Incorporating dynamic mode decomposition and domain adversarial training for cross-domain state of health estimation of lithium-ion batteries

State of Health (SOH) is a critical indicator for evaluating the degradation level of lithium-ion batteries, yet it cannot be measured directly and must be estimated from operational data. Although notable progress has been achieved in SOH estimation, most existing methods rely on extensive aging experiments to obtain labeled data for target batteries, which require substantial experimental effort and limit practical applicability. This limitation is further aggravated under cross-domain conditions and in offline application scenarios. In this work, we propose a deep-learning-based framework for cross-domain SOH estimation with minimal reliance on labeled target battery data. The framework extracts SOH-related feature sequences from constant current charging curves and employs dynamic mode decomposition to characterize degradation dynamics. Domain adversarial training is further introduced to learn domain-invariant representations using only a small portion of target domain data. The proposed method is validated on battery datasets covering multiple domains, demonstrating reliable SOH estimation performance under significant domain discrepancy. This work highlights the potential of data-driven transfer learning approaches to reduce the need for degradation experiments and to support efficient SOH estimation in both online and offline application scenarios. • Charging-curve features are efficiently derived and SOH-informative. • DMD extracts intrinsic representations strongly correlated with SOH. • A domain-adversarial model learns invariant features with 5% labeled samples. • Extensive cross-domain validation confirms robust SOH estimation.