Scholarlink

<p>Internet of Things (IoT) devices generate massive volumes of Iot time series data through sensors during continuous operation. The exponential growth in data volume has imposed increasingly stringent requirements on the real-time processing and analysis of this data. Particularly in high-load, high-speed data processing scenarios, the timely detection and handling of anomalies become crucial for maintaining system stability, making efficient anomaly detection a key task for ensuring system reliability. Additionally, many unsupervised deep learning methods often result in high false positive rates. To address these challenges, we propose an unsupervised anomaly detection model: dynamic transformer with denoising graph structure learning (DT-DGSL). DT-DGSL leverages the advantages of the Transformer parallel structure, utilizing the Anomaly Transformer variant and self-attention mechanism to independently learn the spatio-temporal characteristics of sensor data, which enables more accurate reconstruction of the original sequences. We also tackle subtle anomalies in complex datasets and false anomalies from environmental noise by introducing a layer of radial basis function (RBF) neurons and employing window-based dynamic graph structure learning, respectively, to reduce the model’s false positive rate. Experimental results on five public datasets show that DT-DGSL has improved the average <i>F</i>1 score by 20.29% and achieved state-of-the-art results on multiple datasets. Moreover, DT-DGSL demonstrates excellent real-time performance, which is suitable for IoT applications.</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

DT-DGSL: dynamic transformer using denoising graph structure learning for IoT time series anomaly detection

  • 期刊论文

  • 发表日期：

    2025年7月23日

查看全文
  • Peng You,
  • Xinxin Wang,
  • Peng Chen,
  • Lei Wu,
  • Juan Chen,
  • Xi Li,
  • Shengke Zeng,
  • Huangyining Gao

摘要

Internet of Things (IoT) devices generate massive volumes of Iot time series data through sensors during continuous operation. The exponential growth in data volume has imposed increasingly stringent requirements on the real-time processing and analysis of this data. Particularly in high-load, high-speed data processing scenarios, the timely detection and handling of anomalies become crucial for maintaining system stability, making efficient anomaly detection a key task for ensuring system reliability. Additionally, many unsupervised deep learning methods often result in high false positive rates. To address these challenges, we propose an unsupervised anomaly detection model: dynamic transformer with denoising graph structure learning (DT-DGSL). DT-DGSL leverages the advantages of the Transformer parallel structure, utilizing the Anomaly Transformer variant and self-attention mechanism to independently learn the spatio-temporal characteristics of sensor data, which enables more accurate reconstruction of the original sequences. We also tackle subtle anomalies in complex datasets and false anomalies from environmental noise by introducing a layer of radial basis function (RBF) neurons and employing window-based dynamic graph structure learning, respectively, to reduce the model’s false positive rate. Experimental results on five public datasets show that DT-DGSL has improved the average F1 score by 20.29% and achieved state-of-the-art results on multiple datasets. Moreover, DT-DGSL demonstrates excellent real-time performance, which is suitable for IoT applications.