Scholarlink

<p>Multivariate time series anomaly detection remains a critical challenge in domains requiring high-dimensional temporal pattern recognition, from industrial IoT to financial systems. While deep learning approaches like Autoencoders (AEs) and Generative Adversarial Networks (GANs) have shown promise, they exhibit complementary weaknesses. AEs often tend to overfit, while GANs frequently encounter mode collapse and unstable training dynamics. We present SAE-GAN, an integrated framework that combines a stacked attention autoencoder with adversarial training to overcome these limitations. The proposed architecture introduces three key innovations: (1) hierarchical attention mechanisms for simultaneous temporal and cross-feature dependency modeling, (2) adversarial regularization to prevent overfitting while maintaining reconstruction fidelity, and (3) a stabilized training mechanism that balances generator-discriminator dynamics. Extensive evaluations on three benchmark multivariate time series datasets demonstrate that SAE-GAN consistently outperforms state-of-the-art methods in both detection accuracy and computational efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SAE-GAN: integrating stacked attention autoencoder and generative adversarial networks for multivariate time series anomaly detection

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Shimin Sun,
  • Xiangyun Liu,
  • Li Han,
  • Ze Wang,
  • Yong Zhu

摘要

Multivariate time series anomaly detection remains a critical challenge in domains requiring high-dimensional temporal pattern recognition, from industrial IoT to financial systems. While deep learning approaches like Autoencoders (AEs) and Generative Adversarial Networks (GANs) have shown promise, they exhibit complementary weaknesses. AEs often tend to overfit, while GANs frequently encounter mode collapse and unstable training dynamics. We present SAE-GAN, an integrated framework that combines a stacked attention autoencoder with adversarial training to overcome these limitations. The proposed architecture introduces three key innovations: (1) hierarchical attention mechanisms for simultaneous temporal and cross-feature dependency modeling, (2) adversarial regularization to prevent overfitting while maintaining reconstruction fidelity, and (3) a stabilized training mechanism that balances generator-discriminator dynamics. Extensive evaluations on three benchmark multivariate time series datasets demonstrate that SAE-GAN consistently outperforms state-of-the-art methods in both detection accuracy and computational efficiency.