Abstract: Mapping the raw time series data to a modality space effectively is a critical problem in time series similarity search. This paper puts forward a method of Feature Points Transform Symbolization(FPTS) for time series, which combines ideology of symbolic time series and piecewise linear representation. It can extract the shape modality of time series, hold the speciality of extreme value points under dimensionality reduction and noises removing, and support the similarity measure based on Dynamic Time Warping(DTW) distance, which overcomes the limitation of precise matching of traditional symbolization method. Experiments prove the accuracy and efficiency of the method.

Key words: time series, similarity search, symbolization, feature points, Dynamic Time Warping(DTW) distance

摘要： 将时序数据有效地映射到特征空间是时间序列相似性搜索的一个关键问题。文章结合时间序列符号化思想与分段线性表示中分段点选取的思想，提出一种基于特征点转换的时间序列符号化方法FPTS。该方法能有效提取序列的形状特征，在降维和除噪的同时保留序列的极值点特性，支持基于动态时间弯曲距离的相似性度量，克服传统的符号化方法受限于精确匹配的缺陷。实验证明了该方法的准确性和高效性。

关键词: 时间序列, 相似性搜索, 符号化, 特征点, 动态时间弯曲距离

CLC Number: 

• TP274