Interactive fusion of local and global degradation representations for rapid estimation of lithium-ion battery state-of-health

Accurate estimation of lithium-ion battery state-of-health (SOH) is the basis for BMS to ensure the battery safety and prolong its lifespan along the long-term cycling operation. Regarding feature extraction for battery aging state analysis, traditional convolutional neural networks (CNNs) are advantageous in capturing local aging information, but they have deficiencies in capturing long sequence dependencies. On the other hand, Transformers exhibit superiority in capturing long sequence dependencies, while they inevitably suffer from detail loss in local aging information. This study proposes an end-to-end battery SOH rapid estimation algorithm using a concurrent structured fusion of convolutional operations and Transformer self-attention mechanism. In this paper, feature consistency between the CNN and Transformer is realized by applying an Interactive Feature Fusion Block (IFFB) which can effectively integrate and retain battery local and global aging representation. The proposed model relies only on short term charging segments to establish a mapping relationship from the global and local features to the SOH by capturing the dynamic features within and between operating variables. Two public battery datasets with different battery types and aging paths were adopted to validate the proposed model. Simulation results show that the proposed model's performance demonstrates the averaged Root Mean Square Error (RMSE) on LFP test batteries and LCO test batteries to be 0.1874 % and 0.2853 %, respectively.