Variational Graph Attention Networks With Self-Supervised Learning for Multivariate Time Series Anomaly Detection

Cyber-physical systems (CPSs) are multidimensional and complex systems that integrate computing, networking, and physical environment, including various sensors that can produce vast multivariate time series data. Recently, scholars have proposed many unsupervised anomaly detection methods to monitor and alarm potential faults of CPSs. However, current methods still have some limitations. First of all, it is challenging to capture spatial-temporal relationships explicitly. In the second place, the anomaly detection model based on the variational autoencoder (VAE) may not use the latent variable and often falls into a kind of collapse, which is encouraged by the Kullback-Leibler (KL) divergence term. In the end, more information about the anomalies is needed. This article proposes a novel anomaly detection method called variational graph attention networks with self-supervised learning (VGATSL). We use stacked spatial-temporal graph attention (STGAT) networks to capture the temporal and feature correlations. Next, a VAE is constructed based on bidirectional gated recurrent units (BiGRUs), and the von Mises-Fisher (vMF) distribution is used as its latent space. Indeed, this distribution places latent representation on the surface of a unit hypersphere to allow for uninformed priors, thus preventing the model collapse. Additionally, self-supervised learning is introduced to simulate abnormal behaviors and to help the model form more explicit normality boundaries. Extensive experiments on three public real-world datasets demonstrate that VGATSL outperforms the baseline methods. We also illustrate that VGATSL has superior interpretability and flexibility.