature, μ → Dynamic viscosity of the fluid, d → Radius of the dust particles, B0 → Magnetic field, k → Permeability of the porous medium, σ → Electrical conductivity of the fluid, φ → Volume fraction, Cp , Cm → Specific heat of fluid and dust particles respectively, τT → Thermal equilibrium time, τv → Relaxation time of the dust particles, q  → Space and temperature dependent internal heat generation (non-uniform heat source/sink), α → Thermal conductivity, ν → Kinematic viscosity of the fluid, Tw , T∞ → Temperature at the wall and at

24

D. Dey and B. Chutia

a large distance respectively, (T, Tp ) → Temperature of fluid and dust particles μl respectively, S = bρk → Permeability parameter, e = φ1 → Volume fraction p −

parameter for dust phase, = Pr =

ρϑCp α

1 1−φ

→ Volume fraction parameter, for fluid phase,

→ Prandtl number, Ecx =

2 uw (Tw −T∞ )Cp

→ Eckert number in x-direction,

2 vw (Tw −T∞ )Cp

→ Eckert number in y direction, ω → Density ratio. A*, B*→ Ecy = Coefficient of space and temperature dependent internal heat generation/absorption respectively, A, D → Constants, Uw (z) → Stretching velocity along x direction, vw (z) → Stretching   velocity along y direction, c → Stretching rate, η → Similarity ρp variable, ρr = ρ → Relative density, Rex → Local Reynolds number, Bi = ν hf → Biot number, C f x → Friction factor along x direction, C f y → Friction c α factor along y direction, λ → Ratio of the velocities in y and x directions, , h f → Convective heat transfer.

References 1. Saffman, P.G.: On the stability of laminar flow of a dusty gas. J. Fluid Mechan. 120–128 (1962) 2. Vajravelu, K., Nayfeh, J.: Hydromagnetic flow of a dusty fluid over a stretching sheet. Int. J. Nonlinear Mechan. 27, 937–945 (1992) 3. Gireesha, C.S., Ramesh, B.J., Abel, G.K., Bagewadi, M.S.: Boundary layer flow and heat transfer of a dusty fluid flow over a stretching sheet with non-uniform heat source/sink. Intern. J. Multiphase Flow 37, 977–982 (2011) 4. Sarangi, G.C., Mishra, S.K.: Boundary layer flow and heat transfer of a dusty fluid over a stretching sheet. Eng. Sci. Technol. Internat. J. 5, 2250–3498 (2015) 5. Dey, D.: Dusty hydromagnetic oldryod fluid flow in a horizontal channel with volume fraction and energy dissipation. Internat. J. Heat Technol. 34(3), 415–422 (2016) 6. Dey, D.: Hydromagnetic oldroyd fluid flow past a flat surface with density and electrical conductivity stratification. Latin Am. Applied Res. 47(2), 41–45 (2017) 7. Kumar, K.G., Gireesha, B.J., Gorla, R.S.R.: Flow and heat transfer of dusty hyperbolic tangent fluid over a stretching sheet in the presence of thermal radiation and magnetic field. Int. J. Mechan. Mater. Eng. (2018). DOI: https://doi.org/10.1186/s40712-018-0088-8. 8. Dey, D.: Viscoelastic fluid flow through an annulus with relaxation, retardation effects and external heat source/sink. Alexandria Eng. J. 57(2), 995–1001 (2018) 9. Dey, D., Boruah, A.J.: Dusty memory fluid through a horizontal channel with energy transfer. AIP Conf. Proce. 2061(1), 020006 (2019) 10. Wang, C.Y.: The three-dimensional flow due to a stretching sheet. Phy. Fluids 27, 1915–1917 (1984) 11. Mallikarjuna, H.B., Jayaprakash, M.C., Mishra, R.K.: Three-dimensional boundary layer flow and heat transfer of a fluid particle suspension over a stretching sheet embedded in a porous medium. Nonlin. Eng. 8, 734–743 (2019) 12. Mohaghegh, A.B., Rahimi, M.R.: Three–dimensional Stagnation–point flow and heat transfer of a dusty fluid toward a stretching sheet. J. Heat Transfer 138(11), 112001 (2016) 13. Ahsan, S., Khare, R.K., Paul, A.: Effect of volume fraction on non-Newtonian flow of dusty fluid between two oscillating parallel plate. Internat. J. Emerging Technol. Innovat. Eng. 1(7), 2394–6598 (2015) 14. Dey, D.: Non-Newtonian effects on hydromagnetic dusty stratified fluid fow through a porous medium with volume fraction. Proce. Nat. Acadamy Sci. India Sect. A: Phys. Sci. 86(1), 47–56 (2016)

Modelling of Two-Phase Fluid Flow Over …

25

15. Rath, B.K., Ganesh, V., Jagannadham, N., Dash, D.K.: Effect of longitudinal velocity of the particle of the dusty fluid with volume fraction in the incompressible fluid. Internat. J. Eng. Sci. Invent. 5(12), 15–18 (2016) 16. Dey, D.: Dusty Jeffrey fluid flow in a rotating system with volume fraction and hall effect: an analytical approach. Adv. Model. Anal. A, 70–75 (2018)

Boundary Layer Flow and Its Dual Solutions Over a Stretching Cylinder: Stability Analysis Debasish Dey, Rupjyoti Borah, and Bhagyashree Mahanta

Abstract Analysis of magnetohydrodynamics (MHD) boundary layer viscous fluid flow with variable thermal conductivity and mass transfer over a stretching cylinder has been done. The joule heating due to the weak magnetic field is encountered here and the thermal conductivity is taken to be a linear function of temperature. Numerical solutions are built up for momentum, energy and concentration equations using “MATLAB built-in bvp4c solver technique”. Impact of different flow parameters on the flow velocity, temperature and concentrations profiles have been discussed graphically. It is found that the dual solutions exist due to stretching cylinder and there is an unexpected increase of heat at the surface of the cylinder. Stability analysis of dual solutions is implemented to characterize the linearly stable and physically realizable solution. Keywords MHD · Variable thermal conductivity · Mass transfer · Dual solutions · Stretching cylinder · Stability analysis

1 Introduction Analysis of stability analysis aims at various fluids to study their flow performance and the development of turbulence due to these instabilities. As there is an evolution in technology, the boundary layer steady viscous flow with mass transfer over a stretching cylinder has drawn numerous scientists and engineers due to its extensive industrial applications in metal mining, extrusion, wire drawing, retreating of copper wire etc. The study on stability analysis has added substantial attention due D. Dey · R. Borah (B) · B. Mahanta Department of Mathematics, Dibrugarh University, Dibrugarh 786004, Assam, India e-mail: [email protected] D. Dey e-mail: [email protected] B. Mahanta e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_3

27

28

D. Dey et al.

to high demands in industries and it has been done by many researchers. Recently, the importance of dual solutions and their stability in scientific fields of Newtonian and non-Newtonian fluids over different stretching/shrinking surfaces are described in [1–6]. Based on numerous works done in the field of MHD, variable thermal conductivity etc., many researchers have extended their ideas to draw interesting conclusions. In [7], studied the nature of MHD boundary layer flow over a stretching cylinder with variable thermal conductivity. In [8], demonstrated the unsteady hydromagnetic flow of dusty non-Newtonian fluid through a porous medium, which is bounded by two parallel plates. Also, the visco-elastic fluid flow past an infinite flat surface with the influence of magnetic field by considering the Oldroyd fluid model is investigated in [9]. In [10], analysed the heat and mass transfers effect on the boundary layer flow of nanofluid over a linear stretching surface. Our present work is all motivated by the above-referenced work and its cosmic appliances in the industrial field. The execution on stability analysis of dual solutions of MHD flow near stagnation point over a stretching cylinder with effects of variable thermal conductivity together with mass transfer is presented in our work taking the assistance of MATLAB built-in bvp4c solver methodology. Further, the graphical demonstration is shown for several parameters on various profiles. Also, tabulations on the parameters of its first and second solutions are arranged accordingly.

2 Mathematical Formulation We consider steady, incompressible, viscous and two-dimensional boundary layer MHD flow over a stretching cylinder with the influence of heat and mass transfer, which is shown in Fig. 1. Let R be the radius of the cylinder that immerse in viscous fluid. Further, thermal conductivity k(T ) is assumed to vary linearly with temperature. Joule heating due to weak magnetic field B0 (no electric field is applied and the induced magnetic field is neglected) is also considered. The free stream velocity is assumed to be a linear function of x only i.e., Ue = bx where b ≥ 0. Again, stretching Fig. 1 Flow model and coordinate system

Boundary Layer Flow and Its Dual Solutions …

29

velocity Uw , wall temperature Tw and wall concentration Cw are assumed to vary linearly with the distance from the stagnation point i.e., UW (x) = ax, TW (x) = T∞ + x and Cw (x) = C∞ + x, where a is the constant such that its positive value resembles stretch at the surface, T∞ and C∞ are the free stream temperature and concentration, respectively, and x is the coordinate measured along with the cylinder. The governing equations are: ∂ ∂ (r u) + (r v) = 0 ∂x ∂r   2 ∂u σ B02 1 ∂u ∂u ∂ u u + + +v =υ (u − Ue ) ∂x ∂r ∂r 2 r ∂r ρ   ∂T 1 1 ∂ ∂T σ B02 2 ∂T +v = k(T )r + u u ∂x ∂r ρc p r ∂r ∂r ρc p  2  ∂C ∂ C ∂C 1 ∂C +v =D u + ∂x ∂r ∂r 2 r ∂r

(2.1)

(2.2)

(2.3)

(2.4)

The relevant boundary conditions are u = Uw = ax, v = 0, T = Tw , C = Cw at r = R u → Ue = bx, T → T∞ , C → C∞ as r → ∞

(2.5)

where, ρ, υ, c p , σ and D are the density, kinematic viscosity, specific heat, electric charge density and mass diffusivity respectively, u and v the velocity components along x and r directions respectively. We introduce the following similarity variables (following [7]):  η=

   T − T∞ a r 2 − R2 υa C − C∞ ,ψ = Rx f (η), θ (η) = , φ(η) = 2υ R 2 Tw − T∞ Cw − C∞ (2.6)

where, ψ is the stream function and u = r1 ∂ψ and v = − r1 ∂ψ . For liquid metal, the ∂r ∂x thermal conductivity k changes with temperature in linear manner in the range from 0ν F − 400ν F approximately [11]. Following [6], the variable thermal conductivity ∞ is the small can be written in the form of k(T ) = k∞ (1 + εθ ), where ε = kwk−k ∞ parameter. Substituting (2.6) into Eqs. (2.1)–(2.4), we get the following set of ordinary differential equations: (2 + 2K η) f  + f f  − ( f  )2 + 2K f  − M 2 ( f  − A) + A2 = 0

(2.7)

30

D. Dey et al.

(1 + εθ)(2 + 2K η)θ  + ε(2 + 2K η)θ 2 + (1 + εθ)K θ  + Pr( f θ  + M 2 Ec f 2 ) = 0

(2.8) (2 + 2K η)φ  + 2K φ  + Sc( f φ  − f  φ) = 0

(2.9)

and the boundary conditions become f (0) = 0, f  (0) = 1, θ(0) = 1, φ(0) = 1; f  (η) → A, θ(η) → 0, φ(η) → 0 : η → ∞

(2.10) where prime represents the differentiation with respect to η and η = 0 resembles the surface of the  cylinder. μc σ B2 Uw2 1 K = R 2υ , Pr = k∞p , M 2 = ρa0 , Ec = c p (Tw −T , Sc = Dυ and A = ab are the a ∞) curvature parameter, Prandtl number, Hartmann number, Eckert number, Schmidt number and velocity ratio parameter respectively. In this study, we have used the dimensionless quantities skin friction coefficient (C f ), local Nusselt number (Nux ) and local Sherwood number (Shx ) to determine the skin friction (resistance at the surface), rate of heat transfer and mass accumulation rate. These physical quantities can be defined as below: Cf =

 μ ∂u ∂r r =R ρUw2

, Nux = −

x

 ∂T

∂r r =R

k(Tw − T∞ )

and Shx = −

x

 ∂C

∂r r =R

D(Cw − C∞ )

.

Using the similarity transformations, we get 1 1 1 1 −1 −1 √ C f Rex2 = f  (0), √ Nux Rex 2 = −θ  (0), √ Shx Rex 2 = −φ  (0) 2 2 2

where Rex =

Uw x υ

(2.11)

is the local Reynolds number.

3 Stability Analysis To characterize which solutions will be stable and physically realizable, a stability analysis is executed. The stability analysis is very important to know the flow behaviour of different fluid. To test this nature, the unsteady governing equations , ∂ T & ∂C in (2.1)–(2.4). Where t of this present problem are considered by adding ∂u ∂t ∂t ∂t denotes the time. We introduce a new dimensionless time variable τ = at. Therefore, we use the following new similarity variables:  η=

   T − T∞ a r 2 − R2 υa ,ψ = Rx f (η, τ ), θ (η, τ ) = , 2υ R 2 Tw − T∞

Boundary Layer Flow and Its Dual Solutions …

φ(η, τ ) =

C − C∞ Cw − C∞

31

(3.1)

and hence using these similarity transformations in the time-dependent equations, we get the following non-linear ordinary differential equations:   ∂ 2 f (η, τ ) ∂ 2 f (η, τ ) ∂ 3 f (η, τ ) 2 ∂ f (η, τ ) + f (η, τ ) + 2K − M − A ∂η3 ∂η2 ∂η2 ∂η  2 2 ∂ f (η, τ ) ∂ f (η, τ ) − − A2 = 0 − (3.2) ∂η ∂η∂τ   ∂ 2 θ (η, τ ) ∂θ (η, τ ) 2 ∂θ (η, τ ) (1 + εθ )(2 + 2K η) + ε(2 + 2K η) + K (1 + εθ ) ∂η2 ∂η ∂η 2  ∂ f (η, τ ) ∂ f (η, τ ) ∂θ (η, τ ) − Pr θ (η, τ ) + M 2 Ec Pr + Pr f (η, τ ) ∂η ∂η ∂η ∂θ (η, τ ) − Pr =0 (3.3) ∂τ

(2 + 2K η)

∂ 2 φ(η, τ ) ∂φ(η, τ ) ∂φ(η, τ ) + 2K + Sc f (η, τ ) ∂η2 ∂η ∂η ∂φ(η, τ ) ∂ f (η, τ ) φ(η, τ ) − Sc =0 − Sc ∂η ∂τ

(2 + 2K η)

(3.4)

The relevant boundary conditions are: ∂ f (0, τ ) = 1, θ (0, τ ) = 1, ∂η ∂ f (η, τ ) → A, θ (η, τ ) → 0, φ(η, τ ) → 0 : η → ∞ φ(0, τ ) = 1; ∂η f (0, τ ) = 0,

(3.5)

Now, to solve Eqs. (3.2)–(3.4), we assume the following perturbations (separation of variables method) those are taken to check the stability of the steady flow solution f (η) = f 0 (η), θ (η) = θ0 (η) and φ(η) = φ0 (η) that satisfy the governing Eqs. (2.1)– (2.5). Following [12], we have taken as f (η, τ ) = f 0 (η) + e−ωτ F(η, τ ); θ (η, τ ) = θ0 (η) + e−ωτ G(η, τ ); φ(η, τ ) = φ0 (η) + e−ωτ H (η, τ ), where, ω is an unknown eigen-value parameter and F(η, τ ), G(η, τ ) and H (η, τ ) are small related to steady flow solutions f (η) = f 0 (η), θ (η) = θ0 (η) and φ(η) = φ0 (η) respectively. Substituting this portrayal into Eqs. (3.2)–(3.4) and the solutions of the steady governing equations are obtained by setting the time variable τ = 0. Hence the

32

D. Dey et al.

functions F = F0 (η), G = G 0 (η) and H = H0 (η) will signify the initial decay or growth of the solutions of the perturb equation. Thus, we have to solve the following set of eigen-value problems: (2 + 2K η)F0 + f 0 F0 + F0 f 0 + 2K F0 − M 2 F0 − 2 f 0 F0 + ωF0 = 0

(3.6)

(1 + εθ0 )(2 + 2K η)G 0 + ε(2 + 2K η)G 0 θ0 + K (1 + εθ0 )G 0 + K εG 0 θ0 + 2ε(2 + 2K η)θ0 G 0 + 2M 2 Ec Pr f 0 F0 + Pr f 0 G 0 + Pr F0 θ0 − Pr f 0 G 0 − Pr F0 θ0 + Pr ωG 0 = 0 (3.7) (2 + 2K η)H0 + 2K H0 + Sc f 0 H0 + ScF0 φ0 + ωScH0 − Sc f 0 H0 − ScH0 φ0 = 0 (3.8) the boundary conditions will reduce to F0 (0) = 0, F0 (0) = 0, G 0 (0) = 0, H0 (0) = 0; F0 (η → 0, G 0 (η) → 0, H0 (η) → 0 : η → ∞. (3.9) The value of ω can be determined by solving the set of linearized Eqs. (3.7)–(3.9). The stability of the steady flow solutions depends on the least eigen-value ω1 such that if the least eigen-value is found to be positive then an initial decay of disturbances occurs and the flow becomes stable and it is physically more realistic, whereas negative least eigen-value gives an initial growth of disturbances with time and the flow behave like unstable. Following [13], we assume that the boundary condition F0 (η) → 0 as η → ∞ transforms to a new boundary condition F0 (0) = 1.

4 Results and Discussion The numerical method “MATLAB built-in bvp4c solver technique” is used to solve the resulting ordinary differential equations. The outcome of this work highlights the influences of the curvature parameter K, Schmidt number Sc, magnetic parameter M and small temperature parameter ε. In this paper, we have considered the Prandtl number Pr = 1. In this investigation, we have found that dual solutions exist only a certain region of the surface due to stretching of the cylinder and unexpected increase of heat at the surface of the cylinder. One of them is a steady solution, which is denoted by the first solution and another one is unsteady solution, which is represented by the second solution. Also, the weak uniform magnetic field is applied, which reduces the flow speed over the cylinder and increases the heat transfer rate. The variations of the skin friction coefficient f  (0) with curvature parameter K for various values of magnetic parameter M are plotted, which is shown in Fig. 2. It

Boundary Layer Flow and Its Dual Solutions …

33

-0.5

First Solution Second Solution

M = 0.0, 0.4,0.8

-1

f ''(0)

-1.5 K c = 3.21 -2

K c =3.712

K c = 3.561

-2.5

-3

0

0.5

1

1.5

2

2.5

3

3.5

4

K

Fig. 2 Sketch of skin friction coefficient against K for various values of M when A = 0.1, Pr = 1, Ec = 1, Sc = 1 and ε = 0.2

is seen that the dual solutions exist up to a critical value of the curvature parameter K i.e., K = K c and there is no solution exists outside this region. It is also observed that the magnitude of the skin friction coefficient reduces with the increase of magnetic parameter, but the critical value of the curvature parameter enhances. The values of skin friction coefficient f  (0) are computed for different values of curvature parameter K, which is shown in Table 1. From this table, it is observed that both the solutions reduce with the curvature parameter K. The obtained results for steady flow case (first solutions) are in good agreement with the results of [7] and this is depicted by Table 1. The rate of mass accumulation boosts (during both steady and unsteady cases) with the curvature parameter K (see Table 2). Again, for fixed value of curvature parameter K, the Schmidt number Sc helps to raise the mass accumulation rate of the fluid flow. The influences of the flow parameters on velocity distributions f  (η) are depicted through Figs. 3 and 4. In Fig. 3, it is perceived that fluid flow experiences deceleration with curvature parameter K. In steady flow (no variation with time), fluid reaches its free-stream region more rapidly than unsteady flows (second solution). Application of magnetic field creates Lorentz force and it helps to retard the fluid and as a consequence more and more energy will remain stored. This physical phenomenon Table 1 Values of skin friction coefficient f  (0) for different values of curvature parameter K when M = 0.2, A = 0.1, Pr = 1.0, Sc = 0, Ec = 0.2 and ε = 0.2

Values of K

Results of [7] Present results First solution

First solution

Second solution

0.0

−0.68667

−0.6871

−1.1557

0.2

−0.72698

−0.7259

−1.2530

0.3

−0.74680

−0.7467

−1.3026

34 Table 2 Values of Sherwood number −φ  (0) for different values of curvature parameter K and Schmidt number Sc when M = 0.2, A = 0.1, Pr = 1.0, Ec = 0.2 and ε = 0.2

D. Dey et al. Values of K

Values of Sc

First solution

Second solution

0.2

1.0

0.7199

0.8007

0.8207

0.8942

0.3 0.0 0.2

0.8742

0.9433

0.0

0.2506

0.2506

0.5

0.5889

0.5936

1.0

0.8300

0.7982

Fig. 3 Velocity distribution f  (η) against η for Pr = 1, Sc = 1, M = 0.1, Ec = 0.1 A = 0.1 & ε = 0.2

Fig. 4 Velocity distribution f  (η) against η for Pr = 1, Sc = 1, K = 0.5, Ec = 0.1, A = 0.1 & ε = 0.2

Boundary Layer Flow and Its Dual Solutions …

35

Fig. 5 Temperature distribution θ(η) against η for ε = 0.2, Pr = 1.0, Ec = 0.1, A = 0.1, M = 0.1, and Sc = 1.0

is in perfect agreement with our result in time-independent case, but during timedependent cases, an opposite behaviour is noticed (see Fig. 4). Figure 5 is portrayed for temperature distribution under the influence of curvature parameter K. The first solution (steady case) reduces as increase of the curvature parameter K, whereas a reverse case is perceived in the second solution (unsteady case), but magnetic parameter helps to increase the temperature of the fluid flow, which is observed in Fig. 6. It is also concluded that the thickness of the momentum boundary layer of the first solution is greater than that of the second solution (unsteady solution) (see Fig. 6). The concentration distributions for various values of curvature parameter K and Schmidt number Sc are depicted in Figs. 7 and 8. From Fig. 7, it is noticed that both

Fig. 6 Temperature distribution θ(η) against η for ε = 0.2, Pr = 1.0, A = 0.1, K = 0.5, Ec = 0.5 and Sc = 1.0

36

D. Dey et al.

Fig. 7 Concentration distribution θ(η) against η for A = 0.1, Pr = 1, ε = 0.2, M = 0.1, Ec = 0.1 and Sc = 1.0

Fig. 8 Concentration distribution θ(η) for Pr = 1.0, A = 0.1, K = 0.5, ε = 0.2, M = 0.1 & Ec = 0.1

the solutions (first and second solutions) reduce as the enlarge of curvature parameter K. It is also observed that the concentration intensity of the second solution (unsteady case) is comparatively higher than that of the first solution (steady solution). Again, the Schmidt number helps to increase the concentration level of the solutions (see Fig. 8). In this analysis, all the figures satisfy the far-field boundary conditions asymptotically of the problem. All the figures display the existence of dual solutions up to a certain region only and support the variations of skin friction coefficient. From this

Boundary Layer Flow and Its Dual Solutions …

37

study, we have obtained two types of solutions one is linearly stable (first solution), which is physically achievable and the other one is physically not realizable.

5 Conclusion The following results are highlighted from this investigation: 1.

2. 3. 4. 5.

From the stability analysis, it is observed that the first solution is stable as there is an initial decay of disturbances on the flow and the flow becomes physically realizable, while the second solution is unstable as the disturbances will affect the flow at certain region and it is physically not acceptable. It is found that the dual solutions exist up to critical values of the curvature parameter K. The curvature parameter K assists to reduce the speed of the flow and temperature for steady case. The magnetic parameter M helps to increase the temperature of the fluid flow. The curvature parameter K and the Schmidt number Sc help to raise the concentration level of the fluid flow.

Future Scopes: There are several scopes to extend this work in multi-directions. Some of the possible scopes are listed below: a.

b.

c.

This work can be extended by considering different flow configurations such as shrinking surfaces, rotating spheres, etc. due to their important applications in engineering and industrial areas. In the modern time, the non-Newtonian fluids (there not exists a linear relationship between shear stress and rate of deformation) have greater applications in industrial processes and medical sciences in comparison to the Newtonian fluids. This flow configurations and the concept of stability analysis can be used in various non-Newtonian fluids. This present results may also be applied in various impulsion instruments for aircraft, missiles, satellites, etc.

References 1. Awaludin, I.S., Weidman, P.D., Ishak, A.: Stability analysis of stagnation-point flow over a stretching/shrinking sheet. AIP Adv. 6(4) (2016) 2. Kamal, F., Zaimi, K., Ishak, A., Pop, I.: Stability analysis on the stagnation-point flow and heat transfer over a permeable stretching/shrinking sheet with heat source effect. Int. J. Numer. Methods Heat Fluid Flow 28(11), 2650–2663 (2018) 3. Ahmed, A., Siddique, J.I., Sagheer, M.: Dual solutions in a boundary layer flow of a power law fluid over a moving permeable flat plate with thermal radiation, viscous dissipation and heat generation/absorption. Fluids 3(1) (2018)

38

D. Dey et al.

4. Chowdhury, M.M.K., Parveen, N.: Stability analysis of the laminar boundary layer flow. GANIT J. Bangladesh Math. Soc. 29, 23–34 (2009) 5. Najib, N., Bachok, N., Arifin, N.M.: Stability of dual solutions in boundary layer flow and heat transfer over an exponentially shrinking cylinder. Indian J. Sci. Technol. 9(48) (2017) 6. Abel, M.S., Datti, P.S., Mahesha, N.: Flow and heat transfer in a power-law fluid over a stretching sheet with variable thermal conductivity and non-uniform heat source. Int. J. Heat Mass Transf. 52(11–12), 2902–2913 (2009) 7. Jahan, S., Sakidin, H., Nazar, R.M.: MHD stagnation point flow over a stretching cylinder with variable thermal conductivity and joule heating. AIP Conf. Proc. 1787 (2016) 8. Dey, D.: Non-Newtonian effects on hydromagnetic dusty stratified fluid flow through a porous medium with volume fraction. Proc. Natl. Acad. Sci. India Sect. A—Phys. Sci. 86(1), 47–56 (2016) 9. Dey, D.: Hydromagnetic Oldroyd fluid flow past a flat surface with density and electrical conductivity stratification. Lat. Am. Appl. Res. 42(2), 41–45 (2017) 10. Narender, G., Govardhan, K., Sarma, G,S.: Heat and mass transfer of nanofluid over a linear stretching surface with viscous dissipation effect. J. Heat Mass Transf. Res. 6, 117–124 (2019) 11. Keys, W.M.: Convective heat and mass transfer. McGraw-Hill, New York (1966) 12. Merkin, J.H.: Mixed convection boundary layer flow on a vertical surface in a saturated porous medium. J. Eng. Math. 14(4), 301–313 (1980) 13. Harris, S.D., Ingham, D.B., Pop, I.: Mixed convection boundary-layer flow near the stagnation point on a vertical surface in a porous medium: Brinkman model with slip. Transp. Porous Media 77(2), 267–285 (2009)

Towards a Conceptual Modelling of Ontologies Chhiteesh Rai, Animesh Sivastava, Sanju Tiwari, and Kumar Abhishek

Abstract Ontologies are considered as a cornerstone for knowledge-based systems and semantic web for reusing and sharing the knowledge by explicit specification of shared conceptualizations. For representing and organizing knowledge, ontology needs a conceptual model to express the real-world objects. The conceptual model explores a series of entities, relationships and attributes to integrate and correlate the knowledge of domain-related data. It allows the layout of architectures for the usage of contents and the ontology provides an initial conceptual level for the knowledge organization. In this paper, the authors have presented a different conceptual model for the management and representation of specific domain knowledge. Keywords Conceptual modelling · Knowledge representation · Ontology · Semantic web

1 Introduction In the last few decades, several ontologies have been designed for the engineering domains [1] such as buildings, constructions, automations and many more. In this domain, ontologies are taken as engineering artefacts and follow a structured manner

C. Rai AKTU, Lucknow, India e-mail: [email protected] A. Sivastava Poornima University, Jaipur, India e-mail: [email protected] S. Tiwari (B) IEEE Member, Kurukshetra, India e-mail: [email protected] K. Abhishek NIT Patna, Patna, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_4

39

40

C. Rai et al.

to determine the correctness and efficiency of the ontology. Ontologies are considered as the backbone of knowledge-based systems as they provide knowledge reuse and sharing with the help of explicit specification of shared conceptualizations [2]. As ontology development has emerged one of the evolving research areas in the engineering domain and a huge range of ontologies have been developed since the past decade. Ontologies are engineering artefacts and their construction should accept a structured method with the relevant metrics to describe the correctness and sufficiency of the ontology. Different methods have been presented for ontology development to instruct the experts and knowledge engineers with the help of common efforts of the development life cycle, which includes requirements specification, conceptualization, knowledge acquisition, implementation, evaluation and maintenance [3–5]. The conceptualizations of ontologies represent an abstract vision of a particular domain through a group of interrelated concepts. It is an essential phase of ontology development to represent the whole ontology as a model. As all the concepts are determined and described, it is required to interlink each other through the relationships and logically presenting as a conceptual model. A poor presentation of conceptualize model causes incomplete expressivity and does not meet with the specified requirements. Generally, ontological conceptualization still suffers from a lack of precise structure and it is mostly presented in an unfixed approach without succeeding any systematic way [3]. A weak conceptual model will drive to a conditional ontology that does not match the requirements with respect to completeness and expressivity regardless of the formality level integrated with coding of the ontology. The ontological conceptualization can be aligned by analyzing domain knowledge sources and exploring, representing and organizing concepts in a platform-independent and neutral format. During the ontology development, it is recommended to generalize the domain conceptualization according to the specification of requirements. Domain experts are strictly following the conceptual model to organize taxonomy into concepts, relations, axioms and individuals. The primary aim of this paper is to present a conceptual model of the proposed ontology. The rest of the paper organized as follows: Sect. 2 presents related work. In Sect. 3, a conceptual model of ontologies has been presented. Section 4 presented the discussion and conclusion.

2 Related Work Ontologies are initially presented as conceptual models that can play a significant role in big data applications by precisely expressed domain knowledge using Web Ontology Language (OWL) [6]. Ontologies can be organized with data management system for large data applications to reconciling data heterogeneity, automate querying, and data processing [7]. Biomedical ontologies, such as the Foundational Model of Anatomy (FMA) and Gene Ontology (GO) [8, 9], have performed a significant role in data integration and annotation in biomedical data management. OWL ontologies are integrated with reasoning rules or defined rules and these rules are

Towards a Conceptual Modelling of Ontologies

41

based on OWL semantics [10] to interpret extra information [11]. Upper-level ontologies, such as the DOLCE and BFO [12, 13], are generally extended to design the domain-specific ontologies. There is extensive use of ontologies in the perspective of the semantic web, and some of them, such as DBpedia and SKOS [14, 15] are primary elements in the application development and Linked Open Data datasets. Ontological models can be evaluated based on the models’ semantic richness and preciseness. The ontologies quality is largely based on the accepted engineering methodology [16] where conceptualization is an expressive activity [17, 18]. Conceptual modelling can be described as “the activity of formally describing some aspects of the physical and social world around us for the purposes of understanding and communication”. Ontology engineering resolves the problems generated in ontology development and makes it useful to its lifetime [19]. The development of ontologies includes a set of exercises performed during ontology conceptualisation, ontology design, ontology implementation and ontology deployment phases [20]. Several methodologies are existing for developing ontologies by [18, 21, 22]. On the other hand, Uschold and King [17] presented four main steps for ontology development: analyze the purpose, ontology development (ontology capture, ontology coding, and organizing existing ontologies), evaluation, and deployment; whilst Pinto and Martins [23], proposed that the ontology engineering task is integrated with the five phases: “specification, conceptualisation, formalisation, implementation, and maintenance”. These all methodologies accept conceptualization as a primary task in ontology engineering. McCusker and his friends [24] describe the conceptual interoperability, provide requirements and use cases for it, and propose the Conceptual Model Ontology (CMO) to meet the defined requirements and use cases. They also introduced the implementation of CMO for swBIG (semantic web Biomedical Informatics Grid), semantic metadata, caBIG (a linked data proxy for cancer Biomedical Informatics Grid) models. Al-Debei and his colleagues [25] presented a conceptual modelling to develop a qualified ontology. They address the modelling paradigms for ontology models and suggested to give extra attention during defining entities especially their features such as properties and instances, which makes richer and accurate conceptualization. To accomplish this they have reengineered the bookstore ontology based on a defined conceptual model. Ameri and his colleagues [3] have proposed a methodology for incremental and systematic ontological conceptualization in the manufacturing domain. This methodology is designed to support the social method for developing an ontology. This research adopted a thesaurus-based approach that integrates the Simple Knowledge Organization System (SKOS) standard [26]. The proposed methodology for conceptualization has mainly three main phases such as development of thesaurus, evaluation of thesaurus and conversion of concept. Verdonck and his friends [27] presented a survey on the existing literature to analyze the research articles related to the ontology-driven conceptual modelling. This survey composed both a systematic review and a systematic mapping study.

42

C. Rai et al.

Erikson and his friends [28] propose an institutional ontology that can provide the design of information framework. This ontology has information of institutional theory and various case studies. They also proposed several guidelines to model institutional reality.

3 Conceptual Modelling of Ontology Ontological engineering needs the standardization and an explanation of a life cycle that moves from requirements generation to maintenance of the proposed domain, as well as techniques and methodologies that execute their development. There are four steps to design an ontology for a particular domain according to the existing literature [4, 5, 29, 30, 32]. Conceptualization is the first step of ontology development methodology. The IDEAL Methodology [31] introduces two concurrent exercises termed as analysis and synthesis in the conceptualization step. The analysis identifies the factual, strategical, tactual knowledge and the domain meta-knowledge. Both modelling exercises use a set of intermediate representations (IRs) to express knowledge. There are two major steps for conceptualizing any domain according to the existing literature. Step #1. To acquire the knowledge of a particular domain and to provide a requirement specification document. Step #2. To conceptualize the requirement specification in a set of intermediate representations (IRs). A conceptual model can be represented in different representational form. We have followed a common conceptualization model [1] by using different shapes for concepts and properties. The conceptual modelling representation is based on entityrelationship models, which frequently offer the well-off semantic representation. There are several possible ways to conceptualize the target world. Conceptualization inferred the hidden knowledge of the real world. Some advanced conceptual models follow the object-oriented rules by accepting the multi-instantiation principle. Object instances and relation instances carry a system-generated, different identity. They can be integrated into specialization and generalization frame using is-a hierarchy. Conceptual modelling provides semantically rich definition of structured datasets. Conceptualization of a domain can be interpreted and presented in a different way as in Figs. 1, 2 and Table 1. A hierarchical structure has been presented in a graphical conceptual model in Fig. 1. This model depicts clearly the structure of classes, subclasses, instances and their relationship with other classes and instances. It can be easily categorized the ontology components by the ontology developer to build any ontology with the help of conceptual model. In Fig. 1, it is described that Fruit and Color are the main classes under Thing class, has_Color describes the relation between Fruit and Color classes and all dotted objects are instances of Fruit and Color classes. This hierarchical structure can be semantically conceptualized in Fig. 2. Fruit (Banana, Apple, Grape).

Towards a Conceptual Modelling of Ontologies

43

Fig. 1 A graphical conceptual model

Red

Apple Banana

Fruit

has_Color

Color

Green

Grape

Yellow

Fig. 2 Semantics-based conceptual model

Table 1 Tabular conceptual model

Classes

Properties

Instances

Thing

–

–

Fruit

has_Color, has_Name, has_Weight

Banana, Apple, Grapes

Color

–

Red, Green, Yellow

Color (Red, Green, Blue). Entities of a domain can be conceptualized in a tabular structure as in Table 1. The conceptual model represents the entities in hierarchical, semantically and in the tabular structure. It is easy to interpret the hidden information of a domain. In this example, it is shown that Fruit is related to Color by has_Color property and Fruit has instances Apple, Banana, Grape so it is inferred that every instance of Fruit can have a color. As Color also have multiple instances so if one concept is related

44

C. Rai et al.

to another concept by property then all instances of that class also inherit all features of that concept. A Conceptual model easily explained it very well.

4 Conclusion Ontology development is a creative task and the automated role of experts with the help of existing ontologies. To design a conceptual model is the initial stage of the ontology development phase. It is useful for ontology engineers as well as users to provide a richer description of a particular domain. This paper presents a study of existing work and different ways to describe the conceptual modelling of ontologies. The presented work will be helpful for beginners to understand the conceptualization of a specific domain. Three ways are discussed in this paper to present the conceptual model. As future work, these different ways can be used to develop theme-based ontology. Acknowledgements This work was not funded by any grant.

References 1. Poveda-Villalón, M., Garcıa-Castro, R.: Extending the SAREF ontology for building devices and topology. In Proceedings of the 6th Linked Data in Architecture and Construction Workshop (LDAC 2018), Vol. CEUR-WS (Vol. 2159, pp. 16–23). (2018) 2. Gruber, T.R.: Toward principles for the design of ontologies used for knowledge sharing? Int. J. Hum Comput Stud. 43(5–6), 907–928 (1995) 3. Ameri, F., Kulvatunyou, B., Ivezic, N., Kaikhah, K.: Ontological Conceptualization Based on the Simple Knowledge Organization System (SKOS), 14(3), 11 (2014, May) 4. Mishra, S., Jain, S.: An intelligent knowledge treasure for military decision support. Internat. J. Web-Based Learn. Teach. Technol. (IJWLTT) 14(3), 55–75 (2019) 5. Mishra, S., Jain, S.: Ontologies as a semantic model in IoT. Internat. J. Comput. Applic., 1–11 (2018) 6. Hitzler, P., Krötzsch, M., Parsia, B., Patel-Schneider, P.F., Rudolph, S.: OWL 2 web ontology language primer. W3C Recommen. 27(1), 123 (2009) 7. Embley, D.W., Liddle, S.W.: Big data—conceptual modeling to the rescue. In International Conference on Conceptual Modeling (pp. 1–8). Springer, Berlin, Heidelberg (2013, November) 8. Rosse, C., Mejino, J.L., Jr.: A reference ontology for biomedical informatics: the foundational model of anatomy. J. Biomed. Inform. 36(6), 478–500 (2003) 9. Ashburner, M., Ball, C.A., Blake, J.A., Botstein, D., Butler, H., Cherry, J.M., Harris, M.A.: Gene ontology: tool for the unification of biology. Nat. Genet. 25(1), 25–29 (2000) 10. Motik, B., Patel-Schneider, P.F., Grau, B.C.: OWL 2 web ontology language: direct semantics. W3C Recommendation, 27 October 2009. (2009). World Wide Web Consortium. https://www. w3.org/TR/owl2-direct-semantics/. last visited March 12, 2012 11. Ferrucci, D., Brown, E., Chu-Carroll, J., Fan, J., Gondek, D., Kalyanpur, A. A., ... Schlaefer, N.: Building watson: an overview of the DeepQA project. AI Magaz. 31(3), 59–79 (2010) 12. Gangemi, A., Guarino, N., Masolo, C., Oltramari, A., Schneider, L.: Sweetening ontologies with DOLCE. In International Conference on Knowledge Engineering and Knowledge Management (pp. 166–181). Springer, Berlin, Heidelberg (2002, October)

Towards a Conceptual Modelling of Ontologies

45

13. B., Ceusters, W., Klagges, B., Köhler, J., Kumar, A., Lomax, J., ... & Rosse, C.: Relations in biomedical ontologies. Genome Biol. 6(5), R46 (2005) 14. Lehmann, J., Isele, R., Jakob, M., Jentzsch, A., Kontokostas, D., Mendes, P.N., ... Bizer, C.: DBpedia–a large-scale, multilingual knowledge base extracted from Wikipedia. Semant. Web 6(2), 167–195 (2015) 15. Miles, A., Bechhofer, S.: SKOS simple knowledge organization system reference. W3C recommendation, 18, W3C (2009) 16. De Nicola, A., Missikoff, M., Navigli, R.: A software engineering approach to ontology building. Inform. Syst. 34(2), 258–275 (2009) 17. Uschold, M., King, M.: Towards a methodology for building ontologies. Workshop on Basic Ontological Issues in Knowledge Sharing (1995) 18. Al-Debei, M.M., & Fitzgerald, G.: OntoEng: a design method for ontology engineering in information systems. ACM SIGPLAN International Conference on Object Oriented Programming, Systems, Languages and Applications, ODiSE Workshop, Orlando, Florida, 1–25 (2009) 19. Gomez-Pérez, A., Corcho, O., Fernández-López, M.: Ontological engineering: with examples from the areas of knowledge management, e-commerce and the semantic web. Fifth edn., Springer (2004) 20. Devedzic, V.: Understanding ontological engineering. Commun. ACM 45(4), 136–144 (2002) 21. Fernández-López, M., Gómez-Pérez, A., Sierra, J.P., Sierra, A.P.: Building a chemical ontology using methontology and the ontology design environment. IEEE Intell. Syst. 14(1), 37–46 (1999) 22. Gruninger, M., Fox, M.S.: Methodology for the design and evaluation of ontologies. In Proceedings of the Workshop on Basic Ontological Issues in Knowledge Sharing, 1–10 (1995) 23. Pinto, H.S., Martins, J.P.: Ontologies: how can they be built? Knowl. Inf. Syst. 6(4), 441–464 (2004) 24. McCusker, J.P., Luciano, J.S., McGuinness, D.L.: Towards an ontology for conceptual modeling. In ICBO (2011, July) 25. Al-Debei, M.M., Asswad, M.M.A., de Cesare, S., Lycett, M.: Conceptual modelling and the quality of ontologies: Endurantism vs. perdurantism (2012). arXiv preprint arXiv:1207.2619 26. W3c - World Wide Web Consortium: Skos Simple Knowledge Organization System Reference (2009). https://www.w3.org/TR/skos-reference/ 27. Verdonck, M., Gailly, F., de Cesare, S., Poels, G.: Ontology-driven conceptual modeling: a systematic literature mapping and review. Appl. Ontol. 10(3–4), 197–227 (2015) 28. Eriksson, O., Johannesson, P., Bergholtz, M.: Institutional ontology for conceptual modeling. J. Informat. Technol. 33(2), 105–123 (2018) 29. Tiwari, S.M., Jain, S., Abraham, A., Shandilya, S.: Secure semantic smart healthcare (S3HC). J. Web Eng. 17(8), 617–646 (2018) 30. Rahul, M., Kohli, N., Agarwal, R., Mishra, S.: Facial expression recognition using geometric features and modified hidden Markov model. Int. J. Grid Util. Comput. 10(5), 488–496 (2019) 31. Gomez-Perez, A., Fernndez-Lpez, M., deVicente, A.: Towards a method to conceptualize domain ontologies. In Proceedings of ECAI96 Workshopon Ontological Engineering, Budapest, pp. 4151 (1996) 32. Mishra, S., Sagban, R., Yakoob, A., Gandhi, N.: Swarm intelligence in anomaly detection systems: an overview. Internat. J. Comput. Applicat., 1–10 (2018)

Mathematical Modelling of Power Law Fluid Flow Through a Pipe and Its Rheology Debasish Dey and Bhagyashree Mahanta

Abstract This paper approaches to give an insight into a numerical investigation on rheology of steady MHD flow of power law fluid through an infinite pipe with variable viscosity. Two types of viscosity models viz., space-dependent viscosity and Reynolds viscosity were used. Conversion of governing equations into non-linear ordinary differential equations by applying some appropriate similarity transformation has been done. Subsequently, these converted similarity equations are numerically solved using MATLAB built-in bvp4c solver method to study the influence of various parameters such as power law parameter and pressure drop in velocity and temperature distribution. Likewise, graphs are marked for the above profiles. It has been observed that pressure drop helps in reducing the shear stress. Also, as power law parameter increases, velocity decreases. Keywords Power law fluid · Pipe · Space-dependent model · Reynolds model · bvp4c method

Nomenclature r, w μ A θ, ρ V0 B0 k Cp

Radial velocity and velocity at z direction Viscosity (variable) Pressure drop Temperature and density of the fluid respectively Constant Magnetic field to the fluid Thermal conductivity Specific heat

D. Dey · B. Mahanta (B) Department of Mathematics, Dibrugarh University, Dibrugarh 786004, Assam, India e-mail: [email protected] D. Dey e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_5

47

48

M n P1 K Ec Pr Re

D. Dey and B. Mahanta

Magnetic field parameter Flow behaviour index Pressure drop parameter Power law parameter Eckert number Prandtl number Reynolds’ number

1 Introduction Fluid flow past a pipe has been extensively considered as a benchmark in the study of fluid dynamics due to its remarkably high potential relevance of applications in industries. Along this study, the model of power law fluid is taken into consideration due to its importance in the study of shear stress in engineering. Various numerical and analytical methods were approached by researchers over the years to study on power law fluid flow past a pipe and the results were astonishingly exceptional. Ameur [1] has investigated the shear thinning flow through a cylindrical channel and studied the drop in pressure due to its sudden expansion. Consequences of transfer of heat on visco- elastic MHD flow on a wavy channel with slip velocity have been explored by Choudhury et al. [2]. The study of viscoelastic fluid has been examined by Choudhury et al. [3] in a porous medium along with the effect of heat and mass transfer. In the presence of magnetic field, Dey [4] carried out a study on convective flow (steady) in a micro-polar fluid. Dey and Deb [5] examined the magnetohydrodynamic flow of power law fluid along with energy dissipation using MATLAB built-in bvp4c method. Using finite difference method, Rytlsena and Shrager [9] carried out the problem of steady non-isothermal power law fluid past a pipe due to sudden expansion. Using the shooting method and taking viscosity (variable), a study on the effects of radiation on a non-Newtonian flow past a spongy sheet has been established brilliantly by Iqbal [6]. An analytical method has been implemented by Khan and Yovanovich [7] to study the nature of fluid flow and the effects of heat transfer in power law fluid past a cylinder. Nazeer et al. [8] have explored notably the effects of space-dependent viscosity with heat transfer inside a pipe using the perturbation method. By using second-law analysis, Saleem [10] has established the result of minimal entropy generation in asymmetric channel. Further, the effects of Power law fluid have been studied by Yao et al. [11] as the fluid flows past a sheet undergoing suction or blowing. Yurusoy and Guler [12] have stunningly presented the steady flow between concentric cylinders with heat transfer using the perturbation technique by taking Teo viscosity models. The effects of heat transfer of a power law fluid along various cross-sections of a pipe have been presented by Zhang and Xu [13]. By studying and reviewing many previous papers, the study on power law fluid flow through a pipe has been explored by very few. As such, influenced by the

Mathematical Modelling of Power Law Fluid Flow Through a Pipe and Its Rheology

49

above great findings, our present works revolves around the study of MHD power law fluid through a pipe by taking space-dependent and Reynolds model (depends on temperature exponentially) using MATLAB built-in bvp4c solver method to find the effects of various parameters graphically as a function of pipe radius. Problem formulation for the analysis of the same has been portrayed in the next section. In Sect. 3, the methodology used for the problem is widely outlined. Results obtained from the above works are noted and shown graphically in Sect. 4, also, discussions are presented thoroughly in it. Finally, in Sect. 5, conclusions of the work are highlighted in a very concise way for the easy understanding, followed by the references.

2 Mathematical Formulation of the Flow Analysis We consider a steady flow with variable viscosity for the flow analysis. The constitutive equations are governed by power law fluid model. Let w be the velocity component along z direction. The flow is guided by Lorentz force and pressure gradient A. Heat is generated into the system through energy dissipation. The governing equations of fluid motion are: du =0 dr         1 dw d dw n dw n 2 ρV0 + = −σ B0 w − A + μ μ dr dr dr r dr   dθ dw n+1 d 2θ ρC p u =μ +k 2 dr dr dr

(2.1) (2.2)

(2.3)

The relevant boundary conditions are: w¯ = 0, θ¯ = θ1 at r¯ = R¯ ¯ d w¯ = 0, ddθr¯ = 0 at r¯ = 0 d r¯

 (2.4)

Following are the introduced similarity transformation: r=

r w μ A R2 θ − θ0 ,w = ,μ = ,A= ,θ = R V0 μ0 μ0 V0 θ1 − θ0

Using the above data in Eqs. (2.1)–(2.3), we have the set of dimensionless equations:      dw ∂ 2 w dw n−1 dμ μ dw n + nK 2 = −Mw − P1 + K + dr dr r dr ∂r dr

(2.5)

50

D. Dey and B. Mahanta

   2  K E c μ dw n+1 d θ dθ 1 = + dr Pr dr Pr Re dr 2 σ B2 R

μ V n−2

0 where, M = V00ρ , P1 = ρAμ , K = 0ρ R0n , Pr = RV0 The appropriate boundary conditions are: w = 0, θ = 1 at r = 1. dw = 0, dθ = 0 at r = 0. dr dr

μ0 C p , k

Ec =

(2.6)

V02 μ0 , Re k(θ1 −θ0 )

=

RV0 . υ0

3 Method of Solution We adopt MATLAB built-in bvp4c solver technique in this problem of the form  y = w(x, y, c) with the applicable boundary conditions θ (y(a), y(b), c) where a ≤ x ≤ b. We have the set of non-linear ordinary differential equations as:

θ −

K E c μ n+1 w Pr

−



  w n − n K w  w n−1 = 0

dμ + μr dr 1 θ  = 0 Pr Re

w  + Mw + P1 − K

(3.1)

The related boundary conditions are: w = 0, θ = 1 at r = 1. w = 0, θ  = 0 at r = 0. The set of Eqs. (3.1) is further reduced to first-order differential equations as follows: Let w = y1 , w  = y2 y1 = w  = y2 y2

  1 2−n M y1 1−n dμ μ 1−n y y + y2 =w = + + P1y2 − K nK 2 nK 2 dr r 

θ = y3 , θ  = y4 y3 = θ  = y4 y4 = θ  = Pr Rey4 − ReK E c μy4n+1 The formulated boundary conditions are: y1 (1) = 0, y3 (1) = 1 y2 (0) = 0, y4 (0) = 0

Mathematical Modelling of Power Law Fluid Flow Through a Pipe and Its Rheology

51

4 Results and Discussion An investigation has been done graphically using a numerical method known as MATLAB built-in bvp4c solver method to solve the final obtained ordinary differential equations. This article deals to study on the rheology of MHD steady flow of power law fluid through an infinite pipe taking the viscosity to be variable. This work emphasises on the impacts of different parameters such as power law parameter and pressure drop on velocity and temperature distribution. In the present study, we have considered M = 1, Pr = 1, Re = 1, E c = 0.1. Further, n < 1 and n > 1 signify shear thinning and shear thickening cases. Also, velocity gradient and shear stresses are calculated for different values of the parameters and are shown in Tables 1, 2, 3, 4, 5 and 6. We have taken two cases of viscosity models as discussed below: Table 1 Velocity gradient for various values of P1 in case of space-dependent viscosity

Table 2 Shear stress for various values of P1 in case of space-dependent viscosity

Table 3 Velocity gradient for various values of K in case of space-dependent viscosity

Table 4 Shear stress for various values of K in case of space-dependent viscosity

w n = 0.5

n = 1.5

P1 = −0.5

8.4353e-26

7.5666e-05

P1 = −1.5

5.6050e-26

7.0936e-06

P1 = −2.5

1.3708e-26

1.3578e-06

w n n = 0.5

n = 1.5

P1 = −0.5

3.5314e-13

2.4669e-36

P1 = −1.5

3.2310e-13

1.8467e-36

P1 = −2.5

2.2706e-13

1.6304e-36

w n = 0.5

n = 1.5

K = 0.1

9.1717e-26

3.3884e-04

K = 0.2

6.8788e-26

3.1766e-05

K = 0.3

3.3729e-26

3.4744e-06

n = 0.5

n = 1.5

K = 0.1

4.2041e-13

6.2372e-06

K = 0.2

3.6409e-13

1.7904e-07

K = 0.3

3.2623e-13

6.4763e-09

w n

52

D. Dey and B. Mahanta

Table 5 Velocity gradient for various values of E c in case of space-dependent viscosity

Table 6 Shear stress for various values of E c in case of space-dependent viscosity

w n = 0.5

n = 1.5

E c = 0.1

9.1717e-26

3.3884e-04

E c = 0.2

6.8788e-26

1.9862e-04

E c = 0.3

1.6694e-26

9.9796e-05

w n n = 0.5

n = 1.5

E c = 0.1

4.2041e-13

6.2372e-06

E c = 0.2

3.6409e-13

2.7992e-06

E c = 0.3

3.1914e-13

9.9694e-07

Case 1: Space-dependent viscosity, Here we take, μ = r . Figure 1 represents the velocity profile for various values K. As the power law parameter (K) varies directly to viscosity, so an increasing behaviour in K raises the viscosity of the fluid and as such fluid motion is decelerated owing to the decrease in the velocity. Also, an increment of boundary layer thickness is clear from the figure when n < 1 (Fig. 1a). A similar deviation is perceived in the variation of velocity profile for values of n > 1 (Fig. 1b). Hence, the physical interpretation of the power law parameter is that, it has the tendency to decrease the fluid velocity causing the boundary layer to increase. Under a steady flow condition, pressure drop varies directly to flow rate. A gradual fall in pressure is noticed as viscosity decreases. Also, if the temperature is lowered,

(a)

(b)

Fig. 1 a and b Velocity distribution against radius (r) when n < 1 and n > 1 respectively for various values of K

Mathematical Modelling of Power Law Fluid Flow Through a Pipe and Its Rheology

53

a shrink in viscosity is noted. As a result, pressure drop degrades the temperature when n < 1 as seen in Fig. 2a. An identical pattern is observed in Fig. 2b when n > 1. Case 2: Reynolds’ Model: Here μ = eαθ , where α is a constant. In Reynolds model, i.e., Fig. 3a, b, acceleration of the fluid i.e., increase in velocity is noted when the value of power law parameter (K) is increased gradually in both the cases of power law index (n). Physically, in this model, the power law parameter tends to decrease the boundary layer owing to the acceleration of the fluid. Also, Fig. 4a, b signifies the effect of pressure drop (P1) on the temperature distribution. As P1 becomes negative, the thermal boundary layer tends to shrink and as a result of which rise in temperature is noted at the centre of the pipe. Velocity gradient across the pipe has been portrayed for various values of pressure drop, power law parameter and Eckert number by Tables 1, 3 and 5, respectively.

