Industrial Fault Diagnosis Model with Covariance Information

Abstract

Fault diagnosis of chemical process is a very important link in the industrial manufacturing process. Timely and accurate detection of faults or potential safety hazards in production process can greatly reduce production accidents. This paper presents a chemical process fault diagnosis system based on RIC(Resnet-Isqrt-Cov) network model. The module for extracting second-order information is embedded in ResNet network to calculate the characteristic covariance matrix after downsampling to extract the correlation between dimensions, which solves the problem that the traditional maximum pool layer or average pool layer is too simple to extract first-order information, resulting in a large loss of effective information of the model. The performance of this model is verified on the standard Tennessee-Hysmans (TE) process. The experimental results show that RIC network with covariance matrix information has superior performance in fault diagnosis (classification) of industrial data.