Physics-Informed Neural SOH Estimation Method for Lithium-ion Battery under Partial Observability and Sparse Sensor Data

Establishment of precise state-of-health (SOH) consists of a vital step to guarantee reliable, secure, and preventive maintenance of electric vehicle (EV) fleets and grid-associated energy storage systems based on lithium-ion batteries. Nevertheless, the exiting SOH estimation approaches basically contemplate the situation where the complete sensor profiles or the full charge/discharge condition are at disposal, which most of the time is not the case owing to the degradation of the sensors, communication losses, or economic factors. This work introduces a novel Physics-Informed Neural Network (PINN) framework to enable robust SOH estimation under partial observability, leveraging minimal sensor data and physics-constrained learning to ensure both accuracy and interpretability. The model relies on an approximation of the PINN architecture able to embed the fundamental degradation mechanisms, such as solid electrolyte interphase (SEI) layer growth and capacity fading, as information that the training process can use through physical regularization. While PINN is semantically mixed with traditional purely data-driven models, it uses a composite loss function that leverages data consistency with underlying electrochemical laws. This makes accurate predictions possible even when there are no further capacity measurements and no complete voltage profiles. The framework is tested in the case of publicly available lithium-ion battery data sets that show under simulated sensor sparsity scenarios, superior generalization and capacity to withstand situations where missing inputs or even sensor noise is introduced. The results of experiments demonstrate the quality of the model that enhances precision of SOH estimation in the conditions of partial observability and also increases the model's interpretability and applicability in real time in various operational settings. Statical results show that the proposed method achieves a mean absolute error (MAE) of 0.008, root mean square error (RMSE) of 0.011, and a coefficient of determination (R²) of 0.96, outperforming baseline models such as CNN–LSTM and hybrid SE-NN by 25–40% in sparse data regimes. Furthermore, the framework exhibits high generalization capability across different chemistries and retains robustness with as little as 50% of the input features. These results underscore the practical potential of the proposed PINN approach for real-time, physically consistent battery health diagnostics in embedded battery management systems