Scholarlink

<p>The increasing complexity of Printed Circuit Board (PCB) designs and the high reliability demands in modern electronics have made defect detection a critical step in the production chain. This study presents a fully functional, low-cost, and modular system for automated soldering defect detection and classification, combining a custom-built motorized XY scanning platform, synchronized optical and thermal imaging, and deep learning models. Multiple object detection architectures were evaluated, including YOLO (v8–v11), RetinaNet, RT-DETR, and Faster R-CNN with ResNet backbones. In addition to testing standard models, a custom-modified version of YOLOv11n was developed, integrating attention and multiscale enhancements (HCAM and FSPPCSP), achieving 91.3% mAP@0.5 at 75 FPS—making it ideal for real-time industrial use. The highest recall (99.3%) and mAP (93.7%) were achieved by Faster R-CNN with ResNet101. Thermal imaging was employed as a complementary inspection modality, enabling the detection of latent defects such as cold joints and thermal stress, which are not visible through RGB alone. This work demonstrates that high-performance, scalable, and cost-effective PCB inspection is achievable using off-the-shelf components, open-source tools, and carefully optimized deep learning models—providing a practical solution for agile and small-scale electronics manufacturing environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated detection and classification of soldering defects in printed circuit boards using deep learning and optical and thermal imaging

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Tiago Diogo,
  • António Ramos,
  • Filipe Pereira,
  • Nuno Araújo,
  • António Lopes

摘要

The increasing complexity of Printed Circuit Board (PCB) designs and the high reliability demands in modern electronics have made defect detection a critical step in the production chain. This study presents a fully functional, low-cost, and modular system for automated soldering defect detection and classification, combining a custom-built motorized XY scanning platform, synchronized optical and thermal imaging, and deep learning models. Multiple object detection architectures were evaluated, including YOLO (v8–v11), RetinaNet, RT-DETR, and Faster R-CNN with ResNet backbones. In addition to testing standard models, a custom-modified version of YOLOv11n was developed, integrating attention and multiscale enhancements (HCAM and FSPPCSP), achieving 91.3% mAP@0.5 at 75 FPS—making it ideal for real-time industrial use. The highest recall (99.3%) and mAP (93.7%) were achieved by Faster R-CNN with ResNet101. Thermal imaging was employed as a complementary inspection modality, enabling the detection of latent defects such as cold joints and thermal stress, which are not visible through RGB alone. This work demonstrates that high-performance, scalable, and cost-effective PCB inspection is achievable using off-the-shelf components, open-source tools, and carefully optimized deep learning models—providing a practical solution for agile and small-scale electronics manufacturing environments.