Scholarlink

<p>Fused Deposition Modeling (FDM) is susceptible to in-process defects, including under-extrusion holes, layer misalignment, and material voids. If these defects are not detected during the process, they can critically undermine the strength and reliability of the part. Deep learning approaches provide a robust and dependable solution for in-process monitoring, correction, and decision-making. This paper introduces two novel end-to-end pipelines for the detection, segmentation, and classification of defects on the Creality Ender-3 V3 KE printer. The first approach employs a CNN-based methodology utilizing YOLOv11n, DeepLabV3 + , and EfficientNet, while the second approach is a transformer-based hybrid architecture leveraging YOLOv11n, SegFormer-B2, and ConvNeXt V2. A Unity 3D graphical interface is utilized for creating informative twins and facilitating real-time decision-making. This work presents the first systematic benchmark comparing CNN and transformer-based multi-stage pipelines for in-situ FDM defect detection, offering actionable guidance for selecting architectures based on application-specific considerations of recall, precision, and inference speed.</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

CNN-based and transformer-based informative twin for defect detection in fused deposition modeling

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Omer Farooq,
  • Shahbaz Khan,
  • Kunwar Faraz Khan,
  • M. Saqib Nazir,
  • Shahid Ikramullah,
  • Mian Ashfaq Ali

摘要

Fused Deposition Modeling (FDM) is susceptible to in-process defects, including under-extrusion holes, layer misalignment, and material voids. If these defects are not detected during the process, they can critically undermine the strength and reliability of the part. Deep learning approaches provide a robust and dependable solution for in-process monitoring, correction, and decision-making. This paper introduces two novel end-to-end pipelines for the detection, segmentation, and classification of defects on the Creality Ender-3 V3 KE printer. The first approach employs a CNN-based methodology utilizing YOLOv11n, DeepLabV3 + , and EfficientNet, while the second approach is a transformer-based hybrid architecture leveraging YOLOv11n, SegFormer-B2, and ConvNeXt V2. A Unity 3D graphical interface is utilized for creating informative twins and facilitating real-time decision-making. This work presents the first systematic benchmark comparing CNN and transformer-based multi-stage pipelines for in-situ FDM defect detection, offering actionable guidance for selecting architectures based on application-specific considerations of recall, precision, and inference speed.