A Comprehensive AI-Driven Framework for Construction Risk Assessment: From Site-Level Prediction to Task-Level Reasoning and Regulatory Integration = AI 기반의 건설현장 위험성 평가 통합 프레임워크: 현장 수준 예측에서 작업 수준 위험 추론 및 규제 연동까지

다국어 초록 (Multilingual Abstract)

The construction industry continues to face high accident rates due to complex work environments and fragmented safety-management practices. Conventional risk assessment remains heavily dependent on the subjective judgment of assessors, while available accident databases are largely categorical and lack the contextual detail necessary for precise analysis. Furthermore, the absence of systematic linkage between risk assessment and safety regulations contributes to documentation-oriented “paper safety,” limiting the practical effectiveness of hazard prevention.
To address these limitations, this dissertation proposes an intelligent, regulation-aware, multi-level risk assessment framework that integrates three core components: (1) a machine learning (ML)–based site-level risk prediction model; (2) a task-level accident scenario inference framework using Large Language Models (LLMs) and Graph Neural Networks (GNNs); and (3) a regulation-integration mechanism that links quantitative risk results to construction safety laws.
First, a site-level risk assessment framework is developed using 13,047 accident cases from the CSI database. Three ML sub-models—predicting accident objects, accident types, and death–injury outcomes—are trained using RF, XGBoost, CatBoost, and LightGBM. After resolving class imbalance with SMOTE, XGBoost shows the best performance, with weighted F1-scores of 0.839 (object), 0.749 (type), and 0.977 (severity). A web-based prototype is implemented to operationalize these predictions for on-site decision support.
Second, to address the lack of contextual detail in structured records, the framework applies LLM-based extraction to obtain fine-grained entities—such as work tasks, equipment, and materials—from unstructured accident narratives. These entities form an accident-correlation network modeled with GNNs. Simple-HGN delivers the highest performance, achieving a 93.6% mAP for accident-object prediction and strong accuracy for accident causes and types. Scenario outputs show close alignment with real accident distributions (JSD < 0.3) and well-calibrated probabilities (ECE < 0.05). For severity prediction, a Lognormal–SVR model yields a MAE of 0.212.
Finally, a Safety Legal Map is developed to link risk predictions with regulatory requirements from OSHA, CTPA, and SAPA, enabling automatic retrieval of relevant legal provisions and documentation obligations alongside inferred accident scenarios.
Collectively, this research provides a unified methodology that advances both academic and practical domains. Academically, it demonstrates how ML, LLMs, and GNNs can be combined to overcome the contextual limitations of accident data and model complex relational patterns. Practically, it establishes the foundation for transforming construction safety management from subjective, post-incident responses to an objective, data-driven, and regulation-aligned system that supports proactive accident prevention.

국문 초록 (Abstract)

건설산업은 복잡한 작업환경과 분절된 안전관리 관행으로 인해 여전히 높은 재해율을 보이고 있다. 기존의 위험성 평가는 평가자의 주관적 판단에 크게 의존하며, 활용 가능한 사고 데이터 또한 대부분 범주형 정보로 구성되어 있어 정밀한 분석에 필요한 작업 맥락이 부족하다. 또한 위험성 평가 결과가 안전보건 법규와 체계적으로 연계되지 않아 문서 중심의 형식적 안전관리를 초래하며 실질적인 사고 예방 효과를 제한한다.
이러한 한계를 해결하기 위해, 본 논문은 (1) 머신러닝(ML) 기반 현장 수준(site-level) 위험 예측 모델, (2) 대규모 언어모델(LLM)과 그래프신경망(GNN)을 이용한 작업 수준(task-level) 사고 시나리오 추론 모델, (3) 정량적 위험결과를 건설안전 법규와 연결하는 규제 통합 메커니즘으로 구성된 지능형·법규연계형 위험성 평가 프레임워크를 제안한다.
우선, CSI 데이터베이스의 13,047건 사고 사례를 활용하여 사고객체, 사고유형, 사망·부상 여부를 예측하는 세 개의 ML 하위모델을 RF, XGBoost, CatBoost, LGBM 알고리즘을 통해 학습시켰다. 그 결과 XGBoost가 가장 높은 성능을 보였으며, 가중 평균 F1-score는 사고객체 0.839, 사고유형 0.749, 사망·부상 0.977을 기록하였다. 또한 현장 실무를 지원하기 위해 웹 기반 프로토타입을 구현하였다.
다음으로, 구조화 사고 데이터의 맥락 부족 문제를 해결하기 위해 사고 데이터의 비정형 텍스트에서 LLM 기반 정보추출을 수행하여 작업, 장비, 자재 등의 세부 맥락 정보를 확보하였다. 이를 통해 사고 상관관계 네트워크를 구축하고 GNN을 활용해 정량적 위험 추론을 수행하였다. 평가된 모델 중 Simple-HGN이 가장 뛰어난 성능을 보였으며, 사고객체 예측에서 93.6% mAP를 기록하고 사고원인과 사고유형 예측에서도 높은 정확도를 달성하였다. 생성된 사고 시나리오는 실제 사고 분포와 높은 유사성을 보였으며(JSD<0.3), 확률 추정에서도 우수한 보정 성능을 나타냈다(ECE<0.05). 사고 심각도는 Lognormal–SVR 모델이 MAE 0.212로 정밀하게 예측하였다.
마지막으로, 산언안전보건법, 건설기술진흥법, 중대재해처벌법 등 주요 건설안전 법령을 기반으로 안전법규 맵을 구축하여 위험성 평가 결과와 법규 요건을 자동으로 연계할 수 있는 체계를 마련하였다. 이를 통해 예측된 사고 시나리오와 함께 관련 안전조치 및 문서요구사항을 즉시 조회할 수 있도록 지원한다.
종합적으로 본 연구는 ML·LLM·GNN·법규연계 기술을 통합하여 사고 데이터의 맥락 부족 문제를 극복하고, 복잡한 사고 관계 구조를 정량적으로 모델링하는 통합적 방법론을 제시한다. 이는 주관적·사후 대응 중심의 안전관리에서 벗어나 객관적·데이터 기반·법규 정합적 예방 중심 체계로의 전환을 가능하게 한다.

