Adversarial Transformer-Based Anomaly Detection for Multivariate Time Series

Abstract

Anomaly detection in multivariate time series is crucial to monitor system status, such as fault detection in industrial systems. However, detecting anomalies in multivariate time series is challenging due to few labels, complex spatiotemporal correlations, and ultrafast detecting demands. Existing anomaly detection methods rarely address these challenges simultaneously. Herein, we design an adversarial transformers-based unsupervised anomaly detection model (ATUAD). In ATUAD, a Transformer-based encoder–decoder is constructed to learn sequence features, and adversarial training is adopted to amplify mild anomalies and enhance the robustness. Besides, we propose a peak-over-threshold-based dynamic threshold mechanism to improve the anomaly detection performance of ATUAD by automatically determining the threshold. In addition, we provide an anomaly explanation method to help ATUAD pinpoint root causes for anomalies. Comparison experiments, ablation studies, and overhead analysis on public datasets show that ATUAD can outperform the state-of-the-art baseline methods.