Sequence to sequence network with Bayesian attention and state transition for self-data-driven remaining useful life estimation

Accurate remaining useful life (RUL) estimation is a prerequisite for achieving predictive maintenance. Most of the existing research on RUL prediction driven by deep learning belongs to the pattern recognition paradigm. Such methods usually require run-to-failure data from multiple entities for training to obtain excellent prediction results. However, the actual industrial scenarios usually fail to meet this requirement. To this end, this paper proposes a sequence to sequence trend prediction network with Bayesian attention and state transition (SeqBAST) for self-data-driven RUL prediction. SeqBAST predicts degradation trend based on historical monitoring data, and RUL can be determined by the time when the predicted trend reaches the set failure threshold. Specifically, Bayesian attention is designed into long short-term memory network to capture the temporal dependencies during the degradation process. In addition, state transition module is constructed for nonlinear degradation trend modeling. Finally, SeqBAST generates the probabilistic estimation of RUL to provide prediction uncertainty. The effectiveness of the proposed method is verified on the data from multiple bearings and engine units.