Cross-Domain Detection of Lithium Battery Electrode Endpoints Based on Structural Optimization and Incremental Learning

Lithium battery electrode endpoint detection is essential for ensuring battery safety and performance. However, in industrial environments, changes in production conditions often lead to domain shifts in X-ray images, significantly degrading the performance of existing detection models. To address this challenge, this paper proposes a cross-domain detection method integrating structural optimization and incremental learning. Structurally, we enhance the YOLO model by introducing cross-layer connections in the neck network to improve multi-scale feature fusion capability. Algorithmically, we incorporate the Elastic Weight Consolidation (EWC) incremental learning strategy, which constrains important parameters to retain source domain knowledge while adapting to target domain distributions, without requiring access to original source data. Experimental results demonstrate that our method achieves 82.2% mAP on the target domain while maintaining 73.1% mAP on the source domain, significantly outperforming traditional transfer learning approaches and effectively mitigating catastrophic forgetting.