Rapid health estimation of in-service battery packs based on limited labels and domain adaptation

Abstract

For large-scale in-service electric vehicles (EVs) that undergo potential maintenance, second-hand transactions, and retirement, it is crucial to rapidly evaluate the health status of their battery packs. However, existing methods often rely on lengthy battery charging/discharging data or extensive training samples, which hinders their implementation in practical scenarios. To address this issue, a rapid health estimation method based on short-time charging data and limited labels for in-service battery packs is proposed in this paper. First, a digital twin of battery pack is established to emulate its dynamic behavior across various aging levels and inconsistency degrees. Then, increment capacity sequences (△Q) within a short voltage span are extracted from charging process to indicate battery health. Furthermore, data-driven models based on deep convolutional neural network (DCNN) are constructed to estimate battery state of health (SOH), where the synthetic data is employed to pre-train the models, and transfer learning strategies by using fine-tuning and domain adaptation are utilized to enhance the model adaptability. Finally, field data of 10 EVs exhibiting different SOHs are used to verify the proposed methods. By using the △Q with 100 mV voltage change, the SOH of battery packs can be accurately estimated with an error around 3.2%.