Industrial Artificial Intelligence Technologies and Applications

The advances in industrial edge artificial intelligence (AI) are transforming the way industrial equipment and machine interact with the real world, with other machines and humans during manufacturing processes. These advances allow industrial internet of things (IIoT) and edge devices to make decisions during the manufacturing processes using sensors and actuators data. Digital transformation is reshaping the manufacturing industry, and industrial edge AI aims to combine the potential advantages of edge computing (low latency times, reduced bandwidth, distributed architecture, improved trustworthiness, etc.) with the benefits of AI (intelligent processing, predictive solutions, classification, reasoning, etc.). The industrial environments allow the deployment of highly distributed intelligent industrial applications in remote sites that require reliable connectivity over wireless and cellular connections. Intelligent connectivity combines IIoT, wireless/cellular and AI technologies to support new autonomous industrial applications by enabling AI capabilities at the edge and allowing manufacturing companies to improve operational efficiency and reduce risks and costs for industrial applications. There are several critical issues to consider when bringing AI to industrial IoT applications considering training AI models at the edge, the deployment of the AI-trained inferencing models on the target reliable edge hardware platforms and the benchmarking of the solution compared with other implementations. The next-generation trustworthy industrial AI systems offer dependability by design, transparency, explainability, verifiability, and standardised industrial solutions to be implemented into various applications across different industrial sectors. New AI techniques like embedded machine learning (ML) and deep learning (DL) capture edge data, employ AI models and deploy them to hardware target edge devices from ultra-low-power microcontrollers to embedded devices, gateways, and on-premises servers for industrial applications. These techniques reduce latency, increase scalability, reliability, and resilience, and optimise wireless connectivity, greatly expanding IIoT capabilities. The book overviews the latest research results and activities in industrial artificial intelligence technologies and applications based on the innovative research, developments and ideas generated by the ECSEL JU AI4DI, ANDANTE and TEMPO projects. The authors describe industrial AI's challenges, the approaches adopted, and the main industrial systems and applications to give the reader a good insight into the technical essence of the field. The articles provide insightful material on industrial AI technologies and applications. The book is a valuable resource for researchers, post-graduate students, practitioners, and technology developers interested in gaining insight into the industrial edge AI, IIoT, embedded machine and deep learning, new technologies, and solutions to advance the intelligent processing at the edge. River Publishers Series in Communications and Networking Front Cover 1 Industrial Artificial Intelligence Technologies and Applications 4 Dedication 6 Acknowledgement 6 Contents 8 Preface 16 List of Figures 20 List of Tables 26 List of Contributors 30 1 Benchmarking Neuromorphic Computing for Inference 34 1.1 Introduction 35 1.2 State-of-the-art in Benchmarking 36 1.2.1 Machine Learning 38 1.2.2 Hardware 40 1.3 Guidelines 42 1.3.1 Fair and Unfair Benchmarking 43 1.3.2 Combined KPIs and Approaches for Benchmarking 44 1.3.3 Outlook : Use-case Based Benchmarking 46 1.4 Conclusion 48 References 49 2 Benchmarking the Epiphany Processor as a Reference Neuromorphic Architecture 54 2.1 Introduction and Background 54 2.2 Comparison with a Few Well-Known Digital Neuromorphic Platforms 57 2.3 Major Challenges in Neuromorphic Architectures 59 2.3.1 Memory Allocation 59 2.3.2 Efficient Communication 61 2.3.3 Mapping SNN onto Hardware 62 2.3.4 On-chip Learning 62 2.3.5 Idle Power Consumption 63 2.4 Measurements from Epiphany 63 2.5 Conclusion 65 References 66 3 Temporal Delta Layer: Exploiting Temporal Sparsity in Deep Neural Networks for Time-Series Data 68 3.1 Introduction 69 3.2 Related Works 70 3.3 Methodology 72 3.3.1 Delta Inference 72 3.3.2 Sparsity Induction Using Activation Quantization 73 3.3.2.1 Fixed Point Quantization 74 3.3.2.2 Learned Step-Size Quantization 75 3.3.3 Sparsity Penalty 77 3.4 Experiments and Results 78 3.4.1 Baseline 78 3.4.2 Experiments 78 3.4.3 Result Analysis 79 3.5 Conclusion 82 References 83 4 An End-to-End AI-based Automated Process for Semiconductor Device Parameter Extraction 86 4.1 Introduction 87 4.2 Semantic Segmentation 89 4.2.1 Proof of Concept and Architecture Overview 89 4.2.2 Implementation Details and Result Overview 94 4.3 Parameter Extraction 97 4.4 Conclusion 101 4.5 Future Work 102 References 102 5 AI Machine Vision System forWafer Defect Detection 106 5.1 