Abstract: Convolutional neural network(CNN) has been widely used in mechanical system fault diagnosis because of its powerful feature extraction and classification capabilities. However, CNN is a typical“black box model”, and the mechanism of CNN’s decision-making is not clear, which not only reduces the credibility of intelligent diagnosis results but also limits the application in fault diagnosis with high-reliability requirements. Facing this problem, the physically meaningful chirplet transform (CT) is introduced into the traditional convolutional layer to formulate the chirplet convolutional layer and Chirplet-CNN with the complete process of using Chirplet-CNN for fault diagnosis. A series of experiments show that Chirplet-CNN has excellent fault diagnosis ability similar to the current state-of-the-art methods, and has outstanding performance in interpretability. It can interpret the frequency band basis for CNN to extract category features and make judgments through spectrum analysis. In addition, the proposed chirplet convolutional layer has good generality and when combined with CNN models of different depths, it can effectively improve the diagnostic accuracy and obtain good interpretation results.

Key words: CNN, interpretability, Chirplet transform, time-frequency transform, deep learning, fault diagnosis