Scholarlink

<p>Real-time quality assurance in Additive Manufacturing remains challenging, as most monitoring frameworks do not directly account for process dynamics on the printer. This work presents the Additive Edge Multi-Agent (AEMA) framework, an acoustic monitoring system that runs entirely on an embedded single-board computer connected directly to a fused deposition modeling printer. AEMA separates the task into three parts: Specialist Agents that analyze short, phase-specific windows of spectral audio features, a Transformer-based Historian that models the full print history, and an Orchestrator that combines their outputs using an attention mechanism to predict the current print condition. Experiments on six print classes (normal print, over-extrusion, under-extrusion, no extrusion, poor layer adhesion, and toolhead failure) show an improvement in accuracy from 87.76% for the baseline multi-agent model to 98.37% for the final AEMA architecture, with an average inference time of about 105&#xa0;ms per 1-s segment on an ODROID-XU4Q board. Findings indicate that a multi-agent, edge-deployed design can support practical real-time defect monitoring in AM.</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

Towards scalable intelligence at the edge: a multi-agent framework for quality assurance in additive manufacturing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月26日

查看全文
  • Michael Olowe,
  • Joao Paulo Jacomini Prioli,
  • Santosh Kumar Parupelli,
  • Salil Desai

摘要

Real-time quality assurance in Additive Manufacturing remains challenging, as most monitoring frameworks do not directly account for process dynamics on the printer. This work presents the Additive Edge Multi-Agent (AEMA) framework, an acoustic monitoring system that runs entirely on an embedded single-board computer connected directly to a fused deposition modeling printer. AEMA separates the task into three parts: Specialist Agents that analyze short, phase-specific windows of spectral audio features, a Transformer-based Historian that models the full print history, and an Orchestrator that combines their outputs using an attention mechanism to predict the current print condition. Experiments on six print classes (normal print, over-extrusion, under-extrusion, no extrusion, poor layer adhesion, and toolhead failure) show an improvement in accuracy from 87.76% for the baseline multi-agent model to 98.37% for the final AEMA architecture, with an average inference time of about 105 ms per 1-s segment on an ODROID-XU4Q board. Findings indicate that a multi-agent, edge-deployed design can support practical real-time defect monitoring in AM.