Noise-Resilient Bearing Fault Diagnosis via Hybrid Attention-Based Multiscale ScConv and Quaternion Transformer

Bearings serve as vital elements within mechanical systems, and their condition significantly influences the safety and operational efficiency of the entire setup. However, existing bearing fault diagnosis methods often exhibit poor performance under high-noise conditions, struggling to effectively extract discriminative fault features. To mitigate this challenge, this paper proposes a noise-robust bearing fault diagnosis approach that integrates multi-scale spatial-channel reconstructive convolution, a hybrid attention mechanism, and a quaternion Transformer model. Specifically, multi-scale convolution is employed to capture features at various scales, and a weighted fusion strategy is utilized to strengthen the model’s adaptability to challenging noisy conditions. In addition, the hybrid attention mechanism combines spatial, channel, and global attention modules, significantly improving the expressiveness of the learned features. The quaternion Transformer module incorporates quaternion-based operations, replacing the conventional multi-head self-attention and feed-forward network, thereby further optimizing both global and local feature extraction capabilities. Experimental validation on bearing fault datasets from both CaseWestern Reserve University and Nanjing Tech University confirms the proposed method’s significant performance improvement, demonstrating enhanced diagnostic accuracy and noise immunity compared to conventional approaches under high-noise scenarios.