This book explores new developments, ideas, and concepts in intelligent edge-embedded technologies for the digitising industry. The work is based on recent research results and activities in edge industrial computing, artificial intelligence (AI), the Industrial Internet of Things (IIoT) and digital twin technologies. Each chapter builds on the research, developments and innovative ideas generated by AI4DI, ANDANTE and TEMPO ECSEL JU, as well as other European research projects.
The evolution towards an environmentally friendly Industry 5.0 brings more industrial edge devices that include embedded intelligence, with enough computing power and sufficiently advanced programming to make operational decisions based on local data and independent of external advice.
These new intelligent edge-embedded devices act on the data they capture or generate about evolving conditions in the industrial process and change the machine's behaviour to optimise the operation and its functionalities. Connected, these intelligent edge IIoT devices can act on information generated across multiple intelligent edge devices and even across various types of intelligent edge devices to create more intelligent behaviours for industrial manufacturing processes.
The embedded intelligence technologies implement artificial intelligence (AI) capabilities into the edge device itself, so the device can learn, analyse, and act autonomously.
Intelligent edge systems implemented on-premises improve industrial manufacturing facilities' outcomes by making instantaneous, autonomous, or semi-autonomous decisions independent of external cloud computing capabilities.
The specific interest here lies in the advancement of the convergence of edge computing and AI technologies in edge industrial application areas. This is examined by introducing the concepts of sustainable industrial-edge AI technologies and industrial-edge AI for sustainability.
Insights from recent research on key AI technologies that support the development of industrial-edge AI applications for the digitising industry are presented. The concept of AI at the edge is introduced, and the edge continuum components and their distribution across the micro-, deep- and meta-edge continuum are explained.
Moreover, the book discusses how to build AI models (e.g., model training and inference) on edge and provides insights into this new interdisciplinary field of intelligent industrial edge from a broader perspective.
The authors examine the technologies and hardware for neuromorphic computing, highlighting emerging in-memory computing techniques, the implementation of resistive synapses for neural networks, neuromorphic reference architectures, and the tools and methodologies for training and mapping neural networks on hardware targets.
Furthermore, the book reviews the core concepts for edge AI advancements in the digitising industry and the impact of AI and digital twins on IIoT. Finally, the book addresses ethical considerations and the trustworthiness of industrial AI systems’ core concepts, as well as the current challenges of AI standardisation in the digitising industry.
This book's target audience includes academics, research scholars, industrial experts, scientists, and postgraduate students working in industrial edge-embedded intelligence hardware, software, and algorithms to add machine learning and deep learning to enhance the industrial edge processing capabilities.
Author(s): Ovidiu Vermesan, Mario Diaz Nava
Series: River Publishers Series in Communications and Networking
Publisher: River Publishers
Year: 2022
Language: English
Pages: 339
City: Gistrup
Front Cover
