Cross-attention-based hybrid feature fusion network for state-of-health estimation of lithium-ion batteries

Estimating the state of health (SOH) of lithium-ion batteries (LIBs) is crucial in a battery management system. To improve the accuracy of SOH estimation, a new method that combines a Gramian angle field (GAF) and multi–model fusion is proposed. First, the GAF is used to encode incremental capacity data into an image, making small differences easier to identify. Second, a Gramian angle field–convolutional neural network–long short-term memory model with a bi-directional cross-attention–based fusion network (GAF-CNN-Fusion-LSTM) is proposed to solve the problem of original feature loss, that is, the partial loss of original temporal feature information during the image conversion process. By introducing a bi-directional cross-attention mechanism, the model enables deep interaction and effective fusion between image and time-series features, thereby improving the accuracy and robustness of SOH estimation. Finally, the proposed method was validated using the NASA and Oxford datasets. On the NASA dataset, the proposed model achieved an average root mean square error (RMSE) of 0.0033, which was 73.4%, 47.6%, 57.1%, and 44.1% lower than those of the CNN-LSTM model, GAF-CNN-LSTM model, direct-concatenation model, and uni-directional attention model from the image branch to the time-series branch, respectively. On the Oxford dataset, the average RMSE was 0.0021. These results demonstrate that the GAF-CNN-Fusion-LSTM model has higher accuracy and stronger robustness.