A Bayesian Data-Driven Framework for Aleatoric and Epistemic Uncertainty Quantification in Remaining Useful Life Predictions
Mudi Jiang, Tianyang Xing, Enrico Zio, Xiaoliang Zhu
Abstract
Remaining useful life (RUL) prediction is a key technology for device prognostic and health management. Due to deficiencies in data and models during the prediction process, the predicted RUL results exhibit various types of uncertainty. However, most RUL prediction models address point estimates or total uncertainty. To this end, this article proposes a Bayesian data-driven RUL framework with aleatoric uncertainty and epistemic uncertainty quantification. First, considering the impact of data inherent noise and model ignorance on prediction uncertainty separately, an algorithm for quantifying aleatoric and epistemic uncertainty of relevance vector machine (RVM) is proposed by Monte Carlo sampling. Then a Bayesian data-driven RUL predictive framework with uncertainty quantification is proposed. An adaptive training set based on the similarity method is adopted to extract units of a training set with features that are similar to the test unit. Finally, the application of the proposed framework is shown on a public turbofan engine dataset commercial modular aero-propulsion system simulation (C-MAPSS) and a case of the once-through steam generator (OTSG) of nuclear power plants (NPPs). The superior prediction performance of the proposed framework is illustrated by comparing it with other state-of-the-art methods.