A synergistic sequence–feature fusion framework for robust battery SOH estimation across real-world random charging profiles

Accurate State-of-Health (SOH) estimation is critical for the safety and economic management of large-scale battery systems. However, random and incomplete charging behaviors in the field degrade the accuracy of conventional models, which typically require fixed voltage windows or full charging cycles. To resolve this limitation, we present a universal SOH estimation framework based on a synergistic sequence–feature fusion architecture, by integrating a boundary-independent feature extraction protocol with a dual-stream neural network. Evaluated on a diverse dataset comprising 320,000 operational segments across six commercial cathode chemistries, including LFP, NMC, and LCO, the proposed method achieves a mean absolute error (MAE) of 1.69% and a minimum error of 1.06%, outperforming all the representative baseline models. Moreover, the error variation across different datasets is less than 1.6%, demonstrating robust cross-chemistry adaptability. This methodology provides a potential solution for intelligent battery management in real-world environments.