关键词: 多变量时间序列, 无监督学习, 图注意力网络, 自注意力, 异常检测

Abstract: In the task of anomaly detection for multivariate time series, the correlations between variables are complex, and such correlation is difficult to be clearly learned by traditional anomaly detection methods. In addition, the correlation between variables is only considered by most models, while time dependencies are not sufficiently addressed. Therefore, a time series anomaly detection model based on High-order Feature Aggregation (HFA) was proposed. Initially, a variable relationship graph was constructed through graph structure learning. Subsequently, the traditional Graph ATtention network (GAT) framework was enhanced by incorporating higher-order neighbor node correlations, enabling more accurate modeling of complex inter-variable relationships. Finally, temporal dependencies within sequences were further captured through the integration of one dimensional convolutional operations with self-attention mechanisms. Experimental evaluations conducted on four public datasets demonstrate that compared with the suboptimal baseline model Anomaly Transformer, the precision of the HFA method is increased by 3.92% on average, and the F1 score is improved by 1.34% on average. Compared with the current mainstream baseline method TopoGDN, the precision of the HFA method is increased by 6.05% on average, and the F1 score is improved by 9.05% on average. The results of the ablation experiment further verify the effectiveness of each module in the model.

Key words: multivariate time series, unsupervised learning, graph attention network, self-attention, anomaly detection