一种时空协同的图卷积长短期记忆网络及其工业软测量应用
作者:
作者单位:

浙江大学控制科学与工程学院

作者简介:

通讯作者:

中图分类号:

TP183

基金项目:

浙江省工业化与信息化融合联合基金(No. U1709211),工业控制技术国家重点实验室自主课题(ICT2021A15)


A Spatio-Temporal Synergistic Graph Convolution Long Short-Term Memory Networks and Its Application for Industrial Soft Sensor
Author:
Affiliation:

Control Science and Engineering, Zhejiang University

Fund Project:

NSFC-Zhejiang Joint Fund for the Integration of Industrialization and Informatization (No. U1709211), the Research Project of the State Key Laboratory of Industrial Control Technology, Zhejiang University, China, under Grant ICT2021A15

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    近年来,软测量技术的发展有效解决了工业过程中对于难以直接测量的质量变量的感知困难,为过程的控制与优化提供了有力保障。通常在含有多个质量变量的过程中,样本间的时序关系和多个质量变量间相互影响的空间关系能够反映过程本身的特性,这种时空特性的挖掘有益于软测量模型性能的提升,传统软测量方法往往局限于对时序关系的学习而并未考虑对质量变量间的空间关系进行有效利用。因此本文提出一种时空协同的图卷积长短期记忆网络(Graph Convolution Long Short-Term Memory Networks, GC-LSTM),应用于工业软测量场景。采用多通道网络结构将图卷积网络的空间关系挖掘能力和长短期记忆网络的时序关系学习能力相结合,对过程进行时空协同学习,实现了软测量应用。具体来说,每条通道用于对每种质量变量进行独立学习;对于过程的时序特性,利用各通道内的长短期记忆网络提取针对不同质量变量的时序特征;对于过程的空间特性,构建质量变量间空间关系的图结构,采用跨通道的图卷积运算将不同通道内不同质量变量的时序特征基于空间关系进行融合,得到兼具过程时空特性的特征,因而在软测量建模中实现了过程时空协同学习与融合。通过燃煤电厂磨煤机的实际生产数据验证了所提出的方法对软测量性能提升的有效性。

    Abstract:

    Recently, the development of soft sensor has shown great superiority in the measurement of unmeasurable quality variables in industrial process, which provides essential basis for the control and optimization of the process. Generally, in process with multiple quality variables, the temporal dependencies among samples and the spatial dependencies among quality variables can well reflect the inner property of a process. Therefore, the mining for spatio-temporal property will be advantageous for the promotion of soft sensor performance, while conventional methods are generally limited in learning temporal dependencies but neglect the usage for the spatial dependences among quality variables. In this paper, we propose a spatio-temporal synergistic Graph Convolution Long Short-Term Memory Networks (GC-LSTM) for the application of industrial soft sensor, which combines the spatial-mining ability of Graph Convolutional Networks and the temporal-mining ability of Long-Short Term Memory through multi-channel network structure. The proposed model adopts spatio-temporal synergistic learning to exploit the inner property of the process, and implements soft sensing. Specifically, each quality variable is learnt independently in each channel. As for the temporal property of a process, Long-Short Term Memory Networks in each channel can extract temporal features for specific quality variable. As for the spatial property of a process, graph is constructed to describe the spatial dependencies among quality variables. Then, graph convolutional operation across channels can fuse the temporal features of different quality variables in different channels based on the spatial information in graph for the extraction of spatio-temporal features. Thus, the proposed model implements spatio-temporal synergistic learning and fusion in the modeling process. Experiments based on real data set from coal mill at coal-fired power plant validate the effectiveness of the proposed model for performance improvement.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2020-07-05
  • 最后修改日期:2021-08-14
  • 录用日期:2020-10-12
  • 在线发布日期: 2020-12-01
  • 出版日期: