基于分层动态时间弯曲的序列相似性度量方法研究

Hierarchical dynamic time warping method for evaluating time series similarity

为了更好地体现时间序列的形态特征,并探索更适合于较长时间序列之间相似性度量的方法,在动态时间弯曲算法的基础上进行改进,提出了基于分层动态时间弯曲的序列相似性度量方法。对时间序列进行多层次分段,并从分段中均匀抽取相对应的层次分段子序列,然后将层次分段子序列抽象为三维空间的点(反映了分段子序列的均值、长度和趋势)进行相似性度量,最后综合各个层次的相似性度量作为结果。实验表明,在参数设置合理的情况下,此方法能获得较高的序列相似性度量准确度和效率。

In order to better reflect the morphological characteristics of time series, and explore more suitable methods for e- valuating a longer time series similarity, this paper proposed a hierarchical dynamic time warping method based on dynamic time warping method. First,it divided the time series into multi-levels and extract uniformly the corresponding subsequence seg- ments. Secondly,it made the sequence segments to the three-dimensional abstraction points （ reflecting the mean length, and trends of the subsequence） for similarity measure. Finally, it integrated all levels of similarity measure as the ultimate result. Experiments reveal that this method can get the higher accuracy and efficiency with the reasonable parameters.