| Battery Pouch Cell Defect Detection System Using YOLOv11 Lightweight Model `Pouch Cell Defect Detection using Lightweight YOLOv11` |
|
Dong-Ju Kang1, Seung-Yong Son1, Doo-Hyun Choi2 |
1School of Electronics Engineering, Kyungpook National University Daegu 41566, Republic of Korea
2School of Electronics Engineering, Kyungpook National University Daegu 41566, Republic of Korea |
Correspondence:
Doo-Hyun Choi,
Email: dhc@ee.knu.ac.kr
|
|
Received: 28 July 2025 • Accepted: 7 October 2025 |
| Abstract |
|
This study introduces a real-time vision-based defect detection system for lithium-ion battery pouch cells, leveraging a lightweight YOLOv11n architecture. A dataset comprising 15,312 high-resolution images (1024×1024 pixels) was constructed from a commercial production line and categorized into five surface defect classes: pass, leakage, pinhole, swelling, and scratch. To mitigate the dataset's inherent class imbalance—with leakage representing only 8.8% of samples—comprehensive data augmentation strategies were employed, including mosaic augmentation (0.8 probability), MixUp (0.2), and random erasing (0.35) to enhance class distribution equilibrium and strengthen model generalization capabilities. The proposed Lightweight YOLOv11n model was developed through systematic optimizations: global channel-width reduction (0.75 ratio), C2K depthwise-separable residual blocks, SPPF-Lite spatial pyramid pooling, C2PSA attention modules, unified 64-channel detection heads, removal of stride-32 detection branch, magnitude-based weight pruning (threshold: 1×10⁻²), and mixed-precision optimization (FP32 to FP16). Comparative experiments with YOLOv11n and YOLOv12n were conducted on an NVIDIA A100 GPU. Lightweight YOLOv11n attained an F1-score of 0.9791, with precision and recall of 0.9781 and 0.9802 respectively, achieving mAP@0.5 of 0.9883 and mAP@50-95 of 0.8456. The model maintained real-time inference capability with a process duration of 9.43 ms (106.1 theoretical FPS), representing a 48.4% reduction compared to YOLOv12n (18.26 ms) while sacrificing only 0.47 percentage points in F1-score. These results demonstrate that Lightweight YOLOv11n provides an optimal trade-off between detection performance and computational efficiency, validating its applicability for inline quality inspection on high-speed manufacturing lines. |
| Keywords:
Object detection, You Only Look Once (YOLO), Lithium-ion battery, Deep learning, Real-time Inspection, Defect detection, Pouch cell |
|
PDF Links
Full text via DOI
Download Citation
