A two-stage data-driven approach to remaining useful life prediction via long short-term memory networks

Abstract

Accurate remaining useful life (RUL) prediction is of great importance for predictive maintenance. With the recent advancements in sensor technology and artificial intelligence, the data-driven approaches to RUL prediction of industrial equipment have gained a lot of attention. However, past researches have not adequately considered the variety of degradation rates and the accumulated information in degradation processes. To deal with this problem, a novel two-stage machine learning approach of RUL prediction is proposed in this paper. A set of nonlinear health indicator functions are constructed to guide the training process of a long short-term memory learner of degradation processes, then a time delay neural network is utilized for RUL prediction. The superiority of the proposed approach in terms of prediction accuracy and conservativeness is demonstrated by a case study of rolling element bearing dataset.