Battery state-of-health prediction using a novel enhanced sparse variational Gaussian process framework

Current lithium-ion battery (LIB) state-of-health (SOH) prognostics primarily focus on offline applications, which fail to meet the real-time adaptability and edge-processing efficiency required by onboard battery management system (BMS). Irregular degradation patterns introduce suboptimal computational overhead in variational inference frameworks, while dynamically reconfigured inducing points lack theoretical stability guarantees. Existing methods also fail to establish interpretable connections with physical aging mechanisms, thereby limiting their generalizability. To address these challenges, we propose an enhanced sparse variational Gaussian process (SVGP) framework that integrates dynamic kernel adaptation and incremental Bayesian compression. The proposed framework models degradation through time-varying kernel decomposition to characterize multi-mechanism aging behaviors. A variational energy-gradient-driven inducing point screening mechanism achieves Pareto optimization between computational efficiency and predictive accuracy. Experimental results demonstrate the framework’s superior performance: (1) with 24,812 data points and 2,400 inducing points, computational complexity is reduced by 99.96%, and memory usage decreases from 897MB (traditional SVGP) to 616KB; and (2) on the National Aeronautics and Space Administration (NASA) dataset , dynamic kernels reduce the prediction root mean square error (RMSE) from 1.89% to 1.32%, achieving a 72.87% reduction in error for capacity "regeneration" phenomena. By unifying LIB aging mechanism modeling with online learning efficiency, this work presents a theoretically grounded and promising solution for embedded SOH prediction. • Proposing a dynamic kernel adaptation mechanism for battery aging modeling • Developing an incremental Bayesian compression algorithm based on gradient evidence • Achieving Pareto optimization and mutation detection improvement of SOH prediction • Designing a framework that unifies LIB aging interpretability with online learning efficiency.