Phys-TSGAIN: A Physics-Informed Generative Imputation for Data Completeness Governance of Lithium-Ion Battery Time Series

Data completeness is a prerequisite for accurate life prediction and health management of Lithium-ion Batteries (LIBs). However, real-world battery time series often suffer from complex missing patterns due to sensor malfunctions, transmission errors, or human oversight. Existing imputation methods, which largely rely on statistical patterns, often fail to maintain physical consistency and lack domain knowledge guidance. To address this, we propose Phys-TSGAIN, a physics-informed generative imputation framework for LIBs data completeness governance. Unlike purely data-driven approaches, this method explicitly embeds electrochemical degradation mechanisms into a Generative Adversarial Network (GAN) through a novel Physics-Informed Knowledge Module and a Dual-Channel Hint Mechanism. Theoretically, we provide rigorous derivations to guarantee the unbiasedness, effectiveness, and convergence of the proposed algorithm. Extensive experiments across five datasets demonstrate that Phys-TSGAIN significantly outperforms traditional and state-of-the-art baselines under various missingness scenarios (MCAR, MAR, and MNAR). Specifically, the method achieves an average RMSE of 0.076 and MAE of 0.069 in downstream prediction tasks, reducing the error by approximately 73% compared to competing methods. Phys-TSGAIN effectively bridges the gap between data-driven learning and physical priors, providing high-quality data support for reliable battery modeling.