Scholarlink

<p>High-performance computing (HPC) systems rely on multivariate time series (MTS) anomaly detection to identify unusual or unexpected behaviors in computing nodes. Effective detection requires robust modeling of spatiotemporal dependencies to distinguish anomalies from noise. However, most existing methods fail to fully capture the changing relationships between different variables or the varying importance of each variable over time, leading to missed or false alarms. To address these challenges, this paper proposes DGAAD, a novel unsupervised anomaly detection framework for multivariate time series. DGAAD first employs a Pearson correlation prior to identify and retain statistically robust edges while filtering out noisy ones. Building upon this prior, it then computes cosine similarity between variable sequences within each time window and applies top-k selection to generate adaptive, window-specific dynamic graphs that capture inter-variable relationships. A graph attention network and a transformer-based reconstruction network are then jointly employed to learn robust representations of normal behavior. Experiments on four large real-world datasets show that DGAAD significantly outperforms previous state-of-the-art (SOTA) methods in key metrics.</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

Dgaad: a novel attention-based model for HPC anomaly detection

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Xu Gao,
  • Yibo Wang,
  • Hang Dong,
  • Hailiang Wang,
  • Zhenyu Li,
  • Xianliang Yang,
  • Jiandong Shang

摘要

High-performance computing (HPC) systems rely on multivariate time series (MTS) anomaly detection to identify unusual or unexpected behaviors in computing nodes. Effective detection requires robust modeling of spatiotemporal dependencies to distinguish anomalies from noise. However, most existing methods fail to fully capture the changing relationships between different variables or the varying importance of each variable over time, leading to missed or false alarms. To address these challenges, this paper proposes DGAAD, a novel unsupervised anomaly detection framework for multivariate time series. DGAAD first employs a Pearson correlation prior to identify and retain statistically robust edges while filtering out noisy ones. Building upon this prior, it then computes cosine similarity between variable sequences within each time window and applies top-k selection to generate adaptive, window-specific dynamic graphs that capture inter-variable relationships. A graph attention network and a transformer-based reconstruction network are then jointly employed to learn robust representations of normal behavior. Experiments on four large real-world datasets show that DGAAD significantly outperforms previous state-of-the-art (SOTA) methods in key metrics.