A fault diagnosis method for critical rotating components in trains based on multi-pooling attention convolution and an improved vision transformer

To address the susceptibility of sensor signals to strong noise interference, this study proposes a high-precision fault diagnosis method based on multi-sensor information fusion. The proposed approach integrates depthwise separable convolution (DSConv) with a multi-pooling attention (MPA) mechanism and an improved Vision Transformer (IVIT). First, continuous wavelet transform is employed to convert raw time-series signals collected from multiple sensors into two-dimensional feature representations, which are then fed into the model using a multi-sensor data-layer feature fusion strategy. Subsequently, channel shuffle and DSConv are utilized to effectively extract local features, while the MPA mechanism enhances the model’s capability to perceive critical features in noisy environments. Finally, the IVIT further strengthens feature extraction and representation capabilities, producing fault diagnosis results. Experimental evaluations on three datasets demonstrate that even under a signal-to-noise ratio of −2, MPAIT-Net outperforms other models, achieving average accuracies of 96.83, 95.00, and 98.92%, respectively. Moreover, the performance of the model on datasets with complex faults and faults of different degrees is studied. The results show that the model exhibits stable performance under different noise levels, showing excellent robustness and generalization ability. The corresponding source code is publicly available at https://github.com/xieph001/MPAIT-Net .