A rolling bearing fault diagnosis method based on feature fusion threshold attention residual network and enhanced transformer under small samples and strong noise

Abstract As an important component of rotating machinery, rolling bearing often operates under strong noise environments, which may cause the system to fail to work normally once a fault occurs; in addition, there is the problem of limited labeled samples of fault data during bearing operation. Therefore, to address the problem of poor fault diagnosis accuracy of rolling bearings under strong noise environments and small sample conditions, this paper proposes a multi-sensor feature fusion threshold attention residual network and convolutional enhancement transformer (MFFTARN-CET) method. First, a GS-SSWT method is proposed, which converts the acoustic and vibration signals into two-dimensional time-frequency maps to retain the time-frequency information. Then, a multi-channel feature fusion block is designed, which fully exploits the similarity relationship of multi-sensor data with different sizes of convolutional layers. Meanwhile, the representational capability of the network is improved by learning the correlation and importance between different channels through a squeeze-and-excitation network mechanism. Second, the fused features are input into MFFTARN-CET for training, and the outputs are fused based on feature weighting to ensure the full utilization of multi-sensor signals. Third, a hybrid adaptive loss is designed to allow the method to adaptively adjust the contribution of different loss components during the training process through a gradient magnitude dynamic weight adjustment strategy. Finally, the effectiveness and superiority of the MFFTARN-CET method are verified using two rolling bearing datasets.