知识图谱在复杂信息整理，事故归因推理分析等多个应用场景中扮演着重要角色， 随着智慧矿山理念被提出，利用此技术构建煤矿智能信息库成为了研究焦点。目前，由 于煤矿数据的复杂性，难以对其进行统一高效的收集，导致无法构建相对完整的井下不 安全行为知识体系。单一模态的不安全行为实体展示存在单调、不完整、难以理解等问 题，导致井下工人难以丰富、全面的理解行为安全体系。针对上述问题，本文提出两种

面向煤矿井下不安全行为知识图谱构建方法。主要内容如下：

    （1）针对煤矿数据难以统一高效的收集，无法构建相对完整的井下不安全行为知 识体系等问题，本文提出构建煤矿井下不安全行为知识图谱。首先，利用相关文献及煤矿安全规则构建数据集。其次，提出一种基于 RoBERTa-BiLSTM-MLP-CRF 命名实体识别方法。此外，设计了基于知识三元组的依存句法树结构，并根据该结构对井下不安全 行为领域的知识资源进行了抽取与表示。最后，利用 Neo4j 进行知识存储，构建煤矿井 下不安全行为知识图谱。实验结果表明，在自建煤矿井下不安全行为数据集下，所提命名实体识别模型的准确率、召回率、F1 值分别达到 77.2%、80.6%和 78.9%。

    （2）针对单一模态的不安全行为实体展示存在单调、不完整、难以理解等问题，本文提出构建煤矿井下不安全行为视频-文本多模态知识图谱。首先，基于自建多模态数据集与标注信息，提出 RoBERTa 联合抽取模型获取文本实体，采用人工干预的方式获取视频实体。其次，提出基于 LCS 融合对齐方法，利用最长公共子序列计算相似度，并进行匹配链接。最后，利用 Neo4j 进行知识存储，并利用 KgBuilder 多模态知识图谱构 建工具实现对视频节点的嵌入，从而构建煤矿井下不安全行为视频-文本多模态知识图谱。实验结果表明，在自建煤矿井下不安全行为数据集下，所提融合对齐方法的准确率、召回率、F1 值分别达到 85.7%、88.6%和 87.1%。

    （3）搭建了煤矿井下不安全行为知识库管理系统，实现了煤矿井下不安全行为相关文本信息查询以及视频展示。

       Knowledge graph plays an important role in many application scenarios such as complex information collation, accident attribution reasoning and analysis, and with the concept of smart mine being proposed, the construction of coal mine intelligent information database using this technology has become the focus of research. At present, due to the complexity of coal mine data, it is difficult to collect them uniformly and efficiently, resulting in the inability to construct a relatively complete knowledge system of unsafe underground behavior. There are problems such as monotony, incompleteness, and incomprehension of the entity display of unsafe behaviors in a single modality, which makes it difficult for underground workers to understand the behavior safety system in a rich and comprehensive way. In order to solve the above problems, this paper proposes two methods for constructing knowledge graphs based on unsafe behaviors in coal mines. The main contents are as follows:

         (1) In view of the problems that it is difficult to collect coal mine data in a unified and efficient manner, and it is impossible to construct a relatively complete knowledge system of unsafe underground behaviors, this paper proposes to construct a knowledge graph of unsafe behaviors in coal mines. Firstly, the dataset was constructed by using relevant literature and coal mine safety rules. Secondly, a named entity recognition method based on RoBERTa-BiLSTM-MLP-CRF was proposed. In addition, a dependent syntactic tree structure based on knowledge triples was designed, and the knowledge resources in the field ofunderground unsafe behaviors were extracted and represented according to this structure. Finally, Neo4j was used for knowledge storage to construct a knowledge graph of unsafe behaviors in coal mines. Experimental results show that the accuracy, recall and F1 values of the proposed named entity recognition model are 77.2%, 80.6% and 78.9%, respectively, under the dataset of unsafe behavior in self-built coal mines.

       (2) In view of the monotonous, incomplete and incomprehensible problems in the display of unsafe behaviors in a single modality, this paper proposes to construct a video-text multimodal knowledge graph of unsafe behaviors in coal mines. Firstly, based on the self-built multi-modal dataset and annotated information, a RoBERTa joint extraction model was constructed to obtain text entities, and manual intervention was used to obtain video entities. Secondly, a fusion alignment method based on LCS is proposed, which uses the longest common subsequence to calculate the similarity and perform matching linking. Finally, Neo4j was used for knowledge storage, and the KgBuilder multimodal knowledge graph construction tool was used to embed video nodes, so as to construct a video-text multimodal knowledge graph of unsafe behaviors in coal mines. The experimental results show that the accuracy, recall and F1 values of the proposed fusion alignment method are 85.7%, 88.6% and 87.1%, respectively, under the dataset of unsafe behaviors in self-built coal mines.

       (3) The knowledge base management system of unsafe behaviors in coal mines has been built, and the text information query and video display related to unsafe behaviors in coal mines have been realized. It enables underground workers in coal mines to better understand the knowledge of unsafe behaviors, so as to enhance safety awareness and reduce accidents.

      This paper constructs a knowledge graph of unsafe behaviors in coal mines, which provides strong support for employee safety management in coal mines by using the prior knowledge of unsafe behaviors in coal mines, and then improves the efficiency of underground safety management.