Introduction and Background 106 5.2 Machine Vision-based System Description 108 5.3 Conclusion 111 References 112 6 Failure Detection in Silicon Package 114 6.1 Introduction and Background 115 6.2 Dataset Description 116 6.2.1 Data Collection and Labelling 117 6.3 Development and Deployment 118 6.4 Transfer Learning and Scalability 119 6.5 Result and Discussion 120 6.6 Conclusion and Outlooks 122 References 122 7 S2ORC-SemiCause: Annotating and Analysing Causality in the Semiconductor Domain 124 7.1 Introduction 125 7.2 Dataset Creation 126 7.2.1 Corpus 126 7.2.2 Annotation Guideline 126 7.2.3 Annotation Methodology 127 7.2.4 Dataset Statistics 128 7.2.5 Causal Cue Phrases 128 7.3 Baseline Performance 129 7.3.1 Train-Test Split 129 7.3.2 Causal Argument Extraction 129 7.3.3 Error Analysis 130 7.4 Conclusions 132 References 132 8 Feasibility ofWafer Exchange for European Edge AI Pilot Lines 136 8.1 Introduction 137 8.2 Technical Details and Comparison 138 8.2.1 Comparison TXRF and VPD-ICPMS Equipment for Surface Analysis 138 8.2.2 VPD-ICPMS Analyses on Bevel 141 8.3 Cross-Contamination Check-Investigation 142 8.3.1 Example for the Comparison of the Institutes 142 8.4 Conclusiion 144 References 145 9 A Framework for Integrating Automated Diagnosis into Simulation 146 9.1 Introduction 146 9.2 Model-based Diagnosis 148 9.3 Simulation and Diagnosis Framework 151 9.3.1 FMU Simulation Tool 151 9.3.2 ASP Diagnose Tool 153 9.4 Experiment 154 9.5 Conclusion 158 References 160 10 Deploying a Convolutional Neural Network on Edge MCU and Neuromorphic Hardware Platforms 162 10.1 Introduction 162 10.2 Related Work 163 10.3 Methods 164 10.3.1 Neural Network Deployment 164 10.3.1.1 Task and Model 165 10.3.1.2 Experimental Setup 165 10.3.1.3 Deployment 166 10.3.2 Measuring the Ease of Deployment 168 10.4 Results 169 10.4.1 Inference Results 169 10.4.2 Perceived Effort 170 10.5 Conclusion 170 References 171 11 Efficient Edge Deployment Demonstrated on YOLOv5 and Coral Edge TPU 174 11.1 Introduction 174 11.2 Related Work 175 11.3 Experimental Setup 176 11.3.1 Google Coral Edge TPU 176 11.3.2 YOLOv5 178 11.4 Performance Considerations 178 11.4.1 Graph Optimization 178 11.4.1.1 Incompatible Operations 178 11.4.1.2 Tensor Transformations 179 11.4.2 Performance Evaluation 180 11.4.2.1 Speed-Accuracy Comparison 180 11.4.2.2 USB Speed Comparison 183 11.4.3 Deployment Pipeline 184 11.5 Conclusion and Future Work 185 References 185 12 Embedded Edge Intelligent Processing for End-To-End Predictive Maintenance in Industrial Applications 190 12.1 Introduction and Background 191 12.2 Machine and Deep Learning for Embedded Edge Predictive Maintenance 192 12.3 Approaches for Predictive Maintenance 194 12.3.1 Hardware and Software Platforms 195 12.3.2 Motor Classification Use Case 196 12.4 Experimental Setup 196 12.4.1 Signal Data Acquisition and Pre-processing 197 12.4.2 Feature Extraction, ML/DL Model Selection and Training 198 12.4.3 Optimisation and Tuning Performance 200 12.4.4 Testing 202 12.4.5 Deployment 203 12.4.6 Inference 205 12.5 Discussion and Future Work 206 References 207 13 AI-Driven Strategies to Implement a Grapevine Downy MildewWarning System 210 13.1 Introduction 210 13.2 Research Material and Methodology 212 13.2.1 Datasets 212 13.2.2 Labelling Methodology 213 13.3 Machine Learning Models 213 13.4 Results 216 13.4.1 Primary Mildew Infection Alerts 216 13.4.2 Secondary Mildew Infection Alerts 217 13.5 Discussion 218 13.6 Conclusion 219 References 220 14 On the Verification of Diagnosis Models 222 14.1 Introduction 222 14.2 The Model Testing Challenge 225 14.3 Use Case 227 14.4 Open Issues and Challenges 231 14.5 Conclusion 234 References 234 Index 238 About the Editors 240 Back Cover 244 River,Publishers,Series,in,Communications,and,Networking The advances in industrial edge artificial intelligence (AI) are transforming the way industrial equipment and machines interact with the real world, with other machines and humans during manufacturing processes. These advances allow Industrial Internet of Things (IIoT) and edge devices to make decisions during the manufacturing processes using sensors and actuators.Digital transformation is reshaping the manufacturing industry, and industrial edge AI aims to combine the potential advantages of edge computing (low latency times, reduced bandwidth, distributed architecture, improved trustworthiness, etc.) with the benefits of AI (intelligent processing, predictive solutions, classification, reasoning, etc.).The industrial environments allow the deployment of highly distributed intelligent industrial applications in remote sites that require reliable connectivity over wireless and cellular connections. Intelligent connectivity combines IIoT, wireless/cellular and AI technologies to support new autonomous industrial applications by enabling AI capabilities at the edge and allowing manufacturing companies to improve operational efficiency and reduce risks and costs for industrial applications.There are several critical issues to consider