Comparative Machine Learning for Operating-Mode Classification of Pantograph-Arc Events in Railway Condition Monitoring

Pantograph arcing in electrified railway systems is influenced by vehicle operating state because traction, coasting, regenerative braking, and rheostatic braking produce different electrical interactions among the overhead supply, pantograph, onboard filter, traction converter, and braking circuits. Distinguishing these conditions is important for condition-monitoring systems that must separate operating-state changes from arc-related disturbances. This study develops and evaluates a leakage-safe framework for classifying pantograph-arc operating mode as braking or traction/coasting using 13 independent 3 kV DC recordings acquired from a Trenitalia E464 locomotive. Seven recordings represented braking and six represented the source dataset’s composite traction/coasting category. Each recording contributed seven equally weighted 200 ms arc-centered windows, producing 91 analysis windows while retaining the complete recording as the independent unit. The primary analysis used pantograph voltage, pantograph current, and filter voltage; braking-rheostat current was excluded to prevent a direct operating-mode shortcut. Four engineered-feature classifiers and two temporal networks were compared under nested leave-one-recording-out validation. The selected RBF-SVM achieved event-level balanced accuracy of 0.9167, a Macro-F1 of 0.9212, a Matthews correlation coefficient of 0.8539, an ROC-AUC of 1.0000, and a Brier score of 0.0267. A held-out-recording perturbation analysis identified voltage-skewness, current-distribution, and signed-energy descriptors as quantitative contributors, supporting a distributed multivariate interpretation. In matched sensitivity runs, removing filter voltage reduced Brier loss by 0.0059, but the 95% bootstrap interval crossed zero (−0.0004 to 0.0169; exact paired p = 0.3796), so the apparent perfect classification does not establish channel dispensability. These results support operating-mode-aware condition monitoring while emphasizing the small sample, absence of synchronized force or camera validation, and need for external validation before operational use.