• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 人工智能与数据挖掘 • 上一篇    下一篇

一种基于Attention-GRU和iForest的周期性时间序列异常检测算法

王腾1,焦学伟2,高阳2   

  1. (1.中国电信股份有限公司江苏分公司,江苏  南京 210037;
    2.南京大学计算机科学与技术系,江苏 南京 210000)
  • 收稿日期:2019-04-26 修回日期:2019-06-20 出版日期:2019-12-25 发布日期:2019-12-25

An anomaly detection algorithm based on
Attention-GRU and iForest for periodic time series

WANG Teng1,JIAO Xue-wei2,GAO Yang2   

  1. (1.Jiangsu Branch,China Telecom Corporation Limited,Nanjing 210037;
    2.Department of Computer Science and Technology,Nanjing University,Nanjing 210000,China)
  • Received:2019-04-26 Revised:2019-06-20 Online:2019-12-25 Published:2019-12-25

摘要:

对数据中异常模式的检测(异常检测)是数据分析领域一个非常重要的研究方向,尤其对时间序列的异常检测是其中的一个难点。目前,关于时序数据异常检测的研究有很多,如利用滑动窗口、小波分析、概率图模型、循环神经网络等不同技术来进行检测,但是在处理问题时还存在或多或少的不足,无法保证实际工程中的实时效率和准确性。针对周期性时间序列异常检测问题,提出一种基于Attention-GRU和iForest的异常检测算法,根据带有注意力机制的循环神经网络构建模型,实现预测长序列数据,保证工程检测效率,并通过iForest建立正常数据波动区间,避免了使用假设检验造成的误差。经实践验证,该算法能够提高实际工程中的周期性时序数据异常检测效率,并有较好的召回率和准确率。

关键词: 时间序列, 周期序列, 异常检测, 循环神经网络, 孤立森林

Abstract:

Abnormal patterns detection in data (abnormal detection) is a very important research direction in the field of data analysis, and especially the anomaly detection of time series is one of the difficulties. At present, there are many researches on anomaly detection of time series data. Different techniques such as sliding window, wavelet analysis, Probabilistic Graphical Model (PGM), and Recurrent Neural Network (RNN) are used to detect the abnormal patterns. However, these techniques still have some deficiencies in processing this problem, which cannot guarantee real-time efficiency and accuracy in real engineering. In this paper, an anomaly detection algorithm based on Attention-GRU and iForest is proposed for periodic time series. The model is constructed based on Attention-GRU to predict long data sequence and ensure the detection efficiency in engineering. iForest is used to establish a normal data fluctuation range, avoiding the error caused by hypothesis testing. Practice verification shows that this model can improve the anomaly detection efficiency of periodic time series data in actual engineering, and has better recall rate and accuracy.


 

Key words: time series, periodic sequence, anomaly detection, recurrent neural network (RNN), iForest