基于南水北调工程巡检专报结合专家经验构建巡检知识图谱概念模型，在此基础上利用实体关系联合抽取框架进行巡检知识抽取，并以Neo4j图数据库为载体进行巡检知识图谱可视化。基于巡检知识图谱进行南水北调工程巡检信息推荐，利用BERT（Bi-directional encoder representation from transformers）预训练模型设计BERT孪生网络，通过知识检索及字符串相似度计算等技术，实现关联工程风险信息推荐，以辅助巡检人员进行工程风险等级诊断。通过实验评估了知识抽取及字符串相似度计算模型的准确性，知识抽取F1 值达到88.42%，字符串相似度计算 F1 值为86.00%。该方法可提高南水北调工程风险管理能力、推动工程运维的数字化发展。

There are many risks in the operation of the South-to-North Water Transfers Project, and the inspection work of the South-to-North Water Transfers Project is of great significance to ensure the safety and stability of the project. Traditional project inspection methods mainly rely on manual experience and have a low degree of digitalization. Due to the uneven professional level of inspection personnel, it is difficult to form a unified record standard, which in turn leads to redundant inspection special report information, among which the accessibility of effective information is poor, making the traditional project inspection method inefficient.As a new generation of information technology, knowledge graph is a powerful tool for knowledge organization and management. In order to alleviate the limitations of traditional project inspection methods and improve the efficiency of project inspection, knowledge graph with deep learning technology was combined, and using knowledge graph to empower intelligent inspection of the South-to-North Water Transfers Project was proposed. Specifically, the inspection knowledge graph was constructed based on the project inspection text, and an project risk information recommendation method was designed based on the inspection knowledge graph. In the process of building the inspection knowledge graph, the conceptual model of the inspection knowledge graph is defined based on expert experience, and on this basis, the entity relationship joint extraction framework is used to extract the structured triplet knowledge from the unstructured project inspection text, and the knowledge visualization is carried out with the Neo4j graph database as the carrier. The inspection knowledge graph clearly presents inspection information such as engineering sites, parts, risk events, and disposal measures in the form of entity-relationship-entity triples, and supports knowledge visualization and knowledge retrieval, which alleviates the limitation of poor accessibility of effective information in inspection reports. In the project risk information recommendation method, the Bert pre-training model and twin network framework were used to design the Bert twin network, and recommends the project risk information of the entities and entities associated with the current part in the inspection knowledge graph to the inspectors by calculating the string similarity of the part entities, and assists the inspectors in the project risk level diagnosis. The quality of the knowledge graph and the effectiveness of the method are evaluated experimentally. The experimental results show that the average F1 value of various relational triples extracted is 88.42%, and the knowledge extraction results have high accuracy, and the quality of the knowledge graph is considered to be reliable. The F1 value of the candidate entity ranking model designed reaches 86%, which is higher than that of the traditional Jaccard algorithm and Word2Vec model.In general, the Bert twin network designed shows good performance in the string similarity calculation task, and the project risk information recommendation results based on patrol inspection knowledge graph are basically reliable. Knowledge graph and deep learning technology were introduced into the intelligent application of project inspection, which realizes the deep correlation and effective use of inspection knowledge, which can provide reference significance for improving the operation and maintenance efficiency of the South-to-North Water Transfers Project and strengthening the risk management capability of the project.