主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (6): 171-186.doi: 10.16381/j.cnki.issn1003-207x.2024.1370cstr: 32146.14.j.cnki.issn1003-207x.2024.1370

• • 上一篇    下一篇

基于主动学习随机Kriging模型的在线质量设计方法

杜晨1, 林成龙1,2, 石雨葳1, 马义中1()   

  1. 1.南京理工大学经济管理学院,江苏 南京 210094
    2.安徽工业大学管理科学与工程学院,安徽 马鞍山 243032
  • 收稿日期:2024-08-13 修回日期:2025-01-18 出版日期:2026-06-25 发布日期:2026-05-22
  • 通讯作者: 马义中 E-mail:yzma-2004@163.com
  • 基金资助:
    国家自然科学基金重点项目(71931006);国家自然科学基金面上项目(71871119);国家自然科学基金面上项目(72471119);江苏省研究生科研与实践创新计划项目(KYCX23_0533)

An Online Quality Design Method Using Active Learning-Based Stochastic Kriging Model

Chen Du1, Chenglong Lin1,2, Yuwei Shi1, Yizhong Ma1()   

  1. 1.School of Economics and Management,Nanjing University of Science and Technology,Nanjing 210094,China
    2.School of Management Science and Engineering,Anhui University of Technology,Maanshan 243032,China
  • Received:2024-08-13 Revised:2025-01-18 Online:2026-06-25 Published:2026-05-22
  • Contact: Yizhong Ma E-mail:yzma-2004@163.com

摘要:

针对未知随机噪声导致的质量特性波动较大的问题,提出了一种基于主动学习随机Kriging模型的在线质量设计方法。在线质量设计利用主动学习随机Kriging模型的主动学习能力,构建噪声响应下的期望改进准则来挑选新样本点,并完成模型的在线更新以进一步提升模型预测精度。实现过程中,首先在初始试验设计基础上完成随机Kriging模型构建,再利用主动学习方法不断更新随机Kriging模型,最后根据构建的期望二次质量损失模型进行质量设计。数值算例及仿真案例结果表明:相比采用主动学习常Kriging建模和离线随机Kriging建模的方法,所提方法能获得更加精确的响应曲面模型;且结合期望二次质量损失函数进行参数优化后,所提方法在同样的仿真计算资源下可有效过滤噪声,获得更为稳健的参数设计值。同时,从经济性视角提出了一种质量成本评价方法,结合历史生产数据可对获得的参数设计方案进行评价,为管理者提供决策依据。

关键词: 稳健参数设计, 随机Kriging模型, 主动学习, 在线质量设计, 质量成本

Abstract:

To address the low robustness of quality characteristics caused by unknown random noise, an online quality design method based on the active learning stochastic Kriging model is proposed. The online quality design method uses the active learning characteristic to select new sample points through the expected improvement criterion formulated under noisy response, then updates the model to improve prediction accuracy. The stochastic Kriging model is first constructed based on the initial design in the implementation process, then updated using active learning method, and finally, quality design is achieved through the expected quality loss function model. Numerical examples and simulation results demonstrated that the proposed method can obtain more accurate response surface model compared with methods using the ordinary Kriging model and offline stochastic Kriging. The proposed method can filter the noise and get more robust parameter design solutions with the same computation resources by utilizing the expected quality loss function model. Lastly, a quality cost evaluation method is proposed from an economic perspective, which evaluates the parameters of the design solutions obtained based on historical production data, providing managers with valuable information for decision-making.

Key words: robust parameter design, stochastic Kriging model, active learning, online quality design, quality cost

中图分类号: