Enhanced bearing fault diagnosis under strong noise: A deep residual network with combined attention mechanisms

In industrial manufacturing, bearing condition is critical for safety, but bearings often operate in harsh and complex environments. To improve bearing fault diagnosis accuracy under strong noise interference, this paper proposes an enhanced deep residual network integrating combined attention mechanisms (MSCKE-CFMSS-ResNet50). The model adopts adaptive pooling weights and multi-scale selectable convolution kernels in the channel dimension for effective information aggregation; in the spatial dimension, it groups feature maps, infers individual attention maps, and enhances information exchange via feature map shuffling. Embedding this attention structure into a residual network strengthens feature extraction and fault classification performance. Validated on a public rolling bearing fault dataset, the model outperforms existing attention mechanisms and conventional residual networks by 1.9%-3.82% in accuracy under -4dB to -10dB noise conditions. It also surpasses several state-of-the-art models, with ablation studies confirming the effectiveness of multi-scale convolution kernels and channel shuffling. This demonstrates superior diagnostic accuracy, robustness, and strong adaptability for bearing fault diagnosis in high-noise environments.