《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (10): 3214-3220.DOI: 10.11772/j.issn.1001-9081.2024101438

• 数据科学与技术 • 上一篇    

基于频域增强图变分学习的时间序列异常检测

夏雨禾1,2, 王晓东1,2, 何启学1,2()   

  1. 1.中国科学院 成都计算机应用研究所,成都 610213
    2.中国科学院大学,北京 100049
  • 收稿日期:2024-10-12 修回日期:2024-12-09 接受日期:2024-12-13 发布日期:2024-12-23 出版日期:2025-10-10
  • 通讯作者: 何启学
  • 作者简介:夏雨禾(1999—),女,四川蓬溪人,硕士研究生,主要研究方向:异常检测、机器学习、大数据分析
    王晓东(1973—),男,四川乐山人,研究员,主要研究方向:网络工程
    何启学(1978—),男,四川富顺人,高级工程师,主要研究方向:数据挖掘、人工智能。Email:heqixue@casit.com.cn
  • 基金资助:
    四川省科技成果转移转化示范项目(2023ZHCG0005)

Time series anomaly detection based on frequency domain enhanced graph variational learning

Yuhe XIA1,2, Xiaodong WANG1,2, Qixue HE1,2()   

  1. 1.Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610213,China
    2.University of Chinese Academy of Sciences,Beijing 100049,China
  • Received:2024-10-12 Revised:2024-12-09 Accepted:2024-12-13 Online:2024-12-23 Published:2025-10-10
  • Contact: Qixue HE
  • About author:XIA Yuhe,born in 1999, M. S. candidate. Her research interestsinclude anomaly detection, machine learning, big data analysis.
    WANG Xiaodong, born in 1973, research fellow. His researchinterests include network engineering.
    HE Qixue,born in 1978, senior engineer. His research interestsinclude data mining, artificial intelligence.

摘要:

时间序列异常检测是时间序列分析领域的重要研究课题。由于现实工业场景中的多变量时间序列具有复杂的时空依赖性和随机性,现有许多针对单一依赖性建模的异常检测方法无法有效学习数据特征。此外,忽略频域信息会导致模型特征表示不全面。针对上述问题,提出基于频域增强图变分学习网络的时间序列异常检测模型FeGvL(Frequency-domain enhancement Graph-variational Learning)。首先,在分块操作后,通过自注意力建模时间维度上的依赖关系;其次,将频域增强后的图关系特征映射到潜在空间;最后,利用图聚合注意力网络提取实体间的特征,并结合时间依赖实现具有泛化性的变分重构。在PSM(Pooled Server Metrics)、SWaT(Secure Water Treatment)和WADI(WAter DIstribution)公共数据集上的实验结果表明,FeGvL的F1值均高于GDN(Graph Deviation Network)、TranAD(Transformer-based Anomaly Detection)和GReLeN(Graph Relational Learning Network)等7个先进的异常检测方法,且FeGvL的平均F1值与次优模型GReLeN相比提高了1.7个百分点。可见,所提模型能够有效捕获时空依赖性,提供表征能力,具有较高的异常检测精度。

关键词: 异常检测, 多元时间序列, 注意力机制, 变分自编码器, 频域增强

Abstract:

Time series anomaly detection is an important research topic in the field of time series analysis. Due to complex spatio-temporal dependencies and randomness of the multivariate time series in real industrial scenarios, many existing anomaly detection methods for single dependency modeling cannot learn data features effectively. In addition, ignoring frequency domain information will lead to incomplete model feature representation. To address the above problems, a time series anomaly detection model based on frequency domain enhanced graph variational learning network — FeGvL (Frequency-domain enhancement Graph-variational Learning) was proposed. Firstly, after the block operation, the dependency in time dimension was modeled by self-attention. Secondly, the graph relationship features after frequency domain enhancement were mapped to the latent space. Finally, the graph aggregation attention network was used to extract features between entities, and the temporal dependency was combined to achieve generalized variational reconstruction. Experimental results on public datasets PSM (Pooled Server Metrics), SWaT (Secure Water Treatment) and WADI (WAter DIstribution) show that the F1 value of FeGvL is higher than those of seven advanced anomaly detection methods such as GDN (Graph Deviation Network), TranAD (Transformer-based Anomaly Detection), and GReLeN (Graph relational Learning Network), and the average F1 value of FeGvL is 1.7 percentage points higher than that of the second-best model GReLeN. It can be seen that the proposed method can capture spatio-temporal dependencies effectively, provide representation capabilities, and has high anomaly detection accuracy.

Key words: anomaly detection, multivariate time series, attention mechanism, Variational AutoEncoder (VAE), frequency domain enhancement

中图分类号: