Abstract: Time series anomaly detection is an important research topic in the field of time series analysis. Due to the complex spatiotemporal dependencies and randomness of multivariate time series in real industrial scenarios, many existing anomaly detection methods for single dependency modeling cannot effectively learn data features. In addition, ignoring frequency domain information will also lead to incomplete model feature representation. To address the above problems, a time series anomaly detection model FeGvL (Frequency-domain enhancement Graph-variational Learning) based on frequency domain enhanced graph variational learning networks was proposed. First, after the block operation, the dependency in the 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. The performance of the proposed model was verified on three public datasets. The F1 value of FeGvL was higher than that 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 was 1.7 percentage points higher than that of the second-best model GreLeN. Experimental results show that the proposed method can effectively capture spatiotemporal dependencies, provide representation capabilities, and has high anomaly detection accuracy.

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

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，忽略频域信息也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进行实体间的特征提取，结合时间依赖实现具有泛化性的变分重构。在3个公共数据集上验证所提模型的性能，FeGvL的F1值均高于GDN（Graph Deviation Network）、TranAD（Transformer-based Anomaly Detection）、GreLeN（Graph relational Learning Network）等7个先进的异常检测方法，平均F1值与次优模型GreLeN相比提高了1.7个百分点。实验结果表明，所提模型能够有效捕获时空依赖性，并且提供表征能力，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