An Efficient Segment Anything Model Adaptation Method for Electrode Overhang Analysis in Lithium-Ion Battery Manufacturing

With the global trend towards carbon neutrality, Lithium-ion battery has been one of the leading contributors for the energy storage. During its manufacturing process, the overhang analysis at the electrode winding stage is critical for ensuring the energy density and cell uniformity, which first requires accurately identifying and separating each electrode. In the related literature, customized segmentation models are usually built for specific scenarios, overlooking the high diversity of imaging modalities, battery types, etc. Recently, some works begin to explore the potential of segment anything model for zero-shot predictions in industrial applications. However, it has been pretrained on public datasets of natural RGB images with a large discrepancy of characteristics, and directly fine-tuning it requires massive well-annotated battery cell images and considerable computations. To deal with the challenge and pioneer the utilization of segment anything model in Lithium-ion battery manufacturing, an efficient downstream adaptation method is proposed in this paper to achieve accurate and generalized segmentation of electrodes. The encoder and decoder are jointly updated with less than 3% of original model parameters. Extensive experiments show that the proposed method with only 16 images far outperforms the basic segment anything model and its variants with full fine-tuning.