Fault diagnosis for high-speed train braking system based on disentangled causal representation learning

Data-driven methods have shown a great potential in diagnosing ongoing faults in high-speed trains (HSTs). However, lacking enough interpretability, data-driven methods have not been widely considered in practical operation of HST. In recent years, the rapid development of the causal discovery technology provides an effective way to improve the model interpretability. In this work, based on disentangled causal representation learning (DCRL), an effective and interpretable fault diagnosis framework is proposed for HST braking system. Independent potential factors of the high-dimensional monitoring data are extracted by the DCRL based on factor analysis. A stable and clear causal network in the factor space is obtained based on causal discovery, and the information irrelevant to fault diagnosis can be eliminated by feature selection. With logistic regression as the fault diagnosis model, the risk importance ranking of the monitoring features can be obtained. Compared with most commonly used methods, the method proposed in this paper has high interpretability and application value, which is more conducive to the subsequent fault location and troubleshooting. Based on real monitoring data of a HST braking system, it is justified that the effectiveness of the fault diagnosis model can be significantly improved by DCRL. Moreover, the applicability of the proposed method is also discussed.