(a)

(b)

Fig. 2 a and b Temperature distribution against radius (r) when n < 1 and n > 1 respectively for various values of P1

(a)

(b)

Fig. 3 a and b Velocity distribution against radius (r) when n < 1 and n > 1 respectively

54

D. Dey and B. Mahanta

(a)

(b)

Fig. 4 a and b Temperature distribution against radius (r) when n < 1 and n > 1 respectively

Velocity gradient or rate of deformation plays an important role in fluid flow and in drag formation. Zero velocity gradient indicates that flow is undisturbed during fluid flow and its higher values portray high variation with flow displacement variables. Physically, it also signifies the deformation rate due to the application of force per unit area. The rate of deformation across the pipe gradually decreases for shear thinning and shear thickening fluid for various values of the above parameters. Thus the application of stress reduces the deformation rates. Research on fluid dynamics is incomplete without the analysis of shear stress or viscous drag. This shear stress is responsible for the damage at the surface due to the presence of friction between fluid layers. Further, research has been done to check the parameters which help in reducing the shear stress. In this paper, we have seen that shear stress steadily decreases as there is a rise in the P1, K and E c as seen in Tables 2, 4 and 5 respectively. Pressure drop helps to reduce the shear stress at the surface. Moreover, the velocity forces gradually fall at the surface.

5 Conclusion An investigation on steady MHD flow of power law fluid on space-dependent viscosity model and Reynolds model through an infinite pipe has been widely carried out numerically. Few points based on our study are highlighted below: • In space-dependent viscosity model, a drop in velocity can be witnessed as the power law parameter increases. In a similar context, a drop in temperature is noted as there is degradation in pressure. • In Reynolds model, acceleration of the fluid is noted as power law parameter is increased. Likewise, a rise in pressure drop leads to the rise in temperature.

Mathematical Modelling of Power Law Fluid Flow Through a Pipe and Its Rheology

55

Future Scope: Power law fluid flow is used mainly on industrial purposes. So, collaborative research may be done with the industrial engineering department.

References 1. Ameur, H.: Pressure drop and vortex size of power law fluid flow in branching channels with sudden expansion. J. Appl. Fluid Mech. 11(6), 1739–1749 (2018). https://doi.org/10.29252/ jafm.11.06.28831 2. Choudhury, R., Deb, H.R., Dey, D.: Effects of MHD flow and heat transfer of a visco- elastic fluid in a vertical channel with slip velocity. Int. J. Appl. Eng. Res. 6(3), 331–343 (2011) 3. Choudhury, R., Dhar, P., Dey, D.: Visco- elastic MHD flow through a porous medium bounded by horizontal parallel plates moving in opposite direction in presence of heat and mass transfer. Int. J. Comput. Appl. 97(16), 6–12 (2014). https://doi.org/10.5120/17089-7622 4. Dey, D.: Mixed convective MHD Micro- polar flow in a porous medium with radiation absorption. Int. J. Math., Eng. Manage. Sci. 4(2), 381–399 (2019). https://doi.org/10.33889/IJMEMS. 2019.4.2-031 5. Dey, D., Deb, H.R.: Hydromagnetic flow of power law fluid in a porous medium with energy dissipation: a numerical approach. Italian J. Eng. Sci.: TechnicaItaliana 61(1), 130–134 (2018). https://doi.org/10.18280/ijes.620211 6. Iqbal, Z.: Effect of time dependent viscosity and radiation efficacy on a non- Newtonian fluid flow. Heliyon 5(2), e01203. https://doi.org/10.1016/j.heliyon.2019.e012013 7. Khan, W., Yovanivich, M.M.: Fluid flow and heat transfer in power law fluids across circular cylinders: analytical study. https://doi.org/10.1115/1.2241747 8. Nazeer, M., Ahmed, F., Saleem, A., Saeed, M., Naveed, S., Shaheer, M.: Effects of constant and space dependent viscosity on Eyring- Powell Fluid in a pipe: comparison of the perturbation and explicit finite difference method. A J. Phys. Sci. 7(11), 961–969. https://doi.org/10.1515/ zna-2019-0095 9. Rytlsena, K., Shrager, G.: Non- isothermal flow of power law fluid in a pipe with sudden expansion. J. Phys. Conf. Series 1128(1), 012027. https://doi.org/10.1088/1742-6596/1128/1/ 012027 10. Saleem, N.: Entropy production in peristaltic flow of a space dependent viscosity fluid in asymmetric channel. Thermal Sci. 22(00), 164–164 (2017). https://doi.org/10.2298/TSCI16 1020164S 11. Yao, C., Li, B., Lu, J.: Power law fluids over a viscous sheet with mass suction/ flowing: multiple solutions. AIP Adv. 9, 115121 (2019). https://doi.org/10.1063/1.5129862 12. Yurusay, M., Guler, O.F.: Perturbation solution for MHD flow of a non- Newtonian fluid between concentric cylinders. Int. J. Appl. Mech. Eng. 21(1), 199–211. https://doi.org/10. 2478/ijame-2019-0013 13. Zhang, H., Xu, T.: Study on heat transfer on non- Newtonian power law fluid in pipes with different cross sections. Procedia Eng. 205, 3381–3388 (2017). https://doi.org/10.1016/j.pro eng.2017.09.845

NeurolncRNA: A Database of LncRNAs Associated with Human Neurodegenerative Diseases Aniruddha Biswas, Aishee De, Kumaresh Singha, and Angshuman Bagchi

Abstract Long non-coding RNAs (lncRNAs) correspond to a large class of functional RNA molecules that play important roles in the onset of neurodegenerative diseases in humans. A systematic collection and summary of such deregulated lncRNAs associated with human neurodegenerative diseases are essential for a proper understanding of the mechanisms of the disease onset. In this work, we built a database to store various sequence-based features of such lncRNAs having direct associations with the onset of neurodegenerative diseases. The database will help in understanding the characteristics of such lncRNAs. The database contains data related to around 100 lncRNAs linked to various neurodegenerative diseases. To this end, different data were collected to integrate the diverse interactions of the functional elements present in lncRNAs associated with disease onset. In this work, each association includes lncRNA and name of the disease, GC content, genomic context, location, the original reference and additional annotation information. The database provides a user-friendly interface to conveniently browse data. With the rapidly increasing interest in lncRNAs, it will significantly improve our understanding of lncRNA deregulation in neurodegenerative diseases and has the potential to be a timely and valuable resource. Keywords lncRNA · Neurodegenerative diseases · Disease association · lncRNA deregulation

A. Biswas · A. De · K. Singha JIS College of Engineering, Kalyani, India e-mail: [email protected] A. Bagchi (B) Department of Biochemistry and Biophysics, University of Kalyani, Kalyani, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_6

57

58

A. Biswas et al.

1 Introduction A non-coding RNA (ncRNA) is an RNA molecule with immense potentials to be involved in different disease pathways. However, as opposed to their nomenclature, some of the ncRNAs are translated to yield small polypeptides (having less than 100 amino acids). These biologically relevant hidden peptides are required for various cellular processes. ncRNAs are categorized into two main types according to the length of their nucleotides: small ncRNAs and long non-coding RNAs (lncRNAs). lncRNAs are defined as transcripts greater than 200 nucleotides with no apparent open reading frames (ORFs) (Matsumoto and Nakayama 2018 [1]). They do not have protein-coding potentials. The human genome produces more than 10,000 lncRNAs, and the number of newly discovered lncRNAs is increasing rapidly. They are expressed in differentiated tissues pertaining to neurodegenerative diseases (Johnson et al. 2012 [2]). lncRNAs play an important role in the prognosis and diagnosis of human neurodegenerative diseases (Zhang et al. 2019 [3]). When compared to small ncRNAs, lncRNAs are less evolutionarily conserved at the sequence level. They have been classified into five biotypes according to their closeness to protein-coding genes: sense, antisense, intergenic, intronic and bidirectional (Ponting et al. 2009 [4]; Gibb et al. 2011 [5]). lncRNAs act as important molecules in regulating human neurodegenerative diseases. They interact with other macromolecules like RNA, DNA and protein resulting in many important phenotypes of cancer. The important functions of lncRNAs include epigenetic regulation of gene expression, regulation of chromatin structure to name a few. Human nucleus contains the maximum number of lncRNAs (Derrien et al. 2012 [2]). Many lncRNAs have been identified, which play critical roles in the development and progression of neurodegenerative diseases in humans (Gibb et al. 2011 [5]). This can lead to the identification of therapeutic targets of human neurodegenerative diseases. The majority of the works pertaining to lncRNAs are directed towards their genetic aspects. There are also different databases available, which store lncRNA data (Fritah et al. 2014 [6]). In the present scenario, an attempt was made to collect the various sequence-based features of lncRNAs, associated with the onset of neurodegenerative diseases [7], along with their cellular location preferences and genomic contexts. There are numerous databases of lncRNAs and all of them store different information about the lncRNAs and their other characteristics. However, there is practically no specific database available that would amalgamate the features of lncRNAs and how the features are linked to the various types of neurodegenerative diseases. We, therefore, tried to compile such data and develop a database so that the user can get first-hand information about the lncRNAs and their associations with different types of neurodegenerative diseases. We believe our newly developed database might aid in the purpose of analyzing the involvements of lncRNAs in different types of neurodegenerative diseases.

NeurolncRNA: A Database of LncRNAs …

59

2 Materials and Methods 2.1 Data Extraction from Literature To provide a comprehensive collection of experimentally validated interactions with lncRNAs, the available publications were curated manually to collect the data. The keywords ‘LncRNA’, ‘lncRNA’, ‘long non coding RNA’, ‘Long non-coding RNA’, and ‘Long ncRNA’ were used to search the PubMed [8] database up to August 1, 2019. The species mentioned in publications were limited to human and the disease as ‘Neurodegenerative Disease’, ‘Alzheimer’, ‘Parkinson’,’Huntington’,’stroke’, ‘plexiform neurofibroma’, ‘neurofibromatosis’,’ intracranial aneurysm’, ‘spinocerebellar ataxia’, ‘Prader-Willi syndrome’, ‘angelman syndrome’, ‘fragile X syndrome’, ‘pituitary adenoma’, ‘meningioma’, ‘Spinal muscular atrophy’, ‘Prion disease’, ‘Motor neurone diseasese’, ‘Post-Synaptic Signaling’, ‘Oxidative Stress’. So far, nearly 7000 research articles were retrieved using the aforementioned keywords.

2.2 Retrieval of Sequence Information of the lncRNAs The sequences of the different lncRNAs were retrieved from different databases, viz., ‘deepBase’, ‘LNCipedia’, ‘lncRNAdb’, ‘LncRNAWiki’, ‘LncBook’, ‘MONOCLdb’, ‘NONCODE’, ‘lncRNome’, ‘NRED’, ‘C-It-Loci’, ‘MiTranscriptome’, ‘slncky Evolution Browser’ [10, 11, 12]. There were around 1000 sequences in lncRNAdb and approximately 0.1 million many sequences in lncBook. The retrieved sequences were analyzed to extract different sequence-based features, which mainly focus on neurodegenerative diseases.

2.3 Extraction of Sequence-Based Features of the Collected lncRNAs The collected nucleotide sequences of the lncRNAs were analyzed by an in-house pipeline to extract the different features pertaining to the sequences. The sequencebased features included the percentages of the GC contents of the lncRNAs, and their genomic contexts.

2.4 Methodology of Development The website has been created using HTML, Java Script. The relational database was developed using MySQL as backend. The web interfaces have been developed to

60

A. Biswas et al.

execute the SQL queries with an abstraction. It contains a user-friendly search field to extract information dynamically. This search can be made on any field and fetches information synchronously in a tabular form. Each column in the field can be sorted alphabetically with a single click. The response time is very quick. The URL for the lncRNA database for neurodegenerative diseases is https://lncrna2db.com.

3 Results and Discussion 3.1 Categorization of lncRNAs and their Literature Details The details of the lncRNAs were collected from research publications. The different lncRNAs were listed and categorized as per the different types of neurodegenerative diseases. The Gene Symbols of lncRNAs and the PubMed ids of the corresponding literatures referring to the lncRNAs were also collected. This way a systematic classification and characterization of the different lncRNAs involved in neurodegenerative disease were done.

3.2 Different Sequence-Based Features of lncRNAs We used an in-house pipeline to calculate the different sequence-based features of lncRNAs. Among the sequence-based features, we identified the %GC content, the length of the lncRNAs, and the types of the lncRNAs. The types would represent their appearances from the chromosome; it would reflect whether the lncRNAs are obtained from the intergenic location or the intronic locations, etc. The following parameters are identified and displayed on the website against each lncRNA responsible for various neurodegenerative diseases.

lncRNA

Gene symbol

GC%

Length

Chromosome

Genomic context

PMID

Binding partner

Sequence

Disease name

NeurolncRNA: A Database of LncRNAs … 61

62

A. Biswas et al.

The lengths of the lncRNAs were found to be varying from 200 nucleotides (the minimum as obtained in case of BCYRN1) to 10,194 nucleotides (the maximum as observed in case of MIAT, lncRNA for neurodegenerative disease Alzheimer’s). In the neurodegenerative disease’s type, viz., ‘Alzheimer’, ‘Parkinson’, ‘Huntington’, ‘stroke’, ‘plexiform neurofibroma’, ‘neurofibromatosis’, ‘intracranial aneurysm’, ‘spinocerebellar ataxia’, ‘Prader-Willi syndrome’, ‘angelman syndrome’, ‘fragile X syndrome’, ‘pituitary adenoma’, ‘meningioma’ the minimum lengths of the lncRNAs were found to be starting from over 200 nucleotides. The maximum lengths of the lncRNAs were found to be highly variable ranging from 2000 to 10000 bp.

4 Conclusion and Future Scope The dysregulation of lncRNAs plays a critical role in the onset of human neurodegenerative diseases [9]. They have become novel potential molecules for neurodegenerative disease prognosis, diagnosis and treatment. These lncRNAs associated with neurodegenerative diseases have been identified by experiments, which have been conducted over the past couple of years. Therefore, this database will provide researchers with a vital resource for research in neurodegenerative diseases. The database provides a comprehensive resource on lncRNA deregulation in various human neurodegenerative diseases. As next-generation sequencing technology would become cheaper and accessible, more lncRNAs are expected to be discovered. These newly annotated lncRNA sequences will be incorporated into the database. In the future, it is planned to implement structure-based classification information of the lncRNAs. The database has the potential to become a timely resource for lncRNA transcript information and annotation. This database will become a hot cake for researchers whose focus is on lncRNA with neurodegenerative diseases. Website is expected to be functional as a mobile application in near future. More records to be updated as per availability of new lncRNA sequence found in the vast biological database sites. Acknowledgements Authors are thankful for the generous help of Bioinformatics Infrastructural Facility (BIF) (BT/BI/25/001/2006), University of Kalyani, UGC-SAP-DRS-II, DST-FIST-II for support. Conflict of Interest None.

References 1. Matsumoto, A., Nakayama, K.I.: Hidden peptides encoded by putative non coding RNAs. Cell Struct. Funct. 43(1), 75–83 (2018) 2. Johnson, R., Derrien, T., Bussotti, G., Tanzer, A., Djebali, S., Tilgner, H., Guernec, G., Martin, D., Merkel, A., Knowles, D.G., et al.: The GENCODE v7 catalog of human long noncoding

NeurolncRNA: A Database of LncRNAs …

3. 4. 5. 6. 7.

8. 9.

10. 11. 12.

63

RNAs: analysis of their gene structure, evolution, and expression. Genome Res. 22, 1775–1789 (2012) Zhang, T., Hu, H., Yan, G., Wu, T., et al.: Long non-coding RNA and breast cancer. Technol. Canc. Res. Treat. 18, 1–10 (2019) Ponting, C.P., Oliver, P.L., Reik, W.: Evolution and functions of longnoncoding RNAs. Cell 136, 629–641 (2009) Gibb, E.A., Brown, C.J., Lam, W.L.: The functional role of long noncodingRNA in human carcinomas. Mol Cancer 10, 38 (2011) Fritah, S., Niclou, S.P., Azuaje, F.: Databases for lncRNAs: a comparative evaluationof emerging tool. RNA 20, 1655–1665 (2014) Rashad Hussain 1,*, Hira Zubair 2, Sarah Pursell 1 and Muhammad Shahab 2.: Neurodegenerative diseases: regenerative mechanisms and novel therapeutic approaches. Brain Sci. 8(9), 177 (2018). https://doi.org/https://doi.org/10.3390/brainsci8090177 https://www.ncbi.nlm.nih.gov/pubmed/ Dong Youa, Hong Youb: Repression of long non-coding RNA MEG3 restores nerve growth and alleviates neurological impairment after cerebral ischemia-reperfusion injury in a rat model (Jan, 2019). https://doi.org/https://doi.org/10.1016/j.biopha.2018.12.067 https://bigd.big.ac.cn/lncbook/index https://www.gold-lab.org/clc https://www.gold-lab.org/

Affective State Analysis Through Visual and Thermal Image Sequences Satyajit Nayak, Vivek Sharma, Sujit Kumar Panda, and Satarupa Uttarkabat

Abstract This paper presents a contactless system based on twin channels of thermal and visual image sequences to register the affective states of an individual during Human–Computer Interaction (HCI). The negative affective states such as stress, anxiety, and depression in students have raised significant concerns. The first phase obtains the dominant emotional state by an ensemble of cues from visual and thermal facial images using a newly proposed cascaded Convolutional Neural Network (CNN) model named as EmoScale. The second phase clusters a sequence of the obtained emotional states using a trained Hidden Markov model (HMM) as one of the three affective states anxiety, depression, and stress. We perform fivefold crossvalidation of EmoScale on our self-prepared dataset. The performance of the second phase is compared with a standard Depression Anxiety Stress Scale (DASS) and the results are found to be promising. Keywords Emotion Recognition · Affective states · SCNN · Facial Expression · HCI

S. Nayak IIT, Kharagpur, India S. Nayak (B) · V. Sharma · S. K. Panda Gandhi Institute For Technology, BPUT, Odisha, India e-mail: [email protected] V. Sharma e-mail: [email protected] S. K. Panda e-mail: [email protected] S. Uttarkabat NIT, Rourkela, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_7

65

66

S. Nayak et al.

1 Introduction The introduction of affective factors to HCI has gained much popularity in affective computing [1]. Attempts are made to develop smart health-aware artificial intelligent HCI that can perceive, understand, and regulate affective health of an individual [2]. Specifically, affective state recognition is the most critical phase in affective cycle and has been the primary focus of HCI researchers. The DASS is a standard subjective means to quantify the amount of three negative affective states, viz., depression, anxiety, and stress. Stress: It may be defined as a feeling of affective tension [3]. It can manifest from any event or thought that presents frustration, anger, or nervousness. Anxiety: It is a mental condition that causes increased alertness, fear, etc. [3]. It also manifests as physiological changes, such as increased pulse rate, rapid respiration rate, sweating, and fatigue. Depression: It is a mental condition that imbibes feelings of negativity, sadness, and loss of attentiveness in daily activities [3]. We find from these definitions that most of these negative affective states are closely related to the emotional states of an individual such as anger, disgust, fear, and sadness. For example, stress primarily manifests anger and disgust, anxiety manifests fear, while depression manifests sadness [4]. We selected the visual and thermal images of the face, inspired from the work of Basu et al. [5]. Facial expressions are the principal attributes through which an individual can publicly interact. In recent years, automatic recognition of facial expressions from images and videos has been an important domain of research. Facial expression-based approach has become popular due to the availability of relevant training and testing data, their universal properties, and their use in practical scenarios [6, 7]. However, these methods cannot distinguish between real and posed expressions. This is a limitation of using the visual cues exclusively, which can reduced by the use of facial thermal cues. Facial surface temperature distribution provides reliable information about the various emotional states of an individual. This is because every emotion manifests a change in the facial surface temperature due to the blood flow on the peripheral arteries. Khan et al. [8] have proposed a classification of pretended and evoked facial expressions using thermal and visual images. They have conducted multivariate tests and linear discriminate analysis for the purpose.

2 Design of the Experiment We conducted a pilot study with ten subjects to adjust the experimental conditions such as ambient illumination, temperature and orientation of the screen, etc. The emotion eliciting video stimuli are also standardized during the pilot study. The entire experiment was conducted in an isolated room of area 5 × 5 m with controlled illumination to avoid external influences. A customized wooden box has been designed to conceal the thermal and visual cameras. A 24-inch touchscreen monitor was affixed on the front end of the wooden box, and the CPU was kept on the back

Affective State Analysis Through Visual and Thermal Image Sequences

67

Fig. 1 Experimental setup

end along with other peripherals shown in Fig. 1. The temperature and illumination of the room was maintained at 22 o C and 150 lux, respectively, to resolve any bias on the thermal and visual images, respectively. The participants were instructed to sit and relax on a height-adjustable chair while they watch the video clips. There were 30 males and 21 females ranging in age from 18 to 40 (μ=27.118, σ =5.743). Table 1 shows the statistics for all the 51 subjects.

3 Emotion Classification In this work, we propose the cascade of CNNs which we named as the Emotionet [9]. The first network of the cascade is for face detection, while the second network classifies the state of emotion. The transition of the networks occurs via a cropping operation, where the coordinates returned by Network 1 is used to crop the input image to obtain the face. Hence, we have two such Emotionets trained: one for the visual modality while the other one for the thermal modality as shown in Fig. 2. However, the net is trained separately for each modality, which yields a different set of weights for each case. The output node of each Emotionet has a Softmax activation function which gives the marginal density of being in a specific emotion by the concerned modality. Once we obtain the marginal density of emotions from both the modalities, we obtain the joint distribution as a product of marginal distributions.

68

S. Nayak et al.

Fig. 2 Architecture for proposed deep network

3.1 First Network: Face Detection The first network in the cascade takes in input images of size 1280 × 720 × 3, containing the human face. In this model, we used the notation as C L , PL , and FC L for the convolution, pooling, and fully connected layers, where L denotes the layer number. The first convolution layer C1 has 64 learnable filters of size 5 × 5 with a stride of 3, to convolve through the input image to produce the activation feature maps. The second layer P2 is a pooling layer of size 3 × 3 with a stride of 2 [10]. The third layer C3 has 64 filters of size 5 × 5 with a stride of 3. The fourth layer P4 is a pooling layer of size 3 × 3 with a stride of 2. The next layer C5 has 128 filters of size 5 × 5 and a stride of 3. This layer is followed with a pooling layer P6 of size 3 × 3 with a stride of 2. The next two layers, FC7 and FC8, are fully connected with five output nodes in the final layer. The first two output nodes x, y denote the left-most coordinates of the face region window. Similarly, the next two output nodes w, h denote the width and height, respectively, of the face window. The last node P indicates the probability of the presence of a face in the input image.

3.2 Second Network: Emotion Classification The output at the first network is probability of the presence of a face, along with the four coordinates. If P is equal or less than 0.5, the network assumes that a face is not found and hence it inputs at the next frame. When the face is found, it is cropped based on the estimated coordinates, and transferred to the next network. The first layer of the second network is a convolutional layer C9 having 64 filters of size 5 × 5 and a stride of 3. The next layer P10 is a pooling layer of size 3 × 3 with a stride of 2. The subsequent layer C11 has 128 filters of size 5 × 5 and a stride of 3.

Affective State Analysis Through Visual and Thermal Image Sequences

69

The high number of filters is set to encode the different within-class variations of the face with respect to the emotions. The next layer P12 is a pooling layer of size 3 × 3 with a stride of 2. The following layer C13 has 64 filters of size 5 × 5 and a stride of 3. The next layer P14 has a size of 5 × 5 and a stride of 3. The last fully connected layer FC16 has 3072 nodes at the input with seven at the output. Each output node corresponds to the probability of having one of the seven emotional states such as anger, disgust, fear, happy, neutral, sad, and surprise. We used separate training sets for the two networks. For training the first network, we collect 800 images containing the face and 900 images devoid of the face to obtain positive (P=1) and negative (P=0) training samples, respectively. The images were randomly sampled from the data collected in the experiment. The learning uses a backpropagation method where stochastic gradient descent (SGD) is used. The SGD method optimizes the weights by computing derivatives after reflecting the differences between desired and calculated outputs. The error function for this network is the mean squared error for each prediction as φ1 =



(x − xˆi )2 + (y − yˆi )2 + (w − wˆ i )2 + (h − hˆ i )2 + (P − Pˆi )2

(1)

i

Here, the ˆ denotes the predicted values for a given epoch. We use a learning rate λ = 0.001. For the training of the second network, we use the same protocol, except for the cost function and the training images. Here we define the cost function as φ2 =

 (E − Eˆ i )2

(2)

i

We used 350 training images for each class with the labels Anger (1), Disgust (2), Fear (3), Happy (4), Neutral (5), Sad (6), and Surprise (7). As explained earlier, an individual can have a mixture of emotions at any given instance. As such, the output of each emotion net is a normalized score in the range [0, 1]. If the scores are P(Vk ) and P(Tk ) for the visual and thermal images, respectively, then the final combined score P(E k ) is the product of these given as P(E k ) = P(Vk )P(Tk )

(3)

4 Negative Affect Recognition We find that the primary negative affective states are anxiety (Φ), depression (Δ), and stress (σ ) based on the pioneer work of DASS [11]. From the Emotionet, we obtain a normalized score of a specific emotion through the visual image as P(VE ), and

70

S. Nayak et al.

thermal image P(TE ) for a particular instance E. Since the output layer of Emotionet has a Softmax non-linearity, the output layer provides the normalized intensities of the emotion. If we view the normalized score as a probability of being in an affective states A, where A = [Φ Δ σ ] is given as P(Ak |VEk ). P(Ak |TEk ). The posterior probability P(Ak |E k ) can be obtained from the Bayes’ theorem as P(Ak |E k ) =

P(E k |Ak )P(Ak ) P(E k )

(4)

We use an Hidden Markov Model (HMM) to cluster the sequence of emotions. The model is shown in Fig. 3. For a given frame of visual and thermal face image, we obtain the intensities of being in a specific emotion. The state with the maximum intensity is selected as the dominant emotion E k for that frame k. Thus, we obtain the

Fig. 3 Proposed model

Affective State Analysis Through Visual and Thermal Image Sequences

71

sequence of emotional transitions. The seven emotional states E k are the observables of our HMM, given as O = [A D F H N S U] (5) The affective states form the hidden states of our model, as clearly they are unobserved. We define the four hidden states as H = [Δ α σ ]

(6)

4.1 Training the HMM The training process involves finding the HMM parameters λ = {A, B, π } to maximize the probability, i.e., P(O|λ) among the given probability distributions values. The HMM can be built from the given sequence of observations. We use the standard Baum–Welch forward–backward algorithm for the purpose. From the experimental data of DASS, we find the following parameters: π = (0.21, 0.21, 0.32) ⎞ 0.52 0.19 0.29 A = ⎝0.21 0.56 0.23⎠ 0.44 0.18 0.38

(7)

⎛

⎞ 0.19 0.09 0.05 0.05 0.05 0.42 0.15 B = ⎝0.11 0.11 0.33 0.11 0.11 0.11 0.11⎠ 0.50 0.20 0.06 0.06 0.06 0.06 0.06

(8)

⎛

(9)

4.2 Clustering of Emotion Sequences Let us consider we have a sequence of emotions E = E 1 , E 2 , . . . , E N . The problem is to cluster such sequences in three clusters such as depression (Φ), anxiety (Δ), and stress (σ ). The reason of using HMM for clustering over superior methods such as k-means or Gaussian mixtures is that we have a time-series data instead of a feature vector. We use a finite mixture model as f (E) =

4 

f j (E|λ j )w j

(10)

j=1

Here, E denotes a sequence of emotions and w j is the weight given to the jth model λ j . The density function f j (E|λ j ) is computed for the forward part of the forward–

72

S. Nayak et al.

Table 1 Confusion matrix of HMM classifier Φ Φ Δ σ

9 1 1

Δ

σ

1 8 1

1 2 14

Table 2 Performance fivefold cross-validation of Emotionet network 2 Set 1 Set 2 Set 3 Set 4 Visual Thermal

93.45 90.32

93.24 88.67

91.21 87.58

89.76 91.12

Set 5 92.54 87.76

backward algorithm. If A1 , A2 , A3 represents the transition matrices of each HMM, then the resultant transition matrix A can be represented in block-diagonal form as ⎛ ⎞ A1 0 0 A = ⎝ 0 A2 0 ⎠ 0 0 A3

(11)

The initial state probabilities form the relative weights w j of the mixture components. Given this composite HMM, the first step is to initialize the parameters of the model. Next, we update them using the Baum–Welch algorithm to maximize the likelihood. The confusion matrix used for our experimental purpose is detailed in Table 1.

5 Experimental Results Since the paper has two phases, we used the information from the emotion-elicited videos containing the faces to analyze phase I, while using the DASS scores to validate the phase II. The performance fivefold cross-validation of Emotionet Network 2 is detailed in Table 2.

6 Conclusion In this paper, we have proposed an HCI subsystem for affect recognition from thermal and visual facial image sequences. We achieve this by detecting the face, identifying the emotional expression and then clustering a sequence of emotions as stressed, anxious, or depressed. The proposed framework has two phases: the first being the

Affective State Analysis Through Visual and Thermal Image Sequences

73

classification of emotions from an ensemble of visual and thermal image cues, while the second stage uses such a sequence to cluster the affective state. The first phase is achieved using a two-stage cascaded CNN, named as the Emotionet. The second phase of the framework has been validated using the psychometric scores of the DASS-21 scale. The framework can be extended to develop a real-time system to estimate the automatic affective states with appropriate feedback during HCI.

References 1. Zeng, Z., Pantic, M., Roisman, G.I., Huang, T.S.: A survey of affect recognition methods: audio, visual, and spontaneous expressions. IEEE Tran. Pattern Analy. Mach. Intell. 31(1), 39–58 (2008) 2. Zheng, W.-L., Liu, W., Lu, Y., Lu, B.-L., Cichocki, A.: Emotionmeter: a multimodal framework for recognizing human emotions. IEEE Trans. Cyber. 49(3), 1110–1122 (2018) 3. Lovibond, P.F., Lovibond, S.H.: The structure of negative emotional states: comparison of the depression anxiety stress scales (dass) with the beck depression and anxiety inventories. Behav. Res. Therapy 33(3), 335–343 (1995) 4. Rymaszewska, J., Kiejna, A., Hadry´s, T.: Depression and anxiety in coronary artery bypass grafting patients. Euro. Psych. 18(4), 155–160 (2003) 5. Basu, A., Dasgupta, A., Thyagharajan, A., Routray, A., Guha, R., Mitra, P.: A portable personality recognizer based on affective state classification using spectral fusion of features. IEEE Trans. Affect. Comput. 9(3), 330–342 (2018) 6. Happy, S., Routray, A.: Automatic facial expression recognition using features of salient facial patches. IEEE Trans. Affec. Comput. 6(1), 1–12 (2014) 7. Nayak, S., Happy, S., Routray, A., Sarma, M.: A versatile online system for person-specific facial expression recognition. In: TENCON 2019-2019 IEEE Region 10 Conference (TENCON). IEEE, pp. 2513–2518 (2019) 8. Khan, M.M., Ward, R.D., Ingleby, M.: Classifying pretended and evoked facial expressions of positive and negative affective states using infrared measurement of skin temperature. ACM Trans. Appl. Percept. (TAP) 6(1), 6 (2009) 9. Marquez, E.S., Hare, J.S., Niranjan, M.: Deep cascade learning. IEEE Trans. Neural Netw. Learn. Syst. 29(11), 5475–5485 (2018) 10. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1–9 (2015) 11. Lovibond, S.H., Lovibond, P.F.: Manual for the depression anxiety stress scales. Psychology Foundation of Australia (1996)

An Ensemble Deep Learning Method for Diabetes Mellitus N. Komal Kumar, D. Vigneswari, Rahul J. Reynold, Jojo Josy, and Jerin C. Prince

Abstract Diabetes mellitus is a chronic condition that results in too much sugar in the blood. They are of different types, which include Type 1, Type 2, Prediabetes and Gestational diabetes. Hybrid machine algorithms can be designed in a way to identify and predict the type of diabetes among patients. The aim of this paper is to develop an ensemble deep learning method for diabetes mellitus (DM). The work is classified into two stages, first stage involves data preprocessing and classification, the preprocessed data are fed to multiple classifiers, resulting in a classified output. In the second stage, the classified output is fed to the multi-layer perceptron for optimized output. In the evaluation, the proposed hybrid EB-MLP hybrid classifier achieved an accuracy of 92.3%, precision 92.4%, recall 87.4% and execution time 0.028 s. which outperformed other classifiers under analysis. Keywords Deep learning · Ensemble technique · Diabetes mellitus · Classification

N. K. Kumar (B) · R. J. Reynold · J. Josy · J. C. Prince Department of Computer Science and Engineering, St. Peter’s Institute of Higher Education and Research, Avadi, Chennai, India e-mail: [email protected] R. J. Reynold e-mail: [email protected] J. Josy e-mail: [email protected] J. C. Prince e-mail: [email protected] D. Vigneswari Faculty of Engineering and Technology, Sri Ramachandra, Sri Ramachandra Institute of Higher Education and Research, Porur, Chennai, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_8

75

76

N. K. Kumar et al.

1 Introduction Diabetes mellitus is a lifelong condition that affects the body to use the energy in it. Normally, the body has the ability to break down the sugar and carbohydrates into a unique sugar called glucose. Body burns the glucose to get the energy which are store to carryout work. To burn the glucose, the body needs the catalyst cell called insulin, an hormone that is present in the bloodstream. When the person is subjected to diabetes mellitus, either the case occurs, either the pancreas does not produce the required amount of insulin to burn glucose to produce energy or the produced glucose is not sufficiently utilized. Type 1 diabetes is called insulindependent diabetes, where the affected pancreas does not produce enough insulin, resulting in the risks such as diabetic retinopathy, diabetic neuropathy or even diabetic nephropathy. Type 2 diabetes is also called adult-onset diabetes, which are caused due to obese and overweight. Type 2 diabetes is not severe compared with Type 1, where the insulin produced by the pancreas is not sufficient for the body’s needs. There is no cure for diabetes; it can be reduced by weight reduction, increasing nutritious diet and exercise. Gestational diabetes usually occurs during pregnancies in women, where the risks of an unborn baby are greater than the mother. Anti-diabetic agents can be used to solve chronic disease, some of them include Biguanides, glipizide, glyburide, pioglitazone, sitagliptin, saxagliptin, exenatide and liraglutide [1]. The majority of the present anti-diabetic agents, however, exhibit numerous side-effects. In addition, insulin therapy is related to weight gain and hypoglycemic events. Hence, discovering of anti-diabetic drug is a research challenge [2, 3]. With the rapid development in medical diagnosis, machine learning and prediction algorithms play a paramount role in identifying and predicting diabetes mellitus (DM).

2 Related Works This section briefly discusses the recent works on incremental Diabetes Mellitus and machine learning classifiers. D. Gregori [4] suggested various data mining techniques such as classification and rule mining approaches where diabetes can be monitored and taken care. The work proposed in [5] used an association rule mining algorithm in extracting the risk patterns in Type 2 diabetes using a Tehran fluid and glucose database study. The work revealed many risk patterns subjected to Type 2 diabetics in patient under study. Further, the risk patterns are found using the temporal patterns in multivariate time series data [6]. The work in [7] designed a fuzzy classifier for diabetes disease with a modified artificial bee colony, where the fuzzy classifier produced effective results compared to the other classifiers. G. Robertson [8] proposed an AIDA diabetic simulator combined with artificial neural networks for predicting blood glucose. The work produced simulated results of glucose level in blood. S. Belciug [9] proposed an automated medical diagnosis by error-correction

An Ensemble Deep Learning Method for Diabetes Mellitus

77

learning for artificial neural networks using Bayesian paradigm. The work significantly changed the way automated medical diagnosis works. A. Ozcift [10] carried out a comprehensive to increase the performance of machine learning algorithms by a classifier ensemble with rotation forest. The results obtained outperformed than other classifiers under study. S. Bashir [11] proposed a medical decision support application using a novel multi-weighted multi-layer classifier framework. X.H. Meng [12] compared three different data mining models for predicting diabetes, C5.0 classifier produced higher accuracy than ANN. The authors in [13] made a study on screening of type 2 diabetes and risk factors using decision tree classifier, the decision tree classifier screened the demographics of various patients subjected to diabetes mellitus and made decision based upon the symptoms. A. Worachartcheewan [14] predicted metabolic syndrome using random forest classifier, where the various metabolic syndromes are classified on their symptoms. The work by L.F. Chen [15] proposed an optimized particle swarm optimization for feature selection. S. Mani [16] proposed to type 2 risk forecasting using Electronic Medical Records. The results forecasted several risk factors causing type 2 diabetes.

3 Dataset Description We experimented with our proposed method on PimaIndian dataset from kaggle database [18], for Type 1 and Type 2 Diabetes Mellitus. The dataset contained nine attributes and 768 observations. Dataset contained attributes like glucose concentration, blood pressure, triceps skinfold thickness and secretion of insulin for every 2 h, BMI of the patient, Diabetes Pedigree Function (DPF), age of the patient and a class variable (1 or 0). Table 1 shows the attributes and definitions subjected to the study. Table 1 Attribute and definition of diabetes mellitus dataset S. No

Attribute

Definition

1

Pregnancies

Number of times pregnant

2

Glucose

Plasma glucose concentration for 2 h in an oral glucose tolerance test

3

Blood pressure

Diastolic blood pressure (mm Hg)

4

Skin thickness

Triceps skinfold thickness (mm)

5

Insulin

2-h serum insulin (mu U/ml)

6

BMI

Body mass index (weight in kg/(height in m)ˆ2)

7

Diabetes Pedigree Function

Diabetes pedigree function

8

Age

Age (in years)

9

Outcomes

1 or 0

78

N. K. Kumar et al.

4 Proposed Methodology This section describes the proposed methodology. Several machine-learning classifiers [19] are used in the prediction of various diseases. It is not the fact that these classifiers produce accurate results, combining many classifiers can achieve a greater accuracy [20] than a single classifier. We propose an ensemble-based deep learning technique for diabetes mellitus. The works focus on two phases, learning and optimizing, learning phase deals with the fact that the dataset containing preprocessed values are subjected to a cross-validation technique [21] such as K-fold method, where the dataset is divided into a training set and a test set, the test set is fed to multiple classifiers such as Random Forest, k-NN, Decision Tree, CART, and C4.5. Each classifier produces hypotheses, thus making an optimized dataset. In the second phase, the optimized dataset is again subjected to K-fold validation for dividing as an optimized training and a test set, the optimized test set is fed to a multilayer perceptron, which produces many outcomes, further the outcomes are optimized as final outcomes (Fig. 1). Algorithm 1: Optimized dataset Input: Dataset (D) Output: Optimized Dataset (Dop ) begin Let D = {D1 ,D2 ,D3 ,….Dn ) be the dataset values Test = K-fold(D) Classout = classifier (Test) Let h1(x), h2(x), h3(x), h4(x), h5(x) be the hypothesis produced by the classifiers Dout = Append(h1(x), h2(x), h3(x), h4(x), h5(x)) end Let DMLP = (oc1 ,oc2 ,oc3 ….) (continued)

Training Dataset Classifiers

Dataset

Hypothesis

Test Dataset New Training Dataset New Dataset Optimized Outcome

Multi-Layer Perceptron

Fig. 1 EB-MLP framework

New Test Dataset

An Ensemble Deep Learning Method for Diabetes Mellitus

79

(continued) Algorithm 1: Optimized dataset DMLP(optimized) = selection(oc1 ,oc2 ,oc3 ….) end

Algorithm 2: EB-MLP Input: Optimized dataset (Dout ) Output: Optimized outcome (DMLP(optimized) ) begin Let Dout be an optimized dataset Dout (test) = K-fold(D) Dtest = Dout (test) DMLP = MLP(Dtest ) Let DMLP = (oc1 ,oc2 ,oc3 ….) DMLP(optimized) = selection(oc1 ,oc2 ,oc3 ….) end

5 Results and Discussions The main aim of the proposed method is to establish an optimized outcome from the multilayer perceptron for diabetes mellitus. We analyzed the proposed method on five important classifiers metrics such as precision, recall, accuracy, sensitivity and specificity [22]. We applied classifiers to the entire dataset and selected dataset; the results obtained by the classifiers for the entire dataset are shown in Fig. 2, the results obtained by the selected dataset are shown in Fig. 3., and the execution time of entire and selected datasets are shown in Fig. 4. The confusion matrices of the five classifiers are obtained from the output of the classifiers; the performance metrics such as precision, recall, and accuracy for entire dataset are computed [23–25]. K-NN achieved an accuracy of 76.2%, which is slightly lesser than the decision tree classifier which achieved 77.7%, CART achieved an accuracy of 63.2% whereas random forest and C4.5 classifier achieved 67.8% and 78.6%, respectively. Random forest achieved precision of 83.9%, which is lesser than C4.5 classifiers, k-NN achieved 75.5% precision, which is lesser than the precision of CART classifier 81.2%, and C4.5 classifier achieved a precision of 84.5%, which is greater than the other classifiers under comparison. The execution times of the classifier of entire dataset are obtained, the decision tree outperformed than other classifiers taking only 0.046 s, whereas random forest took 0.086 s, k-NN took 0.048 s slightly higher than the decision tree classifier. The performance metrics of the selected dataset is compared with the metrics of entire dataset containing the

80

N. K. Kumar et al.

Fig. 2 Performance metrics comparison of classifiers

Fig. 3 Performance metrics comparison of multilayer perceptron with ensemble

proposed EB-MLP hybrid classifier. EB-MLP outperformed all general classifiers achieving higher accuracy, precision, recall values than the five classifiers under comparison and with less execution time. The performance metrics comparison for the selected dataset is shown in Fig. 3. EB-MLP achieved an accuracy of 92.3%, k-NN achieved 86.5%, random forest achieved 75.4%, decision tree achieved 85.1% and C4.5 managed to achieve 80.2%. the precision and recall values of random forest

An Ensemble Deep Learning Method for Diabetes Mellitus

81

Fig. 4 Execution time

are 84% and 84.2%, k-NN are 76.9% and 77.0%, decision tree are 73.8% and 82.8%, CART are 82.6% and 83.6%, C4.5 are 85.5% and 83.9% and EB-MLP are 92.4% and 87.4% respectively. EB-MLP hybrid classifier executed in 0.028 s, which is lesser than the five classifiers, random forest took 0.076 s, decision tree took 0.031 s, k-NN classifier took 0.036 s, CART took 0.041 s and C4.5 took 0.059 s for producing the result.

6 Conclusion and Future Direction An ensemble-based deep learning approach is proposed in this paper. The main idea of this paper is to optimize the dataset by ensemble learning method with deep learning methodology for diabetes mellitus. The performance metrics of the proposed hybrid classifier EB-MLP is compared with the existing classifiers and the results were obtained. EB-MLP outperformed with an accuracy of 92.3%, precision of 92.4%, and recall of 87.4% with less execution time of only 0.028 s. Hybrid algorithms can be constructed with the help of hybrid ensemble learning methods in the future.

References 1. Krentz, A.J., Bailey, C.J.: Oral antidiabetic agents: current role in type 2 diabetes mellitus Drugs, 65(3), 385–411 (2005) 2. Tsave, O., Halevas, E., Yavropoulou, M.P., KosmidisPapadimitriou, A., Yovos, J.G., Hatzidimitriou, A. et al.: Structure-specific adipogenic capacity of novel, well-defined ternary Zn(II)Schiff base materials. Biomolecular correlations in zinc induced differentiation of 3T3-L1 pre-adipocytes to adipocytes. J InorgBiochem, 152(Nov 2015), 123–137 (2015)

82

N. K. Kumar et al.

3. Gregori, D., Petrinco, M., Bo, S., Rosato, R., Pagano, E., Berchialla, P. et al.: Using data mining techniques in monitoring diabetes care. The simpler the better? J. Med. Syst. 35(2), 277–281 (2011) 4. Ramezankhani, A., Pournik, O., Shahrabi, J., Azizi, F., Hadaegh, F.: An application of association rule mining to extract risk pattern for type 2 diabetes using Tehran lipid and glucose study database. Int. J. Endocrinol. Metab. 13(2) (2015) 5. Batal, I., Fradkin, D., Harrison, J., Moerchen, F., Hauskrecht, M.: Mining recent temporal patterns for event detection in multivariate time series data. KDD, pp. 280–288 (2012) 6. Beloufa, F., Chikh, M.A.: Design of fuzzy classifier for diabetes disease using modified artificial bee colony algorithm. Comput. Meth. Programs Biomed 112(1), 92–103 (2013) 7. Robertson, G., Lehmann, E.D., Sandham, W.A., Hamilton, D.J.: Blood glucose prediction using artificial neural networks trained with the AIDA diabetes simulator: a proof-of-concept pilot study. J. Electr. Comput. Eng., 681786:1–681786:11 (2011) 8. Belciug, S., Gorunescu, F.: Error-correction learning for artificial neural networks using the Bayesian paradigm. Application to automated medical diagnosis. J. Biomed. Inform. 52, 329– 337 (2014) 9. Ozcift, A., Gulten, A.: Classifier ensemble construction with rotation forest to improve medical diagnosis performance of machine learning algorithms. Comput. Methods Progr. Biomed. 104(3), 443–451 (2011) 10. Bashir, S., Qamar, U., Khan, F.H.: IntelliHealth: a medical decision support application using a novel weighted multi-layer classifier ensemble framework. J Biomed. Inform. 59(Feb 2016), 185–200 (2016) 11. Meng, X.H., Huang, Y.X., Rao, D.P., Zhang, Q., Liu, Q.: Comparison of three data mining models for predicting diabetes or prediabetes by risk factors. Kaohsiung J. Med. Sci. 29(2), 93–99 (2013) 12. Habibi, S., Ahmadi, M., Alizadeh, S.: Type 2 diabetes mellitus screening and risk factors using decision tree: results of data mining. Glob. J. Health Sci. 7(5), 304–310 (2015) 13. Worachartcheewan, A., Shoombuatong, W., Pidetcha, P., Nopnithipat, W., Prachayasittikul, V., Nantasenamat, C.: Predicting metabolic syndrome using the random forest method. Sci. World J. 2015, 581501 (2015) 14. Chen, L.F., Su, C.T., Chen, K.H.: An improved particle swarm optimization for feature selection. Intell. Data Anal. 16(2), 167–182 (2012) 15. Mani, S., Chen, Y., Elasy, T., Clayton, W., Denny, J.: Type 2 diabetes risk forecasting from EMR data using machine learning. AMIA AnnuSympProc 2012(2012), 606–615 (2012) 16. American Diabetes Association Diagnosis and classification of diabetes mellitus Diabetes Care, 32(Suppl. 1), S62-S67 (2009) 17. https://www.kaggle.com/uciml/pima-indians-diabetes-database 18. Komal Kumar, N., Lakshmi Tulasi, R., Vigneswari, D.: An ensemble multi-model technique for predicting chronic kidney disease. Int. J. Elect. Comput. Eng. 9(2), 1321–1326 (2019) 19. Komal Kumar, N., Vigneswari, D., Roopa Devi, B.A.S.: MSO–MLP diagnostic approach for detecting DENV serotypes. Int. J. Pure Appl. Mathemat. 118(5), 1–6 (2018) 20. Komal Kumar, N., Vigneswari, D., Kavya, M., Ramya, K., Lakshmi Druthi, T.: Predicting non- small cell lung cancer: a machine learning paradigm. J. Comput. Theor. Nanosci. 15(6/7), 2055–2058 (2018) 21. Komal Kumar, N., Vigneswari, D., Vamsi Krishna, M., Phanindra Reddy, G.V.: An optimized random forest classifier for diabetes mellitus. In: Abraham, A., Dutta, P., Mandal, J., Bhattacharya, A., Dutta, S. (Eds.), Emerging Technologies in Data Mining and Information Security. Advances in Intelligent Systems and Computing, vol. 813. pp. 765–773 (2019) 22. Vigneswari, D., Komal Kumar, N., Ganesh Raj, V., Gugan, A., Vikash, S.R.: Machine learning tree classifiers in predicting diabetes mellitus. In: 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS), pp. 84–87 (2019) 23. Komal Kumar, N., Vigneswari, D.: Hepatitis- infectious disease prediction using classification algorithms. Res. J. Pharm. Tech. 12(8), 3720–3725 (2019)

An Ensemble Deep Learning Method for Diabetes Mellitus

83

24. Kumar, N.K., Tulasi, R.L., Vigneswari, D.: Investigating dengue outbreak in Tamil Nadu, India. Indonesian J. Elect. Eng. Comput. Sci. 18(1), 502–507 (2020) 25. Kumar, N.K., Sindhu, G.S., Prashanthi, D.K., Sulthana, A.S.: Analysis and prediction of cardio vascular disease using machine learning classifiers. In: 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS), Coimbatore, India, pp. 15–21 (2020)

Memory Fluid Flow Past a Vertical Circular Cylinder and Its Energy Transfer Debasish Dey, Ashim Jyoti Baruah, and Rupjyoti Borah

Abstract Time dependent solution stream satisfying constitutive equation of memory fluid around a vertical round cylinder has been considered. Flow is guided by buoyancy force and Lorentz force. Free or natural convection appears in the flow system due to density variation caused by Buoyancy force. Thermal energy is generated into the system due to energy dissipation. Suitable similarity transformation and numerical technique “MATLAB bvp4c solver technique” has been used to carry out the numerical solutions. Keywords Memory fluid · Lorentz force · Buoyancy · Circular cylinder

1 Introduction Fluid flow using constitutive model of Walters liquid has been used extensively in engineering and industrial works. This encourages various scientists or researchers to study fluid flow through various geometries. Recently, memory fluid flow with heat transfer has been analysed in various works [1–7]. Choudhury and Deka [8] have made an attempt to study the boundary layer separation of visco-elastic fluid flow around a circular cylinder. Veerakrishna et al. [9, 10] have carried out the solution of hydromagnetic second grade fluid flow including the influences of radiation and Hall current. For MHD flow of nanoparticles, study of heat and mass transfer was done by Tausee Mohyud-din et al. [11]. Chopra and Mittal [12] have examined the numerical simulations of flow past a circular cylinder. The objective of our work is to study the boundary layer flow of memory fluid around vertical circular cylinder in presence of buoyancy force, Lorentz force and energy dissipation.

