Leakage-Resistant and Uncertainty-Aware Remaining Useful Life Prediction: An Engine-Disjoint Benchmark and Deployable Application on NASA C-MAPSS

Abstract

Remaining useful life (RUL) models are frequently ranked by point error while two deployment-critical questions remain unresolved: whether evaluation isolates complete assets and whether the reported prediction is accompanied by calibrated uncertainty. We present an engine-disjoint and leakage-resistant benchmark on all four NASA C-MAPSS subsets, together with a bilingual research application. Causal features use only the current and preceding cycles. Complete engine identities are separated during grouped validation and conformal calibration. Ridge regression, histogram gradient boosting (HGB), random forests, and extremely randomized trees are evaluated at the official test endpoint. The best RMSE values are 19.112, 27.895, 17.989, and 28.298 cycles on FD001–FD004, respectively. No single feature/model combination dominates all subsets. At nominal 90% coverage, split-conformal intervals attain 0.990, 0.892, 1.000, and 0.948 coverage, whereas conformalized quantile regression attains 0.960, 0.741, 1.000, and 0.915. Conditional analysis exposes severe CQR undercoverage for FD002 engines whose true RUL exceeds 90 cycles. Sensor corruption further reveals that a fixed calibration layer is not shift-proof: 5% standardized noise raises RMSE from 27.80 to 39.33 on FD002 and from 29.16 to 45.61 on FD004, while split-conformal coverage falls to 0.757 and 0.819. We therefore recommend reporting engine-level separation, marginal and conditional coverage, and corruption stress tests as a minimum reliability protocol. The supplied application returns a point estimate, two 90% intervals, and an explicit research-use warning. All data, code, trained artifacts, figures, and numerical tables are provided for reproduction.