Ensemble-learning approach for the classification of Levels Of Geometry (LOG) of building elements. (January 2022)

• Record Type: Journal Article
• Authors: Abualdenien, Jimmy; Borrmann, André
• Is Part Of: Advanced engineering informatics. Volume 51(2022)
• Publication Date: 2022-01
• Subjects: Building Information Modeling (BIM) -- Level Of Development (LOD) -- Level Of Geometry (LOG) -- Geometric Complexity -- Ensemble models; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2021.101497
• Languages: English
• ISSNs: 1474-0346

Abstract

Abstract: The provision of geometric and semantic information is among the most fundamental tasks in BIM-based building design. As the design is constantly developing along with the design phases, there is a need for a formalism to define its maturity and detailing. In practice, the concept of Level of Development (LOD) is used to specify what information must be available at which time. Such information is contractually binding and crucial for different kinds of evaluations. Numerous commercial and open-source BIM tools currently support the automatic validation of semantic information. However, the automatic validation of the modeled geometry for fulfilling the expected detailing requirements is a complex and still unsolved task. In current practice, domain experts evaluate the models manually based on their experience. Hence, this paper presents a framework for formally analyzing and automatically checking the Level Of Geometry (LOG) of building information models. The proposed framework first focuses on generating a LOG dataset according to the popular LOD specifications. Afterwards, multiple geometric features representing the elements' complexity are extracted. Finally, two tree-based ensemble models are trained on the extracted features and compared according to their accuracy in classifying building elements with the correct LOG. Measuring the modeling time showed a 1.88–2.80-fold increase between subsequent LOGs, with an 8–15-fold increase for LOG 400 compared to LOG 200. The results of classifying the LOG indicated that the combination of 16 features can represent the LOG complexity. They also indicated that the trained ensemble models are capable of classifying building elements with an accuracy between 83% and 85%.