D. Dey (B) · R. Borah Department of Mathematics, Dibrugarh University, Dibrugarh 786004, India e-mail: [email protected] A. J. Baruah Department of Mathematics, Namrup College, Dibrugarh, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_9

85

86

D. Dey et al.

2 Mathematical Formulation A time independent memory fluid stream around straight up circular tube has been taken with heat transfer influences. Lorentz force is generated into the system due to the introduction of crossway magnetic field. Apart from Lorentz force in the body force, other forces present in the system are pressure gradient, viscous and visco-elastic forces. Heat is generated into the system because of energy dissipation. In geometry, the axis of the tube is taken along z-axis, r measured along radial direction. Using Boussinesq approximation and boundary layer assumptions, the leading equations of stream are:





2 

u  ∂∂ux  + v  ∂u − v ∂∂ yu2 + ∂ y



∂v  ∂u  =−   ∂x ∂y ∂u  ∂ 2 u  ∂ x  ∂ y 2  

∂u  ∂ 2 v  ∂ y  ∂ y 2  

(2.1) 3 

+ u  ∂ x∂ ∂uy 2 + v ∂∂ yu3  2 B −σ ρ0 U − u − gβ T − T∞ − U ddUx  = 0 k0 ρ

+

3



      2   2   ∂u  ∂ 2 u   ∂u  ∂ 3 u  v ρC p u  ∂∂Tx  + v  ∂∂Ty  − k ∂∂ yT2 − μ ∂u + k + u 0   2   2 ∂y ∂y ∂y ∂y ∂x ∂y =0

(2.2)

(2.3)

and the restrictions imposed on the boundary are y  = 0 : u  = 0, v  = V0 , T  = Tw &y  → ∞ : u  → U, T  → T∞

(2.4)

where, y  = r −a, U = 2U∞ sinθ is the velocity of uniform stream and x  is measured along cross radial direction at the surface, u , v and T  velocity components along x  and y directions and temperature of fluid respectively. = 0 & v  − ∂ψ = 0 and following Now we initiate a stream function ψ; u  + ∂ψ ∂y ∂x similarity variables 

va ψ= θ 2U∞ f (η), η = y  2U∞ B2a M =σ 0 , 2ρU∞ Gr =



2U∞ ∗ T − T∞ 2U∞ k0 , T (η) = , ,e = va Tw − T∞ ρva

2 μC p 4U∞ μ gβ(Tw − T∞ ) , Ec = 2 , Pr = 2 4U∞ k a (Tw − T∞ )k

(2.5)

Using (2.5) into (2.2) and (2.3), we get following ordinary differential equations: f



2



− ff = f



        − e 2 f f − f 2 − f f iv + M 1 − f + Gr T ∗

(2.6)

Memory Fluid Flow Past a Vertical Circular Cylinder and Its Energy Transfer

         − Ec f 2 T ∗ + Pr f T ∗ = Ece f f 2 − f f f

87

(2.7)

Restrictions on boundaries in terms of similarity variables are: η = 0 : f  = 0, f = s, T ∗ = 1&η → ∞ : f  → 1, T ∗ → 0

(2.8)

In Eqs. (2.6)–(2.8) primes indicate the differentiations with respect to η.

3 Method of Solution Numerical results of (2.6) and (2.7) with the restriction (2.8) are carried out using MATLAB constructed bvp4c solver procedure. The solver bvp4c takes a finite difference code. The user have to put the resulting ordinary differential equations in terms of first order differential equations by introducing new variables as follows: y1 = f, y2 = f  , y3 = f  , y4 = f  , y5 = T ∗ , y6 = T ∗ Then, the set of first order differential equations is: y1 = y2 , y2 = y3 , y3 = y4 , y4   2 y2 − y1 y3 − Gr y5 − M(1 − y2 ) − y4 + 2ey2 y4 − ey32 = ; ey1   y5 = y6 , y6 = eEc y2 y32 − y1 y3 y4 − Ecy32 − Pr y1 y6 . The two-point boundary conditions (2.8) of these ordinary differential Eqs. (2.6) and (2.7) are coded in the following ways: y1(2) − 1, y1(3), y1(5). y0(2), y0(1) − s, y0(5) − 1;

4 Results and Discussion The results as we obtained show the influences of e (visco elastic parameter), Ec (Eckert number), Gr (Grashoff number for thermal diffusion), M (magnetic parameter), Pr (Prandtl number) and s (suction parameter) on velocity component and temperature profile of the fluid flow. Also, the numerical values of shear stress (σ) when Pr = 1.0 and Ec = 0.05 for varying values of M and Gr are shown in tabular

88

D. Dey et al.

form. For graphical results we have considered values of the parameters as shown in the figures. Figures 1 and 2 display the influences of viscoelastic and magnetic field parameters on the primary velocity component distribution against displacement variable η. Stream experiences an acceleration with visco-elasticity e and deceleration with magnetic parameter M. Graphical representation of speed allocation over distance parameter is portrayed by Fig. 3 and we can see that acceleration in stream with Fig. 1 Velocity distribution for a variety of ‘e’ when s = 1

Fig. 2 Velocity distribution for a variety of ‘M’ when s=1

Memory Fluid Flow Past a Vertical Circular Cylinder and Its Energy Transfer

89

Fig. 3 Velocity distribution for a variety of ‘Gr’ when s=1

Grashoff number(representative of buoyancy force). Figures 4 and 5 show the influences of Pr and e on the heat allocation over distance variable. Motion is encountered with reduction in heat with Prandtl number and visco-elasticity. Figure 6 depicts the temperature profile of the fluid stream against η for various values of Eckert number. It has a negative effect on the temperature distribution of system. Figure 7 represents the distribution of velocity against η for a variety of s. An acceleration in stream is seen with s. Fig. 4 Temperature distribution for a variety of ‘Pr’ when s = 1

90

D. Dey et al.

Fig. 5 Temperature distribution for a variety of ‘e’ when s = 1

Fig. 6 Temperature distribution for a variety of ‘Ec’ when s = 1

Figure 8 and 9 shows the influences of Prandtl number and magnetic parameter on the temperature gradient of the fluid flow. Prandtl number has a retarding effect on the temperature change across the flow region. Enhancement in temperature gradeint is observed with the rise of Pr number for η > 0.4. From the Fig. 9 it is observed that temperature is enhanced with M, but the change of behavior is seen with the increasing values of M after η > 1. Table 1 shows the influence of M and Gr on shear stress,σ when Pr = 1.0 and Ec = 0.05. It is noticed that with the increasing values of M the shear stress, σ decreases for e = 0.2 but increases for e = 0.5. But for increasing values of Gr the shear stress is increased and is not effected by the change in the value of e.

Memory Fluid Flow Past a Vertical Circular Cylinder and Its Energy Transfer

91

Fig. 7 Velocity distribution for a variety of suction parameter s

Fig. 8 Temperature gradient for variety of Pr

5 Conclusion Major points from the above work are highlighted as below: • Acceleration in fluid motion is seen during the rise in viscoelasticity and Grashoff number • Decrease in speed is seen during the enhancement of strength of Lorentz force (Magnetic parameter) • Rise in temperature is noticed due to the growth of magnetic parameter.

92

D. Dey et al.

Fig. 9 Temperature Gradient for variety of M

Table 1 Numerical values of shear stress (σ ) when Pr = 1.0 and Ec = 0.05 E

M

σ

Gr

σ

0.2

0.5

0.78256

−0.5

−0.12632

1.0

0.74078

−0.2

0.0009

2.5

0.57658

0.5

0.22984

0.5

0.3934

−0.5

0.37345

1.0

0.4280

−0.2

0.4059

2.5

0.5200

0.5

0.5

0.41935

• Temperature fall is observed with Prandtl number. Future Scope The problem is solved numerically using arbitrary values of thermo-physical properties. In the future, the work may be extended by using themo-physical properties of the fluid used in industries.

References 1. Choudhury, R., Dey, D.: Free convective elastico-viscous fluid flow with heat and mass transfer past an inclined porous plate in slip flow regime. Latin Am. Appl. Res. 42(4), 327–332 (2012) 2. Dey, D.: Non-newtonian effects on hydromagnetic dusty stratified fluid flow through a porous medium with volume fraction. Proc. Natl. Acad. Sci., India, Sect. A 86(1), 47–56 (2016) 3. Dey, D.: Viscoelastic fluid flow through an annulus with relaxation, retardation effects and external heat source/sink. Alexandria Eng. J. 57(2), 995–1001 (2018)

Memory Fluid Flow Past a Vertical Circular Cylinder and Its Energy Transfer

93

4. Dey, D., Baruah, A.: J, Dusty memory fluid through a horizontal channel with energy transfer. AIP Conf. Proc. 2061(1), 020006 (2019) 5. Dey, D., Baruah, A.J.: Visco-elastic effects on nano-fluid flow in a rotating system in presence of hall current effect. In: Emerging Technologies in Data Mining and Information Security, pp. 575–585 6. Dey, D.: Modelling and analysis of bio-convective nano-fluid flow past a continuous moving vertical cylinder. In: Emerging Technologies in Data Mining and Information Security, pp. 331– 340 7. Dey, D., Baruah, A.J.: Free convective flow of two immiscible memory fluids in an inclined channel with energy dissipation. Model., Measur. Control B 87(2), 63–67 (2018) 8. Choudhury, R., Deka, B.: MHD visco-elastic fluid flow and heat transfer around a circular cylinder. WSEAS Trans. Fluid Mech. 12, 98–107 (2017) 9. Veera, K.M., Subba Reddy, G., Chamkha, A.J.: Hall effects on unsteady mhd oscillatory free convective flow of second grade fluid through porous medium between two vertical plates. Phys. Fluids 30, 023106 (2018). https://doi.org/10.1063/1.5010863 10. Veera, K.M., Chamkha, A.J.: Hall effects on unsteady MHD flow of second grade fluid through porous medium with ramped wall temperature and ramped surface concentration. Phys. Fluids 30, 053101 (2018). https://doi.org/10.1063/1.502554 11. Syed, T.M.D., Khan, U., Ahmed, N., Rashidi, M.M.: A study of heat and mass transfer on magnetohydrodynamic (MHD) flow of nanoparticles. Propul. Power Res. 7(1), 1–102 (2018). https://doi.org/10.1016/j.jppr.2018.02.001 12. Chopra, G., Mittal, S.: Numerical simulations of flow past a circular cylinder. J. Phys.: Conference series, 822, 15th Asian Congress of Fluid Mechanics. Malaysia (2016)

A Comparative Study on Financial Market Forecasting Using AI: A Case Study on NIFTY Bhaskar Mondal, Om Patra, Ashutosh Satapathy, and Soumya Ranjan Behera

Abstract Financial markets are fundamentally volatile, nonparametric, nonlinear, and unpredictable. It implies that investors have to contend with time-series on the commercial market, which is non-stationary, noisy, and unpredictable, making it hard to anticipate future indexes. Presently, without additional statistical details, we can extract features from a broad data framework. In this paper, algorithms such as ARIMA, Facebook’s Prophet Algorithm, Support Vector Regressor, Long ShortTerm Memory, and Gated Recurrent Unit are used to forecast NIFTY based on historical data available and compared the accuracy of prediction. In our empirical analysis, all the NIFTY 50 daily high-recurrence exchange information were used, for instance. It was noted that an increase in data count increases predictive efficiency. This demonstrates that deep learning exploits transaction data’s nonlinear characteristics and can serve as a reliable financial index forecast for stock-market investors. The result obtained was contrasted with each of the algorithms we used, and the recurrent neural systems were found to outflank the current models.

B. Mondal Department of Computer Science & Engineering, National Institute of Technology Patna, Patna, India e-mail: [email protected] O. Patra (B) · A. Satapathy · S. R. Behera School of Computer Science & Engineering, Xavier University Bhubaneswar, Bhubaneswar, India e-mail: [email protected] A. Satapathy e-mail: [email protected] S. R. Behera e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_10

95

96

B. Mondal et al.

1 Introduction In the last decade, interest in the use of artificial neural networks (ANNs) for forecasting has resulted in an enormous increase in research activities. ANN offer a potential alternative approach to traditional linear statistical methods [8, 13]. Prediction of the financial market is an on-demand domain and is a well-known problem for researchers. Many researchers have suggested different strategies to solve this problem, and various theories for this problem space are available. One such theory is named the Efficient Market Hypothesis (EMH) [3] Efficient Market Hypothesis expresses that mutual costs reflect all crucial data, and it is not very easy to beat the market. There is likewise another hypothesis called Random Walk Theory [7], which expresses that adjustments in stock costs happen arbitrarily, and these irregular movements cannot be predictable. In addition to these hypotheses, two different trading techniques were created for the prediction of the stock market. Such techniques are termed fundamental analysis and technical analysis. Over the past few decades, several ANN and hybrid models are being proposed to outperform traditional statistical approaches. This paper evaluates the effectiveness of neural network designs recognized in predictions as dynamic and reliable. The frameworks are multi-layer perceptron (MLP), dynamic artificial neural network (DAN2), and hybrid neural networks that generate new input variables using generalized self-regressive conditional heteroscedasticity (GARCH) [5, 6]. CNN was seen to be outperforming the other models. Although trained with NSE data, the network was able to predict NYSE. This was likely because the two stock markets share some of the traditional internal dynamics. The results obtained were compared with the linear model because ARIMA is a univariate time series prediction and therefore cannot recognize underlying dynamics in different time series. Wang, Zhang, and Guo proposed a new method for forecasting stock price using neural wavelet denoising back propagation network. For the forecasting of stock prices, a functional algorithm was developed. Wavelet transform is a valuable pre-processing technique for data that can be combined with other forms of prediction, such as statistical and other AI models [10]. A hybrid solution incorporating ARIMA, BPNN, and ESM has been proposed to be the most efficient of all three models since the use of a single classic model can not produce precise stock price index forecasts. Genetic algorithm determines the weight of the proposed hybrid model [11]. This paper showed that the use of more robust nonlinear techniques, such as recurrent neural networks, to construct predictive financial models has seen significant increases in the accuracy and performance recorded on benchmark datasets in recent times. Advanced statistical techniques achieve higher performance compared to conventional analytical methods, and the use of advanced techniques, such as recurrent neural networks, was important because they were capable of a more accurate predictive financial model showing interactions and curvature. The sub-sections below provide a brief discussion of some of the classical and advanced statistical models used to determine Indian financial markets.

A Comparative Study on Financial Market Forecasting Using AI …

97

2 Preliminaries This section outlines the various techniques used in this paper for forecasting financial markets, which incorporates ARIMA, SVR, Facebook’s Prophet, RNN.

2.1 Auto-Regressive Integrated Moving Average (ARIMA) ARIMA is widely used for forecasting time series data in economics and finance. It can be seasonal and non-seasonal. The non-seasonal A R I M A(k, d, q) model can also be represented with three following parameters; k the number of auto-regressive terms, d the number of nonseasonal differences, and q the number of moving-average terms which is presented as Eq. 1 [12]: X t = φ1 X t−1 + φ2 X t−2 + φk X t−k + · · · + θ1 εt−1 + θ2 εt−2 + · · · + θq εt−q + · · · + εt .

(1) In this paper, non-seasonal A R I M A(k, d, q) [2] model is used.

2.2 Facebook’s PROPHET PROPHET is open-source time series data forecasting model for in Python, and R. PROPHET is made public through Facebook’s Core Data Science team. PROPHET is configured for commercial forecasting applications on Facebook. For efficient time series and seasonal handling, PROPHET architecture has its unique data frame. The data frame consists of two foundation columns. One of these columns is ds , and it stores a time series of dates in this column. The second column is y, and the corresponding values for the time series are stored in the data frame. The system, therefore, operates smoothly on a seasonal time series and offers some options for managing the dataset seasonality. Such opportunities are annual, weekly, and regular. As these options are presented, the data analyst may select the granularity of time available in the dataset for the forecast model [9].

2.3 Support Vector Regression (SVR) In this paper, Vapnik’s suggested standard SVR algorithm is used, which uses the s-insensitive loss function for predictive regression problems. The SVR algorithm refers to the transformation function O from the initial Input Space to the higher dimensional Space F feature to the original data points [1]. In this

98

B. Mondal et al.

new space, we are building a linear model Eq. 2 in the unique area which corresponds to a non-linear model as Eq. 3. N   y= (νi − νi∗ ) · xi , x + β

(2)

i=1

 y=

N  (νi − νi∗ ) · ϕ(xi ), ϕ(x) + β

(3)

i=1

Lagrange function is created, and once saddle point conditions have been applied, the solution is obtained as Eq. 4  y = f (x) =

l  (ν − νi∗ )K (xi , x) + β

(4)

i=1

Here, the two parameters are ν, and ν∗, and the term K (xi , x) represents the inner product between ϕ(X i ) and ϕ(x), known as the kernel function. In this paper, linear-, polynomial-, and radial-based functions (RBF) are used.

2.4 Recurrent Neural Network (RNN) RNN is widely used in time series modeling due to its unique, deep structure in the temporal dimension. If we unroll the RNN model over time, then its structure can be demonstrated in the RNN, the input value of the tth day xt = tth day xt = (xt1 , . . . , xtn ) is the n-vector indicating the characteristics. The following equations route the Eqs. 5 and 6 iterate the algorithm [5]: st = tanh(U xt + W st−1 + β)

(5)

ot = tanh(V st + c).

(6)

If st is the hidden state determine depended on it’s preceding value st−1 and the present input xt , Ot are the yield that could be considered predictable. U, V, W are respective input-to-hidden, hidden-to-output, and hidden-to-hidden weight matrices measured in the training process. Compositions β and c are unilateral.

A Comparative Study on Financial Market Forecasting Using AI …

2.4.1

99

Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU)

In this paper, LSTM and GRU are used which can be formulated as Eqs. 7 and 8: i t = σ (Ui xt + wi st−1 + bi )

(7)

Ct = tanh(Uc xt + Wc st−1 + bc ).

(8)

In this paper, the number of hidden layers and units are 3 and 128, respectively. In the intermediate layers, tanh is used as an activation function, and the sigmoid function is used in the final layer to produce predictive probabilities. In Python, the “Keras”; package is used to fit the deep learning models described above [4].

3 Methodology This study uses NIFTY50 data for the experimental application to analyze the effectiveness of our prediction approach as an instance. In this study, the stock prices used include the daily prices from November 2015 to November 2019. The training set consists of the first 80% of the sample, while the remaining 20% was used as the test set during training to avoid overfitting. All the data was acquired from https:// www1.nseindia.com/website.

3.1 Data Pre-processing The MinMaxScaler transforms the characteristics into a given range by scaling up each feature. This estimator scales and translates each element independently in such a way as to be between zero and one in a given training range. MinMaxScaler works better and is often used as an alternative to unit variance, zero mean, scaling in cases where the distribution is not Gaussian, or the standard deviation is small. The standardization is given as Eq. 9: X scaled =

(X − min(X )) × (max(X ) − min(X )) + min(X ), (max(X ) − min(X ))

where min(X) and max(X) are the ranges of feature (X).

(9)

100

B. Mondal et al.

4 Experimental Approach The data were standardized to improve the performance as comparators of LSTM, GRU, PROPHET, ARIMA, and SVR. The sample dataset was divided into two data files to start model building—check (20%) and training and cross-validation (80%). The withheld test dataset was used solely for testing the performance of the built classification models. This approach provides some insight into model success in real-world settings. Dataset training and test were used to develop models for each classifier type.

5 Result Analysis The implemented paradigm is assessed in hardware with Intel 6 Core 2.6GHz with 9 MB Cache, 16 GB RAM processor, and 6 GB NVIDIA GeForce RTX2060 graphics card. Python (v.3.5.2) and Tensorflow Keras package is used for implementation and result analysis. The “closing price” feature is used as the target value that we need to assess the efficiency of forecasts. The data was spat into two parts, that is, for training and testing. Performance metrics of the time series provide essential information on the predictive model capability expected to be made by the predictions. Figure 1 shows the projection and the original values to check the 12-day prediction. We first trained our system as a test scenario with the previous data values and used the performance evaluation metrics to test predictive accuracy by comparing real stock data with the expected values. We look at our model’s effectiveness forecast using the following statistical indicators: Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-Squared (R 2 ).

Fig. 1 True and predicted values of 12 days using a ARIMA, b Prophet, c linear SVR, d polynomial SVR, e GRU, f RBF SVR, g LSTM

A Comparative Study on Financial Market Forecasting Using AI …

101

5.1 Root Mean Square Error (RMSE) Residuals are the distance between the data points and regression line. RMSE shows how these residues are spread out. It tells you how the data is clustered around the best fit axis which is given as Eq. 10. RMSE is widely used for validation of experimental results throughout climatology, regression analysis, and forecasting. The calculated values are tabulated in Table 1.  (10) RSME = ( f − o)2 , where f is the predicted value and o is the actual value.

5.2 Mean Absolute Percentage Error (MAPE) MAPE is the statistical measure of forecasting system’s precision. This accuracy is calculated as a percentage and can be measured as the average absolute percent error for each time minus the actual values separated by the actual costs. If At is the real value and Ft is the forecast value, the following shall be given as Eq. 11. The calculated values are tabulated in Table 1. MAPE =

5.3

 n  1   At − Ft  . n t−1  At 

(11)

R2

The range is between 0 and 1. Coefficient of R 2 is similar to the coefficient of correlation given as Eq. 12. The calculated values are tabulated in Table 1.    (n( x y) − ( x) y)2   . R2 =   2    n x − ( x)2 n y 2 − ( y)2

(12)

This shows that the use of more robust non-linear techniques, such as deep learning, exploits the nonlinear characteristics of transaction data and can serve as a reliable financial index forecast for stock market investors. The result obtained was contrasted with each of the algorithms we used, and the Gated Recurrent Unit (GRU) was found to outperform the current models and there has been a significant increase in accuracy and performance recorded on the benchmark datasets in recent times.

102

B. Mondal et al.

Table 1 Calculated values of RSME, MAPE, R 2 Method RMSE MAPE LSTM GRU PROPHET ARIMA Linear SVR Poly SVR RBF SVR

173.56964313 81.72976432 307.24106215 631.49456772 146.39340918 149.63906385 133.96070348

0.96339823 0.59289791 2.32860629 5.21533862

R2 0.99542459 0.99661207 −18.29236635 −80.50161652 0.88317847 0.87794100 0.90217842

6 Conclusion Given the nonlinearity of the financial time series, making accurate stock market forecasts is a very challenging task. Technical analysts, however, insist that some of the current values may somehow predict future prices. This paper present a comparative study of forecasting accuracy among few AI techniques such as LSTM, GRU, PROPHET, ARIMA, and SVR. These techniques can be used to effectively forecast the stock market movement. The main challenge in stock market prediction systems is that time series-based algorithms cannot be identified using historical stock data as they are influenced by some variables, including government policy decisions, consumer expectations, and so on. Data from different sources are therefore required in decision-making, and pre-processing of data is a complex task for data mining. These are significant limitations that need to be addressed in the future by implementing advanced forecasting techniques on the stock market.

References 1. Awad, M., Khanna, R.: Support Vector Regression, pp. 67–80. Apress, Berkeley, CA (2015) 2. Ediger, V., Akar, S.: ARIMA forecasting of primary energy demand by fuel in Turkey. Energy Policy 35(3), 1701–1708 (2007) 3. Fama, E.F.: The behavior of stock-market prices. J. Bus. 38(1), 34–105 (1965). http://www. jstor.org/stable/2350752 4. Gers, F.A., Schmidhuber, J., Cummins, F.: Learning to forget: continual prediction with LSTM. Neural Comput. 12(10), 2451–2471 (2000) 5. Guresen, E., Kayakutlu, G., Daim, T.U.: Using artificial neural network models in stock market index prediction. Expert Syst. Appl. 38(8), 10389–10397 (2011) 6. Hiransha, M., Gopalakrishnan, E.A., Menon, V.K., Soman, K.P.: NSE stock market prediction using deep-learning models. Procedia Comput. Sci. 132, 1351–1362 (2018) (International Conference on Computational Intelligence and Data Science) 7. Malkiel, B.: A Random Walk Down Wall Street: The Time-Tested Strategy for Successful Investing, 9th edn. W. W. Norton (2007) 8. Mondal, B.: Artificial Intelligence: State of the Art, pp. 389–425. Springer International Publishing, Cham (2020)

A Comparative Study on Financial Market Forecasting Using AI …

103

9. Taylor, S.J., Letham, B.: Forecasting at scale. Am. Stat. 72(1), 37–45 (2018) 10. Wang, J.Z., Wang, J.J., Zhang, Z.G., Guo, S.P.: Forecasting stock indices with back propagation neural network. Expert Syst. Appl. 38(11), 14346–14355 (2011) 11. Wang, J.J., Wang, J.Z., Zhang, Z.G., Guo, S.P.: Stock index forecasting based on a hybrid model. Omega 40(6), 758–766 (2012) (Special Issue on Forecasting in Management Science) 12. Zhang, G.: Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing 50, 159–175 (2003) 13. Zhang, G., Patuwo, B.E., Hu, M.Y.: Forecasting with artificial neural networks: the state of the art. Int. J. Forecast. 14(1), 35–62 (1998)

Hydromagnetic Oscillatory Couette Flow of a Visco-Elastic Fluid with Dust in a Channel with Radiative Heat Hridi Ranjan Deb

Abstract A theoretical study of the oscillatory Couette flow of visco-elastic fluid with non-conducting dust is under consideration. The motion of fluid along with dust particles through a porous medium is under constant pressure gradient along with radiative heat transfer in a vertical channel. In this study, second-order fluid characterizes the non-Newtonian fluid and also considering dusty fluid model into account. A uniform magnetic field of strength B0 is applied perpendicular to the flow region. The left wall of the channel is fixed and right wall of the channel is oscillating with time. The governing equations are solved by applying perturbation method. The impact of visco-elastic parameter on velocity of fluid, dust particles and skin friction have been shown graphically along with the other non-dimensional parameter involved in this study. Keywords Oscillatory flow · Dusty fluid · Visco-elastic · Heat transfer · Similarity transformation · Skin friction

1 Introduction In Industry and technology we generally come across different types of polymers, paints, DNA suspension and others from chemical industry which do not follow linear relationship between rate of strain and shear stress these are termed as non-Newtonian fluid. There are different catagories of non-Newtonian fluid and visco-elastic fluid is one of them. A theoretical model was put forward by Coleman and Noll [2] for visco-elastic fluid and which is named as Second order fluid by assuming that the stress is more sensitive to recent deformation than to deformation which occurred in the distant past. σ = − p I + μ1 A1 + μ2 A2 + μ3 (A1 )2

(1)

H. R. Deb (B) Silchar Collegiate School, Silchar, Assam, India 788003 e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_11

105

106

H. R. Deb

where σ is the stress tensor, An (n = 1, 2) are the kinematic Rivlin-Ericksen tensors. The material coefficients viscosity (μ1 ), cross-viscosity (μ3 ) are taken as positive and visco-elasticity (μ2 ) as negative [1]. Saffman [3] has proposed a dusty fluid model to analyse the stability of laminar flow of dusty gas. Flow of visco-elastic fluid containing particulate suspension has attracted many scientist because of its application in the field of engineering scientist applied problems that are concerned with powder technology, aircraft icing, lunar ash flows. nuclear reactors and many others. Gupta and Gupta [4] considered the unsteady flow of a dusty visco-elastic fluid through channel with volume fraction. The effects of MHD convective flow of dusty viscous fluid with fraction in porous medium have been investigated by Ibrahim et al. [5]. Also, Makinde and Chinyoka [6] studied the MHD transient flows and heat transfer of dusty fluid in a channel with variable physical properties and Navier slip condition. MHD free convection flow of a visco-elastic (Kuvshiniskitype) dusty gas through a semi infinite plate moving with velocity decreasing exponentially with time and radiative heat transfer have been analysed by Prakash et al. [7], Singh and Singh [8] have considered the effects of MHD on Convective flow of Dusty viscous fluid with volume fraction. Sivraj and Kumar [9] studied the unsteady MHD dusty visco-elastic fluid couette flow in an irregular channel with varying mass diffusion. Prakash and Makinde [10] studied the MHD oscillatory couette flow of dusty fluid in a channel filled with a porous medium with radiative heat and buoyancy force. In the present analysis, an attempt has been made to extend the problem studied by Prakash and Makinde [10] to the non Newtonian fluid as put forward by Coleman and Noll [2] and Coleman and Markovitz [1] and is associated with dust as proposed by Saffman [3].

2 Mathematical Formulation In Cartesian system, the two dimensional unsteady Couette flow of a visco-elastic (second order fluid) through a porous medium along with dust particles in a vertical channel is considered. The channel walls are separated by a distance a has been considered. A uniform magnetic field of strength B0 is applied perpendicular to the flow region. In Cartesian system, x-axis is taken along the centre of the channel and distance in transverse direction is represented by y-axis. The equations of the vertical wall are considered as y = 0, which is fixed and y = a, oscillates with time t, where t > 0. For investigating the governing fluid motion the following assumptions are considered: (1) (2)

The dust particles are solid, uniform in size, spherical, do not conduct electrically. During the motion of the fluid, the dust particles are evenly spread and the number density N 0 is remains fixed.

Hydromagnetic Oscillatory Couette Flow of a Visco-Elastic Fluid …

(3) (4)

(5) (6)

107

Chemical reaction and the interrelationship between the particles have been neglected. The channel walls are impermeable coupled with parallel flow assumption the effects of convective term in the model momentum and energy equations have been neglected. For visco-elastic fluid the magnitude of fluid velocity is less so viscous dissipation is neglected. In the direction of x-axis, the vertical channel is assumed to be infinite in length so only pressure is the function of x.

Considering Boussinesq fluid model and the above assumptions, the governing equations are 1 ∂P ∂ 2u ∂ 3u v1 N0 K 0 ∂u =− (u p − u) − u+ + v1 2 + v2  2 ∂t ρ∂x K ρ ∂y ∂t∂ y −

σe B02 u + gβ(T − T0 ) ρ

(2.1)

∂u p = K 0 (u − u p ) ∂t

(2.2)

k ∂2T ∂T 1 ∂q = − ∂t ρc p ∂ y 2 ρc p ∂ y

(2.3)

The initial boundary conditions are: ⎫ u(y, 0) = 0, u(a, t) = U (1 + εeiωt ), u(0, t) = 0 ⎬ (2.4) u p (y, 0) = 0, u p (a, t) = U (1 + εeiωt ), u p (0, t) = 0, ⎭ T (y, 0) = T f , T (a, t) = Tw = T0 + (T f − T0 )(1 + eiωt ), T (0, t) = T0 , where u, up —velocities of fluid and dust particles in the x- direction, νi = μρi where i = 1, 2, t —time, ω—frequency of oscillation, T, T f —fluid temperature and the initial fluid temperature, T 0 , Tw—the left and right wall temperature, P—fluid pressure, g—acceleration due to gravity, q—radiative heat flux, σe —conductivity of the fluid, k—thermal conductivity, K—permeability porous medium, β—coefficient of volume expansion, K 0 —Stokes constant, ρ—fluid density, D—average radius of dust particles, cp —specific heat at constant pressure, μe is the magnetic permeability and H o is the intensity of magnetic field. The radiative heat flux is given by Cogley et al. [11], as fluid assumed to be of low density and optically thin ∂q = 4α 2 (T0 − T ) ∂t

(2.5)

108

H. R. Deb

where α is the constant of mean radiation absorption. We introduce the following non-dimensional variables: x=

up T − T0 x y u tU aP , y = ,u = ,θ = ,P = ,up = ,t = a a U T f − T0 a vρU U

Darcy number (Da) = aK2 , Particle mass parameter (M) = K 0νa 2 , Reynolds number (Re) = Uνa , K0 a2 Particle concentration parameter (l) = N0ρν , Visco-elastic parameter (d) = νaν2 U1 ,  2 2 vρc Prandtl number (Pr) = k p , Radiation parameter (N) = 4αk a , Grashof number (Gr) =

gβ(T −T0 )a 2 , vU

Hartmann number (H) =



a 2 σe B02 , vρ

Porous medium shape factor (s) =



1 . Da

where U is the velocity of the mean flow. The non-dimensionl form of governing equations are Re

∂P ∂ 2u ∂ 3u ∂u =− + 2 +d − (s 2 + H 2 + l)u + lu p + Gr θ ∂t ∂x ∂y ∂t∂ y 2 ∂u p = u − up ∂t

(2.7)

∂ 2θ ∂θ = 2 + N 2θ ∂t ∂y

(2.8)

ReM

Re Pr

(2.6)

The corresponding boundary conditions are u(y, 0) = 0, u(1, t) = 1 + εeiωt , u(0, t) = 0 u p (y, 0) = 0, = u p (1, t) = 1 + εeiωt , u p (0, t) = 0

(2.9)

θ (y, 0) = 1, θ (1, t) = 1 + εeiωt , θ (0, t) = 0,

3 Method of Solution For periodic flow, let − ∂∂ Px − λ = εeiωt , u(y, t) = u 0 (y) + εu 1 (y)eiωt , u p (y, t) = u p0 (y) + εu p1 (y)eiωt , θ (y, t) = θ0 (y) + εθ1 (y)eiωt where λ-constant steady flow pressure gradient.

 (3.1)

Hydromagnetic Oscillatory Couette Flow of a Visco-Elastic Fluid …

109

Substituting the values from Eq. (3.1) in to Eqs. (2.6)–(2.9), the following equations are obtained d 2u0 − a12 u 0 = −λ − Gr θ0 dy 2 a42

(3.2)

d 2u1 − a22 u 1 = −1 − Gr θ1 dy 2

(3.3)

u p0 = u 0

(3.4)

u1 1 + ReMiω

(3.5)

d 2 θ0 + N 2 θ0 = 0 dy 2

(3.6)

d 2 θ1 + a32 θ1 = 0 dy 2

(3.7)

u p1 =

The corresponding boundary conditions are  u 0 = u p0 = u 1 = u p1 = 0, θ0 = 0, θ1 = 0 at y = 0 u 0 = u p0 = u 1 = u p1 = 1, θ0 = 1, θ1 = 1 at y = 1 wher e a12 = s 2 + 2l + H 2 , a22 = s 2 + l + H 2 − a32 = N 2 − iωRe Pr, a42 = 1 + iωd

i 1+ReMiω

(3.8)

+ iωRe,

Skin Friction (ST-1) for the fluid at y = a is given by ST − 1 =

∂ 3u ∂ 2u + d ∂ y2 ∂t∂ y 2

(3.9)

and Skin Friction (ST-2) for the dusty particles at y = a is given by ST − 2 =

∂ 3u p ∂ 2u p + d ∂ y2 ∂t∂ y 2

(3.10)

The rate of heat transfer i,e, Nusselt number across the channel wall at y = a is given by Nu = N cot(N ) + a3 cot(a3 )εeiωt

(3.11)

110

H. R. Deb

4 Results and Discussion In this study, the main object is to investigate the effects of visco-elastic parameter in combination with other flow parameters involved in the study The effects of other parameters has been discussed by Prakash and Makinde [10]. The visco-elastic effect is exhibited through the non-zero values of dimensionless parameter d. The corresponding Newtonian fluid flow behaviour is obtained by setting d = 0. In order to understand the flow behavior for both fluid and dust particles the results are presented graphically. Figures 1, 2, 3 and 4, represents the flow behavior of velocity profile of fluid phase and dust phase versus displacement. x 10 5

2 1.8

d=0, Gr=8 d=-.03, Gr=8 d=0, Gr=4 d=-.03, Gr=4

1.6 1.4 1.2

u

1 0.8

0.6 0.4 0.2 0

0.1

0

0.2

0.3

0.4

0.5

y

0.6

0.7

0.8

0.9

1

Fig. 1 Velocity (u) versus y for Re = 0.2, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0.1, s = 0.1 x 10 5 d=0, Gr=8 d=-.03, Gr=8 d=0, Gr=4 d=-.03, Gr=4

up

2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0

0

0.1

0.2

0.3

0.4

0.5

y

0.6

0.7

0.8

0.9

1

Fig. 2 Velocity (up ) versus y for Re = 0.2, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0.1, s = 0.1

Hydromagnetic Oscillatory Couette Flow of a Visco-Elastic Fluid …

111

Figures 1 and 2 represents the velocity profile with the variation of Grashof number (Gr) and visco-elastic parameter (d). Grashof number (Gr) is the ratio of the buoyancy force to viscous force and also it studies the natural convection. So, rising in the Grashof number (Gr) also increases the effect of buoyancy force and reduces the effect of viscous force and velocity of fluid in Fig. 1 and velocity of dust particles in Fig. 2 also increases. Again, Reynolds number (Re) is the ratio of inertial force to the viscous force. So, increases in Reynolds number (Re) reduces the viscous drag and which leads to increase in velocity of fluid in Fig. 3 and velocity of dust particles in Fig. 4. Also it is observed from the Figs. 1, 2, 3 and 4, that the change in visco-elastic parameter from d = 0 to d = −0.03 decelerates the flow. So, the velocity of the Newtonian fluid is more in comparison to non-Newtonian fluid. x 104

18 16

d=-.03, Re=.2 d=-.03, Re=2 d=-.03, Re=.2 d=0, Re=2

14 12

u

10 8 6 4 2 0

0

0.1

0.2

0.3

0.4

0.5

y

0.6

0.7

0.8

0.9

1

Fig. 3 Velocity (u) versus y for Gr = 4, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0.1, s = 0.1

up

4

18 x 10 16 14 12 10 8 6 4 2 0 0.1 0

d=-.03, Re=.2 d=-.03, Re=2 d=0, Re=.2 d=0, Re=2

0.2

0.3

0.4

0.5

y

0.6

0.7

0.8

0.9

1

Fig. 4 Velocity (up ) versus y for Gr = 4, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0.1, s = 0.1

112

H. R. Deb

On the boundary surfaces, the impact of fluid flow are measured by skin friction. Figs. 5, 6, 7, 8, 9 and 10, depicts the skin friction with the variation of flow parameters involved in the problem. Here, two types of fluid flow is considered (i) for heated plate Gr < 0 (ii) for cooled plate Gr > 0. The effect of Hartmann number (H) on the skin friction for dusty fluid (ST-1) and dust particles(ST-2) are shown in Figs. 5 and 6. It is observed from both the figures that when Hartmann number intensifies skin friction for both dusty fluid (Fig. 5) and dust particles (Fig. 6) experienced by both heated plate and cooled plates gradually decreases. Also, it is observed from both the Figs. 5 and 6 that the magnitude of skin friction of visco-elastic fluid is more than the Newtonian fluid for both the cases (Gr < 0) and (Gr > 0). Figures 7 and 8, depicts skin friction of dusty fluid (ST-1) and dust particles (ST2) with the variation of particle mass parameter. It is illustrated from these figures that due to increase in particle mass parameter (M) skin friction for both dusty fluid (Fig. 7) and dust particles (Fig. 8) decreases gradually as we move away from the 6 0 x 10

d=0, d=-0.03, d=0, d=-0.03,

-1

Gr=4 Gr=4 Gr=-4 Gr=-4

ST-1

-2 -3 -4 -5 -6 3

3.2

3.4

3.6

3.8

4

H

4.2

4.4

4.6

4.8

5

Fig. 5 Shearing stress (ST-1) at right wall against for Re = 0.2, l = 0.1, t = 1, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0.1, s = 0.1 x 10 6

0

d=0, d=-0.03, d=0, d=-0.03,

-1

Gr=4 Gr=4 Gr=--4 Gr=-4

ST-2

-2 -3 -4 -5 -6

3

3.2

3.4

3.6

3.8

4

H

4.2

4.4

4.6

4.8

5

Fig. 6 Shearing stress (ST-2) at right wall against for Re = 0.2, l = 0.1, t = 1, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0.1, s = 0.1

Hydromagnetic Oscillatory Couette Flow of a Visco-Elastic Fluid …

113

x 10 6

5.4 5.35

ST-1

5.3

d=0, Gr=4 d=-0.03, Gr=4 d=0, Gr=-4 d=-0.03, Gr=-4

5.25 5.2 5.15 5.1 5.05 0

0.2

0.4

0.6

0.8

1

1.2

M

1.4

1.6

1.8

2

Fig. 7 Shearing stress (ST-1) at right wall against M for Re = 0.2, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, λ = 0.1, ω = 1, ε = 0.1, s = 0 x 10 6

5.4 5.3

d=0, Gr=4 d=-0.03, Gr=4 d=0, Gr=-4 d=-.03, Gr=-4

5.2

ST-2

5.1 5 4.9 4.8 4.7 4.6 4.5 4.4

0

0.2

0.4

0.6

1.2

1

0.8

M

1.4

1.6

1.8

2

Fig. 8 Shearing stress (ST-2) at right wall against M for Re = 0.2, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, λ = 0.1, ω = 1, ε = 0.1, s = 0 7.5

x 106 d=0, Gr=4 d=-0.03, Gr=4 d=0, Gr=-4 d=-0.03, Gr=-4

7

ST-1

6.5 6 5.5 5 0.1

0.2

0.3

0.4

0.5

s

0.6

0.7

0.8

0.9

1

Fig. 9 Shearing stress (ST-1) at right wall against s for Re = 0.2, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0.1

114

H. R. Deb 7.5

x 10

6

d=0, d=-0.03, d=0, d=-0.03

7

Gr=4 Gr=4 Gr=-4 Gr=-4

ST-2

6.5 6 5.5 5 0.1

0.2

0.3

0.4

0.5

s

0.6

0.7

0.8

0.9

1

Fig. 10 Shearing stress (ST-2) at right wall against s for Re = 0.2, l = 0.1, t = 1, H = 5, Pr = 6, N = 0.7, M = 0.1, λ = 0.1, ω = 1, ε = 0

plate. Also, skin friction in case of visco-elastic fluid is less than the Newtonian fluid for both (Gr < 0) and (Gr > 0). The influence of porous medium shape factor parameter (s) on the skin friction for dusty fluid (ST-1) and dust particles (ST-2) is observed in Figs. 9 and 10. It is noticed from these figures that due to increase in porous medium shape factor parameter (s) skin friction for both dusty fluid (Fig. 9) and dust particles (Fig. 10) gradually increases for both (Gr < 0) and (Gr > 0) as we move away from the plate. Also, it is observed from the Figs. 7, 8, 9 and 10, that skin friction in case of visco-elastic fluid is less than the Newtonian fluid for both (Gr < 0) and (Gr > 0). The expression (3.11) is independent of visco-elastic parameter (d). Therefore, it can be conclude that visco-elastic parameter (d) has no effect on the rate of heat transfer i,e, on Nusselt number. It is observed that obtained results are in this investigation are in good agreement as the results obtained by Prakash and Makinde [10].

5 Conclusion In this study, an analytical solution of hydromagnetic oscillatory couette flow of a visco-elastic fluid with dust through a porous medium in a channel with radiative heat has been considered. The key points of from the present investigation are highlighted below: 1. 2.

The velocity of visco-elastic fluid gets retarded with the rise in visco-elastic parameter. The skin friction of Newtonian fluid shows a rising trend than visco elastic fluid against s and M for dusty fluid (ST-1) and dust particles (ST-2).

Hydromagnetic Oscillatory Couette Flow of a Visco-Elastic Fluid …

3. 4.

115

In case of heated plate (Gr < 0) the shearing stress is dominant in nature as compared to cooled plate (Gr > 0) against s and M. The visco-elastic parameter has no effect on rate of heat transfer i,e, Nusselt number.

References 1. Coleman, B.D., Markovitz, H.: Incompressible second order fluid. Adv. Appl. Mech. 8, 69 (1964) 2. Coleman, B.D., Noll, W.: An approximation theorem for functional with applications in continuum mechanics, Archs ration. Mech Analysis. 6, 355 (1960) 3. Saffman, P.G.: On the stability of laminar flow of a dusty gas. J. Fluid Mech. 13(1), 120 (1962) 4. Gupta, R.K., Gupta, K.: Unsteady flow of a dusty visco-elastic fluid through channel with volume fraction. Indian J. Pure Appl. Math. 21, 677 (1990) 5. Ibrahim, S., Yusuf, M.W., Uwanta, I.J., Iguda, A.: MHD effects on convective flow of dusty viscous fluid with fraction in porous medium. Aust. J. Basic Appl. Sci. 4, 6094 (2010) 6. Makinde, O.D., Chinyoka, T.: MHD transient flowsand heat transfer of dusty fluid in a channel withvariable physical properties and Navier slip condition. Comput. Math. Appl. 60, 660 (2010) 7. Prakash, O., Kumar, D., Dwivedi, Y.K.: MHD free convection flow of a visco-elastic (Kuvshiniski type) dusty gas through a semi infinite plate moving with velocity decreasing exponentially with time and radiative heat transfer. AIP Adv. 1, 022132 (2011) 8. Singh, N.P., Singh, A.K.: MHD effects on convective flow of dusty viscous fluid with volume fraction. Bull. Inst. Math. Academia Sinica 30, 141 (2002) 9. Sivraj, R., Kumar, B.R.: Unsteady MHD dusty visco-elastic fluid Couette flow in an irregular channel with varying mass diffusion. Int. J. Heat Mass Transf. 55, 3076–3089 (2012) 10. Prakash, O., Makinde, D.: MHD oscillatory Couette flow of dusty fluid in a channel filled with a porous medium with radiative heat and buoyancy force. Latin Am. Appl. Res. 45, 185 (2015) 11. Cogley, A.C.L., Vinvent, W.C., Gilees, S.E.: Differential approximation for radiative transfer in a non grey gas near equilibrium. Am. Inst. Aeronaut. Astronaut. 6, 551 (1968)

Hard Exudates Detection: A Review Satya Bhushan Verma and Abhay Kumar Yadav

Abstract Hard exudates in the retina are the white or yellowish-white small deposits with sharp boundaries that appear as waxy, glistening or shiny surfaces. The hard exudates are located in the outer layer of retina and in deep to the retinal vessels. The damage on the retina of the eye, termed retinopathy may give us a clue to the vision injury or vision loss. Retinopathy is also evident in diabetic patients or in hypertension. This paper presents a complete review of some latest methods for detecting hard exudates in retinal images. This paper will help the researcher in studying the newest technique of various diabetic retinopathy and hypertensive retinopathy screening methodologies. The previously proposed techniques to detect hard exudates and retinal exudates are discussed in this present paper and the automated identification of hard exudates in hypertensive retinopathy is also discussed in the proposed paper. Keywords Hard exudates · Retinopathy · Retinal images

1 Introduction The retina is a thin tissue layer located at the backside of the eye that converts light to nerve signals and sends those nerve signals to the brain for understanding. The term retinopathy refers to the damage in the retina of the eye, which may lead to vision impairment or vision loss. Retinopathy is also seen in hypertension or diabetic patient [1]. Continuous high blood pressure makes the retina’s blood vessel walls thicken and narrow, and this places pressure on optic nerve and causes vision problems. Retinopathy can cause various lesions on the retina [2]. White or yellowish-white minor deposits with a sudden change are called hard exudates, this appears as shiny, waxy, or some sparkly. This is organized as individual S. B. Verma · A. K. Yadav (B) Babasaheb Bhimrao Ambedkar (Central University) University, Lucknow, Lucknow, India e-mail: [email protected] S. B. Verma e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_12

117

118

S. B. Verma and A. K. Yadav

Fig. 1 A human eye with diabetic retinopathy. a Hard exudates. b Soft exudates or CWS

points, sheets, confluent patches, or in rings or horseshoes shape close areas of retinal edema. Until now, no proper treatment for this pathology has been accepted, and unfortunately, even after being detected hard exudates remain unresolved with scant or no recovery for the patient. Hence early detection of exudates is considered to be a vital aspect in saving the patient vision. Computerized analysis of fundus image is in use for over two decades, and this computational process does not seem to slow down at all [3] (Fig. 1). The image processing plays an important role in any automated system [4]. Various types of preprocessing methods are used in any automated system [5]. It gives a noninvasive method for the detection of various retinal diseases such as hypertensive retinopathy, diabetic retinopathy, etc. The detection result will help out to take the fast decision for automatic referrals to ophthalmologists. Detection of contributing signs of a diseased retina from the fundus image helps in early diagnosis of the disease and necessary treatment can be carried on further thus reducing vision loss of the victim. Plenty of works have been proposed to separate exudates from the fundus background. First, image segmentation approaches for exudates detection have been based on thresholding, clustering, mixture modeling, region growing, morphology, classification techniques and many others. Usually, exudates detection approaches begin with preprocessing and removal of the optic disc. Any of the first detection methods provide a set of possible candidates belonging to the class of exudates. Finally, range or procedure of classification is applied based on the features computed from every sample results in only exudates. In this paper, our focus is on comparative analysis of detection of exudates and difference from other focused zones being present in the retinal image. Exudates are present as the common early clinical signs of diabetic retinopathy [6]. Hence, their detection would be considered an essential contribution to the screening tasks [7].

Hard Exudates Detection: A Review

119

Moreover, exudates detection could assist to take the first step in the direction of a whole observing and classifying the disease.

