In the context of energy transition, accurate state of health (SOH) estimation of lithium-ion batteries is essential for safe system operation and lifetime management. To address the problems that fragmented charging data cannot fully characterize battery degradation and that significant sample distribution discrepancies across operating conditions reduce model generalization, this paper proposes a cross-domain SOH estimation method based on physics-data dual-feature fusion. First, degradation-sensitive physical features, including peak voltage, accumulated capacity, dQ/dV value at the peak, and peak time ratio, are extracted based on full-life-cycle data analysis and battery degradation mechanisms. Meanwhile, fragmented charging voltage curves are used as inputs, and a convolutional neural network (CNN) is employed to extract data-driven features. Subsequently, physical and data features are fused to construct a degradation representation with both mechanistic interpretability and data representation capability. Maximum mean discrepancy (MMD) is introduced to align the feature distributions between the source and target domains, thereby improving the cross-condition transfer capability of the model. To validate the proposed method, lithium-ion battery degradation experiments were conducted under four charging and discharging conditions. Full-life-cycle data from 39 NMC 18650 cells were collected, and 12 cross-condition transfer tasks were constructed. Experimental results show that the proposed method achieves the lowest estimation errors in all 12 tasks, with an root mean square error (RMSE) of 0.022 3 and an mean absolute percentage error (MAPE) of 2.05%. Compared with the CNN model using only data features, the average RMSE and MAPE are reduced by approximately 52.6% and 53.0%, respectively. In comparisons with long short-term memory (LSTM), gate recurrent unit(GRU) , Transformer, domain-adversarial neural network(DANN), and typical domain adaptation methods, the proposed method shows better overall performance and demonstrates certain transferability on the MIT cross-batch dataset. The results indicate that collaborative modeling of physical and data features, together with MMD-based domain alignment, can effectively improve the accuracy and generalization capability of cross-condition SOH estimation for lithium-ion batteries.