Scholarlink

<p>Anomaly detection is crucial in time series analysis for identifying abnormal events. To address the limitations of traditional methods in integrating spatiotemporal correlations and modeling normal patterns, we propose a Time Series Anomaly Detection Model Based on Spatio-Temporal Feature Fusion (TADST). First, the Spatio-Temporal Feature Fusion Network (STF) combines temporal convolutional networks and graph attention influence networks to capture temporal dynamic dependencies and attribute correlations respectively, facilitating joint spatiotemporal feature modeling. Then, the Time Series Reconstruction Network (TSR) employs a multi-layer encoder-decoder architecture to learn the normal sample distribution and amplify discrepancies between reconstructed and anomalous data. Finally, the Anomaly Detection Mechanism (ADM) identifies anomalies by fitting the tail distribution of reconstruction deviations. When the anomaly score exceeds a predefined threshold, the mechanism updates the parameters of the Generalized Pareto Distribution, keeping the detection criteria adaptive. Experiments demonstrate that the proposed TADST achieves state-of-the-art results on five publicly available datasets.</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

TADST: reconstruction with spatio-temporal feature fusion for deviation-based time series anomaly detection

  • 期刊论文

  • 发表日期：

    2025年2月8日

查看全文
  • Bin Yang,
  • Tinghuai Ma,
  • Huan Rong,
  • Xuejian Huang,
  • Yubo Wang,
  • Bowen Zhao,
  • Chaoming Wang

摘要

Anomaly detection is crucial in time series analysis for identifying abnormal events. To address the limitations of traditional methods in integrating spatiotemporal correlations and modeling normal patterns, we propose a Time Series Anomaly Detection Model Based on Spatio-Temporal Feature Fusion (TADST). First, the Spatio-Temporal Feature Fusion Network (STF) combines temporal convolutional networks and graph attention influence networks to capture temporal dynamic dependencies and attribute correlations respectively, facilitating joint spatiotemporal feature modeling. Then, the Time Series Reconstruction Network (TSR) employs a multi-layer encoder-decoder architecture to learn the normal sample distribution and amplify discrepancies between reconstructed and anomalous data. Finally, the Anomaly Detection Mechanism (ADM) identifies anomalies by fitting the tail distribution of reconstruction deviations. When the anomaly score exceeds a predefined threshold, the mechanism updates the parameters of the Generalized Pareto Distribution, keeping the detection criteria adaptive. Experiments demonstrate that the proposed TADST achieves state-of-the-art results on five publicly available datasets.