Applying novelty detection to identify model element to IFC class misclassifications on architectural and infrastructure Building Information Models
Abstract Ensuring the correct mapping of model elements to Industry Foundation Classes (IFC) classes is fundamental for the seamless exchange of information between Building Information Modeling (BIM) applications, and thus achieve true interoperability. This research explored the possibility of employing novelty detection, a machine learning approach, as a way to detect potential misclassifications that occur during current ad hoc and manual mapping practices. By training the algorithm to learn the geometry of BIM elements for a given IFC class, outliers are detected automatically. A framework for leveraging multiple BIM models and training individual one-class SVM's was formulated and tested on four IFC classes. Performance results demonstrate the classification models to be robust and unbiased. The algorithms developed thus can be leveraged to check the integrity of IFC data, a prerequisite for BIM-based quality control and code compliance. Highlights The correct mapping of BIM elements to IFC classes is critical for IFC based interoperability. A framework is formalized for applying novelty detection to automate the checking of misclassifications. One-class SVM's are trained and tested on two architectural and two infrastructure IFC classes. Performance metrics indicate robust and unbiased models with high accuracy and true negative rates. Novelty detection is a superior approach to outlier detection in identifying misclassifications of BIM to IFC associations.