Interpretable bearing fault diagnosis based on ensemble learning with improved ResNet

The operational condition of the bearing is critical for ensuring the safe and reliable performance of mechanical equipment. A fault diagnosis model based on ensemble learning and improved ResNet, which is called BSpline-Attention-ResNet with ensemble learning (EBARN) is proposed to resolve the challenge of inadequate accuracy and limited interpretability inherent in a singular fault diagnosis model under varying operational conditions. First, the Convolutional Block Attention Module (CBAM) is strategically incorporated into the architecture prior to the execution of the residual connection within the final residual block. Then, the activation function following the residual connection is refined by replacing the conventional ReLU activation with a B-spline activation function. Finally, an effective ensemble strategy for the improved ResNets is proposed, optimizing the weight distribution among base learners using the geyser algorithm. This optimization ensures that the aggregated classification results from the ensemble yield superior overall performance. To validate the effectiveness and interpretability of EBARN, a series of experiments were conducted using the Case Western Reserve University (CWRU) and Rotor datasets. These included ablation studies, comparative analyses, and masking tests. The experimental results demonstrate that EBARN exhibits superior diagnostic performance and generalization capabilities for rolling bearing fault diagnosis relative to existing methods. Moreover, comparisons of channel outputs before and after masking reveal that EBARN effectively captures key fault features and demonstrates its interpretability.