Enhanced CCA-based Fault Detection with Deep Neural Networks in High-Speed Train Traction Systems

This paper proposes an innovative fault detection method for nonlinear systems by combining deep neural networks (DNNs) with canonical correlation analysis (CCA). Unlike traditional linear CCA-based fault detection methods, this approach initially uses DNNs to extract advanced and complex feature extractions from raw data, followed by using these features as inputs for CCA. By computing the correlation coefficients and continually adjusting the parameters of the DNN, the maximum correlation in the data, resulting in the optimal residuals to enhance the accuracy and efficiency of fault detection. Addressing the issue that the assumption of independence among features after DNN feature extraction may no longer hold, a kernel density estimation method is utilized to determine the threshold. Finally, the proposed fault detection method is applied to the simulation of a high-speed train traction system, verifying the effectiveness of the method.