Estimating the health status of lithium-ion batteries using features extracted from relaxation voltage

Abstract Accurate estimation of the State of Health (SOH) for lithium-ion batteries (LIBs) is critical for ensuring system safe operation. However, the presence of capacity recovery during battery aging, coupled with experimental uncertainties, introduces fluctuations and redundancies in the features extracted from post-discharge relaxation voltage curves, which in turn adversely affect the accuracy of SOH prediction. For this reason, this paper proposes a SOH estimation framework that integrates relaxation voltage features, autoencoder-based feature enhancement, and a transformer-inspired dual-residual network. Firstly, five physically original features are extracted from the relaxation voltage curve, including open-circuit voltage (OCV), first-second relaxation voltage data (FSRVD), integral area (IA), Dynamic Time Warping (DTW) distance, and Wasserstein (WAS) distance. Next, the autoencoder is designed to compress and denoise these original features, producing three-dimensional enhanced features (AE1–AE3) that improve feature stability and representation. Then, a transformer-inspired neural network with dual residual connections is constructed to enhance model depth and training stability. Finally, SOH estimation was performed under different discharge rates and SOC conditions. Experimental results show that the proposed method achieves favorable prediction accuracy under different conditions. Compared to original features, the enhanced features significantly improve SOH estimation, yielding an average Root Mean Square Error (RMSE) reduction of 23.33%, which provides valuable insights into the development of advanced battery management systems.