Intelligent Edge-Embedded Technologies for Digitising Industry
Dedication
Acknowledgement
Contents
Preface
List of Figures
List of Tables
List of Contributors
List of Abbreviations
1 Industrial AI Technologies for Next-Generation Autonomous Operations with Sustainable Performance
1.1 Industrial AI
1.1.1 Challenges of Industrial AI versus Consumer AI
1.1.2 Sustainable AI
1.2 Capabilities Spectrum of Industrial AI
1.3 The Industrial AI Spectrum
1.3.1 Narrow AI vs. General AI
1.3.2 Weak AI vs. Strong AI
1.3.3 Basic AI vs. Super AI
1.3.4 Red AI vs. Green AI
1.4 AI Problem Solving Domains
1.4.1 Expert Systems
1.4.2 Machine Vision
1.4.3 Robotics
1.4.4 Biomimicry
1.4.5 Genetic and Evolutionary Algorithms
1.4.6 Generative AI
1.4.7 Artificial Swarm Intelligence
1.4.8 Natural Language Processing
1.4.9 Machine Learning
1.4.10 Neural Networks
1.4.11 Automated Planning and Plan Recognition
1.4.12 AI for the Metaverse
1.5 Edge AI Continuum
1.6 Symbolic AI – ML Continuum
1.7 Logic-based AI: Knowledge Representation and Reasoning
1.8 Hardware/Software Technology Stack
1.8.1 ML Methods and Techniques
1.8.2 Neural Networks Architectures
1.8.3 Industrial Embedded AI/ML
1.8.4 On-device ML Applications Enabling True Edge Computing
1.8.5 Machine Learning on Embedded Devices
1.8.6 Embedded ML Development Flow in Industrial Setting
1.9 Summary
References
2 Technology and Hardware for Neuromorphic Computing
2.1 Mobile Devices Call for Efficient Neuromorphic Computing
2.2 Neuromorphic Hardware Enables Next Generation AI
2.3 Building Neuromorphic Hardware
2.3.1 Approach to Realise the Emerging Technologies
2.3.2 Approach to Derive the Hardware Architectures and Designs
2.3.3 Approach Related to Neuromorphic Algorithms and Applications
2.4 Positioning Within the Neuromorphic Computing Landscape
2.5 Targeted Use Cases and Application Domains
2.5.1 Food – Food Classification
2.5.2 Automotive – Object Recognition and Sound Localization
2.5.3 Digital Industry – Pattern Recognition (Keyword Spotting)
2.5.4 Consumer – Coaching Biomechanical Assistance (Running)
2.5.5 Medical Health – Medical Image Denoising
2.6 Neuromorphic Hardware Technologies Being Developed
2.7 Conclusion
References
3 Tools and Methodologies for Training, Profiling, and Mapping a Neural Network on a Hardware Target
3.1 Introduction
3.1.1 Edge Computing Benefices and Challenges
3.1.2 Artificial Neural Networks (ANNs) and Spiking Neural Networks (SNNs)
3.2 State-of-the-art of key aspects of Neural Networks
3.2.1 ANN and SNN Hardware Aware Design
3.2.2 Sparsity
3.2.3 ANN-to-SNN Conversion
3.2.4 Surrogate Gradient Descent
3.2.5 Neural Engineering Object (Nengo) Simulator
3.3 NN Transformation: Temporal Delta Layer
3.3.1 Temporal Delta Layer: Training Towards Brain Inspired Temporal Sparsity for Energy Efficient Deep Neural Networks
3.3.2 Related Works
3.3.3 Methodology
3.3.3.1 Delta inference
3.3.3.2 Activation quantization to induce sparsity
3.3.3.3 Fixed point quantization
3.3.3.4 Learned step-size quantization
3.3.3.5 Sparsity penalty
3.3.3.6 Proposed algorithms
3.3.4 Experiments and Results
3.3.4.1 Baseline
3.3.4.2 Experiments
3.3.4.3 Accuracy v/s Activation sparsity
3.4 NN Compiler for Dedicated Inference Accelerator Hardware
3.4.1 Compiler Components
3.4.2 ONNX Parser
3.4.3 Hardware Architecture Representation
3.4.4 Mapper
3.4.5 Mapping Strategy
3.4.6 Mapping of Deep Spiking NN Architectures to Digital SNN Inference Devices
3.5 Simulator/Profiler
3.6 Conclusions
3.6.1 On NN Model Transformation
3.6.2 On NN Compiler for Dedicated Inference Accelerator Hardware with Analog In-Memory Computing Conclusion
3.6.3 Simulator/Profiler
References
