土木工程施工工人安全知识智能推荐研究

刘佳静

摘要

土木工程施工安全问题一直是困扰业界、学者和社会的重大难题。安全培训是提高施工工人安全知识认知水平、操作技能水平和安全意识的有效手段之一。然而，现有安全培训方法忽视了工人学习风格、技能水平等个性化特征对安全培训效果的影响。由此，无法以个体施工工人乐于接受的方式适应性呈现其需要并缺乏的安全知识。为此，本文提出了基于自适应学习的施工工人安全知识智能推荐方法，主要研究工作如下：

（1）基于自然语言处理的施工工人安全知识领域模型的构建。结合施工安全标准规范、事故调查报告等安全知识载体，构建了面向施工工人安全知识学习的语义本体模型。为从包含施工安全知识的文本中自动抽取安全知识三元组，以CASREL 模型为基础，集成对比学习思想提出了一种融合无监督学习和有监督学习的CL-CASREL模型，精确率、召回率和F1分数分别提高了9.9%、5.0%和7.0%。采用图数据库Neo4j 对抽取的安全知识三元组进行存储和更新，构建了可视化的安全知识知识图谱。由此，实现了施工工人安全知识领域模型的构建，为安全知识适应性智能推荐提供了数据基础。

（2）基于机器视觉的施工工人安全知识学习者模型的构建。阐述了教育领域学习者模型中个性化特征的常见构成，结合施工工人特色特征，明确了工人学习者模型的自适应源，主要包括工种类型、学习风格、认知水平和技能水平。对于自适应源的特征描述和获取，采用所罗门学习风格量表获取施工工人的学习风格，采用安全知识测试题考核施工工人的认知水平。利用视频帧中蕴含的时空信息，提出了基于机器视觉和深度学习的方法监测施工工人的技能水平。由此，实现了施工工人安全知识学习者模型的构建与动态更新，为安全知识适应性智能推荐提供了依据。

（3）基于多模态数据的施工工人安全知识自适应引擎的构建。建立了自适应引擎安全知识推荐的框架流程，描述了不同施工工人特征与培训形式、培训路径及培训内容推荐间的关系。针对施工工人的认知需求和学习风格差异，构建了多模态安全知识智能推荐算法，包括基于K-BERT 的文本匹配的培训素材推荐，基于KSGRAF的图像文本匹配的培训素材推荐，以及基于DOLG 的图像匹配的培训素材推荐。通过融合领域模型包含的安全知识语义信息，数据和知识融合驱动的算法（K-BERT和K-SGRAF）表现出比数据驱动的算法（BERT和SGRAF）更优越的性能。由此，实现了施工工人安全知识自适应引擎的构建，可支持安全知识的智能推荐与适应性呈现。

（4）施工工人安全知识自适应学习系统的设计与实现。基于施工工人安全知识领域模型、学习者模型和自适应引擎的构建，对施工工人安全知识自适应学习系统的移动端和网页端进行了设计、开发与实现。对比采用自适应学习系统和采用非自适应学习系统工人的安全知识学习状态的变化，验证了基于自适应学习的安全知识智能推荐框架体系对工人的安全知识认知水平和技能水平的提升效用。

本研究成果对减少施工现场不安全行为、提高工地安全管理水平、促进建筑业的高质量发展具有重要意义。

关键词：不安全行为；安全培训；自适应学习；知识图谱；深度学习；多模态数据

Abstract

Safety in construction has always been a major challenge plaguing practitioners,scholars, and society. Safety training is one of the effective means to improve safety knowledge, operation skills and safety awareness of workers. Existing safety training methods ignore the influence of individual characteristics, such as learning style and skill level, on the effect of safety training. As a result, the safety knowledge that individual workers need and lack cannot be adaptively presented in a receptive way. To this end, this research proposes an adaptive learning-based approach for an intelligent recommendation of safety knowledge for workers in construction. The main research contents are as follows:

(1) Construction of worker safety knowledge domain model based on natural language processing. Based on the analysis of construction safety standards, accident reports and other safety knowledge carriers, a semantic ontology model for worker safety knowledge learning is constructed. To automatically extract safety knowledge triples from texts containing construction safety knowledge, based on the CASREL model, a CL-CASREL model integrating unsupervised learning and supervised learning is proposed. The experimental results show that the precision, recall and F1-score are increased by 9.9%, 5.0% and 7.0%, respectively; Neo4j is then used to store and update the extracted safety knowledge triples to construct a visual safety knowledge graph. As a result, the construction of the worker safety knowledge domain model is realized, which provides a data basis for the adaptive intelligent recommendation of safety knowledge.

(2) Construction of worker safety knowledge learner model based on computer vision. The typical composition of personalized characteristics in the learner model in the field of education and the distinctive features of workers are expounded. Therefore, the adaptive source of the worker learner model is clarified, including the type of work, learning style, cognitive level and skill level. For the feature description and acquisition of the adaptive source, the Solomon Learning Style Scale is used to obtain workers’ learning style, and the safety knowledge test is applied to assess workers’ cognitive level. Making use of the spatiotemporal information contained in video frames, an approach based on computer vision and deep learning is proposed to monitor the workers’ skill level. As a result, the construction and dynamic update of the worker safety knowledge learner model is realized, which provides a basis for the adaptive intelligent recommendation of safety knowledge.

(3) Construction of worker safety knowledge adaptive engine based on multimodal data. A framework process for adaptive engine safety knowledge recommendation is first established to describe the relationship between the characteristics of workers and the recommendation of training form, training path and training content. Considering the differences in workers’ cognitive need and learning style, intelligent recommendation algorithms for safety knowledge of different modalities is developed, including K-BERT for the recommendation of text training materials based on text, K-SGRAF for the recommendation of image training materials based on text, and DOLG for the recommendation of image training materials based on an image. By fusing the safety knowledge semantic information contained in the domain model, data and knowledge fusion-driven algorithms (K-BERT and K-SGRAF) exhibit superior performance than datadriven algorithms (BERT and SGRAF). In this way, the construction of worker safety knowledge adaptive engine is realized, which can support the intelligent recommendation and adaptive presentation of safety knowledge.

(4) Design and application of worker safety knowledge adaptive learning system. Based on the construction of worker safety knowledge domain model, learner model and adaptive engine, the mobile and web terminal of the worker safety knowledge adaptive learning system are designed, developed, and applied. Comparing changes in safety knowledge learning status of workers with adaptive learning system and workers with nonadaptive learning system, it verifies that the adaptive learning-based safety knowledge intelligent recommendation framework improves workers’ cognitive level and skill level.

This research is of great significance for reducing unsafe behaviors on construction sites, improving the level of safety management, and promoting the high-quality development of the construction industry.

Keywords: Unsafe behavior; Safety training; Adaptive learning; Knowledge graph; Deep learning; Multi-modal Data