A cell screening method for lithium-ion battery grouping based on pre-trained data-driven model with multi-source time series data

Cell Screening with multi-source time series data for lithium-ion battery (LIB) grouping is a challenging task in the production of LIB pack. Currently, most of these cell screening methods adopt a plain data fusion strategy that does not consider the relationship between different sources in the multi-source time series data. Then, these methods sort cells with supervised models which need a large amount of labeled data to guarantee the screening performance. In this paper, we propose a cell screening method for LIB grouping based on the pre-trained data-driven model with multi-source time series data. Our method is more effective in feature extraction and less reliant on labeled data. The screening model in our method is pre-trained on a large unlabeled dataset for the cell screening relevant tasks to improve its feature extraction ability on multi-source time series data. Then, we replace the task head of the pre-trained screening model and fine-tune it on a small labeled dataset to adapt for the cell screening task. Experiments based on real-world production data of 18650 battery verify the effectiveness of our proposed method. The results show that our method can reach the top-1 accuracy of 95.8%, which outperforms other compared data-driven methods and is comparable to the performance of the same model structure with supervised learning.