TimeAutoAD: Autonomous Anomaly Detection With Self-Supervised Contrastive Loss for Multivariate Time Series

Abstract

Multivariate time series (MTS) data are becoming increasingly ubiquitous in networked systems, e.g., IoT systems and 5G networks. Anomaly detection in MTS refers to identifying time series which exhibit different behaviors from normal status. Building such a system, however, is challenging due to a few reasons: i) labels for anomaly cases are usually unavailable or very rare; ii) most existing approaches rely on manual model-design and hyperparameter tuning, which may cost a huge amount of labor effort. To this end, we propose an autonomous anomaly detection technique for multivariate time series data (TimeAutoAD) based on a novel self-supervised contrastive loss. Specifically, we first present an automatic anomaly detection pipeline to optimize the model configuration and hyperparameters automatically. Next, we introduce three different strategies to augment the training data for generating pseudo negative time series and employ a self-supervised contrastive loss to distinguish the original time series and the generated time series. In this way, the representation learning capability of TimeAutoAD can be greatly enhanced and the anomaly detection performance can thus be improved. Extensive empirical studies on real-world datasets demonstrate that the proposed TimeAutoAD not only outperforms state-of-the-art anomaly detection approaches but also exhibits robustness when training data are contaminated.