A Dual-Discriminator Generative Adversarial Network for Anomaly Detection.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Multivariate time series anomaly detection is crucial across finance, aerospace, and security.
• Challenges include fuzzy anomaly definitions, complex patterns, and scarce abnormal data.
• Existing autoencoder (AE) and generative adversarial network (GAN) methods face issues like overfitting and data quality dependence.

Purpose of the Study:

• To propose a novel dual-discriminator GAN for enhanced time series anomaly detection.
• To address limitations of existing AE and GAN-based anomaly detection algorithms.
• To improve the practical deployment of GANs in anomaly detection tasks.

Main Methods:

• A novel GAN with a dual-discriminator structure is proposed.
• The generator produces reconstructions, and pseudo-labels categorize data based on reconstruction error.
• Two discriminators enforce distinct loss criteria for normal and potentially anomalous data reconstructions.

Main Results:

• The dual-discriminator GAN effectively constrains the generator, preserving normal data information while discarding anomalous data.
• Experimental results on benchmark datasets show superior performance compared to advanced anomaly detection methods.
• The model demonstrates strong performance on practical transformer data.

Conclusions:

• The proposed dual-discriminator GAN offers a robust solution for multivariate time series anomaly detection.
• This approach mitigates overfitting and reduces reliance on high-quality training data.
• The method shows significant potential for real-world applications, including transformer data analysis.