化工学报 ›› 2021, Vol. 72 ›› Issue (3): 1529-1538.DOI: 10.11949/0438-1157.20201748

• 过程系统工程 • 上一篇    下一篇

基于分位数回归CGAN的虚拟样本生成方法及其过程建模应用

陈忠圣1,2(),朱梅玉1,2,贺彦林1,2,徐圆1,2,朱群雄1,2()   

  1. 1.北京化工大学信息科学与技术学院,北京 100029
    2.智能过程系统工程教育部工程研究中心,北京 100029
  • 收稿日期:2020-12-02 修回日期:2020-12-09 出版日期:2021-03-05 发布日期:2021-03-05
  • 通讯作者: 朱群雄
  • 作者简介:陈忠圣(1994—),男,博士研究生,zschen@mail.buct.edu.cn
  • 基金资助:
    国家自然科学基金项目(61973024)

Quantile regression CGAN based virtual samples generation and its applications to process modeling

CHEN Zhongsheng1,2(),ZHU Meiyu1,2,HE Yanlin1,2,XU Yuan1,2,ZHU Qunxiong1,2()   

  1. 1.College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China
    2.Engineering Research Center of Intelligent PSE, Ministry of Education of China, Beijing 100029, China
  • Received:2020-12-02 Revised:2020-12-09 Online:2021-03-05 Published:2021-03-05
  • Contact: ZHU Qunxiong

摘要:

针对复杂工业过程因难以检测变量或因时间上和经济上成本因素导致的建模样本稀缺问题,提出了一种将分位数回归(quantile regression)嵌入到条件生成式对抗网络(conditional generative adversarial network, CGAN)的虚拟样本生成方法QRCGAN。首先,在标准CGAN“生成器-判别器”两元对弈结构中嵌入回归器,使模型不仅具备标签样本生成能力,同时也具备处理回归预测问题的能力。其次,以分位数回归神经网络实现回归器,连同判别器和生成器进行同步对抗训练。当模型到达Nash平衡时,在分位数回归神经网络回归器的作用下,生成器能够产生落在一定置信区间的新样本。然后,利用Kullback-Leibler(KL)散度评估生成样本的质量。最后,通过标准函数数据和实际化工过程数据验证所提方法的有效性。

关键词: 虚拟样本生成, CGAN, 分位数回归, 数据稀缺, 软测量, 深度学习

Abstract:

Complex industry processes suffer from the problems on data scarcity of training samples which are collected for modeling, as a result of inaccessibility of difficult-to-measure variables or high cost in time or economy. To tackle the issues, we proposed a novel virtual sample generation embedding quantile regression into conditional generative adversarial networks (QRCGAN). First of all, a regression is embedded in the standard CGAN “generator-discriminator” two-element game structure, so that the model not only has the ability to generate label samples, but also has the ability to handle regression prediction problems. Secondly, the regressor is implemented by the quantile regression neural network (QRNN), together with the discriminator and generator for simultaneous adversarial training. Once the model reaches the Nash equilibrium, with the help of the QRNN regressor, the generator can generate new samples that fall within a certain confidence interval. Moreover, the Kullback-Leibler (KL) divergence was used to evaluate the quality of the generated samples. Finally, the effectiveness of the proposed method is verified by standard function data and actual chemical process data.

Key words: virtual sample generation, CGAN, quantile regression, data scarcity, soft sensing, deep learning

中图分类号: