Abstract
Accurate state of health (SOH) estimation is essential for effective lithium-ion battery management, particularly under fast-charging conditions with a constrained voltage window. This study proposes a hybrid deep neural network (DNN) learning model to improve SOH prediction. With approximately 22,000 parameters, the model effectively estimates battery capacity by combining local feature extraction (convolutional neural networks [CNNs]) and global dependency analysis (self-attention). The model was validated on 222 lithium iron phosphate (LFP) batteries, encompassing 146,074 cycles, with limited data availability in a state of charge (SOC) range of 80%-97%. Trained on fast-charging protocols (3.6C-8C charge, 4C discharge), it demonstrates high predictive accuracy, achieving a mean absolute percentage error (MAPE) of 3.89 mAh, a root-mean-square error (RMSE) of 4.79 mAh, and a coefficient of determination (R(2)) of 0.97. By integrating local and global analysis, this approach significantly enhances battery aging detection under fast-charging conditions, demonstrating strong potential for battery health management systems.