基于图神经网络的时序信号异常检测方法  

A time series signal anomaly detection methodbased on graph neural network

作　　者：王婕婷 张泽珑 李飞江 钱宇华 WANG Jieting;ZHANG Zelong;LI Feijiang;QIAN Yuhua(Institute of Big Data Science and Industry,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing,Ministry of Education,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学大数据科学与产业研究院,山西太原030006 [2]山西大学计算智能与中文信息处理教育部重点实验室,山西太原030006

出　　处：《西北大学学报(自然科学版)》2025年第2期343-354,共12页Journal of Northwest University(Natural Science Edition)

基　　金：国家自然科学基金(62306170,62106132);山西省科技重大专项(202201020101006);山西省基础研究计划(202103021223026,20210302124271);山西省科技创新人才团队专项(202304051001001)。

摘　　要：高维时序数据异常检测是指从多元时间序列中识别出偏离整体模式或偏离预期行为的样本点的过程。在高维时序数据中,传感器间潜在的关联关系对于预测或检测任务的性能具有较大影响。图神经网络是一种基于节点的近邻关系学习节点表征的深度模型,能够有效建模传感器间的复杂关联。然而,现有基于图神经网络的异常检测方法大多依赖于单一的相似度度量来捕捉传感器间的关系,不能很好地学习传感器间的依赖关系。此外,在阈值选择时,现有方法使用正常数据中的最大异常得分作为切割阈值,限制了异常事件发生时的检测能力,从而造成较低的召回率。综上,提出了一种基于图神经网络的时序信号异常检测方法,根据传感器的特有特征使用多种相似度度量集成进行图结构学习;其次,将图结构学习方法与图神经网络相结合得到异常得分;最后,通过区间搜索法最优化F-measure指标寻找最优异常切割阈值。在两个真实传感器数据集上进行的实验表明,该方法比基准对比方法取得了较高的F 1值和召回率。Anomaly detection in high-dimensional time series data refers to the process of identifying sample points that deviate from the overall pattern or expected behavior from a multivariate time series.In high-dimensional time series data,the potential correlation between sensors has a significant impact on the performance of prediction or detection tasks.Graph neural networks are a deep model that learn node representations based on node proximity relationships,which can effectively model complex correlations between sensors.However,existing anomaly detection methods based on graph neural networks mostly rely on a single similarity measure to capture the relationships between variables and cannot learn the dependencies between variables well.In addition,when selecting thresholds,existing methods use the maximum anomaly score in normal data as the cutting threshold,which limits the detection ability when abnormal events occur,resulting in lower recall rates.In summary,this paper proposes a temporal signal anomaly detection method based on graph neural networks,which integrates multiple similarity measures based on the unique features of sensors for graph structure learning.Then,the graph structure learning method is combined with the graph neural network to obtain anomaly scores.Finally,the interval search method was used to optimize the F-measure index and find the optimal anomaly cutting threshold.Experiments on two real sensor datasets showed that our method achieved higher F 1 values and recall compared to the benchmark comparison method.

关 键 词：时序异常检测 图结构学习 注意力机制 相似度计算 区间搜索法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]