Scholarlink

<p>Weld discontinuities are major sources of structural failure because they promote crack initiation, accelerate fatigue propagation, and reduce the reliability of welded components. This study presents a scalable deep learning framework for automated weld quality assessment using three You Only Look Once version 11 (YOLOv11) variants (Model-N, Model-M, and Model-L) to detect and classify Bad Weld, Good Weld, and Defect regions. Under identical training conditions, architectural scaling consistently improved optimization convergence, feature representation, and localization accuracy. Validation performance increased from mAP@0.5 = 0.623 and mAP@0.5:0.95 = 0.385 for Model-N to 0.722 and 0.471 for Model-L, with the largest gains achieved for the challenging Defect class. Evaluation on an independent test dataset demonstrated that Model-M achieved the best generalization, attaining Precision = 0.704, Recall = 0.665, mAP@0.5 = 0.732, and mAP@0.5:0.95 = 0.462, while Model-L delivered the highest localization precision and the most reliable qualitative detection of subtle and clustered defects. Inference times ranged from 12.5&#xa0;ms/image (Model-N) to 41.3&#xa0;ms/image (Model-L), with Model-M providing the optimal balance between detection accuracy and computational efficiency (36.1&#xa0;ms/image) for real-time industrial deployment. The proposed framework enables accurate weld quality assessment, early defect localization, and reliable structural failure prevention, providing an effective foundation for intelligent manufacturing, predictive maintenance, and automated quality assurance.</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

Scalable Deep Learning for Intelligent Weld Quality Assessment Toward Structural Failure Prevention

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Khier Sabri

摘要

Weld discontinuities are major sources of structural failure because they promote crack initiation, accelerate fatigue propagation, and reduce the reliability of welded components. This study presents a scalable deep learning framework for automated weld quality assessment using three You Only Look Once version 11 (YOLOv11) variants (Model-N, Model-M, and Model-L) to detect and classify Bad Weld, Good Weld, and Defect regions. Under identical training conditions, architectural scaling consistently improved optimization convergence, feature representation, and localization accuracy. Validation performance increased from mAP@0.5 = 0.623 and mAP@0.5:0.95 = 0.385 for Model-N to 0.722 and 0.471 for Model-L, with the largest gains achieved for the challenging Defect class. Evaluation on an independent test dataset demonstrated that Model-M achieved the best generalization, attaining Precision = 0.704, Recall = 0.665, mAP@0.5 = 0.732, and mAP@0.5:0.95 = 0.462, while Model-L delivered the highest localization precision and the most reliable qualitative detection of subtle and clustered defects. Inference times ranged from 12.5 ms/image (Model-N) to 41.3 ms/image (Model-L), with Model-M providing the optimal balance between detection accuracy and computational efficiency (36.1 ms/image) for real-time industrial deployment. The proposed framework enables accurate weld quality assessment, early defect localization, and reliable structural failure prevention, providing an effective foundation for intelligent manufacturing, predictive maintenance, and automated quality assurance.