Scholarlink

<p>Ensuring accurate quality prediction under data scarcity remains a critical challenge in manufacturing systems, where frequent changes in materials and processes limit the availability of labeled data and degrade model performance. This study proposes a two-stage transfer learning framework designed to maintain the predictive accuracy in copper foil manufacturing, where gel-time prediction is a key quality indicator. The framework combines adaptive instance reweighting with a threshold-based maintenance strategy that triggers model updates once the prediction accuracy falls below the operational requirements. Using industrial datasets collected from 2018 to 2022, the proposed approach was benchmarked against conventional retraining and feature-based transfer-learning methods. The results demonstrate that the framework restores accuracy from 65% to 90% and from 70% to 85% within a ±5-second tolerance, outperforming the baselines under domain shift conditions. Furthermore, the scenario analysis highlights tangible industrial benefits, including reduced scrap rates and improved process stability. This study presents a deployable and cost-effective solution for robust quality prediction, providing practical decision-support insights for manufacturing systems that face data scarcity issues.</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

A two-stage transfer learning framework for quality prediction under data scarcity: an industrial case study in copper foil manufacturing

  • 期刊论文

  • 发表日期：

    2026年1月23日

查看全文
  • Chien-Chih Wang,
  • Chun-Kai Wei

摘要

Ensuring accurate quality prediction under data scarcity remains a critical challenge in manufacturing systems, where frequent changes in materials and processes limit the availability of labeled data and degrade model performance. This study proposes a two-stage transfer learning framework designed to maintain the predictive accuracy in copper foil manufacturing, where gel-time prediction is a key quality indicator. The framework combines adaptive instance reweighting with a threshold-based maintenance strategy that triggers model updates once the prediction accuracy falls below the operational requirements. Using industrial datasets collected from 2018 to 2022, the proposed approach was benchmarked against conventional retraining and feature-based transfer-learning methods. The results demonstrate that the framework restores accuracy from 65% to 90% and from 70% to 85% within a ±5-second tolerance, outperforming the baselines under domain shift conditions. Furthermore, the scenario analysis highlights tangible industrial benefits, including reduced scrap rates and improved process stability. This study presents a deployable and cost-effective solution for robust quality prediction, providing practical decision-support insights for manufacturing systems that face data scarcity issues.