130 / 2025-05-20 15:28:02

WKAConvNet: An Explainable Wavelet Kolmogorov–Arnold Convolutional Network for Intelligent Fault Diagnosis

convolutional neural network,fault diagnosis,Kolmogorov-Arnold network

Special Sessions > SSL 01 Advanced signal processing enabled intelligent fault diagnosis and measurement

Convolutional neural networks (CNNs), with the powerful capability of data mining and feature adaptive learning, have been widely applied in mechanical fault diagnosis. However, many existing CNN-based models suffer from insufficient explainability in extracted features, making the results are not reliable. To address the challenge, a wavelet Kolmogorov–Arnold convolutional (WKAConv) layer is proposed by letting the learnable wavelet Kolmogorov–Arnold kernel as the convolution kernel. Based on this, a novel wavelet Kolmogorov–Arnold convolutional network (WKAConvNet) is proposed by replacing the first convolutional layer of the traditional CNNs with the WKAConv layer, thereby achieving the extraction of ante-hoc explainable features. Experiments are conducted on the planetary gearbox dataset to verify the effectiveness of the proposed method, and the results indicate that it achieves the best performance compared with the existing advanced CNNs. Besides, in the model explanation part, the ante-hoc explainability of WKAConvNet are also demonstrated through visualizing feature maps after the WKAConv layer.