基于FastICA和G-G聚类的多元时序自适应分段  

Adaptive Segmentation of Multivariate Time Series with FastICA and G-G Clustering

作　　者：王玲[1,2] 李泽中 WANG Ling;LI Ze-zhong(School of Automation and Electrical Engineering,University of Science and Technology Beijing,Beijing 100083,China;Key Laboratory of Knowledge Automation of Industrial Processes of Ministry of Education,School of Automation and Electrical Engineer-ing,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]北京科技大学自动化学院,北京100083 [2]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

出　　处：《电子学报》2023年第5期1235-1244,共10页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62076025,No.61572073)。

摘　　要：现有多元时间序列的分段方法主要通过检测时序数据统计特性或形状的变化情况,并以此为依据对分段点的位置进行“硬划分”.然而,这些分段方法无法对两个分段之间的过渡区间长度进行准确估计,且普遍需要人为预先设置参数,在高维且噪声较强的情况下分段效果较差.本文针对现有分段方法存在的诸多不足,提出一种基于FastICA(Fast Independent Component Analysis)和G-G(Gath-Geva)模糊聚类的多元时序自适应分段方法 .该方法利用FastICA进行特征提取,采用DW(Durbin-Watson)指数自动选取高信噪比的主成分,并根据最小描述长度(Minimum Description Length,MDL)设计基于G-G模糊聚类的自适应分段模型,实现对于多元时间序列的“软划分”.基于多种领域的真实数据集实验结果表明:与现有主流的分段方法相比,本文方法在上述数据集上的平均F1和MAE(Mean Absolute Error)可分别提升8.4%~16.8%和3.06%~6.56%.The existing segmentation methods detect the statistical or shape changes of multivariate time series,and perform crisp segmentation on the location of change points.However,these methods fail to estimate the length of the tran-sition interval between two segments,cannot accurately segment multivariate time series with high dimension,strong noise,and need to set parameters in advance.To address such matters,an adaptive multivariate time series segmentation method based on FastICA(Fast Independent Component Analysis)and G-G(Gath-Geva)clustering is proposed.In this method,the key features of multivariate time series are extracted via FastICA,and DW(Durbin-Watson)criterion is used to automatical-ly select main components with high signal-to-noise ratio.According to the minimum description length(MDL),an adap-tive multivariate time series segmentation model based on G-G clustering is designed,which is able to perform soft segmen-tation of multivariate time series.The experimental analysis is carried out on real datasets in many different fields.Com-pared with state-of-art benchmarks,the average F1 and MAE(Mean Absolute Error)of the proposed method on the above-mentioned datasets improve 8.4%~16.8%and 3.06%~6.56%,respectively.

关 键 词：多元时间序列 自适应分段 快速独立主成分分析 Gath-Geva聚类 最小描述长度 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]