Railway equipment maintenance faces challenges from large-scale, heterogeneous data and complex faults. This paper proposes a train profiling method using data mining and a tagging system. Multi-source distributed heterogeneous data from PHM and MRO systems are integrated into a six-dimensional hierarchical tagging scheme for structured, semantic representation. An end-to-end framework performs preprocessing, tag generation, and balanced association rule mining using a multi-minimum-support MsEclat algorithm. The mined patterns train a BP neural network optimized by PSO and DE for fault prediction, improving accuracy and convergence. Experiments show high-quality rule mining, with training and validation MSE reduced by 25.9% and 38.0%, respectively, and convergence speed increased by 83.3%. The method provides an intelligent, end-to-end solution for rail transit maintenance using heterogeneous data.