Enhanced few-shot state-of-health estimation for lithium-ion batteries via Masked Autoencoder

Accurately estimating the state-of-health (SOH) of lithium-ion batteries (LIBs) is crucial for optimizing performance, ensuring operational safety, and enabling predictive maintenance in battery management systems. With the widespread adoption of LIBs, a large amount of field data has been generated, yet current data-driven SOH estimation methods often fail to fully utilize it due to the lack of labeled data. To address this, we propose a method based on semi-supervised learning to exploit large-scale unlabeled data for accurate SOH estimation. A generative unsupervised model, the Masked Autoencoder (MAE), is pre-trained on unlabeled field charging data to automatically extract latent representations related to SOH. The model is then fine-tuned with a small amount of labeled data. Experimental results show that using only 20 % of the labeled data usually required for supervised learning, the method achieves an RMSE of 2.14 %. The latent representation extraction capability of the MAE is validated via incremental capacity (IC) analysis, which explains the 14 % improvement in estimation accuracy (RMSE of 1.84 %) when using data from a specific voltage range (3.8–3.9 V). Furthermore, experiments demonstrate that even with only 21.33 min of charging data—consisting of only charge quantity and voltage signals—the model can still achieve a competitive RMSE of 1.94 %. This work introduces a novel approach for SOH estimation using large-scale, unlabeled field data and provides valuable insights for battery management in the era of artificial intelligence.