Abstract: Objectives To improve the accuracy and robustness of SOC estimation for large-capacity marine LiFePO4 batteries, and solve problems of traditional methods such as inaccurate estimation in the voltage plateau region and sensitivity to noise and initial errors. Methods An experimental platform was built using EVE Energy LF280K LiFePO4 batteries. Current and voltage data under marine conditions (HPPC, FUDS, 0.5C constant current, 0.5P constant power) were collected. A dual-observation SOC estimation method fusing AEKF and iTransformer was proposed: with the second-order RC model combined with AEKF as the physical basis, iTransformer captures temporal features to generate SOC soft measurement values. These values and measured voltage form a dual-observation vector, and joint estimation is realized via AEKF’s adaptive noise estimation and dynamic Kalman gain weighting. Results Under multiple conditions, the method has a MAX error of 0.42% and an RMS error of 0.31%, outperforming the AH method, standalone AEKF and iTransformer. It maintains stable estimation even with abnormal current/voltage jumps and 15% SOC initial error, showing strong robustness. Conclusions This method combines model-driven physical constraints and data-driven nonlinear fitting ability. It solves issues like estimation deviation in the voltage plateau, noise and initial errors, provides a high-accuracy, high-robustness engineering solution for marine BMS SOC estimation, and has important application value for electric ship technology development.