Small-Sample Wear-Rate Prediction and a Conditional-Quantile-Informed Maintenance Decision Interface for High-Speed Train Brake Pads Using Sparse Life-Cycle Records

High-speed train brake pads are safety-critical consumable components whose wear condition affects braking safety, maintenance scheduling, and component life utilization. This study assesses the predictive potential and limitations of sparse life-cycle maintenance records for brake-pad wear-rate estimation when continuous degradation trajectories are unavailable. A dataset of 54 valid brake-pad life-cycle records was constructed, including replacement date, accumulated mileage, wear per 10,000 km, installation position, motor/trailer classification, and train identifier. Linear regression, ridge regression, random forest, Gaussian process regression, and quantile regression were evaluated under leave-one-out cross-validation. Random forest produced the numerically best aggregate LOOCV point-prediction metrics, with a mean absolute error of 0.0823 mm/10,000 km, a root mean square error of 0.1030 mm/10,000 km, and an R2 of 0.6008; relative to the global-mean baseline, MAE and RMSE were reduced by 39.0% and 36.8%, respectively. The P10–P90 quantile interval achieved an empirical coverage of 62.96% and an average width of 0.2165 mm/10,000 km, with 11 observations exceeding the P90 upper bound, indicating undercoverage in the upper tail. Feature-importance analysis showed pronounced temporal stratification. Removing replacement year/month reduced random forest LOOCV R2 from 0.6008 to 0.2440, whereas leave-one-train-ID-out grouped validation yielded R2 = 0.5687; expanding-window chronological validation was materially weaker (R2 = −0.0590 for the full retrospective feature set). These results indicate that the temporal descriptors capture dataset-specific temporal structure and may partly act as statistical proxies for unrecorded time-varying conditions; they should not be interpreted as physical determinants, and prospective temporal transferability remains unverified. As a secondary application of the prediction outputs, the point prediction and upper conditional quantile were linked with current pad thickness, the allowable thickness limit, and planned operating mileage, to illustrate a preliminary maintenance-screening interface. This interface is intended as a preliminary maintenance-screening procedure rather than a validated replacement policy. The results show that sparse maintenance records contain record-level predictive information for brake-pad wear-rate estimation, while temporal dependence, upper-tail undercoverage, and the absence of field decision-outcome validation currently limit direct operational use.