Early Fault Detection in Safety Critical Systems Using Complex Morlet Wavelet and Deep Learning

Abstract

Automated Fault Detection and Diagnosis (FDD) plays an important role in health monitoring of safety critical systems. Typically, critical industrial processes involve voluminous number of sensors that are capable of assessing the system’s working condition and health. Time series analysis of sensor measurements can be used to predict potential system errors before the damage is irreparable. Predictive analysis is quintessential to reduce the system downtime and the costs associated. A major challenge in FDD is to detect faults much before the full-length time series is available, such that reliable predictions are achieved early in time. Thus, Early Classification of Time Series (ECTS) has to deal with the trade-off between accuracy and earliness, unlike conventional approaches that handle only accuracy. This paper proposes Complex Morlet Wavelet (CMW)-based time–frequency analysis for ECTS in a Deep Learning (DL) framework that combines Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) Network. The proposed approach is validated using the publicly available Tennessee Eastman Process (TEP) dataset. Results demonstrate that CMW in combination with hybrid CNN-LSTM architecture outperforms the state-of-the-art approaches for ECTS. The scheme is benefited by the DL architecture that combines CNN and LSTM, rather than these architectures considered individually. The proposed approach is able to achieve superior joint accuracy-earliness optimization when compared to time domain and frequency domain analyses considered separately, conventional Short Time Fourier Transform (STFT)-based time–frequency analysis, and other state-of-the-art time–frequency approaches.