4 Using FeFETs as Resistive Synapses in Crossbar-based Analog MAC Accelerating Units
4.1 Introduction and Background
4.2 Requirements of Crossbar Structure on eNVMs
4.3 Synapse Design
4.3.1 Conventional Design
4.3.2 Gate-Cascaded FeFETs
4.3.3 Exploration Results
4.4 Conclusion
References
5 Emerging In-memory Computing for Neural Networks
5.1 Memory Technologies
5.1.1 Volatile Memories
5.1.2 Non Volatile Memories
5.2 In-memory Architecture
5.2.1 Computational Domain
5.2.1.1 Mixed signal approach
5.2.1.2 Digital approach
5.2.2 Target Network Quantization
5.2.2.1 Floating point architectures
5.2.2.2 Fixed-point architectures
5.2.2.3 Binarized architectures
5.2.2.4 Flexible precision architectures
References
6 Artificial Intelligence Advancements for Digitising Industry
6.1 AI at the Edge in Industrial Processes
6.2 A pan-European AI Framework for Manufacturing and Process Technology
6.3 AI Technologies
6.4 AI Application Areas
6.4.1 Automotive
6.4.2 Semiconductor
6.4.3 Industrial Machinery
6.4.4 Food and Beverage
6.4.5 Transportation
6.5 AI Technology Roadmap for Digitising Industry
6.6 Conclusion
References
7 Impact of AI and Digital Twins on IIoT
7.1 Introduction to the Hexa-X Project
7.2 An Ecosystem Concept for Digital Twins in IIoT
7.3 Digital Twins for Emergent Intelligence
7.4 Network-aware Digital Twins for Local Insight Generation
7.5 AI at the Intersection between DTs and HMI in Industrial IoT
7.6 Conclusion
References
8 Lesson Learnt and Future of AI Applied to Manufacturing
8.1 Introduction
8.2 IoT Enabled by Machine Learning
8.3 Machine Learning at the Edge
8.3.1 Applications of EdgeML in Industrial IoT
8.3.2 Challenges in EdgeML
8.4 Federated Learning – A Solution to Train ML Models
8.4.1 Applications for Federated Learning in Industrial IoT
8.4.2 Federated Learning Scenarios
8.4.3 Challenges in Federated Learning
8.4.4 Frameworks and products for leveraging Federated Learning
8.5 Reducing Complexity of RX Processing
8.6 Enhancing Reliability by Multi-Connectivity in the Uplink
8.7 Communications in an “Embodied Artificial Intelligence” Future
8.8 Embodied Artificial Intelligence
8.9 High Integration as a Central Technological Driver
8.10 Conclusion
References
9 Ethical Considerations and Trustworthy Industrial AI Systems
9.1 Introduction
9.2 Ethics and Responsible AI in Industrial Environments
9.3 Requirements for Industry-Grade AI
9.4 Industrial AI Challenges
9.4.1 Complexity
9.4.2 Use of Natural Resources
9.4.3 Pollution and Waste
9.4.4 Energy
9.5 Ethical Considerations for Digitising Industry
9.5.1 AI Trustworthiness
9.5.2 Bias and Fairness
9.5.3 Transparency
9.5.4 Accountability
9.5.5 Explainability
9.5.6 Control
9.5.7 Human-Machine Interaction and Manipulation of Behaviour
9.5.8 Autonomous Industrial Systems
9.5.9 Machine Ethics
9.5.10 Automation and Employment
9.6 AI and the Future Digitising Industry
9.7 Ethical Guidelines for AI in Industrial Environments
9.8 Recommendations for Ethical AI in Industrial Environments
9.9 Conclusion
References
10 Current Challenges of AI Standardisation in the Digitising Industry
10.1 Introduction
10.2 International Principles
10.3 Role of AI Standardisation in Digitising Industry
10.4 Challenges Associated with AI Deployments in Industrial Environments
10.5 AI Standardisation Needs in Industrial Automation
10.6 Standardisation of Security and Safety in AI Systems
10.7 The Global AI Standards Landscape and Standardisation Activities
10.7.1 CEN-CENELEC
10.7.2 ETSI
10.7.3 IEC
10.7.4 ISO
10.7.5 IEEE
10.7.6 IETF
10.7.7 ITU-T
10.8 AI Certification
10.9 Recommendations for an AI Standardisation Roadmap
10.10 Conclusion
References
Index
About the Editors
Back Cover