2 Literature Review Long et al. [8] proposed an algorithm for hard exudates detection. Initially, they make an automatic retinal image preprocessing method by using the active threshold technique and fuzzy C-means clustering (FCM) technique, then they used a classifier named support vector machine. The proposed algorithm contains four stages, first is preprocessing, second is the localization of the optic disc, the third stage is the determination of candidate hard exudates by using dynamic threshold in the grouping with global threshold, which is based on the FCM, and last stage is feature extraction. They used DIARETDB1 [9] and the e-ophtha EX retinal image database. On e-ophtha EX database, they achieved average sensitivity 76.5%, PPV 82.7%, and F-score 76.7%, and on the DIARETDB1 retinal database they achieved average sensitivity 97.5%, specificity 97.8%, and accuracy 97.7%. A fascinating perspective to the color retinal image processing to human recognition was given by Saeed et al. [10]. In their work, a completely automated system was presented. To make retinal images invariant of scale and rotation, they used SIFT in their method. Moreover, they used ICGF-based preprocessing method. Individual features of the retina were evaluated based on the existence of diabetic retinopathy. Their perspective was such that their results display that the exudates were detected from a database with specificity 96.1%, 96.9% sensitivity and the accuracy was 97.38%. Another interesting approach was of Marupally et al. [11], who put forward a semi-automated quantification technique for the purpose of hard exudates detection in photographs of color fundus that were diagnosed for diabetic retinopathy. They used color fundus photographs of 30 eyes were taken from 30 samples, 21 were from men and 9 women. They were able to develop a semi-automated algorithm to quantify the area covered by HEs. Two different methodologies were given by them (i) top-hat filtering with second-order statistical filtering and (ii) color fundus images thresholding. They were able to detect about 60–90% area under the Hard Exudates in 13 images and about 90–100% in other 17 images. Kaur et al. [12] managed to describe different works that were needed for the automatic identification of HEs and CWSs in retinal images for detection of diabetic retinopathy and images classification based on support vector machine (SVM). They evaluated their system on a big data records having 129 images of human retina. Their proposed method achieved sensitivity 96.9%, specificity 96.1% and accuracy 97.38% for the detected exudates from a database. A systematic approach was given by Narang et al. [13], which proposed an algorithm for hard exudates detection using a Lifting Wavelet Transform (LWT). This technique was based on image enhancement method and Support Vector Machine. They described various image processing techniques, which were expected to play an important part role in the diagnosis of

120

S. B. Verma and A. K. Yadav

retinal diseases due to diabetic retinopathy. A detailed comparative study with the Support Vector Machine classifiers gave good results. A training set of 40 data values and testing set of 20 data values are used. Training data gave accuracy for SVM of about 92.77% and the testing data gave the accuracy of 89.9%. Another quite interesting approach was given by Benzamin et al. [14], which implemented an algorithm based on using deep learning for hard exudates detection in given fundus images of the retina. Dataset downloaded from IDRiD was used in this method. They effectively detected Hard Exudates with an accuracy of 98.6% by using deep learning model. Rokade et al. [15] proposed an algorithm that consisted of seven different steps. They performed some easy and simple operations on the image (such as grayscale conversion or normalization of color). Moreover, some complex and advanced algorithms (such as decomposition and reconstruction of Haar wavelet). However, their approach was able to give a very low accuracy of about 22.48% only. This algorithm had a disadvantage of not being effective due to its very low accuracy Kavitha et al. [16] uses the thresholding approach based on histogram. Their technique considered histogram local minima for this operation. They ignored the small changes in the histogram and considered the last values to be the threshold value. This threshold value was applied to the image in order to detect exudates with the optic disk of the eye. Since the blood vessels converge at the optic disk. So, they were able to detect the converging point of the blood vessels using this method and the sharp bright area where this intersection point occurs were considered as the optic disk, and the other regions were declared as exudates. They were able to obtain an accuracy of about 89%, sensitivity of about 92.87% and the predictive value of 96.03% was achieved. Punnolil et al. [17] proposed an easy thresholding method that was used for detection of the hard exudates with application of an appropriate threshold level, the obtained pre-processed image was accompanied and had a 0.97 value threshold value on which segmentation was performed. The binary images were closed morphologically and were opened only for the removal of optic disk and the blood vessels. They were able to achieve sensitivity of about 96.89% and specificity of 97.15% is achieved, however obtained accuracy was not reported by them. Garcia et al. [18] proposed an automatic exudates detection technique in fundus images. They normalized each image and a combination of global and adaptive thresholding was proposed for the segmentation of the candidate EX. In order to improve the performance of the proposed algorithm, innovative post-processing of the image was done to eliminate the noisy regions. They were able to achieve mean sensitivity of nearly 92.1% and mean positive predictive value of 86.4% was acquired. Considering the image-based criterion, their system obtained a mean sensitivity of about 100%, mean specificity of about 70.4% and mean accuracy of 88.1% were recorded. Al Sariera et al. [19] proposed an exudates detection and classification method. For detection, masking of the optic disc (OD) is done of the presented retinal image and then the segmentation is done based on morphological reconstruction techniques and also on the thresholding value for the detection of the bright patches that would further lead to hard exudates. For the purpose of classification of different features like that

Hard Exudates Detection: A Review

121

of color, size, and texture are extracted from the obtained segmented subject regions dependent on these functions and the regions are determined by using multi-layered perceptron neural network (MLP). Their proposed model measured the performance depending on average sensitivity, specificity, and accuracy. They were able to obtain 86.9%, 94.8% and 93.0% for sensitivity, specificity, and accuracy respectively. Wang et al. [20] presented a mathematical approach that merges statistical classification method with local window-based verification strategy and brightness adjustment procedure. They were able to perform them based on the color information, which states lesions presence detection by utilizing MDD classifier based on statistical pattern recognition techniques. They were successful in obtaining accuracy of about 100% accuracy in detecting all lesions present in the retinal images. They managed to obtain 70% accuracy in correct classification of retinal images as normal, which were free from any lesions. Ruba et al. [21] proposed algorithms based on the SVM classifier for the effective identification and segmentation of exudates. Preprocessing was done initially to remove the noises. Gabor and GLCM algorithms were used for feature extraction from the images. Depending on the feature values, the extracted images were classified as normal or disease affected. Their given classifier presented 82% accuracy. For optical disk segmentation, the obtained results were 99.63% accuracy and for exudates, segmentation results were 99.35% accurate. Sopharak et al. [22] proposed the sets of different morphological operators that were used optimally adjusted for exudates detection on DR patients, which have non-dilated pupil with low-contrast images. The operator automatically detected exudates were validated by comparing with expert ophthalmologists’ hand-drawn ground truths. Validations were successful and the sensitivity of 80% and specificity of 99.5% were obtained for exudates detection. Nugroho et al. [23] focused separation of optic disc from exudates as exudates have similar characteristics with optic disc. In order of removing false positive in exudates, detection is done. Finding the small area of the optic disc that is enlarged to cover its total area was used for the removal of optic disc. A high pass filter is used in filtering the green channel that consists of all necessary information for detection of exudates. After that, proper segmentation of exudates is done by morphological and thresholding operations. The proposed approach in exudates detection was able to obtain accuracy of 99.99%, sensitivity of 90.15% and specificity of 99.99% respectively. Manoj Kumar et al. [24] proposed an efficient diagnostic model for identification of feature extraction in DR. The identification of the optic disk was done by combining the high disk intensity properties with the blood vessel convergence in a cost function. On considering the algorithm on different images of the database with different illumination characteristics, contrast and different DR stages were able to successfully detect rate 90% for hemorrhage. Afterward on many other features, the obtained has given out about 95% detection rate for microaneurysm, 95% detection rate sensitivity and 94% detection specificity for hard exudates identification and were able to achieve 97% success rate for the optic disk localization process. Imani et al. [12] proposed an exudates segmentation algorithm based on dynamic thresholding and morphological operators applied in retinal images, and they utilized

122

S. B. Verma and A. K. Yadav

Morphological Component Analysis (MCA) algorithm. They achieved specificity of 83.5% and sensitivity of 99.8%. Harangi et al. [25] used multiple active contours and they used region-wise classification, they were able to achieve sensitivity of 92%, specificity of 68% and sensitivity of 87%, specificity of 86% on different databases. Pereira et al. [17] use ant colony optimization approach and they achieved an accuracy of 97.8%; sensitivity of 80.8% and specificity of 99.1% Wisaeng et al. [26] proposed an algorithm to detect exudates by employing composite features based on morphological mean shift algorithm. Their algorithm begins with the pre-normalization of the given retinal image, followed by contrast enhancement, then noise removal, and followed the localization of the optical disk. Then, a complete segmentation using mean shift is done, which provides a different set of exudates and non-exudates candidates separately. In the end, a detailed classification based on the mathematical morphology algorithm (MMA) procedure was presented to contain only exudates pixels and eliminate other lesions. The optimal value parameters of the MMA increased the accuracy results by 13.10%. Their methods were able to achieve an average sensitivity of 98.4%, specificity of around 98.13, and accuracy of 98.35% in the process of detection of exudates.

3 Conclusion In the proposed paper, a detailed comparative study has been performed on the procedures of automatic detection of hard exudates algorithms. It highlights the significant improvement, achieved by the different algorithms given by different researchers. Based on this paper, we are able to conclude that a very little research has been done on the usage of evolutionary algorithms to detect the hard exudates. Usage of evolutionary algorithms will serve as a possible direction for future work in this field. This paper will help the researcher to understand the presently available procedure and methods for the detection of hard exudates and it will help them in pursuing their research forward.

References 1. Hard Exudates as a vascular disease. https://www.columbiaeye.org/education/digital-refere nce-of-ophthalmology/vitreous-retina/retinal-vascular-diseases/hard-exudates 2. Niemeijer, M., van Ginneken, B., Russell, S.R., Suttorp-Schulten, M.S.A., Abràmoff, M.: Automated detection and differentiation of drusen, exudates, and cotton-wool spots in digital color fundus photographs for diabetic retinopathy diagnosis. Invest. Ophthalmol. Vis. Sci. 48(5), 2260–2267 (2007). https://doi.org/10.1167/iovs.06-0996 3. Kavitha, M., Palanib, S.: Hierarchical classifier for soft and hard exudates detection of retinal fundus images. J. Intell. Fuzzy Syst. 27, 2511–2528 (2014). https://doi.org/10.3233/ifs-141224 (IOS Press) 4. Verma, S., Chandran, S.: Contactless palmprint verification system using 2-D Gabor filter and principal component analysis. Int. Arab J. Inf. Technol. 16(1) (2019)

Hard Exudates Detection: A Review

123

5. Chandran, S., Verma, S.B.: Touchless palmprint verification using shock filter SIFT I-RANSAC and LPD IOSR. J. Comput. Eng. 17(3), 2278–8727 (2015) 6. Dhiravidachelvi, E., Rajamani, V., Janakiraman, P.A.: Identification of hard exudates in retinal images. Biomed. Res. (2017) 7. Klein, R., Klein, B.E., Moss, S.E., Davis, M.D., DeMets, D.L.: The Wisconsin epidemiologic study of diabetic retinopathy VII. Diabetic nonproliferative retinal lesions. Ophthalmology 94, 1389–1400 (1987) 8. Long, S., Huang, X., Chen, Z., Pardhan, S., Zheng, D.: Automatic Detection of Hard Exudates in Colour Retinal Images Using Dynamic Threshold and SVM Classification: Algorithm Development and Evaluation. https://doi.org/10.1155/2019/3926930 9. Database DIARETDB1—Standard Diabetic Retinopathy Database for retinal images. https:// www.it.lut.fi/project/imageret/diaretdb1/index.html 10. Saeed, E., Szymkowski, M., Saeed, K., Mariak, Z.: An approach to automatic hard exudate detection in retina color images by a telemedicine system based on the d-eye sensor and image processing algorithms 11. Marupally, A.G., Vupparaboina, K.K., Peguda, H.K., et al.: Semi-automated quantification of hard exudates in colour fundus photographs diagnosed with diabetic retinopathy. BMC Ophthalmol. 17, 172 (2017). https://doi.org/10.1186/s12886-017-0563-7 12. Kaur, I., Kaur, N., Tanisha, Gurmeen, Deepi: Automated identification of hard exudates and cotton wool spots using biomedical image processing. Int. J. Comput. Sci. Technol. 7(4) (2016) 13. Narang, A., Narang, G., Singh, S.: Detection of hard exudates in colored retinal fundus images using the Support Vector Machine classifier. In: 2013 6th International Congress on Image and Signal Processing (CISP), Hangzhou, pp. 964–968 (2013). https://doi.org/10.1109/CISP.2013. 6745304 14. Benzamin, A., Chakraborty, C.: Detection of hard exudates in retinal fundus images using deep learning. In: 2018 Joint 7th International Conference on Informatics, Electronics & Vision (ICIEV) and 2018 2nd International Conference on Imaging, Vision & Pattern Recognition (icIVPR), Kitakyushu, Japan, pp. 465–469 (2018). https://doi.org/10.1109/iciev.2018.8641016 15. Rokade, P., Manza, R.: Automatic detection of hard exudates in retinal images using Haar wavelet transform. Int. J. Appl. Innov. Eng. Manag. 4, 402–410 (2015). ISSN 2319-4847 16. Kavitha, D., Shenbaga Devi, S.: Automatic detection of optic disc and exudates in retinal images. In: Proceedings of 2005 International Conference on Intelligent Sensing and Information Processing, pp. 501–506 (2005) 17. Punnolil, A.: A novel approach for diagnosis and severity grading of diabetic maculopathy. In: 2013 International Conference on Advances in Computing, Communications and Informatics (ICACCI), pp. 1230–1235, 22–25 Aug 2013 18. Garcia, M.: Detection of hard exudates in retinal images using a radial basis function classifier. Ann. Biomed. Eng. 37(7), 1448–1463 (2009). https://doi.org/10.1007/s10439-009-9707-0 19. Al Sariera, T.M., Rangarajan, L., Amarnath, R.: Detection and classification of hard exudates in retinal images. J. Intell. Fuzzy Syst. 38, 1943–1949 (2020). https://doi.org/10.3233/JIFS190492 (IOS Press) 20. Wang, H., Hsu, W., Goh, K.G., Lee, M.L.: An effective approach to detect lesions in color retinal images. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 181–186 (2000) 21. Ruba, T., Ramalakshmi, K.: Identification and segmentation of exudates using SVM classifier. In: IEEE International Conference on Innovations in Information Embedded and Communication Systems ICIIECS, pp. 1–6 (2015) 22. Sopharak, A., Uyyanonvarab, B., Barman, S., Williamson, T.H.: Automatic detection of diabetic retinopathy exudates from non- dilated retinal images using mathematical morphology methods. Comput. Med. Imaging Graph. 32, 720–727 (2008) (Elsevier) 23. Nugroho, H.A., Oktoeberza, K.Z.W., Adji, T.B., Bayu, S.M.: Segmentation of exudates based on high pass filtering in retinal fundus images. In: ICITEE IEEE, pp. 436–441 (2015) 24. Manoj Kumar, S.B., Manjunath, R., Sheshadri, H.S.: Feature extraction from the fundus images for the diagnosis of diabetic retinopathy. In: Emerging Research in Electronics, Computer Science, pp. 240–245. IEEE (2015)

124

S. B. Verma and A. K. Yadav

25. Harangi, B., Hajdu, A.: Automatic exudate detection by fusing multiple active contours and regionwise classification. Comput. Biol. Med. 54, 156–171 (2014) 26. Wisaeng, K., Sa-Ngiamviboo, W.: Exudates detection using morphology mean shift algorithm in retinal images. IEEE Access 7, 11946–11958 (2019). https://doi.org/10.1109/ACC ESS.2018.2890426

Analysis of Visco-elastic Fluid Flow Over an Inclined Cylinder: A Numerical Approach Ardhendu Sekhar Khound, Debasish Dey, and Rupjyoti Borah

Abstract In this problem, we have considered a steady boundary layer flow of visco-elastic fluid over an inclined porous circular cylinder. A uniform magnetic field (B0 ) is applied in the transverse direction to the flow. The dimensional governing equations are converted into non-dimensional equations by using some suitable nondimensional parameters. Further these non-dimensional equations are reduced to nonlinear ordinary differential equations which are solved by using Bvp4c technique. The results are discussed graphically for various values of flow parameters. Keywords Visco-elastic · Boundary layer flow · Porous medium · Circular cylinder · Radiation · Chemical reaction · Bvp4c method

Nomenclatures x: y: V: u: v: ue : g: B0 : a: K: k0 :

Coordinate in the direction of surface motion. Coordinate in the direction of surface motion, Velocity component of the fluid. Velocity component in x-direction. Velocity component in y-direction. velocity outside the boundary layer region. gravitational force. Uniform magnetic field. radius of the cylinder. Visco-elastic parameter. short-memory coefficient.

A. S. Khound (B) Department of Mathematics, Luit Valley Academy, Jorhat 785001, Assam, India e-mail: [email protected] D. Dey · R. Borah Department of Mathematics, Dibrugarh University, Dibrugarh 786004, Assam, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_13

125

126

M: N: Pr: Sc: h: kc : ρ: ν: α: β: β ∗: θ: η: ψ: ϕ: λ1 : λ2 : μ0 :

A. S. Khound et al.

Magnetic parameter. Radiation parameter. Prandtl number. Schmidt number. Chemical reaction parameter. Chemical reaction rate. Density of the fluid. Kinematic viscosity. Thermal diffusion. Coefficient of thermal expansion. Coefficient of concentration expansion. non-dimensional temperature. Porosity parameter. Stream function. non-dimensional concentration. mixed convection parameter for heat transfer. mixed convection parameter for heat transfer. Dynamic viscosity.

1 Introduction The fluids are divided into two types depending on their characteristics, i.e. Newtonian and non-Newtonian fluids. Newtonian fluid is a fluid which has linear relation between Stress and rate of strain. But, in case of non-Newtonian fluid, there is a nonlinear relation between stress and rate of strain where viscosity is dependent on shear rate or history of shear rate. Further, one of the category of non-Newtonian fluid is visco-elastic fluid (posses both of viscous and elastic property). Because of its uses in various industries such as polymer solution, suspension, paints, cosmetic products etc., many researchers and scientists have conducted their research to examine this fluid. Relaxation, retardation, Hall current effects etc. on visco-elastic fluid flow through various geometrical structures with entropy generation have been investigated by Dey and Khound [1–5]. Merkin [6] has investigated the free convective boundary layer on isothermal horizontal circular cylinder. Widodo and Ingham [7] have examined the free surface fluid flow in an arbitrary shaped in a channel. Mixed convection boundary layer flow from a horizontal circular cylinder in viscoelastic fluid has been investigated by Nazar et al. [8] and Anwar et al. [9]. Hsiao [10] has studied the mixed convection for viscoelastic fluid past a porous wedge. Widodo et al. [11] have done the mathematical modelling and numerical solution of iron corrosion problem based on condensation chemical properties. Ahmed et al. [12] and Ghosh and Shit [13] have discussed the MHD mixed convection flow of viscoelastic fluid and mass transfer from an infinite

Analysis of Visco-elastic Fluid Flow Over an Inclined Cylinder …

127

vertical porous plate with chemical reaction in presence of a heat source. Hussanan et al. [14] have examined the unsteady boundary layer MHD free convection flow in a porous medium with constant mass diffusion and Newtonian heating. Viscoelastic fluid flow with the presence of magnetic field past a porous circular cylinder has been investigated by Widodo et al. [15]. Dey [16] has studied the visco-elastic fluid flow through an annulus with relaxation, retardation effects and external heat source/sink.

2 Mathematical Formulations A steady mixed convective visco-elastic fluid flow over an inclined porous circular cylinder has been examined. The coordinate system is described in the following figure (Fig. 1). The flow is taken along y-axis which is the axis of the cylinder. A uniform magnetic field B = (0, 0, B0 ) is applied in the transverse direction to the flow which generates Lorentz force. We assume the magnetic Reynold number to be very small in order to neglect the induced magnetic field. The ambient fluid is moving with a uniform velocity. The temperature and concentration of the fluid on the surface of the cylinder is Tw and Cw respectively and that of ambient fluid is assumed to be T∞ and C∞ respectively. Using the above assumption and the boundary layer approximation [17] we formulate the governing equations for the problem as follows: Equation of continuity: ∂u  ∂v  +  =0  ∂x ∂y Momentum equation:

Fig. 1 Geometry of the problem

(1)

128

A. S. Khound et al.            ∂u e + v  ∂ 2 u  − 1 σ B 2 + υ u  + u  + gβ(T − T∞ )+ sin x  u  ∂∂ux  + v  ∂u = u ∗   e e 2 ρ a 0 k ∂y ∂x ∂y ⎛   ⎞ gβ (C − C∞ ) 3  3   2  ∂ u ∂ u ∂ u ∂u + v  3 −  ⎜ u ⎟  ∂ y 2 ∂ y ∂ x ∂ y ⎟ ∂ x ∂ y k0 ⎜ ⎜ ⎟  ⎟ ρ ⎜  2  ⎝ ∂u ∂ u ⎠ +  2 ∂x ∂y

(2)

Energy equation: u

 ∂T  ∂2T  ∂q   ∂T + v = α − ∂x ∂ y ∂ y 2 ∂ y

(3)

For a optically thin  ∞ heat transfer (Cogley et al. [18]) is given by,   fluid, the radiative = 4b2 T  − T∞ , wher e, b2 = 0 K λw dedTλ h dλ , K λw is absorption coefficient and eλ h is Plank’s function. The energy equation for species concentration together with first order chemical reaction (rate of reaction is proportional to concentration) ∂q  ∂ y

u

   ∂C  ∂ 2C   ∂C + v = D − kc C  − C ∞   2 ∂x ∂y ∂y

(4)

The corresponding boundary conditions are as follows: u  = v  = 0, T  = Tw , C  = Cw at y = 0, ∂u  u  = u e , = 0, T  = T∞ , C  = C∞ at y → ∞ ∂ y

3 Method of Solution The following non dimensional parameters are used into the above equations from (1) to (4) to make them non-dimensional: √ y √ u v x , y = Re , u = , v = Re , a a U∞ U∞   u x  T  − T∞ C  − C∞ υ ue = e ,θ = ϕ= , Pr = , U∞ Tw − T∞ Cw − C∞ α 4b2 a kc a υ U∞ a N= , Sc = , h = , Re = U∞ D U∞ υ x=

The dimensionless equations are obtained as follows:

Analysis of Visco-elastic Fluid Flow Over an Inclined Cylinder …

129

∂v ∂u + =0 ∂x ∂y

(5)

e u ∂∂ux + v ∂u = u e ∂u + v ∂ u − (M + η)(u + u e ) + [λ1 θ + λ2 φ] sin(x) ∂y ∂ x 3 ∂y 2  2   2  3 ∂ u K u ∂ ∂x∂uy 2 + v ∂∂ yu3 − ∂u + ∂∂ux ∂∂ yu2 ∂ y ∂ x∂ y

(6)

2

u

∂θ ∂θ 1 ∂ 2θ +v = − Nθ ∂x ∂y Pr ∂ y 2

(7)

u

∂φ 1 ∂ 2φ ∂φ +v = − hφ ∂x ∂y Sc ∂ y 2

(8)

The corresponding boundary conditions are as follows: u = v = 0, θ = 1, φ = 1 at y = 0 ; u = u e (x), ∂u = 0, θ = 0, φ = 0 at y → ∞ ∂y We use the following similarity transformation to solve the non-dimensional equations from (5) to (8) with respective boundary conditions ψ = x f (x, y), θ = θ (x, y), φ = φ(x, y), where ψ is the stream function defined by u = ∂ψ & v = − ∂ψ ∂y ∂y and substitute u e = sin(x) [6] we get:   ∂f sin x sin x cos x − (M + η) − + x ∂y x   2 2 3 4 ∂ f ∂f ∂ f sin x ∂ f −K 2 + [λ1 θ + λ2 φ] − f 4 − x ∂ y ∂ y3 ∂y ∂ y2  2  ∂ f ∂3 f ∂ f ∂4 f ∂ f ∂4 f ∂2 f ∂3 f − Kx − + − ∂ x∂ y ∂ y 3 ∂ x ∂ y4 ∂ y ∂ x∂ y 3 ∂ y 2 ∂ x∂ y 2   ∂ f ∂2 f ∂ f ∂2 f − = ∂ y ∂ x∂ y ∂ x ∂ y2   1 ∂ 2θ ∂ f ∂θ ∂ f ∂θ ∂θ − N θ = x − + f Pr ∂ y 2 ∂y ∂y ∂x ∂x ∂y   1 ∂ 2φ ∂ f ∂φ ∂ f ∂φ ∂φ − hφ = x − + f Sc ∂ y 2 ∂y ∂y ∂x ∂x ∂y

∂3 f ∂2 f + f − ∂ y3 ∂ y2



∂f ∂y

2

and the boundary conditions become f =

∂f ∂f sin x = 0, θ = 1, φ = 1 at y = 0; = , ∂y ∂y x

(9)

(10)

(11)

130

A. S. Khound et al.

∂2 f = 0, θ = 0, φ = 0 at y → ∞ ∂ y2 All the equations from (9) to (11) are reduced to highly non-linear ordinary differential equations at the lower stagnation point (x ≈ 0).   f  + f f  − f 2 + 1 − (M + η) f  − 1 + (λ1 θ + λ2 ϕ)   2  − K 2 f  f  − f f  − f  =0

(12)

1  θ + f θ − Nθ = 0 Pr

(13)

1  φ + f φ  − hφ = 0 Sc

(14)

and the corresponding boundary conditions are: f (0) = f  (0) = 0, θ (0) = 1, φ(0) = 1 ; f  (∞) = f  (∞) = 0, θ (∞) = 0, φ(∞) = 0

sin x , x

Using BVP4C technique, the above equations from (12) to (14) are solved.

4 Results and Discussion From Figs. 2, 3 and 4 we describe the velocity profile against the displacement variable for different values of non-dimensional parameters involved in the equations. Fluid motion is decelerated (Fig. 2) with the increase of magnetic parameter (M). Fig. 2 Velocity against y for Pr = 1, K = 1.5, η = 0.1, λ1 = 1, λ2 = 1, Sc = 0.7, h = 1.8, N = 1.5

M=0.5 M=0.8 M=1.5

1.2

f'(\y)

1 0.8 0.6 0.4 0.2 0

0

0.5

1

1.5

2

2.5

y

3

3.5

4

4.5

Analysis of Visco-elastic Fluid Flow Over an Inclined Cylinder …

131

1.4

Fig. 3 Velocity against y for Pr = 1, M = 2.5, η = 0.1, λ1 = 1, λ2 = 1, Sc = 0.7, h = 1.8, N = 1.5

K=2 K=4 K=6

1.2

f'(y)

1 0.8 0.6 0.4 0.2 0

0

0.5

1

1.5

2

2.5

y

3

3.5

4

8

Fig. 4 Velocity against y for M = 2.5, K = 1.5, η = 0.1, λ1 = 1, λ2 = 1, Sc = 0.7, h = 1.8, N = 1.5

4.5

5

Pr=1 Pr=1.2 Pr=2

7 6

f'(y)

5 4 3 2 1 0 -1

0

0.5

1

1.5

2

2.5

y

3

3.5

4

4.5

5

Physically, it may be interpreted like that the grouping of Lorentz force and viscosity makes the system thicker and as a result velocity decrease. Again, visco-elasticity (K) helps to reduce the velocity of the fluid motion (Fig. 3). The Prandtl number (Pr) and Schmidt number (Sc) have a negative impact on fluid velocity which is noticed in Figs. 4 and 5 respectively. It is due to the fact that increase in both Prandtl number (Pr) and Schmidt number (Sc) enhances the resistance of the fluid to deform during the application of forces. The temperature and concentration profiles for various values of the flow parameter are represented from Figs. 6, 7, 8 and 9. In Fig. 6, we have noticed that the temperature of the system increases with the increase of the Prandtl number (Pr). Physically it is possible due to the fact that the increase of Prandtl number reduces the thermal diffusivity of the system and as a result temperature rises. The Radiation parameter (N) has a negative impact on the temperature of the system (Fig. 7). It is due to the fact that when radiation parameter (N) increases the system release

132

A. S. Khound et al. 8

Fig. 5 Velocity against y for Pr = 1, M = 2.5, K = 1.5, η = 0.1, λ1 = 1, λ2 = 1, h = 1.8, N = 1.5

Sc=1 Sc=1.5 Sc=1.8

7 6 5

f'(y)

4 3 2 1 0 -1

0

0.5

1.5

1

2

2.5

y

3

3.5

4

4.5

1

Fig. 6 Temperature against y for M = 2.5, K = 1.5, η = 0.1, λ1 = 1, λ2 = 1, Sc = 0.7, h = 1.8, N = 1.5

5

Pr=1 Pr=3 Pr=5

θ (y)

0.5

0

-0.5

0

0.5

1

1.5

2

2.5

y

3

3.5

4

4.5

5

1

Fig. 7 Temperature against y for Pr = 1, M = 2.5, K = 1.5, η = 0.1, λ1 = 1,λ2 = 1, Sc = 0.7, h = 1.8

0.9

N=2 N=5 N=7

0.8 0.7

θ (y)

0.6 0.5 0.4 0.3 0.2 0.1 0

0

0.5

1

1.5

2

2.5

y

3

3.5

4

4.5

5

Analysis of Visco-elastic Fluid Flow Over an Inclined Cylinder …

133

1

Fig. 8 Concentration against y Pr = 1, M = 2.5, K = 1.5, η = 0.1, λ1 = 1,λ2 = 1, h = 1.8, N = 1.5

Sc=1 Sc=1.2 Sc=1.25

0.9 0.8

φ (y)

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Fig. 9 Contration against y Pr = 1, M = 2.5, K = 1.5, η = 0.1, λ1 = 1,λ2 = 1, Sc = 0.7, N = 1.5

0

0.5

1

1.5

2

2.5

y

3

3.5

4

1

4.5

5

h=1 h=3 h=5

0.9 0.8 0.7

φ (y)

0.6 0.5 0.4 0.3 0.2 0.1 0

0

0.5

1

1.5

2

2.5

y

3

3.5

4

4.5

5

heat energy rapidly as a result temperature of the system increases. The concentration (deposition) is enhanced (Fig. 8) with the increase of the Schmidt number (Sc). Physical reasoning behind this is that the mass diffusion rate decreases with the increase of Schmidt number (Sc). When the chemical reaction parameter (h) increases the reaction rate of the system is also increased. As a result the concentration of the system decreases (Fig. 9).

134

A. S. Khound et al.

5 Conclusions Some of the important results are given below: i. ii. iii. iv. v.

The fluid velocity can be controlled by increasing the magnetic parameter and visco-elastic parameter. Prandtl number and Schmidt number have a negative impact on fluid flow. We may reduce the temperature of the system by increasing the radiation parameter and decreasing the Prandtl number respectively. Schmidt number has a positive impact on the concentration of the system. The concentration of the system may be controlled by increasing the cahmical reaction parameter.

References 1. Dey, D., Khound, A.S.: Relaxation and retardation effects on free convective visco-elastic fluid flow past an oscillating plate. Int. J. Comput. Appl. 144(9), 34–40 (2016) 2. Dey, D., Khound, A.S.: Hall current effects on binary mixture flow of Oldroyd-B fluid through a porous channel. Int. J. Heat Technol. 34(4), 687–693 (2016) 3. Dey, D., Khound, A.S.: Analysis of thin film flow of Oldroyd-B nanofluid in an oscillating inclined belt with convective boundary conditions. Int. J. Math. Arch. 9(7), 142–150 (2018) 4. Dey, D., Khound, A.S.: Entropy generation of Oldroyd-B fluid through various geometrical structures. Int. J. Res. Advent Technol. 6(9), 2358–2365 (2018) 5. Dey, D., Khound, A.S.: Free convective Oldroyd fluid flow through an annulus under transverse magnetic field using modified bessel functions. Int. J. Heat Technol. 37(1), 41–47 (2019) 6. Merkin, J.H.: “Free Convective Boundary Layer on Isothermal Horizontal Circular Cylinder”, ASME/AIChE, heat transfer conference. St, Louis (1976) 7. Widodo, B., Wen, X., Ingham, D.B.: The free surface fluid flow in an arbitrary shaped in a channel. J. Eng. Anal. Bound. Elem., 299–308 (1997) 8. Nazar, R., Amin, N., Pop, I.: Mixed convection boundary layer flow from a horizontal circular cylinder in micropolar fluids: case of constant wall temperature. Int. J. Numer. Methods Heat Fluid Flow, 86–109 (2003) 9. Anwar, I., Amin, N., Pop, I.: Mixed convection boundary layer flow of a viscoelastic fluid over a horizontal circular cylinder. J. Non-Linear Mech., 814–821 (2008) 10. Hsiao, K.L.: MHD Mixed convection for viscoelastic fluid past a porous wedge. J. Non-Linear Mech., 1–8 (2010). Elsevier 11. Widodo, B., Fatahillah, A., Rahayuningsih, T.: Mathematical modelling and numerical solution of iron corrosion problem based on condensation chemical properties. Aust. J. Basic Appl. Sci. 5(1), 79–86 (2011) 12. Ahmed, N., Sharma, D., Deka, H.: MHD mixed convection and mass transfer from an infinite vertical porous plate with chemical reaction in presence of a heat source. Appl. Mech. Sci., 1011–1020 (2012) 13. Ghosh, S.K., Shit, G.C.: Mixed convection MHD flow of viscoelastic fluid in porous medium past a hot vertical plate. J. Mech., 262–271 (2012) 14. Hussanan, A., Ismail, Z., Khan, I., Hussein, A., Shafie, S.: Unsteady boundary layer MHD free convection flow in a porous medium with constant mass diffusion and Newtonian heating”, J. Eur. Phys., 1–16 (2014) 15. Widodo, B., Siswono, G.O., Imron, C.: Viscoelastic fluid flow with the presence of magnetic field past a porous circular cylinder. Int. J. Mech. Prod. Eng. 3(8), 123–126 (2015)

Analysis of Visco-elastic Fluid Flow Over an Inclined Cylinder …

135

16. Dey, D.: Viscoelastic fluid flow through an annulus with relaxation, retardation effects and external heat source/sink. Alex. Eng. J., 1–7 (2017) 17. Zaib, A., Kasim, A.R.M., Mohammad, N.F., Shafie, S.: Unsteady MHD mixed convection stagnation point flow in micropolar fluid on vertical surface in a porous medium with Soret and Dufour effects. Heat Trans. Res., 603–620 (2013) 18. Cogley, A.C.L., Vincenti, W.G., Giles, E.S.: Differential approximation for radiative heat transfer in Non-linear equations grey gas near equilibrium. Am. Inst. Aeronaut. Astronaut. 6, 551–553 (1968)

Convolution Neural Network-Driven Computer Vision System for Identification of Metanil Yellow Adulteration in Turmeric Powder Dipankar Mandal, Arpitam Chatterjee , and Bipan Tudu

Abstract Identification of adulteration in food products is a challenging task. This paper presents a deep convolutional neural network (CNN)-based classification model for classification between adulterated and unadulterated turmeric powder. Metanil yellow is used as adulterant in this work. The consolidated results show that the presented CNN model can provide up to 98.5% classification accuracy. Therefore, the presented computer vision technique can be a possible alternative approach to adulterant identification in turmeric powder. Keywords Convolution neural network · Classification · Adulterant detection

1 Introduction Different spices are commonly used in India and other Asian countries. Turmeric is one of the most commonly used spices, which not only increase food value and taste of the foods but also contribute toward healing of different diseases due to its medicinal values. It has been found helpful for healing diseases related to digestion, respiration, burn and wound recovery, etc. Turmeric belongs to ginger family and commonly used in powdered form. The yellow color of turmeric is resulted from the presence of polyphenol component, namely curcumin [1]. Despite the healing power, the powdered turmeric is often adulterated with artificial yellow color to attract customers and reducing the price. Such adulteration gives D. Mandal Department of Applied Electronics & Instrumentation Engineering, Future Institute of Engineering & Management, Kolkata, India e-mail: [email protected] A. Chatterjee (B) Department of Printing Engineering, Jadavpur University, Kolkata, India e-mail: [email protected]; [email protected] B. Tudu Department of Instrumentation & Electronics Engineering, Jadavpur University, Kolkata, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_14

137

138

D. Mandal et al.

a better appearance compromising the quality. Prolong consumption of adulterated food ingredients can cause different health hazards even to the extent of hormonal disorders and cancers. Identification of adulterants in food and food ingredients therefore is a challenging field of research. Metanil yellow is one of the commonly used adulterants, which is added to enhance the yellow appearance of ground turmeric [1]. Currently, the adulterant detection process is mostly chemical and instrumental in nature. For example, gas chromatography (GC), liquid chromatography (LC), Terahertz spectroscopy and FT-NIR spectroscopy are some such processes [2]. These processes can detect the adulterant very accurately but are expensive, time taking and invasive in nature. They are also limited in terms of mobility of test procedure. Computer vision can be a possible alternative to such stationary mechanisms, which is also less expensive and can be made portable with the latest technical upgrades. Computer vision is becoming a popular tool for the food engineering domain where it has been applied to the detection of defects, identification and sorting of grades and classification of species qualitatively, however, its application to adulterant detection in food is not that popular yet. In the presented the deep convolutional neural network (CNN) [3] has been employed for classification between with and without adulterant turmeric powder. CNN is powered by several advantages over conventional statistical and parametric classifiers like logistic regression, K-nearest neighbor (KNN), support vector machine (SVM), naïve Bayes classifier, etc. [4]. One of the major advantages is feature extraction by convolution layers. However, CNN needs considerable time-consuming in its training stage, but once the model is trained, it has shown a better classification rate with lower degree of overfitting. The results of the presented work show that the CNN-based classification can achieve up to 98.5% accuracy in adulterant detection. The result has been compared with the classification result obtained with two other classifiers, namely, support vector machine (SVM) and naïve Bayes classifier. In those classifiers, the feature set was made by color channel statistics at different color modes namely, RGB, HSV and Lab followed by dimension reduction using principal component analysis (PCA). The comparative results clearly show that the simple CNN architecture can noticeably outperform the other two classifiers. The presented technique can be considered as an initial step toward the development of a portable detection instrument, which can be used portably across different food chains and it can be even developed as an appbased system due to the recent advent of mobile device-based technologies. Based on the performance, the presented method is considered as potential alternative to conventional testing mechanisms.

Convolution Neural Network-Driven Computer Vision System …

139

2 Experimentation 2.1 Sample Preparation The experiments were made by turmeric powder prepared in-house. Washed turmeric roots were dried and blended to form the powder as described in [2]. The Metanil yellow, which is used as adulterant in this work, was procured from market. It was added to the prepared turmeric powder in two different percentages by weight, i.e. 5 and 20% to produce adulterated turmeric powder samples.

2.2 Image Acquisition An in-house imaging chamber as shown in Fig. 1 was made with a wooden box and LED light panels to keep the illumination modulation under control. The dimension of the chamber is 0.5ft × 0.5ft × 1.5ft and the internal walls were covered with translucent papers to have controlled reflection from the walls. It is important to note that CNN can handle illumination variation but considering our application illumination has been stabilized. A 15MP camera was fitted to the chamber that interfaced to the computer using USB 2.0 connectors. The prepared with and without

Fig. 1 The imaging chamber (top) and the example of captured image (bottom)

140

D. Mandal et al.

adulterated samples were placed in a small flat container and distributed uniformly. Total 1000 images comprising 500 unadulterated and 500 adulterated (250 of each 5% and 20% Metanil yellow) were captured. Some samples of the captured pictures are shown in Fig. 1.

2.3 CNN Model The CNN models comprise of convolution layer, pooling and flattening layers. The number of layers engaged in each category depends on the criticality of the applications. There are additional types of layers as well, e.g. batch normalization and dropout, which are helpful to avoid over-fitting and under-fitting problems. In this work, the sequential CNN model is engaged using TensorFlow backend [5] and Keras [6] libraries. The structure is given in Fig. 2 and the parameters are shown in Table 1.

Image preprocessing

Image Database

Convoluti on layer (Conv_2d) 32 filters of 3x3

Max_poolin g layer Pool size 2x2

Flattening layer

Dense layer 1 ReLU activation

Dense layer 2 sigmoid activation

Decision

Fig. 2 The CNN model for classification between adulterated and pure turmeric powder

Table 1 The CNN model layers and parameters Layer (type)

Number of filters

Conv2d_1

32

Max_pooling2d_1

Filter dimension

Output shape

3×3

148 × 148 × 32

2×2

74 × 74 × 32

Flatten_1

175232

Dense_1

128

Dense_2

1

Number of parameters 896 0 0 22,429,824 129

Convolution Neural Network-Driven Computer Vision System …

141

Fig. 3 Processed images; (top row) unadulterated/pure and (bottom row) adulterated turmeric powder

Step 1: Image preprocessing. This includes two major steps; image resampling and image augmentation. The resampling includes two steps; discarding the background and cropping image in a specified pixel dimension of 600 × 600 × 3. Some examples of preprocessed images are shown in Fig. 3. It is to be noted that CNN can potentially handle images with backgrounds but that undoubtedly increases the processing time and complexity of the network. In our case, the resampling was done to make the detection much faster with a low complexity and lower degree of connectivity in the model. On the other hand, the augmentation was made to generate more images with the existing images. It is important for our case as we had only 500 test samples of each class. This augmentation can be easily achieved without much coding and processing necessities using excellent image preprocessing libraries of Keras. The library helps to augment data during each epoch by applying different transforms like rotation, rescaling, etc. on the existing training image data. This library also helps to deal with the labels, which eliminate the tedious labeling task as well. This augmented set was fed to the CNN with further resizing at 150 × 150 × 3. Step 2—Convolution layer (conv_2d). Convolution layer is a powerful mechanism to extract convolved features from the images. The kernel dimension in our case is 3 × 3 with a stride of (1 × 1), which resulted in output shape of convolution layer as 148 × 148 × 3. This layer performs a dot product between the kernel coefficients and the input image followed by a activation map for the specified number of filters, which is 32 in our case. Through these filters, the CNN model learns the activation pattern for the filters with a particular type of image. The activation function of the layer is RelU, which is nonlinear function expressed as Eq. 1, which shows that the output of RelU function is the same as input z value where z is positive else it is 0. R(z) = max(0, z)

(1)

142

D. Mandal et al.

Step 3—Pooling layer (max_pooling2d). This layer pools out the specified statistical values from the output of conv_2d layer, which further reduces the matrix dimension depending on the filter size specified in the pooling layer. As in our case, the matrix size is 2 × 2 means the output of the pooling layer will be half of the conv_2d layer row and column dimension of the conv_2d layer output. However, the third dimension remains same as 3 since that is for the color channel (R, G and B). The statistics in our case is max value in the 2 × 2 coverage of every pixel, which actually gives the name max_pooling. Step 4—Flattening layer (flatten_1). This is layer is a simple one, which is used to arrange the output data of pooling layer in a column vector format to make it suitable for feeding into the following artificial neural network (ANN) architecture. Step 5—Dense layer 1 (Dense_1). This is the first fully connected layer of our architecture. This layer consists of neurons/nodes, which take the flattened output as input and pass the same to its connected next layer (which is Dense_2 or output layer in our case) in a weighted manner to minimize the loss function at the output. In our case, 128 nodes at this layer are found to be optimal on a trial and error basis. The loss function is binary_crossentropy in our case since we have a binary classification task and it is expressed as Eq. 2. H p (q) = −

N 1  yi · log( p(yi )) + (1 − yi ) · log(1 − p(yi )) N i=1

(2)

where, y is the predicted class for i-th sample and p(yi ) represents the probability of ith sample of being 1. The weight adjustment is achieved by the adaptive moment optimization (Adam) algorithm, which combines features of momentum and RMSProp. The weight update equation in this algorithm can be represented as Eqs. (3)–(6). j

j

j

wt+1 = wt + wt

(3)

vt j wt = −η √ · gt st + ε

(4)

st = β2 · st−1 − (1 − β2 ) · gt2

(5)

vt = β2 · vt−1 − (1 − β1 ) · gt

(6)

where, β1 and β2 are hyperparameters with value commonly set as 0.9 and 0.99, respectively. η is the initial learning rate, vt and st are the exponential average gradient and average of squared gradient in the direction of w j . gt is the gradient at tth time along w j . j is the index of connection weight between dense layers. Step 6—Dense layer 2 (Dense_2). This is the final output layer consisting of a single node, which can give either class 1 or 0 depending on the probability resulted in the

Convolution Neural Network-Driven Computer Vision System …

143

output. The activation function in this layer is sigmoid, which provides nonlinear mapping of data.

3 Results and Discussion The CNN model was realized using Spyder Anaconda editor in the python platform. The Keras libraries were used to realize the sequential CNN model. The entire dataset comprising images of pure (unadulterated) and adulterated turmeric powder samples were divided into training and testing set with a 80:20 splitting. This split again was smoothly achieved using Keras image preprocessing libraries. Since we had a small number of images (total 1000) for the experiments, the validation set was not considered. Also the preprocessing of the images especially the background removal also helped to achieve the training goal in a much smaller number of epochs. It has been found that only 10 epochs are enough for our work, which indeed reduced the training time to a great extent. Some of the testing results as appeared in the realized implementation are shown in Fig. 4. The model performance metrics and confusion matrix are presented in Tables 2 and 3, respectively. Table 4 portrays a comparison of the presented method and two established classifiers, namely, SVM and naïve Bayes classifiers. Tables 3 and 4 show the potential of the presented CNN model toward the classification of adulterated and unadulterated turmeric powder samples. As the tables show, up to 98.5% overall accuracy has been achieved, which is visibly higher than two other classifiers under comparison. Similarly, the high precision and recall rate confirm the model efficiency toward the classification task. Finally, the FMeasure confirms the high accuracy rate of the presented CNN model. The high FMeasure value also confirms a good balance between recall and precision, which confirms the model potential in terms of sensitivity and specificity [7].

4 Conclusion The work presented a CNN-based classification between adulterated and unadulterated turmeric powder where Metanil yellow has been used as adulterant. The preprocessed images of both classes were used for experiments and a CNN model with a single convolution and pooling layer was engaged. The results as evaluated using confusion metric measures are found to be potential and can give up to 98.50% accuracy with considerable precision and recall. This can outperform some of the established classifiers also. The work is however in preliminary stage, which can be further extended to multiclass classification, improvement of the performance in terms of accuracy and a prediction model, which can predict the possible amount of

144

D. Mandal et al.

Fig. 4 Examples of CNN output window (adult stands for adulterated and original stands for unadulterated samples) Table 2 Performance metrics Accuracy = (TP + TN)/Total population Precision = TP/(TP + FP) Recall = TP/(TP + FN)

FMeasure = (2 * Precision * Recall)/(Precision + Recall)

(TP—true positive, TN—true negative, FP—false positive and FN—false negative) Table 3 Confusion matrix

Pure Pure adulterated

Adulterated

99

1

2

98

98.01% Precision 99% Recall 98.5% Accuracy FMeasure = 98.50

Convolution Neural Network-Driven Computer Vision System …

145

Table 4 Comparison between CNN, KNN and SVM Classifier

Precision (%)

Recall (%)

Accuracy (%)

FMeasure

Naïve Bayes

74.75

88.30

82.25

80.96

SVM

90.59

86.72

88.25

88.61

CNN

99.00

98.01

98.50

98.50

adulterant mixed with the turmeric powder. This can be further developed as a handheld portable device for on-spot detection of adulteration and even can be extended toward development of mobile app, which can be used by the end users.

References 1. Dhakal, S., Chao, K., Scimidt, W.: Evaluation of turmeric powder adulterated with Metanil yellow using FT-Raman and FT-IR spectroscopy. J. Foods. 5, 36 (2016) 2. Kar, S., et al.: FTNIR spectroscopy coupled with multivariate analysis for detection of starch adulteration in turmeric powder. Food Addit. Contam. Part A (2019). https://doi.org/10.1080/ 19440049.2019.1600746 3. Mathieu, M., Henaff, M., LeCun, Y. Fast training of convolutional networks through ffts. In: Proceedings of the International Conference on Learning Representations (ICLR) (2014) 4. Shai, B.-D., Shai, S.-S., Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press (2014) 5. Martín, A., et. al.: TensorFlow: Large-scale machine learning on heterogeneous systems (2015). Software available from tensorflow.org 6. Chollet, F.: Keras. https://github.com/fchollet/keras (2015) 7. Buckland, M.K., Gay, F.: Confusion matrix visualization, Intelligent information processing and web mining. In: Advances in Soft Computing, 25. Springer, Berlin, Heidelberg (2004)

Identification of Biologically Relevant Biclusters from Gene Expression Dataset of Duchenne Muscular Dystrophy (DMD) Disease Using Elephant Swarm Water Search Algorithm Joy Adhikary and Sriyankar Acharyya Abstract A gene expression dataset contains expressions of different genes at different conditions of a disease. The biclustering procedure explores only those subsets of genes, which have similarities in expression behaviors across some subsets of conditions. These subsets of genes and conditions form sub-matrices, known as biclusters. Here, for the first time, the biclustering approach is applied to Duchenne Muscular Dystrophy (DMD) disease dataset. This paper presents a meta-heuristic-based method for identifying biclusters from the dataset related to Duchenne Muscular Dystrophy (DMD) disease. It identifies shifting and scaling pattern-based biclusters considering different objectives together. For this purpose, Elephant Swarm Water Search Algorithm (ESWSA) and a proposed variant of ESWSA, named Chaotic Move Elephant Swarm Water Search Algorithm (CMESWSA) have been implemented. The proposed method (CM-ESWSA) has been able to recognize the shifting and scaling pattern-based biclusters of better quality. To determine the efficiency of ESWSA and CM-ESWSA, statistical testing and benchmark analysis have also been done and it shows that the proposed method outperforms the basic ESWSA. Keywords Biclustering · Elephant Swarm Water Search Algorithm

1 Introduction DNA microarray technologies are used to analyze the activity level of many genes under different conditions in an efficient manner. The data generated from this technology is a matrix where rows represent genes and columns represent conditions [1, 2]. Genes are situated in the chromosomes of every cell in an organism. Each J. Adhikary · S. Acharyya (B) Department of Computer Science and Engineering, Maulana Abul Kalam Azad University of Technology, BF-142, Sector-1, Salt Lake, Kolkata, West Bengal, India e-mail: [email protected] J. Adhikary e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_15

147

148

J. Adhikary and S. Acharyya

gene produces protein in its expression, which controls different activities of the concerned organism. Due to some external effects, protein production rate of a gene varies across different physiological stages or conditions [3]. In gene expression analysis, biclustering is a standard local clustering approach, which finds subsets of genes having similarity in expression behavior across subsets of conditions in a gene expression matrix. These statistically significant sub-matrices contain coexpressed genes and there is a strong possibility of the existing biological relationship among these coexpressed genes [2, 4, 5]. At first, Cheng and Church (CC) [6] proposed a model for identifying significant biclusters using greedy search heuristics. The metric Mean Squared Residue (MSR), which measures the quality of biclusters is used to identify biclusters having shifting pattern [6], but MSR is not capable to identify the biclusters with searching pattern. Mukhopadhyay et al. [7] have introduced another metric, called Scaling Mean Squared Residue (SMSR) to identify shifting and scaling pattern biclusters separately. Huang et al. [4] proposed a meta-heuristic approach, named Condition Based Evolutionary Biclustering (CBEB) based on Evolutionary Algorithm (EA) using MSR metric and a predefined threshold. Pontes et al. [2] introduced Evolutionary Biclustering based on Expression Patterns (Evo-Bexpa) [2]. They used a cost function incorporating four different objectives (quality, volume, overlapping amount and gene variance). This cost function is widely used to recognize shifting and scaling pattern of bicluster simultaneously. Duchenne Muscular Dystrophy (DMD) disease causes progressive weakness and decay in muscles resulting in premature death. Decrease in protein, called dystrophin, causes this disease and it happens because of mutation in the gene (DMD), which codes particular protein [3]. Biclustering is a NP-hard problem [4, 6] and meta-heuristic methods are suitable to find a near-optimal solution to this type of problem in reasonable time. This work has applied a population-based meta-heuristic namely, Elephant Swarm Water Search Algorithm (ESWSA) [8] and proposed a new variant of ESWSA, namely, Chaotic Move Elephant Swarm Water Search Algorithm (CM-ESWSA). In the proposed variant, the chaotic behavior is used in search for providing better exploration to avoid pre-mature convergence. The proposed variant (CM-ESWSA) is validated by benchmark analysis and statistical testing [9]. For the first time, here, meta-heuristic algorithms are used to identify biologically relevant biclusters from Duchenne Muscular Distrophy (DMD) [3] dataset. The enhanced exploration property of the proposed variant helps capture the biclusters with optimum cost. The rest of the paper is organized as follows. Section 2 briefly describes the problem and its formulation. Section 3 describes the state-of-the-art method ESWSA and proposed method CM-ESWSA. Section 4 describes the experimental results. Section 5 concludes the paper and mentions some future works.

