Automating 3D As-Built BIM with Deep Learning and Knowledge-Driven Graphs

The accuracy and efficiency of as-built building information modelling (BIM) are critical for modern construction and renovation projects. Traditional methods for creating as-built BIM models often depend on manual processes, which are not only time-consuming but also prone to errors, leading to models that may lack precision and detail. This study introduces an approach that leverages deep learning and knowledge-driven graph construction to automate BIM reconstruction. This method systematically segments building elements from 3D point clouds using semantic instance segmentation. Point cloud segments of each component are then analyzed for precise geometric and spatial features, transforming them into nodes within a comprehensive connectivity graph. This graph illustrates not only the spatial and geometric relationship but also the topological relationships between building components through nodes and edges, which are categorized into structural connections. The approach is further refined through knowledge-driven inference, which addresses gaps and inaccuracies in point cloud data, ensuring a detailed and accurate representation of component features and topological relationships. The results show notable gains in both overall BIM model correctness and semantic segmentation accuracy. Specifically, over 80% of the point clouds conform to a 10 mm tolerance threshold, significantly enhancing the automation and precision of BIM reconstructions.