Clash context representation and change component prediction based on graph convolutional network in MEP disciplines.

Clash management, including clash detection and resolution, is critical for construction project success. While clash detection has been automatically conducted by Building Information Modeling (BIM) tools, the clash resolution process is still being conducted manually. This process includes predicting clash change components and determining change methods. Previous studies predict clash change components based on single clash features while ignoring the interdependency between building components and clashes. This paper proposes to use a graph convolutional network (GCN) to improve clash change component prediction by considering the interdependency nature of building components. Instead of assuming each clash is independent, the GCN model uses a component dependency graph as the input to determine whether to make changes to a particular component based on both 1) its features and 2) its impacts on related components, which is transferred through the spatial dependency structure. To further decide what kind of spatial dependency is more informative for clash change component prediction, this study compares different dependency settings. We experiment on the BIM model of a real-world project and then use the Kruskal-Wallis test to compare model accuracy in different set-tings. The result shows that integrating dependency information into consideration can significantly improve model prediction accuracy, and the clash relation is the most critical factor for clash change component prediction.