Lithium-ion Battery Health State Prediction based on Adaptive Fine-tuning Learning Approach

Grasping battery decline patterns from existing battery data and predicting the early life of other batteries by means of fine-tuning has become a hot research topic in transfer learning. However, rigidly freezing specific layers of a neural network for fine-tuned learning is not sufficient to address the trade-off between training efficiency and accuracy. To this end, this study proposes an optimization method that flexibly adjusts the parameters of the freezing layer of the neural network according to the difference in data distribution between the source and target domains. Firstly, the data processing and feature engineering work including data cleaning, outlier processing, and feature selection are completed in the pre-processing stage of the data. Secondly, a CNN-CBAM-BiLSTM model is built to learn the recession patterns of known cells on the source domain. Finally, the maximum mean difference is used to quantify the difference in distribution between the source and target domain data, thus freezing the neural network layers at different depths. The results show that the method is capable of adaptive fine-tuned learning based on the distribution of the given samples, with a prediction error within 3%, and has great potential for application.