Physics-Informed Mamba Network for Perturbation-Robust State of Health Estimation of Lithium-Ion Batteries

Abstract

Accurate state-of-health (SOH) estimation is essential for the safe and reliable operation of lithium-ion battery management systems. However, data-driven models may be sensitive to local fluctuations in input health features, resulting in unstable prediction trajectories under random perturbations. To address this issue, this paper proposes a physics-informed Mamba framework for perturbation-robust SOH estimation. Six health features are extracted from incremental capacity curves and used as time-series inputs to the Mamba network. The positive association between IC peak height and SOH is formulated as a gradient-based physical consistency loss and jointly optimized with the data-fitting loss, thereby constraining the local response direction of predicted SOH. Experiments on the NASA and CALCE datasets show that PI-Mamba achieves accuracy comparable to Mamba without additional perturbations. When Gaussian perturbations are introduced into the normalized health features, PI-Mamba errors increase more gradually, while local fluctuations in some prediction trajectories are reduced. At the highest tested perturbation level, the average RMSE on the CALCE dataset remains below 1%, and error growth on the NASA dataset is lower than that of Mamba. These results indicate that the proposed method reduces sensitivity to random feature perturbations and improves prediction stability without substantially affecting baseline estimation accuracy.