2024 VOLUME 4
RESEARCH ARTICLE
Time series anomaly detection based on GAN-VAE
AUTHOR
Shibo Liao, Chunshan Liu*, Yongxiang Xia, Haicheng Tu
ABSTRACT
Time-series anomaly detection is crucial for identifying unusual patterns that deviate from expected behavior in temporal data, enabling early intervention in diverse fields such as finance, healthcare, and cybersecurity. This paper proposes a novel anomaly detection method based on deep learning models including Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). The proposed GAN-VAE model combines the strengths of GAN and VAE to effectively capture the distribution of time series data and optimize sequence mapping in latent space, achieving high accuracy in the reconstruction of normal time-series. Anomalies can then be detected by identifying abnormally large re-construction errors. To enhance the convergence of the GAN-VAE model in training, a se-quential training method is proposed that trains the encoder, decoder, and discriminator in an alternating fashion. The effectiveness of the proposed anomaly detection method is verified through real-world time series datasets..
KEYWORDS
Anomaly detection; GAN-VAE; Sequential training