Enhanced Battery Health Estimation Using a Multi-Kernel ARD-GPR Framework with Bayesian Optimization for Li-Ion Batteries

The estimation of battery aging plays a crucial role in electric vehicle applications. Currently, battery health status, including capacity estimation, has attracted extensive research and innovation. Gaussian process regression (GPR) is a prominent non-parametric modeling approach that can be used to evaluate battery degradation mechanisms while quantifying uncertainty. The kernel functions of GPR and their related length scale parameters are relatively sensitive to the input distribution, However, they are often fixed during GPR applications, which may not be suitable for the battery’s dynamic circumstances. In this study, an automatic relevance determination (ARD)-based GPR with a constructed kernel function is proposed to ensure the estimation performs as promised, where the ARD is applied to capture five health indicators (HIs) during the battery charging-discharging process. Through Bayesian optimization for the parameter identifications for varying time windows, the estimated multiple dimensional length scales of ARD-GPR show the dependence of the degradation process during the various phases. According to the cross-validation results under a complex environment, the proposed multi-kernel ARD-GPR can provide a precise estimation for experimental battery cells with an average RMSE of 0.0195, which is 30.4% lower than the conventional methods. Moreover, the proposed multi-kernel ARD-GPR locates the degradation level by attributing the variation of ARD length scales under various battery degradation circumstances, which can offer an insight into the battery degradation mechanism.