工业工程

工业工程 ›› 2023, Vol. 26 ›› Issue (3): 143-150.doi: 10.3969/j.issn.1007-7375.2023.03.016

• 系统建模与优化方法 • 上一篇    下一篇

考虑特征学习的IPSO-LSTM晶圆加工周期预测

张蓝天, 石宇强   

  1. 西南科技大学 制造科学与工程学院，四川 绵阳 621010
  • 收稿日期:2021-12-09 发布日期:2023-07-08
  • 作者简介:张蓝天(1996-),女,四川省人,硕士研究生,主要研究方向为生产系统性能预测、系统仿真
  • 基金资助:
    四川省教育厅科研资助项目(18ZA0497)

Wafer Cycle Time Prediction of IPSO-LSTM Considering Feature Learning

ZHANG Lantian, SHI Yuqiang   

  1. College of Manufacturing Science and Engineering, Southwest University of Science and Technology, Mianyang 621010, China
  • Received:2021-12-09 Published:2023-07-08

RichHTML

16

PDF

1596

摘要： 为了推动大数据技术在制造车间的应用，针对复杂产品晶圆制造过程中海量制造数据时序性、强噪音影响加工周期预测精度的问题，提出考虑特征学习的改进粒子群优化长短期记忆网络 (improved particle swarm optimization-long short term memory, IPSO-LSTM) 的加工周期预测方法。采用降噪自编码器和稀疏自编码器联合构建深度自编码器，增强特征学习能力和抗噪能力；运用IPSO优化LSTM参数，克服时间依赖性，提升预测模型性能。实例验证了所提方法的预测精度优于传统机器学习方法，其平均绝对误差低于3%；并分析特征学习方法的有效性，将支持向量回归和多层感知器等传统方法加入特征学习方法，R2分别提高了1.46%、1.05%，为晶圆加工周期的有效预测提供新的解决方法。

关键词: 粒子群优化算法, 生产周期, 自动编码器, 长短时记忆网络, 特征学习

Abstract: In order to promote the application of big data technology in manufacturing workshops, aiming at the problem that the temporality and strong noise of massive manufacturing data have an effect on prediction accuracy of cycle time in complex wafer manufacturing process, a cycle time prediction method based on improved particle swarm optimization-long short term memory (IPSO-LSTM) considering feature learning is proposed. A combination of denoising auto-encoder and sparse auto-encoder are used to construct a deep auto-encoder to enhance the abilities of feature learning and noise resisting. IPSO algorithm is used to optimize the parameters of LSTM, overcoming time dependence and improving the performance of the prediction model. Experiments verifies that the prediction accuracy of the proposed approach is better than that of traditional machine learning methods, of which the average absolute error is less than 3%. The effectiveness of the feature learning method is analyzed, which is introduced into traditional prediction methods, such as support vector regression and multi-layer perceptron, resulting an R-square being increased by 1.46% and 1.05%, respectively, and providing a new solution for effective prediction of wafer processing cycle time.

Key words: particle swam optimization algorithm, production cycle time, auto-encoder, long short-term memory network, feature learning

中图分类号: