Scholarlink

<p>The volume of data from internet of things (IoT) systems has increased dramatically, so anomaly detection for multivariate time series data collected by IoT systems is becoming particularly important. However, due to the lack of anomaly labels and the dimensional complexity of the data, it is a challenge to build an effective detection model with high accuracy and strong robustness. This paper proposes an unsupervised anomaly detection model based on heterogeneous graph neural networks, Parallel Heterogeneous Graph Learning based Unsupervised Anomaly Detection and Explanation via Cross-channel Feature Fusion. The model achieves outstanding anomaly detection accuracy and generalization through two instances of cross-channel feature fusion. Firstly, in terms of nonlinear feature extraction, we utilize temporal convolutional network connected with squeeze-and-excitation network for temporal dependency extraction and enhancement of crucial features. We employ heterogeneous graph neural network and graph attention network to extract local and global features in parallel. After cross-channel feature fusion, Transformer is embedded for contextual information extraction. Secondly, after extracting linear features using vector autoregression, we perform secondary cross-channel feature fusion of both nonlinear and linear features, to achieve better robustness. Lastly, we incorporate an interpretation module. We compare the performance of the proposed model with 10 baseline methods on 5 public datasets. The experimental results show that the proposed model can achieve 0.930 in terms of F1-score, which is superior to ten state-of-the-art methods by 4.83%. Our source code is available at: <a href="https://github.com/xiqinghui/PUC.git">https://github.com/xiqinghui/PUC.git</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

Parallel heterogeneous graph learning based internet of things multivariate time series anomaly detection and explanation via cross-channel feature fusion

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Qinghui Xi,
  • Xi Li,
  • Peng Chen,
  • Juan Chen,
  • Xianhua Niu,
  • Lei Xu

摘要

The volume of data from internet of things (IoT) systems has increased dramatically, so anomaly detection for multivariate time series data collected by IoT systems is becoming particularly important. However, due to the lack of anomaly labels and the dimensional complexity of the data, it is a challenge to build an effective detection model with high accuracy and strong robustness. This paper proposes an unsupervised anomaly detection model based on heterogeneous graph neural networks, Parallel Heterogeneous Graph Learning based Unsupervised Anomaly Detection and Explanation via Cross-channel Feature Fusion. The model achieves outstanding anomaly detection accuracy and generalization through two instances of cross-channel feature fusion. Firstly, in terms of nonlinear feature extraction, we utilize temporal convolutional network connected with squeeze-and-excitation network for temporal dependency extraction and enhancement of crucial features. We employ heterogeneous graph neural network and graph attention network to extract local and global features in parallel. After cross-channel feature fusion, Transformer is embedded for contextual information extraction. Secondly, after extracting linear features using vector autoregression, we perform secondary cross-channel feature fusion of both nonlinear and linear features, to achieve better robustness. Lastly, we incorporate an interpretation module. We compare the performance of the proposed model with 10 baseline methods on 5 public datasets. The experimental results show that the proposed model can achieve 0.930 in terms of F1-score, which is superior to ten state-of-the-art methods by 4.83%. Our source code is available at: https://github.com/xiqinghui/PUC.git.