Pantograph Arc Detection for Condition Monitoring of 3-kV DC Railway Infrastructure

Pantograph arcing is both a vehicle current-collection problem and a railway-infrastructure condition-monitoring problem because repeated loss of electrical contact can accelerate wear of the overhead contact wire and pantograph strip, degrade traction power quality, and increase maintenance demand. This study develops a leakage-safe diagnostic framework for 3-kV DC railway operation using 13 independent high-frequency recordings from the public Trenitalia E464 pantograph-arcing dataset. Because the repository does not provide synchronized optical/contact-force ground truth, the machine-learning target is consistently treated as a physics-guided candidate interval rather than an independently verified arc label. Pantograph voltage, pantograph current, filter voltage, and braking-rheostat current were sampled at 50 kSa/s and transformed into 207 event-preserving analysis windows. A total of 849 candidate features were extracted across time, frequency, time-frequency, nonlinear, and physics-informed electrical domains. The strongest leave-one-recording-out configuration was Extra Trees with frequency-domain features, with mean event-level accuracy of 0.9936, balanced accuracy of 0.9952, Macro-F1 of 0.9932, MCC of 0.9874, ROC-AUC of 0.9994, and PR-AUC of 0.9989. Ten-repeat grouped five-fold validation, with complete recordings retained as groups, produced a mean Macro-F1 of 0.9910 (SD 0.0193) across 50 grouped test folds. Five hundred recording-grouped bootstrap resamples yielded a Macro-F1 mean of 0.9893 with a 95% confidence interval of 0.9694–1.0000. A dedicated guard audit found zero candidate-interval overlap in all 138 retained normal 100 ms feature windows. Sensitivity analysis showed that 100 ms spectral features were materially more stable than 20 ms features, while 25% and 50% candidate-overlap thresholds produced nearly identical performance. The dataset does not contain long-duration independently verified arc-free operation, chainage/GPS catenary position, or synchronized contact-force measurements; consequently, the results are interpreted as proof-of-concept electrical screening of candidate current-collection disturbances rather than fleet-wide ground-truth arc detection.