化工学报 ›› 2021, Vol. 72 ›› Issue (9): 4830-4837.DOI: 10.11949/0438-1157.20210357

• 过程系统工程 • 上一篇    下一篇

复杂化工过程调控操纵策略的深度学习方法

唐晓婕(),杨博,李宏光()   

  1. 北京化工大学信息科学与技术学院,北京 100029
  • 收稿日期:2021-03-09 修回日期:2021-05-10 出版日期:2021-09-05 发布日期:2021-09-05
  • 通讯作者: 李宏光
  • 作者简介:唐晓婕(1996—),女,硕士研究生,sweetxj123@163.com

Deep learning approaches to complex chemical process control manipulating strategies

Xiaojie TANG(),Bo YANG,Hongguang LI()   

  1. College of Information Science & Technology, Beijing University of Chemical Technology, Beijing 100029, China
  • Received:2021-03-09 Revised:2021-05-10 Online:2021-09-05 Published:2021-09-05
  • Contact: Hongguang LI

摘要:

现代复杂化工过程生产运行记录了大量的过程调控时序数据,如何提取其中有价值的调控操纵经验和规则,对于提升过程运行智能化水平具有重要意义。时间序列聚类是一种挖掘历史调控操纵序列的有效方法,然而由于实际工况经常与历史数据出现偏差,使得重构准确的过程调控操纵策略出现困难。为此,本文提出了一种基于数据碎片化深度学习的过程调控操纵提取方法,采用基于Levenshtein距离的凝聚层次时序聚类获取不同过程扰动状态类别,提取对应的有效操纵序列进行碎片化处理,采用卷积神经网络对调控操纵策略进行深度学习和重构。将此方法在工业换热器过程上进行了应用,获得了满意的结果,表明所提出的方法能够克服常规操纵序列挖掘的工程应用适应性差、对数据源依赖性强等缺点。

关键词: 操纵策略, 层次聚类, Levenshtein距离, SAX符号化, 卷积神经网络

Abstract:

Production operations of modern chemical processes record a large amount of process control manipulating temporal data. How to extract valuable manipulating experiences and rules is of great significance to enhance process operation intelligent level. Previous research results have shown that time series clustering is an effective method for mining historical control manipulating sequences. However, practical working conditions often deviate from the historical data, making it difficult to reconstruct accurate process control manipulating strategies. In response to this problem, this paper proposes a process control manipulation extraction method based on deep learning of data fragmentation, using agglomerative hierarchical temporal clustering based on Levenshtein distance to obtain different process disturbance state classes, extracting corresponding effective manipulating sequences for fragmentation, and employing convolutional neural networks for deep learning and reconstructions of manipulating strategies. The approach is demonstrated by industrial heat exchanger processes, in which, the experiment shows that the proposed approach is able to overcome the poor adaptability and strong dependence on data sources of conventional control manipulating sequence mining methods in practical applications, achieving satisfactory results.

Key words: manipulation strategies, hierarchical clustering, Levenshtein distance, SAX symbolizations, convolutional neural networks

中图分类号: