<p>Accurate estimation of the state of health (SOH) of lithium-ion batteries (LIBs) is crucial for ensuring the safe and stable operation of energy storage systems. However, existing data-driven methods suffer from reduced prediction accuracy and reliability due to their reliance on empirically selected hyperparameters and complex deep learning models, as well as the presence of significant noise and outliers in raw data. To address these challenges, this paper proposes a SOH estimation framework based on a multi-stage feature optimization and an improved grey wolf optimizer for back propagation neural network (IGWO-BP). First, twelve health features (HFs) were extracted from the charging data of two battery datasets. Subsequently, the four optimal HFs were selected using Pearson, Spearman, and Kendall correlation coefficients, followed by dimensionality reduction via kernel principal component analysis (KPCA) to reduce computational complexity. Furthermore, the IGWO was developed by improving the convergence factor and position update formula of the traditional GWO, while incorporating Levy flights, a Logistic-Tent chaotic map, and a dimension learning-based hunting (DLH) search strategy. This algorithm was then employed to optimize the initial weights and thresholds of the BP neural network. Experimental results demonstrate that the proposed IGWO-BP model achieves superior performance across various training set ratios. Specifically, the root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) of the model are all maintained below 1%. Comparative analysis confirms that the proposed method significantly outperforms other methods, providing an effective solution for LIB SOH estimation.</p>