The impedance of the battery contains rich information about the internal structure and health state of a battery. Existing impedance-based state-of-health (SOH) estimation methods often rely on machine-learning algorithms to capture the complex relationship between the impedance and the SOH.

In this study, for the first time, a novel hybrid SOH estimation method is proposed, where an unscented Kalman filter (UKF) is used in combination with polynomial regression models. The battery degradation trend and the relationship between the health indicator (HI) and SOH are used to predict the state and measurement variables, respectively.

Three model structures, including polynomial regression (RP), Gaussian process regression (GPR) and feedforward neural network (FNN), are used to capture these two relationships. Open-source datasets containing 51 cells following various aging-paths are used for validation. The estimator can reach an error between 1.4 to 1.7% SOH regardless of the state-of-charge (SOC) level.

The estimator is then evaluated under non-ideal conditions. The estimator is shown to successfully adapt to (1) a large mismatches between the training and testing SOC, (2) previous unseen cell types and (3) insufficient training data. Estimation errors of about 2 to 3% can still be achieved amid these uncertainties albeit within a more limited SOC range and model structures. Comparing the three model structures suggests that the GPR can achieve the lowest estimation error while the PR can achieve slightly higher errors with a much simpler structure. The FNN can achieve similar estimation errors as the PR but is more complex and more sensitive to a reduction in the training data.

Overall, the proposed SOH estimation strategy shows good accuracy and robustness against non-ideal conditions. By correcting offline trained models using online measurements through an adaptive filter, the estimator successfully overcomes cell-to-cell variation, variations in the aging path, SOC mismatch, unseen cell types and insufficient training data.