Bayesian neural networks for uncertainty quantification in remaining useful life prediction of systems with sensor monitoring

Many machine learning (ML) algorithms have been developed over the past two decades for prognostics and health management (PHM) of complex engineering systems. However, most of the existing algorithms tend to produce point estimates of a variable of interest, for example the equipment’s remaining useful life (RUL). The point estimation of the RUL often neglects the uncertainty inherent in model parameters and/or the uncertainty associated with data inputs. Bayesian Neural Networks (BNNs) have shown a lot of promise in obtaining credible intervals for model parameters, thus accounting for the uncertainties inherent in both the model and data. This paper proposes a deep BNN model with the Monte Carlo (MC) dropout method to predict the RUL of engineering systems equipped with sensors and monitoring instruments. The model is tested on NASA’s Turbofan Engine Degradation Simulation Dataset and the results are discussed and analyzed. It is revealed that the method can produce highly accurate predictions for RUL distribution parameters in safety critical components.