Self-Supervised Variational Graph Autoencoder for System-Level Anomaly Detection

Unsupervised anomaly detection (AD) methods, either reconstruction based or prediction based, determine anomalies based on residuals. Occasional mutations in a single variable can cause the residuals to exceed the limits. Indeed, such mutations are not variations in the operating mechanism of the system. Thus, system-level anomalies are challenging to characterize. Complex networks (or 'graphs') are well adapted for modeling and characterizing the laws of evolution of complex systems. However, most industrial scenarios are without graphs. Hence, a self-supervised variational graph autoencoders (SS-VGAE) method is proposed. First, the multisource sensor dynamic graph is constructed through detrended cross-correlation analysis (DCCA). Second, target and self-supervised learning tasks are designed. The target task is to reconstruct the input graph structure to minimize the reconstruction loss. The self-supervised task is to learn the optimal center of the hypersphere in the latent space such that the mean features are gathered toward the center as much as possible. Multitask joint optimization allows high- and low-dimensional space features to be considered simultaneously, thereby improving the reliability of anomaly scores. Then, the distribution of anomaly scores is calculated and integrated into a system health indicator (HI). The system HI is more applicable to assist decision making. Finally, the superiority of the proposed method, namely, better detection accuracy and robustness, is demonstrated by nuclear personal computer transient analyzer (PCTRAN) simulation data and Skoltech anomaly benchmark (SKAB) data. Last but not least, systematic anomalies are found to make the correlation between the variables stronger.