Fault Magnitude Prognosis in Chemical Process Based on Long Short-Term Memory Network

Abstract

This paper presents a long range process fault prognosis system using long short-term memory (LSTM) network. Data from historical process operation with faults present are used to train LSTM networks. During process monitoring, a principal component analysis (PCA) model developed from normal historical process operation data is used to detect the presence of a fault. Once a fault is detected, reconstruction based fault diagnosis is used to diagnosis the detected fault. Then the trained LSTM network corresponding the diagnosed fault is use to provide long range fault magnitude forecast. The proposed method is applied to a simulated continuous stirred tank reactor (CSTR) and is compared with fault prognosis using extreme learning machine (ELM). The results show that the proposed fault prognosis method based on LSTM network can achieve excellent long range prognosis performance.