Accurate State of Health Estimation of Lithium-Ion Batteries Based on L-BERT Large Model

Accurate state of health (SOH) estimation of lithium-ion batteries is critical to ensuring the safety and improving performance of battery systems. This paper proposes a SOH estimation model, L-BERT, which combines a long short-term memory (LSTM) network with Bidirectional Encoder Representations from Transformers (BERT). LSTM captures local temporal dependencies, while BERT focuses on long-range sequence interactions. Given the difficulty of traditional feature engineering, this paper uses the time intervals of current and voltage in charging segment as input features, which are easier to obtain and more sensitive to battery dynamic degradation. In model fine-tuning, Low Rank Adaptation (LoRA) is applied to improve model training efficiency and significantly reduce computational cost. The proposed method is evaluated on a large dataset from Xi'an Jiao tong University (XJTU). The validation results show that the proposed L-BERT model exceeds traditional LSTM and CNN models. Specifically, L-BERT's SOH estimation had a root mean square error of 0.01480, a mean absolute percentage error of 0.01199, and a coefficient of determination of 0.9256. The proposed method provides a technical route that balances accuracy and stability for SOH estimation under different operating conditions.