Research on DUAL-ADGAN Model for Anomaly Detection Method in Time-Series Data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time-series anomaly detection is crucial in manufacturing and cybersecurity.
• Generative Adversarial Networks (GANs) show promise but face challenges like unstable training and detection inaccuracies.

Purpose of the Study:

• To propose DUAL-ADGAN (Dual Anomaly Detection Generative Adversarial Networks), a novel model to address limitations in GAN-based time-series anomaly detection.
• To enhance the stability, accuracy, and detection rate of anomaly detection in time-series data.

Main Methods:

• Utilized Wasserstein distance for stable generator training in the data reconstruction module.
• Incorporated a data prediction module to improve discrimination between normal and abnormal samples and reduce missed detections.
• Introduced Fence-GAN loss to align the discriminator with the anomaly detection target, minimizing false positives.

Main Results:

• Achieved an average F1 score of 0.881, surpassing nine other baseline models.
• Demonstrated improved training stability compared to traditional GAN networks.
• Effectively reduced both anomaly miss detection and false detection rates.

Conclusions:

• DUAL-ADGAN offers a more stable and accurate approach to time-series anomaly detection.
• The model successfully overcomes key challenges associated with GANs in anomaly detection tasks.
• DUAL-ADGAN provides a reliable solution for identifying anomalous data in time-series.