A Hybrid Method for Robust Estimation of Lithium-Ion Battery Health Across Random Voltage Windows

In many real-world applications of batteries, the partial charging/discharging involves random voltage windows. However, data-driven state-of-health (SOH) estimation strongly relies on specific voltage windows to extract effective health indicators, hence its performance degrades outside these voltage windows. To robustly estimate SOH across random voltage windows, this article proposes a new hybrid method, which integrates an online-updating empirical degradation model (EDM) with multiple Gaussian process regression (GPR) models in the framework of an extended Kalman filter (EKF). Specifically, each GPR model is trained over a distinct voltage window of 0.05 V, and only those models within the partial curve will be activated during online estimation. Then, an adaptive weight assignment mechanism is proposed: the variance of activated GPR models is used to compute their credibility, while the EKF adaptively adjusts the weights of GPR and EDM based on the credibility. This mechanism ensures robust SOH estimation under any voltage window: activated GPR models provide accurate SOH estimation when they are credible, while the EDM dominates the SOH estimation when GPR models are incredible. Experimental results on different battery chemistries and operating conditions demonstrate that the method can accurately estimate SOH even over an arbitrary voltage window of 0.05 V.