Proactive approach to enhancing safety management using deep learning classifiers for construction safety documentation

Construction accidents remain a major global concern, highlighting the need for proactive safety measures during the design phase. Design for Safety (DfS) reports, which aim to identify and mitigate hazards during the design stage, play a crucial role in preventing accidents. However, their manual classification in South Korea is time-consuming and error-prone, reducing their effectiveness. Thus, this study presents a deep learning-based classification framework to automatically classify DfS reports based on accident types, facility types, and work types. Three deep learning architectures, Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM model, were developed and trained on a comprehensive dataset of 1513 DfS reports collected from government databases and construction companies' internal records. An extensive model evaluation was conducted using key performance metrics, including accuracy, precision, recall, F1 score, and Area Under the Receiver Operating Characteristic Curve (AUC ROC). The results demonstrated that the CNN model achieved the highest performance across all tasks, with an accuracy of 71 % for accident type classification, 46 % for facility type classification, and 64 % for work type classification. Additionally, CNN outperformed LSTM and CNN-LSTM in terms of precision, recall, and F1 score across all categories, indicating its superior capability in extracting spatially correlated textual patterns. Beyond model performance, this study contributes to construction safety by introducing an automated, data-driven approach to DfS report analysis, which facilitates proactive hazard identification, enhances risk prioritization, and optimizes resource allocation during the design phase.