With the growing adoption of Building Information Modeling (BIM), specialized applications have been developed to perform domain-specific analyses. These applications need tailored information with respect to a BIM model element’s attributes and relationships. In particular, architectural elements need further qualification concerning their geometric and functional ‘subtypes’ to support exact simulations and compliance checks. BIM and its underlying data schema, the Industry Foundation Classes (IFC), provide a rich representation with which to exchange semantic entity and relationship data. However, subtypes for individual elements are not represented by default and often require manual designation, leaving it vulnerable to errors and omissions. Existing research to enrich the semantics of IFC model entities employed domain-specific rule sets that scrutinize their legitimacy and modify them, if and when necessary. However, such an approach is limited in their scalability and comprehensibility. This study explored the use of 3D geometric deep neural networks originating from computer vision research. Specifically, Multi-view CNN(MVCNN) and PointNet were investigated to determine their applicability in extracting unique features of door (IfcDoor) and wall (IfcWall) element subtypes, and in turn be leveraged to automate subtype classifications. Test results indicated MVCNN as having the best prediction performance, while PointNet’s accuracy was hampered by resolution loss due to selective use of point cloud data. The research confirmed deep neural networks as a viable solution to distinguishing BIM element subtypes, the critical factor being their ability to detect subtle differences in local geometries.

随着建筑信息模型（BIM）的日益普及，已经开发出专用的应用程序来执行特定领域的分析。这些应用程序需要有关BIM模型元素的属性和关系的定制信息。特别是，建筑元素需要有关其几何和功能“子类型”的进一步鉴定，以支持精确的模拟和一致性检查。BIM及其基础数据模式（行业基础类（IFC））提供了丰富的表示形式，可以用来交换语义实体和关系数据。但是，默认情况下未表示单个元素的子类型，并且经常需要手动指定，因此容易出现错误和遗漏。现有的丰富IFC模型实体语义的研究采用了特定于领域的规则集，该规则集审查其合法性并在必要时对其进行修改。但是，这种方法的可扩展性和可理解性受到限制。这项研究探索了源自计算机视觉研究的3D几何深度神经网络的使用。具体来说，研究了多视图CNN（MVCNN）和PointNet，以确定它们在提取门的独特特征中的适用性（IfcDoor）和wall（IfcWall）元素子类型，并进而利用它们来自动化子类型分类。测试结果表明MVCNN具有最佳的预测性能，而PointNet的准确性由于选择性使用点云数据而导致的分辨率损失而受到阻碍。这项研究证实了深度神经网络是区分BIM元素亚型的可行解决方案，关键因素是它们能够检测局部几何形状的细微差异。