北京科技大学 自动化学院； 工业过程知识自动化教育部重点实验室， 北京 100083
收稿日期:2019-12-13出版日期:2020-12-04发布日期:2020-12-04通讯作者:黄健(1990-)，女，河北省廊坊市人，讲师，现主要从事工业过程监测与故障诊断研究.电话（Tel.）：010-62332780；E-mail：jianhuang@ustb.edu.cn.基金资助:国家自然科学基金项目(61903026, 67673053)，中国博士后科学基金项目(2019M660462)，中央高校基本科研业务费 (FRF-TP-18-103A1)Online Weighted Slow Feature Analysis Based Fault Detection Algorithm
HUANG Jian,YANG XuSchool of Automation and Electrical Engineering; Key Laboratory of Knowledge Automation for Industrial Processes of the Ministry of Education, University of Science and Technology Beijing, Beijing 100083, China
Received:2019-12-13Online:2020-12-04Published:2020-12-04摘要/Abstract
摘要： 在工业过程监测中，传统的过程监测方法无法提取过程的动态信息，且进行特征选择时没有突出在线故障特征.针对此问题，提出基于在线加权慢特征分析(OWSFA)的故障检测算法.采用慢特征分析(SFA)算法提取过程的本质动态特征；基于正常数据估计出特征阈值，根据松弛系数挑选出在线特征中超过阈值的嫌疑故障特征；引入权重系数，进一步构造基于在线加权的嫌疑故障特征统计量.将提出的OWSFA算法在数值系统和Tennessee Eastman过程进行仿真验证，证实了所提算法的故障检测效果优于主成分分析和SFA算法.OWSFA算法根据故障信息，在线构造加权统计量，加强了动态故障特征在监测模型中的表达.
关键词: 慢特征分析, 故障检测, 特征加权, 动态过程, 化工过程
Abstract: In industrial process monitoring, traditional process monitoring methods fail to extract process dynamic information and the online fault-related information is not fully utilized in feature selection. To solve these problems, a fault detection method based on online weighted slow feature analysis (OWSFA) is proposed. First, the slow feature analysis(SFA)algorithm is utilized to extract the dynamic features. The threshold of slow feature is estimated based on benchmark data. The online features which exceed the threshold based on the slack factor are selected as the suspected fault features. Then, the monitoring statistics are built based on the weighted suspected fault features by introducing the weights. The proposed OWSFA algorithm is applied in a numerical system and the Tennessee Eastman process, which proves that the proposed method has superiorities compared with principal component analysis and the SFA method. According to the faulty information, the OWSFA algorithm generates the online weighted statistics to enhance the fault features in the monitoring model.
Key words: slow feature analysis (SFA), fault detection, weighted feature, dynamic process, chemical process
PDF全文下载地址:
点我下载PDF
