Interpretable convolutional neural network with multilayer wavelet for Noise-Robust Machinery fault diagnosis

Convolutional neural networks (CNNs) are being utilized for mechanical fault diagnosis, due to its excellent automatic discriminative feature learning ability. However, the poor interpretability and noise robustness of CNNs have plagued both academia and industry. Since traditional signal analysis technology has a sound theoretical basis and physical meaning, it motivates us to use signal processing theory to improve the interpretability and performance of the CNN algorithm. To this end, this paper proposes a multilayer wavelet attention convolutional neural network (MWA-CNN) for noise-robust machinery fault diagnosis. This framework aims to learn discriminative fault features from the wavelet domain, which allows the model to obtain better interpretability and superior performance than conventional time-domain-based CNNs. The proposed Discrete Wavelet Attention Layer (DWA-Layer) is used to map time domain signals to wavelet space, and obtain valuable information through the learnable convolutional layer. By alternately using DWA-Layer and convolutional layer for signal decomposition and feature learning, the proposed framework actually embeds a similar multi-resolution analysis algorithm in CNN. This helps integrate physics-based knowledge into the CNN. Finally, the frequency attention mechanism is proposed to enhance the ability of MWA-CNN to obtain fault-related features from different frequency components. Experiments on high-speed aeronautical bearing and motor bearing datasets prove that the proposed method has excellent fault diagnosis ability and noise robustness. The visual analysis of the attention mechanism contributes to the interpretability of CNN in the field of fault diagnosis.