Identification of Biologically Relevant Biclusters …

149

2 Problem Description 2.1 Biclustering Biclustering is a two-mode clustering technique [6], which discovers different pattern-based sub-matrices in Gene Expression Matrix, where, rows imply a subset of genes and columns imply a subset of conditions. These sub-matrices are biologically relevant, because they have a similarity in expression behavior. A sub-matrix is considered as a bicluster if the entries of rows follow a similarity in expression behavior across some of the conditions (columns). Let B be a bicluster consisting of a set of genes (g) and a set of conditions (c). Each element in B is represented by bi j (bi j ∈ B), i ∈ g and j ∈ c. Several kinds of biclusters are available based on gene expression patterns. This paper considers shifting and scaling pattern-based bicusters [2, 10]. For example, suppose M is a 5 × 4 data matrix, where rows imply a set of genes and columns imply a set of conditions . From data matrix (M), a shifting and scaling pattern-based bicluster (B 1 ) is selected, where the genes are and the conditions are . The mathematical notation of bi j in B 1 is bi j = Π j × αi + βi (αi = scaling coefficient and βi = shifting coefficient). In the pattern of B 1 = = , = = and = = . ⎛

Data matrix (M) is:

c1 g1 ⎜ ⎜ 97 ⎜ g 2 ⎜ 153 ⎜ g 3 ⎜ 110 ⎜ g 4 ⎝ 214 g 5 414

c2 28 103 80 144 105

c3 215 208 143 291 236

⎞ c4 177 ⎟ ⎟ ⎟ 293 ⎟ ⎟ 197 ⎟ ⎟ 304 ⎠ 202

Bicluster, B 1 (shifting and scaling pattern-based bicluster): ⎛

Mathematical notation of B 1 is:

⎛

⎞ c1 c4 g 1 ⎝ 97 177 ⎠ g 2 153 293 ⎞

c41 c1 g 1 ⎝ 20 × 4 + 17 40 × 4 + 17 ⎠ g 2 20 × 7 + 13 40 × 7 + 13

2.2 Initial Population In this work, each sub-matrix represents a candidate solution and the length of each candidate solution (one dimensional) is equal to the sum of number of rows and columns. Table 1 represents a randomly generated candidate solution. The bits 0’s

150

J. Adhikary and S. Acharyya

Table 1 Complete structure of a candidate solution g1

g2

g3

g4

g5

c1

c2

c3

c4

1

1

0

0

0

1

0

0

1

and 1’s are randomly selected for positions in rows and columns. In this case, bits 1 and 0 indicate the corresponding gene or condition is selected and not selected respectively in bicluster. After generating candidate solutions, cost of each solution has been calculated and then sorted on cost.

2.3 Cost Function In this research, the cost function has four different objectives, they are—transposed virtual error, volume, overlapping amount and gene variance [2]. Minimization of this cost function is capable of identifying shifting and scaling pattern-based bicluster [1, 2]. The cost function (C(B)) may contain weighted terms as follows. C(B) =

V E t (B) 1 + Ws × V ol(B) + Wov × Overlap(B) + Wvar × 1 + GeneV ar (B) V E t (M)

(1)

Transposed Virtual Error (V E t ) is used to measure the quality of a bicluster. It is used to identify the degree of relationship among genes [1, 2]. It is represented in the following equations|J |

ρi =

ρι =



bιj =

V E t (B) =

1  bi j |J | j=1

(2)

ρi − μ p σp

(3)

bi j − μc j σc j

(4)



|I | |J | 1 abs bιj − ρι i=1 j=1 |I | · |J |



(5)

Bicluster volume (V ol) is defined in terms of the number of genes and conditions present in a bicluster. A bicluster of smaller volume has greater chances to express a perfect pattern (low value of V E t ) of a bicluster [1, 2]. It is defined by equation: V ol(B) =

−ln(|I |) ln(|I |) + wg

+

−ln(|J |) ln(|J |) + wc

(6)

Identification of Biologically Relevant Biclusters …

151

The overlapping term (Overlap(B)) computes how many times an element of a bicluster has appeared in other biclusters [2]. Larger value of Overlap(B) enhances the chances of occurring erroneous pattern. Overlap(B) is defined in Eq. (7).  Overlap(B) =

i∈I, j∈J

W (bi j )

|I | · |J | · (n b − 1)

(7)

Gene variance (GeneV ar (B)) of a bicluster is defined by the mean of variances of all gene expressions in a bicluster [2]. High gene variance indicates that bicluster may not be significant. Gene variance can be expressed by Eq. (8). |I |

GeneV ar (B) =

|J |

 2 1 bi j − μgi |I | · |J | i=1 j=1

(8)

3 Proposed Method Meta-heuristic optimization techniques are generalized, robust and not problemspecific like heuristics, that is why, they are widely used. This research has used a meta-heuristic method (ESWSA) and its variant (CM-ESWSA) to identify biclusters.

3.1 Elephant Swarm Water Search Algorithm (ESWSA) The algorithm ESWSA [8] is a nature-inspired algorithm based on Swarm Intelligence. Elephants search water being divided into several groups/swarms. When a group of elephants is able to find some water resources, they try to communicate with other groups and confirm how far it is from the current position. The smaller the cost, the nearer the water resource is. In local water search, the nearest (best) location of the water source found by a group of elephants in its journey so far corresponds to the local best solution of the group. In global water search, the nearest location of water source experienced by all groups/swarms in their journey so far corresponds to the global best solution. The local water search and the global water search movements of the groups are expressed by Eqs. (9) and (10) respectively. These searching strategies are controlled by a context switching probability p [8]. Vi = Vi ∗ w + rand(1, d)  (Pbesti − X i );

(9)

Vi = Vi ∗ wt + rand(1, d)  (Gbest − X i );

(10)

152

J. Adhikary and S. Acharyya



wmax − wmin w = wmax − tmax

 ∗t

(11)

In Eqs. (9) and (10), rand(1, d) refers to a d-dimensional array of random values within [0, 1]. Here,w is an inertia weight strategy, named Linear Decreasing Inertia Weight (LDIW), expressed by Eq. (11). This strategy is used in both types of water search (local and global). Vi is the velocity of ith group of elephants and X i is the current position of ith group of elephants. The best elephant group finds water resources faster than others. The local best solution of ith elephant group is defined by Pbest i . The best solution among all local best solutions is the global best solution. It is defined by Gbest. In Eq. (11), t and tmax indicate iteration and total number of iterations, respectively. Here, wmax and wmin are control parameters.

3.2 Chaotic Move Elephant Swarm Water Search Algorithm (CM-ESWSA) In ESWSA, Linear Decreasing Inertia Weight (LDIW) strategy is used in both search processes (local and global water search). It is used to improve the performance of water search capability of elephants. This inertia strategy does not perform well in case of global water search, because w is decreased linearly from wmax to wmin (0.9 to 0.4) in a systematic manner. It is responsible for causing pre-mature convergence, which shows poor exploration capability of ESWSA. In the modified variant of ESWSA, named as Chaotic Move Elephant Swarm Water Search Algorithm (CMESWSA), the concept of logistic map (a type of chaotic map) is used to provide better exploration in global search process. wt = μ ∗ Z k ∗ (1 − Z k ) where, Z k ∈ (0, 1) and value of μ is 4.

(12)

Identification of Biologically Relevant Biclusters …

153

4 Numerical Experiment The methods have been implemented in a machine with Pentium Dual-Core CPU and a 4 GB memory; the software environment is Microsoft Windows 7 and the platform is Matlab R2012b. Section 4.1 analyzes the results obtained on benchmark functions. Section 4.2 describes convergence graphs and Sect. 4.3 analyzes the results obtained from statistical testing. Section 4.4 analyzes the results of real gene expression dataset (disease DMD).

4.1 Analysis of the Results on Benchmark Functions This section validates and compares the performance of ESWSA and CM-ESWSA. The size of the population is 10 and maximum number of iterations is 100. 50 independent runs have been taken in case of each benchmark function. The experiments

154

J. Adhikary and S. Acharyya

Table 2 Comparative performance between ESWSA and CM-ESWSA for benchmark functions (best cost is bolded in each row) Mean

Std Dev

Median

Min

Max

Sphere (F1) ESWSA

6.53e−07 4.23e−06 2.74e−10

8.05e−15

2.99e−05

CM-ESWSA

2.17e−08 8.33e−08 3.80e−11

4.75e−15

4.48e−07

Zakharov (F2) ESWSA

3.0555

12.1420

3.00e−04

2.60e−11

70.312

CM-ESWSA

0.7724

4.8386

8.58e−06

9.92e−12

34.1404

Dixon-Price (F3) ESWSA

0.1392

0.2854

0.0387

3.52e−05

1.904

CM-ESWSA

0.2197

0.4464

0.0316

1.47e−05

1.867

Shifted Sphere (F4) ESWSA

2.7090

4.2002

0.7603

5.438e−04 17.0089

CM-ESWSA

1.7956

2.9297

0.5825

9.779e−05 16.9792

Ackley (F5) ESWSA

1.161

2.7740

0.2574

7.919e−05 18.4332

CM-ESWSA

2.627

4.7431

4.66e−04

3.06e−07

18.1850

Shifted Rotated Griewank’s Function without Bounds (F6) ESWSA

0.0871

0.1412

0.0325

2.28e−07

0.6216

CM-ESWSA

0.0534

0.0783

0.0150

6.74e−07

0.3232

Shifted Rotated Rastrigin’s Function (F7) ESWSA

1.4031

3.1526

0.9950

2.257e−10 16.0

CM-ESWSA

1.0330

1.588

0.9950

1.101e−13 8.9546

have been conducted on four unimodal (having one local optimum) and three multimodal (more than one optimum) benchmark functions [9]. Experimental results in benchmark functions are given in Table 2, where the best value (minimized cost) has been bolded in each row. In Table 2, CM-ESWSA has outperformed ESWSA in all metrics (mean, standard deviation, median, min, max) for F1, F2, F4 and F7 benchmark functions. Moreover, it performs better than ESWSA in case of most of the metrics for F3, F5 and F6. So, the overall performance of CM-ESWSA is better in benchmark functions.

4.2 Convergence Analysis This section illustrates the convergence characteristics of ESWSA and CM-ESWSA applied to benchmark functions. In Fig. 1 and Fig. 2, convergence graphs are given for benchmark function F1 and F3, respectively. Here, X axis represents the number

Identification of Biologically Relevant Biclusters …

155

Fig. 1 Convergence graph for sphere (F1) benchmark function

Fig. 2 Convergence graph for dixon-price (F3) benchmark function

of iterations and Y axis represents the cost. In both cases, cost decreases gradually and steadily for CM-ESWSA and it reaches earlier to the most promising part of the search region than ESWSA.

4.3 Statistical Testing Wilcoxon’s rank sum (non-parametric) [9] test pair-wise compares the metrics. It is conducted on five benchmark functions (F1, F2, F5, F6 and F7). In Table 3, in all cases, the final outcome (p-value) is less than 0.05, which validates the performance of CM-ESWSA. Table 3 Results of statistical testing (ESWSA vs. CM-ESWSA) p-value

F1

F2

F5

F6

F7

1.2e−04

0.0164

1.5e−04

0.0464

0.0187

156 Table 4 Results on real dataset (DMD)

J. Adhikary and S. Acharyya Dataset

Algorithm

Number of biclusters

Mean of C(B)

DMD

ESWSA

100

344.2309

CM-ESWSA

100

244.1437

4.4 Results on Real Life Problem Here, Duchenne Muscular Dystrophy (DMD) [3] disease dataset has been used. It has 54,675 genes and 22 samples. ESWSA and CM-ESWSA have extracted 100 biclusters from DMD dataset. Table 4 represents the mean of cost (C(B)) of 100 biclusters. Mean of C(B) is minimum (bolded) in CM-ESWSA, so it has discovered better quality biclusters compared to ESWSA. This result justifies the superiority of CM-ESWSA.

5 Conclusion Biclustering is a well-known problem in computational genetics. For the first time, it is used to analyze DMD dataset, e.g. gene expression data and to find subgroups of genes that show similar responses under a subset of conditions. Therefore, it plays an important role in gene therapy. This research has proposed CM-ESWSA, a variant of ESWSA to perform better than other variants on various benchmark functions. To identify the best quality biclusters, the modified variant of ESWSA, i.e., CMESWSA is applied to DMD dataset. In this comparison, CM-ESWSA has performed better than other variants in finding the best quality biclusters.

References 1. Pontes, B., Giráldez, R., Aguilar-Ruiz, J.S.: Biclustering on expression data: a review. J. Biomed. Inform. 57, 163–180 (2015) 2. Pontes, B., Giráldez, R., Aguilar-Ruiz, J.S.: Configurable pattern-based evolutionary biclustering of gene expression data. Algorithms Mol. Biol. 8(1), 4 (2013) 3. Biswas, S., Acharyya, S.: Identification of disease critical genes causing Duchenne muscular dystrophy (DMD) using computational intelligence. CSI Trans. ICT 5(1), 3–8 (2017) 4. Huang, Q., Tao, D., Li, X., Liew, A.: Parallelized evolutionary learning for detection of biclusters in gene expression data. IEEE/ACM Trans. Comput. Biol. Bioinform. 9(2), 560–570 (2011) 5. Tanay, A., Sharan, R., Shamir R.: Discovering statistically significant biclusters in gene expression data. Bioinformatics 18, S136–S144 (2002) 6. Cheng, Y., Church, G.M.: Biclustering of expression data. Intell. Syst. Mol. Biol. 8(2000), 93–103 (2000)

Identification of Biologically Relevant Biclusters …

157

7. Mukhopadhyay, A., Maulik, U., Bandyopadhyay, S.: A novel coherence measure for discovering scaling biclusters from gene expression data. J. Bioinform. Comput. Biol. 7(05), 853–868 (2009) 8. Mandal, S., Saha, G., Paul, R.K.: Recurrent neural network-based modeling of gene regulatory network using elephant swarm water search algorithm. J. Bioinform. Comput. Biol. 15(04), 1750016 (2017) 9. Jordehi, A.R.: Enhanced leader PSO (ELPSO): a new PSO variant for solving global optimisation problems. Appl. Soft Comput. 26, 401–417 (2015) 10. Pontes, B., Giráldez, R., Aguilar-Ruiz, J.S.: Measuring the quality of shifting and scaling patterns in biclusters. In: IAPR International Conference on Pattern Recognition in Bioinformatics, pp. 242–252 (2010)

Shadow Detection Using DenseUNet Satyajeet Singh, Sandeep Yadav, Antoreep Jana, and Seba Susan

Abstract The objective of this paper is shadow detection. We propose a novel semantic segmentation approach based on DenseUNet. The UNet is an encoder– decoder architecture consisting of an encoder and a decoder. The encoder is used for downsampling the image and the decoder for upsampling the image to produce the required output. Our proposed architecture consists of the basic UNet model in which a Dense network connects the encoder and the decoder. Outputs from each of the four encoder downsampling units are connected to each of the four decoder upsampling units through the Dense network. The proposed architecture was trained and evaluated on the SBU shadow image dataset and the results assert the effectiveness of our method as compared to the state of the art. Keywords Shadow detection · DenseUNet · Neural network

1 Introduction Shadows are an inherent part of any illuminated scene. The obstruction of light by an object, emitted by a light source, results in the formation of a shadow. Shadows provide us with various useful information like the shape of the object, the position of the light source and illumination conditions. Apart from this useful information, shadows are a hindrance in a lot of image processing tasks like object segmentation S. Singh · S. Yadav · A. Jana · S. Susan (B) Department of Information Technology, Delhi Technological University, Shahbad Daulatpur, Bawana Road, Delhi 110042, India e-mail: [email protected] S. Singh e-mail: [email protected] S. Yadav e-mail: [email protected] A. Jana e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_16

159

160

S. Singh et al.

[1] and document image processing [2]. Algorithms used in photography, image editing, object detection and classification can be improved if the images are shadowfree and the first step in this direction is to detect the shadows. Therefore, automatic detection of shadow is a worthwhile task. A number of approaches have been studied for shadow detection. Physics-based illumination invariant approaches [3, 4] are effective in high-quality images. Image processing-based approaches [5, 6] are based on pixel manipulation. Such methods may not work well with all types of images. On the other hand, statistical inferencing techniques look promising but the results are still far from perfect [7]. The performance and results of these methods depend on the quality and quantity of training images. The shadow detection problem can be defined as the task of labeling each pixel as either shadow or non-shadow. Likewise, in semantic segmentation, each pixel is classified into one of the classes [8]. The similarity in the nature of the two tasks prompted us to approach the task of shadow detection as a semantic segmentation problem. In this paper, we have presented a novel deep learning architecture for automatic shadow detection. More specifically, a DenseUNet architecture is proposed. The organization of this paper is as follows. A brief review of shadow detection methods in literature is given in Sect. 2, the proposed DenseUNet architecture and methodology are presented in Sect. 3, the experimentation and results are described in Sect. 4, and the final conclusion is given in Sect. 5.

2 A Brief Review of Shadow Detection Shadows in image segmentation tasks were incidentally found segregated by histogram-based image thresholding in [9] and binary clustering of images in [10]. Both binarization schemes used the non-extensive entropy with Gaussian gain that exhibited a skewed response towards the darkest pixels in the image that included the shadow region. However, shadow detection has emerged over time as a dedicated task. It is some of these works that we review in this section. Finlayson et al. [11] have proposed a method that is based on the formation of illumination-invariant images from the given image. An illumination invariant image is an image in which the variation due to changes in the illumination is masked and the presence of shadow in the image is removed. Deb et al. [5] have proposed a method for image shadow removal by converting the RGB image into YCbCr color space. The method works by first converting the RGB image into YCbCr color space, then the average intensity of the Y channel is calculated and using this, thresholding operation is performed to form the shadow mask. Murali et al. [6] have proposed a method for image shadow removal by converting the RGB image into CIELA*B* color space. The method works by first converting the RGB image in the LA*B* color space, then in order to form the binary image shadow mask, the average of the intensities in the L, A* and B* planes are taken and a thresholding operation is applied. Maciej Gryka et al. [12] have proposed a method of image shadow removal by dividing the image into

Shadow Detection Using DenseUNet

161

patches. In this method, a feature vector is computed which is based on the information contained in a particular patch and this feature vector is used as input for different supervised learning algorithms. Reddy et al. [13] have proposed a method for image shadow removal by performing fuzzy logic operations on the image. The algorithm works by grouping the shadow pixels into four disjoint sets based on the S-curve, which is popularly used for modeling membership degrees.

3 Proposed Method 3.1 Intuition The task of shadow detection can be modeled as a semantic segmentation problem because, in this task, we are required to classify each pixel of the input image as either a shadow pixel or a non-shadow pixel. In semantic segmentation, we aim at labeling each image pixel to a particular class to which it belongs. This particular similarity in the nature of the two tasks prompted us to approach this problem as a semantic segmentation problem. The most renowned architecture used for semantic segmentation is UNet [14]. The UNet architecture has been primarily used before for biomedical semantic segmentation problems.

3.2 Basic UNet Architecture The UNet is an encoder–decoder architecture in which the encoder downsamples the image and the decoder upsamples the image to give the final output. The UNet is named so because of its “U” shaped structure. The UNet consists of three parts: the downsampling unit (encoder), the bottleneck section and the upsampling unit (decoder). The encoder is the part of the network, which is responsible for mapping the inherent features of the image into a smaller dimensional feature space. The bottleneck provides a smooth transition from the encoder to decoder. The decoder uses the encoding provided by the encoder network to upsample the image and provides the required output (Fig. 1).

3.3 DenseUNet Architecture DenseUNet architectures have already been used for biomedical semantic segmentation [15]. We took inspiration from these architectures and designed our own DenseUNet architecture. Our proposed DenseUNet architecture consists of the basic UNet model in which a Dense network connects and the encoder and decoder (Fig. 2). The

162

S. Singh et al.

Fig. 1 Basic UNet architecture

encoder has four downsampling blocks. Each downsampling block consists of two convolutional layers followed by a max-pooling layer. The encoder is followed by the bottleneck layer that comprises two convolutional layers. The bottleneck layer is followed by the decoder layer, which has four upsampling blocks. Each upsampling block consists of one deconvolutional layer (transposed convolution) followed by a concatenation layer. The concatenation layer is a part of the Dense network, which is connected to the encoder units. The Dense network connects the outputs from each of the four encoder downsampling units to each of the four decoder upsampling units. In order to connect all the decoder layers together we first reshape the outputs of decoder layers into the shape of particular encoder unit, which is under consideration. Then all the reshaped outputs are stacked together to generate the segmented output.

3.4 Preprocessing of Shadow Images For preprocessing the dataset, we load the RGB shadow image and grayscale shadow mask and resize them to dimensions (128, 128). The RGB shadow image is then normalized in the range [0, 1]. We perform binary thresholding on the shadow mask using the thresholding function  f (x) =

0, x < 127 1, x ≥ 127

(1)

Shadow Detection Using DenseUNet

163

Fig. 2 DenseUNet architecture used in our experiments

3.5 Training For training the model, we used Adam optimizer with a batch size of 8 images of size (128, 128). The training of the model was done in two phases on the SBU dataset [16] training images, which consists of 4085 images. The training images were divided into two sets, 4000 images in the training set and 85 images in the validation set. The binary cross-entropy loss function was used. In the first phase, we trained the model for 100 epochs using a learning rate of 0.001. In the second phase, we trained the model for 100 epochs using a learning rate of 0.0001. We used NVIDIA Tesla P100 GPU for a training time of 1 h. After training the network for two phases, we noticed that the loss is not decreasing anymore hence we decided to move forward with testing the model.

4 Results and Discussions 4.1 Testing The testing was performed on the 638 SBU dataset test images. For testing purposes, we used the following metrics, where TN is the number of True Negatives, TP is the number of True Positives, FP is the number of False Positives and FN is the number of False Negatives. Accuracy =

TP + TN TP + TN + FP + FN

(2)

164

S. Singh et al.

F1 score =

2 × precision × recall precision + recall

Recall =

TP TP + FN

Precision =

TP TP + FP

(3) (4) (5)

4.2 Comparison to the State of the Art For comparison purpose of the proposed method, we implemented the following shadow detection methods: 1. 2. 3.

YCbCr—Deb et al. [5]—Application of Image processing operations in YCbCr color space. LA*B*—Murali et al. [6]—Application of Image processing operations in LA*B*color space. Patch-based Random Forest by Gryka et al. [12]—Patch-based partition of image and supervised learning.

The results of shadow detection are shown in Figs. 3, 4, 5, 6, 7, 8 and 9 for three instances of shadow images. The proposed DenseUNet architecture yields the most visually cognizant results depicting the shadow region. The observation also tallies with the accuracy readings in Table 1 where the proposed method achieves higher scores for all the metrics than all other comparison methods. Figures 10 and 11 show a couple of failure cases where our algorithm does not perform up to the mark.

Fig. 3 a, b, c Three shadow images from the dataset

Shadow Detection Using DenseUNet

165

Fig. 4 a, b, c Original masks for three shadow images

Fig. 5 a, b, c Predicted masks using LA*B* (Murali et al. [6]) for three shadow images

Fig. 6 a, b, c Predicted masks using YCbCr (Deb et al. [5]) for three shadow images

4.3 Failure Cases As we can see from Figs. 10 and 11, our model fails in the cases where the objects in the image are too dark as compared to the shadow itself, giving erroneous results.

166

S. Singh et al.

Fig. 7 a, b, c Predicted masks using Patch-based Random Forest (Gryka et al. [12]) for three shadow images

Fig. 8 a, b, c Predicted masks using basic UNet for three shadow images

Fig. 9 a, b, c Predicted masks using DenseUNet for three shadow images

Shadow Detection Using DenseUNet

167

Table 1 Results of shadow detection for the proposed method and the state of the art (the UNet results are highlighted in bold) Method

Accuracy

F1-score

Recall

Precision

YCbCr (Deb et al. [5])

0.7352

0.5845

0.9753

0.4173

LA*B* (Murali et al. [6])

0.8931

0.7316

0.7630

0.7028

Patch-Based Random Forest (Gryka et al. [12])

0.7899

0.7271

0.6710

0.7935

UNet

0.9428

0.8484

0.8379

0.8592

DenseUNet

0.9439

0.8554

0.8688

0.8423

Fig. 10 a Original image. b Original mask. c Predicted mask

Fig. 11 a Original image. b Original mask. c Predicted mask

5 Conclusion UNet-based neural network architectures have been widely used in medical image segmentation. Realizing that shadow detection is nothing but a semantic segmentation problem, motivated us to propose a DenseUNet network architecture, which provided us with very promising results in this domain. The proposed architecture was trained and evaluated on the SBU shadow image dataset and the results assert

168

S. Singh et al.

the effectiveness of our method as compared to the state of the art. The extension of this project to the shadow removal problem forms the future scope of our work.

References 1. Susan, S., Verma, O.P., Swarup, J.: Object segmentation by an automatic edge constrained region growing technique. In: 2012 Fourth International Conference on Computational Intelligence and Communication Networks, pp. 378–381. IEEE (2012) 2. Susan, S., Rachna Devi, K.M.: Text area segmentation from document images by novel adaptive thresholding and template matching using texture cues. Pattern Anal. Appl. 1–13 (2019) 3. Finlayson, G., Hordley, S., Lu, C., Drew, M.: On the removal of shadows from images. IEEE Trans. Pattern Anal. Mach. Intell. 28(1), 59–68 (2006) 4. Finlayson, G., Drew, M., Lu, C.: Entropy minimization for shadow removal. Int. J. Comput. Vis. 85, 35–57 (2009) 5. Deb, K., Suny, A.H.: Shadow detection and removal based on YCbCr color space. SmartCR 4(1), 23–33 (2014) 6. Murali, S., Govindan, V.K.: Shadow detection and removal from a single image using LAB color space. Cybern. Inf. Technol. 13(1), 95–103 (2013) 7. Chen, Q., Zhang, G., Yang, X., Li, S., Li, Y., Wang, H.H.: Single image shadow detection and removal based on feature fusion and multiple dictionary learning. Multimed. Tools Appl. 77(14), 18601–18624 (2018) 8. Susan, S., Kumar, A.: Auto-segmentation using mean-shift and entropy analysis. In: 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom), pp. 292–296. IEEE (2016) 9. Susan, S., Singh, R., Kumar, A., Kumar, A., Kumar, A.: Segmentation of dark foreground objects by maximum non-extensive entropy partitioning. Int. J. Appl. Res. Inf. Technol. Comput. 9(1), 67–71 (2018) 10. Susan, S., Agarwal, M., Agarwal, S., Kartikeya, A., Meena, R.: Binary clustering of color images by fuzzy co-clustering with non-extensive entropy regularization. In: 2016 2nd International Conference on Contemporary Computing and Informatics (IC3I), pp. 512–517. IEEE (2016) 11. Finlayson, G.D., Hordley, S.D., Drew, M.S.: Removing Shadows from Images. In: Heyden, A., Sparr, G., Nielsen, M., Johansen, P. (eds.) Computer Vision—ECCV 2002. Lecture Notes in Computer Science, vol. 2353. Springer, Berlin, Heidelberg (2002) 12. Gryka, M., Terry, M., Brostow, G.J.: Learning to remove soft shadows. ACM Trans. Graph. (TOG) 34(5), 1–15 (2015) 13. Reddy, R.P.K., Nagaraju, C.: Low Contrast image Shadow removal by using Fuzzy logic technique. Int. J. Pure Appl. Math. 114(10), 55–63 (2017) 14. Ronneberger, O., Fischer, P., Brox, T.: U-net: convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and ComputerAssisted Intervention, pp. 234–241. Springer, Cham (2015) 15. Li, X., Chen, H., Qi, X., Dou, Q., Fu, C.-W., Heng, P.-A.: H-DenseUNet: hybrid densely connected UNet for liver and tumor segmentation from CT volumes. IEEE Trans. Med. Imaging 37(12), 2663–2674 (2018) 16. Vicente, T.F.Y., Hou, L., Yu, C.-P., Hoai, M., Samaras, D.: Large-scale training of shadow detectors with noisily-annotated shadow examples. In: European Conference on Computer Vision, pp. 816–832. Springer, Cham (2016)

Map Merging for an Autonomous Multi-robotic System Arnab Mandal and Chintan K. Mandal

Abstract Map-merging technique is a well-studied and challenging domain in the field of autonomous mobile robotics. It is not always possible to create a map only using a single robot. Thus, the most common practice is to collect maps of individual robots of the region it has been able to reveal, also known as sub-maps and merge all of them to create a single map. In this paper, we propose an algorithm for merging sub-maps acquired from different autonomous mobile robots. We also show that the merging of the maps is dependent on the degrees of freedom of the robots. Keywords Autonomous robots · Map merging · Motion planning

1 Introduction Autonomous mobile robots is a popular field of research in the domain of robotics [2]. One of the challenges is making the robot aware of its surroundings which can include static/moving bodies, its relative location to some object, etc. A robot can easily move in an environment if it is aware of its surroundings with the help of a map of the surroundings. However, a robot having no knowledge of its surroundings builds a map by itself by sensing the environment by using its sensors like a camera, distance sensors, wheel odometers, etc. (Fig. 1a). This process of generating the maps and simultaneously informing the robot about its location w.r.t. the map is known as SLAM (simultaneous localization and mapping). In a disaster or a war zone, it is not always possible to build a global map of the whole area by a single robot. In The authors are grateful to Dr. Ankur Mandal, IISER Mohali, Department of Physics for his valuable suggestions with the mathematics of the map-merging process. A. Mandal · C. K. Mandal (B) Department of Computer Science and Engineering, Jadavpur University, Kolkata, India e-mail: [email protected] A. Mandal e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_17

169

170

A. Mandal and C. K. Mandal

(a) A Robot Specification

(b) Communication structure of multi-robot system

Fig. 1 Robot structure

such areas, it is easier to build a map using multiple robots in a Cooperative-SLAM (C-SLAM) framework [5]. For a multi-robotic system, an acquired map of the region a robot explores is referred to as a sub-map, which are combined to form a single map. There are several approaches for map merging in C-SLAM with respect to accuracy and computation time, e.g., “Robot-to-robot measurements,” “Point feature matching,” and “Spectra-based map matching” [1, 4]. Tungadi et al. [6] discusses a probabilistic Haar-based place recognition approach in order to merge the maps. Zhou et al. [8] discuss a map-merging algorithm with unknown initial correspondence and a k-d tree is used to optimize the computation. Konolige [3] discusses a distributed map-merging technique where the map-merging decision is made by different type of patches from their features. In this paper, we propose a method to merge maps, acquired from different robots, to create a single map for C-SLAM. To illustrate our method, we have assumed that the robots move in an environment containing homogeneous/3D regular distinctly colored geometric landmarks.

2 Proposed Method In this section, we illustrate the steps for merging the sub-maps into a single map. This method requires the understanding of the map-merging characteristics [7] and their adaptation in this work.

Map Merging for an Autonomous Multi-robotic System

171

2.1 Map-Merging Characteristics and Its Adaptation (1) Reference frame: The reference frame refers to the axis which features a map. In our work, the reference frame of each map is the geo-magnetic frame of Earth obtained from the measurements taken through a on-board compass sensor (Fig. 2a). Each robot, having different starting points, has its own (Fig. 2a). The position of a robot, Ri , at instance t in its own reference frame is given by Rti (X tRi , YtRi ). (2) Type of the maps: There are different types of maps, e.g., grid maps, feature maps, and metric maps. For our work, we use the feature map, which are metric maps with images as a feature. Each robot Ri generates a map, referred to as a j Ri i , Y j,M ), sub-map. The position of a landmark in the sub-map is given as Mi (X Rj,M i i where j represents the object index and Mi denotes the sub-map obtained by Ri . (3) Decision of map merging: The map-merging decision is verified through the known map overlaps based on the common landmarks among the sub-maps. The minimum number of overlapped landmarks is dependent on the number of degrees of freedom of the robots, e.g., a differential drive robot has only two types of movements: a. linear motion, i.e., moving forward or backward b. rotation w.r.t to its vertical axis with a certain angle, e.g., moving left/right is rotation of the robot of 90◦ clockwise/anti-clockwise, respectively. Thus, as the degrees of freedom is two(2), the minimum number of common landmarks required are two(2). This will be further analyzed in Sect. 3. (4) Structure of the environment: The environment can be either regular (i.e., indoor environment, artificial test bed created by human, etc.) or as in a disaster environment. In our work, we choose a structured indoor environment. (5) Enabler of the map merging: The enabler is a combination of an “information exchanger” during map merging and the map-merging decision for orienting the sub-maps. The relative position/frames of the robot will be calculated using the common homogeneous objects or landmarks, after exchanging information of the sub-maps.

2.2 The Map-Merging Method Let there be a set of n robots, R = {R1 , R2 , . . . , Rn }, ordered such that two adjacent robots have two common landmarks and their corresponding sub-maps be M1 , M2 , . . . Mn . The robots move a d distance either forward/backward and detect a landmark/object at a distance s from it. Thus, the position of the robot, Ri , in its corresponding sub-map Mi with respect to its own reference frame (as set at the onset of its movement w.r.t. the geo-magnetic poles) at a given specific time stamp, t, is given by the following equations: Ri + d sin θt−1 X tRi = X t−1

(1)

172

A. Mandal and C. K. Mandal Ri YtRi = Yt−1 + d cos θt−1

(2)

the positions of the landmark j in the sub-map Mi are obtained through the visual sensors (e.g., camera, infrared sensors, etc.) which is given by the following equations: Ri i + st sin θt (3) XM j = X Y jMi = Y Ri + st cos θt

(4)

Equations 1–4 together define the sub-map for Ri . The robots send their individual sub-maps to the central server (Fig. 1b), where they are merged. In the first layer of the central server (Fig. 2b), n sub-maps are merged to form n  = n/2 number of sub-maps for the second layer, which are further merged to form n  /2 sub-maps and so on till the global map is not formed. If M1 , M2 , . . . Mn be the sub-maps, M12 represents the merged map generated from M1 and M2 in layer 2; M1234 is represented as the merged map generated from M12 , M34 in layer 3, and so on. In each layer, the number of maps to be merged is reduced by a factor of log2 . Suppose there are n i and n j landmarks present in Mi and M j , respectively, there should exist at least two common landmarks, l1 and l2 , among themselves. If there are more than two common landmarks, one can choose any two landmarks. The merging process of the sub-maps is done by taking two sub-maps at a time, Mi and M j . Thus, the final merged map will have a reference frame as the geo-magnetic frame of Earth, while its origin will be the same as the origin of R j . Among these common landmarks, l1 is used for translation of the sub-map M j w.r.t Mi . The points of M j are translated with respect to the translation factor, which is based on the coordinate of the common landmark, l1 in Ri and R j . X  and Y  are the translated coordinates. ⎡ ⎤ ⎡ ⎤⎡ ⎤ X 1 0 X X ⎢  ⎥ ⎣ ⎣ Y ⎦ = 0 1 Y ⎦ ⎣ Y ⎦ 1 0 0 1 1

R

(5)

R

where X = (X lR1 i − X l1 j ) and Y = (Yl1Ri − Yl1 j ) the translation factor along xaxis and y-axis, respectively. The other landmark, l2 , is used for rotating the points of map M j , i.e., (X  , Y  ) w.r.t Mi . The points of M j is rotated w.r.t. the rotation angle φ, which is the difference between l2 of M j and l2 of Mi . X  and Y  are the final coordinates after rotation. ⎤⎡  ⎤ ⎡  ⎤ X X cos φ sin φ 0 ⎢ ⎢  ⎥ ⎣ ⎥ ⎣ Y ⎦ = − sin φ cos φ 0 ⎦ ⎣ Y  ⎦ 0 0 1 1 1 ⎡

(6)

Hence, the merged maps are due to combination of translation and rotation of M j w.r.t Mi w.r.t. landmarks l1 and l2 , respectively.

Map Merging for an Autonomous Multi-robotic System

(a)

(b)

173

(c)

Fig. 2 a Reference frame robot. b Map merging of sub-maps in the central server. c Map-merging algorithm ⎡ ⎤ R1 R2 ⎡ ⎤ ⎤ ⎡  ⎤ cos φ sin φ 0 ⎢ 1 0 X l1 − X l1 ⎥ X X ⎢  ⎥ ⎣ R R ⎥⎣ ⎦ ⎣ Y ⎦ = − sin φ cos φ 0 ⎦ ⎢ ⎣ 0 1 Yl 1 − Yl 2 ⎦ Y 1 1 1 0 0 1 1 0 0 1 ⎡

(7)

We give a schematic view of the central server (Fig. 2b) to describe the merging procedure. The robots send the collected information about the sub-maps of the arena covered by them. These sub-maps are then merged, two at a time, as discussed below to get the global map of the arena. In Fig. 2c, we show the steps necessary for combining two sub-maps, Mi and M j , from two robots, Ri and R j . The environmental data (odometer data, distance sensor data, etc.) is used to build the sub-maps and merge them in the “merge-map” step. Step (1) The position and orientation of robots, Ri , R j , and their corresponding landmark(s) are found. Step (2) A Common Landmark Array (CLMA) is constructed of the common landmarks from sub-maps Mi and M j of Ri and R j , respectively. Step (3) The M j is rebuilt by translating all points of M j with respect to Mi , by fixing the l1 in the CLMA. Step (4) The points of M j are rotated w.r.t the angle between l1 and l2 by fixing the reference frame of Ri and R j to form Mi, j .

3 Map Merging and Degrees of Freedom: A Mathematical Analysis R1 A complete map is the union of the objects in the sub-maps: (l1,M , . . . , lnR11,M1 ) ∪ 1 Rn Rn Ri . . . ∪ (l1,Mn , . . . , ln n ,Mn ); l j,Mi is the jth object in the reference frame of Ri while Ri Ri (l1,M , . . . , ln,M ) are the landmarks in the sub-map Mi w.r.t global reference frame i i

174

A. Mandal and C. K. Mandal

covered by the robot Ri . The coordinates of the objects are denoted as (X lGj , YlGj ) in the global frame of reference, which forms the complete map. Let us consider two robots R1 and R2 to demonstrate the merging process. The coordinates of the objects in the sub-map M1 : M1R1 from R1 is represented by 1 1 (X lRi ,M , YliR,M ) ∈ M1R1 . The coordinates of these objects are in the frame of R1 . 1 1 These can also be represented by vectors, where the tip of the vector is at the position of the object and the tail is at the origin (0, 0) of the reference frame of R1 :  1 1 2 2 ai = X lRi ,M eˆ + YliR,M eˆ . Similarly, (X lRi ,M , YliR,M ) ∈ M2R2 in the frame of R2 1 x 1 y 2 2  is denoted as the vector b = X R2 eˆ + Y R2 eˆ . In addition, let the two common i

li ,M2 x

li ,M2 y

R1 R2 , l1,M ) object pairs identified by R1 and R2 in the global reference frame be (l1,M 1 2 R1 R2 , l ), respectively. and (l2,M 2,M2 1

Proposition 1 The relative position of the objects is same as that in the original R1 R2 map in global reference frame: If the locations of object pair (l1,M ⇐⇒ l1,M ) and 1 2 R1 R2 (l2,M1 ⇐⇒ l2,M2 ) obtained by R1 and R2 , respectively, for l1 and l2 coincide, then the same operation provides the merged map. Proof Consider the objects and their corresponding coordinates in global reference frame in sub-maps M1G and M2G be [a1 , a2 , a3 ] and [b1 , b2 , b3 ], respectively . G G = M1,2 = M1G ∪ M2G . Since (l1,M1 , l1,M2 ) The composite map is given by Mcomposite and (l2,M1 , l2,M2 ) pairs are the same objects, the coordinates a1 = b1 and a2 = b2 are same, which are in the global frame of reference. For simplicity, let us take the reference frame for R2 as same to the global reference frame. One gets the sub-map from R1 by arbitrary translation and rotation of M1G . The merging of the two sub-maps is done as follows: R1 R2 1. The difference between the position vectors of the objects l1,M and l1,M is 1 2     d = (a1 + R1 )ei θ − b1 . We subtract d from the position vectors obtained from R1   a2 + R1 )eiθ − (a1 + R1 )eiθ + b1 b1 + (a2 − a1 )eiθ =a2 a2 − d(   a3 + R1 )eiθ − (a1 + R1 )eiθ + b1 b1 + (a3 − a2 )eiθ =a3 a3 − d(

(8) (9)

2. Translate these vectors one more time to set origin of rotation as (0, 0), we  subtract a1 from all position vectors.     a1 = 0 a1 = a1 −  a2 = b1 + (a2 − a1 )eiθ − a1  = (a2 − a1 )eiθ   a3 = b1 + (a3 − a1 )eiθ − a1 = ((a3 − a1 )eiθ )

(10)

Map Merging for an Autonomous Multi-robotic System

175

Subtracting a1  (= b1 ) from the position vectors of the objects obtained by R2   b1 = b1 − a1 = b1 − b1 = 0

(11)

  b2 = b2 − a1 = b2 − b1

(12)

  b1 = b3 − a1 = b3 − b1

(13)

 3. Rotate the vector a2 by an angle φ to coincide with the vector   a2 eiφ = b2

(14)

 a1 )eiφ = b2 (a2 −

(15)

eiφ =

(b2 − b1 ) = e−iθ (a2 − a1 )e−iθ

(16)

R1 , Operating on the position vector of object l3,M1 obtained by R1 , i.e., l3,M 1  R1  iφ (a3 )e = (a3 − a1 ). Therefore, the relative position vectors of the object l3,M1 R1 R2 R2 and l1,M : a3 − a1 , l3,M and l1,M : b3 − b1 which are same as that in the global 1 2 2 reference are obtained.

Each sub-map can be thought of as a two-dimensional rigid body. The merging of the sub-maps is equivalent to fixing the position and orientation of one rigid body with respect to the other. As the degrees of freedom of a sub-map is 2, a translation and a rotation on the plane are necessary and sufficient for merging. Suppose there are n common landmarks are present in Mi and M j . Among these, one is needed for translation on the two maps and the second one is needed for rotation purpose. So, minimum number of landmarks needed for merging two maps are two, which is also the degrees of freedom of the objects/landmarks in the represented maps. If a map is represented in three dimension, then the number of points for merging two maps will be three.

4 Demonstration of the Map-Merging Process For our experiment, we use two differential drive robots consisting of two active gear wheels and one passive ball castor wheel, along-with a camera, compass, and distance sensor. The instructions to the robot are move forward, turn left, move backward, turn right, stop and get environmental data. The wheels do not have any slips, as the experiment was done on a smooth surface. The distance traveled by the robot with instruction forward/backward is π d. The test arena described in Fig. 3a consists of five homogeneous, distinctly colored football cones.

176

A. Mandal and C. K. Mandal

(a) Arena

(b) Sub-map M1 by R1

(c) Sub-maps M2 by R2 .

Fig. 3 Test arena and sub-maps of robots R1 and R2

(a) Sub-maps M1 (open circle) and M2 (asterisk) after translating the green object to (0,0) point

(b) Merged Map after rotation using the blue object from translated R1 and translated R2 .

Fig. 4 Merging of sub-maps M1 and M2 by translation and rotation

The robots (R1 and R2 ) generate the sub-maps M1 (open circle) and M2 (asterisk), respectively. Sub-map M1 consists of green, red, and blue landmarks while M2 consists of green, blue, yellow, and sky landmarks. Among M1 and M2 greenl1 and bluel2 are the common landmarks. With the above operations, one obtains the merged map. The whole process is shown in Fig. 4.

5 Conclusion In our proposed model, we propose a method for merging sub-maps into single map which is dependent on the degrees of freedom of the robot. In this work, we have proposed the system for robots where there are no slips. The method can also be modeled for robots having slips using Kalman filters. The map-merging process, proposed in the work can be done parallel, as merging of the maps in each level is mutually exclusive among themselves.

Map Merging for an Autonomous Multi-robotic System

177

References 1. Andersone, I.: The characteristics of the map merging methods: a survey. Sci. J. Riga Tech. Univ. Comput. Sci. 41(1), 113–121 (2010) 2. Frese, U., Hirzinger, G.: Simultaneous localization and mapping-a discussion. In: Proceedings of the IJCAI Workshop on Reasoning with Uncertainty in Robotics, pp. 17–26. Seattle (2001) 3. Konolige, K., Fox, D., Limketkai, B., Ko, J., Stewart, B.: Map merging for distributed robot navigation. In: Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003) (Cat. No. 03CH37453), vol. 1, pp. 212–217. IEEE (2003) 4. Lee, H.C., Lee, S.H., Lee, T.S., Kim, D.J., Lee, B.H.: A survey of map merging techniques for cooperative-slam. In: 2012 9th International Conference on Ubiquitous Robots and Ambient Intelligence (URAI), pp. 285–287. IEEE (2012) 5. Topal, S., Erkmen, I., Erkmen, A.M.: A novel multirobot map fusion strategy for occupancy grid maps. Turk. J. Elect. Eng. Comput. Sci. 21(1), 107–119 (2013) 6. Tungadi, F., Lui, W.L.D., Kleeman, L., Jarvis, R.: Robust online map merging system using laser scan matching and omnidirectional vision. In: 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 7–14. IEEE (2010) 7. Yan, Z., Jouandeau, N., Cherif, A.A.: A survey and analysis of multi-robot coordination. Int. J. Adv. Robot. Syst. 10(12), 399 (2013) 8. Zhou, G., Lu, J., Wan, C.Y., Yarvis, M.D., Stankovic, J.A.: Body Sensor Networks. MIT Press, Cambridge, MA (2008)

On Some Basic Graph Invariants of Unitary Addition Cayley Graph of Gaussian Integers Modulo n Joy Roy and Kuntala Patra

Abstract The Unitary addition Cayley graph of Gaussian integers modulo n denoted by G n [i] has a vertex set Zn [i]. Any two vertices, x = a1 + ib1 , y = a2 + ib2 of G n [i], are adjacent if and only if gcd(N (x + y), n) = 1, where N (a + ib) = a 2 + b2 is the norm of any element a + ib in Zn [i]. First, we find two important results of unitary addition Cayley graph of Gaussian integers modulo n by considering Zn [i] and Un [i] as the vertex set. Using the results, we obtain clique number, chromatic number of the graph G n [i]. Lastly, domination number of unitary addition Cayley graph of Gaussian integers modulo n is also found in this paper. Keywords Cayley graph · Gaussian integers · Clique number · Domination number

1 Introduction Dejter and Gudici [1] first introduced the unitary Cayley graph with vertex set Zn , the ring of integers modulo n, n > 1 and the two vertices a and b of Zn are adjacent if and only if a − b ∈ Un , where Un is the set of all units of Zn . Koltz and Sender [2] further established more results on the unitary Cayley graph. Some more different structures, properties and results of unitary Cayley graphs have been studied extensively by Akhtar et al. [3] and Boggers et al. [4]. Grynkiewicz et al. [5, 6] have introduced the concept of addition Cayley graphs denoted by G  = Cay + (, B), where  is an abelian group and B is a subset of . The addition Cayley graph induced by B on G is an undirected graph having the vertex set  and the edge set {ab | a + b ∈ B, a, b ∈ }. J. Roy (B) Department of Mathematics, Assam Don Bosco University, Sonapur 782402, Assam, India e-mail: [email protected] K. Patra Department of Mathematics, Gauhati University, Guwahati 781014, Assam, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_18

179

180

J. Roy and K. Patra

Sinha et al. [7] modified the addition Cayley graph to unitary addition Cayley graph by taking  = Zn and B = Un and established some interesting results and properties of unitary addition Cayley graph. Roy and Patra [8] generalised the concept of unitary addition Cayley graph to unitary addition Cayley graph of Gaussian integers modulo n, G n [i] by replacing the vertex set Zn to Zn [i], the set of Gaussian integers modulo n, and the edge set Un to Un [i]. The norm of an element a + ib in Zn [i] is defined as N (a + ib) = a 2 + b2 . An element c + id in Zn [i] will be a unit if and only if gcd(N (c + id), n) = 1. In G n [i], any two vertices a + ib and c + id are adjacent, whenever gcd(N ((a + c) + i(b + d)), n) = 1. Examples of G n [i] for n = 2, 3 and 4 are displayed in Figs. 1, 2 and 3. The paper concentrates on determining clique number and chromatic number of unitary addition Cayley graph of Gaussian integers modulo n. Moreover, domination number of the above-mentioned graph is also studied. Before going to our main results, some basic concepts and preliminary results are discussed in Sects. 2 and 3 which will serve as basic tools in developing the main results given in Sect. 4.

Fig. 1 G 2 [i]

Fig. 2 G 3 [i]

Fig. 3 G 4 [i]

On Some Basic Graph Invariants of Unitary Addition …

181

2 Preliminaries A graph G(V, E) is a pair, where V is a non-empty set and E is a set of unordered pairs of elements of V . The elements of the set V are called vertices and these unordered pairs are the edges of of the graph G. The degree of a vertex v ∈ V , denoted by deg(v) is the number of edges incident at v. If degree of each vertex is equal, say r in G, then G is called r -regular graph. If the vertex set V of a graph G can be partitioned in to two subsets V1 and V2 such that every edge of V1 joins with V2 , then the graph G is called bigraph or bipartite graph. If G contains every edge joining V1 and V2 , then G is called a complete bipartite graph. A graph is said to be complete graph if all its vertices are adjacent to each other. A complete graph with n vertices is denoted by K n . The chromatic number χ (G) is defined as the minimum n for which G has ncolouring. A graph G is n-colourable if χ (G)  n and n-chromatic if χ (G) = n. Clique of a graph G is the complete subgraph of the graph G and the cardinality of the maximal complete subgraph is called the clique number, denoted by ω(G). If the graph G with n vertices forms a maximal complete subgraph with p vertices, then the clique number will be ω(G)  p. In general, we have χ (G)  ω(G)  p. Let  be a group and B be a subset of  not containing the identity of . Assume B −1 ={b−1 | b ∈ B} = B. The Cayley graph Cay(, B) is an undirected regular graph of degree |B| having a vertex set V (X  ) =  and edge set E(X  ) = {ab | ab−1 ∈ B}, where a, b ∈ . The set of Gaussian integers denoted by contains Z[i] which contains all complex numbers of the form a + ib, where a and b are integers and an element a + ib is a unit in Z[i] if and only if N (a + ib) = 1. The only units of Z[i] are 1, −1, i and −i. Let , n ∈ N be the principal ideal generated by n in Z[i], the factor ring Z[i]/ is isomorphic to Zn [i] = {a + ib | a, b ∈ Zn } and it is called the ring of Gaussian integers modulo n. Some basic invariants of G n [i] are studied by Roy and Patra [8]. We present some of the results without proofs from the above-mentioned paper. Lemma 1 [8] Total number of elements in Zn [i] is n 2 . Lemma 2 [8] Total number of unit elements in Zn [i], (i) (ii) (iii) (iv) (v)

|Un [i]| = 22n−1 , when n = 2r , r ∈ N. |Un [i]| = n 2 − 1, when n ≡ 3(mod 4). |Un [i]| = (n − 1)2 , when n ≡ 1(mod 4). |Un [i]| = n 2 − n, when n = k 2 and k is an odd prime. |Un [i]| = |Un 1 [i]|.|Un 2 [i]|, for n = n 1 n 2 , where n 1 and n 2 are distinct primes.

Theorem 1 [8] Let m = a + ib be any vertex in G n [i], then

182

J. Roy and K. Patra

⎧ 2r −1 2 , i f n = 2r , r ∈ N. ⎪ ⎪ ⎪ ⎪ ⎪ 22r −1 (n 1 − 1)2 , i f n = 2r n 1 , r ∈ N and n 1 ≡ 1(mod 4). ⎪ ⎪ ⎪ 2r −1 ⎪ ⎪ (n 1 2 − 1), i f n = 2r n 1 , r ∈ N and n 1 ≡ 3(mod 4). ⎨2 deg(m) = (n − 1)2 , i f n ≡ 1(mod 4) and gcd(N (m), n) = 1. ⎪ ⎪ ⎪ n 2 − 1, i f n ≡ 3(mod 4) and gcd(N (m), n) = 1. ⎪ ⎪ ⎪ ⎪ ⎪n 2 − 2n, i f n ≡ 1(mod 4) and gcd(N (m), n) = 1. ⎪ ⎪ ⎩ 2 n − 2, i f n ≡ 3(mod 4) and gcd(N (m), n) = 1. Theorem 2 [8] G n [i] is a complete bipartite graph if and only if n = 2r , r ∈ N. Corollary 1 [8] G n [i] is a bipartite graph if and only if n = 2r n 1 , r ∈ N and n 1 ≡ 1(mod 4) or n 1 ≡ 3(mod 4).

