Application of Deep Learning in Process Fault Diagnosis, Modelling and Optimization

Abstract

Artificial Neural Network (ANN) has gained great interest in industrial applications due to their supremacy in modelling complex process behaviour. Applications of ANN include process modelling, optimization and fault diagnosis. However, pure data-driven approaches that use only observations to approximate the system dynamics may not be ideal candidates, especially when substantial physics-based knowledge is available but not utilized. This study refers to such knowledge as Domain Knowledge (DK). This M.Sc. research is focused on fault diagnosis, modelling and optimization of industrial plants using ANN techniques. This thesis is mainly divided into two parts. First, Chapter 3 outlines an empirical study of the Long Short-Term Memory (LSTM) network, one of the most well-known Recurrent Neural Network (RNN) algorithms. During this study, we can further understand how the LSTM algorithm works to capture temporal features. Then, we propose adding an Autoregressive Integrated Moving Average with eXogenous input (ARIMAX) on top of the LSTM network to handle faults related to the closed-loop system. The proposed framework is applied to the Wind Turbine System (WTS) to detect blade and pitch system faults. The second part mainly focuses on incorporating DK into the neural network structure. Hence, Chapter 4 describes the Physics-Informed Neural Network (PINN) framework based on Ensemble Sequential Learning and the Mixture Density Network (ESL-MDN). The proposed model estimates the components that build the cost function and constraints of the Differential Evolution (DE) optimizer. Finally, the proposed method has been validated by data collected from Reverse Osmosis Water Desalination (ROWD) plant to demonstrate a reduction in energy consumption, achieved by the optimization strategy under different scenarios.