Predicting Remaining Useful Life with Uncertainty Using Recurrent Neural Process

Recently deep learning based remaining useful life (RUL) prediction approaches have gained increasing attention due to their scalability and generalization ability. Although deep learning based approaches can obtain promising point prediction performance, it is not easy for them to estimate the uncertainty in RUL prediction. In this paper, a recurrent neural process model is proposed to address the prognostics uncertainty problem based on deep learning. Compared with the original neural process model, a recurrent layer is added to extract sequential information from input sliding windows. The RUL prediction problem can be considered as finding a regression function mapping the sliding window input to its corresponding RUL. By obtaining the distribution over the regression functions, the recurrent neural process is able to model the probability distribution of the RUL. As a probabilistic model, stochastic variational inference and reparameterization trick is applied to learn the parameters of the model. The proposed method is validated through the C-MAPSS turbofan engine dataset.