基于深度学习的地铁施工隐患文本分类与检索研究

叶成

摘要

施工隐患排查与整改是地铁施工安全管理的重要环节，其目的在于通过现场安全巡视，识别、记录可能导致施工安全事故的现场安全隐患，并加以跟踪、整改。隐患排查系统的开发，有利于规范隐患排查工作、建立健全隐患排查治理长效机制。隐患排查活动积累了大量记录安全隐患的文本数据，沉积、闲置在排查系统中。为了有效管理、利用海量非结构化隐患文本数据，以期用于揭示隐患分布规律、预判安全风险来源、为治理同类隐患提供参考，首先需要解决隐患文本分类与高效检索的问题。为此，本文探索基于深度学习的施工现场安全隐患文本分类与检索研究，主要研究工作包括：

首先，本文阐释了地铁施工安全隐患的含义和类别，结合文献调研和专家访谈，分析了地铁施工隐患排查的工作机制与存在问题，并据此指出隐患文本分类与检索研究的必要性。当前研究主要基于文本挖掘和关键字匹配实现隐患管理，存在隐患分类效果及鲁棒性不足、隐患数据检索效率低下等问题。

其次，结合文本表示与深度学习等相关理论技术，搭建了基于改进BERT（Bidirectional Encoder Representation from Transformers）结构的分类模型，实现了端到端的安全隐患分类。使用某地铁隐患排查系统积累的隐患文本数据，进行了与其他五种分类方法的对比研究，结果表明该方法有相对更好的隐患分类效果与泛化能力。

最后，提出了基于知识图谱的隐患数据检索框架：采用本体与语义网络进行隐患知识表示，结合直接映射、深度神经网络等实现知识抽取，进以构建隐患知识图谱并存储在Neo4j图数据库中。实验结果表明，基于图结构的隐患数据组织方式，能够很好地处理内部关联复杂的隐患数据，实现高效的结构化检索。

本文的研究，提出了基于改进BERT模型的分类方法、基于知识图谱的结构化检索方法，实现地铁施工隐患文本分类与数据高效检索，为集成系统的开发与应用提供支持。同时，本研究也可以为立足于深度学习、知识图谱技术下的建筑领域文本处理、数据检索及管理提供借鉴。

关键词：安全隐患；隐患分类；隐患检索；深度学习；知识图谱

Abstract

Investigation and rectification of hazards is an important part of metro construction safety management. Its purpose is to identify and record potential safety hazards that may cause construction safety accidents through site safety inspections, and to track and rectify them. The development of hazards troubleshooting system is conducive to standardizing hazards investigation work and establishing and improving long-term mechanism for safety hazards investigation and treatment. Safety hazards investigation activities have accumulated a large amount of text data that records safety hazards conditions on site, which are deposited and idle in the investigation system. In order to effectively manage and use massive unstructured hidden danger text data, with a view to revealing the distribution of hidden dangers, predicting the source of security risks, and providing references for rectifying similar hidden dangers, we first need to solve the problems of hidden danger text classification and efficient retrieval. Therefore, this dissertation explores the text classification and retrieval research of safety hazards on the construction site based on deep learning. The main research work are as follows:

Firstly, the meaning and types of safety hazards in metro construction are explained. Combining literature research and expert interviews, the working mechanism and existing problems of metro construction hazards investigation are analyzed, and the necessity of classification and retrieval research of hidden dangers are raised. The current research is mainly based on text mining and keyword matching to realize hazards management. Existing studies have problems such as poor classification effect, insufficient robustness, and low retrieval efficiency of safety hazard data.

Secondly, combined with text representation and deep learning and other related theoretical technologies, a classification model based on the modified BERT (Bidirectional Encoder Representation from Transformers) structure is built to achieve end-to-end classification of safety hazards. Using the text data accumulated by a metro safety management information system, a comparative study with 5 other classification methods was carried out. The results show that the proposed method has relatively better hazards classification effect and generalization ability.

Finally, a hazard data retrieval framework based on knowledge graph is proposed: the ontology and semantic network are used to express hazards knowledge, direct mapping, deep neural network, etc. are combined to realize knowledge extraction, and then to construct hidden knowledge graph and store it in Neo4j graph database. The experimental results show that the hazard data organization based on graph structure can process hazard data with complex internal associations and achieve efficient structured retrieval.

This dissertation proposed a classification method based on modified BERT model and a structured retrieval method based on knowledge graph in order to realize text classification and data retrieval of safety hazards in metro construction, and provide support for the development and application of integrated systems. Besides, this research can also provide a reference for text processing, data retrieval and management in the field of architecture based on deep learning and knowledge graph technology.

Key Words：Safety hazards; Hazard classification; Hazard retrieval; Deep learning; Knowledge graph