化工学报 ›› 2020, Vol. 71 ›› Issue (7): 3172-3179.DOI: 10.11949/0438-1157.20191581

• 过程系统工程 • 上一篇    下一篇

基于循环自动编码器的间歇过程故障监测

高学金1,2,3,4(),刘腾飞1,2,3,4,徐子东1,2,3,4,高慧慧1,2,3,4,于涌川1,2,3,4()   

  1. 1. 北京工业大学信息学部,北京 100124
    2. 数字社区教育部工程研究中心,北京 100124
    3. 城市轨道交通北京实验室,北京 100124
    4. 计算智能与智能系统北京市重点实验室,北京 100124
  • 收稿日期:2019-12-25 修回日期:2020-02-05 出版日期:2020-07-05 发布日期:2020-07-05
  • 通讯作者: 于涌川
  • 作者简介:高学金(1973—),男,博士,教授, gaoxuejin@bjut.edu.cn
  • 基金资助:
    国家自然科学基金项目(61803005);北京市自然科学基金项目(4172007);山东省重点研发计划项目(2018CXGC0608);北京市教育委员会资助项目

Intermittent process fault monitoring based on recurrent autoencoder

Xuejin GAO1,2,3,4(),Tengfei LIU1,2,3,4,Zidong XU1,2,3,4,Huihui GAO1,2,3,4,Yongchuan YU1,2,3,4()   

  1. 1. Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China
    2. Engineering Research Center of Digital Community, Ministry of Education, Beijing 100124, China
    3. Beijing Laboratory for Urban Mass Transit, Beijing 100124, China
    4. Beijing Key Laboratory of Computational Intelligence and Intelligent System, Beijing 100124, China
  • Received:2019-12-25 Revised:2020-02-05 Online:2020-07-05 Published:2020-07-05
  • Contact: Yongchuan YU

摘要:

针对间歇过程数据非线性、动态性特征,提出一种基于循环自动编码器(recurrent autoencoder,RAE)的过程故障监测方法。采用长短时记忆(long short-term memory,LSTM)循环神经网络构建自动编码器建立监控模型,相比传统自动编码器,其能有效挖掘时序样本间的动态关联信息。该方法首先利用批次展开与变量展开相结合的三步展开方法将间歇过程数据展开成二维,并通过滑动窗采样得到模型输入序列;然后使用LSTM构建自动编码器,重构输入序列。进一步,利用重构误差构造平方预测误差(squared prediction error, SPE)统计量实现在线监测。最后将所提方法应用于青霉素发酵仿真和重组大肠杆菌发酵过程监测,结果表明,该方法能及时监测到故障,具有较好的监测性能。

关键词: 算法, 动态建模, 神经网络, LSTM, 过程监测, 循环自动编码器

Abstract:

Aiming at the non-linear and dynamic characteristics of intermittent process data, a process fault monitoring method based on recurrent autoencoder (RAE) is proposed. To establish monitoring model, an autoencoder is constructed by using long short-term memory (LSTM) recurrent neural network. Compared with the traditional autoencoder, the proposed method can effectively extract the dynamic correlation information between time series samples. Firstly, a three-step expansion method combining batch expansion and variable expansion are used to expand the batch process data into two dimensions, and input sequences for modeling is obtained by sliding window sampling. Then, LSTM is used to reconstruct the input sequences to train an autoencoder model. Moreover, the Squared prediction error (SPE) statistics are constructed based on reconstruction error to achieve on-line monitoring. Finally, the proposed method is applied to penicillin fermentation process for simulation experiment and recombinant Escherichia coli fermentation process monitoring. The results show that the method can detect faults in time and has better monitoring performance.

Key words: algorithm, dynamic modeling, neural networks, LSTM, process monitoring, recurrent autoencoder

中图分类号: