Welcome to visit The Chinese Journal of Process Engineering, Today is

The Chinese Journal of Process Engineering ›› 2022, Vol. 22 ›› Issue (7): 970-978.DOI: 10.12034/j.issn.1009-606X.221399

• Research Paper • Previous Articles     Next Articles

Abnormal condition detection in chemical process based on PCA-SVDD

Yang LIN,  Yadong HE,  Zhuang YUAN,  Chuanpeng WU,  Chengdong GOU,  Chuankun LI*   

  1. State Key Laboratory of Safety and Control for Chemicals, SINOPEC Research Institute of Safety Engineering Co., Ltd., Qingdao, Shandong 266071, China
  • Received:2021-11-30 Revised:2022-01-12 Online:2022-07-28 Published:2022-08-02

基于PCA-SVDD的化工过程异常工况检测

林扬, 何亚东, 袁壮, 武传朋, 苟成冬, 李传坤*   

  1. 化学品安全控制国家重点实验室,中石化安全工程研究院有限公司,山东 青岛 266071
  • 通讯作者: 李传坤 lick.qday@sinopec.com
  • 基金资助:
    国家自然科学基金;中国石化重大科技项目

Abstract: Due to the large number of hazardous materials and serious accident consequences, it has been attracting continuous attentions in the abnormal detection of the chemical plant. Although many detection methods are proposed in the literature, the actual anomaly detection is subject to two challenges. On the one hand, the advanced distributed control system (DCS) can provide massive information about the real-time operating statue of the device, but it also results in a high-order features of training dataset. On the other hand, there is scarce abnormal data in the establishment of abnormal training samples along with the continuous improving device reliability. As a response, this work proposes a PCA-SVDD-based abnormal condition detection method in chemical process under no abnormal data by combing principal component analysis (PCA) and support vector data description (SVDD). Firstly, PCA is employed to reduce the dimensionality by decomposing training samples, which consist of normal data, into the principal subspace and the residual subspace. Then, according to the target type data, an anomaly detection model based on SVDD is established. Further, the Gaussian kernel function is introduced to improve the anomaly detection effect. Finally, the normal data in Tennessee-Eastman (TE) process data is used as training samples to validate the PCA-SVDD. And the two indicators of detection precision (DP) and detection time (DT) are employed to characterize the effect of detection model. In contrast, traditional PCA and SVDD is introduced to carry on anomaly detection of TE process under the same condition. By the comparison, it concludes that the PCA-SVDD-based abnormal condition detection method proposed in this work has better detection effect (higher accuracy and less detection time). In summary, PCA-SVDD can realize the early warning of abnormal working conditions without abnormal data in the chemical process, and has certain significance to ensure the smooth operation of the device.

Key words: PCA, SVDD, chemical process, Abnormal condition detection

摘要: 化工过程往往涉及易燃易爆、高温高压、有毒有害等物料或介质,异常工况检测对保证装置平稳运行具有重要意义。随着装置可靠性和自动化水平的不断提高,异常数据变得匮乏,给异常工况检测提出新的挑战。针对上述问题,本工作融合主成分分析(Principle Component Analysis, PCA)和支持向量数据描述(Support Vector Data Description, SVDD)两种方法,提出一种基于PCA-SVDD的化工过程异常工况检测方法。通过PCA将正常数据分解为主元子空间和残差子空间,然后将主元子空间作为目标类型数据,引入高斯核函数建立SVDD异常检测模型,最后采用田纳西-伊斯曼(Tennessee-Eastman, TE)过程数据对方法进行验证,并与传统PCA和SVDD的异常检测结果进行对比。结果表明,本工作所提方法具有更优的检测效果,可实现化工过程异常工况的早期预警,对保证装置的平稳运行具有一定意义。

关键词: 主成分分析, 支持向量数据描述, 化工过程, 异常工况检测