MFAResNet: multi-sensor fusion attention residual network for interpretable rolling bearing fault diagnosis

Abstract To address the challenges of low diagnostic accuracy and poor interpretability in fault diagnosis under variable conditions with small samples from multiple sensors, this paper proposes a novel multi-sensor fusion attention residual network. First, multi-directional sensor signals are fused using principal component analysis, and 28 time-frequency modal features are innovatively extracted to capture subtle fault signatures across diverse operating conditions. The 1D convolutional neural network is enhanced with residual blocks and a multi-channel attention mechanism, effectively mitigating gradient vanishing and overfitting while emphasizing discriminative features. To overcome the ‘black-box’ nature of deep learning models, the SHapley Additive exPlanations method is incorporated to quantify the contribution of each feature, enabling visualization of the model’s decision logic and identification of critical fault indicators. Experimental results demonstrate that the proposed method achieves superior diagnostic accuracy and generalization performance on small-sample, variable-condition datasets, while offering improved interpretability. These advantages make the approach well-suited for practical bearing condition monitoring and maintenance decision-making.