多元时间序列的相似性匹配

常用多元时间序列相似性匹配方法难以在高效刻画局部形态特征的同时考虑各变量间的相关信息. 针对此问题，提出一种动态窗口内多维拟合分段方法. 基于序列的局部形态特征抽象出各变量维度上拟合线段的倾斜角及持续时间，组成模式表示矩阵，并借助一种多元模式距离实现序列的相似性模式匹配. 与主成分分析法、基于点分布特征的匹配法对不同数据规模的数据集进行对比，验证了该方法的有效性，特别对于多变量、不等时间跨度的中等规模多元时间序列相似性匹配具有较好的效果.

Matching Similar Patterns for Multivariate Time Series

With ordinary methods, it is difficult to take relational information between variables while match the local shape of multivariate time series efficiently. To deal with the problem, we propose a multidimensional fitting piecewise method based on dynamic window to segment multivariate time series. Secondly, the inclination angle and time span of a fitting segment in a certain variable dimension are used to construct a feature pattern matrix. A multivariate pattern distance is used to measure similarity between the series. Finally, by comparison with principal component analysis and the matching method based on point distribution for three different data sets, we obtain preferable results, showing that the proposed method is more efficient, especially
for the medium sized time series with multivariate and varying time span.