Top

A Reliability Gated NeuralODE–Transformer Framework for Fault Precursor Mining in Railway Signalling Equipment

Authors

Christian Bertoni
Faculty of Information Technology, University of Padua, Padua, 35122, Italy
Filippo Caputo
Faculty of Information Technology, University of Padua, Padua, 35122, Italy

Files

Abstract

Early warning for railway signalling maintenance requires consistent interpretation of electrical measurements, equipment events, and environmental observations acquired at different times. This paper proposes the NeuralODE–Transformer framework, a multisource fusion architecture combining neural ordinary differential equations with causally masked attention. Source-specific continuous states represent evolution between observations, while reliability gates reduce the contribution of ageing or incomplete measurements. A precursor head converts fused representations into event risk scores and supports inspection of the evidence preceding an alarm. The evaluation examines score discrimination, event timing, uncertainty, and the engineering requirements of precursor assessment. The score-based evaluation compares six configurations on 1,200 labelled samples containing 180 positive cases. At a fixed threshold of 0.50, the proposed configuration yields 162 true positives and 31 false positives, corresponding to precision, recall, and F1 scores of 83.94%, 90.00%, and 86.86%, respectively. The attention configuration yields an F1 score of 79.27% under the same class distribution. A separate event set has a median warning lead of 33 minutes and supports inspection of heterogeneous precursor trajectories. Missingness and component removal profiles define criteria for subsequent architectural assessment under incomplete or changing evidence. The study contributes a model specification, numerical score analysis, and a validation protocol; assessment of the complete trained architecture requires equipment-level investigation. Operational assessment requires training and testing on independently audited railway records.

Details