Remaining Useful Life Prediction Considering Data and Model Uncertainties

Remaining useful life (RUL) prediction plays a great part in prognostics health management (PHM). Recently, machine learning (ML) methods have been widely used in RUL prediction due to their scalability and generalization. However, most of the research focuses on RUL point estimation and neglects the uncertainties that are caused by data noise and modeling inaccuracy. To get more reliable and practical results, an RUL interval prediction method with data and model uncertainty quantification is proposed. Firstly, a series of neural network models with various parameters and structures are constructed and they are trained with bootstrapped data. After that, the outputs of the models are fused to form new time series labels which contain data and model uncertainties, and a new training set consisting of sequential features and new labels is generated. Finally, a heteroscedastic neural network (HNN) is trained to capture both uncertainties as well as output RUL point estimation. The effectiveness of the proposed method is validated on CMAPSS datasets.