when introducing AI to industrial IoT applications considering training AI models at the edge, the deployment of the AI-trained inferencing models on the target edge hardware platforms, and the benchmarking of solutions compared to other implementations.Next-generation trustworthy industrial AI systems offer dependability in terms of their design, transparency, explainability, verifiability, and standardised industrial solutions can be implemented in various applications across different industrial sectors.New AI techniques such as embedded machine learning (ML) and deep learning (DL), capture edge data, employ AI models, and deploy these in hardware target edge devices, from ultra-low-power microcontrollers to embedded devices, gateways, and on-premises servers for industrial applications. These techniques reduce latency, increase scalability, reliability, and resilience; and optimise wireless connectivity, greatly expanding the capabilities of the IIoT.This book provides an overview of the latest research results and activities in industrial AI technologies and applications, based on the innovative research, developments and ideas generated by the ECSEL JU AI4DI, ANDANTE and TEMPO projects.The authors describe industrial AI's challenges, the approaches adopted, and the main industrial systems and applications to give the reader extensive insight into the technical nature of this field. The chapters provide insightful material on industrial AI technologies and applications.This book is a valuable resource for researchers, post-graduate students, practitioners, and technoloyg developers interested in gaining insight into industrial edge AI, the IIoT, embedded machine and deep learning, new technologies, and solutions to advance intelligent processing at the edge.The Open Access version of this book, available at http://www.taylorfrancis.com, has been made available under a Creative Commons Attribution-Non-Commercial (CC-BY-NC) 4.0 International License. The advances in industrial edge artificial intelligence (AI) are transforming the way industrial equipment and machines interact with the real world, with other machines and humans during manufacturing processes. These advances allow Industrial Internet of Things (IIoT) and edge devices to make decisions during the manufacturing processes using sensors and actuators. Digital transformation is reshaping the manufacturing industry, and industrial edge AI aims to combine the potential advantages of edge computing (low latency times, reduced bandwidth, distributed architecture, improved trustworthiness, etc.) with the benefits of AI (intelligent processing, predictive solutions, classification, reasoning, etc.). The industrial environments allow the deployment of highly distributed intelligent industrial applications in remote sites that require reliable connectivity over wireless and cellular connections. Intelligent connectivity combines IIoT, wireless/cellular and AI technologies to support new autonomous industrial applications by enabling AI capabilities at the edge and allowing manufacturing companies to improve operational efficiency and reduce risks and costs for industrial applications. There are several critical issues to consider when introducing AI to industrial IoT applications considering training AI models at the edge, the deployment of the AI-trained inferencing models on the target edge hardware platforms, and the benchmarking of solutions compared to other implementations. Next-generation trustworthy industrial AI systems offer dependability in terms of their design, transparency, explainability, verifiability, and standardised industrial solutions can be implemented in various applications across different industrial sectors. New AI techniques such as embedded machine learning (ML) and deep learning (DL), capture edge data, employ AI models, and deploy these in hardware target edge devices, from ultra-low-power microcontrollers to embedded devices, gateways, and on-premises servers for industrial applications. These techniques reduce latency, increase scalability, reliability, and resilience; and optimise wireless connectivity, greatly expanding the capabilities of the IIoT. This book provides an overview of the latest research results and activities in industrial AI technologies and applications, based on the innovative research, developments and ideas generated by the ECSEL JU AI4DI, ANDANTE and TEMPO projects. The authors describe industrial AI's challenges, the approaches adopted, and the main industrial systems and applications to give the reader extensive insight into the technical nature of this field. The chapters provide insightful material on industrial AI technologies and applications. This book is a valuable resource for researchers, post-graduate students, practitioners, and technoloyg developers interested in gaining insight into industrial edge AI, the IIoT, embedded machine and deep learning, new technologies, and solutions to advance intelligent processing at the edge