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

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Reconstruction-based methods for multivariate time series anomaly detection often struggle with overfitting and generalization.
• Existing models face challenges in effectively distinguishing normal from anomalous data points.

Purpose of the Study:

• To propose the AOST model for improved multivariate time series anomaly detection.
• To enhance the distinction between normal and anomalous data and improve model robustness and generalization.

Main Methods:

• The AOST model utilizes a Transformer framework with an integrated outlier suppression attention mechanism.
• A dual-decoder generative adversarial architecture is employed for consistent data distribution learning.
• A novel anomaly scoring strategy based on longitudinal differences is introduced.

Main Results:

• The AOST model demonstrated superior performance on SWaT, WADI, SMAP, and PSM datasets.
• Achieved an average F1 score of 88.74%, outperforming existing state-of-the-art methods.
• The outlier suppression attention and dual-decoder architecture enhanced sensitivity and robustness.

Conclusions:

• The proposed AOST model effectively advances multivariate time series anomaly detection.
• The integration of adversarial learning and outlier suppression offers significant improvements in accuracy and generalization.
• AOST provides a robust solution for critical engineering applications requiring reliable anomaly detection.