Scholarlink

<p>To ensure the safety of industrial systems and reduce downtimes, fault diagnosis must be accurate and timely. A graph neural network-based method introduced in this paper is referred to as the Knowledge Graph-based Exogenous Intervention Graph Neural Network (KG-EIGNN), which exploits sensor data and expert knowledge to better model how faults are determined and interpreted. The suggested solution will build a spatiotemporal graph using the data of multi-sensors and embody semantic relations particular to the area into a knowledge graph. It uses a knowledge-directed message-passing scheme and a hybrid learning rule to address data imbalance and enhance generalization: it combines supervised learning and contrastive learning. Many experiments on two real-world datasets allow concluding that KG-EIGNN outperforms baseline models when working with the Three-Phase Flow (TFF) and Nuclear Power System (NPS) datasets. KG-EIGNN recorded the best accuracy of 96.14% on TFF and 89.39% on NPS, beating other models like IAGNN (92.29% on TFF and 87.34% on NPS) and PKT-MCNN (88.58% on TFF and 86.50% on NPS). It further achieved an AUROC of 0.96 (TFF) as well as 0.89 (NPS), which showed strong capability of classification. The experimental outcomes validate the fact that KG-EIGNN would be scalable and intelligent in providing a solution to real-time fault diagnosis in a complicated industrial setting.</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

KG-EIGNN: a knowledge graph based graph neural network for fault detection in industrial systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月19日

查看全文
  • Wen Bo,
  • Chen Ju,
  • Jingjing Hu,
  • Qingyuan Yu,
  • Yan Zhang,
  • Changyou Zhang

摘要

To ensure the safety of industrial systems and reduce downtimes, fault diagnosis must be accurate and timely. A graph neural network-based method introduced in this paper is referred to as the Knowledge Graph-based Exogenous Intervention Graph Neural Network (KG-EIGNN), which exploits sensor data and expert knowledge to better model how faults are determined and interpreted. The suggested solution will build a spatiotemporal graph using the data of multi-sensors and embody semantic relations particular to the area into a knowledge graph. It uses a knowledge-directed message-passing scheme and a hybrid learning rule to address data imbalance and enhance generalization: it combines supervised learning and contrastive learning. Many experiments on two real-world datasets allow concluding that KG-EIGNN outperforms baseline models when working with the Three-Phase Flow (TFF) and Nuclear Power System (NPS) datasets. KG-EIGNN recorded the best accuracy of 96.14% on TFF and 89.39% on NPS, beating other models like IAGNN (92.29% on TFF and 87.34% on NPS) and PKT-MCNN (88.58% on TFF and 86.50% on NPS). It further achieved an AUROC of 0.96 (TFF) as well as 0.89 (NPS), which showed strong capability of classification. The experimental outcomes validate the fact that KG-EIGNN would be scalable and intelligent in providing a solution to real-time fault diagnosis in a complicated industrial setting.