长短滑窗慢特征分析与时序关联规则挖掘的过渡过程识别

PDF下载HTML

Title:

Transition process identification based on the long and short sliding windowed slow feature analysis and time series association rule mining

作者:

刘金平¹, 匡亚彬¹, 赵爽爽¹, 杨广益²

1. 湖南师范大学 信息科学与工程学院, 湖南 长沙 410081;
2. 湖南省计量检测研究院 信息中心, 湖南 长沙 410014

Author(s):

LIU Jinping¹, KUANG Yabin¹, ZHAO Shuangshuang¹, YANG Guangyi²

1. College of Information Science and Engineering, Hunan Normal University, Changsha 410081, China;
2. Information Center, Hunan Institute of Metrology and Test, Changsha 410014, China

关键词:

过程监测; 过渡过程识别; 慢特征分析; 同步频繁树; 时序关联规则挖掘; 稳态工况; 长短滑窗; 多模态工况

Keywords:

process monitoring; identification of transition process; slow feature analysis; synchronize frequent tree; temporal association rule mining; steady state condition; long and short sliding window; multimode condition

分类号:

TP29

DOI:

10.11992/tis.202205048

摘要:

工况过渡过程与异常状态（故障）的数据特性极为相似。如果不对过渡过程加以辨识，极易导致过程监测系统频繁误报警，进而可能引发不适当的人工操作而严重破坏生产的稳定性。本文提出一种基于长短滑窗慢特征分析（slow feature analysis, SFA）与时序关联规则挖掘的过渡过程识别方法。首先，依据稳态工况和过渡工况在时间跨度上的差异性，提出一种长短滑窗与SFA相结合的多工况过程建模方法，将工况状态细分为多个稳态阶段与过渡阶段，并分别建立相应的离线SFA模型；然后，提出一种多时序多时间区间的同步频繁树构建方法，挖掘每种状态转变在多个时间序列与多个时间区间内的关联规则，以实现工况过渡过程的准确辨识。针对田纳西伊斯曼（Tennessee Eastman, TE）过程生成一组包含多模态相互转变的过程数据对所提方法进行实验验证，结果表明所提方法能够在频繁发生过程转变的过程数据中有效识别过渡过程，降低故障误报率，提高过程监测水平。

Abstract:

The transition process of working condition has similar data characteristics with process anomalies or faults. Thus, it is prone to cause frequent false alarms in the process monitoring if the transition process cannot be identified, which may lead to inappropriate manual operations and consequently destroy production stability. This paper proposes a transition process identification method based on the long and short sliding windowed slow feature analysis (SFA) and time series association rules mining approach. First, a multi-mode process modeling method based on the SFA associated with long and short sliding window processing is proposed according to the time-span difference between steady and transition conditions. The working condition state is divided into multiple sub-stages of steady or transition states, and then corresponding offline SFA models are established, respectively. Then, a synchronous frequent tree construction method with multi-time series and multi-time intervals is proposed, which can mine the association rules of each transition state in multi-time series and multi-time intervals, so as to accurately identify transition processes. Based on the Tennessee Eastman (TE) process, a set of process data, including all modal transitions, is established to verify the proposed method. The results show that the proposed method can effectively identify the transition process with frequent transitions, reducing the false alarm rate and improving the level of process monitoring.