Simulation-driven fault diagnosis for track circuits using multi-scale convolution and transformers under imbalanced data conditions

With the increasing complexity of railway systems, diagnosing faults in ZPW-2000A track circuits poses significant challenges, especially under imbalanced data conditions. This paper introduces a novel simulation-driven fault diagnosis framework combining multi-scale convolution, transformer encoders, and LDAM loss to enhance classification accuracy. The method extracts short- and long-term features from time-series data, effectively addressing data imbalance through gated convolution and optimised classification boundaries. Simulation experiments on 1,600 samples demonstrated a 98.75% accuracy, significantly outperforming existing approaches. Ablation studies validated the contribution of each module. The findings show potential for real-time fault detection and improved railway safety. Future work includes validating the model on real-world data and optimising its computational efficiency. This research contributes to simulation and process modelling by offering a robust, scalable solution for fault diagnosis, aligning with industrial needs and advancing the state-of-the-art in imbalanced data analysis.