Multivariate Time series Anomaly Detection based on reconstructed differences using Graph Attention Networks

Abstract

Today, many real-world applications generate the amount of multivariate time series data. Monitoring those data and detecting some meaningful events early is important. As one of those tasks, interest in anomaly detection has grown. In recent research, some authors conducted anomaly detection in multivariate time series data by using graph attention networks to capture relationships among series and timestamps respectively. And another author suggested some connections between timestamps called Spatio-temporal connections. In this paper, we combine two ideas jointly and propose another multivariate time series anomaly detection method using series differences between adjacent timestamps. By using the proposed method, we conduct anomaly detection on two public datasets and compare the performance with other models. Also, to check for the possibility of operation on the edge environment, we measure the throughput of our proposed method in the IoT edge gateway that has restricted resources.