Bayesian Deep Ensembles for Reliable Remaining Useful Life Estimation Under Domain Shift

Abstract

Predictive Maintenance (PdM) has emerged as a cornerstone of Industry 4.0, leveraging sensor data to forecast the Remaining Useful Life (RUL) of critical machinery. While deep learning models have achieved state-of-the-art accuracy in RUL estimation, they predominantly function as deterministic point estimators. These models often fail to generalize effectively when subjected to domain shift—changes in operating conditions or fault modes that deviate from the training distribution. Furthermore, standard deep learning approaches typically exhibit overconfidence in their predictions on out-of-distribution data, posing significant safety risks in high-stakes industrial environments. This paper proposes a Bayesian Deep Ensemble (BDE) framework designed to enhance reliability and quantify uncertainty in RUL estimation under domain shift. By aggregating predictions from multiple probabilistic neural networks, the proposed method captures both aleatoric uncertainty (inherent data noise) and epistemic uncertainty (model ignorance). We demonstrate that BDEs not only improve predictive accuracy on the NASA C-MAPSS dataset under transfer learning scenarios but also provide calibrated uncertainty estimates that can serve as reliable indicators for manual intervention. The results suggest that ensemble-based Bayesian methods offer a robust alternative to single-model architectures for safety-critical PdM applications.