A lightweight multi-scale channel-attention residual network for rolling bearing fault diagnosis

Reliable health monitoring of rolling bearings is essential for ensuring the safe operation of rotating machinery. However, existing deep learning-based diagnostic methods still face several challenges under non-stationary operating conditions, including inadequate multi-scale feature extraction, limited sensitivity to weak fault signatures, performance degradation under cross-condition variations, and the difficulty of balancing diagnostic accuracy with computational efficiency. To address these issues, we propose a multi-scale channel-attention residual network, termed MSCA-ResNet, for rolling bearing fault diagnosis under variable operating conditions, aiming to improve diagnostic performance while maintaining a compact computational footprint. The proposed network first uses a parallel multi-scale convolutional stem to extract time-domain features with different receptive fields from one-dimensional preprocessed signal segments. A compact channel-attention mechanism is incorporated into the residual backbone to adaptively recalibrate channel responses and enhance discriminative feature representation. The training strategy integrates loss regularization and data augmentation techniques, whose contributions are further evaluated through ablation studies. Experiments on the Paderborn bearing dataset show that MSCA-ResNet achieves an in-distribution diagnostic accuracy of 95.98% and an inner-race fault recall of 91.09%, improving the latter by 28.92 percentage points compared with ResNet1D. Cross-condition experiments and supervised Paderborn-to-CWRU adaptation further evaluate its performance under operating-condition variations and supervised cross-dataset adaptation scenarios. Notably, MSCA-ResNet contains 240.9 K trainable parameters and has a model size of approximately 0.965 MB, demonstrating a compact computational footprint under the evaluated benchmark settings. These results indicate the effectiveness of MSCA-ResNet under the evaluated benchmark protocols.