Deep learning-based fault diagnostics and prognostics for industrial systems

Abstract

With an increasing adoption of sensors and Internet-of-Things (IoT) in complex industrial applications, new opportunities arise for developing data-driven prognostics that can translate large volume and heterogeneous data into reliable predictions to support optimal operations and maintenance decisions. In this research, we developed novel diagnostics and prognostics models based on advanced deep learning architectures to achieve two major objectives: (1) identify both precedent and unprecedented failures through an end-to-end learning approach; (2) learn the transferable and interpretable health condition-features from multimodal sensor data to enable the accurate fault diagnosis and prognosis by considering the impact of the varying operating conditions. In the first study, we designed a Variational Autoencoder (VAE)-based framework for novelty detection and known faults classification. In the second research, we develop a Variational Domain-Adversarial Neural Network (VDANN) framework to achieve cross-domain diagnosis. In the third research, we developed two real-time remaining useful life (RUL) prediction frameworks considering the impact of the dynamic operating conditions. In the cases of discrete operating conditions, we show that a Nonlinear Autoregressive Neural Network (NARNET)-based degradation model can can achieve higher RUL prediction accuracy as compare to traditional stochastic models that do not consider the unit-specific operating condition patterns. In the cases where operating conditions are continuous, we designed a deep learning-based framework that entails a random forest-based classification and Long Short-Term Memory (LSTM) prediction, and demonstrated its superior accuracy in RUL prediction in a precision manufacturing process, i.e., ion etch milling process. Based on these results, we show that the deep learning-based approaches have superior performance of failure detection, classification, and prediction by handling large-scale multimodal sensor data in advanced industrial systems.--Author's abstract