Interpretable quadratic convolutional residual neural network for bearing fault diagnosis

Bearing fault diagnosis plays an important role in ensuring operational reliability of rotating machinery and maximizing economic efficiency. However, in complex and changing operating environments, rolling bearings are susceptible to ambient noise, which weakens the effectiveness of fault diagnosis. Existing fault diagnosis models are lacking in interpretability and transparency of the diagnostic process. In response to this challenge, first, we developed a quadratic convolutional neural network (QCNN) utilizing the recently introduced quadratic neurons. Due to their superior feature extraction capabilities, these neurons effectively capture bearing feature signals. Second, we innovatively integrated residual neural networks (ResNet) with QCNN. Leveraging the strengths of ResNet in deep feature propagation and learning, this integration markedly enhances the model's accuracy in diagnosing bearing faults. Additionally, this study introduces the application of the layer-wise relevance propagation algorithm to bearing fault diagnosis, thereby improving the model's interpretability. We enhance the transparency and interpretability of the model's decision-making process by deeply analyzing input signal contributions. Additionally, we visualize the attention maps generated by the QCNN. This helps users better understand how the model identifies bearing fault features and the reasons behind its decisions. Our model combines high diagnostic accuracy with algorithmic interpretability, offering robust and transparent technical support for practical applications in the industrial diagnostics domain. Our code to get the link is https://github.com/wangyuecqupt/Bearing-Diagnostic-Model.