Scholarlink

Additive manufacturing enables the creation of complex geometries for various applications, such as dental, medical, prototyping, and aerospace components. Despite its advantages, the printing process can encounter errors due to its complexity and numerous influencing factors, necessitating real-time anomaly detection and classification. This work introduces a non-destructive method for defect detection by monitoring layer-wise image data of the L-PBF process using a YOLOv11 object detection model to ensure component quality. A novel approach combines multiple grayscale images from a single print sequence into a 3-channel image, incorporating information from the build area in the powder bed to enhance defect classification reliability. The determination of various defect classes is also a key aspect of this work, crucial for informing subsequent interventions and decisions. Tested on a dataset generated with a Trumpf TruPrint 1000 machine, this method achieved visually promising results.

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-Based Defect Detection in Laser Powder Bed Fusion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Cindy Buhl,
  • Faiza Waheed,
  • Ulrich Göhner

摘要

Additive manufacturing enables the creation of complex geometries for various applications, such as dental, medical, prototyping, and aerospace components. Despite its advantages, the printing process can encounter errors due to its complexity and numerous influencing factors, necessitating real-time anomaly detection and classification. This work introduces a non-destructive method for defect detection by monitoring layer-wise image data of the L-PBF process using a YOLOv11 object detection model to ensure component quality. A novel approach combines multiple grayscale images from a single print sequence into a 3-channel image, incorporating information from the build area in the powder bed to enhance defect classification reliability. The determination of various defect classes is also a key aspect of this work, crucial for informing subsequent interventions and decisions. Tested on a dataset generated with a Trumpf TruPrint 1000 machine, this method achieved visually promising results.