Related Experiment Videos
STGAD: Self-temporal generative adversarial framework with transformer attention for unsupervised multivariate
Xiao Liao1, Wei Deng1, Hongyue Ma1
1State Grid Information and Telecommunication Group Co., Ltd., Beijing, China.
Plos One
|May 21, 2026
Summary
This study introduces STGAD, a novel framework for unsupervised anomaly detection in multivariate time series. It enhances system reliability by accurately identifying anomalies using a dual-score generative adversarial approach.
Area of Science:
- Cyber-physical systems
- Time series analysis
- Machine learning
Background:
- Unsupervised anomaly detection is crucial for cyber-physical systems.
- Existing methods struggle with adversarial stability, temporal modeling, and score calibration.
Purpose of the Study:
- To present STGAD, a dual-score generative-adversarial framework for anomaly detection and localization in multivariate time series.
- To address limitations in existing anomaly detection techniques.
Main Methods:
- STGAD utilizes a WGAN-GP critic with a Transformer encoder for temporal modeling.
- A stochastic generator models normal temporal patterns under adversarial supervision.
- Anomaly scores are derived from residuals and critic feedback, fused, and thresholded adaptively.
Main Results:
- STGAD demonstrates strong and consistent performance across five benchmark datasets (SMD, SMAP, MSL, SWaT, MIT-BIH).
- The framework shows effectiveness in server monitoring, aerospace telemetry, industrial control, and ECG signal analysis.
- Ablation studies confirm the benefits of temporal modeling, adversarial learning, and dual-score fusion.
Conclusions:
- STGAD offers an effective solution for unsupervised anomaly detection in multivariate time series.
- The proposed framework improves reliability in complex cyber-physical systems.
- The dual-score fusion and adaptive thresholding contribute to robust anomaly detection.