An Outlier Suppression and Adversarial Learning Model for Anomaly Detection in Multivariate Time Series

TL;DR

This paper introduces AOST, a new model that improves anomaly detection in multivariate time series by combining adversarial learning and an outlier suppression mechanism.

Contribution

The novel AOST model integrates adversarial learning and outlier suppression in a Transformer framework for better anomaly detection.

Findings

AOST achieves an average F1 score of 88.74% on benchmark datasets.

The model outperforms existing state-of-the-art methods in anomaly detection.

The dual-decoder GAN architecture improves data distribution learning and robustness.

Abstract

Multivariate time series anomaly detection is a critical task in modern engineering, with applications spanning environmental monitoring, network security, and industrial systems. While reconstruction-based methods have shown promise, they often suffer from overfitting and fail to adequately distinguish between normal and anomalous data, limiting their generalization capabilities. To address these challenges, we propose the AOST model, which integrates adversarial learning with an outlier suppression mechanism within a Transformer framework. The model introduces an outlier suppression attention mechanism to enhance the distinction between normal and anomalous data points, thereby improving sensitivity to deviations. Additionally, a dual-decoder generative adversarial architecture is employed to enforce consistent data distribution learning, enhancing robustness and generalization. A novel anomaly scoring strategy based on longitudinal differences further refines detection accuracy. Extensive experiments on three public datasets—SWaT, WADI, SMAP, and PSM—demonstrate the model’s superior performance, achieving an average F1 score of 88.74%, which surpasses existing state-of-the-art methods. These results underscore the effectiveness of AOST in advancing multivariate time series anomaly detection.

Figures