Accurate state of health estimation of lithium-ion batteries based on deep learning and polarization features

Accurate state of health (SOH) estimation of lithium-ion batteries is crucial for ensuring the safety of electric vehicles. However, factors such as operating conditions and C-rates can influence battery degradation, which is typically accompanied by voltage fluctuations and increased polarization. Existing studies predominantly rely on direct utilization of voltage and current data for SOH estimation, neglecting the polarization features, thereby compromising estimation accuracy. To address this, this study proposes a deep learning model (CBS-Net) based on polarization features for SOH estimation. This model integrates convolutional neural networks (CNN), bidirectional gated recurrent units (BiGRU), squeeze-and-excitation network, and residual connections. It achieves SOH estimation by inputting only three features: initial charging voltage, initial discharging voltage, and the average voltage change rate. Publicly available lithium-ion battery datasets are employed to train and validate the accuracy and generalization ability of the CBS-Net. Results demonstrate that the coefficient of determination ( R 2 ) > 0.99, with both root mean square error (RMSE) and mean absolute error (MAE) < 1%, indicating excellent estimation accuracy. Moreover, during generalization validation, the model maintains an R 2 > 0.98 with RMSE and MAE < 3%, confirming its strong generalization ability. In comparisons with single models such as CNN and BiGRU, CBS-Net exhibits the best estimation accuracy. Taking CNN as an example, the RMSE and MAE of the CBS-Net model are reduced by 75% and 76%, respectively. This study provides a novel model framework and an effective health features selection method for battery SOH estimation. • A novel model combining CNN, BiGRU, SE, and residual connections is proposed. • The novel polarization features are used for SOH estimation. • Accurate estimation of SOH can be achieved only by voltage data. • The method exhibits high accuracy, with RMSE and MAE < 1%