Integrating predictive uncertainty into maintenance decision-making: A risk-aware framework based on TPA-CV-UQ for optimal cost–availability trade-offs

Abstract

Predictive maintenance requires not only accurate degradation prediction but also maintenance decisions that explicitly account for predictive uncertainty. However, most existing approaches focus primarily on improving forecasting accuracy, while the uncertainty associated with future degradation is rarely incorporated into maintenance decision-making. Consequently, prediction quality does not necessarily translate into optimal maintenance performance. This study proposes a unified decision-centric predictive maintenance framework, termed TPA-CV-UQ, which integrates Temporal Prediction Analysis (TPA), uncertainty quantification (UQ), and cross-validated probability calibration (CV) into a coherent maintenance decision process. The proposed framework models equipment degradation through interpretable temporal dynamics and generates calibrated predictive probability distributions for estimating failure risk. Instead of relying solely on deterministic point forecasts, maintenance actions are adaptively triggered according to the probability of exceeding predefined failure thresholds, enabling risk-aware preventive maintenance and replacement decisions. Comprehensive experiments compare the proposed approach with Gradient Boosting, Naïve forecasting, and Exponential Smoothing using both prediction-oriented and decision-oriented evaluation metrics. Results demonstrate that TPA-CV-UQ achieves competitive prediction accuracy while providing better-calibrated probabilistic forecasts. More importantly, the proposed framework consistently attains the lowest maintenance cost together with the highest failure-aware availability, yielding the best overall cost–availability trade-off among the competing approaches. These findings demonstrate that maintenance effectiveness depends not only on prediction accuracy but also on how predictive uncertainty is incorporated into maintenance decisions. The proposed framework provides an interpretable, robust, and computationally efficient solution for predictive maintenance under uncertain operating conditions and offers a practical decision-support tool for industrial asset management.