Gamma-Process-Informed XGBoost for Fleet-Scale Remaining Useful Life Estimation of Railway Wheels: An Application to 444 Vehicles

Predictive maintenance of railway wheelsets requires remaining useful life (RUL) estimates and treatment of model limitations. This study develops a Gamma-process-informed XGBoost surrogate for an ONCF fleet of 444 vehicles, 3552 candidate wheel positions, and 14 inspection campaigns. Flange width (Fw), flange height (Fh), and flange steepness (qR) were measured using a CALIPRI C42 gauge. After date normalisation and series-level IQR filtering, 3296 candidate series remained per indicator; Gamma eligibility retained 2366 Fw, 1628 Fh, and 2445 qR series. RUL quantiles were evaluated from the discrete Gamma first-passage distribution, with Monte Carlo used only for convergence checking. Vehicle-level median RUL was 13 months for Fw, 3 months for Fh, and 7 months for qR. RUL below six months occurred for 25/402, 283/344, and 100/404 vehicles, respectively. Under vehicle-grouped five-fold cross-validation, surrogate-label reconstruction yielded Pearson r values of 0.968, 0.759, and 0.964, with MAEs of 4.16, 1.67, and 2.82 months. Ablation confirmed strong target-feature coupling, while calibration and bootstrap diagnostics revealed substantial uncertainty. The framework supports model-conditional fleet screening, not validated real-world RUL forecasting or automatic maintenance decisions, until prospective temporal validation links scores to maintenance outcomes.