MTAD RF: Multivariate Time-series Anomaly Detection based on Reconstruction and Forecast

Abstract

Anomaly detection in multivariate time series is an important research direction, which helps to improve the security of industrial systems by detecting abnormally unreliable devices. Multivariate time series (MTS) anomalies not only need to pay attention to the time correlation between different time series but also need to consider the abnormal changes in the relationship between different variables. Once the influence relationship between two variables that influence each other is ignored, it will likely lead to false positives or false positives. At the same time, the degree of influence between different time series or different features is also inconsistent, just like what happened recently have radically different influences on the present. Furthermore, most of the existing models are weak in detecting no abnormality. To tackle these issues, in this paper, we propose a new model of multivariate time series anomaly detection based on reconstruction and forecast, named MTAD RF. First, we capture the temporal and feature correlations of MTS through two parallel GAT layers, and at the same time distinguish the influence degree between different time series or different features based on attention coefficients. Second, we leverage the generative power of VAE and the single-step forecast power of MLP to jointly detect known and unknown anomalies based on reconstructed and predicted models. Major practical implications of the proposed approach is missing. Finally, anomalies are detected and explained based on temporal and feature anomaly scores. Experiments demonstrate that our model outperforms current state-of-the-art methods on 4 real-world datasets, with an average F1 score of about 95% and excellent anomaly diagnostic ability.