Scholarlink

<p>Defects such as over-extrusion, stringing, and warping commonly occur in Fused Filament Fabrication (FFF) 3D printing, leading to material waste and substantial time loss when print jobs span several hours. Although supervised computer vision systems can detect specific categories of FFF flaws, they typically require large annotated datasets for each defect type and degrade when the printer, material, or lighting changes, motivating the search for monitoring approaches that rely less on task-specific retraining. This paper presents <i>FA3D</i>, a fault-aware monitoring framework that uses pre-trained Vision-Language Models (VLMs) within an agentic perceive-reason-act loop for zero-shot defect detection on FFF printers. In this work, “fault-aware” refers to the system’s capacity to identify printing faults, evaluate their impact, and trigger appropriate intervention through a VLM-based perception module, a rule-based stop/continue mechanism, and human-in-the-loop decision support. We conduct a comparative zero-shot evaluation of four state-of-the-art VLMs, including Gemini 2.5 Flash Image, GPT-4o, Qwen 2.5-VL-32B, and GLM-4.5V, on a balanced dataset of 140 print images, and analyze precision, recall, F1-score, and F2-score, where the F2-score is emphasized to reflect the higher cost of missed defects. Gemini 2.5 Flash Image achieves the best overall performance with an F2-score of 0.901 and a recall of 97.1%, while GPT-4o reaches 100% recall at the cost of frequent false alarms, and GLM-4.5V reaches perfect precision but misses 63.8% of defects. The study positions VLM-based zero-shot monitoring as a practical complement to supervised detectors for low-volume and rapidly changing print scenarios, while clearly delineating its current limitations in terms of dataset breadth, prompt sensitivity, and dependence on external APIs.</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

FA3D: a fault-aware monitoring framework for 3D printing using vision-language models (VLMs)

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Fatimah Alali,
  • Saad Ezzini,
  • Jumana Alqudah,
  • Brahim Brahmi

摘要

Defects such as over-extrusion, stringing, and warping commonly occur in Fused Filament Fabrication (FFF) 3D printing, leading to material waste and substantial time loss when print jobs span several hours. Although supervised computer vision systems can detect specific categories of FFF flaws, they typically require large annotated datasets for each defect type and degrade when the printer, material, or lighting changes, motivating the search for monitoring approaches that rely less on task-specific retraining. This paper presents FA3D, a fault-aware monitoring framework that uses pre-trained Vision-Language Models (VLMs) within an agentic perceive-reason-act loop for zero-shot defect detection on FFF printers. In this work, “fault-aware” refers to the system’s capacity to identify printing faults, evaluate their impact, and trigger appropriate intervention through a VLM-based perception module, a rule-based stop/continue mechanism, and human-in-the-loop decision support. We conduct a comparative zero-shot evaluation of four state-of-the-art VLMs, including Gemini 2.5 Flash Image, GPT-4o, Qwen 2.5-VL-32B, and GLM-4.5V, on a balanced dataset of 140 print images, and analyze precision, recall, F1-score, and F2-score, where the F2-score is emphasized to reflect the higher cost of missed defects. Gemini 2.5 Flash Image achieves the best overall performance with an F2-score of 0.901 and a recall of 97.1%, while GPT-4o reaches 100% recall at the cost of frequent false alarms, and GLM-4.5V reaches perfect precision but misses 63.8% of defects. The study positions VLM-based zero-shot monitoring as a practical complement to supervised detectors for low-volume and rapidly changing print scenarios, while clearly delineating its current limitations in terms of dataset breadth, prompt sensitivity, and dependence on external APIs.