Exploratory Cross-System Recognition of Candidate Pantograph Arc Intervals in DC Railways

Cross-system deployment of pantograph-arc classifiers is difficult because railway voltage, rolling stock, operating state, and measurement distributions differ. This exploratory study evaluates transfer from 13 Trenitalia 3 kV DC recordings to six Metro de Madrid 1.5 kV DC recordings, comprising 207 source and 60 target modelling windows. The labels are outputs of an automatic multichannel electrical screening procedure and have not been independently adjudicated; the study therefore measures recognition of detector-defined candidate intervals, not confirmed physical arcs. Seven strategies were evaluated under distinct target-information regimes with source-only hyperparameter selection and recording-disjoint target folds. The complete Madrid recording under testing contributed neither labelled nor unlabelled data to its own adaptation pipeline. The dimensionless physics-derived representation had the highest recording-balanced target macro-F1 (0.883; 95% recording-block bootstrap CI 0.683–1.000), balanced accuracy (0.896), and Matthews correlation coefficient (0.803), and the lowest Brier score (0.054). Its macro-F1 exceeded raw direct transfer by 0.633. The finding was stable for tested current bases: fold-specific 90th–99th percentile bases produced macro-F1 values of 0.883–0.900, and a source-only 95th-percentile base produced 0.937. When one target candidate event at a time was conservatively relabelled as normal, the selected representation remained first-ranked, but its macro-F1 ranged from 0.783 to 0.983 and its Brier score from 0.009 to 0.109, illustrating dependence on uncertain labels. CORAL reduced an unnormalised 32-dimensional PCA covariance norm from 19.611 to 0.096 but achieved a macro-F1 of 0.361. No pairwise Brier comparison remained significant after Holm correction. These six-recording results demonstrate a leakage-controlled exploratory comparison, not general cross-railway robustness; expert adjudication and independent multi-railway validation are required before operational use.