Distance-Based Embedding Learning for Remaining Useful Life Estimation

Remaining useful life (RUL) estimation is a core content of Prognostics and Health Management (PHM). Easy access to sensor data allows data-driven methods to achieve better results than physical or expert modeling methods in complex systems. However, the lack of run-to-failure data remains a problem. To maximize the use of failure samples, in this paper, long short term memory (LSTM) network is utilized to learn a robust embedding. Then RUL of a sample is obtained by calculating the distance between embedding of failure and embedding of the sample. An end-to-end model is trained to achieve the aforementioned functions. Experiments on C-MAPSS dataset show competitive results especially in subsets of multiple conditions.