Remaining Useful Life Prognostics for Lithium-ion Battery Based on Gaussian Processing Regression Combined with the Empirical Model

Abstract


 
 
 Data-driven techniques based on Bayesian framework like Gaussian Process Regression (GPR) can not only predict the lithium-ion battery Remaining Useful Life (RUL), but also provide the uncertainty representation. However, it is always difficult to choose the covariance function of GPR and the confidence bound is usually large if the training data are not enough. In order to solve this problem, a combining method is proposed, it is a prognostic framework based on GPR model combined with Empirical Model (EMGPR) to realize the lithium-ion battery RUL prediction. EMGPR has the advantages of predicting the tendency and uncertainty management for RUL estimation. The modeling process of EMGPR consists of two steps. The self-deterministic part, which reflects the real physical process of battery degradation, is approximated by the empirical model. And the disturbance part, which is caused by random noise such as measurement and environment noise, is expressed by the GPR model. In application, two key factors of EMGPR are focused. Firstly, the prediction result is not accurate enough if the training data are not very reliable. In this case, more reliable training data should be selected optimized. Secondly, the characteristic of the disturbance is involved to determine the kernel function of GPR model. With this EMGPR framework, the RUL result is estimated with uncertainty representation, as well, the covariance function of GPR is easy to choose. Experiments with NASA PCoE and CALCE battery data show the satisfactory result can be obtained with the EMGPR approach.