WavFormer: An Interpretable Wavelet-Constrained Transformer for Industrial Acoustics Diagnosis

With the rapid advancement of sensing and computing technology, diagnosing faults in rotary machines has shifted from traditional signal processing-based methods to intelligent deep learning methods. Despite the emergence of backbone models like convolutional neural network, recurrent neural network, graph neural network, and transformer, the limited interpretability of deep learning methods hinders its acceptance and adoption by industrial users. In this study, we present an interpretable wavelet-constrained transformer (WavFormer) for diagnostic task to extract the local features and calculate the global information. We apply dual tree complex wavelet constraint that conforms to approximate shift invariance to the transformer network, which improves model performance while reduces the number of parameters. Furthermore, we explore the Einstein summation for matrix multiplication in frequency band blending after wavelet transform to reduce computational complexity and accelerate convergence speed. Considering the necessity of non-contact measurement in certain scenarios, we utilize acoustics signals to verify the effectiveness of our method. Experiments results show a significant improvement compared to others. Besides, it is found that the WavFormer is interpretable through class activation mapping.