基于加权动态时间弯曲的多元时间序列相似性匹配方法  

Multivariate Time Series Similarity Matching Method Based on Weighted Dynamic Time Warping Algorithm

作　　者：叶燕清[1] 杨克巍[1] 姜江[1] 葛冰峰[1] 豆亚杰[1] 

机构地区：[1]国防科学技术大学信息系统与管理学院,长沙410073

出　　处：《模式识别与人工智能》2017年第4期314-327,共14页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.71671186;71571185;71501182)资助~~

摘　　要：针对常用方法忽略变量相关性和局部形状特性问题,提出基于加权动态时间弯曲的多元时间序列相似性匹配方法(CPCA-SWDTW).首先,在原加权动态时间弯曲算法基础上,引入形态因子,提出基于形态特征的加权动态时间弯曲算法(SWDTW).然后,提取多元时间序列的主成分作为模式表示,消除变量间的相关性,同时将方差贡献率作为相应主成分的权重.在此基础上,运用SWDTW,度量多元时间序列间的相似度.最后,通过相似性搜索实验表明,CPCA-SWDTW具有较好的准确性和鲁棒性.敏感性分析说明CPCA-SWDTW在一定程度上受到权重函数参数的影响.In most of the current methods, the close correlation between variables and the shape characteristics of time series is neglected. In this paper, a similarity matching method for multivariate time series is proposed based on combined principal component analysis method and a shape-based improved weighted dynamic time warping algorithm（CPCA-SWDTW）. Firstly, a shape coefficient is introduced and a shape based weighted dynamic time warping（SWDTW） algorithm is presented. Next, the principal components of multivariate time series are extracted as the representation, and thus the variable correlations can be eliminated. Besides, the variance devoting rate of each principal component is considered as the weight of each series. On the basis of the proposed representation, SWDTW is used to measure the similarity between transformed multivariate time series. Finally, the resuhs of similarity search experiment show that CPCA-SWDTW is more efficient and show that CPCA-SWDTW can be affected robust. Moreover, the parameter sensitivity analysis experiment by the parameters in weight function to some extent.

关 键 词：多元时间序列 相似性匹配 共同主成分分析 加权动态时间弯曲 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]