Spatial-Temporal Graph Neural Network Based Anomaly Detection

Multivariate time series anomaly detection is an important task in the monitoring system. In practical applications, an efficient and accurate anomaly detection method is particularly important. Recently, the method of anomaly detection based on prediction has made significant progress, but there are still limitations. This paper proposes a paradigm for multivariate time series anomaly identification based on pre-training. The strategy of pre-training is to use Transformer’s encoder to learn the dense vector representation of multiple time series through autoregressive task, so as to enhance the predictability of time series. In the prediction module, we learn the feature dependence of time series through graph attention network, and design an interactive tree structure that takes full advantage of the unique characteristics of time series to capture its time dependence. In addition, our method is well interpretable and allows users to infer the root cause of exceptions. We have proved the effectiveness of our model through extensive experiments. It is significantly superior to the most advanced model in three real data sets.