Scholarlink

<p>Additive manufacturing (AM) faces persistent challenges in mitigating process-induced errors, which often lead to part failure, material waste, and operational inefficiencies. Current solutions struggle with real-time adaptability, generalization across hardware configurations, and balancing rapid corrections with system stability. This study presents a deep learning framework that corrects errors in real time during additive manufacturing by combining reactive adjustments with global consensus-driven corrections. A hybrid dataset of 1.45 million images—combining open-source (Brion and Pattinson in Apollo - University Cambridge Repository, 2022, <a href="https://doi.org/10.17863/CAM.84082">https://doi.org/10.17863/CAM.84082</a>) and proprietary data from diverse printers, materials, and failure modes—trains a shared backbone attention network to simultaneously predict critical metrics such as flow rate, speed of printing, Z-axis displacement, and extruder temperature directly from individual real-time images captured during the printing process with an overall accuracy of 85.3%. The system employs mode thresholding to filter transient noise and proportional control to apply context-aware adjustments, achieving 89% reduction in print failures across complex geometries like the Eiffel Tower and multi-material setups. Key innovations include a two-stage correction workflow enabling quasi-real-time inference (0.3&#xa0;s per prediction) with hardware-agnostic deployment on electronics and hybrid data integration to address bias in lab-controlled datasets, improving generalization to real-world print farm conditions. The framework bridges the gap between high-frequency reactive systems and delayed global corrections, offering a scalable solution for enhancing AM reliability.</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

Intelligent real-time error correction in additive manufacturing via context-aware deep learning

  • 期刊论文

  • 发表日期：

    2025年6月21日

查看全文
  • Prateek Mehta,
  • Vivek Kamat,
  • Daniel Ewing,
  • Shekhar Bhansali

摘要

Additive manufacturing (AM) faces persistent challenges in mitigating process-induced errors, which often lead to part failure, material waste, and operational inefficiencies. Current solutions struggle with real-time adaptability, generalization across hardware configurations, and balancing rapid corrections with system stability. This study presents a deep learning framework that corrects errors in real time during additive manufacturing by combining reactive adjustments with global consensus-driven corrections. A hybrid dataset of 1.45 million images—combining open-source (Brion and Pattinson in Apollo - University Cambridge Repository, 2022, https://doi.org/10.17863/CAM.84082) and proprietary data from diverse printers, materials, and failure modes—trains a shared backbone attention network to simultaneously predict critical metrics such as flow rate, speed of printing, Z-axis displacement, and extruder temperature directly from individual real-time images captured during the printing process with an overall accuracy of 85.3%. The system employs mode thresholding to filter transient noise and proportional control to apply context-aware adjustments, achieving 89% reduction in print failures across complex geometries like the Eiffel Tower and multi-material setups. Key innovations include a two-stage correction workflow enabling quasi-real-time inference (0.3 s per prediction) with hardware-agnostic deployment on electronics and hybrid data integration to address bias in lab-controlled datasets, improving generalization to real-world print farm conditions. The framework bridges the gap between high-frequency reactive systems and delayed global corrections, offering a scalable solution for enhancing AM reliability.