Empirically Calibrated Multi-Fidelity Fusion with Conformal Prediction Intervals for Reliability Assessment of Aerospace Dormant Components

Reliability prediction of aerospace dormant components requires fusing natural-storage observations at the operating temperature with accelerated-storage testing data at elevated temperatures. Existing scalar-weight fusion methods apply a global weight that cannot reflect the time-varying trustworthiness of the accelerated branch as Arrhenius extrapolation distance grows. Physics-based fusion propagates accelerated-test scatter through least squares but leaves the dominant error source—the degradation-model form itself—unaccounted for, and no method in either class verifies the coverage of its intervals. This paper proposes an empirically calibrated multi-fidelity fusion that selects a mechanism-specific natural-branch degradation model by the corrected Akaike information criterion and augments the accelerated-branch variance with an additive model-form term fitted from natural-storage residuals. This term turns the fusion weight into a continuous, time-varying diagnostic that detects Arrhenius misspecification from training data alone and falls back safely to the natural-only estimate. Prediction intervals are calibrated by split-conformal prediction on a disjoint simulated population, giving finite-sample, distribution-free coverage, and the remaining-storage-life interval follows from the band’s first-passage time. On a 1000-run varying-truth simulation, the calibrated band attains 95.5% trajectory coverage at the narrowest band width among six methods; on the torsion-bar case, the fusion reaches a held-out RMSE of 0.045 N·m and a remaining-life interval of 10.4–12.6 years. The model-form variance ratio provides a single-number regime diagnostic across all cases.