Interpreting accident reports by integrating a heterogeneous graph neural network and factor analysis

Occupational safety in the construction industry is one highly prioritized concern around the globe. Accident reports are considered valuable recourses preserving information about corresponding risk factors. Many efforts in the literature have demonstrated that deep learning models are readily applicable to processing and analyzing narrative reports. However, the heterogeneous semantic information was rarely considered. This research utilizes knowledge graph-based accident analysis to provide a machine-assisted approach for construction accident report interpretation. To validate the proposed approach, this research labels 320 crane-related accident reports from the US OSHA database and develops a Crane Safety Knowledge Graph (CSKG) as a case study. Then, a Heterogeneous Graph Attention Network (HAN) is trained to explore the accident features and the importance of various risk factors. Through mapping and clustering the accident data points, the results reveal the capability of the proposed approach to learn the accident patterns and generate safety rules for construction cranes.