Explainable Temporal Analytics for Railway Compressor Failure Prediction and Anomaly Detection Shubham Gupta

Abstract

Railway compressor systems are essential to the pneumatic functions that keep metro operations safe and uninterrupted, yet unexpected degradation of these systems remains a persistent source of service disruption and unplanned maintenance costs. This study proposes an Explainable Temporal Data Analytics (ETDA) framework for early failure prediction and anomaly detection in railway compressor systems, using the publicly available MetroPT-3 dataset [1,2], which contains 15,169,480 one-second observations from 15 analogue and digital sensors of a metro Air Production Unit collected between February and August 2020. Rather than treating sensor readings as independent samples, the proposed framework transforms instantaneous measurements into rolling temporal features (mean, standard deviation, rate of change, range, and cross-sensor correlation), derives a persistence-aware anomaly score, and estimates multi-horizon failure risk, coupled with Shapley-based feature attribution to explain which behavioural changes drive the estimated risk. Chronological validation is used throughout to avoid temporal leakage. The framework was evaluated against five conventional classifiers and detected all four documented air-leak failures, with a mean anomaly lead time of 6.8 hours. At the practically useful 6-hour warning horizon, the framework achieved a 91.4% F1-score and a PR-AUC of 0.903, exceeding the strongest conventional baseline (CatBoost, 88.3% F1-score) by 3.1 percentage points. Ablation analysis confirmed that temporal features, anomaly persistence, and the composite risk-integration layer each contribute statistically significant improvements (Wilcoxon signed-rank test, p < 0.05). Explainability analysis identified motor-current and pressure-related rolling features as the dominant drivers of predicted failure risk, results that are consistent with the physical behaviour of the compressor. The results presented here show that the integration of temporal analytics, persistent anomaly detection, and explainable early-warning estimation has the potential to furnish a more informative, actionable, and helpful foundation for railway maintenance decision-making than approaches based on instantaneous sensor measurements.