Convolutional neural networks (CNNs) are widely used in intelligent fault diagnosis for their strong feature extraction and nonlinear representation abilities. However, their black-box nature makes decision-making processes and classification rationale unclear. To address this, a Chirplet transform-driven convolutional neural network, termed CTNet, is proposed in this paper. Specifically, the physically meaningful Chirplet transform is integrated into conventional convolutional layers to form an interpretable Chirplet convolution (CTConv) module. This module incorporates four learnable parameters, that is, bandwidth, time shift factor, frequency factor, and chirprate to significantly enhance the dynamic adaptability of the convolutional kernels. The effectiveness and superiority of the proposed CTNet are validated on three bearing datasets (public, laboratory, and engineering application) in terms of fault diagnosis accuracy, noise robustness, universality, and interpretability. Experimental results demonstrate that: (1) the CTNet achieves higher fault diagnosis accuracy than other state-of-the-art models; (2) when integrated with other classical network architectures like LeNet, AlexNet, and ResNet, the CTConv module maintains excellent generalization performance. Finally, cumulative frequency band analysis for CTNet decision tracing verifies that the model effectively focuses on frequency bands containing prominent fault characteristics within the original vibration signals.
Zhang et al. (Tue,) studied this question.