Bayesian Neural Network Based Method of Remaining Useful Life Prediction and Uncertainty Quantification for Aircraft Engine

Abstract

Remaining useful life (RUL) prediction is a key component of reliability evaluation and conditional-basedmaintenance (CBM). In the existing prediction methods, neural networks (NNs) are widely used because of the high accuracy. However, most of the traditional NNs prediction methods only focus on accuracy without the ability in handling the problem of uncertainty, where the generalization of the method is limited and their application to practical application are challenging. In this paper, an efficient prediction method based on the Bayesian Neural Network (BNN) is proposed. Network weights are assumed to follow the Gaussian distribution, based on which they can be updated by Bayes’ theorem and the confidence interval (CI) is consequently derived. The method is verified on the C-MAPSS data set published by NASA and the degradation starting point is determined via change point detection method. The experimental results demonstrate that the method performs well in prediction accuracy with the capability of the uncertainty quantification, which is critical for the condition monitoring of complex systems.