基于DTW的时间序列相似度量方法的优化  

Optimization of Time Series Similarity Measurement Based on DTW

作　　者：刘美 王全民 LIU Mei;WANG Quanmin(Faculty of Information Technology,Beijing University of Technology,Beijing 100020)

机构地区：[1]北京工业大学信息学部,北京100020

出　　处：《计算机与数字工程》2023年第4期814-819,938,共7页Computer & Digital Engineering

基　　金：北京市自然科学基金项目(编号:4202004)资助。

摘　　要：现有的时间序列相似度量方法难以兼顾微观形状与宏观结构两方面的相似性,同时存在易受异常值、位移拉伸影响等问题。针对上述情况,提出具有更高精度与鲁棒性,更小时间复杂度的优化DTW与面积距离结合的算法。首先在DTW中引入分段聚合与相似度阈值,前者能降低数据维度,减少DTW的计算量,后者能够过滤数据中的异常值,提高算法的鲁棒性;其次提出面积距离度量,描述整体序列中的起伏变化,发现序列的整体相似性;从微观形状与宏观结构两方面表述时序序列,能够进一步提升算法精度。在UCR的30个不同数据集上进行的实验表明,该方法与现有相似度量方法相比,具有更高的精度与鲁棒性。The existing similarity measurement methods of time series are difficult to take into account the similarity of micro shape and macro structure,and are easily affected by outliers and displacement stretching.In view of the above situation,an algorithm combining DTW with area distance is proposed,which has higher accuracy,robustness and less time complexity.Firstly,segmented aggregation and similarity threshold are introduced into DTW,the former can reduce the data dimension and reduce the calculation amount of DTW,and the latter can filter the outliers in the data and improve the robustness of the algorithm.Secondly,the area distance measure is proposed to describe the ups and downs of the whole sequence and find the overall similarity of the sequence,the temporal sequence can be expressed from the micro shape and macro structure further improve the accuracy of the algorithm.Experiments on 30 different data sets of UCR show that the proposed method has higher accuracy and robustness than the existing similarity measurement methods.

关 键 词：时间序列 相似性量度 动态时间弯曲 面积距离 相似度阈值 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]