Scholarlink

<p>Real-time defect detection is vital in Fused Deposition Modeling (FDM) for ensuring the quality of 3D-printed objects and enabling autonomous control systems that can reduce material waste, improve reliability, and minimize human intervention. This work presents the first applied demonstration of a real-time dual-camera defect-detection system on a low-cost embedded platform for multi-angle defect detection during active 3D printing. A custom dataset of four common extrusion defects: stringing, spaghetti, under-extrusion, and over-extrusion, was developed and used to fine-tune lightweight YOLOv11n and YOLOv12n models through transfer learning. The models were further optimized via multi-objective hyperparameter tuning using the Optuna framework, adapted for low-power embedded inference to balance accuracy, localization precision, and speed with dual cameras for multi-angle monitoring. The results demonstrated that the optimized models jointly improve detection accuracy and localization precision while maintaining real-time performance. The system reliably identifies all four extrusion-related defects and validated a perception-to-action pipeline linking camera inference to printer-firmware response, confirming closed-loop feasibility. These findings establish a practical foundation for AI-augmented additive manufacturing and paves the way toward autonomous, closed-loop 3D and 4D printing systems.</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

Real-time defect monitoring in material extrusion 3D printing using optimized YOLO models

  • 期刊论文

  • 发表日期：

    2025年11月12日

查看全文
  • Abdul Rahman Sani,
  • Abbas Z. Kouzani,
  • Ali Zolfagharian

摘要

Real-time defect detection is vital in Fused Deposition Modeling (FDM) for ensuring the quality of 3D-printed objects and enabling autonomous control systems that can reduce material waste, improve reliability, and minimize human intervention. This work presents the first applied demonstration of a real-time dual-camera defect-detection system on a low-cost embedded platform for multi-angle defect detection during active 3D printing. A custom dataset of four common extrusion defects: stringing, spaghetti, under-extrusion, and over-extrusion, was developed and used to fine-tune lightweight YOLOv11n and YOLOv12n models through transfer learning. The models were further optimized via multi-objective hyperparameter tuning using the Optuna framework, adapted for low-power embedded inference to balance accuracy, localization precision, and speed with dual cameras for multi-angle monitoring. The results demonstrated that the optimized models jointly improve detection accuracy and localization precision while maintaining real-time performance. The system reliably identifies all four extrusion-related defects and validated a perception-to-action pipeline linking camera inference to printer-firmware response, confirming closed-loop feasibility. These findings establish a practical foundation for AI-augmented additive manufacturing and paves the way toward autonomous, closed-loop 3D and 4D printing systems.