首页|融合显隐式关联的图卷积多维时序异常检测方法

融合显隐式关联的图卷积多维时序异常检测方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
针对重构类多维时序异常检测方法对信息物理系统组件间的耦合关系提取能力不足以及建模时易遗漏信息的问题,提出一种融合显隐式关联的图卷积多维时序异常检测方法.利用改进的余弦相似度,提取序列间可用距离度量的显式关联关系.设计基于多头自注意力机制的关联关系提取模块,学习序列间的隐式关联关系.整合显隐式关系,将关系融合图和原始数据共同输入到基于图卷积网络的自编码器中,进行结合时间和空间依赖性的多维时序重构.根据训练好的模型输出的待测数据重构结果计算异常的分数,进而结合自适应阈值来选择算法进行异常检测.4个公开数据集上的实验结果表明,所提方法比相关经典和时效性方法在F1-Score上具有明显的提升,且可以通过输出关联权重矩阵的方式对异常事件进行解释分析.
Graph Convolutional Anomaly Detection Method for Multivariate Time Series With Fusion of Explicit and Implicit Associations
The existing limitations in extracting coupling relationships and information omission in reconstruction-based anomaly detection methods for cyber-physical systems are well known.Therefore,graph convolutional method based on explicit and implicit associations fusion for multivariate time series anomaly detection is proposed.The improved cosine similarity function extracts explicit associations measured by distance.An association extraction module employing the multi-head self-attention mechanism captures implicit associations.The fusion of two kinds of associations creates a graph inputted into a graph-convolutional auto-encoder with raw data for reconstructing time series,incorporating temporal and spatial dependencies.Anomaly scores derived from the reconstruction are used for detecting anomalies through the adaptive threshold selection algorithm.Experimental results on four openly available datasets have confirmed that the proposed method outperforms the state-of-the-art and latest methods in terms of F1-Score.The experimental results also demonstrate that the proposed method can interpret and analyze abnormal events by outputting the association weight matrix.

cyber-physical systemsmultivariate time series anomaly detectionmulti-head self-attention mechanismgraph convolutional networkexplicit and implicit associations extraction

张光耀、高欣、张云凯、刘婧、叶平

展开 >

北京邮电大学人工智能学院,北京 100876

中国电力科学研究院有限公司,北京 100192

南瑞集团有限公司,南京 211106

信息物理系统 多维时序异常检测 多头自注意力机制 图卷积神经网络 显隐式关联关系提取

国家电网有限公司科技项目

5700-202227226A-1-1-ZN

2024

10.13190/j.jbupt.2023-107

北京邮电大学学报
北京邮电大学

北京邮电大学学报

CSTPCD北大核心
影响因子：0.592
ISSN：1007-5321
年,卷(期)：2024.47(3)
  • 1