A visual interpretation-based intelligent fault diagnosis framework for switch machines based on sound signal
Switch machines are critical components in rail transit signal systems, ensuring the safe, efficient, and smooth operation of trains. However, due to the harsh working conditions, they are the most failure-prone among all ground signal devices. Existing studies mainly pursue higher diagnostic accuracy via specialized models, often at the expense of interpretability, diagnostic transparency, and adaptability to different signal representations. In this article, a visual interpretation-based intelligent fault diagnosis framework is developed for switch-machine fault diagnosis using acoustic signals. A flexible time-frequency transformation module is used to convert signals into time-frequency representations, which are then input into a convolutional neural network for fault classification. Heatmaps obtained by the gradient-weighted class activation mapping (Grad-CAM) method are employed to visually explain the fault diagnosis results. To enhance both accuracy and interpretability, an entropy-regularized composite loss function is introduced by combining cross-entropy loss with a Grad-CAM-derived two-dimensional local information entropy term. This design encourages the model to form more compact fault-related saliency distributions while maintaining classification accuracy. Experimental results on the constructed laboratory dataset show that the framework achieves high diagnostic accuracy under different time–frequency representations, with several configurations reaching 100% accuracy under the current evaluation protocol. These results demonstrate the feasibility of visual interpretation for acoustic-based switch-machine fault diagnosis.