Scholarlink

<p>Ceramic firing—a cornerstone skill in materials science curricula—presents a persistent pedagogical challenge: cycles often exceed eight hours, creating substantial lag between student actions and feedback. We developed an LSTM-Attention early warning framework to flag likely defects mid-process. Our dataset comprises 1,000 firing cycles from a university kiln, each recorded across 144 time steps with temperature, pressure, and gas flow monitored. Student operators produced ~ 40% defects, well above industrial norms. We evaluated prediction at four checkpoints: 25%, 50%, 75%, and 100% completion. At 50%—leaving four hours for intervention—accuracy reaches 79.3%, outperforming Random Forest by 4.9% points and MLP by 6.9 points (AUC-ROC: 0.852). Temporal modeling proves especially valuable under constrained observation. Attention weights cluster meaningfully around the 573&#xa0;°C quartz transition zone (~ 29% combined heating and cooling passages) and the 1000–1200&#xa0;°C sintering phase (~ 54%), aligning with ceramic processing theory and enabling instructors to pinpoint problematic intervals. These findings suggest data-driven prediction could support a shift from retrospective grading toward real-time guided instruction, though single-facility data and threshold recalibration requirements for industrial deployment remain limitations.</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

LSTM-based early warning system for ceramic firing defects: a time-series approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月14日

查看全文
  • Yixin Zhao,
  • Yun Zhang

摘要

Ceramic firing—a cornerstone skill in materials science curricula—presents a persistent pedagogical challenge: cycles often exceed eight hours, creating substantial lag between student actions and feedback. We developed an LSTM-Attention early warning framework to flag likely defects mid-process. Our dataset comprises 1,000 firing cycles from a university kiln, each recorded across 144 time steps with temperature, pressure, and gas flow monitored. Student operators produced ~ 40% defects, well above industrial norms. We evaluated prediction at four checkpoints: 25%, 50%, 75%, and 100% completion. At 50%—leaving four hours for intervention—accuracy reaches 79.3%, outperforming Random Forest by 4.9% points and MLP by 6.9 points (AUC-ROC: 0.852). Temporal modeling proves especially valuable under constrained observation. Attention weights cluster meaningfully around the 573 °C quartz transition zone (~ 29% combined heating and cooling passages) and the 1000–1200 °C sintering phase (~ 54%), aligning with ceramic processing theory and enabling instructors to pinpoint problematic intervals. These findings suggest data-driven prediction could support a shift from retrospective grading toward real-time guided instruction, though single-facility data and threshold recalibration requirements for industrial deployment remain limitations.