<p>Lithium-ion batteries (LIBs) are widely used in renewable energy storage, electric vehicles (EVs), and portable devices, owing to high energy density, favorable cycle performance, and environmental friendliness—supporting energy transition and carbon neutrality. However, internal and external factors induce irreversible aging, leading to gradual declines in capacity and performance. State of health (SOH) is a core metric to characterize aging and remaining functionality. Thus, accurate SOH estimation is essential for battery management systems (BMS) to track aging states, guide strategies for mitigating degradation, predict remaining useful life (RUL), and improve system reliability. This paper proposes a ResAttention-Transformer with Gramian hybrid field encoding (RATGH) method, breaking through the two major bottlenecks in existing SOH estimation: the defects in multi-source time series fusion and the insufficiency in long-range degradation modeling. To address the problem of multi-source data fusion defects, a Gramian hybrid field (GHF) encoding is designed: by fusing the mean operation of Gramian angular summation field (GASF) and Gramian angular difference field (GADF), the three-variable time series of current, voltage, and temperature are synchronously mapped into RGB images, uniformly retaining capacity degradation trend features and dynamic thermal anomalies (e.g., temperature surges). Compared with single-field encoding, it significantly improves <i>R</i>² by 10.2% (reaching 97.6%). To tackle the challenge of insufficient long-range degradation trajectory tracking, a ResAttention-Transformer structure is constructed: the residual attention branch embeds a self-attention mechanism in the ResNet block to dynamically focus on key degradation features (such as abnormal voltage fluctuations) for enhancing local perception; the transformer encoder models cross-cycle global dependencies through multi-head attention to accurately capture the slow-changing degradation trends. Verification on the Toyota-MIT-Stanford LFP and XJTU NCM datasets shows that: the RMSE is as low as 0.331%, which is 29.08% lower than that of ResNet; the MAPE is 0.251%, and <i>R</i>² reaches 0.992; the error stability across chemical systems is only 1.89 times, significantly outperforming the benchmark models. Additionally, T-SNE visualization is used to enhance the interpretability of the proposed model, providing a new paradigm of high-precision and strong-generalization SOH estimation for BMS.</p>