An Interpretable Wavelet-Based Spectrum Augmentation Network for Intelligent Fault Diagnosis

Deep learning (DL) is becoming a vital technology in the field of fault diagnosis. However, a critical challenge in DL is the lack of interpretability, which limits the wide applicability of DL-based methods in the manufacturing industry. To address this challenge, a wavelet-based spectrum augmentation network (WSANet) is proposed in this article, which is an interpretable fault diagnosis model built from the frequency perspective. A wavelet-based spectrum augmentation layer (WSALayer) is designed as the first layer of the network. In this layer, the convolutional theorem is first utilized to obtain frequency components through learnable wavelet filtering. Then, frequency components containing fault information are augmented via multiscale weight calculation. Finally, the augmented spectrum is used in the backbone network for fault diagnosis. The proposed method is evaluated on both simulated data and real bearing fault data. The results show that the proposed method exhibits higher diagnosis accuracy, better robustness to noise, and improved generalization ability. Furthermore, the feature visualization and theoretical analysis of the WSALayer illustrate its interpretability.