计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (20): 122-127.DOI: 10.3778/j.issn.1002-8331.1706-0252

• 模式识别与人工智能 • 上一篇    下一篇

时间序列多尺度异常检测方法

陈  波1,刘厚泉1,赵志凯2   

  1. 1.中国矿业大学 计算机科学与技术学院,江苏 徐州 221116
    2.中国矿业大学 信息与控制工程学院 物联网感知矿山研究中心,江苏 徐州 221008
  • 出版日期:2018-10-15 发布日期:2018-10-19

Method of multi-scale anomaly detection in time series

CHEN Bo1, LIU Houquan1, ZHAO Zhikai2   

  1. 1.School of Computer Science and Technology, China University of Mining and Technology, Xuzhou, Jiangsu 221116, China
    2.IoT Perception Mine Research Center, School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, Jiangsu 221008, China
  • Online:2018-10-15 Published:2018-10-19

摘要: 针对时间序列异常模式检测大多采用线性模式分割方式的局限性,研究了在Haar小波变换多尺度特征的基础上,结合时间序列模式分割技术,提出一种时间序列多尺度异常检测方法。该方法首先通过小波变换压缩时间序列,把时间序列分解在不同的尺度上;再利用二次回归模型将分解后的时间序列分割成可变长度的模式序列,计算模式异常值;最后重构原时间序列,检测原时间序列中的异常模式。实验结果表明,该方法可以有效地检测异常,而且可以从不同分解级数的压缩时间序列中检测多尺度异常模式。

关键词: 时间序列, 小波变换, 二次回归模型, 多尺度, 异常模式

Abstract: The limitations of the linear pattern segmentation are mostly used for anomaly pattern detection of time series, a multi-scale anomaly detection method in time series based on multi-scale feature of Haar wavelet transform and pattern segmentation technique of time series is proposed. In this method, time series is compressed and decomposed on different scales by wavelet transform, and then time series of decomposed is segmented into variable length patterns by quadratic regression model. The anomaly value of all patterns are calculated. Finally, the original time series is reconstructed and the anomaly patterns in the original time series are detected. The experimental results show that this method can detect effectively the anomaly and multi-scale abnormal patterns can be detected from the compressed time series with different decomposition stages.

Key words: time series, wavelet transform, quadratic regression model, multi-scale, outlier pattern