基于知识图谱增强的水库大坝安全监测智能问答技术研究OA
Knowledge graph-enhanced intelligent question answering technology for reservoir dam safety monitoring:a case study of the"Thousand Hands"large model for reservoir dam safety
针对大坝安全监测问答中意图识别不准、知识检索分散、回答依据不足等问题,提出一种基于灰狼优化算法(Grey Wolf Optimizer,GWO)、多模型融合与知识图谱增强的大语言模型(Large Language Model,LLM)问答技术.首先,构建大坝安全监测知识图谱本体,建立"监测对象—监测指标—判别依据—病害机制—处置措施"知识组织模式,涵盖监测指标、测点对象、病害机制、规范条文和处置措施等实体及其关系;其次,设计面向业务场景的18类问答意图体系,采用"提问对象_提问目的"的统一命名方式,支撑定义解释、阈值查询、异常判别、原因分析、风险分级和处置建议等任务;然后,融合规则模型、BERT/SVM模型和LLM提示模型进行意图识别,并采用GWO求解最优融合权重,以提高复杂业务场景下意图识别的准确性和稳定性;最后,依据识别意图调用知识图谱查询模板,检索监测指标、病害机制、规范条文和处置措施等知识,并结合检索结果由大语言模型生成包含结论、依据和建议的可追溯回答.在所构建的大坝安全监测问答数据集上的实验结果表明,GWO融合模型的答案准确率和F1值分别达到0.925和0.923,问答准确率和可追溯率分别达到0.928和0.906.结果表明,该方法能够有效提升意图识别准确性、知识检索针对性和回答可追溯性,可为水库大坝安全"千手"大模型的智能问答、异常研判和预警处置提供技术支撑.
Aiming at the problems of inaccurate intention recognition,scattered knowledge retrieval,and insufficient supporting evidence in dam safety monitoring question answering,a large language model(LLM)question answering method based on grey wolf optimization(GWO),multi-model fusion,and knowledge graph enhancement was proposed.First,a domain knowledge graph for dam safety monitoring was constructed,and a knowledge organization model of"monitoring object-monitoring indicator-discrimination basis-failure mechanism-disposal measure"was established,covering monitoring indicators,monitoring point objects,failure mechanisms,regulatory provisions,disposal measures,and their relationships.Second,an 18-category business-scenario-oriented question-answering intent system was designed using the unified naming scheme of"question object_intention purpose",supporting definition and interpretation,threshold queries,anomaly discrimination,cause analysis,risk classification,and disposal recommendations.Then,a rule-based model,a BERT/SVM model,and an LLM prompting model are fused for intention recognition,and the GWO is employed to search for the optimal fusion weights,thereby improving the accuracy and stability of intention recognition in complex business scenarios.Finally,according to the identified intention,the corresponding knowledge graph query template is invoked to retrieve monitoring indicators,failure mechanisms,regulatory provisions,and disposal measures,and the LLM combines the retrieved knowledge to generate traceable answers containing conclusions,supporting evidence,and recommendations.Experimental results on the constructed dam safety monitoring question-answering dataset show that the GWO-based fusion model achieves an answer accuracy of 0.925 and an F1-score of 0.923,while the question-answering accuracy and traceability rate reach 0.928 and 0.906,respectively.The results indicate that the proposed method can effectively improve intention recognition accuracy,the relevance of knowledge retrieval,and answer traceability,and can provide technical support for intelligent question answering,anomaly diagnosis,and early-warning decision-making in the"Thousand Hands"large model for reservoir dam safety.
李双平;杨晨雨;张斌;刘祖强;郑俊星;张辛;管林杰;叶果;汤寒;李永华
长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉华中科技大学土木与水利工程学院,430074,武汉||武汉喻见时空智能科技有限公司,430074,武汉长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉华中科技大学土木与水利工程学院,430074,武汉||武汉喻见时空智能科技有限公司,430074,武汉长江空间信息技术工程有限公司(武汉),430010,武汉||湖北省水利信息感知与大数据工程技术研究中心,430010,武汉||国家大坝安全工程技术研究中心,430010,武汉
建筑与水利
大坝安全监测意图识别知识图谱大语言模型灰狼优化算法多模型融合智能问答
dam safety monitoringintention recognitionknowledge graphlarge language modelgrey wolf optimizermulti-model fusionintelligent question answering
《中国水利》 2026 (18)
50-60,11
湖北省科技计划项目(2025BEB025).
评论