3 Some Properties of Gaussian Integers The set of Gaussian integers defined as Z[i] = {a + ib|a, b ∈ Z} is the generalisation over the set of integers Z and this set has many similarities with that of Z such as both have primes, both are principal ideal domains, both satisfy division algorithms and norms which makes them Euclidean domains and any element in Z or Z[i] can be uniquely factorised in to primes, i.e., both of them enjoys unique prime factorisation. Despite so many similarities, the main difference that lies in them is in their factor rings. We know that if n is prime then Zn is a field. But for the same n, Zn [i] may not be a field. The reason behind this is if n is a sum of two squares, i.e., n ≡ 1(mod 4), then Zn [i] will never be a field. When n ≡ 3(mod 4), then Zn [i] will be a field. To discuss it in more detail, let us have a look on some important theorems. Theorem 3 [9] If a is a positive integer larger than 1, then Z[i]/ ∼ = Z a [i]. Theorem 4 [9] If a and b are relatively prime integers, then Z[i]/ ∼ = Z a 2 +b2 [i]. A non-zero Gaussian integer a + ib can be factorised as   v  a + ib = i d · σm u m · σm m · pmem · (1 + i)n . where |σm |2 , |σm |2 are primes in Z congruent to 1 modulo 4, pm is a prime in Z congruent to 3 modulo 4 and d, u m , vm , em and n all are non-negative integers for which |σm |  |σm |. Theorem 5 [9] If a and b are integers both not zero, and Rn = Z[i]/n , then the following holds:

On Some Basic Graph Invariants of Unitary Addition …

183

Z[i]/ = Zs1 ⊕ Zs2 ⊕ Zt [i] ⊕ Z2n/2 [i], f or even n. Z[i]/ = Zs1 ⊕ Zs2 ⊕ Zt [i] ⊕ Rn , f or odd n. where s1 = |σm |2u m , s2 = |σm |2vm and t =



pmem .

If n ≡ 1(mod 4), then n = a 2 + b2 = (a + ib)(a − ib), for some a, b ∈ Z, then Z[i]/ = Z[i]/ ⊕ Z[i]/< a − ib> ∼ = Z[i]/a 2 + b2 ⊕ Z[i]/ 2 2 ∼ a + b = Zn ⊕ Zn . Thus, for any prime n, Zn is a field but Zn ⊕ Z n will contain zero divisors. Hence, Zn [i] is not a field for n ≡ 1(mod 4). But if n is a prime that is congruent to 3 modulo 4, then n is irreducible in Zn [i] and then Zn [i] is a field.

4 Clique Number and Chromatic Number of G n [i] In this section, we determine the clique number and chromatic number of unitary Cayley graph of Gaussian integers modulo n, G n [i]. First, we prove the following lemma. Lemma 3 If n ≡ 1(mod 4) or n = 4q + 1, q ∈ N, then the unitary addition Cayley graph of Gaussian integers modulo n with the vertex set consisting of non-zero zero divisors of Z n [i] is a complete bipartite graph. Proof When n ≡ 1(mod 4), then n = a 2 + b2 can be factorised as a product of two irreducible elements or Gaussian primes (a + ib) and (a − ib). Thus, and are the only maximal ideals in Z[i]. Also, and are the only maximal ideals in Zn [i] and elements of these maximal ideals are the zero divisor elements. But according to our definition of adjacency of two elements of a graph, none of the elements of are adjacent with each other and it is same with the elements of . But every non-zero elements of will be adjacent to every non-zero elements of . Hence, they will form a complete bipartite graph. Lemma 4 If n ≡ 1(mod 4) or n = 4q + 1, q ∈ N, then the unitary addition Cayley graph of Gaussian integers modulo n with the vertex set U (Zn [i]) is a regular graph. Proof When n ≡ 1(mod 4), then n = a 2 + b2 and there are two maximal ideals and each containing n − 1 elements. Then by Lemma 3 the set of non-zero zero divisors forms a complete bipartite graph. If we take vertex set as the set of unit elements Un [i] then the degree of each vertex will be n 2 − 2n − 2(n − 1) − 1 = n 2 − 4n + 3. Hence, it is a regular graph of degree n 2 − 4n + 3. Theorem 6 For n = 2r where r ∈ N, χ (G n [i]) = ω(G n [i]) = 2. Proof By Theorem 2, when n = 2r , then G n [i] is a complete bipartite graph. So, χ (G n [i]) = ω(G n [i]) = 2. The result can be extended to any finite product.

184

J. Roy and K. Patra

Corollary 2 For n = 2r n where r ∈ N and n is either of the form 4q + 1,q ∈ N or 4q + 3, q ∈ N ∪ {0}, then χ (G n [i]) = ω(G n [i]) = 2. Proof By Corollary 1, when n = 2r n 1 , r ∈ N and n 1 is any odd prime of the form are 4q + 1 or 4q + 3, then G n [i] is a bipartite graph. So, χ (G n [i]) = ω(G n [i]) = 2. g  β Corollary 3 For n = 2r i=1 n iαi hj=1 n j j where r ∈ N and n i ≡ 1(mod 4) and n j ≡ 3(mod 4), then χ (G n [i]) = ω(G n [i]) = 2. Proof The whole vertex set of G n [i] can be partitioned into two parts V1 and V2 one containing the vertices of even norm and the other containing the vertices with odd norm. By using the definition of G n [i], we get a bipartite graph. Hence, χ (G n [i]) = ω(G n [i]) = 2. Theorem 7 For n = p m where p is a prime of the form 4q + 3 q, m ∈ N ∪ {0}, then 2m 2m−2 χ (G n [i])  ω(G n [i])  ( p −2p ) + 1. Proof When n = 4q + 3, q ∈ N ∪ {0} and n is prime, then Z n [i] is a field and has only one zero divisor, namely, 0 + i0. Let n = p n , where p is of the form 4q + 3, where q ∈ N, then the number of zero divisors in Z n [i] will be p 2n−2 and hence the number of unit elements will be p 2n − p 2n−2 . If the additive inverses are removed, 2n 2n−2 then there will be ( p −2p ) number of unit elements along with 0 + i0 they will form a maximal complete subgraph. Hence, χ (G n [i])  ω(G n [i]) 

( p2n − p2n−2 ) 2

+ 1.

Theorem 8 For n = p m , where p is a prime of the form 4q + 1 q, m ∈ N, then m m−1 2 χ (G n [i])  ω(G n [i])  ( p −4p ) + 2. Proof When n is a prime of the form 4q + 1, q ∈ N, then Zn [i] ∼ = Zn ⊕ Zn . Let n = p m then Zn [i] ∼ = Zn ⊕ Zn = Z pm ⊕ Z pm . Now number of unit elements in Z pm m m−1 is p m − p m−1 . Removing the additive inverses we get p −2p distinct unit elements.

So, in Z pm ⊕ Z pm we get ( p −4p ) unit elements which will be adjacent to each other. Again when n is of the form 4q + 1 then by Lemma 3 non-zero zero divisors will m m−1 2 form a complete bipartite graph. Hence, ( p −4p ) number of unit elements along with a copy of K 2 will form a complete subgraph in G n [i]. Hence, χ (G n [i])  m m−1 2 ω(G n [i])  ( p −4p ) + 2. m

m−1 2

Theorem 9 For n = n 1 n 2 where n 1 = n 2 and they are of the form 4q + 3, q ∈ (n 2 −1)(n 2 −1) + 2. N ∪ {0}, then χ (G n [i])  ω(G n [i])  1 4 2 Proof If n = n 1 n 2 , where n 1 , n 2 are of the form 4q + 3, q ∈ N ∪ {0}, then there are two maximal ideals generated by and . So, an element a + ib will be a zero divisor element in Zn [i] iff both a and b are multiples of either n 1 or n 2 . We now partition the zero divisor set in to two parts. The first part contains the zero divisors having both the real and imaginary part multiples of n 1 and the second part

On Some Basic Graph Invariants of Unitary Addition …

185

contains the zero divisors having both the real and imaginary part are multiples of n 2 . Now the elements of these two partition forms a bipartite graph. Again by Chinese (n 2 −1)(n 2 −1) remainder theorem Zn [i] ∼ number = Zn 1 [i] ⊕ Zn 2 [i]. Thus, there will be 1 4 2 of unit elements along with a copy of K 2 forms a maximal complete subgraph of (n 2 −1)(n 2 −1) G n [i]. Thus, χ (G n [i])  ω(G n [i])  1 4 2 + 2. Corollary 4 For n = n 1 n 2 . . . n r where n 1 , n 2 , . . . , n r all are distinct primes of the (n 2 −1)(n 22 −1) ... (nr2 −1) + 2. form 4q + 3, q ∈ N ∪ {0}, then χ (G n [i])  ω(G n [i])  1 2r Theorem 10 For n = n 1 n 2 where n 1 and n 2 are distinct prime such that n 1 = 4q + 3,q ∈ N ∪ {0} and n 2 = 4q + 1,q ∈ N, then χ (G n [i])  ω(G n [i])  (n 21 −1)(n 2 −1)2 + 3. 8 Proof If n = n 1 n 2 where n 1 = 4q + 3, q ∈ N ∪ {0} and n 2 = 4q + 1, q ∈ N, then by Chinese remainder theorem Zn [i] ∼ = Zn 1 [i] ⊕ Zn 2 [i] ∼ = Zn 1 [i] ⊕ (Zn 2 ⊕ Zn 2 ). (n 21 −1) · (n 2 −1)2 number of unit elements will form a complete subgraph. Again Thus, 8 the zero divisor elements of Zn 1 [i] are of the form a + ib where both a and b are multiples n 1 and by Lemma 3 the non-zero zero divisor elements of Zn 2 [i] will form a complete bipartite graph and all the vertices of this bipartite graph will be adjacent to all the zero divisor elements of Zn 1 [i] and all the zero divisors all together (n 2 −1) · (n −1)2 number of unit elements along will form several copies of K 3 . Thus, 1 8 2 with one copy of K 3 will form a maximal complete subgraph of G n [i]. Therefore, (n 2 −1)(n −1)2 χ (G n [i])  ω(G n [i])  1 8 2 + 3. Corollary 5 For n = ms where m = n 1 n 2 . . . n r and s = n 1 .n 2 . . . n t also n i ’s, n j ’s are of the form 4q + 3,q ∈ N ∪ {0} and 4q + 1, respectively, q ∈ N, then χ (G n [i])  (n 2 −1)(n 22 −1) ... (nr2 −1) · (n 1 −1)2 · (n 2 −1)2 ... (n s −1)2 + 3. ω(G n [i])  1 2r +2t Theorem 11 For n = n 1 n 2 where n 1 and n 2 are distinct primes of the form 4q + 1, 2 2 q ∈ N, then χ (G n [i])  ω(G n [i])  (n 1 −1)16(n 2 −1) + 4. Proof If n = n 1 n 2 , where n 1 and n 2 both are of the form 4q + 1, q ∈ N then by Chinese remainder theorem Zn [i] ∼ = Zn 1 [i] ⊕ Zn 2 [i] ∼ = (Zn 1 ⊕ Zn 1 ) ⊕ (Zn 2 ⊕ Zn 2 ). (n 1 −1)2 · (n 2 −1)2 number of unit elements will form a complete subgraph. Now Thus, 16 by Lemma 4.1 the non-zero zero divisor elements of Zn 1 [i] as well as Zn 2 will form a complete bipartite graphs. Again taking one copy of K 2 from each bipartite graphs 2 2 forms a complete subgraph K 4 . Thus, (n 1 −1) 16· (n 2 −1) number of unit elements along with a K 4 forms a maximal complete subgraph. Therefore, χ (G n [i])  ω(G n [i])  (n 1 −1)2 (n 2 −1)2 + 4. 16 Corollary 6 For n = n 1 n 2 . . . n r where n 1 , n 2 , . . . , n r are all distinct primes of 2 2 ... (nr −1)2 + 4. the form 4q + 1,q ∈ N, then χ (G n [i])  ω(G n [i])  (n 1 −1) (n 2 −1) 22r

186

J. Roy and K. Patra

5 Domination Number of G n [i] A dominating set of a graph G is a subset D of the vertex set of G with the property that every vertex not in D is adjacent to atleast one vertex of D. The domination number of G, denoted by Domn(G), is given by the cardinality of a minimum dominating set of G. Theorem 12 The domination number of G n [i] is ⎧ ⎪ ⎨1, i f n ≡ 3(mod 4) Domn(G n [i]) = 2, i f n is even ⎪ ⎩ 3, i f n ≡ 1(mod 4) Proof (i) When n is an odd prime and n ≡ 3(mod 4), then Z n [i] is a field and hence 0 + i0 will adjacent to all the unit elements. So, Domn(G n [i]) = 1. (ii) When n is even, then G n [i] is a bipartite graph. So, Domn(G n [i]) = 2. (iii) When n is an odd prime and n ≡ 1(mod 4), then 0 + i0 will be adjacent to all the unit elements and the non-zero zero divisors will form a complete bipartite graph. So, Domn(G n [i]) = 3.

6 Conclusion When n is an odd prime of the form 4q + 1, q ∈ N, then we get two maximal ideals and which forms an induced complete bipartite graph in G n [i] and the remaining regular elements forms an induced regular graph of degree n 2 − 4n + 3 in G n [i]. The clique number and chromatic number of G n [i] is 2 only when n is an even number. Moreover, clique number and chromatic number of G n [i] will vary depending on different prime factorisation of n. The domination number of G n [i] are 1, 2 and 3 when n ≡ 3(mod 4), n ≡ 2(mod 4) and n ≡ 1(mod 4), respectively.

References 1. Dejter, I.J., Giudici, R.E.: On unitary Cayley graphs. J. Comb. Math. Comb. Comput. 18, 121–124 (1995) 2. Klotz, W., Sander, T.: Some properties of unitary Cayley graphs. Electron. J. Comb. 14, # R45 (2007) 3. Akhtar, R., Boggess, M., Jackson-Henderson, T., Jimenez, I., Karpman, R., Kinzel, A., Pritikin, D.: On the unitary Cayley graph of a finite ring. Electron. J. Comb. 16(1), # R117 (2009) 4. Boggess, M., Jackson-Henderson, T., Jimenez, I., Karpman, R.: The structure of unitary Cayley graphs. J. Summer Undergrad. Math. Sci. Res. Inst. (2008)

On Some Basic Graph Invariants of Unitary Addition …

187

5. Grynkiewicz, D., Lev, V.F., Serra, O.: Connectivity of addition Cayley graphs. J. Comb. Theory Ser. B 99(1), 202–217 (2009) 6. Grynkiewicz, D., Lev, V.F., Serra, O.: The connectivity of addition Cayley graphs. Electron. Notes Discrete Math. 29, 135–139 (2007) 7. Sinha, D., Garg, P., Singh, A.: Some properties of unitary addition Cayley graphs. Notes Number Theory Discrete Math. 17(3), 49–59 (2011) 8. Roy, J., Patra, K.: Some aspects of addition Cayley graph of Gaussian integers modulo n. MATEMATIKA 32(1), 43–52 (2016) 9. Dresden, G., Dymacek, W.M.: Finding factors and factor rings over the Gaussian integers. Math. Assoc. Am. Mon. 112 (2005)

A Brief Survey of Token Curated Registries Jaspreet Kaur and B. Visveswaraiah

Abstract A list is a simple and efficient way for organizing data. Traditional lists have higher chances of unfairness, manipulation, and completely give misinformation because these are managed by the individuals or central authorities. Unlike a traditional list, Token Curated Registry or list also known as TCR is completely decentralized blockchain-based platform which uses intrinsic token along with an incentive system to produce an accurate list curation. This paper gives a brief survey about this decentralized list along with their different types, applications, and various open challenges. Keywords Blockchain · Token Curated Registry or list (TCR) · Decentralized structure · Token · Incentive system

1 Introduction A list is a simple and efficient way of organizing data. There are various lists which we use in day-to-day life such as grocery lists, to-do lists, lists of top movies, top restaurants, etc. Traditional lists have higher chances of unfairness, manipulation, and completely give misinformation because these are managed by the individuals or central authorities. These lists can be manipulated and corrupted by online bots to provide higher incentives to the intruders [1]. So, we require a decentralized technology in the economic web that can overcome these limitations. One such technology is blockchain. Fundamentally, blockchain is a decentralized, distributed, shared, and immutable database ledger that stores any type of data and transactions J. Kaur (B) Department of Computer Science & Engineering, Indian Institute of Technology Jodhpur, Jodhpur, India e-mail: [email protected] B. Visveswaraiah Department of Electrical Engineering, Indian Institute of Technology Jodhpur, Jodhpur, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_19

189

190

J. Kaur and B. Visveswaraiah

across a peer-to-peer (P2P) network. It is an increasing list of assets called blocks which are linked using cryptography to provide data authentication and integrity. It can efficiently store data in a verifiable and permanent way and also allows the majority of participants economically handle the process of verifying transactions [2]. Integration of blockchain into economic web creates various opportunities, which is far away from the recent perception of production where users would retain the value of production. Nowadays, blockchain can be used as an interaction layer in economic web (the web of humans behaving as economic agents) [3]. Tokens are a critical component in this economic web. When a user joins to the network, they have to invest some assets in the exchange of tokens, and chances of success of their applications increases the value of their tokens [3]. Different types of tokens are used depending on various use cases such as security token, asset token, utility token, cryptocurrencies, or payment tokens. Currently, there are various use cases available of token model based on decentralized and incentive mechanisms. The most reputed use of token model is bitcoin or cryptocurrency where token awards give to the miner in the exchange of cryptocurrency. Another most popular token model is Ethereum. Today, it is the most usable platform for creating decentralized applications and smart contracts. So, developers and researchers thought to take advantage of these advanced technologies (blockchain, token model, incentive mechanism) for solving all of the issues or limitations of current web economics. Finally, They came with a new idea, i.e., Token Curated Registry (TCR). The rest of the paper is organized as follows. Section 2 describe the concept and working of TCR followed by the different types of TCR and advantages of TCR in Sects. 3 and 4. In Sect. 5, we give various applications of TCR. Section 6 describe various challenges in TCR along with existing possible solutions. Finally, the conclusion has to be given with future directions of this technology.

2 TCR and It’s Working TCR (list) [4] is a completely decentralized blockchain-based platform in which the product or output is a high-quality list (most valuable and authentic list). It uses the concept known as the wisdom of the crowd and uses an intrinsic token to assign curation rights equivalent to the relative token weight. It creates a distributed network of users maintained by a functional architecture of incentives. So, as long as there are individual or companies which would desire to be curated into a given list, a market can exist which fulfill their desire along with incentive mechanism and aligned toward curating a high-quality list. The basic structure of TCR (shown in Fig. 1) has three types of users depending upon their different interests, incentives, and interaction pattern toward the registry. Users are consumers, candidates, and token holders. These allow the TCR to be self-sustaining. A brief description of these as follows.

A Brief Survey of Token Curated Registries

191

Fig. 1 a General flow of TCR. b Real example of TCR as adChain registry [5]

2.1 Consumers Consumers are the users who desire high-quality information or lists from the internet so that they make their decision in the right direction. In the case of “The Top Business Schools,” the consumers would be parents or students seeking an authentic list of potential colleges. Let a consumer is making decisions about top colleges on the basis of a list which ranks colleges based on their performance within 10 years, if the information of any school in traditional lists are corrupted with misinformation, then the demand for that school in correctly implemented TCR will be high.

2.2 Candidates Candidates are the people, colleges, restaurants, things, etc., that desire to be placed on the list in order to get the attention and consideration of consumers. A school on a list of colleges which has high performance within 10 years will likely see greater application volume than it would were it not on the list. So that, it may be able to raise the reputation of school on that basis.

2.3 Token Holders They are the vital components in maintaining the incentive structure based on their self-interests. They hold tokens for financial gain (as high as possible). They achieve this by simply curating the registry and maintain the balance between entry and rejection of candidates into the registry. If consumers find the list to be inaccurate, they will not use it, candidates will not pursue to be on it, and the value of the token

192

J. Kaur and B. Visveswaraiah

will drop and token value will be high in the reverse state. Anyone can become a token holder by purchasing the tokens on an exchange. It allows the list to update consistently and remain useful in the longer term.

2.4 The Incentive System of Token Curated Registries To make a TCR work, candidates must make an initial deposit nominated in the registry’s intrinsic token to be considered for listing. If a candidate is valid and accepted as a listee. They keep this deposit and may withdraw it at any time. If a candidate is invalid, then its application may be challenged by token holders along with a same initial deposit. If the application would be rejected, its deposit is lost and divided up as a reward among token holders who participated in the voting process. But if the application would be accepted, its deposit remains along with them and challengers lost their deposit to the participating candidate. Candidates will not apply to the registry who believe they would obviously be rejected, as this would result in a financial loss for them.

3 Types of TCRs We classify various token curated registries based on the different parameters such as design pattern, data, and implementation metrics.

3.1 TCR Classification Based on Design Pattern 3.1.1

Unordered TCR

These are lists with a basic set of actions such as entering and exiting the list. These can be finite or infinite. Focal point parameters must be chosen more accurately for better results because these lists are related to game theory [6].

3.1.2

Ordered TCR

The ordered TCR as the name suggests, manage order between various listee by adding an index to each entry in the list and adding additional features to stake and swap indexes. Finally produce a list as queues, rank, or instruction sets. Real-world example as Token Curated Playlists that use tokens to set the playlist of fan-favorite DJ at a music event [6].

A Brief Survey of Token Curated Registries

3.1.3

193

Graded TCR

It is similar to the ordered TCR except an additional feature gives rise to the cryptoeconomic, dual application of ranking and reputation systems where objects can compare with related objects. In this, we extend TCR to dynamic voting of the listee [6].

3.1.4

Layered TCR

The layered TCR is an interesting concept. They replace the binary membership approach of TCRs, with a step-by-step increase in rights and responsibilities as the user adds more services and support. The user only lives in one layer at a time. In order to move to the next layer, the listing must take some sort of responsibilities or criteria. It can be use for access control, governance, identity, reputation, and more [6, 7].

3.1.5

Nested TCR

It is also pronounced as TCR of TCRs. Here, it uses a pointer to linked TCRs or listings. We can represent any type of graph, network, and relationship using this type of TCR. For example, tree-based chat dialogue [6].

3.1.6

Combinatorial TCR

This TCR has the ability to support combinations or permutations of items in a single listing. Each list can be implemented as an array of items which can be either ordered (indexing) or unordered (set completeness) [6].

3.1.7

Framework-Based TCR

In basic TCRs, a user that risks capital must initiate the challenge. In fTCRs, challenges are instantiated by changes in the framework. This method allows us to challenge the conditions which are suitable for the selection of a listee. This method provides objective accounting and reduces app cost [8].

194

J. Kaur and B. Visveswaraiah

3.2 TCR Classification Based on Data 3.2.1

Finite Element TCR

These are lists with a finite number of elements that can be included in them, e.g., coffee shops TCR in a particular area. This can be implemented in token economics with the help of fixed token supply and reparameterization technique [9].

3.2.2

Continuously Growing TCR

These types of TCR depends on the fact that they start out with a large pool of possible initial listees, and the rate of creating possible listees is greater than exploding them,. e.g., human population. If you want to create a TCR today, then you have a large number of members, and the rate of birth is higher than the rate of death. It means, there will be continuous growth going into the future [9].

3.2.3

Data Accrual TCR

DA TCRs want to accumulate data with zero rent-seeking and it uses one token across all TCRs. In this, we use a pending pool where anyone can apply as a listee, without having to deposit tokens. Token holders can then take these data points and submit those to DA TCRs where they get the reward if it gets accepted [9].

3.3 TCR Classification Based on Implementation Metrics 3.3.1

Token Curated Registries 1.0

This is the basic or first version of TCR. It can maintain a decentralized list with the help of crowd decision. The three users are perform their functionality very well as described above. But this earlier version of TCR has several challenges as vote splitting, registry poisoning, and many more. So to overcome these limitations, enhance versions of TCR are developed as versions 1.1 and 2.0 [4].

3.3.2

Token Curated Registries 1.1

In TCR 1.0, there are several challenges which limit the usability of these lists. One such challenge is vote splitting which can be solved in this upgraded version of TCR 1.1 by adding new parameters called MINORITY_BLOC_SLASH. It is the percentage of tokens in the losing voting block disbursed to the winning block, as additional rewards [10].

A Brief Survey of Token Curated Registries

3.3.3

195

Token Curated Registries 2.0

In this TCR, we solve the issue related to the misbehaving inequality. In earlier TCR, this issue can be challenged by the parameter MIN_DEPOSIT. But what if this parameter did not exist? This parameter has been replaced with trust pools, where deposit value can anywhere between zero and entire token supply. For where the inequality is true, it can be expected to take the greatest V_action (set of actions). It is useful for determining the situations when a listed entity may be bribed to take an inaccurate action. The V_stake (value of stake) for any listee is equal to the size of its trust pool [10].

4 Advantages of TCR There are various benefits given below by which TCR is more valuable than the traditional list. 1. The first or most important property of TCRs is its decentralized nature which protects consumers from obtaining manipulative information. TCR is based on wisdom of crowd. 2. Transparency and Trust can be achieved in TCR due to the features of blockchain. TCRs store all transactions on a blockchain for making an authentic decision. The entire history of the list can be tracked for detecting any misbehavior. 3. TCRs uses incentive mechanism to keep information in the list on an ongoing basis instead of creating a single final list or we say it maintains Up-to-date Information. 4. In TCR, every user follows incentive mechanism to increase the value of their product or asset, and penalty is also given to the users for providing any misinformation. This mechanism provides better security in TCR. 5. Self-regulation is an essential property to an economic web. It can lead to automation of any machine which reduces human intervention. TCR produces self-regulating, decentralized, and high-quality lists unlike older traditional lists. 6. At the initial level, users have no rights or permissions in TCR. As users add up the value or stack, they gain rights for those services as vote or challenge any application or read and write files/directories to any file systems. 7. TCR uses public key for identity of users which provide pseudo anonymity. The level of identity can be increased for better user security such as adding email or social media profiles. 8. Token holders use different consensus mechanism as DPOS (Delegated Proof of Stake) for challenging or voting to an application. So that, they can increase the value of their tokens as well as reputation in that system.

196

J. Kaur and B. Visveswaraiah

5 Applications or Use Cases of TCR As already discussed TCRs need three participants for their survival as an incentive structure, the purpose of a TCR is to create a valid list of something which is accepted by the end consumer. Token holders have economic incentive to curate the list and applicant benefits by being on the list. TCR is playing a major part in blockchain ecosystem and is used for different purposes. Here, we present some of the applications of TCR (Due to the page/space limit).

5.1 Journalism Civil Registry is a TCR for journalists and citizens. The aim of Civil is to create a platform for journalism that is free from misleading ads and to build up the trust between citizens and news reporting. It limits the need for third or centralized parties. They maintain the list of Newsrooms (publications) and use Civil constitution on challenging the applicant. Their application to the registry will outline the newsrooms mission, roster, intended community, how it plans to direct that funding [11].

5.2 Regulatory With the creation of ICOs, there is need for transparency of self-regulatory organizations has increased. Messari is a TCR-based platform which maintains a list featuring high-quality, legitimate crypto companies, and creating a self-regulatory organization for the industry. They include basic details about crypto companies, such as their mission statement, the nature of the products or services that they offer, and an explanation of the technology they use [12].

5.3 Ocean Protocol Ocean protocol aims to equalize the opportunity to access data, so that AI practitioners can create value from it. It is a decentralized data-sharing protocol, where TCR is used as an incentive mechanism for user identity by maintaining a whitelist of good actors in the network, where actors do not include consumer of data/service. Various stakeholders are data keepers, curators, providers, and verifiers. It is also used to check the authenticity of the data [13].

A Brief Survey of Token Curated Registries

197

5.4 Deconet In the emerging forms of economics, Deconet offers knowledge workers and investors within the knowledge economy, by providing a crowd-curated marketplace for serving the coder community and clients that require custom software development. The TCR curates a featured list of assets, where assets denote a digital resource that provides utility without further input from the knowledge worker [14].

5.5 Advertising AdChain registry provides whitelist of websites from which advertisers can read to assess whether or not to service inbound bid requests on ad opportunities, it is a smart contract on the ethereum blockchain which stores domain names accredited as non-fraudulent by adToken holders [5].

5.6 Medical MedCredits is a decentralized platform that connects patients to physicians for healthcare services. It maintains a physician registry in which doctors seek entry in order to be granted physician in the ecosystem including access to patient case data [15].

5.7 Userfeeds.io It is a content ranking a nd reputation system inspired by the ideas of liquid democracy and Google’s PageRank algorithm. It provides a token-based discovery layer similar to TCR where token holders become content curators and can use their token holdings to endorse the content on the web and back other curators [3].

5.8 Mapping FOAM is a distributed, decentralized, and permissionless protocol which is designed to provide spatial applications that bring geospatial or location data to blockchains and provide a consensus-driven map of the world. It uses the token mechanisms and incentives method as elements of protocol and allows the distributed users to coordinate and interact to provide an authentic and secured proof of location [16].

198

J. Kaur and B. Visveswaraiah

5.9 Video Paratii is an infrastructure that provides decentralized video applications. The TCR list is formed for intaking reputable videos and excluding non-reputable videos, TCR uses the propose-challenge mechanism and dynamic staking for curating the list [17].

5.10 Real Estate On imbrex, users will maintain ownership of their real estate data. It is a user-curated real estate marketplace [18].

5.11 Employment In this paper, the authors provide a TCR hierarchical structure for employment so that every authentic person with good qualification should receive a job [19].

6 TCR’S Open Challenges and Their Existing Solutions Problems in TCR are subjective to their use-cases; however, here we mention some of the general problems that may occur and ways to move ahead from them.

6.1 Chicken-Egg Problem Blockchain is a social signaling system. If no one is on the registry, then no one wants to apply to the registry; if no one applies, then no one is on the registry. We can solve this by building an incentive structure for early applicants and give them an additional stake in the system [20].

6.2 Passive Token Holding/Speculation TCRs are valued for their reputation, i.e., every token holder should have a stake in the decision-making process which increases the value of the list. So speculation questions the legitimacy of the list being curated. Here, we processed the initial distribution of tokens that is critical. Tokens should be initially distributed to entities

A Brief Survey of Token Curated Registries

199

who particularly care about curation task and have the expertise for them. For that, we could use vote delegation mechanism, where token holders are not rewarded if they don’t vote and they can delegate someone else can vote on their behalf who have required expertise [4, 21].

6.3 Coin-Flipping This problem arises either the voters have no incentive to choose sides or due to the task of scrutinizing the listee. it is already taken care of by incentivizing the majority block and this does not have a major impact assuming there is an even distribution of choices from coin-flippers [4].

6.4 Parametric Issues This is an implementation issue. The parameters in TCR are changed through the consensus among the token holders. This problem will be resolved as per the TCR development [4].

6.5 Order/Ranking in TCRs Ordering/Ranking in TCR is done in the same way as entries/responsibilities are being added to the applicants, where anyone can upvote/downvote/challenge any listee. This process should be done by authorized persons; otherwise, it creates very harmful situations. The solution to this is to attach the valuable resources to the voting person as money, reputation, identification, etc.

6.6 Subjective Curation TCRs should only be used when there is an objective idea over the list being curated. But what if for, e.g., if the list is about favorite genre of movies then every individual choice is subjective to his own, and then we convert this subjective problem to an objective one and these are incentives by majority blocks [20].

200

J. Kaur and B. Visveswaraiah

6.7 Economic Attacks As in TCRs, if the token holders are bribed to admit/reject an applicant or if large token holders collude to form a 51% attack or Sybil attack, then this can be resolved by the integration of identity into the system such that curators don’t just put money but also their reputation in stake [20].

6.8 Spam Attack When there are more numbers of the entries being added in a TCR which is economically beneficial to the curators is taken care of by the incentive mechanism of TCR where the reputation is valued, but in this, the demand of TCR tokens decreases thereby decreasing the overall value of curators [21, 22]. To resolve this issue, curators use some threshold or limit value at the number of entries.

6.9 Registry Poisoning This is an external attack not on the list but on the users, e.g., when a college gets listed it gets the reputation of the best college, this creates an opportunity for it to increase its tuition fee thus exploiting the user. This can be overcome by the regular check of token holders on any such malicious activities [4].

6.10 Real Time Coordination Problem In TCR, There should be cheap and effective coordination signals among honest players. But real-time coordination is almost impossible that’s why curators do this after a particular time interval.

6.11 Publicly Available Data TCR is working with only publicly available data. What if the data is stored privately?

6.12 Nesting of TCRs There is no standard for TCR yet. So, nothing to say about this. There is no smart contract platform which can do it [22].

A Brief Survey of Token Curated Registries

201

7 Conclusion and Future Work In this paper, we provide a basic study of the token curated list that is decentralized, token, and incentive-based blockchain registry which solves many problems of the current economic web and produces a high-quality curation list. It is the future of decentralized list. Currently, it has many practical implementations as adchain running at ethereum smart contract. In future, it will change the history and we will see more practical implementations on different blockchain platforms. In the early phase, it has various challenges or limitations but these will be solved by the technical development of TCR.

References 1. Christopher, T.: What are Token Curated Registries and decentralized lists? https:// hackernoon.com/what-are-token-curated-registries-and-decentralized-lists-d33fa42ba167 (2019). Accessed 4 July 2019 2. Zheng, Z., et al.: Blockchain challenges and opportunities: a survey. Int. J. Web Grid Serv. 14(4), 352–375 (2018) 3. Olpinski, M.: Building ‘Google For The Economic Web’ on The Ethereum Blockchain. https:// blog.userfeeds.io/building-google-for-the-economic-web-on-the-ethereum-blockchainde27cb3d23b (2016). Accessed 7 Oct 2018 4. Goldin, M.: Token-Curated Registries 1.0. https://medium.com/@ilovebagels/token-curatedregistries-1-0-61a232f8dac7 (2017). Accessed 8 Aug 2018 5. adChain: Introducing the adChain Registry! https://medium.com/metax-publication/ introducing-the-adchain-registry-cc5b8b831a7e (2017). Accessed 4 Oct 2018 6. Lockyer, M.: Token Curated Registry (TCR) Design Patterns. https://hackernoon.com/tokencurated-registry-tcr-design-patterns-4de6d18efa15 (2018). Accessed 8 Sept 2018 7. McConaghy, T.: The Layered TCR. https://blog.oceanprotocol.com/the-layered-tcr56cc5b4cdc45 (2018). Accessed 9 Aug 2018 8. Clark, R.: Framework-based Token Curated Registries. https://hackernoon.com/frameworkbased-token-curated-registries-9691e83c2c4c (2018). Accessed 11 Sept 2018 9. Robtoy, C.: The Different Types of Token Curated Registries. https://medium.com/ @coltonrobtoy/different-types-of-token-curated-registries-e1f6fc3ec1fc (2018). Accessed 12 Aug 2018 10. Goldin, M.: Token Curated Registries 1.1, 2.0 TCRs, new theory, and dev updates. https:// medium.com/@ilovebagels/token-curated-registries-1-1-2-0-tcrs-new-theory-and-devupdates-34c9f079f33d (2017). Accessed 22 Aug 2018 11. Iles, M., Ruiz, L., Coolidge, M., Vuong, N., Bode, N.: The Civil white paper. In: CIVIL Corp. https://civil.co/white-paper/ (2018) 12. Selkis, R., McArdle, D.: Messari Token-Curated Registry. In: messari.io. https://messari.github. io/tcr/whitepaper.pdf (2018) 13. Pon, B., McConaghy, T., Gossen, D.: Ocean protocol: a decentralized substrate for AI data & services. In: Ocean Protocol Foundation. https://oceanprotocol.com/tech-whitepaper.pdf (2018) 14. Draper, A.: Deconet: decentralizing the knowledge economy. In: Attention Inc. https://deco. network/whitepaper.pdf (2018) 15. Todaro, J., Praver, M.: MedX Protocol: launch unstoppable medical apps. In: Medical Exchange Protocols Ltd. https://medxprotocol.com/pdfs/medx-protocol--project-slides.pdf (2018)

202

J. Kaur and B. Visveswaraiah

16. King, R.J., Josefsson, K., Zavyalova, K.: FOAM—the consensus driven map of the world. In: Foamspace Corp. https://foam.space/publicAssets/FOAM_Whitepaper.pdf (2018) 17. Ana, F., Gerbrandy, J.: Paratii: a distributed attribution protocol and peer-to-peer video streaming engine. In: paratii.video. https://paratii.video/dist/assets/images/Paratii%20White %20Paper%20-%20v0.3%20-%20June%202018.pdf (2018) 18. King, S.: IMBREX white paper: a decentralized real estate data exchange & real estate transaction application. In: imbrex corp. https://about.imbrex.io/white-paper.html (2018) 19. Kaur, J., Visveswaraiah, B., Kalra, S.: EmployNet: a trusted tokenized employment platform for evolving economy. In: ResearchGate (2019). https://doi.org/10.13140/RG.2.2.10740.09609 20. Wang, Q.: Token-Curated Registries: Value Proposition and Challenges. https://medium. com/@QwQiao/token-curated-registries-value-proposition-and-challenges-5911466ed862 (2018). Accessed 17 Sep 2018 21. Ramsundar, B., et al.: Tokenized Data Markets. arXiv:1806.00139 (2018) 22. Arkratos Blockchain Solutions: AMA: All you need to know about Token curated Registry (TCR). https://medium.com/arkratos/ama-all-you-need-to-know-about-token-curatedregistry-tcr-dea6ab4b9ba6 (2018). Accessed 25 Sept 2018

Comparative Study of Effective Augmentation Method for Bangla ASR Using Convolutional Neural Network Md. Raffael Maruf, Md. Omar Faruque, Md. Golam Muhtasim, Nazmun Nahar Nelima, Salman Mahmood, and Md. Maiun Uddin Riad

Abstract Data scarcity is the main obstacle to get top-notch accuracy in neural network-based automatic speech recognition (ASR). To solve this problem, using data augmentation is a very familiar phenomenon nowadays. But in Bangla ASR, this technique is seldom used. This paper illustrates the benefits of data augmentation. In this work, the most effective augmentation methods have been explored for Convolutional Neural Network (CNN)-based Bangla word recognition system. A few types of augmentation methods have been implemented in this task, e.g., time stretching, background noise injection and pitch shifting. This experiment is performed on Linear Prediction Cepstral Coefficients (LPCC) and Mel Frequency Cepstral Coefficients (MFCC) feature extraction methods for comparative analysis. In LPCC extraction, slow-down augmentation is the most efficient one, whereas, in MFCC, positive pitch shifting augmentation is the hugely successful one at raising the accuracy rate. Overall, time-stretching is the most effective augmentation method that has consistently given better accuracy in both features. Contrariwise, noise injection is a less effective method in both cases. The consequence of using all these augmentation techniques is the escalation of accuracy up to 27.27%. To the best Md. R. Maruf · Md. O. Faruque (B) · Md. G. Muhtasim · N. N. Nelima · S. Mahmood · Md. M. U. Riad Department of Electrical & Electronic Engineering, Ahsanullah University of Science and Technology, Dhaka 1208, Bangladesh e-mail: [email protected] Md. R. Maruf e-mail: [email protected] Md. G. Muhtasim e-mail: [email protected] N. N. Nelima e-mail: [email protected] S. Mahmood e-mail: [email protected] Md. M. U. Riad e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_20

203

204

Md. R. Maruf et al.

of our knowledge, this research paper is a pioneer work to inquire about the most effective augmentation method for Bangla ASR. Keywords Convolutional neural network · Bangla word recognition system · Automatic speech recognition

1 Introduction Automatic speech recognition refers to a technology to utilize the voice with the computer and resemble the conversation. ASR technologies nowadays emphasize using voice as an interface for workplace automation hubs like lighting, audio-visual equipment, camera, security system, etc. Real-time ASR and AI systems perform to monitor trading room conversations and other financial advice conversations. ASR technology is exposing speech pathology system where patients can utilize the speech rehabilitation programs. Voiceprint biometric system becomes stabilized because of the speaker recognition procedure. While Bangla is the seventh most widely spoken language with 250 million speakers worldwide, the development of the Bangla ASR system trails behind others due to the limited number of research works conducted in this area, and this work seeks to fill in the gap. This paper proposes a convolutional neural network (CNN)-based speech recognizer for Bangla frequently used command words and digits. But it requires an enormous amount of data to train a convolutional neural network to achieve significant accuracy. The solution to this problem is to collect more data or manipulating the existing data to increase the size of the dataset. In this paper, we explored the effective audio data augmentation methods for MFCC and LPCC features to maximize prediction accuracy.

2 Literature Review Hybrid Deep Neural Network (DNN)—Hidden Markov Model (HMM)-based speech recognition system showed high accuracy. But CNN-based ASR outperformed DNN by reducing the error rate up to 10% in this paper [1]. Nithya Davis et al. [2] performed different data augmentation techniques on the same dataset, where it was found that LPCC-based augmentation had better performance than other augmentation methods. With data augmentation, it was found that it may increase accuracy by more than 5%. In this paper [3], ‘SpecAugment’ augmentation method was applied to the training dataset of a speech recognition system, where it had converted the ASR system from an over-fitting to an under-fitting problem. The achieved performance was 6.8% WER in LibriSpeech on test other without

Comparative Study of Effective Augmentation Method …

205

the use of a language model. Reference [4] worked on LPCC- and MFCC featurebased isolated Bangla word recognition using several classifiers, i.e., nearest neighbor (DTW), neural net (NN), word-based HMM (HMMw) and phoneme-based HMM (HMMp). They generated synthetic samples from real samples. They reported a significant increase in isolated word detection rate in all classifiers trained with both the samples compared to training with just the real samples. In this paper [5], a performance comparison between three different CNN models was shown for Bangla short speech command. The dataset was used with MFCC feature extraction in one model and raw in the other and the third approach used transfer learning where the MFCC model showed better performance than the others. Reference [6] presented MFCC feature-based Bangla speech recognition system using convolution neural network. They demonstrated an increase in classifier performance by varying dropout rate and using data augmentation techniques. They reported an accuracy of 86.058% for isolated speech to text conversion in the CNN model. An acoustic model for Bangla digit recognition was proposed [7] using DBN, which is a probabilistic generative ANN with feature detectors comprise of multiple layers. Comparing with the other prominent methods, the proposed model showed satisfactory accuracy and outnumbered the rest.

3 Methodology 3.1 Data Collection The dataset used in this research work consists of Bangla digit sequences and short speech commands. Eleven unique words uttered by 306 different speakers were recorded in a room environment containing both male and female speakers between the ages of 7 and 60 with a duration of 3 s. 60% of the whole dataset was used for training purposes. 20% of the dataset was allocated for validation and the rest for the test. Table 1 shows the isolated speeches that comprise the used dataset.

3.2 Data Augmentation Syntactic data for audio can be generated by applying various techniques. The techniques that have been used in the proposed work are listed below. 1.

2.

Time stretching: Time-stretch an audio series by a ‘Stretch factor’ while keeping the pitch unchanged. If factor >1, then the signal speeds up. If factor 0. For negative shifting that

206

Md. R. Maruf et al.

Table 1 Isolated speech words in the dataset

3.

Bengali words

English translation

Phonetic representation

One

Ek

Two

Dui

Three

Tin

Four

Char

Five

Pach

Start

Shuru

End

Shesh

Come

Asho

Go

Jao

Right

Dane

Left

Bame

simulates female voice, factor 20 and age 30

ANOVA or F test

Test multiple groups at the same time

More than 2 samples

Chi-Square test

Check if observed patterns (O) of data fit some given distribution (E) or not

Two categorical variables from a sample

calculated through the standard normal distribution “Z”:    X − μ   |Z| =  σ  The Z-score transformation is a linear transformation with μ = 0 and σ = 1. It is used for feature scaling. A normality test is used to check if a distribution is Gaussian or not. The normal distribution is symmetric about μ. So, the area to the left of μ is equal to the area to the right of μ. We have used hypothesis testing as a statistical analysis in this study. A hypothesis test estimates two mutually exclusive statements about a population to ascertain which statement is best supported by the trial data. The critical parameter of hypothesis testing is the null hypothesis (H0 ), and the alternative hypothesis (Ha) that directly contradicts H0 . The confidence factor (α) is used to decide whether to accept or reject an H0 . The value of “α” is usually kept as 0.05 or 5%, as 100% accuracy is impossible to achieve whether to accept or reject an H0 . Popular, widely used hypothesis testing methods, a short description, and the required sample size are demonstrated in Table 3. A hypothesis test can be either a one-tailed test or a two-tailed test. For each of the testing methods, resulting probability value (P-value) is compared with “α” to accept or reject a null hypothesis. But it suffers from type-I error (false positive) or type-II error (false negative) [13, 14]. “Shapiro–Wilk”, “D’Agostino’s Kˆ2”, “Anderson–Darling” test calculate P-value to decide if a sample looks like Gaussian (P-value > α = 0.05) or not (P-value < α = 0.05). Covariance (COV(x, y)) is a property of a function to retain its form when its variables are linearly transformed. It helps to measure correlation (rxy ) that measures the strength of the linear relationship between two variables. corr (x, y) = C O V (x, y)/(σ x ∗ σ y), wher e − 1 < r < +1 The “Sign” of “r” shows the direction of the relationship among two variables x and y. Table 4 shows the meaning of different |r| values. If two variables/features are strongly correlated, then we selected any one of them during feature selection. Pearson’s correlation coefficient is used to summarize the strength of the linear

Comparing Performance of Ensemble-Based Machine Learning Algorithms … Table 4 Statistical analysis methods on the selected datasets [17]

|r| value

Meaning

0.00–0.2

Very weak

0.2–0.4

Weak to moderate

0.4–0.6

Medium to substantial

0.6–0.8

Very strong

0.8–1

Extremely strong

259

Table 5 Statistical analysis methods on the selected datasets [8, 13, 14] No

Methods

Purpose

1

Mean, standard deviation, skewness

Distribution test

2

t-test, z-test, F-test, Chi-square

Hypothesis test

3

Shapiro–Wilk, D’Agostino’s Kˆ2, and Anderson–Darling test

Normality test

4

Covariance, correlation

Association test

5

Histogram, Swarm, Violin, Bee Swarm, Joint, Box, Scatter

Distribution plot

relationship between two variables in normal distribution and spearman’s correlation is used to calculate the non-linear relationship between two variables [12–14] (Table 5).

3.4 Model Training and Testing In this study, we have selected the following ensemble learning methods, such as bagging, boosting, and voting for multi-class classification analysis as described in Table 6. The description of the ensemble learning methods is described as follows [11, 14]—a. Bagging/Bootstrap Aggregation: It works well with the algorithms that have high variance, such as decision tree (DT), random forest (RF). In this method, the ensemble model tries to improve prediction accuracy and decrease model variance by combining predictions of individual models trained over randomly generated Table 6 Ensemble learning methods [11–14]

Ensemble learning methods

Ensemble algorithms

Bagging

Bagged decision tree

Bagging

Random forest

Bagging

Extra trees

Boosting

AdaBoost

Boosting

Stochastic gradient boosting

Voting

Voting ensemble

260

A. Chatterjee et al.

training samples. The final prediction of the ensemble model is given by calculating the average of all predictions from the individual estimators, b. Boosting: It combines several weak base learners, trained sequentially over multiple iterations of training data, to build powerful ensemble. During the training of weak base learners, higher weights are assigned to those learners which were misclassified earlier, and c. Voting: In this method, multiple models of different types are constructed on some straightforward statistical methods, such as mean or median and they are used to combine the predictions. This prediction serves as the additional input for training to make the final prediction. In our study, for the “voting” ensemble-based classification method we used seven classifiers as follows—(a.) SVM with kernel “RBF” or non-linear kernel, (b.) SVM with linear kernel, (c.) Gaussian NB, (d.) Decision Tree Classifier, (e.) Random Forest Classifier, (f.) KNeighbors Classifier, and (g.) Linear Discriminant Analysis. For other methods, such as “Bagging”, and “Boosting” we used decision tree models with n_estimators = [50, 100, 150, 200], and random_state = 7. The best value of the “n_estimators” was obtained following the “grid-search” method. The steps used to train and test an ensemble machine learning model are described below: [8] • Load data • Data pre-processing following the below steps: – – – – –

remove missing value from the loaded data encode categorical features check distribution of data and features correlation analysis among features and feature scaling if required shuffle the data

• Split data for training and testing (80:20) with some random state • Ensemble learning method selection as described in Table 6 based on classification problem statement • K-fold cross-validation on data (in our study, K = 5) • Evaluate model performance with metrics as described in Sect. 3.5 • Perform model tuning with “grid search” parameter optimization technique where required. N.B: a. The selection of learning rate (α): if too small then slow convergence in gradient descent (GD) and if too large then slow convergence in GD or GD may diverge. b. Let; “m” training samples have “n” features. If, too many features (m < = n), then delete some features or use regularization with regularization factor “λ”. c. If “λ” is too large then the algorithm fails to eliminate overfitting, or even sometimes underfit and GD fails to converge. “λ” (∞) increases to lead high bias and decreases to lead high variance. d. Underfitting results in high bias and overfitting leads to high variance.

Comparing Performance of Ensemble-Based Machine Learning Algorithms …

261

g. Gradient descent follows convex optimization technique with upper bound (L) and lower bound (μ) on curvature f: μId ≤ ∇ 2 f (x) ≤ L Id , wher e ∇ 2 f (x) is the H essian, μ > 0 and L = Li pschit zcontinious

3.5 Model Evaluation Developed ensemble learning methods for classification are evaluated with below metrics [8, 12–14]. Classification metrics: accuracy score, classification report, and confusion matrix. Classification report includes precision, recall and F1-score. A confusion matrix is a table with two dimensions “actual” and “predicted” and both the dimensions have “true positives (TP)”, “true negatives (TN)”, “false positives (FP)”, “false negatives (FN)”. • • • •

TP—both actual class and predicted class of data point is 1. TN—both actual class and predicted class of data point is 0. FP—actual class of data point is 0 and predicted class of data point is 1. FN—actual class of data point is 1 and predicted class of data point is 0. Formulas for calculating classification metricss are stated as below: (T P + T N ) TP , Pr ecision(P) = , T P + FP + FN + T N (T P + F N ) TP Recall (R) or Sensitivit y(S) = , T P + FN PR TN , F1scor e = 2 ∗ Speci f icit y = (1 − Sensitivit y) = T N + FP P+R Accuracy =

Accuracy tells how close a measured value is to the actual one. Precision determines how close a measured value is to the actual one. Recall or sensitivity defines the total number of positives (actual) returned by the ensemble machine learning model.

3.6 Assessment of Body Composition BMI has been used to categorize different weight groups in adults of 20 years or older, both male and female: [4, 8] • Underweight: BMI < 18.5 • Normal weight: BMI is 18.5–24.9

262

A. Chatterjee et al.

• Overweight: BMI is 25–29.9 • Obese: BMI is 30 or more.

4 Results and Discussion Analysis of “BMI” dataset with 500 records reveals that index (body composition) has a strong correlation with BMI as depicted in Fig. 1. BMI column was externally added during data pre-processing and later, was removed during model training due to high correlation. The dataset has six classes for classification under index field—extremely weak, weak, normal, overweight, obesity, and extreme obesity. In multiclass classification, the “voting”-based ensemble outperformed other methods as described in Table 6 with an accuracy score of 89% as depicted in Fig. 2. We added extra feature “body_composition” to the “insurance” dataset with 1338 records based on the BMI feature, and the feature classifies the records among four classes—underweight, normal weight, overweight, and obese. We encoded the categorical features such as sex, smoker, and region. There is a strong correlation between “smoking” and “charge” with |r| = 0.79. Smoking grows negative health behavior in humans. Negative health behavior has a great impact on weight change, obesity, and overweight. So, excess smoking does not only create an active negative impact on health but also creates a passive negative impact on economic position. We have used insurance data for multi-class classification analysis.

