Facility management companies typically struggle with the building data that they receive upon project handover after construction completion. Unstructured, incomplete, or uncertain data, large document size and non-conformance to operational requirements are amongst the most common challenges. Restructuring, altering, or even remodelling of the digital building data is tedious, time-consuming, and often results in a considerable loss of information, higher project costs and lower operational efficiencies. Methods bundled under the term of semantic enrichment attempt to imitate the trained eye of architects and engineers to deduce implicitly available information in BIM models. Machine learning approaches have proven to handle such tasks similarly well or even better than rule-based inference methods. Geometric deep learning recently shows promising results for classification and segmentation tasks of non-Euclidian data such as 3D meshes or graphs. Capturing geometric features from Building Information Modeling (BIM) object geometries in the building unlocks further semantics and thereby adds to provide the required information needed during building operation. The work formulated in this thesis starts from the observation of the described industry need, embeds it into a well-defined scientific context and proposes a novel toolset leveraging the power of Graph Convolutional Neural Network (GCN)s. Aiming to contribute to solving significant bottlenecks in BIM information processing, the work is validated with the industry partner Siemens AG, and implementation possibilities are discussed. Using project handover data such as as-planned BIM models and as-built point clouds, the outcome of this project can be summarized in three contributions. A certainty estimate on the semantic correctness of BIM objects, a performance assessment on the correctness of the as-planned BIM geometries and finally a prediction of a semantic label for items unavailable in the as-planned BIM but visible in the as-built point cloud. The designed and/or trained GCN architectures proved feasible and show average prediction accuracies of 0.80 for classification and 0.77 for semantic segmentation tasks. The results and validation talks with the industry partner show that the proposed solutions could significantly increase the amount of digitally available building information and thereby lead to a more efficient and data-driven building operation.     «

Facility management companies typically struggle with the building data that they receive upon project handover after construction completion. Unstructured, incomplete, or uncertain data, large document size and non-conformance to operational requirements are amongst the most common challenges. Restructuring, altering, or even remodelling of the digital building data is tedious, time-consuming, and often results in a considerable loss of information, higher project costs and lower operational ef...     »