An optimized detection model for copper particles in carbon powder of lithium-ion batteries

Abstract Electric vehicle power batteries face internal short circuits due to trace copper particles from anode debris and equipment wear. These particles dissolve on the cathode and form dendrites, piercing the separator. Detecting copper particles in carbon powder is critical for battery safety. This study proposes Cu-YOLO (Copper-You Only Look Once), an improved YOLOv8n (You Only Look Once version 8 nano) model for detecting copper particles in complex carbon powder datasets. Key enhancements include: the LCGC (Lightweight Channel Grouping Cascade module) in the backbone network, which reduces computational cost while maintaining feature extraction capability; the HS-FPN (High-level Screening Feature Pyramid Network) in the neck network, which enhances multi-scale feature fusion for small-target detection; and the ESCH (Efficient Shared Convolution Head module), which optimizes detection heads through shared convolution layers to reduce parameters. Experimental results show that Cu-YOLO achieves an [email protected]:0.95 (mean Average Precision at Intersection over Union thresholds of 0.5 to 0.95) of 69.5%, a 1.4% improvement over YOLOv8n, while reducing parameters by 58.9% and GFLOPs (Giga Floating-point Operations Per Second) by 45.7%. These results demonstrate its effectiveness in detecting trace copper particles in lithium-ion battery carbon powder.