Scholarlink

<p>Cold-start quality prediction remains a critical challenge in multi-machine manufacturing systems, where newly commissioned equipment suffers from severe data scarcity and pronounced non-independent and identically distributed (non-i.i.d.) heterogeneity across nominally similar machines. In such environments, naïve data pooling or black-box transfer learning often leads to negative transfers and low practitioner trust, limiting deployment in real production settings. To address this challenge, this study proposes a weak-consistency data fusion framework with human-in-the-loop learning for interpretable cold-start quality prediction in heterogeneous manufacturing systems. Rather than enforcing full distributional alignment, the proposed framework enables safe cross-machine data augmentation by combining variability-based machine screening, interquartile range (IQR) overlap filtering, categorical process validation, and structured expert intervention. A physics-informed hierarchical stacking ensemble was further integrated with a three-cycle human-in-the-loop refinement process guided by SHAP and permutation importance, ensuring transparency and operational relevance. The framework was validated using real production data from a copper-clad laminate manufacturing plant with seven parallel impregnation lines. Under a severe cold-start scenario with only 384 initial samples from a newly commissioned machine, the proposed approach safely augments the training dataset by 7.6 times using data screened from a compatible legacy machine. Experimental results demonstrate that the proposed method achieves an R² of 85.7%, an MAPE of 2.14%, and an 87% prediction achievement rate within the plant’s ± 5&#xa0;s quality tolerance, representing a statistically significant improvement of 17% points over the cold-start baseline (<i>p</i> &lt; 0.01). The results confirm that weak-consistency data fusion, combined with structured human oversight, provides a practical and interpretable solution for cold-start quality prediction in multi-machine manufacturing systems, supporting reliable deployment in high-mix, low-volume production environments.</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

Cold-start quality prediction in multi-machine manufacturing systems via weak-consistency data fusion with human-in-the-loop learning

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Chien-Chih Wang,
  • Hong-Gu Shen,
  • Chun-Hua Chien

摘要

Cold-start quality prediction remains a critical challenge in multi-machine manufacturing systems, where newly commissioned equipment suffers from severe data scarcity and pronounced non-independent and identically distributed (non-i.i.d.) heterogeneity across nominally similar machines. In such environments, naïve data pooling or black-box transfer learning often leads to negative transfers and low practitioner trust, limiting deployment in real production settings. To address this challenge, this study proposes a weak-consistency data fusion framework with human-in-the-loop learning for interpretable cold-start quality prediction in heterogeneous manufacturing systems. Rather than enforcing full distributional alignment, the proposed framework enables safe cross-machine data augmentation by combining variability-based machine screening, interquartile range (IQR) overlap filtering, categorical process validation, and structured expert intervention. A physics-informed hierarchical stacking ensemble was further integrated with a three-cycle human-in-the-loop refinement process guided by SHAP and permutation importance, ensuring transparency and operational relevance. The framework was validated using real production data from a copper-clad laminate manufacturing plant with seven parallel impregnation lines. Under a severe cold-start scenario with only 384 initial samples from a newly commissioned machine, the proposed approach safely augments the training dataset by 7.6 times using data screened from a compatible legacy machine. Experimental results demonstrate that the proposed method achieves an R² of 85.7%, an MAPE of 2.14%, and an 87% prediction achievement rate within the plant’s ± 5 s quality tolerance, representing a statistically significant improvement of 17% points over the cold-start baseline (p < 0.01). The results confirm that weak-consistency data fusion, combined with structured human oversight, provides a practical and interpretable solution for cold-start quality prediction in multi-machine manufacturing systems, supporting reliable deployment in high-mix, low-volume production environments.