Abstract: This study addresses the engineering requirements for lithium-ion battery state-of-health (SOH) estimation in large-scale energy storage systems, where both high estimation accuracy and low computational complexity are essential. To meet these requirements, a local voltage fitting (LVF) model based on a second-order equivalent circuit was proposed and validated using both laboratory and public datasets. The proposed LVF model directly utilizes the voltage response during the constant-current charging phase and constructs a mapping relationship between the terminal voltage and charging time, in which the circuit parameters and SOH are treated as coupled fitting coefficients. Unlike conventional data-driven models that rely on large training datasets and complex hyperparameter tuning, the LVF model is built based on clear physical principles and requires only a small amount of easily accessible charging data. To improve the fitting robustness and general applicability, the model employs segmented voltage data within the state-of-charge (SOC) range 0.30, 0.90, avoiding the low- and high-SOC regions, where the voltage curve exhibits strong nonlinearity and is difficult to approximate accurately. In this study, experimental validation was conducted using two identical 18650-type lithium iron phosphate (LiFePO₄) batteries, denoted as LP1 and LP2. The effects of different SOC segmentation strategies on the estimation accuracy and computation time were compared based on the equivalent circuit parameters obtained from hybrid pulse power characteristic (HPPC) tests. Results show that when applying the LVF model within the SOC range 0.30, 0.90, the root mean square error (RMSE) of SOH estimation for LP1 and LP2 reached 1.249×10⁻³ and 2.678×10⁻³, respectively, while the corresponding fitting times were 7.46 and 6.86 s. These results indicated that the proposed LVF model achieved an excellent balance between accuracy and computational efficiency. To further evaluate its performance, the LVF model was compared with two representative data-driven algorithms: the back propagation (BP) neural network and the gradient boosting regression trees (GBRT). Using the LP2 dataset for training and the LP1 dataset for testing, the comparison demonstrated that the LVF model achieved a maximum relative error (MRE) of 9.489×10⁻³, an RMSE of 1.249×10⁻³ and MAPE of 0.1037%. Compared with the BP and GBRT algorithms, the MRE decreased by 57.64% and 53.94%, respectively, whereas the RMSE decreased by 83.12% and 79.12%, respectively. To validate its generalizability, the LVF model was applied to an open-access NCM battery dataset from Tongji University. Six batteries (N11, N13, N14, N24, N25, and N27) were used to estimate the SOH of the target battery (N12) under consistent and inconsistent conditions. When the batteries exhibited good consistency, the data-driven methods achieved comparable accuracy; however, under poor consistency, their estimation errors increased significantly, whereas the LVF model maintained an RMSE of approximately 4.856×10⁻³. This demonstrates the robustness, insensitivity to battery consistency, and adaptability of the model across different Li-ion chemistries. In summary, the proposed LVF model realizes an accurate, physically interpretable, and computationally efficient SOH estimation using only partial constant current–voltage data. This reduces the dependence on complete charge profiles and large-scale datasets while maintaining high estimation precision. The model has been validated on both LiFePO₄ and NCM batteries, confirming its versatility across chemistries. Given its low computational cost, independence of data consistency, and ease of parameter acquisition, the LVF model offers a promising solution for real-time SOH estimation and condition monitoring in large-scale energy storage systems.