An Interpretable CNN With Wavelet Group Policy Embedded for Intelligent Fault Diagnosis
Guangjie Han, Jianhang Chen, Li Liu, Zhen Wang, Fan Zhang, Yilixiati Abudurexiti
Abstract
Interpretable convolutional neural networks (CNNs) are key to reliable industrial fault diagnosis by elucidating model decision-making processes and extracting high-dimensional data features. Presently, interpretable CNNs include time-frequency transformation methods in convolutional layers, but their hyper-parameter setting (e.g. window size, overlap, type of wavelet) depends largely on expert knowledge. This paper focuses on the problem of wavelet type selection, we propose a Wavelet Shrinkage Convolutional Network (GP-WSCN) based on a group policy to solve this problem. Initially, GP-WSCN creates a pre-trained network with wavelet convolution layers and soft-threshold learning, quickly providing a basic prior for diverse wavelet waveforms. This prior knowledge is combined with reinforcement learning, allowing GP-WSCN to independently select suitable wavelet convolution kernels, reducing expert dependency. The pre-trained network is then repurposed and fine-tuned to selected kernels for swift GP-WSCN deployment in fault diagnosis tasks. Experimental results confirm the diagnostic precision and interpretability of GP-WSCN, as proven through multiple trials.