高级检索
    黄训华, 张凤斌, 樊好义, 席亮. 基于多模态对抗学习的无监督时间序列异常检测[J]. 计算机研究与发展, 2021, 58(8): 1655-1667. DOI: 10.7544/issn1000-1239.2021.20201037
    引用本文: 黄训华, 张凤斌, 樊好义, 席亮. 基于多模态对抗学习的无监督时间序列异常检测[J]. 计算机研究与发展, 2021, 58(8): 1655-1667. DOI: 10.7544/issn1000-1239.2021.20201037
    Huang Xunhua, Zhang Fengbin, Fan Haoyi, Xi Liang. Multimodal Adversarial Learning Based Unsupervised Time Series Anomaly Detection[J]. Journal of Computer Research and Development, 2021, 58(8): 1655-1667. DOI: 10.7544/issn1000-1239.2021.20201037
    Citation: Huang Xunhua, Zhang Fengbin, Fan Haoyi, Xi Liang. Multimodal Adversarial Learning Based Unsupervised Time Series Anomaly Detection[J]. Journal of Computer Research and Development, 2021, 58(8): 1655-1667. DOI: 10.7544/issn1000-1239.2021.20201037

    基于多模态对抗学习的无监督时间序列异常检测

    Multimodal Adversarial Learning Based Unsupervised Time Series Anomaly Detection

    • 摘要: 时间序列异常检测旨在发现对应时序特征中不符合一般规律的特异性模式,是机器学习领域重要的研究方向之一.然而,现有的时序异常检测方法大多为单模态学习,忽略了时序信息在多模态空间上不同特征分布的关联性和互补性,不能充分利用已有信息进行有效地模式挖掘,从而造成检测效果差等问题.为此,提出了一种基于多模态对抗学习的无监督时间序列异常检测模型.首先,将原始时间序列转换至频域空间,构造多模态时间序列表示.其次,提出多模态生成对抗网络模型,针对多模态时间序列,实现正常时序信息关于时域和频域特征分布的无监督联合学习.最后,通过将异常检测问题转化为时间序列在时域和频域空间的重构度量问题,从时域空间和频域空间2个方面度量时间序列的异常值,实现更有效的异常检测.在时间序列数据集合UCR和MIT-BIH中的6个真实数据集的实验结果表明,在异常检测任务上相较于传统单模态异常检测方法,提出方法在AUC和AP这2个性能指标上最高分别提升了12.50%和21.59%,证明了方法的有效性.

       

      Abstract: Time series anomaly detection is one of the most important research directions in machine learning, which aims to find the patterns that deviate significantly from the normal behavior of time series. However, most of the existing methods for anomaly detection of time series are based on single-modality feature learning, which ignores the relevance and complementarity of the characteristic distribution of time series in multi-modality space, and consequently fails to make full use of the existing information for learning. To alleviate the above problems, in this paper, we present a time series anomaly detection model based on multimodal adversarial learning. Firstly, we convert the original time series into the frequency domain to construct multi-modality time series representation. Then, based on the constructed multi-modality representation, we propose a multimodal generated adversarial network model to learn normal data’s distributions in time domain and frequency domain jointly. Finally, by modeling the anomaly detection problem as the data reconstruction problem in time domain and frequency domain, we measure the anomaly score of time series from both the time domain and frequency domain perspectives. We verify the proposed method on the time series data sets of UCR and MIT-BIH. Experimental results on the 6 data sets of UCR and MIT-BIH show that, compared with the state-of-the-arts, the proposed method improves the AUC and AP metrics of anomaly detection performance by 12.50% and 21.59% respectively.

       

    /

    返回文章
    返回