A Novel Fault Diagnosis Approach Based on Correlation Analysis and Improved Two-Stream Neural Networks

Industrial process data are usually multivariate, nonlinear, and highly dimensional. Deep learning based on variable correlation are widely used in nonlinear dynamical systems monitoring of industrial process, and most of the existing studies have focused on extracting data with single features. In this paper, from the perspective of improving detection accuracy and reducing detection time, a novel fault diagnosis approach based on correlation analysis and improved two-stream neural networks is proposed to address the problems that traditional neural networks cannot extract the data temporal and spatial features simultaneously. The novel algorithm mainly adopts long short-term memory and convolutional neural networks to extract the data features respectively. And the canonical correlation analysis algorithm is used as the activation function to train the two-stream neural networks in depth. Relevant experiments are designed for comparative analysis, and the results show that the proposed algorithm fully improves the detection accuracy and data smoothing. The method was applied to the fault detection process of nonlinear dynamical traction systems and the fault detection rate was improved by 5%.