The construction industry suffers from workplace accidents, including injuries and fatalities, which represent a significant economic and social burden for employers, workers, and society as a whole. The existing research on construction accidents heavily relies on expert evaluations, which often suffer from issues such as low efficiency, insufficient intelligence, and subjectivity. However, expert opinions provided in construction accident reports offer a valuable source of knowledge that can be extracted and utilized to enhance safety management. Today this valuable resource can be mined as the advent of artificial intelligence has opened up significant opportunities to advance construction site safety. Ontology represents an attractive representation scheme. Though ontology has been used in construction safety to solve the problem of information heterogeneity using formal conceptual specifications, the establishment and development of ontologies that utilize construction accident reports are currently in an early stage of development and require further improvements. Moreover, research on the exploration of incorporating deep learning methodologies into construction safety ontologies for predicting construction safety incidents is relatively limited. This paper describes a novel approach to improving the performance of accident prediction models by incorporating ontology into a deep learning model. First, a domain word discovery algorithm, based on mutual information and adjacency entropy, is used to analyze the causes of accidents mentioned in construction reports. This analysis is then combined with technical specifications and the literature in the field of construction safety to build an ontology encompassing unsafe factors related to construction accidents. By employing a Translating on Hyperplane (TransH) model, the reports are transformed into conceptual vectors using the constructed ontology. Building on this foundation, we propose a Text Convolutional Neural Network (TextCNN) model that incorporates the ontology specifically designed for construction accidents. We compared the performance of the TextCNN model against five traditional machine learning models, namely Naive Bayes, support vector machine, logistic regression, random forest, and multilayer perceptron, using three different data sets: One-Hot encoding, word vector, and conceptual vectors. The results indicate that the TextCNN model integrated with the ontology outperformed the other models in terms of performance achieving an impressive accuracy rate of 88% and AUC value of 0.92.

建筑业经常发生工伤事故，包括伤亡，给雇主、工人和整个社会带来重大的经济和社会负担。现有的施工事故研究严重依赖专家评估，往往存在效率低、智能化程度低、主观性强等问题。然而，建筑事故报告中提供的专家意见提供了宝贵的知识来源，可以提取和利用这些知识来加强安全管理。如今，随着人工智能的出现，这种宝贵的资源可以被开采，为提高建筑工地安全提供了重要机会。本体论代表了一种有吸引力的表示方案。尽管本体已被用于施工安全，通过形式化的概念规范来解决信息异构问题，但利用施工事故报告的本体的建立和开发目前还处于发展的早期阶段，需要进一步改进。此外，将深度学习方法纳入施工安全本体以预测施工安全事件的探索研究相对有限。本文描述了一种通过将本体纳入深度学习模型来提高事故预测模型性能的新方法。首先，采用基于互信息和邻接熵的领域词发现算法来分析施工报告中提到的事故原因。然后将该分析与施工安全领域的技术规范和文献相结合，构建包含与施工事故相关的不安全因素的本体。通过采用超平面翻译 (TransH) 模型，使用构建的本体将报告转换为概念向量。在此基础上，我们提出了一种文本卷积神经网络（TextCNN）模型，该模型结合了专门为建筑事故设计的本体。我们使用三种不同的数据集：One-Hot 编码、词向量和概念向量，将 TextCNN 模型与五种传统机器学习模型（即朴素贝叶斯、支持向量机、逻辑回归、随机森林和多层感知器）的性能进行了比较。结果表明，与本体集成的 TextCNN 模型在性能方面优于其他模型，达到了令人印象深刻的 88% 的准确率和 0.92 的 AUC 值。