Fault Diagnosis Method of Train Transmission Systems Based on Multi-sensor Data Fusion and Multi-channel Convolutional Attention Network

To accurately diagnose the train transmission system faults, this paper proposes a fault diagnosis method based on multi-sensor data fusion and multi-channel convolutional attention network (MCCANet). Firstly, multi-source signals collected from key components of the transmission system are converted into multi-channel data utilizing the proposed data fusion method, effectively integrating the fault status information of all components. Next, MCCANet is proposed to process multi-channel sensor data, which combines one-dimensional convolutional kernels with attention mechanisms to extract features of data, using average pooling and global average pooling for dimensionality reduction to capture key fault features effectively. Finally, to address the issue of generalization under cross-working conditions, MCCANet serves as the base model, and probabilistic voting is applied to the predictions from several trained models. Probabilistic voting leverages the probability distributions predicted by these models, enhancing the accuracy and robustness of the results. Fault diagnosis experiments on the train transmission system dataset of 2024 PHM-Beijing data challenge demonstrate that the proposed method achieves high accuracy with small samples and across working conditions, proving its effectiveness.