Scholarlink

<p>Modern industrial cyber-physical systems, IoT infrastructures, and intelligent monitoring platforms continuously generate high-dimensional multivariate time-series streams at large scale, imposing requirements on effective anomaly detection. However, current anomaly detection methods still struggle to capture local anomaly signals in multivariate time series, which are often weak and easily masked by noise. This limitation leads to several challenges: Existing methods have limited capability to detect localized anomalies; anomalies with small magnitudes or short-term abrupt changes tend to be overlooked; and in high-noise environments, the model may confuse noise with true anomalies. To address these challenges, we propose a <Emphasis Type="Underline">Dual</Emphasis>-branch collaborative framework driven by <Emphasis Type="Underline">P</Emphasis>robabilistic <Emphasis Type="Underline">S</Emphasis>parse Attention and <Emphasis Type="Underline">C</Emphasis>omposite <Emphasis Type="Underline">P</Emphasis>rior Distribution (Dual-PSCP). First, the framework employs a probabilistic sparse (ProbSparse) attention mechanism to model global dependencies over long sequences, focusing on the more important time-series features and reducing interference from redundant and noisy information. Meanwhile, it uses statistical distribution modeling to characterize the local fluctuation properties of the time series, amplifying potential anomaly signals from a statistical perspective. Then, by computing a deviation metric between the two branches, we combine this result with the reconstruction error to measure the degree of deviation of the series from the normal pattern. Finally, evaluation experiments on five public datasets show an average F1-score of 0.949, demonstrating the superiority of our method, and an efficiency comparison demonstrates that it is suitable for deployment in computationally intensive monitoring systems. The code is available at: <a href="https://github.com/IntelligentComp-Lab/Dual-PSCP">https://github.com/IntelligentComp-Lab/Dual-PSCP</a>.</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

Dual-PSCP: dual-branch time-series anomaly detection via probabilistic sparse attention and composite prior distributions

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Yishui Wang,
  • Jiayi Lu,
  • Bin Zhou,
  • Yin Zhang,
  • Xun Li,
  • Jie Yuan,
  • Lidong Wang

摘要

Modern industrial cyber-physical systems, IoT infrastructures, and intelligent monitoring platforms continuously generate high-dimensional multivariate time-series streams at large scale, imposing requirements on effective anomaly detection. However, current anomaly detection methods still struggle to capture local anomaly signals in multivariate time series, which are often weak and easily masked by noise. This limitation leads to several challenges: Existing methods have limited capability to detect localized anomalies; anomalies with small magnitudes or short-term abrupt changes tend to be overlooked; and in high-noise environments, the model may confuse noise with true anomalies. To address these challenges, we propose a Dual-branch collaborative framework driven by Probabilistic Sparse Attention and Composite Prior Distribution (Dual-PSCP). First, the framework employs a probabilistic sparse (ProbSparse) attention mechanism to model global dependencies over long sequences, focusing on the more important time-series features and reducing interference from redundant and noisy information. Meanwhile, it uses statistical distribution modeling to characterize the local fluctuation properties of the time series, amplifying potential anomaly signals from a statistical perspective. Then, by computing a deviation metric between the two branches, we combine this result with the reconstruction error to measure the degree of deviation of the series from the normal pattern. Finally, evaluation experiments on five public datasets show an average F1-score of 0.949, demonstrating the superiority of our method, and an efficiency comparison demonstrates that it is suitable for deployment in computationally intensive monitoring systems. The code is available at: https://github.com/IntelligentComp-Lab/Dual-PSCP.