融合稀疏图注意力的多元时间序列异常检测方法  

Multivariate time series anomaly detection method with fusion of ProbSparse graph attention

作　　者：衡红军[1] 代栋炜 HENG Hong-jun;DAI Dong-wei(School of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学计算机科学与技术学院,天津300300

出　　处：《计算机工程与设计》2025年第3期841-849,共9页Computer Engineering and Design

基　　金：国家自然科学基金联合基金项目(U1333109)。

摘　　要：为解决时序数据中时空依赖关系不明确而导致多元时间序列异常检测效果较差的问题,提出一种基于稀疏图注意力网络的异常检测模型PSGAT-AD(ProbSparse graph attention networks anomaly detection)。采用卷积神经网络(convolutional neural networks,CNN)提取时间戳上下文信息并使用全局时间戳编码和Transformer位置编码增强序列之间的联系。利用稀疏自注意力关注重要的时间戳与特征,通过自注意力蒸馏(self-attention distillation)降低输入规模,使重要的特征更加突出,以学习时间和特征两个维度的复杂依赖关系,提升表示学习质量。通过构建基于预测和重构的综合损失函数,对模型参数进行优化。将综合损失误差作为异常得分实现异常判定。实验结果表明,PSGAT-AD模型在4个公开数据集上的F1得分提升1.47%~6.52%。To solve the problem of unclear spatiotemporal dependence in time series data,which leads to poor anomaly detection results in multivariate time series,an anomaly detection model PSGAT-AD(ProbSparse graph attention networks anomaly detection) based on sparse graph attention networks was proposed.Convolutional neural networks(CNN) were used to extract timestamp context information.The global timestamp embedding and Transformer position embedding were used to enhance the connection between sequences.The sparse self-attention was used to focus on important timestamps and features,and the input size was reduced through self-attention distillation(self-attention distillation) to make important features more prominent to learn the complex dependence of the two dimensions of time and features.Improvement represented the quality of learning.Model parameters were optimized by constructing a joint loss function based on prediction and reconstruction.The comprehensive loss error was used as an abnormality score to realize abnormality determination.Experimental results show that the F1 score of the PSGAT-AD model on four public data sets is increased by 1.47%-6.52%.

关 键 词：异常检测 多元时间序列 图注意力网络 时间戳编码 稀疏自注意力 自注意力蒸馏 综合损失误差 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]