A deep learning model for process fault prognosis

Early fault detection and fault prognosis are crucial functions to ensure safe process operations. Fault prognosis can detect and isolate early developing faults as well as predict fault propagation. To promptly detect potential faults in process systems, it is important to examine the fault symptoms as early as possi-ble. In recent years, fault prognosis approaches have led to the remaining useful life prediction. Therefore, in a process system, advancing prognosis approaches will be beneficial for early fault detection in terms of process safety, and to predict the remaining useful life, targeting the system's reliability. In data-driven models, early fault detection is regarded as a time-dependent sequence learning problem; the future data sequence is predicted using the previous data pattern. Studying recent years' research shows that a recurrent neural network (RNN) can solve the sequence learning problem. This paper proposes an early potential fault detection approach by examining the fault symptoms in multivariate complex process systems. The proposed model has been developed using the Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) approach to forecast the system parameters for future sampling windows' recognition and an unsupervised One-class-SVM used for fault symptoms' detection using forecasted data window. The performance of the proposed method is assessed using Tennessee Eastman process time -series data. The results confirm that the proposed method effectively detects potential fault conditions in multivariate dynamic systems by detecting the fault symptoms early as possible. (c) 2021 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.