With the widespread use of lithium-ion batteries (LIBs) in electric vehicles, it is essential for battery management systems (BMS) to accurately estimate the state of health (SOH) of the battery. This study constructs a regression estimation model applicable to the SOH of LIBs by combining multi-feature signal analysis with a data-driven approach. First, the collected data were smoothed using advanced filtering methods to extract input features characterising capacity degradation properties from differential thermal voltammetry (DTV) curves, incremental capacity analysis (ICA) curves, meantime decay (MTD) curves, and terminal thermal analysis (TTA) curves. Then, the extracted features are subjected to feature fusion using principal component analysis (PCA) to obtain strongly correlated features with a cumulative contribution of 98% or more as inputs to the data-driven model and the hyperparameters of the group method of data handling (GMDH) network are optimised using simulated annealing genetic algorithm (SAGA). Finally, model construction, validation, and comparison are carried out using NASA and Oxford battery datasets. The RMSE between the actual measured value and the estimated value is obtained to be less than 0.3%, and the R ² value is as high as 0.99. The results demonstrate that the approach can efficiently catch the phenomenon of capacity regeneration. It is valid and robust for SOH estimation for different types of LIBs.

Estimating lithium-ion battery state of health (SOH) across operating conditions is challenging because operating profiles, measurement characteristics, and degradation trajectories differ among domains. This study proposes a physics-informed transfer learning (PITL) framework that combines a common health descriptor space, parameter function decomposition, and PINN–DeepHPM regularization. Open-circuit voltage reconstruction is used before feature extraction for dynamic profiles, whereas constant-current data are processed directly in the same descriptor space. During target-domain adaptation, selected shared-representation layers are frozen and the predictor and dynamic constraint parameters remain trainable. In the predefined Cell_1# tests, the resulting feature transfer configuration reduced RMSE relative to full fine-tuning by 23.53% for 21700_Dynamics and 61.78% for CEPREI. In leave-one-cell-out validation, PITL and full fine-tuning had comparable mean RMSEs on 21700_Dynamics, whereas full fine-tuning was more accurate on CEPREI. Freezing approximately 48% of the parameters reduced adaptation time and peak GPU memory. Thus, partial freezing can lower adaptation cost, but its accuracy benefit is target dependent.