Multi-physics data and model feature fusion for lithium-ion battery capacity estimation by transformer-based deep learning

Accurate and reliable estimation of the available capacity of lithium-ion batteries (LIBs) is of vital importance for battery management systems (BMS) to assess battery health status and optimize charge/discharge strategies. Although multi-physical characteristics offer more comprehensive health information, efficiently and accurately extracting, analyzing, and integrating these heterogeneous sources remains a major challenge in practical scenarios. To address this issue, this study proposes a Transformer-based modeling framework that integrates multi-physical features for capacity estimation. Specifically, degradation patterns of LIBs are analyzed from three external physical domains—electrical, thermal, and mechanical. Data-based and model-based health features are respectively extracted from the constant current–constant voltage (CCCV) charging phase and the voltage relaxation phase. Based on these features, an output-level feature fusion architecture is developed using a Transformer-based model to improve estimation accuracy. Experimental validation using both constant-temperature and variable-temperature aging datasets demonstrates that the proposed method performs well across multiple evaluation metrics. Compared with input-level fusion approaches, the mean absolute error (MAE) on the test sets is reduced by 6.09% and 47.81%, respectively. Relative to capacity estimation using individual feature types separately, the proposed fusion method achieves MAE reductions of 42.29% and 29.70%, respectively. Furthermore, cross-validation experiments further verify the model’s generalizability across different batteries, confirming its potential and advantages for practical applications.