• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2018, Vol. 54 ›› Issue (16): 62-69.doi: 10.3901/JME.2018.16.062

• 特邀专栏:制造物联网与智能制造服务技术 • 上一篇    下一篇

基于函数型数据分析的半导体生产过程监控

黎敏1, 谢玄2, 陈泽2, 杨孟瑶2, 杨德斌2, 蒋靖3   

  1. 1. 北京科技大学钢铁共性技术协同创新中心 北京 100083;
    2. 北京科技大学机械工程学院 北京 100083;
    3. 北京中油瑞飞信息技术有限责任公司 北京 100007
  • 收稿日期:2017-10-12 修回日期:2018-05-03 出版日期:2018-08-20 发布日期:2018-08-20
  • 通讯作者: 黎敏(通信作者),女,1980年出生,博士,教授,博士研究生导师。主要研究方向为产品质量建模与监控、信号处理与模式识别。E-mail:limin@ustb.edu.cn
  • 基金资助:
    省部共建耐火材料与冶金国家重点实验室开放基金资助项目(G201704)。

Monitoring of Semiconductor Manufacturing Process Based on Functional Data Analysis

LI Min1, XIE Xuan2, CHEN Ze2, YANG Mengyao2, YANG Debin2, JIANG Jing3   

  1. 1. Collaborative Innovation Center of Steel Technology, University of Science and Technology Beijing, Beijing 100083;
    2. School of Mechanical Engineering, University of Science and Technology Beijing, Beijing 100083;
    3. Beijing Richfit Information Technology Co., Ltd., Beijing 100007
  • Received:2017-10-12 Revised:2018-05-03 Online:2018-08-20 Published:2018-08-20

摘要: 提出一种基于函数型数据分析的非平稳生产状态下的多批次生产过程监控方法。引入函数型数据分析方法将三维数据中的每个变量沿时间方向进行函数拟合,从而将三维离散数据矩阵转化为二维函数矩阵。对各个变量曲线求取二阶导数,消除非平稳生产状态所导致的均值波动现象,增加建模的准确性。在此基础上,利用函数型主成分分析方法对各个变量的导数曲线进行特征提取,并使用基于权重的主成分融合方法,获得函数型主成分矩阵。利用支持向量数据描述方法对函数型主成分矩阵进行监控,并将其应用于工业半导体生产过程的监控中。结果表明,基于函数型数据分析的监控方法与传统方法相比,具有更低的漏检率,验证了新方法的有效性。

关键词: 非平稳生产过程, 过程监控, 函数型数据分析, 函数型主成分分析

Abstract: A monitoring method based on functional data analysis(FDA) is proposed to improve the quality of batch process in non-stationary condition. Each variable of the three-dimensional data is transformed into smooth functions by FDA method. After that, the three-dimensional data can be described as a two-dimensional matrix. In order to improve the accuracy of monitoring model, the second derivative of smooth functions is calculated to eliminate mean fluctuation caused by non-stationary manufacturing process. Features of the second derivative of each variable are extracted by functional principal component analysis(FPCA) to obtain FPCs. A model is built by the support vector data description method(SVDD) to monitor FPCs and the model is applied to an industrial semiconductor manufacturing process. The results show that the new monitoring method has the lowest missed rate compared with conventional methods, so the effectiveness of the proposed method is validated.

Key words: functional data analysis, functional principal component analysis, non-stationary manufacturing process, process monitoring

中图分类号: