基于增量动态时间弯曲的时间序列相似性度量方法

Similarity Measure for Time Series Based on Incremental Dynamic Time Warping

针对动态时间弯曲方法计算时间过长的问题,提出增量动态时间弯曲来度量较长时间序列之间的相似性。首先利用动态时间弯曲方法对历史时间序列数据进行相似性度量,得到相应的历史最优弯曲路径和路径中各元素的累积距离代价。其次,通过逆向弯曲度量方法完成当前序列数据的相似性度量,结合历史数据信息找到与历史弯曲路径相交且度量时间序列距离为当前最小值的新路径,进而实现增量动态时间弯曲的相似性度量。该方法不仅具有良好的度量质量,还具有较高的时间效率。数值实验表明,对于大部分时间序列数据集,新方法的分类准确率和计算性能要优于经典动态时间弯曲。

To address the issues on the over expensive time cost,an incremental dynamic time warping（IDTW） to measure the similarity between two time series was proposed.First of all,dynamic time warping（DTW） was used to measure similarity of the past time sequences and retrieves the best warping path and the cumulated distance cost of each element in the warping path.Next,after computing the similarity between the two current time series by backward warping method,a new warping path intersects with the past one was obtained and its warping distance was minimal.Finally,the incremental dynamic warping method was realized to measure similarity.The new method not only has the good quality to measure the similarity but also is efficient to compute.The numerical experiments demonstrate that the classification accuracy and computing performance of IDTW are better than DTW.