Influence of operational parameters on wheel-flat identification using axle-box acceleration: A full-scale experimental study

Early detection of wheel-flat is critical for railway safety and cost-effective maintenance, as undetected defects accelerate the degradation of both rolling stock and infrastructure. This study investigates wheel-flat identification using three time-domain features—Root Mean Square (RMS), Crest Factor (CF), and Wheel Flat Index (WFI)—computed from field axle-box acceleration signals. A full-scale experimental campaign was conducted on a dedicated closed railway circuit, under controlled and repeatable conditions, systematically varying defect size (10–45mm), measurement axis (longitudinal, lateral, and vertical), number of defects per wheel (up to four), vehicle speed (up to 60km/h), axle load, and sensor location relative to the defective wheel. Sensitivity analysis shows that CF and WFI are highly responsive to defect severity, while RMS better captures the cumulative effect of multiple flats. Measurement-axis analysis indicates that the lateral component, in addition to vertical, provides informative signatures for defect identification, whereas axle load variations have limited influence. Sensor-location analysis further reveals that defect-induced excitations propagate through the axle and bogie structure. A K-Nearest Neighbour (KNN) classifier validates the diagnostic capability of the proposed features through repeated stratified hold-out validation and cross-validation-based hyperparameter selection. Statistical analysis and threshold-based classification evaluation identify 20mm as the minimum defect size ensuring consistent detection across the investigated conditions. Binary classification accuracies above 90% are achieved in all scenarios except measurements acquired on the adjacent wheelset. These results demonstrate the potential of lightweight feature-based approaches for scalable on-board monitoring and predictive maintenance.