Physics-informed neural networks for degradation diagnosis of lithium-ion batteries via electrochemical impedance spectroscopy

Abstract

Lithium-ion batteries (LIBs) inevitably experience performance degradation during long-term service, driven by numerous interrelated coupled mechanisms. Relying solely on state of health (SOH) is insufficient for effective degradation diagnosis. Extracting degradation modes (DMs) provides more comprehensive guidance for understanding degradation mechanisms, optimizing battery design, and developing control strategies. Electrochemical impedance spectroscopy (EIS), as a non-invasive diagnostic tool, offers valuable internal kinetic information and serves as an effective supplement to time-domain data. In this study, a general and high-accuracy impedance decoupling and parameter approximation method is proposed to efficiently extract key health indicators (HIs) associated with battery degradation from EIS. Two physics-informed neural networks (PINNs) with different architectures are constructed to map HIs to SOH and DMs. To address the challenge of acquiring labeled aging data, the proposed PINNs achieve accurate SOH and DMs prediction using only 20 % of early-stage aging data for model training. The effectiveness and generalization of the proposed method are systematically validated using both widely adopted open-access datasets and in-house datasets. The results indicate that PINN framework consistently achieve root-mean-square errors below 2 % in estimating SOH and DMs across varying cell types, SOC levels, and aging states, demonstrating strong robustness, adaptability, and predictive reliability. This work provides a novel PINN design strategy for online degradation diagnosis based on early-stage EIS data, offering significant theoretical insights and practical value for battery management.