Data-augmented SOH estimation for lithium-ion batteries under small-sample conditions: A hybrid STL-transformer-TimesNet approach

Accurate assessment of state of health (SOH) is essential for optimizing the safety and lifespan of lithium-ion batteries. However, existing data-driven models often fail to adequately capture battery degradation patterns under small-sample conditions and exhibit limited capability in modeling complex temporal dependencies. To overcome these limitations, this paper proposes a novel framework that deeply integrates few-sample generative data augmentation with interpretable multi-modal temporal decomposition for accurate SOH estimation. First, to mitigate the scarcity of training data, an improved generative adversarial network (GAN) is designed to reconstruct the distribution of battery degradation features, thereby expanding the sample space and effectively characterizing battery aging behavior. Second, a collaborative framework for SOH estimation is introduced based on Seasonal-Trend decomposition using Loess (STL). In this framework, the trend and seasonal components are modeled separately using TimesNet, while the residual component is processed with iTransformer. Finally, a generative-augmented multi-modal temporal collaborative framework is proposed, which deeply integrates small-sample data augmentation with interpretable temporal decomposition to estimate battery SOH under data-limited conditions. Experimental results demonstrate that the estimated SOH using the developed method show high estimation accuracy, with root mean square error all within 1 %, substantially outperforming existing data-driven approaches such as LSTM, CNN, and Transformer.