목차 (Table of Contents)

• DEDICATION iii
• ACKNOWLEDGEMENT iv
• ABSTRACT vii
• CONTENTS x
• List of Table xiv
• List of Figure xvi
• Chapter 1. Introduction １
• 1.1. Research Background １
• 1.2. Problem Statement ８
• 1.3. Research Objectives １１
• 1.4. Dissertation Outline １３
• Chapter 2. Literature Review １６
• 2.1. Data-Driven Risk Assessment in Construction Safety １６
• 2.1.1. Conventional and Experience-Based Risk Assessment １６
• 2.1.2. Data-Driven and AI-Based Accident Prediction １７
• 2.1.3. Graph Neural Network-Based Accident Prediction ２２
• 2.2. Regulation-Integrated Safety Management Framework ２４
• 2.2.1. Legal and Regulatory Frameworks for Construction Safety ２４
• 2.2.2. Integration of Regulatory Knowledge ２６
• 2.3. Limitations of Existing Studies ２９
• Chapter 3. Machine Learning-Based Site-Level Risk Assessment Framework ３１
• 3.1. Data Preparation ３２
• 3.1.1. Collection of Construction Accident Data ３２
• 3.1.2. Data Cleaning ３３
• 3.1.3. Data Sampling ３４
• 3.2. Development of Site-Level Risk Assessment Framework ３７
• 3.2.1. Optimum Model Selection for Site-Level Risk Assessment ３７
• 3.2.2. Model Evaluation ４０
• 3.3. Experimental Results and Web-based Prototype Design ４２
• 3.3.1. Result of Site-Level Risk Assessment ４２
• 3.3.2. Discussions of Site-Level Risk Assessment ４５
• 3.3.3. UI and UX Design of Site-Level Risk Assessment ５０
• Chapter 4. Graph Neural Network-Based Task-Level Risk Assessment Framework ５４
• 4.1. Development of Accident Correlation Network ５７
• 4.1.1. Data Processing ５７
• 4.1.2. Accident Correlation Network Design ５９
• 4.1.3. NER Scheme and Prompt Design ６２
• 4.1.4. LLM-based NER Process and Results ６９
• 4.2. Data Processing and Model Selection ８０
• 4.2.1. Heterogeneous Accident Correlation Network Structure ８０
• 4.2.2. Model Selection for Accident Scenario Modeling ８４
• 4.3. Model Training ８８
• 4.3.1. Initial Node Embedding ８８
• 4.3.2. Optimal Model Selection ８９
• 4.4. Development of Task-Level Risk Assessment Framework ９３
• 4.4.1. Probability of Occurrence and Severity Calculation ９３
• 4.4.2. Evaluation of Generated Scenarios Diversity ９５
• 4.4.3. Evaluation of Scenario Distribution Similarity and Probability Calibration １０２
• 4.4.4. Severity Prediction of Accident Scenarios １０６
• Chapter 5. Integration of Safety Regulations with Risk Assessment １１４
• 5.1. Construction Safety Legal Data Collection １１５
• 5.1.1. Target Legal Data Selection １１５
• 5.1.2. Legal Data Collection Process １１８
• 5.2. Hierarchical Mapping of Legal Data １２０
• 5.2.1. Legal Data Linking Process １２０
• 5.2.2. Limitations of the Linked Legal Map １２２
• 5.2.3. Legal Unit Decomposition １２４
• 5.3. Filtering and Classification of Construction Safety Legal Units １２６
• 5.3.1. Legal Unit Filtering Process １２６
• 5.3.2. Legal Unit Classification Process １２６
• 5.3.3. Legal Unit Filtering and Classification Results １２８
• Chapter 6. Discussion １３０
• 6.1. Practical Implications for On-site Deployment １３０
• 6.1.1. Unified Operational Pipeline for Risk Assessment and Legal Compliance １３０
• 6.1.2. UI and UX Design for Risk and Compliance Integration １３４
• 6.2. Empirical Evidence and Comparative Evaluation １３７
• 6.2.1. Reliability of Conventional Risk Assessment Framework １３７
• 6.2.2. Comparative Analysis of Task-Level Risk Assessment １４１
• 6.3. Evaluation of Regulatory Matching Quality １４５
• Chapter 7. Conclusions １５０
• 7.1. Summary of Research Objectives and Achievements １５０
• 7.2. Research Contributions １５２
• 7.3. Improvement Opportunities and Future Research １５４
• Appendix A. Legal Unit Filtering Criteria １５７
• Appendix B. Legal Unit Integration Map １６１
• Bibliography １７１
• 국문 초록 １８４