Robust Physics-Informed Neural Network for State of Health Estimation of Lithium-Ion Battery

To address the key gap in the robustness of state of health (SOH) estimation of lithium-ion batterys (LIBs), this study proposes a physics-informed neural network (PINN) framework, which integrates the battery capacity degradation equation. By extracting features from part of the charging data, this architecture combines data-driven learning with physics-based regularization through a hybrid loss function. The noise-free and noisy conditions (5% Gaussian perturbation) were evaluated on the MIT dataset. This method has comparable accuracy to pure data-driven (MAE differs 5.1%), while significantly improving stability. The experimental results show that the robustness to noise has been improved by 20 times: when facing noise disturbances, the MSE of PINN only increases by 0.7%, which is superior to the 14.6% degradation of data-driven methods. Error distribution analysis shows that the variance is significantly reduced and the error distribution is more convergent. This work provides the first quantitative robustness benchmark for SOH estimation based on PINN, offering a practical solution for reliable battery management under real-world uncertainties.