Scholarlink

<p>The exponential growth in volume and dimensionality of time series data has escalated anomaly detection from a conventional analytical task to a High-Performance Computing (HPC)-scale challenge. However, conventional reconstruction-based methods suffer from two key limitations: susceptibility to noise contamination in collected data and an over-reliance on contextual semantics at the expense of other vital time series characteristics. To address these challenges, we propose D3C–a novel deep learning framework integrating Diffusion models, Dual attention mechanisms, and Dual-view Contrastive representation learning. Our approach leverages a Transformer-based diffusion module to disentangle and reconstruct multivariate time series from frequency and trend perspectives, effectively denoising inputs and enhancing anomaly recall and robustness. The reconstructed features then pass through a dual-branch architecture with hierarchical attention and contrastive learning, explicitly amplifying normal-anomaly distinctions. Furthermore, we introduce a specialized loss function to mitigate reconstruction-induced information loss. Extensive experiments on five real-world benchmarks demonstrate that D3C achieves state-of-the-art performance, with strong robustness under heavy noise and practical potential for parallelized deployment on HPC platforms.</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

Toward robust anomaly detection in noisy time series via diffusion-driven denoising and disentanglement

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Xiaorui Huang,
  • Hancheng Xiao,
  • Zhixia Zeng,
  • Zhipeng Qiu,
  • Weifu Zhu,
  • Shi Zhang,
  • Ruliang Xiao

摘要

The exponential growth in volume and dimensionality of time series data has escalated anomaly detection from a conventional analytical task to a High-Performance Computing (HPC)-scale challenge. However, conventional reconstruction-based methods suffer from two key limitations: susceptibility to noise contamination in collected data and an over-reliance on contextual semantics at the expense of other vital time series characteristics. To address these challenges, we propose D3C–a novel deep learning framework integrating Diffusion models, Dual attention mechanisms, and Dual-view Contrastive representation learning. Our approach leverages a Transformer-based diffusion module to disentangle and reconstruct multivariate time series from frequency and trend perspectives, effectively denoising inputs and enhancing anomaly recall and robustness. The reconstructed features then pass through a dual-branch architecture with hierarchical attention and contrastive learning, explicitly amplifying normal-anomaly distinctions. Furthermore, we introduce a specialized loss function to mitigate reconstruction-induced information loss. Extensive experiments on five real-world benchmarks demonstrate that D3C achieves state-of-the-art performance, with strong robustness under heavy noise and practical potential for parallelized deployment on HPC platforms.