Scholarlink

<p>Knowledge graphs (KGs) are increasingly applied in hydropower communication security, with a focus on equipment maintenance and risk identification. However, studies published after 2023 have reported notable limitations. Existing hybrid approaches also exhibit flaws, among which shallow collaboration between their architectural and pre-training components is a key issue. These limitations collectively impair their ability to meet professional requirements. To address these issues, we construct a hybrid model that combines a mixture-of-experts (MoE) architecture with domain-adaptive pre-training (DAP). A customized pre-training task is designed to capture the characteristics of hydropower communication security, integrating entity alignment and vulnerability classification as subtasks. An expert division mechanism is also developed. Together, these two components form a deep collaborative system, improving both reasoning accuracy and runtime efficiency. Two key innovations are incorporated: (i) a domain-knowledge-enhanced pre-training strategy, which embeds hydropower communication protocol specifications into the pre-training process; and (ii) a dynamic expert allocation mechanism, which adjusts expert activation based on input data features. A multi-dimensional experimental verification system is established to evaluate model performance. The experimental results are explicit: the comprehensive performance score reaches 91.8, entity resolution (ER) accuracy hits 92.3% − 3.4% higher than that of combined KG baselines - runtime efficiency improves by 45%, the average robustness score is 0.870, and the performance decay rate is only 5.2%. DAP contributes 9.4% to overall performance. Real-world maintenance scenario tests show promising results: hazard inference accuracy (IRAcc) is 93%, entity annotation F1 (EA-F1) is 88%, and the scene adaptation coefficient is 0.98. These outcomes confirm effective knowledge integration and efficient reasoning. The model bridges the gap between KG modeling and practical hydropower communication security applications, reduces deployment costs, and effectively supports intelligent maintenance in this domain.</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

Construction of a knowledge mapping model for hydropower communication security based on MoE architecture and domain-adaptive pre-training

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月17日

查看全文
  • Yaqian Qiang,
  • Yu Pan,
  • Minghui Liu,
  • Wen Zhan

摘要

Knowledge graphs (KGs) are increasingly applied in hydropower communication security, with a focus on equipment maintenance and risk identification. However, studies published after 2023 have reported notable limitations. Existing hybrid approaches also exhibit flaws, among which shallow collaboration between their architectural and pre-training components is a key issue. These limitations collectively impair their ability to meet professional requirements. To address these issues, we construct a hybrid model that combines a mixture-of-experts (MoE) architecture with domain-adaptive pre-training (DAP). A customized pre-training task is designed to capture the characteristics of hydropower communication security, integrating entity alignment and vulnerability classification as subtasks. An expert division mechanism is also developed. Together, these two components form a deep collaborative system, improving both reasoning accuracy and runtime efficiency. Two key innovations are incorporated: (i) a domain-knowledge-enhanced pre-training strategy, which embeds hydropower communication protocol specifications into the pre-training process; and (ii) a dynamic expert allocation mechanism, which adjusts expert activation based on input data features. A multi-dimensional experimental verification system is established to evaluate model performance. The experimental results are explicit: the comprehensive performance score reaches 91.8, entity resolution (ER) accuracy hits 92.3% − 3.4% higher than that of combined KG baselines - runtime efficiency improves by 45%, the average robustness score is 0.870, and the performance decay rate is only 5.2%. DAP contributes 9.4% to overall performance. Real-world maintenance scenario tests show promising results: hazard inference accuracy (IRAcc) is 93%, entity annotation F1 (EA-F1) is 88%, and the scene adaptation coefficient is 0.98. These outcomes confirm effective knowledge integration and efficient reasoning. The model bridges the gap between KG modeling and practical hydropower communication security applications, reduces deployment costs, and effectively supports intelligent maintenance in this domain.