基于趋势特征聚类的多元相似时间序列的提取

历史相似时间序列的提取在数据挖掘、工业故障检测以及故障根源分析等领域应用非常广泛。针对工业报警系统中异常根源分析方法存在的问题,提出了一种基于趋势特征聚类的多元相似时间序列的提取方法,可以有效地辅助现场工作人员分析关键变量发生异常变化的根源。首先对多元时间序列进行分段线性表示,获得变量的趋势特征信息;然后采用基于密度峰值聚类分析算法对获得的趋势特征在高维空间中聚类,从而实现历史数据的相似性提取;最后可根据关联变量的幅值变化量分析导致主变量发生异常变化的根源变量。数值仿真和实际工业数据案例验证了所提方法的有效性。

Extraction of Multivariate Similar Time Series Based on Trend Feature Clustering

The extraction of historical similar time series is widely used in data mining, industrial fault detection and fault root cause analysis. In order to assist plant operators in analyzing the root causes of occurring anomalous change for a pivot process variable effectively. A method for extracting multivariate similar time series based on trend feature clustering is proposed. Firstly, the trend features of the multivariate time series are obtained by the piecewise linear representation. Then, the obtained trend features are clustered by the density-peak clustering algorithm in high-dimensional space, and thereby the extraction of historical similar time series is achieved. Finally, the root cause variable can be analyzed according to the amplitude change of the related variable when the pivot process variable changes abnormally. The effectiveness of the proposed method is illustrated using numerical and industrial examples.