Battery state-of-health estimation using CNNs with transfer learning and multi-modal fusion of partial voltage profiles and histogram data

Accurate estimation of battery state of health (SOH) is critical for ensuring safe and reliable operation, enabling health-conscious control, and supporting second-life applications. Existing health indicators (HIs) used in data-driven models have practicality, accuracy, and robustness limitations. For instance, partial voltage or incremental capacity curves may lead to misleading SOH estimations, while histogram-based methods require extensive training data. This paper proposes a multi-modal fusion model that integrates two types of HIs extracted from partial voltage curves recorded during charging and histogram data during operation. By addressing the limitations of both types of HIs, the proposed model achieves superior performance in terms of accuracy and robustness. The proposed model is validated on two representative datasets, achieving a root mean squared percentage error (RMSPE) as low as 0.74 %, reducing estimation error by up to 42 % compared to existing models and requiring 60 % less training data. The results demonstrate the feasibility and advantages of combining HIs from different sources, underscoring the importance of detailed feature analysis in developing data-driven models for battery state estimation.