Explainability-driven fault diagnosis for railway switch machines

Abstract To address the limitations of traditional deep learning in the railway industry caused by its ‘black-box’ nature, a Wavelet-Enhanced Dual-Branch Explainable Network (WDEN) is proposed for railway switch machine fault diagnosis. First, a dual-branch parallel convolutional structure is constructed, enabling the extraction of sound signal features of railway switch machines at different scales through parallel convolution kernels, thereby enhancing the network’s feature representation capability. Subsequently, a Wavelet Weight Initialization (WWI) method is introduced, in which wavelet basis functions are embedded into the convolution kernel weights. This endows the convolution kernels with explicit time–frequency physical meaning, thus achieving efficient and transparent multi-scale feature extraction. Furthermore, the wavelet basis functions in WWI are optimized by eliminating translation parameters to reduce redundant computation and suppress waveform distortion, while adjusting the range of scale parameters to emphasize key time–frequency features. In addition, a minimum-scale threshold constraint strategy is employed to mitigate the vanishing gradient problem. Finally, a Gradient-Based Attention Map (GAM) technique is adopted to project the model’s attention regions onto the raw time-domain waveform of the sound signal, thereby visually revealing the critical evidence underpinning diagnostic decisions and further enhancing model interpretability. Experimental results demonstrate that the proposed method achieves 99.84% fault diagnosis accuracy on the ZDJ9 railway switch machine sound dataset. At the same time, WDEN exhibits a computational complexity of only 6.63M and an inference time of 0.023 s, fully meeting the real-time diagnostic requirements of railway field applications. This study provides both theoretical foundations and technical support for the development of handheld railway switch machine fault diagnosis devices in the railway industry, highlighting its strong potential for engineering applications.