基于多维分段和动态权重DTW的多元时间序列相似性度量方法

Similarity Measurement Method of Multivariate Time Series Based on Multidimensional Segmentation and Dynamic Weighted DTW

针对常用方法无法准确度量多元时间序列相似程度的问题,提出一种基于多维分段和动态权重动态时间弯曲距离的多元时间序列相似性度量方法。首先对多元时间序列进行多维分段拟合,选取拟合段的斜率、均值和时间跨度作为每一段的特征,在对多元时间序列降维的同时也保留了变量之间的相关性;然后提出一种动态权重动态时间弯曲距离度量方法计算多元时间序列特征矩阵之间的距离,避免了直接使用动态时间弯曲距离造成的畸形匹配问题。最终实验结果也验证了该方法在多种类型的数据集上都能取得较高的度量精度,表明了该方法的有效性。

Common methods can not measure the similarity of multivariate time series accurately.In this paper,a similarity measurement method for multivariate time series based on multi-dimensional segmentation and dynamic weighted dynamic time warping distance is proposed.Firstly,multivariate time series are fitted with multi-dimensional piecewise method and the gradient,mean value and time span of each segment are chosen as the feature pattern.While the dimension of multi-dimensional time series is reduced,the correlation between variables is preserved.Then,a dynamic weighted dynamic time warping distance measurement method is proposed to calculate the distance between the feature matrix,which avoids the incorrect matching by the direct use of dy⁃namic time warping.The final experimental results also verify that the method can achieve high measurement accuracy on various types of data sets,and show the effectiveness of the method.