Title: MTS-GAT: multivariate time series anomaly detection based on graph attention networks
Authors: Ling Chen; Yingchi Mao; Hongliang Zhou; Benteng Zhang; ZiCheng Wang; Jie Wu
Addresses: College of Computer and Information, Hohai University, Nanjing, Jiangsu, 211100, China ' Key Laboratory of Water Big Data Technology of Ministry of Water Resources, Hohai University, Nanjing, Jiangsu, 211100, China ' College of Computer and Information, Hohai University, Nanjing, Jiangsu, 211100, China ' College of Computer and Information, Hohai University, Nanjing, Jiangsu, 211100, China ' PowerChina Kunming Engineering Corporation Limited, Kunming, Yunnan, 650051, China ' Center of Networked Computing, Temple University, Philadelphia, PA 19122-6096, USA
Abstract: Anomaly detection using multivariate time series data from sensors can determine whether the equipment is operating normally. However, anomaly detection suffers from inadequate utilisation of spatio-temporal dependencies and unclear explanations of anomaly causes. To improve the accuracy of anomaly detection and rationalise the causes of anomalies, we propose multivariate time series anomaly detection based on graph attention networks (MTS-GAT). MTS-GAT constructs variable and temporal graphs using embedding vector similarity. The nonlinear dependencies of the variable and temporal dimensions are learned through two parallel graph attention layers. Finally, MTS-GAT jointly optimises the prediction-based and reconstruction-based models. Anomalous variables are localised with the anomaly scores computed after the joint optimisation to enhance the interpretability of anomaly detection. Experimental evaluations prove that MTS-GAT outperforms the best baseline approach, GDN. The F1 scores are improved by 2.73%, 3.39%, and 0.9% on SWaT, WADI, and SMD datasets.
Keywords: multivariate time series; anomaly detection; graph neural networks; attention mechanism.
DOI: 10.1504/IJSNET.2023.133812
International Journal of Sensor Networks, 2023 Vol.43 No.1, pp.38 - 49
Received: 22 Jun 2023
Accepted: 13 Jul 2023
Published online: 03 Oct 2023 *