An Interpretable Intelligent Bearing Fault Diagnosis Method Using Damage Dynamics Response

Convolutional neural networks (CNNs) are widely applied in fault diagnosis due to their excellent ability to extract local features and process complex data. However, the decision-making and classification mechanisms of CNNs remain poorly understood. Hence, this article proposes a novel approach integrating damage dynamics (DD) responses in a CNN to enhance model interpretability. A damage-dynamics-based convolutional kernel was designed and employed to construct a DD-CNN for fault diagnosis, identification, and classification. First, kinetic response signals obtained from the bearing DD model were fit using an oscillatory decay function. Second, this function was used to construct an interpretable damage-dynamics-based convolutional kernel for matching the fault signals in the input samples. Finally, a DD-CNN was constructed using the damage-dynamics-based convolutional kernel for fault identification and classification. The experimental results demonstrated that the proposed method revealed the underlying feature extraction logic in the convolutional layer based on the dynamic response to matching damage. Meanwhile, the constructed model achieved an average fault identification and classification accuracy of 99%.