Optimizing the effect of battery relaxation on electrochemical impedance spectroscopy measurement for real-time SOC estimation using transfer learning

Conventional Electrochemical Impedance Spectroscopy (EIS) measurements require extended battery rest periods, restricting real-time use in battery management systems (BMS) and limiting generalization across chemistries. This work introduces transfer learning to significantly shorten rest time and improve model adaptability, enabling more practical and scalable EIS integration for real-world state of charge (SOC) estimations. EIS experiments were performed on 52 Ah Lithium Iron Phosphate (LFP) and 3.6 Ah Nickel Cobalt Manganese (NCM) cells at varying SOCs, with and without rest periods. The transfer learning-based Deep Neural Network (DNN-TL) model achieved high SOC estimation accuracy for LFP cells with mean squared error (MSE) of 0.0063 and mean absolute error (MAE) of 0.0664, improving MSE by 77.58% and MAE by 50.92% compared to standard models. Additionally, only 30% of the original dataset size was needed for retraining. Applying the DNN-TL model trained on LFP data to NCM cells using unrested EIS data resulted in up to 82.08% reduction in MSE and 53.15% in MAE, requiring only 20% of the original data size for retraining. • Enhance Li-ion battery SOC estimation with transfer learning-based neural network. • Electrochemical Impedance Spectroscopy data collected for different battery types. • Reduce data collection burden in real-time SOC estimation using unrested EIS data. • Leverage and transfer electrochemical features learned from rested EIS data. • Improve the generality of data-driven SOC estimator for different battery types.