多变量时间序列例外模式的识别

Identification of Outlier Patterns in Multivariate Time Series

多变量时间序列(MTS)在金融、医学、科学、工程等领域是非常普遍的.本文提出一种在 MTS 中识别异常模式的方法.采用自底向上的分割算法将 MTS 分割成互不重叠的子序列,使用扩展的 Frobenius 范数来计算2个MTS 子序列之间的相似性,通过 K-均值聚类将 MTS 子序列分为若干个类.根据异常模式的定义,从这若干个类中识别出异常模式.在2个实际数据集上进行实验,实验结果验证算法的有效性.

Multivariate time series （MTS） is widely available in many fields including finance, medicine, science and engineering . An approach for identifying outlier patterns in MTS is proposed . By using bottomrup segmentation algorithm, MTS is divided into non-overlapping subsequences. An extended Frobenius norm is used to compare the similarity between two MTS subsequences. K-means algorithm is employed to cluster MTS subsequences into some classes. According to the definitions of outlier patterns , the outlier patterns in MTS can be identified from the classes . Experiments are performed on two real-world datasets： stock market dataset and brain computer interface dataset. The experimental results show the effectiveness of the algorithm.