Objective  As the complexity and operational scale of urban rail transit systems continue to increase, the signal system fault diagnosis faces problems such as heterogeneous protocols from multiple manufacturers, massive on-board log redundancy, low manual analysis efficiency, and insufficient localization accuracy, which seriously affect the operational safety and maintenance efficiency. To address these issues, an intelligent signal fault analysis and localization model based on the CNN (convolutional neural network)-Attention (attention mechanism)-LSTM (long short-term memory) hybrid neural network is proposed.

Method The proposed model performs standardized preprocessing on multi-source on-board log data, and completes structured feature extraction through protocol parsing, timestamp unification, and invalid data filtering. A 1D (one dimensional)-CNN module is constructed to capture local mutation features of signals, a multi-head attention mechanism is utilized to adaptively focus on key fault features and weaken the interference of irrelevant information, and an LSTM module is employed to model the temporal dependency of fault evolution, forming an end-to-end fault diagnosis architecture. The model training strategy is optimized, and the two-stage training and dynamic parameter tuning methods are adopted to enhance the generalization ability and interpretability of the proposed model.

Result & Conclusion  Verification results using measured data from multiple lines of Beijing Subway demonstrate that the proposed model can effectively identify typical signal faults such as MA (movement authority) retraction, position invalidity, and speed measurement invalidity. Therefore, the overall diagnostic accuracy is significantly improved, the false fault alarm rate is greatly reduced, and the distinguishability of key features is obviously enhanced. The model analysis results are completely consistent with manual analysis conclusions, and the fault localization accuracy meets the requirements of engineering applications. The model can effectively replace manual analysis work, significantly shortening fault handling cycle.