Remaining Useful Life Prediction Accounting for Epistemic and Aleatoric Uncertainties

Deep learning (DL) has proven highly effective for predicting the remaining useful life (RUL) of systems. However, many DL-based prognostic models typically produce deterministic RUL predictions, neglecting the associated epistemic and aleatoric uncertainties. In real-world applications, evaluating the confidence level of model predictions is essential for informed decision-making. This study presents a novel Bayesian deep learning (BDL) framework aimed at quantifying the uncertainty in RUL predictions. A major strength of the proposed framework lies in its ability to simultaneously capture two types of uncertainties: 1) epistemic uncertainty, which is related to the uncertainty within the predictioni model itself, and 2) aleatoric uncertainty, which pertains to random variations, such as those from measurement errors and degradation process behavior. To integrate both types of uncertainties, we utilize the Bayes by Backprop technique combined with an ensemble method within the BDL framework. The framework’s effectiveness is validated using literature datasets from the Commercial Modular Aero-Propulsion System Simulation (CMAPSS). The results show that our framework not only delivers reliable prediction intervals but also outperforms traditional models in both prediction accuracy and uncertainty quantification.