A generic physics-informed neural network framework for lithium-ion batteries state of health estimation

Accurate, practical, and robust battery state of health (SOH) assessment is crucial for the efficient and reliable operation of electric vehicles. However, reliable and precise battery SOH estimation remains challenging due to the complexity of battery aging models and the difficulty in parameter identification. In this paper, a general physics-informed neural network (PINN) framework is proposed, which uses a two-branch structure to achieve accurate estimation of SOH. Specifically, the degradation model is established by reconstructing the solid electrolyte interphase (SEI) film growth model using Universal Differential Equations (UDE). The extracted features are input into the first multilayer perception (MLP) branch to generate a preliminary SOH estimation. These features, initial SOH and the partial derivative of SOH to features are innovatively input into the degradation model represented by the second MLP branch, and the SEI growth model is explicitly simulated. Finally, the SEI film volume prediction results are mapped back to SOH. An applied voltage window (approximately 71 %–90 % state of charge range) is designed to extract 13 effective health indicators, making the method applicable to different battery types and charge-discharge protocols. Additionally, 83 batteries and 34,696 data samples including five experimental LFP/graphite batteries and three others with datasets are used to validate proposed approach. The mean absolute percentage error (MAPE) is 0.97 % across all test sets. Compared with other network architectures, the proposed PINN framework demonstrates excellent performance in all normal, small samples and transfer tests.