X-DCFR: Explainable Deep Convolutional Neural Networks Fuzzy Rule Learning for Railway Track Circuit Fault Diagnosis

Deep neural network-based intelligent methods have brought new breakthroughs to fault diagnosis. However, their ‘black box’ nature hinders their implementation and widespread adoption in safety-critical equipment such as railway tracks. The Explainable Convolutional Neural Networks Fuzzy Inference System (X-DCFR) combines the advantages of high-dimensional feature extraction from CNNs and the interpretability of FIS rules, seeking a balance between the ability to express deep-level fault relationships and the interpretability of fault diagnosis mechanisms. A customized training dataset was constructed using a track circuit data simulation system, and the model was trained through a staged strategy consisting of ‘CNN pre-training → independent FIS layer training → joint fine-tuning of the entire network’. Simulation results show that X-DCFR outperforms other algorithms in identifying eight typical fault states in railway tracks, achieving an average accuracy of 96.1%. Furthermore, the model’s excellent generalization ability and robustness were verified on actual equipment, providing an engineering reference for the practical application of deep neural network-based fault diagnosis in railway tracks.