A parallel GRU with dual-stage attention mechanism model integrating uncertainty quantification for probabilistic RUL prediction of wind turbine bearings

Abstract

The accurate probabilistic prediction of remaining useful life (RUL) of bearings plays an important role in ensuring the safe operation of wind turbine maintenance decision making. However, it is still a challenge to improve prediction accuracy and quantify prediction uncertainty, leading to the inability of precision and reliable prediction results. Accordingly, this paper proposed a novel probabilistic RUL prediction method combined the parallel GRUs with dual-stage attention mechanism (PDAGRU) prediction model with the non-parametric uncertainty quantification approach to overcome the limitations. In the PDAGRU model, the dual-stage attention mechanism is developed to improve the capability of degradation information extraction. Meanwhile, the parallel structure can enhance prediction accuracy and help quantify model uncertainty. The proposed uncertainty quantification approach with less prior knowledge can provide probabilistic RUL prediction results based on the kernel density estimation (KDE) and Monte Carlo (MC) dropout. Moreover, a first prediction time (FPT) determination method based on the isotonic regression is developed to more accurately reflect the degradation trajectory of bearings’ RUL. Two cases including simulated data and real-world data are deployed to verify the effectiveness of the proposed method. Compared with other methods, the superiority of the proposed method is verified. The RUL prediction accuracy and interval coverage probability for wind turbine bearings are high to 90.4% and 99.7% respectively.