A Multi-Scale Residual Attention Neural Network with Class-Conditional Cross-Load Alignment for Bearing Fault Diagnosis

Bearing fault diagnosis under varying load conditions is challenged by load-dependent changes in bearing dynamic responses and the resulting cross-load feature distribution shifts. This paper proposes a Class-Conditional Cross-Load Alignment Multi-Scale Residual Attention Neural Network (CCLA-MSRA). In this network, the multi-scale convolution module uses multi-scale convolution kernels to characterize the vibration features at different time scales, achieving the fine representation of fine-grained fault features; the residual attention module highlights the channel feature information related to the fault category, enhancing the discriminative ability of fine-grained fault features. The Class-Conditional Cross-Load Alignment mechanism is designed to align the feature distributions of similar samples, by constraining their means and intra-class variances, reducing the intra-class distribution deviation caused by load changes, and improving the cross-load feature consistency. On the 14-class CWRU task with four-fold cross-load validation, CCLA-MSRA achieved 96.29% Accuracy, 95.46% Macro-P, and 95.89% Macro-R, with Accuracy exceeding the CNN baseline by 3.40 percentage points. On the Paderborn University dataset, CCLA-MSRA achieved 98.93% Accuracy and 98.39% Kappa under the unseen operating condition. Visualization results show that the multi-scale convolution kernels have complementary frequency responses, different fault categories exhibit differentiated responses at different scales, and the deep features present good intra-class aggregation and inter-class separability.