Accurate health state estimation is of great importance for li-ion battery management. The health state is usually characterized by the post-decay capacity. Traditional capacity estimation mainly passively collects three signals of battery voltage, current and temperature, and then extracts the capacity-related features. This requires high integrity and regularity of the charging and discharging curves. The method in this paper extracts five health features based on the low-frequency electrochemical impedance spectroscopy detected during charging, including three new physically significant health features, namely the modified Warburg factor, the pseudo lithium-ion diffusion state and the residual value after its empirical mode decomposition. This bridges the gap between the internal kinetic features and the external aging features of li-ion batteries and has a strong correlation with capacity. The determination coefficient R2 of the li-ion battery capacity estimation model before optimization is less than 0.6. By fusing the health features and considering the correlation between the variable groups from a holistic perspective, the model estimation accuracy can be significantly improved, and R2 can reach 0.935 7. RMSE is 0.374 9 mAh and MAPE is 0.836 2 mAh. This work is supported by the National Natural Science Foundation of China (No. 52177206 and No. 51907005).