摘要
随着“天空地水工”一体化感知体系全面建设,引调水工程安全管理数据呈现出多源异构、规模庞大、动态变化复杂特征,传统的基于单模态数据分析、挖掘方法在工程安全智慧监管场景下面临明显的局限性。融合多模态大模型与知识图谱技术,提出一种“感知—认知—决策”的智能监管范式。基于标准规范、风险应急管理资料、巡检文本及图像、多光谱遥感影像,微调多模态大模型并结合动态提示策略,构建面向工程安全的多模态知识图谱;利用检索增强生成技术及知识图谱的结构化知识,提升大模型在专业领域的可靠性及推理能力;提出多智能体协同的决策链构建方法,通过动态任务编排实现模型能力耦合,赋能工程安全管理中的风险识别、评估及预案生成业务。实验结果表明,本研究方法的多模态知识提取准确性高,可支撑引调水工程安全智慧监管。
With the comprehensive development of the“sky-space-earth-water-project”integrated monitoring system,safety management data of water diversion projects exhibit characteristics of multi-source heterogeneity,large volume,and dynamic complexity.Traditional analysis and mining methods based on single-modality data face significant limitations in the context of intelligent safety supervision.By integrating multimodal large models with knowledge graph technology,an intelligent supervision paradigm of“perception-cognition-decision”is proposed.Based on standards and specifications,risk and emergency management materials,inspection texts and images,and multispectral remote sensing imagery,a multimodal large model is fine-tuned and combined with a dynamic prompting strategy to construct a multimodal knowledge graph for engineering safety.Retrieval augmented generation(RAG)and the structured knowledge within the knowledge graph are employed to enhance the model’s reliability and reasoning capability in specialized domains.A collaborative multi-agent decision chain construction method is introduced,enabling the coupling of model capabilities through dynamic task orchestration to support risk identification,assessment,and contingency planning in safety management.Experimental results show that the proposed method achieves high accuracy in multimodal knowledge extraction,providing effective support for intelligent safety supervision of water diversion projects.
作者
王立虎
刘雪梅
李海瑞
陈晓楠
WANG Lihu;LIU Xuemei;LI Hairui;CHEN Xiaonan
出处
《中国水利》
2025年第11期8-19,共12页
China Water Resources
基金
国家重点研发计划(2024YFC3210802)
国家自然科学基金项目(72271091)
2022年度水利部重大科技项目(SKS-2022029)
河南省科学院科技开放合作项目(220901008)
河南省高等教育重点研发项目(24A520021)。
关键词
智能监管
多模态大模型
多模态知识图谱
安全智慧监管
引调水工程
intelligent supervision
multimodal large model
multimodal knowledge graph
intelligent safety supervision
water diversion projects
