MST-GNN: A Multi-scale Temporal-Enhanced Graph Neural Network for Anomaly Detection in Multivariate Time Series

Abstract

Anomaly detection in time is an important task in many applications. Sensors are deployed in the industrial site to monitor the condition of different attributes or different places in real time, which generate multivariate time series. Recently, many methods were proposed to detect anomalies with multivariate time series, but they focused on the sequence attributes or spatial and temporal correlation, ignoring the characteristic of single sensor time series. In this paper, we propose a novel model MST-GNN that builds each sensor representation from Multi-Scale Temporal (MST) view, and use Graph Neural Network to mine their latent correlation to improve the performance of anomaly detection. In the MST representation, shapelets learning is introduced to extract its distinguishing features, a recurrent-skip neural network is used to extract the local temporal dependence relationship, and the raw data retains the original features of time series. These three features are fused to form the multi-scale temporal-enhanced features. Subsequently, the graph neural network is adopted to capture the potential interdependencies between multivariate time series and obtain the optimal representation of time series. Finally, bias assessment and anomaly detection are carried out. Extensive experiments on real-world datasets show that MST-GNN outperforms other state-of-the-art methods consistently, which provides an effective solution for anomaly detection in multivariate time series.