AttVAE: A Novel Anomaly Detection Framework for Multivariate Time Series

Yi Liu, Yanni Han, Wei An

Abstract: Anomaly detection plays a significant role in building a secure and reliable system. Multivariate time series contain important system information, such as system load and time delay. Temporal-dependent methods like RNNs are usually used for anomaly detection on time series. However, inner correlations of time series have shown great potentials in anomaly detection than temporal-dependent methods. In this paper, we propose a novel anomaly detection framework, namely AttVAE, which utilizes attention mechanisms on multivariate time series. This attention mechanism exploits the inner correlations of different time-series dimensions to discover the robust latent variables. Extensive experiments are conducted on two real-world datasets, and results show that AttVAE achieves the best F1-score at 0.79 and 0.97 compared with existing traditional and sophisticated methods. In addition, the missing and false alarm rate by AttVAE is reduced by \(20\%\) on average compared with the state-of-the-art models.