Predicting unsafe behavior of construction workers in Bangladesh using machine learning : a proactive approach to safety management | |
| Author | Hasan, Md. Rashedul |
| Call Number | AIT Thesis no.CM-26-06 |
| Subject(s) | Construction workers--Safety measures--Bangladesh Construction industry--Data processing--Bangladesh Machine learning |
| Note | A thesis submitted in partial fulfillment of the requirements for the Degree of Master of Engineering in Construction, Engineering and Infrastructure Management |
| Publisher | Asian Institute of Technology |
| Abstract | The construction industry remains one of the most hazardous sectors globally, with high rates of accidents and fatalities, particularly in high-risk activities such as excavation and working at height. In Bangladesh, despite significant economic growth, safety practices are largely reactive and dependent on manual inspections, limiting proactive risk prevention. This study aims to develop a machine learning–based predictive framework to assess unsafe behavior among construction workers, with a focus on unsafe behaviors such as failure to use appropriate personal protective equipment (PPE) and working an excavation without protection (no shoring, no trench box ,no slope).A mixed-method research methodology integrating qualitative and quantitative approaches was adopted. Unsafe behavior factors were identified through literature review and validated by five construction safety experts. Data were collected through a structured questionnaire survey using purposive and stratified sampling across construction projects in Bangladesh. The dataset included physical, administrative, human–behavioral, task–operational, organizational– management, and real-time physiological factors, with reliability confirmed through pilot testing and Cronbach’s alpha.Supervised machine learning models, including Random Forest, XGBoost, and Linear Regression, were developed to predict unsafe behavior classification and probability. The dataset was divided into 80% training and 20% testing sets, with k-fold cross validation applied to ensure robustness. A Gradio-based interface was also developed for real time application.The results show that unsafe PPE behavior is mainly influenced by human factors such as overconfidence, competency, and compliance, whereas excavation without protection is primarily driven by organizational and operational factors, including resource allocation, supervision, and geotechnical hazards. XGBoost achieved 95.40% accuracy for PPE prediction, while Random Forest achieved 92.73% accuracy for excavation without protection. Feature importance, SHAP, and domain ablation analyses further revealed that PPE risks are behavior- and knowledge-driven, whereas excavation risks are management- and control driven. The findings emphasize the need for targeted interventions, including improved training, supervision, resource allocation, and safety culture. |
| Year | 2026 |
| Type | Thesis |
| School | Faculty of Civil and Environmental Engineering (2026) |
| Department | Other Field of Studies (No Department) |
| Academic Program/FoS | Construction Engineering and Infrastructure Management (CEIM) |
| Chairperson(s) | Hadikusumo, Bonaventura H. W. |
| Examination Committee(s) | Bhatt, Ayushman;Sarawut Ninsawat |
| Scholarship Donor(s) | Local Government Engineering Department(LGED), Bangladesh |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2026 |
 | |