Scholarlink

<p>Additive manufacturing (AM) offers design flexibility and shorter lead times in manufacturing newer products. The layer by layer manufacturing approach often introduces defects that can deteroriate part performance, which require reliable inspection to detect quality. Manual inspection often followed in industries are slow and error prone, which thrusts for the need to automate inspection. This work proposes a modular deep learning framework for objection detection and defect analysis components manufactured by fused deposition modelling. A real-world dataset of 950 images from additively manufactured elevator maintainance components, is taken to test the model and evaluate performance.YOLOv12 and Faster R-CNN are used for object detection, CNN for binary defect classification and MobileNet to classify both defect severity(Levels 0–3) and defect type. The model successfully predicted, whether an object was defective, determined its defect severity level, and correctly labeled the defect type with greater accuracy. YOLOv12 yielded slightly better results in comparison to Faster R-CNN for object detection with a precision of around 93.6% and recall of 95.6%. The model accurately identified whether an object was defective, determined its defect severity level, and correctly labeled the defect type with high precision.</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

Deep learning for object recognition and defect analysis in additive manufacturing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月15日

查看全文
  • Chandrbhan Thakur,
  • Santosh Kr. Mishra,
  • S. Silva Sajin Jose,
  • Purushottam Kumar Singh,
  • Ram Krishna Upadhyay

摘要

Additive manufacturing (AM) offers design flexibility and shorter lead times in manufacturing newer products. The layer by layer manufacturing approach often introduces defects that can deteroriate part performance, which require reliable inspection to detect quality. Manual inspection often followed in industries are slow and error prone, which thrusts for the need to automate inspection. This work proposes a modular deep learning framework for objection detection and defect analysis components manufactured by fused deposition modelling. A real-world dataset of 950 images from additively manufactured elevator maintainance components, is taken to test the model and evaluate performance.YOLOv12 and Faster R-CNN are used for object detection, CNN for binary defect classification and MobileNet to classify both defect severity(Levels 0–3) and defect type. The model successfully predicted, whether an object was defective, determined its defect severity level, and correctly labeled the defect type with greater accuracy. YOLOv12 yielded slightly better results in comparison to Faster R-CNN for object detection with a precision of around 93.6% and recall of 95.6%. The model accurately identified whether an object was defective, determined its defect severity level, and correctly labeled the defect type with high precision.