Fig. 1 Correlation heatmap of “BMI” data [8]

Comparing Performance of Ensemble-Based Machine Learning Algorithms …

GradientBoosngClassifier

263

84.5

AdaBoostClassifier

46

ExtraTreesClassifier

87.25

RandomForestClassifier

84.5

BaggingClassifier

87.75

Vong

89 0

20

40

60

80

100

Fig. 2 Classification accuracy (%) of ensemble learning models to classify ‘BMI’ data

During classification, we used “body_composition” as a predicted feature, and the accuracy of ensemble learning models is depicted in Fig. 2 and “GradientBoostingClassifier” has performed the best with 99.9% accuracy, n_estimators = 100, random_state = 7, following grid search method (Fig. 3). For the regression, we used “charges” as predicted feature and performed hypothesis testing with “ANOVA” results to retain Ha = {a significant change between the three age categories (young adults, senior adults, elders) with “BMI”} with a P-value of 0.001, 0.060, and 0.000 respectively. The boxplot (left) in Fig. 4 demonstrates that the obesity risk increases with age, and the mean BMI for the three-age categories is in the obesity range, which is a risk. The regression plot (right) in the same figure depicts how insurance charge increases with obesity. Insurance charge also increases with smoking condition and age as depicted in Fig. 5, and Fig. 6, respectively. “Eating-health-module-dataset” with 11,212 records was processed with ensemble ML classification algorithms to classify records in between four classes “underweight”, “normal weight”, “overweight”, and “obese” under “body_composition” feature and the “BaggingClassifier”, “AdaBoostClassifier”, and

GradientBoosƟngClassifier

99.9

AdaBoostClassifier

83.61

ExtraTreesClassifier

94.5

RandomForestClassifier

99.15

BaggingClassifier

99.81

VoƟng

98.78 75

80

85

90

95

100

105

Fig. 3 Classification accuracy of ensemble ML models to classify “body_composition” insurance data

264

A. Chatterjee et al.

Fig. 4 a Relationship in between “Age category” and “BMI” [8]; b Relationship in between “BMI” and “charges” [8]

Fig. 5 Relationship in between “charges” and smoking condition

Fig. 6 Relationship in between “charges” and “age”

“GradientBoostingClassifier” classifiers performed the best as depicted in Fig. 7 with accuracy = 99.9%, n_estimators = 50, and random_state = 7, following grid search method. The regression analysis in Fig. 8 shows that sweet beverages, economic condition, fast food, sleeping, meat and milk consumption, drinking habit, exercise has a sharp

Comparing Performance of Ensemble-Based Machine Learning Algorithms …

265

GradientBoosƟngClassifier

99.9

AdaBoostClassifier

99.9

ExtraTreesClassifier

85.43

RandomForestClassifier

99.74

BaggingClassifier

99.9

VoƟng

98.51 75

80

85

90

95

100

105

Fig. 7 Classification accuracy of ensemble ML models to classify “body_composition” data

Fig. 8 Regression analysis of selected parameters with “body_composition” in eating health module dataset [18]

266

A. Chatterjee et al.

Fig. 9 Regression analysis of diabetes dataset to show a positive relationship between weight change and age, BMI, blood glucose, and blood pressure

impact on growing obesity in human. The regression analysis of “Pima-Indiansdiabetes-database” dataset with 768 records resulted in a positive dependency on between weight change and discovered factors such as blood glucose, blood pressure, and age as depicted in Fig. 9. We used ensemble ML classification algorithms to classify records among two classes “obese” (1) and “non-obese” (0) under feature column “outcome” and “Voting”-based ensemble method outperformed other classifiers as depicted in Fig. 10. This analysis shows a strong relationship between obesity and diabetes. The regression analysis of “cardiovascular-disease-dataset” with 462 records shows, in Fig. 11, that blood pressure, tobacco consumption, lipid profile, adiposity, family history, obesity, drinking habit, and age have a strong connection with CVDs. In binary classification problems on the used heart dataset, “GradientBoostingClassifier” outperformed other classifiers with accuracy = 70.7%, n_estimators = 50, and random_state = 7, following grid search method as depicted in Fig. 12. From the above data analyses, we observed how different ensemble-based machine learning classification methods are performing on different publicly available health datasets! The “Voting”, and “Gradient Boosting” ensemble classification machine learning method offered consistent performance on every dataset. The identified risk factors of obesity from the above analyses can be summarized as—a. BMI, b. age, c. tobacco consumption, d. sweet beverages, e. economic condition, f. fast food, g. sleeping pattern, h. diet, i. blood pressure, j. blood glucose, k. lipid profile, l. adiposity, m. exercise, and n. family history.

Comparing Performance of Ensemble-Based Machine Learning Algorithms …

GradientBoosƟngClassifier

267

76.38

AdaBoostClassifier

76.21

ExtraTreesClassifier

76.7

RandomForestClassifier

76.5

BaggingClassifier

76.86

VoƟng

78.17 75

75.5

76

76.5

77

77.5

78

Fig. 10 Classification accuracy of ensemble ML models to classify diabetes dataset

Fig. 11 Regression analysis of cardiovascular disease dataset

78.5

268

A. Chatterjee et al.

GradientBoosƟngClassifier

70.7

AdaBoostClassifier

68.82

ExtraTreesClassifier

67.74

RandomForestClassifier

69.1

BaggingClassifier

66.92

VoƟng

66.65 64

65

66

67

68

69

70

71

72

Fig. 12 Classification accuracy of ensemble ML models to classify cardiovascular dataset

5 Conclusion This study aims to identify probable risk factors associated with obesity, after studying existing health datasets, publicly available in “Kaggle”, and “UCI” with statistical and ensemble learning methods. The trouble we found with the size of data, target population, and data integration. The statistical analysis has given a clear indication of the potential risk factors to be addressed and further studied. In the future, the study can be extended with non-convex optimization (artificial neural net, deep learning). Our future study aims to design, develop, test, and evaluate the performance of an intelligent eCoach system for automatic generation of personalized, contextual behavioral recommendations to address health and wellness challenges related to obesity. For the same, we will collect health data related to associated risk factors from controlled participants over time, as identified from this study.

References 1. Butler, É.M., et al.: Prediction models for early childhood obesity: applicability and existing issues. In: Hormone Research in Paediatrics, pp. 358–367 (2018) 2. Singh, B., Tawfik, H.: A machine learning approach for predicting weight gain risks in young adults. In: 2019 10th International Conference on Dependable Systems, Services and Technologies (DESSERT), pp. 231–234 IEEE (2019) 3. Grabner, M.: BMI trends, socioeconomic status, and the choice of dataset. In: Obesity Facts, pp. 112–126 (2012) 4. WHO page. https://www.who.int/news-room/fact-sheets/detail/obesity-and-overweight 5. Csige, I., Ujvárosy, D., Szabó, Z., L˝orincz, I., Paragh, G., Harangi, M., Somodi, S.: The impact of obesity on the cardiovascular system. J. Diabet. Res. (2018)

Comparing Performance of Ensemble-Based Machine Learning Algorithms …

269

6. Gerdes, M., Martinez, S., Tjondronegoro, D.: Conceptualization of a personalized ecoach for wellness promotion. In: Proceedings of the 11th EAI International Conference on Pervasive Computing Technologies for Healthcare, pp. 365–374 (2017) 7. Chatterjee, A., Gerdes, M.W., Martinez, S.: eHealth initiatives for the promotion of healthy lifestyle and allied implementation difficulties. In: 2019 International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob), pp. 1–8. IEEE (2019) 8. Chatterjee, A., Gerdes, M.W., Martinez, S.G.: Identification of risk factors associated with obesity and overweight—a machine learning overview. Sensors 20(9), 2734 (2020) 9. Padmanabhan, M., Yuan, P., Chada, G., Van Nguyen, H.: Physician-friendly machine learning: a case study with cardiovascular disease risk prediction. J. Clin. Med., 1050 (2019) 10. Selya, A.S., Anshutz, D.: Machine learning for the classification of obesity from dietary and physical activity patterns. In: Advanced Data Analytics in Health, pp. 77–97. Springer, Cham (2018) 11. Jindal, K., Baliyan, N., Rana, P.S.: Obesity prediction using ensemble machine learning approaches. In: Recent Findings in Intelligent Computing Techniques, pp. 355–362. Singapore (2018) 12. Schapire, R.E., Freund, Y.: Boosting: foundations and algorithms. In: Kybernetes (2013) 13. Brandt, S.: Statistical and computational methods in data analysis. No. 04; QA273, B73 1976. In: Amsterdam: North-Holland Publishing Company (1976) 14. Sklearn page. https://scikit-learn.org/stable/supervised_learning.html 15. Kaggle data page. https://www.kaggle.com/data 16. Eating-health-module-dataset description. https://www.bls.gov/tus/ehmintcodebk1416.pdf 17. Chatterjee, A., Gerdes, M.W., Martinez, S.G.: Statistical explorations and univariatetimeseries Analysis on COVID-19 datasets to understand the trend of disease spreading and death. Sensors 20(11), 3089 (2020) 18. Python page. https://docs.python.org/

Krylov Subspace Method Using Quantum Computing Vidushi Jain and Yogesh Nagor

Abstract The Krylov subspace algorithm uses iterative methods to solve bulky linear equations. It has a time complexity of O(n2 ) when run on a classical computing machine. The proposed √ algorithm has an exponential speedup over the classical algorithm. It takes O(n k) time complexity when run on a quantum computer, where k is the condition number. In this paper, we will discuss an in-depth comparison between classical Krylov subspace algorithm and quantum Krylov subspace algorithm. A brief introduction of quantum computing is presented. Further, we will understand quantum Krylov subspace algorithm through quantum entanglement feature of quantum computing. The proposed algorithm using quantum phase estimation will provide deterministic solutions to the linear vector subspace. The paper will be closed by a quantum logic gate implementation of the quantum algorithm and will be concluded by the results. Keywords Krylov subspace method · Quantum algorithms · Quantum entanglement · Superposition

1 Introduction Quantum computers are based on the quantum mechanical phenomenon such as entanglement and superposition which in turn enables them to produce an exponential speedup over classical computers. For instance, while factoring large numbers as well as while processing large datasets quantum algorithm works exponentially better than the classical ones [1]. Quantum computing is the field of natural sciences that might change the way we think about computers. It has a wide range of applications from the field of artificial V. Jain (B) · Y. Nagor Department of Computer Science and Engineering, Medi-Caps University, Indore, Madhya Pradesh, India e-mail: [email protected] Y. Nagor e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_27

271

272

V. Jain and Y. Nagor

intelligence to the field of chemistry. Quantum computers can store and operate on exponentially more information than the units of information it possesses. For example, if a classical computer has to find the value of a variable ‘x’ from a very large dataset of an extremely large database, it has to sort through all the datasets, but a quantum computer can do the same instantly as a qubit of quantum computers, in superposition, can be in multiple places simultaneously. Therefore, it provides us a solution toward high-dimensional problems.

1.1 Qubits The information is stored on a quantum computer as qubit which is the fundamental unit of information whose value lies between |0 and |1. The classical computer’s fundamental unit of information is called a bit whose value can be either 0 or 1. The bit system comprises x such that x ∈ {0, 1}. Any integer value can be represented i as the string of these bits in the form N i=0 ki 2 and real numbers can be represented up to some decimal precision. In contrast, the qubits can take any superposition of {|0, |1}, given they satisfy: |ψ = x |0 + y |1, where |x|2 + |y|2 = 1, x, y ∈ C such that C2 is the associated Hilbert space. In other words, |ψ is from the subspace of C2 modulo the equivalence relation (x, y) ∼ (αx, αy) for some non-zero, complex α [2]. A special case of qubits are ‘cbits’ which are a pair of two-state orthonormal vector which can be denoted by the symbols:     0 0 |0 |1 = 1 1 Multi-qubit systems are tensor products of single-qubit systems, like with cbits. The tensor product is represented by matrices [3]. The Hadamard gate is a single-qubit operation that maps the pair of orthonormal two vectors:     |0 − |1 |0 + |1 {|0} to and |1 to √ √ 2 2 thereby creating an equal superposition of the pair. The Hadamard gate also takes a qbit in exactly equal superposition, and transforms it into a 0- or 1-bit [4]. Thus, we can easily structure quantum computation in a deterministic manner rather than a probabilistic one. Mathematically, Hadamard gate,   1 1 1 H= √ 2 1 −1

Krylov Subspace Method Using Quantum Computing

273

1.2 Deutsch-Jozsa Oracle In the Deutsch-Jozsa problem, we can use an oracle to determine some function f: {0,1}n → {0,1} is constant, i.e., 0 on all outputs or 1 on all outputs or is balanced, i.e., it returns 1 for half of the input domain and 0 for the other half. The function f takes n-digit binary value as input [5]. The entanglement property of a quantum computing system provides deterministic solutions, thereby providing an exponential speedup over a classical computing system. With the help of entanglement property, the Deutsch-Jozsa functions f(0) ⊕ f(1) and f(0) ⊕ g(0) are calculated simultaneously. To produce deterministic solutions we use bichromatic fields which produce entangles states of trapped ions [6, 7]. The proposed algorithm uses quantum phase estimation to provide deterministic solutions for the x(k) vectors of the linear subspace A(k) x0 . This research paper comprises related work done on the Krylov subspace algorithm. Then in the next section, classical algorithm is presented. The proposed quantum algorithm for Krylov subspace method is discussed in Sect. 4 with a quantum logic gate realization of the algorithm. Finally, the analysis of the given quantum algorithm is provided.

2 Related Works Basically, in Krylov subspace method we find x(k) vectors for subspace A(k) x0 . The Krylov subspace method uses iterative methods to solve bulky linear equations [8]. Mathematically, Aw = b, where A is a non-singular mxm matrix, b is a m-vector where m is very large. The matrix A generates the Krylov subspace and b is the linear subspace intersected by the images of the sequence:   Kr = b, A2 b, A3 b, . . . Ar−1 b The Krylov method moves toward an approximate value wr from a vector space w0 + Kr by constraining (b − Awr ) ⊥ (Lr ), where Lr is a vector space of dimension r, w0 is an initial assumption to the value of the solution [9]. In Krylov subspace methods, Kr = Kr (A, s0 ), where so = b − Aw  r Kr = r0 , A2 s0 , A3 s0 , . . . Ar−1 s0

274

V. Jain and Y. Nagor

By choosing different subspace Lr and preconditioning the system, we can produce Krylov subspace methods in different forms. The following techniques come out by choosing different subspace Lr : FOM: Lr = Kt GMRES, MINRES: Lr = AKr

2.1 Arnoldi’s Method 1.

Given a Krylov subspace represented by the vectors y1 , y2 , …, ym forming an orthonormal subspace, the algorithm is iterated till m steps [10]. Km = span of{y1 . . . ym } = {c1 v1 + c2 v2 + · · · + cm ym }

2. 3.

At any step i, a breakdown occurs where the Arnoldi’s method projected on the vectors y1 , y2 , …, ym will be exact. The matrix of the vectors y1 , y2 , …, ym is represented by Vm Hermitian matrix which is a n × m matrix.

3 Classical Algorithm The classical algorithm for Krylov subspace method for solving system of equations is presented below which is to be later compared with the quantum algorithm for Krylov subspace method. In the steps given below, M is the preconditioning matrix. The matrix M cannot change from iteration to iteration. It should be symmetric, fixed and positive-definite. The behaviour of the preconditioned conjugate gradient method may become unpredictable if these assumptions on the matrix M are violated. Given a preconditioned conjugate gradient, the classical algorithm constitutes the following steps: 1. 2. 3.

We preprocess then the preconditioner Ad is computed. Initiate: s0 : = b -Aw0 , q0 : = y0 : = M-’s0 Repeat the following steps till convergence: (i) (ii) (iii) (iv) (v) (vi)

x: = A*qi ∝i : = (si , yi )/(x, qi ) wi+1 : = wi + ∝i *qi si+1 : = si − ∝i *xi yi+1 : = M−1 si+1 βi : = (si+1 , yi+1 )/(si , yi )

Krylov Subspace Method Using Quantum Computing

275

Fig. 1 Convergence of the ritz values to the highest eigenvalues

(vii)

qi+1 : = yi+1 + βi *qi

The classical algorithm presented above has a time complexity of O(n2 ). The Krylov subspace method when run on a classical computer requires years of time for a very large vector space (Fig. 1).

4 Proposed Algorithm The algorithm proposed by us finds x(k) vectors for subspace A(k) x0 using quantum entanglement. Hamiltonian simulation techniques are used to apply the unitary operator eiAt to |b for a superposition at different time t [11, 12]. The technique is to decompose |b into the eigenbasis of A and to find eigenvalues λ using the quantum phase estimation. Basically, entanglement of bits is a property by whose virtue we can find the value of one bit of the entangled pair if we know the value of the other. For an entangled bit pair, if the value of the one bit of the pair is 0, then that of the other bit will be 0 as well. We can find x(k) vectors for subspace A(k) x0 by using properties. The proposed algorithm constitutes the following steps as illustrated in Fig. 2: 1.

Find (x)k vectors, subspace Ak x0 ,

276

V. Jain and Y. Nagor

Fig. 2 Quantum logic gate implementation of the proposed algorithm

2.

3.

4.

x0 , Ax0 , AK−1 x0 , …, Ak x0 span of {v1 … vm} = {c1 v1 + c2 v2 + ··· + cm vm } We find eigenvector that lies in the k + 1-dimensional space {v1 , v2 , …, vm } {x(k) } = {cx(k) }: c m (C} is a one-dimensional space. Krylov matrix, Km = {u, Au … Am−1 u} all the other dimensions as well depend on A and the starting vector u. Krylov subspace = {u, Au … Am−1 u}, Km is nested, i.e., K1 ⊂ K2 ⊂ K3 …Kn where K1 = c1 u + 0Au + · · · K2 = u, Au, A(Au), . . . and so on. So, we use various iterations. X = Z1 u + Z2 (Au) + ··· + Zm (Am−1 u) = Km z, where A is non-singular. Replace the action of A over all of Rn with its approximate behaviour over the lower dimensional Km , for m = 1, 2, 3,… Approximate problem using: minx mRn b − Ax2 x = Km z b – Ax = 0

5.

For the solution x = A−1 b We solve n × n Hermitian matrix A and a unit vector b,

1. 2.

|b = bi |i, i = 1 Hamiltonian simulation techniques are used to apply the unitary operator eiAt to |b for a superposition at different time t. The ability to decompose |b into the eigenbasis of A and to find eigenvalues λ using the quantum phase estimation.

Krylov Subspace Method Using Quantum Computing

277



Nj=1 Bj uj λj , j = 1

uj = A−1 |b = |x Ni=1 Bi λ−1 j Example A=rand(50) + 50*eye(50) b=ones(50,1); K=b; For m=1:10 Z=(A*K)\b; sum=K*Z; resid(m)=norm(b-A*xm ) K(:, m+1)=A*k(:, m);

5 Result The block diagram of the given system of quantum Krylov subspace algorithm is presented. The first stage is the quantum phase estimation stage in which we decompose |b into eigenbasis of A and to find eigenvalues λ. The second stage is non-unitary map. Hamiltonian simulation techniques are used to apply the unitary operator eiAt to |b for a superposition at different time t [13]. As all quantum systems are reversible, we have the reverse phase estimation [14]. We can find x(k) vectors for subspace A(k) x0 by using the properties of entanglement and phase estimation. The proposed algorithm uses quantum phase estimation to provide deterministic solutions for the x(k) vectors of the linear subspace A(k) x0 . The given graph shows the time complexities of both the algorithms clearly proving how quantum approach is better than the classical one. Krylov subspace algorithm requires a time complexity √ of O(n2 ) when run on a classical computing machine while it takes O(n k) time complexity when run on a quantum computing machine (Fig. 3).

6 Conclusion In this paper, we discussed a comparison between the existing classical Krylov subspace algorithms and quantum Krylov subspace algorithm through quantum entanglement feature of quantum computing. We created a quantum logic gate realisation of the Krylov subspace algorithm. We were able to achieve an exponential speedup of Krylov subspace algorithm requiring a time √ complexity of O(n2 ) when run on a classical computing machine while it takes O(n k) time complexity when run on a quantum computer, where k is the condition number. The proposed algorithm used quantum phase estimation to provide deterministic solutions for the x(k)

278

V. Jain and Y. Nagor

Fig. 3 Time complexity of quantum versus classical algorithm

vectors of the linear subspace A(k) x0 . The issue with quantum systems is the stability of qubits as a result of which the systems are prone to errors in the form of decoherence. Quantum computers have a wide range of applications ranging from chemical science to natural sciences and have the ability to revolutionize the way we see the world.

References 1. Kais, S.: Introduction to quantum information and computation for chemistry. Adv. Chem. Phys. (2014) 2. Mastriani, M.: Quantum Boolean image denoising. Quantum Inf. Process. (2014) 3. Guzman, H., Maziero, J., Sarandy, M., Céleri, L., Serra, R.: Quantum and classical thermal correlations in the XY spin-1/2 chain. Phys. Rev. 82, 012106 (2010) 4. Pechukas, P., Light, J.C.: On the exponential form of time displacement operators in quantum mechanics. J. Chem. Phys. (1966) 5. Gangopadhyay, S., Behera, B.K., Panigrahi, P.K.: Generalization and demonstration of an entanglement-based Deutsch–Jozsa-like algorithm using a 5-qubit quantum computer. Quantum Inf. Process. 17(7) (2018) 6. Sørensen, A., Mølmer, K.: Entanglement and quantum computation with ions in thermal motion. Phys. Rev. 62(2) (2000) 7. Satyajit, S., Srinivasan, K., Behera, B.K., Panigrahi, P.K.: Nondestructive discrimination of a new family of highly entangled states in IBM quantum computer. Quantum Inf. Process. 17(9) (2018) 8. Beerwerth, R., Bauke, H.: Krylov subspace methods for the Dirac equation (2004). Arxiv 1407.7370 9. Cullum, J., Willoughby, R.A.: Computing eigenvalues of very large symmetric matrices-an implementation of a Lanczos algorithm with no reorthogonalization. J. Computat. Phys. (1981) 10. Guan, X., Noble, C., Zatsarinny, O., Bartschat, K., Schneider, B.: ALTDSE: Arnoldi-Lanczos program to solve the time dependent Schrödinger equation. Comput. Phys. Commun. (2009) 11. Saad, Y.: Numerical Methods for Large Eigenvalue Problems. SIAM (2011)

Krylov Subspace Method Using Quantum Computing

279

12. Cross, A.: The IBM Q experience and QISKit open-source quantum computing software. Bull. Am. Phys. Soc. 63 (2018) 13. Oliveira, D.S., Ramos, R.V.: Quantum bit string comparator: circuits and applications. Quantum Comput. Comput. (2007) 14. Wang, Y., Li, Y., Yin, Z.: 16-qubit IBM universal quantum computer can be fully entangled. npj Quantum Inf. 4 (2018) 15. Ruf, M., Bauke, H., Keitel, C.H.: A real space split operator method for the Klein-Gordon equation. J. Comput. Phys. (2009)

An IoT-Based Cryptographic Algorithm: Mapping the Credentials for Micro-application in Single Column Kishore Kumar, Subhranil Som, Sarvesh Tanwar, and Shishir Kumar

Abstract In today’s world, millions of people reach different websites, and if someone wants to access anything, they need to use the credential principle for highly secured applications. The latest trend of most of the web developers and developers of mobile applications is to use two text fields for both the user id and password. In this paper, an algorithm describes the mapping of the user id and password for the SQL Lite database in a single text field. Besides, the user can put the id with maximum of 40 characters and also place the password of 30 characters in the same textbox. Furthermore, the id is visible to all users, but the user cannot see the next 30 characters of a password which plays a vital protective role. The result of this algorithm thus requires only one textbox for both id and password, and can be saved in a single column; with Android OS mapping id and password, the complexity of space and time will be minimized. Keywords Text field class · Algorithm · IoT · Encryption · Android OS · SQL lite database

K. Kumar (B) · S. Som · S. Tanwar Amity Institute of Information Technology, Amity University, Noida, UP 201313, India e-mail: [email protected] S. Som e-mail: [email protected] S. Tanwar e-mail: [email protected] S. Kumar Dept. of Computer Science and Engineering, Jaypee University of Engineering and Technology, Guna, MP 473226, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_28

281

282

K. Kumar et al.

1 Introduction A significant number of credential encryption schemes were introduced in the last decades. Most of these schemes achieve a high-security standard, but due to their complex mechanism and slow speed make them unusable in real-time applications. Motivated by that, we are proposing a new technique of effective and high-speed text encryption. The input is provided as a one-line plain text editor control in which the text is blurred so that it is not readable, usually by replacing each character with a symbol such as “*” asterisk [1]. Android OS is focused on helping to make the most of the latest technology while at the same time ensuring that users’ privacy and security remain a top priority. Therefore, there are many updates to Android to protect their privacy. Such enhancements have implications for different framework activities or how data can be processed [2]. An algorithm for mapping of id and password in a single-column methodology consists of mechanisms to assure that if we put the id and password in a single column the time and space complexity automatically reduces. Nowadays, millions of people hit several websites and if someone wants to access something they must use the concept of id and password to secure the application, in different textboxes. But using this algorithm only one textbox is required for both id and password. After that, this will be stored in a string, and then several programs are written in different languages that send this data into the database. Simultaneously, in the database also only one column is required for both id and password. This section explains that data handling is important for data service providers and applications to gather large quantities of IoT-related data generated by an enormous number in real time and that must be accessed securely [3]. Here, the authors introduce, in [4, 5], access-control strategy that should be continuously implemented to avoid unauthorized programs from accessing sensitive resources. This includes permission, verification of data processing, privacy protection, security audit and policies at the application layer; security requirements on the network layer inspire the use of authorization, authentication, use of data and signaling data confidentiality systems, signaling policies for integrity protection, etc.; and cryptographic algorithm on the device level includes identity validation of user confidentiality and data integrity protection strategies. The advantages of the proposed algorithm are: • • • • • •

Time complexity will be reduced. Space complexity also gets reduced. Its benefit will directly be seen by digital marketing customers. Users start getting benefits at an early stage. Risk is very less. Confidence gets established in the application.

An IoT-Based Cryptographic Algorithm: Mapping …

283

2 Algorithmic Concept and Description 2.1 Interference and Algorithm Robustness This phase is mainly focused on the logical part, such as database, time complexity, and space complexity. It works on unstructured databases like MySQL, Derby, and Oracle. In real encrypted communications, any natural or intentionally added noise or data loss often distort the transmitted ciphertext. It can also save resources and guard against theft, and make power supply more secure [6].

2.2 Configuration and Review When a user enters id and password, the program can compute the hash value of the password entered and equate it to the value stored. Confidentiality, rather than shielding data from violations of privacy, is a clear example of protection of the information flow. Security is enforced in this model by controlling accesses so that any subject can only access objects which are classified at the same level for which the subject has clearance, or for a lower level [7]. A program takes in some character string and outputs another character string. How does the algorithm now measure the input size? The natural response is that the length of the input string is measured [8]. For operating systems, this non-reversible encryption technique is applied [9]. To be effective in real-world communications, a successful cryptosystem must, therefore, be able to push through the obstacle [10]. The easiest concept can also be implemented using most demanding language, i.e., JavaScript for logical implementation and “WeSetupYourWebViewApp” as an iOS and Android development tool. The algorithm is to learn how to communicate and build in previously unthinkable ways for all sorts of applications to produce faster results and better outcomes [11].

2.3 Usage

2.4 User Interface Input-Textfield usually functions much like other textual input boxes; the only difference is to mask the content to prevent users from reading the password near the input

284

K. Kumar et al.

Fig. 1 Input textfield

field as shown in Fig. 1. Hence, most of the major vendors of databases and applications simply adopt the standard and natively support encryption, leaving the user with the key management [12].

3 Proposed Methodology 3.1 Single-Textfield Performance The algorithm measures the length of the input string. The measurement of an algorithm’s running time involving an input string in terms of the string length includes an input number: it measures the number of digits in that number. This is log(n) (up to a constant). Also, the running time in large O terms generally measures the worst-case running time: the longest it could take for a given length input [8]. The workflow of the proposed framework is shown in Fig. 2.

3.2 Text Encryption Algorithm. The input string is processed by the proposed algorithm as below: • Creating a class, constructor, or method for a class (onCreate()),

Start

TextField

Passcode method

Processing method

Successful Database Comparing both Input and Stored String

End Error Message

Fig. 2 The workflow of the proposed algorithm

An IoT-Based Cryptographic Algorithm: Mapping …

285

• This method accepts the 80 characters in the TextField. Then the passcode encryption method will call up to 40 characters, • A string variable is declared which stores the user’s data input in the TextBox, • Then the program is written in any of the high-level languages that send the data to the database. And, eventually, users get permission to access or deny the application.

3.3 Information Entropy Analysis A fundamental assumption in cryptanalysis is to keep the plaintext secret from intruders trying to get some plaintext information. Adversaries can also be involved, as discussed earlier, and seek to change the message. It is assumed then that cryptography will guarantee the confidentiality of the messages. Adversaries are believed to have complete access to the medium of communication [9].

4 Results and Discussion 4.1 Comparative Study of the Proposed Model Acceptability of the method is to be widely used for very quick processing and encryption of credentials, as shown in Table 1, according to Wikipedia. A string is created in the data encryption process which allows users to encrypt and decrypt Table 1 Experimental result of the frequency of letter Letter

Count (%)

Rank

Letter

Count (%)

Rank

E

12.70

1

M

2.40

14

T

9.05

2

W

2.36

15

A

8.16

3

F

2.22

16

O

7.50

4

G

2.01

17

I

6.96

5

Y

1.97

18

N

6.74

6

P

1.92

19

S

6.32

7

B

1.49

20

H

6.09

8

V

0.97

21

R

5.98

9

K

0.77

22

D

4.25

10

J

0.15

23

L

4.02

11

X

0.15

24

C

2.78

12

Q

0.09

25

U

2.75

13

Z

0.07

26

286

K. Kumar et al.

the data when they need to access it. The protection of this data became a growing problem studied by many experts and researchers [10]. It must be secured so that malicious or unauthorized users do not have access to the data. Here, string management is important for an effective encryption solution and most industry regulations either mandate it or strongly suggest it [12].

4.2 Information Exchange Simulation The outcome is a safety classification that contains both the sensitivity mark and thus when a user or application has to access the encrypted data, they simply need to authenticate that they are allowed to access the data. From there, SQL Lite database uses the master key to decrypt the tablespace key and the data is decrypted using the tablespace key. This cycle is never seen by the end-user; it is invisible to them [12], as shown in Table 1.

5 Conclusion The study described in this paper is related to the fields of requirements and methods. Much work is done in both the fields to study how the TextBox contains both the id and password. The overall concept describes the method and class which help in software upgradation and also by using this code time and space complexity will be maintained. There are several cases where proof of one’s identity is required. Typical situations include signing into a computer, gaining access to an electronic banking account, or withdrawing money from an automated teller. Older systems enforce user authentication using passwords or PINs. Though used effectively in many settings, there are also limitations in those approaches. For instance, someone needs to offer a password to verify the potential to use that password to impersonate it. The interactive method of proof naming systems offers a new form of identification to users.

References 1. MDN Web Docs: Web technology for developers. https://developer.mozilla.org/en-US/docs/ Web/HTML/Element/input/password (2020). Accessed 30 May 2020 2. Android developers: Android Architecture. https://developer.android.com/guide/basics/whatis-android.html (2019). Accessed 5 Dec 2019 3. Lee, G.M., Crespi N., Choi, J.K., Boussard, M.: Internet of Things. In: Bertin, E., Crespi, N., Magedanz, T. (eds.) Evolution of Telecommunication Services. Lecture Notes in Computer Science, vol. 7768. Springer, Berlin, Heidelberg (2013). https://doi.org/10.1007/978-3-64241569-2_13

An IoT-Based Cryptographic Algorithm: Mapping …

287

4. Pathan, A.K., Fadlullah, Z.M., Choudhury, S., et al.: Internet of Things for smart living. Wireless Netw. (2019). https://doi.org/10.1007/s11276-019-01970-3 5. Harbi, Y., Aliouat, Z., Harous, S., et al.: A Review of Security in Internet of Things. Wireless Pers. Commun. 108, 325–344 (2019). https://doi.org/10.1007/s11277-019-06405-y 6. Han, W.: Research of Intelligent Campus System Based on IOT. In: Jin D., Lin S. (eds.) Advances in Multimedia, Software Engineering and Computing, vol. 1. Advances in Intelligent and Soft Computing, vol. 128. Springer, Berlin, Heidelberg (2011). https://doi.org/10.1007/ 978-3-642-25989-0_29 7. Tilborg, V., Henk, A., Jajodia, S. (eds.) Encyclopedia of Cryptography and Security. Springer Science & Business Media, Springer (2011). https://doi.org/10.1007/978-1-4419-5906-5 8. Rubinstein-Salzedo, S.: Cryptography. In: Springer Undergraduate Mathematics Series. Springer Nature Switzerland AG (2018), pp. 85. https://doi.org/10.1007/978-3-319-94818-8 9. Delfs, H., Knebl, H.: Introduction to Cryptography. Springer, Berlin, Heidelberg (2007). https:// doi.org/10.1007/3-540-49244-5 10. Talhaoui, M.Z., Wang, X., Midoun, M.A.: Fast image encryption algorithm with high security level using the Bülban chaotic map. In: J Real-Time Image Proc (2020). https://doi.org/10. 1007/s11554-020-00948-1 11. Industry Spotlight: Internet of Things (IoT): https://www.springernature.com/jp/librarians/ news-events/r-d-news/new-developments-in-it-software/11033012 (2020). Accessed 30 April 2020 12. The Encryption Guide: A Guide to Effective Encryption for Every Business. https://info.tow nsendsecurity.com/ (2020). Accessed 28 May 2020

Identification of Disease Critical Genes in Preeclampsia Using Squirrel Search Algorithm Mohitesh Ch Agarwal, Biswajit Jana, and Sriyankar Acharyya

Abstract The rise in microarray technology motivates researchers of computational biology to apply various computational methods on biological data. Identification of a subset of disease-causing critical genes out of a large set of microarray gene expression dataset is a challenging task in precision medicine. In this work, a metaheuristic algorithm, namely, squirrel search algorithm (SSA) has been applied to select a subset of critical genes in preeclampsia. Three different classifiers, namely, k-nearest neighbor (kNN), support vector machine (SVM), and random forest (RF), have been used for sample classification. Three hybrid algorithms, namely, SSAkNN, SSA-SVM, and SSA-RF, have been proposed and applied to find the diseasecausing genes of preeclampsia disease. A comparison of performance of these newly proposed algorithms has been conducted in this paper. Three different sizes of preeclampsia datasets, i.e. GSE60438, GSE10588, and E-MEXP-1050 have been used here. In different datasets, the percentage of average classification accuracy of these three hybrid algorithms varies between 73.24 and 99.53%. SSA-kNN has the best performance having 80.38–99.53% average classification accuracy considering all datasets. Keywords Gene selection · Meta-heuristics · Squirrel search optimization · K-nearest neighbor · Preeclampsia

M. C. Agarwal · B. Jana · S. Acharyya (B) Dept. of Computer Science and Engineering, Maulana Abul Kalam Azad University of Technology, Nadia, West Bengal, India e-mail: [email protected] M. C. Agarwal e-mail: [email protected] B. Jana e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_29

289

290

M. C. Agarwal et al.

1 Introduction A microarray gene expression dataset contains a large number of genes, where only a few genes are responsible for causing a particular disease. Identification of diseasecausing critical genes from a large number of genes of gene expression dataset in a reasonable amount of time is quite difficult for the research communities. The search space is vast, and traditional deterministic methods are slow in execution. A metaheuristic algorithm is a good choice for getting an approximately correct solution in a reasonable time, i.e. a near-optimal solution. Several meta-heuristic algorithms and classifiers have been used previously for gene selection and sample classification. Li et al. [1] used a combination of genetic algorithm and k-nearest neighbor method in analyzing gene expression data for sample classification. Ben-Dor et al. [2] proposed tissue classification on gene expression profiles. Diaz-Uriarte [3] proposed gene classification on a microarray dataset using random forest and attracted researchers. Shivade and Keerthi [4] proposed a sparse logistic regression-based gene-selection technique and applied to microarray datasets. Shen et al. [5] used particle swarm optimization with support vector machine for gene selection and tumor classification. Biswas et al. [6] used two meta-heuristic algorithms, simulated annealing (SA) and particle swarm optimization (PSO) for gene selection. The kNN classifier has been used for sample classification. Here, in this paper, a nature-inspired meta-heuristic, namely squirrel search algorithm (SSA) has been used to select the subset of genes responsible for preeclampsia disease. Three different types of classifier, namely, KNN, SVM, and RF, have been used for sample classification. Three classifiers have been combined with SSA to produce three hybrid algorithms, namely, SSA-kNN, SSA-SVM, and SSA-RF. The proposed algorithms (SSA-kNN, SSA-SVM, and SSA-RF) have been applied to three preeclampsia datasets having accession numbers, GSE60438, GSE10588, and E-MEXP-1050 to find the preeclampsia-causing genes. A comparative analysis of the performances of SSA-kNN, SSA-SVM, and SSA-RF has also been done in this paper. The rest of the paper is organized as follows. The input gene expression dataset and its expression sub-matrix have been discussed in Sect. 2. Gene selection by sample classification along with methodologies has been discussed in Sect. 3. Section 4 contains the experimental results and analysis, and finally, the conclusion has been made in Sect. 5.

2 Gene Expression Data Generally, a large number (maybe more than 40,000) of genes are present in input microarray gene expression dataset. The objective is to find a very small subset (say 30) of genes from this input dataset (X) that is responsible for the disease preeclampsia. There are n C d possibilities in the search space (where n is the total

Identification of Disease Critical Genes in Preeclampsia …

291

Table 1 Format of the input dataset Gene Normal patient samples G1 G2 Gn

Preeclampsia-affected samples

N1

N2

N g1,1 N g2,1 N gn,1

N g1,2 N g2,2 N gn,2

… Nmax N … … …

N g1,max N N g2,max N N gn,max N

D1

D2

… Nmax D

D g1,1 D g2,1 D gn,1

D g1,2 D g2,2 D gn,2

D … g1,max D D … g2,max D D … gn,max D

number of genes in the input dataset and d is the total number of selected genes present in the subset) out of which one has to search through different subsets of the same size d. Here optimization technique is required to solve this diversified search problem to maximize the classification accuracy. In this large search space, deterministic methods cannot find an optimal solution in a reasonable amount of time. Therefore, meta-heuristic methods are required to deal with this problem. The structure of an input dataset X is described in Table 1. As in the standard form, gene Gk is identified by the kth row. Normal samples (N i represents the ith normal sample) and diseased samples (Dj represents the jth diseased sample) as preset in the columns. Here, k ∈ {1, 2, 3, . . . , n}, i ∈ {1, 2, 3, . . . , max N }, and j ∈ {1, 2, 3, . . . , max D }, where n is the total count of genes, max N is the total count of normal samples, and max D is the total count of diseased samples in the input c dataset X. Every entry (matrix element) is the normalized expression level gk, p of gene Gk at sample p (the class c, i.e. category of sample p may belong to either normal or diseased), p ∈ {1, 2, 3, . . . , M}, where M = max N + max D . The candidate solution S has d distinct gene indices randomly chosen from X index . Here, X index ∈ {1, 2, 3, . . . , n}. Expression sub-matrix X S is formed by selecting rows corresponding to indices in S from X.

3 Gene Selection by Sample Classification Using SSA-SVM, SSA-KNN, and SS-RF Every sample in the dataset X s has been classified through kNN, SVM, random forest classifiers. A sample is properly classified if the computed class tallies with the original class. The main objective of sample classification is that the count of samples that are classified properly in X s is maximum. Classification accuracy CA has been used to find the quality of the candidate solution. It is maximized to find the best possible candidate solution S. CA can be calculated by Eq. 1. CA =

NumberofproperlyclassifiedsamplesinX S × 100% TotalnumberofsamplesinX

(1)

A resampling procedure, namely k-fold cross-validation method, is used here to classify the samples [7]. In k-fold cross-validation, the data is randomly shuffled

292

M. C. Agarwal et al.

followed by splitting into k-groups (here, k = 5). For each unique group, a group is taken as test data, and remaining groups are taken as training data. The state-of-theart implementation of kNN, SVM, and RF available in the Python sk-learn library have been employed here.

3.1 k-Nearest Neighbor (kNN) The kNN [6, 10] is a statistical approach used in the early 1970s in various applications. The basic logic behind kNN is that it finds a group of k samples that are closest to an unknown sample taken from the input dataset. From the group of selected samples k, the majority class decides the class of the unknown sample. As a result, k plays the role of a deciding factor on the performance of the algorithm. kNN combined with squirrel search algorithm approach (SSA-kNN) has been used for sample classification.

3.2 Support Vector Machine (SVM) SVM is a classification method based on the theory of statistical learning. It has been used to assign unlabeled samples to two binary classes, i.e. classifies them whether they are disease-causing or not. SVM finds hyper plane for linearly separated data, such that the distance between training samples and the hyper plane is maximum. When data is nonlinearly separable, the hyper plane is found by mapping the samples to multi-dimensional space [7]. Kernel function has been used for the assignment purpose. SVM is combined with squirrel search algorithm approach (SSA-SVM) and used for sample classification.

3.3 Random Forest Classifier (RF) Random forest (RF) is an ensemble learning-based classification algorithm. It creates decision trees based on the sample data. After that, it makes the prediction from each of decision tree and lastly chooses the best solution using the voting method [8]. RF is combined with squirrel search algorithm approach (SSA-RF) and used for sample classification.

Identification of Disease Critical Genes in Preeclampsia …

293

3.4 Squirrel Search Algorithm (SSA) Squirrel search algorithm (SSA) is a meta-heuristic approach simulated by the foraging behavior of southern flying squirrels [9]. According to Jain et al. [9], the assumptions for the modeling of SSA have been discussed as follows: 1. 2. 3.

4.

The number of squirrels (m) and the number of trees (m) are the same in a forest, and only one squirrel is assigned to each tree. Every squirrel searches for optimal food source through their foraging behavior. Three types of trees, i.e. hickory tree (F S ht ) (hickory nut food source), acorn nuts tree (F S at ) (acorn nut food source), and normal tree (F S nt ) (no food source), are available in the forest. There are one hickory nut food source and three acorn nut food sources available in the forest.

The optimal solution of SSA is the squirrels on F S ht . The pseudo-code of the hybrid squirrel search algorithm is given in Algorithm 1. In Algorithm 1, each squirrel is a candidate solution X S and the R1 , R2, and R3 are the random numbers with the limit of [0, 1]. Pdp is the predator presence probability. The d g , Gc is the gliding distance and gliding constant, respectively.

294

M. C. Agarwal et al.

The pseudo-code for calculating the objective function value has been described in Algorithm 2. A score is used to calculate the total fitness of each candidate solution S. In other words, it is the total number of properly classified sample in X S . The main objective is to maximize the fitness value. If the predicted class and the original class are the same, then the test sample is called properly classified sample, and the score is increased by 1; otherwise, the score remains unaltered.

4 Experimental Results The machine used is HP-Pavilion-15CK069TX, having quad-core processor i58250U with base clock frequency 1.6 GHz, 8 GB RAM. Programs are written in Python 3 in Windows 10 environment. The number of iterations is fixed to 200 for all the observations. All the parameters of SSA have been set according to Jain et al. [9]. The total number of flying squirrels (m) is 50. Problem dimension (d) = 30. The F S ht , F S at , and F S nt have been considered as 1, 3, and 46, respectively. The predator presence probability (Pdp ) and gliding constant (Gc ) have been set to 0.1 and 1.9, respectively. The gliding distance (d g ) is set randomly. Three datasets for the preeclampsia disease, namely, GSE10588.xlsx, GSE60438_Expr.xlsx, E-MEXP-1050.xlsx have been described in Table 2. The 10 independent runs with different seeds were carried out for each observation for SSA-SVM, SSA-kNN, and SSA-RF algorithms. The experimental results of SSA-SVM, SSA-kNN, and SSA-RF have been described in Table 3. For the GSE10588 dataset, all the three methods SSA-SVM, SSA-kNN, and SSA-RF have best CA with 100% accuracy, but SSA-SVM is more appropriate Table 2 Description of preeclampsia datasets S. no.

Dataset accession no.

Collected from (site)

Number of normal samples

Number of preeclampsia samples

Number of genes (n)

1

GSE10588

GEO, NCBI

26

17

32,878

2

GSE60438

GEO, NCBI

42

35

47,323

3

E-MEXP-1050

Array Express

17

18

8,793

Identification of Disease Critical Genes in Preeclampsia …

295

Table 3 Performance comparison between SSA-kNN, SSA-SVM, and SSA-RF Dataset name

Algorithms

Best fitness

Avg fitness

Worst fitness

Fitness SD

Best CA (%)

Average CA (%)

GSE10588

SSA-kNN

43

42.8

42

0.4216

100

99.53

SSA-SVM

43

42.5

42

0.5270

100

98.83

SSA-RF

43

42.3

42

0.4830

100

98.37

SSA-kNN

65

61.9

58

2.5144

84.41

80.38

SSA-SVM

62

56.4

54

2.7568

80.51

73.24

SSA-RF

64

59.9

57

1.9119

83.11

77.79

SSA-kNN

33

30.9

29

1.2866

94.28

88.28

SSA-SVM

31

29.4

28

0.9660

88.57

80.00

SSA-RF

33

29

21

1.4142

94.28

82.85

GSE60438

E-MEXP-1050

when average CA (98.83%) is considered. For the GSE60438 dataset, SSA-kNN has the best CA with 84.41% accuracy. For the E-MEXP-1050 dataset, SSA-kNN and SSA-RF perform equally in terms of best CA (94.28%). SSA-kNN outperforms other hybrid algorithms in terms of average CA, as described in Table 3. The convergence comparison graphs between SSA-SVM, SSA-kNN, and SSARF on GSE10588, GSE60438, and E_MEXP_1050 have been presented in Figs. 1–3. The selected subsets of genes corresponding to the best candidate solution obtained have been presented in Table 4. Many widely known genes including TSC1, HEY2, HOXA5, PBX1, KLF8, XYLT1, SLC30A7, CD14, GSTA, APOA1, CCDC117, FAM164A, ATG7, SLC5A3, and IFNA2 have been identified among the selected genes. These selected genes have direct or indirect influence, causing preeclampsia. The other genes identified in Table 4 need to be investigated clinically.

Fig. 1 Convergence comparison graph on GSE10588 dataset

296

M. C. Agarwal et al.

Fig. 2 Convergence comparison graph on GSE60438 dataset

Fig. 3 Convergence comparison graph on E_MEXP_1050 dataset Table 4 Selected genes obtained from SSA-kNN, SSA-SVM, and SSA-RF Methods

List of selected genes

SSA-kNN

‘HS.546375’, ‘REG1A’, ‘LOC652688’, ‘LCE3D’, ‘CPZ’, ‘LOC100130928’, ‘LOC652845’, ‘LOC650568’, ‘ASNA1’, ‘TRIM25’, ‘NKX2-2’, ‘PTPN22’, ‘LOC729009’, ‘SLFN13’, ‘TSC1’, ‘HEY2’, ‘TMCC3’, ‘TMEM90A’, ‘HOXA5’, ‘COL24A1’, ‘LOC100130005’, ‘PBX1’, ‘PPM1B’, ‘LOC150577’, ‘C8ORF45’, ‘LOC647181’, ‘STK25’, ‘PRR19’, ‘KLF8’, ‘PELO’

SSA-SVM

‘ARL6’, ‘PARP10 , ‘POU3F1’, ‘HS.578080’, ‘XYLT1’, ‘PRMT10’, ‘ZNF322A’, ‘SLC30A7’, ‘LOC391013’, ‘CD14’, ‘LOC647802’, ‘LOC641727’, ‘RFX1’, ‘LOC441666’, ‘HS.583147’, ‘HS.567464’, ‘WWOX’, ‘GSTA1’, ‘LOC338799’, ‘HS.583949’, ‘HS.553278’, ‘FBXO24’, ‘HTR1B’, ‘HS.560984’, ‘DNAJB5’, ‘LOC649037’, ‘CCRL2’, ‘APOA1’, ‘HS.568618’, ‘LOC100128031’

SSA-RF

‘CER1’, ‘CCDC117’, ‘SGCA’, ‘HS.155579’, ‘FAM164A’, ‘POFUT1’, ‘MICALL2’, ‘CDC25A’, ‘LY6G5B’, ‘KIAA1107’, ‘IFI35’, ‘HS.563550’, ‘GAGE6’, ‘TBX1’, ‘LOC647006’, ‘TSTA3’, ‘ATG7’, ‘HS.193767’, ‘TMCC1’, ‘LOC648000’, ‘IGSF2’, ‘HS.145721’, ‘RAB37’, ‘SLC5A3’, ‘LRRC10’, ‘LOC653567’, ‘YY2’, ‘LOC100132957’, ‘IFNA2’, ‘LOC652606’

Identification of Disease Critical Genes in Preeclampsia …

297

5 Conclusion In this paper, three hybrid algorithms, namely, SSA-SVM, SSA-kNN, and SSA-RF, have been proposed. A comparative analysis has been conducted between SSASVM, SSA-kNN, and SSA-RF on three different preeclampsia disease datasets. The experimental results show that SSA-kNN outperforms SSA-SVM and SSARF with respect to average classification accuracy (80.38–99.53%) in all datasets. The convergence analysis shows that the SSA-kNN achieved a better convergence rate than other algorithms. Some of the selected genes are already known to be disease-causing. Other selected genes may be clinically investigated in future.

References 1. Li, L., Darden, T.A., Weingberg, C.R., Levine, A.J., Pedersen, L.G.: Gene assessment and sample classification for gene expression data using a genetic algorithm/k-nearest neighbor method. Comb. Chem. High Throughput Screening 4(8), 727–739 (2001) 2. Ben-Dor, A., Bruhn, L., Friedman, N., Nachman, I., Schummer, M., Yakhini, Z.: Tissue classification with gene expression profiles. In: Proceedings of the Fourth Annual International Conference on Computational Molecular Biology, pp. 54–64 (2000) 3. Díaz-Uriarte, R., De Andres, S.A.: Gene selection and classification of microarray data using random forest. BMC Bioinf. 7(1), 3 (2006) 4. Shevade, S.K., Keerthi, S.S.: A simple and efficient algorithm for gene selection using sparse logistic regression. Bioinformatics 19(17), 2246–2253 (2003) 5. Shen, Q., Shi, W.M., Kong, W., Ye, B.X.: A combination of modified particle swarm optimization algorithm and support vector machine for gene selection and tumor classification. Talanta 71(4), 1679–1683 (2007) 6. Saha, S., Biswas, S., Acharyya, S.: Gene selection by sample classification using k nearest neighbor and meta-heuristic algorithms. In: 2016 IEEE 6th International Conference on Advanced Computing (IACC), pp. 250–255. IEEE (2016) 7. Alba, E., Garcia-Nieto, J., Jourdan, L., Talbi, E.G.: Gene selection in cancer classification using PSO/SVM and GA/SVM hybrid algorithms. In: 2007 IEEE Congress on Evolutionary Computation, pp. 284–290. IEEE (2007) 8. Okun, O., Priisalu, H.: Random forest for gene expression based cancer classification: overlooked issues. In: Iberian Conference on Pattern Recognition and Image Analysis, pp. 483–490. Springer, Berlin, Heidelberg (2007) 9. Jain, M., Singh, V., Rani, A.: A novel nature-inspired algorithm for optimization: squirrel search algorithm. Swarm Evol. Comput. 44, 148–175 (2019) 10. Jana, B., Acharyaa, S.: Critical gene selection by a modified particle swarm optimization approach. In: International Conference on Pattern Recognition and Machine Intelligence, pp. 165–175. Springer, Cham (2019)

An Approach for Bengali News Headline Classification Using LSTM Md. Rafiuzzaman Bhuiyan, Mumenunnessa Keya, Abu Kaisar Mohammad Masum, Syed Akhter Hossain, and Sheikh Abujar

Abstract Headline is called the soul of news. Headline carries a very important meaning. Generally, many people start reading the news after seeing the headlines. It is very important for the user to classify the headlines that he/she preferred. Classifying news type based on their headlines is a problem of text classification which lies under natural language processing (NLP) research. In different languages, there are many works done but none of them observed in Bengali. In our work we tried to visualize our very first approach to solve this problem. A LSTM-based architecture is used for solving this problem. A total of 4580 headlines are trained in our model. Finally, our model gives us 91.22% accuracy. The main challenge of our work is finding the right word vector. As far headlines are made up with different types of words and there is no similarity between any of them, so it is difficult to map them. Keywords Headline classification · Word embedding · Deep learning · LSTM

Md. R. Bhuiyan (B) · M. Keya · A. K. M. Masum · S. A. Hossain · S. Abujar Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh e-mail: [email protected] M. Keya e-mail: [email protected] A. K. M. Masum e-mail: [email protected] S. A. Hossain e-mail: [email protected] S. Abujar e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_30

299

300

Md. R. Bhuiyan et al.

