CDE4301

Table of Contents

Navigate through the sections of this report using the links below

Acknowledgement

We extend our deepest gratitude to our project supervisors, Dr. Jovan Tan and Dr. Elliot Law for their invaluable support, insightful feedback, and encouragement throughout the project.

We also wish to thank Mr Qin Zheng for his steadfast support in providing the essential laboratory resources and space for our experimental data.

Prof. Jovan
Dr. Jovan Tan
Lecturer
Engineering Design & Innovation Centre, NUS
Prof. Elliot
Dr. Elliot Law
Senior Lecturer
Engineering Design & Innovation Centre, NUS
Section 1 Introduction →
Abstract
1.1 First Principle Modelling (FPM)
1.2 Machine Learning (ML)
1.3 Physics-Informed Machine Learning (PIML)
1.4 Objective & Scope
Section 2 Development of Case Study →
2.1 Case Study Selection
2.2 Case Study 2.2.1 Working Principles 2.2.2 Industrial Applications in Control Systems
Section 3 Experimental Data Collection →
3.1 Experimental Set-up
3.2 Sampling of Data Points
Section 4 Architectural Design for PIML →
4.1 Overview
4.2 Recurrent Neural Network 4.2.1 Gated Recurrent Units 4.2.2 Long Short Term Memory
4.3 Neural Ordinary Differential Equation (NODE) 4.3.1 NODE-ED 4.3.2 NODE-ED-Uncoupled 4.3.3 NODE-Context 4.3.4 NODE-Context-Uncoupled
4.4 Performance Metric & Statistical Significance
4.5 Training Methodology 4.5.1 Hyperparameter Optimisation 4.5.2 K-fold Cross-validation 4.5.3 100 repeated K-fold 4.5.4 Bagging
4.6 Analysis of results for RNN Variants
4.7 Analysis of results for NODE variants
4.8 Comparison between RNN and NODE
Section 5 Physics Guided Enhancements →
5.1 Steady State Constraint with Parameter Discovery (SS-PD)
5.2 Steady State Collocation with Parameter Discovery (SS-C-PD)
5.3 Analysis of results for Physics Guided Enhancements
Section 6 Current Limitations & Future Works →
Section 7 Conclusion →
Section 8 Appendix →
Next →