TAGAN: multivariate time series anomaly detection algorithm with attention and generative adversarial network

Abstract

In this paper, a new anomaly detection architecture, TAGAN, is proposed. By combining the reconstruction approach with the prediction approach, TAGAN is used for anomaly detection over multivariate time series. A new loss function based on Wasserstein distance with gradient penalty is introduced in the reconstruction branch, and attention mechanism is introduced in the prediction branch. The performances of the proposed algorithm are tested over four real-world datasets (MSL, SMAP, SMD, and SWaT). Numerical experiments show that the proposed algorithm performs better than that of six anomaly detection algorithms.