1 Introduction Everyday denominations becoming dependent on internet news. Information is now available on internet in all types of entitled language by which people can grab actuality easily [1]. People also depend on internet news rather than outside. Finding effective information is the principle of information technology. News headline is the exalted generalization of text contentment [2]. Different types of news can sort such as sports, computer, Hollywood, Bollywood, music, politics and social science in internet. Users can easily locate the tidings using internet. From the internet news a user can find their interested topics, and by clicking at the topic they can see the internal news. If anyone want to see the playing news they can search by the name and grab the information. As a result, it is commandment to riddle the sentence in many classes in order such that users can have entrance easily [3]. Headline classification combines comprehensive digitized recitation to trace necessary information. News headline classification is like text mining. Text mining is discovering of information that is formerly obscure by gaining data from unstructured text. Recitation significant and pompous information is adjutant to user. Searching information about interested thing in internet is also time-consuming for users. So, news classification can reduce the time decrements from searching news. It can save users time to find the relevant news. Many researchers find various way to find the wanted news by classifying headlines into different classes. Computer-based news ramification is more essential than classification by human being. That is why researchers indicate about the computer-based news headline ramification. Human searching will not be desirable than search engine machine searching. Computers generally brake theme as simple alternation of character lacing. Without processing the information search engine cannot provide useful data. That’s why processing and pre-processing methods happen in text [4]. News headline classification is useful for users. Users can find their wanted news by searching and so many works have been done based on this in different languages. But there is no work done yet based on Bengali news headline. By different class of headline, a user will find their desired news and can understand the internal topic of the news. And that’s why we have to make the headline more meaningful so that people can easily understand and catch the interior tidings of the news. In our paper we select four types of headlines as national, international, science-technology and sports. When people will search about these topics they’ll be able to find the news easily by headlines. We have to classify headlines in this manner such that they can find spontaneously the desired result without wasting their important time. We proposed a long short-term memory (LSTM) model for our problem. It is good for text classification. More than 4500 headlines with four different classes are collected from numerous authentic newspapers using web scraping techniques. After several steps of pre-processing we feed our data into our proposed architecture and train them. Finally, we got a good accuracy and predict the headlines.

An Approach for Bengali News Headline Classification Using LSTM

301

In Sect. 2 we will discuss about the previous works done by others. Section 3 discusses about the methodology of our problem. Section 4 for the result discussion and Sect. 5 for conclusion and future work.

2 Literature Review This news headline classification is an interested area in research field. Popularity of online newspaper is increasing day-by-day. There are several works done on news classification, i.e. financial news classification [5] and classification of short texts [6]. For classifying headlines, some work done on emotion extraction from headlines [7, 8] and another work for classifying headlines from twitter [9]. Classification of news headline is considered as short text classification. The next conversation included how the researchers contributed to each category of news headline. The author Rana et al. [10] discussed about the probabilistic framework for classification of news headline in their paper. After analyzing many papers about this, the researchers concentrate on news classification and news headline classification based on probabilistic method. The probabilistic approach recites that it calculates each word of the headline and does statistical calculation of each word. Using frequency matrix the probabilities of the sample classes are enumerated. Yin et al. [11] thought about the problem of natural propagation of concise text and implemented semantic representation method. They demanded classification of Chinese news headline is the art-of-states on representation and proficiency. Using vector space model (VSM) they use BOW which describes the semantics of text. Deshmukh et al. [3] concentrate on recommendation of news on the interested basis. This moved forward recommending users the news based on their profile where their interest is predefined. For this purpose, SVM is deliberated for intelligence personalization scheme where SVM classifies the headline of the news. They mentioned E-newspaper personalization process as per user’s choice. In this paper the authors are concerned with the subscribed users to provide the information categorized as their interest. This paper aimed to classify information for identifying the popular news groups of the user from desired country [12]. Support vector machine which support data with high dimensions and term frequency-inverse document frequency manner mentioned for classifying the news headline. They studied about the approach of news text. From BBC Urdu and Urdu-point news headline Zaidi and Hassan [1] implement text classification using various machine learning algorithm, in which they find best result from ridge classifier. They create their own dataset collecting from different sites and classify the text based on news headlines, and analyze which algorithm is best for the dataset they created. By using N-Gram language model, Liu et al. [2] explore the internet news headline classification. By calculating probability grade of word string, they find better classification performance. They use N-Gram model which predicts the next item in the form of order. Putting forward N-Gram language model they classify the headline of internet to establish two Yuan classification models. Rana et al. [4] reviewed paper on news headline classification. They reviewed more

302

Md. R. Bhuiyan et al.

researcher’s paper and discussed about the model which is best for what type of data and also prattle about the text pre-processing techniques, document indexing, and feature selection briefly. For pre-process data step-by-step procedures are described and were made easy for the new researchers to able to know about the models and pre-process deftness. Daskalopoulos [13] recount a NL perception agent sketch that is capable of predicting a jolt price stroke due to a news headline event related to a stock beacon. They apply NB method on long data to score the impact of the stock price. This is a domain independent of deputationist which will be considered as sub-module of any machine learning scheme. Agent dialect sample is used to check the intuition of news headline in positive or negative way. NB used to determine the classification of stock price if is in fall or risk. In our paper, we introduced a very fast approach toward news headline classification using LSTM architecture.

3 Methodology In this piece of work for building news headlines classification system we need to do few steps. Data collection is an essential task for any kind of data-driven approach. For our problem we also need lots of data. Further pre-processing is needed to find the actual data for our work. In Sects. 3.1 and 3.2 we will be covering the approach for dealing with our data, Sect. 3.3 for word embedding, Sect. 3.4 will be the model discussion and Sect. 3.5 for our proposed model. For training our model we rely on TensorFlow GPU version. Figure 1 shows our end-to-end process of building an automatic news headline classification.

Fig. 1 Workflow of news headline classification

An Approach for Bengali News Headline Classification Using LSTM

303

Fig. 2 Dataset information

3.1 Data Collection Data is an integral part of data-driven techniques like machine learning or deep learning. The more the data available, the better the output will be. Many headlines were needed for our news headlines classification problem. So, from the authentic news portal, i.e. Prothomalo [14], we have collected headlines using web scraping techniques. A popular well-known library called Beautiful-Soup [15] is used for this. In total, 4580 news headlines contain different classes, i.e. Bangladesh, international, science and technology, sports. Adding them all together we build our dataset. Label each of the items in terms of its corresponding headlines. Finally, our dataset contains two columns, i.e. headlines and news types. Figure 2 shows the amount of data we collected.

3.2 Data Pre-processing The data in our dataset is not well furnished yet. It is very important to have a wellprocessed data for making model better. That’s why we pre-processed our data in a few steps. First, split the texts that made dataset into token and the process is called tokenization. Then we removed the unnecessary characters like “?” “//” from our dataset via regular expressions. Further split the dataset into 80% for training and 20% for testing.

304

Md. R. Bhuiyan et al.

3.3 Word Embedding Texts are represented as a word vector. There are numerous studies for this representation. Word embedding is a powerful way to represent words that can learn from textual data. It is a vector representation of words and can give us a powerful semantic relationship between the words. But as we are working with Bengali, so we need our own word embedding. Sometimes it is good approach to work on pre-trained embedding. So, in our case “bn w2v model” is used for our purpose. Before training, we need to finalize our vocabulary size. For this from the pre-process text we count the vocabulary.

3.4 Model There are several types of deep learning models out there. Each of them is used for a variety of purposes, just like ConvNet is used to handle image data and video processing stuff. Another type of model is used for text classification and language modeling. Recurrent neural network and long short-term memory network (LSTM) are used for text classification. In our problem set LSTM, a type of RNN used for our problem.

3.5 Long Short-Term Memory LSTM is a type of recurrent neural network (RNN). It omits the vanishing gradient problem of RNN and works really well in long sequences of data. From its name we see that LSTM itself has a term called memory; it actually keeps the previous activations into its memory over a long time that RNN cannot apply this technique. LSTM has three gates. For input it uses input gate, it ; for output it uses output gate, ot and a forget gate f t to control the amount of information from the previous memory state. • • • • • •

Input gate, it = σ (W ix x t + W ih ht1 ) Output gate, ot = σ (W ox x t + W oh ht1 ) Forget gate, ft = σ (W fx x t + W fh ht1 ) Process input or input node, gt = tanh (W gx x t + W gh ht1 ) Memory state, ht = tanh (st ).ot Hidden node, st = st−1 .ft + gt .it .

An Approach for Bengali News Headline Classification Using LSTM

305

Fig. 3 Applied LSTM architecture

3.6 Applied Model In this section, the applied model is discussed. A sequential model is used to build the whole architecture. Figure 3 shows our applied model and description as follows: • Input layer: Texts that we are passing in input layer have dimension of 150. • Embedding layer: In our model we use an embedding layer which is 150 dimensional. • LSTM layer: A LSTM unit is used after embedding layer whose dimension is 128. • Activation layer: Activation is some sort of functions that give a corresponding output that are fed from an input. In our task we used a nonlinear activation function called ReLU after embedding layer. • Dropout layer: Overfitting happens frequently in machine learning model. For reducing overfitting, we add dropout into our model. We have used two dropout layers in our model. 0.4 and 0.2 dropout rate was used in our model. • Dense layer: Dense layer is used for classification. We have used 128 units and 4 units of dense layer. The four-unit dense layer is used for classification.

4 Experimental Results and Discussion TensorFlow 1.15 is used for our model. Before training our model we need to finalize some parameters as predefined, which is also known as hyper-parameters. If we do not declare these parameters into our model, then our model is not working well, as well as we cannot give good accuracy. So, it is really important in training time for

306 Table 1 Hyper-parameters settings

Md. R. Bhuiyan et al. Batch size

124

Epoch

50

Learning rate

0.0001

LSTM size

128

getting higher accuracy. In our model we also use some of these parameters. Table 1 shows the hyper-parameters of our model. We used 80% data for training and rest for testing. We compiled our model into Adam [16] optimizer and categorical cross entropy for loss function. Then use.fit(·) method to start training. After training, our model got 91.22% accuracy. Figure 4a shows our training and validation accuracy of our model. We have seen our training accuracy as 91.22% and validation accuracy as 80.01%. Figure 4b shows our training and validation loss of our model. We have seen our training loss is 0.155 and validation loss is 0.622. Table 2 shows the report of our model. Apart from that, our average precision and f1-score are quite good. Table 3 shows confusion matrix of our model. Here, Bangladesh, international, science and technology, sports are represented, respectively.

Fig. 4 Training curves. a Training versus validation accuracy, and b Training versus validation loss

Table 2 Classification report

News type

Precision

Recall

f1-score

Bangladesh

0.78

0.59

0.67

International

0.81

0.67

0.74

Science and technology

0.81

0.65

0.72

Sports

0.77

0.75

0.76

Average

0.80

0.66

0.72

An Approach for Bengali News Headline Classification Using LSTM

307

Table 3 Confusion matrix Bangladesh Bangladesh

International

Science and technology

33

International

36

255

34

22

Science and technology

13

17

131

18

9

22

12

142

Sports

18

Sports

133

21

5 Conclusion and Future Work In our paper, we demonstrate an automatic Bengali news headline classification using LSTM network architecture. This is our first approach for building an automatic system for headlines classification. This model gives us a new dimension toward building headlines classification in Bengali. Although our model does not accrue hundred percent accuracy, even it can predict very well regarding any headlines. It can be used for further news classification task also. The first drawback of our work is the dataset. As deep learning algorithm needs lot of data but for our first work, there is no dataset available. We build our dataset on our own for the web. Making a dataset is not an easy task after all. Another drawback is in Bengali there is not enough word to vector. We have seen in our problems that there were a lot more words present that comes very infrequently as different types of news headlines are written in different context and there is no similarity between any of them. So, we have to improve them for further research. There is no lemmatizer available in Bengali. Our future work will be added more words into word to vector, that’s why the word representation for Bengali will be more convenient to use.

References 1. Hassan, S., Zaidi A.: Urdu News Headline, Text Classification by using different machine learning algorithms (2019) 2. Liu, X., Rujia, G., Liufu, S.: Internet news headlines classification method based on the N-Gram language model. In: 2012 International Conference on Computer Science and Information Processing (CSIP) 3. Deshmukh, R.R., Kirange, D.K.: Classifying news headlines for providing user centered enewspaper using SVM. Int. J. Emerg. Trends & Technol. Comput. Sci. (IJETTCS) 2(3) 2013. www.ijettcs.org 4. Rana, M.I., Khalid, S., Akbar, M.U.: News classification based on their headlines: a review. ISBN: 978–1–4799–5754–5/14/\$26.00 ©2014 IEEE 5. Drury, B., Torgo, L., Almeida, J.J.: Classifying news stories to estimate the direction of a stock market index. Paper presented at the information systems and technologies (CISTI) 6th Iberian conference (2011) 6. Heb, A., Dopichaj, P., Maab, C.: Multi-value classification of very short texts. In: Proceedings of the 31st Annual German Conference on Advances in Artificial Intelligence, pp.70–77. Springer, Berlin (2008)

308

Md. R. Bhuiyan et al.

7. Kirange, D.K., Deshmukh, R.R.: Emotion classification of news headlines using SVM. Asian J. Comput. Sci. Inf. Technol. 104–106 (2012) 8. Jia, Y., Chen, Z., Yu, S.: Reader emotion classification of news headlines. In: Proceedings of the International Conference on Natural Language Processing and Knowledge Engineering (NLP-KE), pp. 1–6 (2009) 9. Dilrukshi, I., De Zoysa, K., Caldera, A.: Twitter news classification using SVM. In: 2013 8th International Conference on Computer Science & Education (ICCSE), pp. 287–291. IEEE (2013) 10. Rana, M.I., Khalid, D.S., Abid, F., Ali, A., Durrani, M.Y., Aadil, F.: News headlines classification using probabilistic approach. VAWKUM Trans. Comput. Sci. 7(1) (2015) 11. Yin, Z., Tang, J., Ru, C., Luo, W., Luo, Z., Ma, X.: A Semantic representation enhancement method for Chinese news headline classification. In: Huang, X., et al. (eds.) NLPCC 2017, LNAI 10619. Springer International Publishing AG, pp. 318–328 (2018) 12. Dadgar, S.M.H., Araghi, M.S., Farahani, M.M.: A novel text mining approach based on TF-IDF and support vector machine for news classification. In: 2nd IEEE International Conference on Engineering and Technology (ICETECH), 17th & 18th March 2016, Coimbatore, TN, India (2016) 13. Daskalopoulos, V.: Stock price prediction from natural language understanding of news headlines (2021) 14. Latest Bangladesh national & local news, photo, video. (n.d.). Prothomalo. https://www.pro thomalo.com/bangladesh 15. Beautiful Soup Documentation — Beautiful Soup 4.9.0 documentation. (n.d.). https://www. crummy.com/software/BeautifulSoup/bs4/doc/ 16. Kingma, D.P., Jimmy L.B.: Adam: a method for stochastic optimization.” ICLR 2015: International Conference on Learning Representations (2015)

Artistic Natural Images Generation Using Neural Style Transfer Atiqul Islam Chowdhury, Fairuz Shadmani Shishir, Ashraful Islam, Eshtiak Ahmed, and Mohammad Masudur Rahman

Abstract Rendering an image in a different form of style is a difficult task in image processing. And the representation of an image is very important as an image contains a lot of information. To represent the image uniquely, the Neural Style Transfer technique plays an important role. Neural Style Transfer refers to a technique of generating new images in the style of another image and generates new images as if they are oil painted images. In this paper, we have modified different types of natural images of Bangladesh with the help of Neural Style Transfer. It is a technique that is used to blend the content image and style image in an optimized way and gives a creative output image that is very interesting and attractive. We have transformed the input image for minimizing its content distance and style distance with the help of the content image and style image using the VGG19 network model. Keywords Neural Style Transfer · Transfer learning · Content image · Style image · VGG19 · ImageNet

A. I. Chowdhury (B) United International University, Dhaka, Bangladesh e-mail: [email protected] F. S. Shishir Ahsanullah University of Science and Technology, Dhaka, Bangladesh e-mail: [email protected] A. Islam · E. Ahmed Daffodil International University, Dhaka, Bangladesh e-mail: [email protected] E. Ahmed e-mail: [email protected] M. M. Rahman Bangladesh University of Engineering and Technology, Dhaka, Bangladesh e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_31

309

310

A. I. Chowdhury et al.

1 Introduction Painting is basically an art which attracts the people of all category. If this painting would be easier, then people will give more attraction in paintings because of its ease and availability. Moreover, image can represent various kinds of information. If any person is interested in painting, then the image conversion is important for impressing the mankind. So our proposal is to give two kinds of images to every person, which are content image (input) and output image (after applying style image to content image). Some people are interested in real images, while some people like modified images. Basically, if we consider the situation of children, they like painting so much. So if we convert a real image into painting, they will get much interest and give attention to the picture and may understand by paying attention to the image. To make it available, Neural Style Transfer helps us basically. Neural Style Transfer is an optimal process which mainly takes a content image and a style image (as reference image), and blend them together to transform the input image into the content image in a “painted” form. Transferring the style from one image to another image is considered as a texture transfer issue. In texture transfer, the main focus is a texture integration from a source image while suppressing the texture combination to maintain the preservation of the content regarding a target image [1]. There are two distance functions that we get from the principal of the Neural Style Transfer, which are the distinction between the two images with their content and the distinction between the two images concerning their style [2]. The “style” means colors, textures, and visual patterns in an image. On the other side, the “content” means the input image which has a higher level macrostructure. The Neural Style Transfer automatically converts the image in a better way because of the VGG19 model. This architecture is useful for extracting the image maps and thus we get a better output. The generated images’ condition relies on the input image and the style image resolution. This Neural Style Transfer is a part of a CNN that was mainly created for the classification of an image. The area to which Neural Style Transfer belongs is basically image processing. It is all about to manipulate an image to adopt the style of another image. The first proposition of Neural Style Transfer was by Gatys et al which worked really well [1]. Their hypothesis was to apply a neural method of Artistic Style for isolating and recombining the image content and image style in the field of natural images. We have modified this for our purpose, that is, to get a generated image of the natural beauty of Bangladesh. And we have got a better result for those content images.

2 Related Study This section will discuss about the baseline study and functionality of Neural Style Transfer technique, VGG19 model, and ImageNet dataset thoroughly.

Artistic Natural Images Generation Using Neural Style Transfer

311

2.1 Overview of Neural Style Transfer Neural Style Transfer takes an image as an input and a style image for getting the styled input image. That means I nput I mage + St yleI mage = Generated I mage(Out put).

(1)

Before this Neural Style Transfer technique, deep learning-based style transfer procedure offered results unparalleled which were previously achieved with computer vision techniques. But that was not good to give an accurate result. After that, a style transfer algorithm was proposed that had the ability to extract the content of semantic image from the target image and render that content of the image concerning the source image [1]. As we know there was much research about this texture transfer system which defines transformation of style from one image to another. A paper discussed texture synthesis, in which there was non-parametric algorithm of large range which did adjustment of photo-realistic textures of natural image with the pixels recomputation of a given source texture [3]. Another paper introduced a mapping system which was correspondent mapping including the intensity of target images’ feature to constrain the approach of texture synthesis [4]. Then another paper improved the algorithm that was used previously with the help of extracting the features of the texture transfer with edge orientation [5]. But all of these papers used low-level target images for their research purpose. Filis et al. carried out a comparison between a computer-generated style transfer and a physically painted style transfer [6]. In Neural Style Transfer, a loss function should be represented to specify what we need to achieve. Basically, there are three types of loss for this optimization problem. They are content loss, style loss, and total variation loss. The most essential issue in automating image transfer to artistic style is the way of modeling and extracting style from a given image. In this way, reconstruction of that image having the intended styles is another important process after acquiring the style representation from the previous stage. Moreover, the image content needs to be preserved during the reconstruction process [7].

2.2 VGG19 Network Model and Transfer Learning VGG19 is a pre-trained CNN model, trained on approximately 1.2 million images from the ImageNet dataset [8]. It is a simple variant of the VGG16 network. This model contains 19 layers and helps to classify the images into the object by 1000 categories. This VGG19 network takes three images as input: the content image, the style image, and a placeholder for the generated image. Figure 1 shows the architecture of the VGG19 network model. In the figure, conv means convolution and FC means fully connected. So, we can say that this VGG19 model is a combination of multiple convNets with trained images of different sizes.

312

A. I. Chowdhury et al.

Fig. 1 Network architecture of VGG19 model [9]

On the other hand, Transfer Learning is a process of machine learning by which a model developed for a process is reused as the starting point of the next model. The benefits of Transfer Learning are that it can speed up its time for the development and train a model by reusing the modules of already developed models [10].

2.3 ImageNet Dataset ImageNet is the largest dataset in the world containing 1000 classes with more than 14 million images. The organization of this dataset follows WordNet hierarchy. There are multiple words and phrases in this WordNet database which are called “synonym set” or “synset.” WordNet database consists more than 100,000 synsets [10]. ImageNet does not have its own images; it only provides the URL of the image. This ImageNet dataset compiles an accurate list of web images for each synset of WordNet [10]. In the recent era, proposition of several methods has been made for authorizing a single network to transfer multiple styles, including a model that conditioned on binary selection units [11]. In the end, we have to create an image that is similar to content image by applying the style image. We need to do this after the transformation of input image with the minimization process of content and style loss.

3 Proposed Methodology Neural Style Transfer is based on texture transfer or re-texturing technique and plays a vital role to make an image into an art. But images with high resolution and a strong textured style are needed to get the best outcome using the process.

Artistic Natural Images Generation Using Neural Style Transfer

313

In Neural Style Transfer, we need two images. They are style image and content image. We need to transcribe the style from the style image and bring to bear upon it into the content image so that we can get a computer-generated output image. Our proposed model follows an algorithm for this technique. The algorithm of our method is given below: Algorithm 1: Neural Style Transfer Technique Load the image; Preprocessing(); Load the model; for each iteration in iteration number do Style-loss(); Content-loss(); Variation-loss(); end First of all, we need a content image regarding a style image and use both of them as an input image. Then we have to preprocess these images for the next operation. After preprocessing, these images need to be run on the VGG19 model. Then we get the generated image as our output. In this full process, the L-BFGS algorithm is needed and we need to save the current generated image at each iteration. For each iteration, we have to calculate the three loss values, which are content loss, style loss, and variation loss. Here we have done this process by 20 iterations and got the better result for the images that we captured by our devices. We have used our captured images for this experiment. The flow diagram of our proposed method is displayed in Fig. 2.

Fig. 2 Flow diagram of our working method

314

A. I. Chowdhury et al.

4 Experimental Analysis This section describes the implementation process and result analysis of our work. We have run our code in Google Colaboratory GPU for the betterment concerning time.

4.1 Implementation Process For this research, we need to know the working process of this Neural Style Transfer techniques, which are given below: • Firstly, we have to set up a network for computing VGG19 layer activations for the content image, style reference image, and generated image at a time. • Then, we have to use the layer activations computed over these three images to measure the loss function that needs to be minimized for achieving style transfer. • Lastly, we need to set up a gradient descent process to minimize this loss function. Then we have to load the VGG19 network model. It takes three images: a style image, a content image, and a placeholder for the generated image. After that, we have to measure the values of content loss, style loss, and variation loss for the image. We have calculated the gradient loss by using the L-BFGS algorithm. L-BFGS works better if there are a large number of parameters in an image. For the computation of the content loss, we only took one top layer which is the “conv2” layer. On the other hand, for the style loss analysis, we used a list of layers that spans both low-level and high-level layers. Then we have added the total variation loss at the end. Equation [12] calculates the total loss function. L total (S, G, C) = αL content (C, G) + β L st yle (S, G).

(2)

Here, • • • •

S = style image, C = content image, G = generated image, and α, β = weights of each type of loss.

Finally, we have calculated the three losses (style loss, content loss, and variation loss). These losses are calculated in each iteration. And the total number of iteration is 20. Each iteration takes approximately 20–21 s to complete.

Artistic Natural Images Generation Using Neural Style Transfer

315

4.2 Result Analysis As we have mentioned earlier, we have done this process using the natural images of Bangladesh. We have used some of the images of beautiful places in Bangladesh and applied the style image to get the generated image. Here we have used one style image that is shown in Fig. 3. This image is collected from [13] which is an image of Stanford Campus modified by AI Van Gogh. We have worked on 50+ images. Many images have low resolution, many of them have high. In Fig. 4, we took the content images as input images which give better output images. And the output takes less time to generate every image.

Fig. 3 The style reference image for our work

Fig. 4 Some content images (left) and output images (right) after being styled with the style reference image

316

A. I. Chowdhury et al.

5 Future Work and Discussion In our paper, we have used Transfer Learning using the VGG19 network model. We have transformed the basic input image by reducing both content and style distances. In future, we will build our network model to implement the task so that we can get better output in the future. Also, our dataset is small, we will collect and capture more images of different locations of Bangladesh and the other country places. We have to do this for the betterment of the artistic style so that people can understand the value of the art and use them for their different purposes. Also, we will try to apply the generative adversarial network model for more realistic style transfer.

6 Conclusion Neural Style Transfer means to apply a style reference image to a content image and get the computer-generated image as output. The model which we have used (VGG19) gave us better and satisfactory results from the perspective of our captured and collected images of beautiful places of Bangladesh. As we experimented with this technique with our dataset, we got better output images that yield to enhance the attraction of those images. We have worked on more than 50 images using transfer learning which results in extremely appealing output with artistic effects. But the resolution of the input images must be high for getting better style-generated images. In future, we hope that many variants and refinements will be possible if we cling with this experiment.

References 1. Gatys, L., Ecker, A., Bethge, M.: Image style transfer using convolutional neural networks. pp. 2414–2423 (2016) 2. Neural style transfer: creating art with deep learning using tf.keras and eager execution. https://medium.com/tensorflow/neural-style-transfer-creating-art-with-deep-learningusing-tf-keras-and-eager-execution-7d541ac31398 3. Efros, A.A., Leung, T.K.: Texture synthesis by non-parametric sampling. In: Proceedings of the Seventh IEEE International Conference on Computer Vision, vol. 2, pp. 1033–1038 (1999) 4. Efros, A.A., Freeman, W.T.: Image quilting for texture synthesis and transfer (2001) 5. Lee, H., Seo, S., Ryoo, S.T., Yoon, K.: Directional texture transfer, pp. 43–48 (2010) 6. Coba, F., Abrahams, R.: Neural style transfer with human comparison (2018) 7. Jing, Y., Yang, Y., Feng, Z., Ye, J., Yu, Y., Song, M.: Neural style transfer: a review (2017) 8. Deep learning toolbox model for vgg-19 network. https://ww2.mathworks.cn/matlabcentral/ fileexchange/61734-deep-learning-toolbox-model-for-vgg-19-network 9. Zheng, Y., Yang, C., Merkulov, A.: Breast cancer screening using convolutional neural network and follow-up digital mammography, p. 4 (2018)

Artistic Natural Images Generation Using Neural Style Transfer

317

10. An introduction to transfer learning in machine learning. https://medium.com/ kansas-city-machine-learning-artificial-intelligen/an-introduction-to-transfer-learningin-machine-learning-7efd104b6026?fbclid=IwAR2JS0t_Xw4ncjgeZnKBX55mVhoDsclPwjaSrOswVnjQdkFgPXktaNRCGo 11. How to use the pre-trained vgg model to classify objects in photographs. https:// machinelearningmastery.com/use-pre-trained-vgg-model-classify-objects-photographs 12. Li, Y., Fang, C., Yang, J., Wang, Z., Lu, X., Yang, M.H.: Diversified texture synthesis with feed-forward networks, pp. 266–274 (2017) 13. Teaching computers to paint like van gogh. https://medium.com/@AlexZhuang/teachingcomputers-to-paint-like-van-gogh-fc4cd7c22dab

Routh-Hurwitz Criterion for Stability: An Overview and Its Implementation on Characteristic Equation Vectors Using MATLAB Aseem Patil

Abstract Stability is one of the most significant system analysis and design factor. Therefore, stability analysis should be studied, understood and properly applied in engineering education systems like control systems. The Routh-Hurwitz stability criterion in control systems is a mathematical method which is reasonable and essential to ensure the stability of an LTI system. The stability criterion of Routh-Hurwitz is a requirement and an effective stability condition. If any control system does not fulfill the requirements, we may conclude that it is dysfunctional. Nonetheless, the control system may or may not be stable if it meets the appropriate criteria. In this paper, the Routh-Hurwitz stability criterion is first discussed in brief and also the implementation of the stability criterion using characteristic equation vectors using the MATLAB software. Keywords Routh-Hurwitz criterion · Stability · Characteristic equations · Array · Routh table · LHP · RHP · Stable · Unstable

1 Introduction The Routh-Hurwitz criterion has lost some of its value considering the computer resources available today, but remains important for practical problems. The method allows to achieve analytical stability conditions with several plant and controller parameters for specific applications. In any case, a remarkable result with significant value remains the Routh-Hurwitz criterion. This criterion is based on the order of the characteristic equation coefficients in the form of an array known as the Routh array. Using this technique one can only tell how many loop mechanism peaks in the LHP, RHP and jω axis are present. Therefore, the number of poles in each segment can be identified by ‘Routh array’. The application of Routh-Hurwitz helps one to demonstrate the conditions necessary and adequate to ensure the stability of the operating stage. Some relation contains only dynamic coefficients while the critical A. Patil (B) Department of Electronics Engineering, Vishwakarma Institute of Technology, Pune, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_32

319

320

A. Patil

mass intervenes in two relations; thus it is theoretically possible to find instances in which the system is unbalanced or balanced irrespective of the value of a variable like mass M or cases in which the system is stable for other values M, depending on the dynamic coefficients values, respectively. Conditions for polynomial for the left half of the plane (LHP) need not be defined and the stability of discrete time systems cannot be evaluated directly. This is where the stability criterion comes into picture.

2 Routh-Hurwitz Stability Criterion The Routh-Hurwitz stability test is a mathematical tool for assessing, if all polynomial roots have negative real parts and are used for LTI systems stability analysis [3]. In different engineering applications, this stability criterion is useful. In several special cases they are discussed on the basis of the Routh array. It is important to decide whether any root of its equation of characteristics lies in the right half of the “S” level in order to maintain the stability of the linear time-invariant system. Hurwitz and Routh published the approach for evaluating the system’s necessary equilibrium conditions independently. It is necessary for all elements in the first column of the Routh array to have adequate stability [2, 5, 6]. The Routh-Hurwitz criterion states that if and only if all the roots in the first column have the same sign and if the same sign does not appear or if a change of sign exists, then the number of changes in sign in that column equals the number in roots in the right half of the s plane, and the number of roots with positive real portions. Routh-Hurwitz criterion has three states of stability: 1. 2. 3.

Stable system: If all the roots in that particular characteristic equation lie on the right side of the “S” plane, then the system is called a stable system. Marginally stable: If all the roots are located on the imaginary axis of the “S” plane, then it is assumed that the structure is slightly stable. Unstable system: If all the system’s roots are on the left half of the “S” axis, then it is assumed that the system is an imperfect one.

2.1 Structure of the Routh-Hurwitz Criterion Let us consider a characteristic equation: a0 s m + a1 s m−1 + · · · + am = 0

(1)

Consider the following steps so as to make a determinant of the coefficients of the characteristic equation: 1.

The first two rows serve as the baseline for determining the third row and further.

Routh Hurwitz Criterion for Stability …

2.

321

Row 1

A0

A2

A4

………………….

Row 2

A1

A3

A5

………………….

The third row is derived from the first two rows as can be seen below:

Row 1

A0

A2

A4

………………….

Row 2

A1

A3

A5

………………….

Row 3

B1

B3

B5

………………….

Here, the values of B1 and B3 are derived by using the values of the first 2 rows in the form of a determinant matrix, as shown in Eqs. (2) and (3).  1  A0 A1  A1  1  A0 B3 = −  A1 A1

B1 = −

3.

 A0 A3 − A1 A2 A2  =− A3  A1  A0 A5 − A1 A5 A4  =−  A5 A1

(2)

(3)

For deriving the 4th row, we shall use the 2nd and 3rd row values.

Row 1

A0

A2

A4

………………….

Row 2

A1

A3

A5

………………….

Row 3

B1

B3

B5

………………….

Row 4

C1

C3

C5

………………….

Here, the values for C1 and C3 are derived by using the values of the 2nd row and the 3rd row in the form of a determinant matrix, as shown in Eqs. (4) and (5).  1  A1 B1  B1  1  A1 C3 = −  B1 B1

C1 = −

 A1 B3 − B1 A3 A3  =− B3  B1  A1 B5 − B1 A5 A5  =−  B5 B1

(4)

(5)

From the 3rd step onwards, we can calculate the number of rows after the 4th row, by following the same process till the nth row. Now we shall try following the same method by using the same steps in coding the above criterion using the MATLAB software package.

322

A. Patil

2.2 Special Cases in the Stability Criterion There are two special cases when it comes to the Routh stability criterion. They can be classified as the following: Case 1: If the first term of each array row is null, while the remaining part of the row is not null. In this case a somewhat minimal value (ε), tends to zero rather than zero, is assumed. We can compute all the elements in the Routh sequence by replacing zero with (ε). After all elements have been determined, we extend the limit to any element that contains (ε). When the limit is overcome by each variable and if we have a positive limiting value, then we must conclude that it is stable, thus the function must be stable under all other circumstances. Case 2: If the entire number of elements in the Routh array is zero. We may say that the system has marginal stable characteristics in this situation. First, let’s understand the physical meaning of zero elements of every row [1]. The physical significance is that the roots of the characteristic equation in the plane are symmetrically positioned. In this case, we must consider the secondary equation in order to explore stability. The auxiliary equation can be constructed with the row elements in the Routh sequence just above the null-line. After the auxiliary equation is found, the auxiliary equation is separated to obtain null-line elements. If the current Routh array generated by auxiliary equations has no sign change, then we conclude that the structure is stable, although the given system is unstable in most of the other cases.

2.3 Advantages and Disadvantages of the Criterion The advantages of using the criterion can be summarized as the following points: 1. 2. 3.

Without solving the equation, we can find the stability of the system. The relative stability of a system can accurately be determined. By this approach, the intersection point for root sites with the imaginary axis can also be calculated [8]. The disadvantages of the criterion can be summarized as the following points:

1. 2. 3. 4. 5. 6.

Only a linear system is applicable for the criterion. The precise position of the poles on the right and left side of the plane S is not given. For any characteristic equation vector, the criterion is only valid for real coefficients. As the device order increases, individual determinants become difficult to solve. The right hand plane (RHP) is not sure how many roots are there. Marginal system stability is difficult to predict.

Routh Hurwitz Criterion for Stability …

323

3 Implementation Using MATLAB The following flowchart shows the process of determining whether a characteristic equation is stable or not. This flowchart will help us in breaking down the problem in hand into steps that can help us in making the model. The process has a total of five steps. The first step is to input the characteristic equation coefficients to the function call while running the code in the command window. When the input has been entered, the decision box checks whether the input added is in the form of an integer or a variable. Since the stability analysis checks the stability using integers as a determinant to find the real value of the equation, as shown in Eqs. (2), (3), (4) and (5), variables will not work. If by chance variables are entered, the process will stop and will ask the user to input an integer coefficient. If the coefficient is an integer, the process continues [2]. The next step is performing stability analysis using the Routh-Hurwitz criterion, by adding the values of the coefficients as a particular row in the same order as the characteristic equation. In the next step, the equation values are analyzed and the first three row values are found. If the first element is zero, then we have to further solve using limits and then check for sign changes by completing the Routh table. After the table is analyzed and the values are found, the characteristic equation in hand, the Routh table, the number of sign changes and stability status of the equation are displayed in the command window (Fig. 1). In making the model, we shall ensure the following points and make each point as a function in the MATLAB code: 1. 2. 3.

4.

Complete Routh table: If the table does not have the first element as 0, then the table does not need to be further solved. Finish incomplete Routh table: If the table has the first element as 0, then we have to solve it further using limits and sign changes. Making the Routh table: After the stability status has been displayed in the command window, the complete Routh table for the equation should get displayed. Number of sign changes: The number of sign changes that take place before reporting the stability status as an output (Fig. 2).

Let us now implement each stability status using an application characteristic equation: A.

For a stable system:

Consider the characteristic equation s4 + 3 s3 + 3 s2 + 2 s + 1 = 0. Let’s solve the equation using the Routh array before we implement it on MATLAB: 1.

Check Routh-Hurwitz stability for required conditions. The equation has positive coefficients of the characteristic polynomial. This follows the specifications of the control system [7].

324

A. Patil

Fig. 1 Flowchart of the model to be made using MATLAB

Fig. 2 Four functions have been made in the same folder

2.

Make the Routh array for the given polynomial s4

1

3

s3

3

2

s2

(3×3)−(2×1) = 73 3   7 3 ×2 −(1×3)

(3×1)−(0×1) 3

s1

7 3

=

1 =

3 3

=1

5 7

(continued)

Routh Hurwitz Criterion for Stability …

325

(continued) s0

3.

1

Ensure that the Routh-Hurwitz stability is adequately conditional.

All the elements in the Routh list first column are positive. Within the first column of the Routh sequence there is no change in signs. The system is thus stable. As we can see, from Fig. 3, we get the same results as we got previously. Hence, a stable system has been identified by the model. B.

For an Unstable System—Special Case 1:

Consider the characteristic equation s4 + 2 s3 + s2 + 2 s + 1 = 0. We shall use the same method that we used for finding the stability status of the previous characteristic equation. 1.

2.

Check Routh-Hurwitz stability for required conditions. The equation has positive coefficients of the characteristic polynomial. This follows the specifications of the control system [7]. Make the Routh array for the given polynomial.

s4

1

1

s3

1

1

1 (continued)

Fig. 3 Implementing the above stable system in MATLAB

326

A. Patil

(continued) (1×1)−(1×1) 1

s2

(1×1)−(0×1) 1

=0

=1

s1 s0

The s3 row elements have a common factor of 2. So, we divide all the elements by 2. 3.

4.

Here, it can be said that the Routh criterion uses a special case, which is the first element in the s2 row is zero. So, replace it with ε and resume the Routh table cycle till the very end. s4

1

1

s3

1

1

s2

ε

s1

(ε×1)−(1×1) ε

s0

1

1

1 =

ε−1 ε

Verifying the condition for Routh-Hurwitz stability. As ε tends to zero, we get

s4

1

1

s3

1

1

s2

0

1

s1

−∞

s0

1

1

Since there are two sign changes in the first column of the table, we can say that the system is unstable. The implementation of the above problem can be seen in Figure 4. C.

For an unstable system—Special Case 2

Consider the characteristic equation s5 + 3 s4 + s3 + 3 s2 + s + 3 = 0 Let’s solve the equation using the Routh array before we implement it on MATLAB: 1.

2.

Check Routh-Hurwitz stability for required conditions. The equation has positive coefficients of the characteristic polynomial. This follows the specifications of the control system [7]. Make the Routh array for the given polynomial.

Routh Hurwitz Criterion for Stability …

327

Fig. 4 Implementing the above unstable system in MATLAB

s5

1

1

1

s4

1

1

1

s3

(1×1)−(1×1) 1

(1×1)−(1×1) 1

=0

=0

s2 s1 s0

The s4 row elements have a common factor of 3. So we divide all the elements by 3. Here, we have a special case, which is, all row s3 elements are zero. So add A(s) of row s4 into the auxiliary equation. A(s) = s 4 + s 2 + 1

(6)

Differentiating Eq. (6) with respect to s, we get d A(s) = 4s 3 + 2s ds

(7)

328

A. Patil

Adding these coefficients in the s3 row, we get S5

1

1

1

S4

1

1

1

S3

2

1

S2

(2×1)−(0×1) 2

S1

(2×1)−(1×1) = 0.5 2 (0.5×1)−(1×2) = −1.5 0.5 0.5

S0

1

=1

= −3

Since there are two sign changes in the first column, we can say that the system is unstable. As we can see in Fig. 5, we have done the same steps as presented using the MATLAB software. Hence, an unstable system using the second special case has been identified.

Fig. 5 Implemented the above unstable system in MATLAB

Routh Hurwitz Criterion for Stability …

329

4 Conclusions For systems, stability acts as a really essential component. The analysis must therefore be conducted with great precision. In this study, a software package called MATLAB was used to code the stability analysis of Routh-Hurwitz criterion from scratch without using any MATLAB built-in functions [4]. This paper discusses all three scenarios of stability status by first solving the question using the Routh table and then implements it on MATLAB so as to understand the relation between the code and the steps mentioned. This analysis technique helps the user to evaluate the complete, conditional and dependent stability of the systems by entering the characteristic equation coefficients conveniently and easily. The code also gives a step-by-step explanation.

References 1. Chen, B., Chou, Y.-C., Cheng, H.H.: Open source ch control system toolkit and web based control system design for teaching automatic control of linear time-invariant systems. Comput. Appl. Eng. Educ. 21(1), 95–112 (2013). https://doi.org/10.1002/cae.20454 2. Ogata, K.: Modern Control Engineering. Prentice-Hall, NJ, Englewood Cliffs (1970) 3. Rao, M.V.C., Rao, P.V.: Some more comments On the Routh-Hurwitz criterion. IEEE Trans. Automat. Contr., AC-20, 714–716 (1975) 4. D’Azzo, J.J., Houpis, C.H.: Linear Control System Analysis and Design. McGraw-Hill, New York (1975) 5. Hurwitz, A.: On the conditions under which an equation has only roots with negative real parts. Math. Ann. 46, 273–284 (1895). Also in Selected Papers on Mathematical Trends in Control Theory, Dover, New York, pp. 70–82 (1964) 6. Mansour, M.: A Simple Proof of the Routh-Hurwitz Criterion, Report #88-04. Institute of Automatic Control & Industrial Electronics, Swiss Federal Institute of Technology (December 1988) 7. Routh, E.J.: A Treatise on the Stability of a Given State of Motion: Particularly Steady Motion. Macmillan (1877) 8. Nise, Norman: Control Systems Engineering, p. 9781118800829. Wiley, ISBN (2015)

Path Finding Algorithms D. D. V. Sai Ashish, Sakshi Munjal, Mayank Mani, and Sarthak Srivastava

Abstract In this survey we have compared five different path finding algorithms, namely, Dijkastra’s algorithm, A* Search, Greedy Best First Search, Breadth First Search and Depth First Search on the basis of the following criteria: Complexity: It refers to the rate at which the execution time grows with respect to the input size. Reliability: It refers to whether the algorithm is able to guarantee the shortest path. Weighted/unweighted: Some algorithms are not suited for weighted graphs and thus have a major limitation. Their respective advantages and disadvantages have also been discussed. Keywords Greedy algorithm · Path finding algorithm · Graph traversal

1 Introduction The paper concerns itself with finding the shortest path between two (or more) vertices in a graph. Path finding algorithms try to determine the best path given some starting and ending node. The usage of pathfinding algorithms is not just limited to navigation systems. The same overarching idea that they explore can be applied to other applications, like analyzing social networks as well. These algorithms have a wide range of applications like Google maps, satellite navigation systems and routing packets over the internet. D. D. V. S. Ashish · S. Munjal · M. Mani (B) · S. Srivastava Department of Computer Science Engineering, Maharaja Surajmal Institute of Technology, Delhi, India e-mail: [email protected] D. D. V. S. Ashish e-mail: [email protected] S. Munjal e-mail: [email protected] S. Srivastava e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 A. E. Hassanien et al. (eds.), Emerging Technologies in Data Mining and Information Security, Advances in Intelligent Systems and Computing 1286, https://doi.org/10.1007/978-981-15-9927-9_33

331

332

D. D. V. S. Ashish et al.

2 Dijkstra’s Algorithm Dijkstra’s algorithm is one of the algorithms for determining the shortest path from a starting vertex to a target vertex in a graph. It creates the shortest path tree connecting the starting vertex to all of the other vertices. One prerequisite for using the algorithm is that the graph should not have any negatively weighted edges. Algorithm ● ●

The graph has the following: ○ Vertices, or nodes, represented by either v or u; ○ (u,v) represents an edge, and w(u,v) represents its weight. The following values are initialized: ○ distance, an array of distance between the source vertex s and each vertex in the graph:

(1) The distance will be recalculated as the algorithm proceeds. ○

●

q, a queue of all vertices in the graph that have not been visited. At the beginning it holds all the vertices and at the end, q will be empty. ○ s, an empty set,it contains all the vertices the algorithm has visited. s will contain all the vertices of the graph when the algorithm finishes. The algorithm proceeds as follows: ○ While q is not empty, pop an unvisited vertex v from q (that is not in s) with the smallest distance(v). In the first iteration, starting vertex s will be selected because distance(s) was set to 0. v is added to s as it has now been visited. ○ Update distance values of all the adjacent vertices of the v as follows: distance(v) = min( dist (v) + weight(u,v) , dist (u) )

(2)

Complexity: The time complexity of Dijkstra’s algorithm is O(Vertices2 ) but if a minimum priority queue is used it drops down to O(Vertices + Edge * log(Vertices)) [1]. Reliability: It guarantees the shortest path. Since it explores every possible vertex without any bias until it finds the target, the algorithm is very reliable in terms of its results albeit a bit slow. Weighted/Unweighted: It can be applied to both weighted and unweighted graphs. However, the weighted graph must only contain non-negative weights as it fails in case of negatively weighted edges. Advantages: Dijkstra works on both directed or undirected graphs. Disadvantages: • It executes blind search and wastes a lot of time while processing. • It cannot handle negative edges.

Path Finding Algorithms

333

Applications: • It is used in Google maps. • It finds applications in IP routing to determine open shortest path first. • Telephone networks use Dijkstra.

3 A* Search A* works similar to Dijkstra, that is by creating a lowest-cost path tree from the starting vertex to the target vertex. The difference that makes A* better is that for each vertex, A* uses a function f (n) that provides a rough cost estimate of a path through that vertex. For path finding, A* algorithm repeatedly examines the most promising unexplored location it has seen. When a location is explored, the algorithm is finished if that location is the goal; otherwise, it makes note of all that location’s neighbors for further exploration. Therefore, A* is a heuristic-based algorithm [3]. A* expands paths that are already less expensive by using the function: f (n) = g(n) + h(n), f (n) = g(n) = h(n) =

(3)

estimated cost of path using vertex n. cost to get to vertex n. cost estimate from n to target. This is the heuristic part of the cost function, so it is a guess.

Heuristics To calculate h(n) we have two methods: ● ●

Calculating the exact value of h(n). Approximate the value of h(n) using heuristics. We will be exploring the approximate methods going forward. Three methods of calculating the heuristic distance are discussed: ○

Manhattan Distance - Sum of absolute values of differences in the coordinates of current cell’s & the target’s x & y respectively, i.e.,

○

Diagonal Distance - Maximum out of absolute values of differences in the coordinates of target’s x & y & the current cell’s x and y respectively, i.e.,

(4)

(5) ○

Euclidean Distance - It is the distance between the target vertex & the current vertex using the distance formula (6)

334

D. D. V. S. Ashish et al.

Algorithm We make two lists – OpenList and ClosedList. ● ● ●

● ●

Initialize the OpenList. Initialize the ClosedList and put the starting vertex on the OpenList (with its f at zero). Repeat the following steps: ○ Look for the vertex which has the lowest f on the OpenList. Refer to this vertex as the current vertex. ○ Switch it to the ClosedList. ○ For each reachable vertex from the current vertex: ■ If on the ClosedList, ignore it. ■ Add it to the OpenList if it isn’t on OpenList. Current vertex should then be made parent of this vertex. Record the f, g & h value of this vertex. ■ If it’s on the OpenList already, check to see if this is a better path. If it is a better path, recalculate the f and g value and change its parent node to the current node. Stop when: ○ The target vertex is added to the ClosedList. ○ The target vertex was not found and the OpenList is empty. Tracing backwards from the target vertex to the starting vertex gives us the path.

Complexity: The worst-case time complexity of A* Search is O(Edge). Reliability: Arguably the best path finding algorithm. It guarantees the shortest path much faster than Dijkstra’s algorithm using heuristics. Weighted/Unweighted: It works on both weighted and unweighted graphs. Moreover, it does not have the same limitation that we found in Dijkstra, i.e. it can handle negative edges as well. It can work for negative weights if there are a finite number of edges with f-value less than that of optimal path cost. Advantages • It’s space–time trade-off is good. • It has optimal efficiency. Disadvantages • The main drawback of the A* algorithm is the memory requirement as the entire OpenList has to be saved. The A* algorithm is severely space-limited. • As A* search algorithm depends heavily on approximations to calculate h, it doesn’t always find the shortest path. Applications: A* is often used in applications such as video games. It also finds applications in diverse problems, including “parsing using stochastic grammars in NLP”.

4 Greedy Best-First Search In BFS, when we are at a vertex, we can traverse to any of the neighboring vertices as the next vertex. Best-first search uses an evaluation function to determine which adjacent is most promising and then explore [2].

Path Finding Algorithms

335

Algorithm Best-First-Search(Graph g, Node s) ● ● ●

●

Generate an empty Priority Queue pq Insert s in pq. While pq is not empty, pop u ○ If u is the target, exit ○ Else for each neighbor v of u ○ If v is unvisited mark v as visited and insert it in pq. ○ Mark u as “Examined" End procedure

Complexity: The time complexity for best-first search is O(n * Log n), where n is number of vertices. In the worst case, all the vertices have to be visited before reaching the target. Performance of the algorithm is dependent on the efficient design of cost function. Reliability: It is a faster, more heuristic-dependent version of A* but it does not guarantee the shortest path. Weighted/Unweighted: It works on both weighted and graphs equally well. The weights may be positive or negative. Advantages: • Best-first search gains the advantage of both BFS and DFS by switching between them. • Best-first search is more efficient than BFS and DFS algorithms. • It is efficient in terms of time complexity. Disadvantages • In the worst-case scenario, it behaves as an unguided depth-first search. • As DFS, it can get stuck in a loop.

5 Breadth-First Search BFS is a building block of many graph algorithms. It is popularly used to test for connectivity and to compute the single source shortest paths of unweighted graphs [4].

336

D. D. V. S. Ashish et al.

Algorithm front starts from the vertex ‘vertices’ and expands outwards during each step. It visits all the vertices in the same depth first before moving to the next depth. In BFS the top down approach is followed where each vertex verifies if all its adjacent vertices are visited or not. The following algorithm is a Single Step Top-Down approach in a conventional BFS algorithm. ● ● ● ●

● ●

front