Wheel-Flat Identification from Axle-Box Acceleration: Cross-Track Generalisation and In-Service Validation

The objective of this study is to assess cross-track generalisation and direct controlled-to-service transfer of wheel-flat classifiers based on three revolution-wise axle-box acceleration indicators: Root Mean Square (RMS), Crest Factor (CF), and Wheel-Flat Index (WFI). Full-scale controlled measurements were analysed across distinct track sections, and cross-track transferability was assessed through nested leave-one-track-out validation of speed-dependent k-nearest neighbour classifiers. Final all-track models were applied without retraining or field-specific adaptation to 403 acquisitions collected during commercial freight operation with a wireless axle-box sensor. The controlled measurements showed that the dynamic manifestation of a fixed wheel-flat is not spatially invariant, while the combined RMS–CF–WFI representation retained useful discrimination across track sections. Cross-track balanced accuracy ranged from 0.841 to 0.980; all-track models achieved mean balanced accuracies of 0.967, 0.937, and 0.931 at 30, 45, and 60 km/h. In service, balanced accuracy was 0.878, specificity 0.984, and binary-decision coverage 0.978 against a manually reviewed signal-based reference. Applicability analysis of the exploratory high-speed extension indicated that speed extrapolation does not necessarily imply feature-space extrapolation, although false-positive acquisitions showed substantial departures from the controlled support. These results support compact, transferable axle-box diagnostics for in-service wheel-flat monitoring.