Fault Diagnosis Based on Modified BiLSTM Neural Network

Authors:

Abstract
Fault diagnosis of industrial processes has recently received widespread attention. Some studies applied neural networks to fault classification algorithms, which directly take the raw data as input of the neural network for training, and then use classifier to classify faults. However, the raw data of industrial usually has high dimension and the variables has linear correlations between each other, which may not only lead to higher computational complexity and cause the neural network difficult to converge. In this paper, a fault diagnosis algorithm based on a modified bidirectional long-short term memory network (BiLSTM) is investigated. The proposed method introduces the linear discriminant analysis (LDA) before the architecture of networks to reduce the dimension of input data. Then the LDA-BiLSTM is tested in Tennessee Eastman Process (TEP) platform. Experiment results show that the dimension reduction procedure can speed up the convergence of neural network, and may improve the classification accuracy. Extensive experimental results indicate that developed algorithm provides better diagnosis performance.

This publication has 13 references indexed in Scilit: