Domain adaptation state of health estimation for lithium-ion batteries accounting for state of charge dependence

Accurate State of Health (SOH) estimation is essential for ensuring the safety, reliability, and long-term performance of electric vehicle batteries. However, the generalization performance of Artificial Intelligence methods depends on the quality and coverage of training data. Acquiring high-quality labeled data with sufficient coverage remains a challenge. To address this, this paper first comprehensively compares and analyzes the State of Charge (SOC) dependence of the mechanical, electrical, and thermal features of batteries under different initial SOC. Features with weak SOC dependence are selected as model inputs, enabling robust SOH estimation in partial charge–discharge scenarios using only fully completed charge–discharge data. This effectively reduces the model’s demand for broad data coverage. Then, this paper proposes a domain adaptation method based on Maximum Mean Discrepancy (MMD), which integrates the Transformer and Kolmogorov–Arnold Network (KAN). Experimental results show that adopting features with weak SOC dependence reduces the relative Mean Absolute Percentage Error (MAPE) by 85.48% for slow charging and 64.33% for fast charging. Moreover, the proposed domain adaptation method achieves a 47.69% relative reduction in MAPE.