An online fine-tuning method for modelling online SOH estimation under multi-stage constant current charging protocol

With the rapid advancement of artificial intelligence, data-driven methods have emerged as a promising approach for estimating the state of health (SOH) of lithium-ion batteries. However, these methods sometimes suffer from poor model generalization due to unavoidable differences arising from variations in the materials' microstructure, even when batteries are used under the same operating conditions. To address this issue and improve model generalization, this study proposes an end-to-end convolutional neural network-transfer learning online (CNN-TLON) method. This end-to-end approach captures the relationship between charging sequences and SOH through pre-training a convolutional neural network model, followed by a transfer learning-based fine-tuning strategy that enables online model updates to adapt to differences between batteries. A multi-stage constant current charging protocol, an effective approach that has been relatively underexplored in previous research, was adopted in this study to conduct cycle aging experiments on four batteries. The estimation results demonstrate that the proposed method achieves a root mean square error (RMSE) of only 23.29 mAh in estimating the cycling maximum capacity. In addition, on this dataset, the proposed method outperforms the traditional machine learning method in all evaluation metrics. Furthermore, the proposed method has been applied to the MIT and Oxford datasets, yielding reliable and accurate capacity estimation results, with an approximately 40 % reduction in RMSE compared to the baseline CNN.