No abstract

Automated parsing of design data will increasingly be a prerequisite for efficient data- and analytics-driven management of building portfolios. The high complexity and low rigidity in BIM model exchange standards such as IFC result in considerable differences in data quality and impede the direct data availability for analytics-based decision support. Mis-or unclassified building elements are a common issue and can lead to tedious manual reworks. At the same time, Scan-to-BIM processes still require considerable manual effort to identify subclass element geometry. In this work, we leverage the benefits of a 3D light-weight, geometric algorithm to autonomously generate meaningful geometric features assisting shape classification in erroneous design models and pre-segmented point clouds. Geometric Deep Learning (GDL) is introduced in two steps; a discussion about the benefits of GCN (Graph Convolutional Networks) is given before a set of experiments on BIM element datasets are conducted. Utilizing explainable AI methods, the GCNs performance is made suitable for human-algorithm interaction. Leveraging element geometry solely, the classification reaches a promising average performance of above 83% for the model healing task with reduced computation time. The encoded geometric knowledge from the design models shows to be helpful in showcase examples of segment classification in point clouds.