Objective Regarding the key components in subway train transmission system—traction motors, gearboxes and axle boxes, and the issues of safety degradation, noise and oil leakage caused by their failures, it is necessary to carry out a research on corresponding fault diagnosis methods.

Method  First, the acceleration signals of various components in traction system are finely processed using RFFT (real fast Fourier transform). By analyzing the unique performance characteristics of abnormal signals in the time-frequency domain, precise differentiation and identification of various types of abnormal signals are achieved. Second, in order to deal with problems of long-range dependency mining and cross-channel correlation analysis in multivariate time series data, a feature extraction backbone network based on the HDMixer deep learning architecture is designed to efficiently capture and accurately analyze the key information in the data. Given the complex interactions among traction motor, gearbox and axlebox data features, an integrated learning model based on deep learning is proposed, which constructs sub-models tailored to different fault types. The outputs of these sub-models are then subjected to ensemble processing, and the best-performing model results of fault category prediction are selected as the final fault judgement basis.

Result & Conclusion The integrated learning model achieves an accuracy of 84.82% in fault diagnosis task of the subway train transmission system, demonstrating high accuracy and stability under various operating conditions.