MCFS: Multi-Channel Causal Feature Selection for Subway Bogie System Fault Diagnosis

In this study, we propose an intelligent diagnosis method based on Multi-Channel Causal Feature Selection (MCFS) to address the problems of high feature complexity and fuzzy component association relationships that commonly exist in multi-channel high-dimensional data in the fault diagnosis of rail transit bogie drive systems. The method aims to provide physically interpretable and optimized feature sets for diagnosis using traditional machine learning models. MCFS draws on the theory of synergistic-unique-redundant decomposition (SURD) of causal relationships and innovatively defines channel-level causality quantitative indexes: redundancy, uniqueness, and synergistic causality. Based on the above indicators, a composite scoring function is constructed to realize the adaptive screening of the channel; and then the key features of each channel are screened by mutual information theory to construct the final optimized feature set. The experimental results on the BJTU-RAO bogie dataset, which contains 51 types of faults in 9 operating conditions, demonstrate that the proposed MCFS method, achieves a diagnostic accuracy of$94.55 \% \pm 1.59 \%$. This represents a significant improvement, and importantly, it reduces the feature dimension from 504 to 120 while significantly decreasing the training time from 7599.16s to 248.76s, thus effectively improving the accuracy, efficiency, and interpretability of fault diagnosis for complex systems.