Scholarlink

<p>Multivariate time series anomaly detection is a critical task in various domains, including industrial systems, healthcare, and network monitoring. While Generative Adversarial Networks (GANs) have demonstrated significant potential due to their powerful unsupervised learning capabilities, existing models are often hindered by several challenges. These include mode collapse, which limits sample diversity; the presence of contaminated data, which weakens the discriminator’s ability to learn normal patterns; and a general neglect of frequency-domain features. To address these limitations, we propose the Exponential Informativeness-Based Bidirectional Memory GAN (EM-BMGAN). This model mitigates mode collapse by introducing an improved loss function based on exponential informativeness. It employs a Multi-Scale Temporal Memory Network (MSTMN) as its generator to suppress the reconstruction of anomalous patterns. Furthermore, EM-BMGAN incorporates a Multi-Scale Adaptive Discrete Cosine Transform Network (MADN) to achieve effective time-frequency feature fusion. Experimental results on multiple public datasets demonstrate that EM-BMGAN not only achieves superior performance but also possesses a more lightweight architecture compared to state-of-the-art Transformer-based models.</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

EM-BMGAN: Exponential informativeness-based bidirectional memory GAN model for multivariate time series anomaly detection

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Jun Tu,
  • Zhilin Zhang,
  • Jieling Wu,
  • Shupu Wu

摘要

Multivariate time series anomaly detection is a critical task in various domains, including industrial systems, healthcare, and network monitoring. While Generative Adversarial Networks (GANs) have demonstrated significant potential due to their powerful unsupervised learning capabilities, existing models are often hindered by several challenges. These include mode collapse, which limits sample diversity; the presence of contaminated data, which weakens the discriminator’s ability to learn normal patterns; and a general neglect of frequency-domain features. To address these limitations, we propose the Exponential Informativeness-Based Bidirectional Memory GAN (EM-BMGAN). This model mitigates mode collapse by introducing an improved loss function based on exponential informativeness. It employs a Multi-Scale Temporal Memory Network (MSTMN) as its generator to suppress the reconstruction of anomalous patterns. Furthermore, EM-BMGAN incorporates a Multi-Scale Adaptive Discrete Cosine Transform Network (MADN) to achieve effective time-frequency feature fusion. Experimental results on multiple public datasets demonstrate that EM-BMGAN not only achieves superior performance but also possesses a more lightweight architecture compared to state-of-the-art Transformer-based models.