欢迎访问过程工程学报, 今天是

过程工程学报 ›› 2021, Vol. 21 ›› Issue (12): 1491-1502.DOI: 10.12034/j.issn.1009-606X.220340

• 系统与集成 • 上一篇    

基于约翰逊转换的鲁棒化工过程监控方法

王骥1， 柳楠1， 胡明刚2， 田文德1*
  

  1. 1. 青岛科技大学化工学院，山东 青岛 266042 2. 青岛诺诚化学品安全科技有限公司，山东 青岛 266071
  • 收稿日期:2020-10-19 修回日期:2021-02-06 出版日期:2021-12-28 发布日期:2022-03-28
  • 通讯作者: 田文德 tianwd@qust.edu.cn
  • 作者简介:王骥(1995-)，男，山东省济南市人，硕士研究生，化学工程专业，E-mail: 1160001576@qq.com；田文德，通讯联系人，E-mail: tianwd@qust.edu.cn.
  • 基金资助:
    山东省重点研发项目

A robust method for chemical process monitoring based on Johnson transformation

Ji WANG1,  Nan LIU1,  Minggang HU2,  Wende TIAN1*   

  1. 1. College of Chemical Engineering, Qingdao University of Science & Technology, Qingdao, Shandong 266042, China 2. Qingdao Nuocheng Chemical Safety Technology Co., Ltd., Qingdao, Shandong 266071, China
  • Received:2020-10-19 Revised:2021-02-06 Online:2021-12-28 Published:2022-03-28
  • Contact: TIAN Wen-de tianwd@qust.edu.cn

PDF

摘要： 化工厂中一个小故障可能导致大事故，从而造成生命财产损失和环境破坏。为了防止小故障演变成大事故，化学工业需要有效的过程监控来及时检测故障和诊断故障原因。传统化工过程监控方法主元分析法(Principal Component Analysis, PCA)假设数据服从高斯分布，实践中有时并不满足该条件。此外，其使用方差、协方差捕捉数据非线性变化时，鲁棒性较差。本工作提出一种改进的主元分析法—基于约翰逊转换的鲁棒过程监控方法。首先引入约翰逊正态转换(Johnson Transformation)使过程数据服从高斯分布；其次使用鲁棒性强的斯皮尔曼相关系数(Spearman Correlation Coefficient)矩阵代替传统主元分析法的协方差矩阵提取特征向量，构造特征空间；最后将过程数据投影到特征空间，使用T2和SPE统计量实施过程监控。将此方法应用于TE过程故障案例，并与PCA和核主元分析法(Kernel Principal Component Analysis, KPCA)对比，验证了此方法的有效性。

关键词: 斯皮尔曼相关系数, TE过程, 约翰逊转换, 过程监控

Abstract: A tiny fault in chemical plants is likely to cause an enormous accident possibly with heavy losses of personnel, property, and environment. Therefore, process monitoring is demanded to timely detect faults and identify fault variables, so as to avoid deterioration of tiny faults into accidents. Nowadays principal component analysis (PCA) is the most widely used method in chemical process monitoring practice with its simplicity and effectiveness. However, it has some drawbacks. First, it roughly assumes process data as Gaussian distribution. But sometimes it is not satisfied. Furthermore, PCA uses variance and covariance (also called Pearson correlation coefficients) as criterion to choose principal components, however they are not robust in capturing nonlinear data variation. To alleviate these problems, an improved PCA-a Johnson transformation based robust method for process monitoring (JSPCA) was proposed in this work. First, Johnson transformation was introduced to make process data obey Gaussian distribution. Second, the Spearman correlation coefficient matrix instead of covariance matrix was established to extract principal components and span feature space. Finally, process data were projected into feature space where T2 and SPE statistics were obtained for process monitoring. The proposed method had its fault detection ability tested in the benchmark TE process with comparison of PCA and KPCA. The results showed that the proposed method had higher fault detection rates than PCA and KPCA when using T2 as detection indicator. However, the proposed method with SPE as detection indicator had higher false alarm rates than PCA and KPCA. As for fault diagnosis ability, the proposed method was tested against fault 5 and fault 10 of TE process and diagnoses fault variables more precisely than PCA and KPCA. The proposed method was better than PCA and KPCA and it was worth promoting.

Key words: Spearman correlation coefficients, TE process, Johnson transformation, process monitoring