Accurate estimation of battery state of health (SOH) is crucial for ensuring the safety and reliability of electric vehicles (EVs). However, the low-quality field data and the scarcity of reliable SOH labels hinder the development of SOH estimation methods. This study proposes a SOH estimation framework based on a patch cross-variate Transformer (PatchCVT) architecture. First, a multifactor correction method is developed for capacity calculation. It improves the reliability of SOH labels under varying operating conditions. Then, a local patching strategy and a cross-variate attention mechanism are designed to capture temporal dependencies in battery degradation as well as interactions among input features. To further enhance the model’s performance, a masked self-supervised pretraining strategy is introduced. It leverages unlabeled data and learns generalizable feature representations. Finally, the framework is validated using 1 year of real-world operational data collected from 41 EVs. Results show that PatchCVT achieves an estimation root-mean-square error (RMSE) of 0.894%, representing the lowest error metrics among all baseline models. This error further decreases to 0.729% after pretraining. Moreover, the framework is extended to cross-domain transfer tasks. A pretrained PatchCVT fine-tuned on target data achieves comparable performance to its supervised-transfer version, with the RMSE differing by only 0.353%. These results underscore its applicability to large-scale field data and offer a viable solution for battery health management.
Wang et al. (Wed,) studied this question.