中国水利

针对水利设施巡检中人工模式效率低、覆盖有限、安全风险高等问题，研发了一套基于多模态大模型的水利场景无人值守智能巡检系统。系统以星逻智能MkY2智能机库与星祺行业级无人机为核心硬件，构建了“感知-执行-调度-平台-智能”五层架构。在智能感知层面，设计了“1个超级大脑+N个领域专家+知识库”的异构协同驱动机制：以多模态大语言模型作为全局语义解析中枢，赋予系统场景理解与模糊决策能力；以坝体裂隙提取、渗漏辨识、水位读取、漂浮物归类等轻量化垂直模型作为精准执行单元；以结构化水利知识库提供可追溯的专业解释依据。该机制打通了通用模型语义泛化与专用模型精细度量之间的技术壁垒，推动监测逻辑从“目标检测”向“态势认知”跃升。系统同时集成了基于改进遗传算法的多机协同调度模块与三层递进式航线规划框架，实现了7×24 h无人值守作业。在水库大坝安全监控与河道常态化巡查两类场景中的试点应用表明：对比传统人工巡检，无人机智能巡检作业效率提升167%，空间覆盖范围从约60%扩大至全域无死角，潜在隐患识别成功率从70%提升至95%以上，作业区安全事件发生概率降低90%，综合运维投入下降约50%，为智慧水利体系建设提供了一项融合通用人工智能与垂直领域知识的新型技术。

To address the limitations of manual inspection in water conservancy facilities, including low efficiency, limited coverage, and high safety risks, this study developed an unattended intelligent inspection system based on multimodal large models for water conservancy scenarios. The system was built around the Skysys MkY2 intelligent hangar and the Xingqi industry-grade drone as core hardware, adopting a five-layer architecture of “perception, execution, scheduling, platform, and intelligence”. At the intelligent perception level, a heterogeneous collaborative driving mechanism consisting of “one super brain + N domain experts + knowledge base” was designed: a multimodal large language model served as the global semantic parsing hub, endowing the system with scene understanding and fuzzy decision-making capabilities; lightweight vertical models(e.g., dam crack extraction, seepage identification, water level reading, and floating object classification) acted as precision execution units; a structured water conservancy knowledge base provided traceable professional explanations. This mechanism bridged the technical gap between the semantic generalization of general-purpose models and the fine-grained metrics of specialized models, elevating the monitoring logic from “target detection” to “situational awareness”. The system also integrated a multi-drone collaborative scheduling module based on an improved genetic algorithm and a three-tier progressive flight planning framework, enabling 7 × 24 unattended operation. Pilot tests in two types of scenarios of reservoir dam safety monitoring and routine river patrols demonstrate that compared with traditional manual inspection, intelligent drone inspection improves efficiency by 167%, expands spatial coverage from about 60% to full area without blind spots, increases the success rate of potential hazard identification from 70% to over 95%, reduces the probability of safety incidents in operational zones by 90%, and cuts comprehensive operational costs by approximately 50%. This system provides a new technological paradigm for smart water conservancy systems, integrating general-purpose artificial intelligence with domain-specific knowledge.