AHWCN: An interpretable attention-guided hierarchical wavelet convolutional network for rotating machinery intelligent fault diagnosis

Intelligent fault diagnosis is becoming increasingly important in modern industries, but the opacity of deep learning models limits their practical use. Current research primarily focuses on local interpretability, rule extraction, or model visualization. However, these approaches are insufficient for fully understanding and trusting diagnostic decisions. To address this problem, this paper proposes an Attention-guided Hierarchical Wavelet Convolutional Network (AHWCN) to enhance the interpretation of the model. The method decomposes signals using a hierarchical wavelet convolution (HWC) layer to reduce noise interference. It dynamically analyzes fault-relevant components using a time–frequency attention module (TFA). Finally, it combines TFA weights with gradients to enhance channel linkage learning. Experiments show that AHWCN achieves accuracies of 98.33% and 98.24% on Datasets A and B, respectively, significantly outperforming existing explainable models. More importantly, the model not only accurately captures fault pulse distributions in the time domain but also effectively localizes fault characteristic frequencies in the frequency domain.