中国机械工程 ›› 2020, Vol. 31 ›› Issue (14): 1686-1692.DOI: 10.3969/j.issn.1004-132X.2020.14.007

• 服务型制造 • 上一篇    下一篇

基于误差注意力的晶圆制造数据异常检测

余石龙1, 鲍劲松1, 李婕1, 张启华2   

  1. 1. 东华大学机械工程学院, 上海, 201600;
    2. 中芯国际集成电路制造有限公司, 上海, 201203
  • 收稿日期:2019-05-09 出版日期:2020-07-25 发布日期:2020-08-26
  • 通讯作者: 鲍劲松(通信作者),男,1972年生,教授、博士研究生导师。研究方向为智能制造、虚拟显示、人机交互。E-mail:bao@dhu.edu.cn。
  • 作者简介:余石龙,男,1994年生,硕士研究生。研究方向为工业大数据、异常检测、无人系统。
  • 基金资助:
    国家自然科学基金资助项目(51435009)

Wafer Manufacturing Data Anomaly Detection Based on Error Attention

YU Shilong1, BAO Jinsong1, LI Jie1, ZHANG Qihua2   

  1. 1. College of Mechanical Engineering, Donghua University, Shanghai, 201600;
    2. Semiconductor Manufacturing International Corp., Shanghai, 201203
  • Received:2019-05-09 Online:2020-07-25 Published:2020-08-26

摘要: 针对晶圆制造数据异常检测过程中异常特征提取难度大且检测效率不高的问题,提出了一种基于误差注意力的晶圆制造数据异常检测方法。在保持数据分布不变的前提下,将晶圆制造数据转化成灰度图像,根据与正常样本的误差对灰度图像生成基于位置的柔性注意力图,增加误差特征的显性表达并略去冗余特征;利用深度学习神经网络LeNet-5模型将注意力图进行卷积训练,得到异常检测的最优化模型。采用晶圆制造数据集与现有方法进行对比,所提方法耗时缩短160%、F2-Score提高3%,证明了所提方法的有效性。

关键词: 注意力机制, 异常检测, 晶圆制造, 特征提取, 卷积神经网络

Abstract: In wafer manufacturing processes, it was difficult and inefficient to extract abnormal data in the anomaly detection. Therefore, an anomaly detection method was proposed based on error attention mechanism. Firstly, the wafer manufacturing data was converted into grayscale images while keeping the data distribution unchanged. Then, based on the errors with the normal sample, a position-based flexible attention map was generated for the grayscale images. The dominant expressions of the error features were developed while the redundant features areomitted. In addition, deep learning neural network model LeNet-5 was used to perform convolution training to obtain an optimal model for anomaly detection. Finally, the wafer fabrication data set was used to compare the performance between the proposed method and the existing methods. The proposed method reduces the time consuming by 160% and increases the F2-Score by 3%, which verifies the effectiveness of the method.

Key words: attention mechanism, anomaly detection, wafer fabrication, feature extraction, convolutional neural network

中图分类号: