An Integrated Method for Lithium-ion Batteries Remaining Useful Life Prediction Based on Unscented Particle Filter and Relevance Vector Machine

Accurate remaining useful life (RUL) prediction for lithium-ion batteries plays a crucial role to prevent accidents from happening. In this paper, we proposed an hybrid method based on unscented particle filter (UPF) and relevance vector machine (RVM), where the genetic algorithm optimized support vector regression (GA-SVR) was uesd to correct the error generated in the prediction process. RVM is first used to extract feature points and initialize degradation model. In order to implement the RUL prediction, UPF is employed to extrapolate trends of the battery degradation capacity, where the unscented Kalman filter (UKF) is applied to generate a proposal distribution as an importance function for PF. In addition, GA-SVR is uesd to correct prediction error. Finally, to verify the effect of the proposed UPF-GASVR method with higher prediction accuracy, NASA battery dataset was used to carry out the test and compare with some classical methods.