面向煤矿设备维护高效知识问答模型的构建为煤矿设备智能运维提供坚实基础，也是煤矿设备智能化进程的重要保障。促进煤矿设备维护知识（Coal Equipment Maintenance Knowledge, CEMK）的高效管理面临着诸多挑战，包括但不限于其结构形式多样、信息体量庞杂、知识来源广泛等方面，限制了相关知识的可利用能力。促进设备维护知识多源共享，挖掘设备维护知识内在关系，是促进设备维护发展的重要研究内容，为了实现煤矿设备维护领域的智能化发展，重要研究方向之一是多源知识共享，并深入探索其中的内在关系。知识问答模型作为一种治理知识数据的先进技术，能够准确、合乎逻辑地理解和解释自然语言，因而在诸多研究领域得到广泛使用。构建煤矿设备维护知识问答模型可以帮助煤矿设备维护人员更快速、准确地解决各种设备维护问题，提高工作效率，降低维护成本，确保设备的安全可靠运行。

本文主要开展了以下工作：

对煤矿设备维护领域的知识进行深入分析和收集，包括煤矿设备的种类、设备零部件、安全规程、常见故障和维护方法等方面的知识。基于本体建模技术构建煤矿设备维护知识的本体模型，通过Protégé软件可以进行层次结构的组织和维护，对实体和关系进行形式化表示，并行存储，确定命名实体识别中的实体范围和关系抽取中的关系类型。

针对命名实体识别，对煤矿设备维护实体类别进行了定义，针对现有命名实体识别模型未考虑到汉字偏旁部首的这类深层次特征，识别效果还有很大的进步空间，本文提出一种基于多文本特征和局部敏感哈希注意力（Multiple Text Features—Locally Sensitive Hash Attention, MTF-LSHA）的命名实体识别模型。经过实验验证，精确率与召回率分别达到93.32%，92.81%。

针对关系抽取任务，使用依赖树来改进关系提取，为分析实体关提供更好的语义指导。同时为了减少在依赖树自动生成情况下的噪声影响，设计了一种依赖驱动和的注意力图卷积网络（Attention-Graph Convolutional Networks, A-GCN）。在煤矿设备维护领域关系抽取数据集上与其他主流模型进行对比试验，F₁值分别提高了7.33%和4.79%。

在命名实体识别和关系抽取的基础之上，通过基于Neo4j图数据库进行知识存储方法，设计了映射规则，搭建了煤矿设备维护可视化知识图谱系统。基于领域多源文本，构建了基于ReliableCEMK-Self-Instruction煤矿设备维护领域数据集，通过三重低秩自适应（Low-Rank Adaptation, LoRA）微调与直接偏好优化，开发了基于知识图谱增强的煤矿设备维护领域大语言问答模型。在煤矿设备对话咨询、煤矿设备专业顾问和维护决策分析三项任务实验中均为最佳，且推理过程用时最短。

The construction of an efficient knowledge Questions and Answers (Q&A) model for coal mine equipment maintenance provides a solid foundation for intelligent operation and maintenance of coal mine equipment, and is also an important guarantee for the process of coal mine equipment intelligence. Promoting the efficient management of coal mine equipment maintenance knowledge faces a number of challenges, including but not limited to its diverse structural forms, large volume of information, and wide range of knowledge sources, which limit the ability to utilize the relevant knowledge. Promoting multi-source sharing of equipment maintenance knowledge and exploring the intrinsic relationship of equipment maintenance knowledge is an important research content to promote the development of equipment maintenance, and in order to realize the intelligent development in the field of coal mine equipment maintenance, one of the important research directions is to share multi-source knowledge and explore the intrinsic relationship in depth. As an advanced technology for governing knowledge data, knowledge question and answer modeling can accurately and logically understand and interpret natural language, and thus has been widely used in many research fields. Constructing a knowledge Q&A model for coal mine equipment maintenance can help coal mine equipment maintenance personnel solve various equipment maintenance problems more quickly and accurately, improve work efficiency, reduce maintenance costs, and ensure the safe and reliable operation of equipment.

This paper mainly carries out the following work:

In the depth analysis and collection of knowledge in the field of coal mine equipment maintenance, including the types of coal mine equipment, equipment parts, safety regulations, common failures and maintenance methods. The ontology model of coal mine equipment maintenance knowledge is constructed based on the ontology modeling technology, which can be organized and maintained in a hierarchical structure through Protégé software, formal representation of entities and relationships, parallel storage, and determination of entity scope in named entity identification and relationship type in relationship extraction.

For named entity recognition, the entity category of coal mine equipment maintenance is defined, for the existing named entity recognition model does not take into account such deep-level features of Chinese character radicals, and there is still a lot of room for progress in the recognition effect, this paper proposes a NER model based on multiple text features and Locally Sensitive Hash Attention (MTF-LSHA). As verified experimentally, the values of Precision and Recall reached 93.32% and 92.81% respectively.

For the task of relationship extraction, a dependency tree is used to improve the relationship extraction and provide better semantic guidance for analyzing entity relations. Meanwhile, in order to reduce the noise effect in the case of automatic generation of dependency trees, a graph convolutional network fusing dependency-driven and attention (A-GCN) is designed. Comparative tests with other mainstream models on the relationally extracted dataset in the field of coal mine equipment maintenance showed an increase in values of F₁ by 7.33% and 4.79% in the respective cases.

Based on multi-source texts, we constructed the ReliableCEMK-Self-Instruction coal mine equipment maintenance domain dataset. By utilizing triple LoRA fine-tuning and direct preference optimization, we developed a large language question-answering model in the coal mine equipment maintenance domain enhanced by a knowledge graph. It was the best in all three task experiments of coal mine equipment dialog consulting, coal mine equipment professional consulting and maintenance decision analysis. At the same time, it took the shortest time for the reasoning process.