A multi-domain degradation signal feature extraction and adaptive feature fusion regression method for state of health prediction in lithium-ion batteries

This paper proposes a multidimensional feature extraction and adaptive feature fusion regression method for state-of-health (SOH) prediction in lithium-ion batteries. This method addresses the high dimensionality and redundancy of raw charge–discharge data and the difficulty of balancing prediction accuracy and computational efficiency. First, based on the two-stage degradation process of charging and discharging, feature curve clusters are constructed in the time, state-of-charge (SOC), and differential response domains. Parameters including the response centre, response width, main inflection point location, and slope contrast are extracted to form 52-dimensional candidate health features. These features are jointly evaluated and filtered by considering correlation, stability, foresight, and stage sensitivity, yielding nine key health features. Furthermore, an adaptive feature fusion regressor (AFFR) is developed to achieve accurate SOH prediction through adaptive sparse filtering, active subset refinement, and structural relaxation refitting. To validate the effectiveness and generalisation capability of the proposed method, this paper conducts experiments using two public datasets from NASA and MIT and performs comparative analyses across multiple battery samples and different training-ratio conditions. The results demonstrate that the proposed method predicts SOH with high accuracy and robustness across different degradation trajectories. The RMSE for typical battery samples is as low as 0.0020, and the overall predictions agree well with the actual SOH degradation trends. These results indicate that the proposed method improves SOH prediction accuracy and computational efficiency while maintaining the physical interpretability of the features and provides a reliable basis for practical online battery health management applications.