化工学报 ›› 2018, Vol. 69 ›› Issue (8): 3528-3536.DOI: 10.11949/j.issn.0438-1157.20180025

• 过程系统工程 • 上一篇    下一篇

基于状态空间主成分分析网络的故障检测方法

董顺1,2, 李益国1,2, 孙栓柱3, 刘西陲1,2, 沈炯1,2   

  1. 1 东南大学能源与环境学院, 江苏 南京 210096;
    2 东南大学能源热转换及其过程测控教育部重点实验室, 江苏 南京 210096;
    3 江苏方天电力技术有限公司, 江苏 南京 211102
  • 收稿日期:2018-01-09 修回日期:2018-04-09 出版日期:2018-08-05 发布日期:2018-08-05
  • 通讯作者: 李益国
  • 基金资助:

    国家自然科学基金项目(51476027);江苏省自然科学基金项目(BK20141119)。

Fault detection method based on state space-PCANet

DONG Shun1,2, LI Yiguo1,2, SUN Shuanzhu3, LIU Xichui1,2, SHEN Jiong1,2   

  1. 1 School of Energy and Environment, Southeast University, Nanjing 210096, Jiangsu, China;
    2 Key Laboratory of Energy Thermal Conversion and Control of Ministry of Education, Southeast University, Nanjing 210096, Jiangsu, China;
    3 Jiangsu Frontier Electric Technology Company Limited, Nanjing 211102, Jiangsu, China
  • Received:2018-01-09 Revised:2018-04-09 Online:2018-08-05 Published:2018-08-05
  • Supported by:

    supported by the National Natural Science Foundation of China (51476027) and the Natural Science Foundation of Jiangsu Province (BK20141119).

摘要:

作为一种经典的方法,主成分分析(PCA)在多元统计过程监控领域得到了广泛的应用。然而,主成分分析及其各种改进方法仅从原始数据中提取了一层特征,缺乏对深层次特征的提取。计算机领域深度学习技术的发展表明了深层次的网络结构有利于数据特征的提取,因此,将主成分分析网络(PCANet)这种深度学习网络结构引入到故障诊断领域,与多元统计过程监控方法进行结合,以增强故障检测效果。在PCANet框架下,针对工业过程数据的动态特征,在网络结构中增加了状态空间模型作为动态层以解决动态性问题。此外,还以故障检测为目标重新设计了输出层。最后,通过在TE过程上的仿真测试验证了该方法用于故障检测的可行性和有效性。

关键词: 过程系统, 主元分析, 算法, 故障检测, 状态空间, 深度学习

Abstract:

As a classical algorithm for feature extraction, principal component analysis (PCA) has been widely used in multivariate statistical process monitoring. However, PCA and its various improved methods extracted from original data only one layer of features but no deep layer features. The development of deep learning technology in computer field indicates that deep network structure is beneficial to extraction of data features. Therefore, principal component analysis network (PCANet), a deep learning network structure, was introduced into fault detection and combined with multivariate statistical process monitoring method to enhance fault detection efficiency. Under framework of PCANet, state space model was added to network structure as dynamic layer to solve dynamic issue of industrial process data. In addition, the output layer was redesigned to use fault detection as target function. Finally, method feasibility and validity for fault detection were verified by simulated testing on the Tennessee Eastman (TE) process.

Key words: process systems, principal component analysis, algorithm, fault detection, state space, deep learning

中图分类号: