|
| 基于大语言模型的水利工程运行维护决策支持智能体构建 |
| Construction of large-language-model-based agents for decision support in the operation and maintenance of water conservancy projects |
| 投稿时间：2025-04-15 修订日期：2026-01-17 |
| DOI： |
| 中文关键词: 多模态大语言模型 智能体 水利工程巡检 图检索增强生成 |
| 英文关键词: multimodal large language model intelligent agent water conservancy project patrol inspection graph retrieval-augmented generation |
| 基金项目:国家重点研发计划资助(2024YFC3210800)；国家自然科学(72271091); 河南省科技厅科技攻关项目(252102210030); 河南省高等学校重点科研项目(25A520006) |
|
| 摘要点击次数: 35 |
| 全文下载次数: 0 |
| 中文摘要: |
| 水利工程巡检是保障重大基础设施安全稳定运行的核心管理任务。当前传统巡检模式存在人工依赖度高、隐患风险辨识准确率低、动态决策能力不足等突出问题。本文构建了融合大语言模型（LLMs）与图检索增强生成（GraphRAG）技术的多智能体协同智能决策框架，通过感知-记忆-通信-规划-行动的模块化架构实现全流程自动化巡检。研究采用近三年多源巡检数据构建多模态数据集，通过领域自适应微调，多模态大模型在设备识别和缺陷检测上F1值分别提升了7.2%和6.9%。进一步建立基于GraphRAG的动态知识图谱系统，通过知识注入技术弥补专业领域知识缺漏，同时利用实体关系推理机制有效抑制模型幻觉现象。实验结果表明，由该方法生成的水利巡检报告经过专家和运维人员的双重评估，能够准确体现水利运维领域的专业知识和技术深度。该研究为水利基础设施智能运维提供了可解释性强、可靠性高的新型技术范式，对推动行业数字化转型具有重要工程应用价值。 |
| 英文摘要: |
| The patrol inspection of water conservancy projects is a core management task for ensuring the safe and stable operation of critical infrastructures. Traditional patrol inspections suffer from high reliance on manual labor, low accuracy in hazard identification, and insufficient dynamic decision-making capabilities. This paper proposes a multi-agent collaborative intelligent decision-making framework that integrates Large Language Models (LLMs) and Graph Retrieval-Augmented Generation (GraphRAG) technologies. Adopting a modular architecture encompassing perception, memory, communication, planning, and action, the framework achieves full automation of the patrol inspection processes. A multi-modal dataset was constructed using multi-source patrol inspection data from the past three years. Domain-adaptive fine-tuning significantly improved the F1 scores of multi-modal large language models in equipment recognition and defect detection by 7.2% and 6.9%, respectively. Furthermore, a dynamic knowledge graph system based on Graph RAG was developed to bridge domain-specific knowledge gaps through knowledge infusion techniques, while simultaneously employing an entity-relation reasoning mechanisms to effectively mitigate model hallucinations. Experimental results demonstrate that patrol inspection reports generated by this method, upon dual evaluation by both domain experts and operational maintenance personnel, accurately reflect the professional expertise and technical depth required in operations and maintenance of water conservancy projects. This research provides a novel, interpretable, and reliable technical paradigm for intelligent operation and maintenance of water infrastructure, holding significant engineering application value for advancing digital transformation within the industry. |
|
查看/发表评论 下载PDF阅读器 |
| 关闭 |