Abstract:Deep learning-based rolling bearing fault diagnosis has been widely used in practical production. In this paper, an ensemble Swin-LE Transformer model with residuals is proposed to address the problems of background noise interference, difficulty in fault feature extraction and insufficient generalisation of the model. A local enhancement module is proposed and applied to the Swin Transformer, named Swin-LE Transformer, to enhance the model fault feature extraction and thus improve the diagnostic accuracy. An ensemble learning architecture with residuals is also proposed based on complete ensemble empirical modal decomposition with adaptive noise (CEEMDAN). The architecture structure uses the original signal as the residual structure for fusion voting to improve the overall generalisation capability of the model. The performance of the Swin-LE Transformer under this architecture is analysed against the original Swin Transformer and networks such as ViT and YOLO through experimental simulations on different bearing datasets, and the results show that the proposed ensemble Swin-LE Transformer achieves an accuracy of 98.62 , which is higher than the original network by 2.4 . Based on this, a multi-noise set was designed to verify the performance of the proposed architecture. The experimental results show that the proposed fusion architecture is still 2.94 higher at SNR-10 dB compared to the single-vote ensemble learning model, demonstrating its improved robustness.

An ensemble Swin-LE model with residuals for rolling bearing fault diagnosis