Abstract
Abstract
To overcome the engineering drawback that existing rolling bearing diagnosis models fail to simultaneously identify fault types and classify fine-grained damage severities, this paper proposes a multi-scale multi-task CNN embedded with bidirectional squeeze-and-excitation attention (BSE-MSTCNN). Built on a shared-branch framework, Its shared block adopts parallel 7 × 1, 5 × 1, 3 × 1 multi-scale convolutions for multi-granularity fault feature extraction. BSE modules embedded in shared and task branches bidirectionally suppress noise and strengthen fault impulses, while residual connections mitigate gradient vanishing in deep training. Two independent branches jointly train fault classification and damage severity evaluation. A dynamic adaptive loss algorithm balances dual-task losses to avoid single-task bias and realize synchronous identification of fault types and damage degrees. Validated on the CWRU and PU bearing datasets via multi-gradient SNR noise experiments, ablation studies, multi-model diagnostic accuracy comparisons, and effectiveness verification of the BSE attention mechanism, the model obtains fault-classification accuracies of 100.00% and 95.25%, as well as damage-severity grading accuracies of 98.94% and 94.62% under SNR = 0.Experimental results prove BSE widens feature gaps of subtle damages and optimizes fine-grained severity classification. This method realizes stable high-precision feature extraction under heavy noise and provides reliable technical support for multi-objective maintenance of industrial rotating bearings.