Accurate and adaptive state of health estimation for lithium-ion battery based on patch learning framework

• The local features extracted from the voltage change curve of the constant current stage are simpler. • The patch learning framework with the global model and the local model can achieve accurate SOH estimation. • The patch learning model has higher estimation accuracy and generalization compared with other algorithms. To solve the low accuracy and adaptivity of state of health (SOH) estimation for lithium-ion batteries, a patch learning framework is proposed in this paper. The global model and the patch model are adaptively combined to improve the adaptability and local tracking ability of SOH estimation. The gate recurrent unit model is selected as the global model for its global description capability, and the convolutional neural network-bidirectional long short-term memory model is chosen as the patch model for its local tracking capability. With the global and patch models, the patch learning model is developed by searching the patch segments until the error is below the allowable value. The proposed method is evaluated and compared with some existing methods on our battery aging dataset and public dataset. The experimental results show that the RMSE of the proposed method for all cells is within 0.72%, the MAE is within 0.59% and the MAPE is within 0.66%, and it has the highest accuracy and the best generalization performance compared with GRU, CNN-BiLSTM, SVR, ELM and LSTM methods.