A Unified Uncertainty-Aware Multi-Task Framework for Robust Remaining Useful Life Prediction Under Distribution Shift

Reliable prognostics in predictive maintenance requires not only accurate remaining useful life (RUL) estimation but also robustness under distribution shift and support for decision-making under uncertainty. However, most existing approaches focus on in-distribution evaluation and treat prediction, uncertainty estimation, and decision-making as separate tasks. This paper proposes a unified uncertainty-aware predictive maintenance framework that integrates sequence modeling, multitask learning, calibrated uncertainty estimation, and cross-dataset robustness evaluation within a single pipeline. The framework combines a sequence-aware multitask architecture with conformal prediction to jointly estimate RUL, failure proximity, and prediction intervals. This approach was evaluated on multiple N-CMAPSS datasets under both in-distribution (ID) and out-of-distribution (OOD) conditions. Although several models show similar performance under ID settings, significant differences appear under the distribution shift. The proposed MT-Conf model achieved the lowest mean OOD RMSE (11.10) and demonstrated stable generalization across eight OOD datasets. Last-cycle analysis showed that sequence-based models significantly reduced the error near failure (for example, MAE = 6.52 GRU vs. 14.30 XGBoost). The framework also provides well-calibrated uncertainty estimates (empirical coverage = 0.973), enabling reliable decisions. Under OOD conditions, uncertainty-aware policies issue only 2.7% of immediate replacement decisions, while maintaining perfect precision. Overall, the results highlight the importance of jointly addressing accuracy, robustness, uncertainty, and decision impact in predictive maintenance.