Advancing Automated Fire Safety Code Compliance: Automatic Annotation of BIM Door Elements

Abstract

Automated code compliance checking often faces challenges due to insufficient semantic information in Building Information Modeling (BIM) models, particularly during the early design stages, when only basic geometrical and alphanumerical information may be available. This study extends previous research on machine-learning-based annotation of ‘Exit’ doors by investigating whether additional BIM-derived features and adapted feature-engineering strategies can support the identification of both ‘Exit’ and ‘Apartment Exit’ doors in early-design residential architectural BIM models. A data-driven workflow was developed using features extracted from BIM models containing basic representations of doors, walls, stairs, and rooms, defined by their approximate geometry, dimensions, locations, and element types. For ‘Exit’ door classification, the model achieved an F1-score of 0.97 and produced correct predictions for the three test buildings used in the evaluation. For ‘Apartment Exit’ door classification, the integration of probability-encoded features improved the F1-score from 0.93 to 0.99 and increased the building-level test scores from [0.18, 0.47, 0.89, 1.00] to [0.82, 0.77, 0.95, 1.00] for the four test buildings, respectively. The results indicate that structured geometrical and alphanumerical information available in early-design BIM models can provide useful input for ML-based semantic enrichment of door elements. The main contributions of this study are: (1) the definition of eight BIM-derived spatial and topological features that can be extracted from basic building-element representations; (2) the adaptation of probability-encoded feature engineering to bounded integer-valued features using neighbor-based mean estimation for unseen values; and (3) the extension of the door-annotation workflow from ‘Exit’ to ‘Apartment Exit’ classification in Belgian high-rise buildings. The proposed approach is intended to support early-stage compliance checking by reducing the need for manual semantic annotation, while further validation on more diverse building datasets is required to assess broader generalizability.