A CNN-based Explainable Fault Diagnosis Model for Gearboxes in Rotating Machinery

Convolutional neural networks(CNN) as a class of deep neural networks are attracting remarkable attention due to their powerful feature extraction capability in various areas such as gearbox fault diagnosis in rotary machinery. Although the identification performance of the CNN has demonstrated a superiority over traditional approaches, it is difficult to explain which parts of the inputs to the CNN are learned by this black box model. Hence, understanding the relationship between the inputs and the deep learning models will help to establish connection to the physical meaning of the fault diagnosis, contributing to the broad acceptance of deep learning models as a trustworthy complement to physical-based reasoning by human experts. In this paper, using Gradient-weighted Class Activation Mapping++ (Grad-CAM++) as the interpreter, the CNNs trained by two-dimensional time-frequency domain signal are interpreted by Grad-CAM++ to show the attention part in these signals by the CNN model.