Chemical process fault diagnosis based on bi-level dynamic IndRNN

Abstract

Diagnosis of easily confused faults can help reduce losses in chemical process production. Traditional independently recurrent neural network (IndRNN) based fault diagnosis method has a classifier, and cannot extract dynamic features. Therefore, a bi-level dynamic IndRNN based fault diagnosis method is proposed for chemical processes. In the first level classifier, a dynamic IndRNN is proposed to simultaneously extract trend and dynamic features. In dynamic IndRNN, IndRNN is used to extract trend features, and attention fully convolutional network based on squeeze-and-excitation and global temporal attention is adopted to extract dynamic features. Confusion ratio is utilized to determine the hardly diagnosed faults. The second level classifier is designed to finely identify hardly diagnosed faults. To enhance fault distinguishability, principal component analysis statistics are introduced to augment the measured vector, and robust standardization is adopted to improve model generalization ability. Simulations on the Tennessee Eastman process demonstrate the effectiveness of the proposed method.