Browsing by Author "Peng, Mengnanlan"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Combined Unsupervised-Supervised Classification for Building Information Modeling Object Assemblies
    (2023-03-20) Peng, Mengnanlan; Pitkänen, Henri; Filoche, Pascal; Perustieteiden korkeakoulu; Bäckström, Tom
    In Building Information Modeling, as the volume of data and the collaborative nature of projects increase, there is a growing need to structure the data to maintain and search for information efficiently. Many Building Information Modeling software requires a classification system to identify and summarize similar objects. In Tekla Structures, one of the most famous Building Information Modeling software, assembly data classification is critical due to corresponding functionalities developments. Tekla Structures contains many assembly data, laying the foundation for retrieving similar assemblies and developing related classification systems. In general, before applying the object's attribute value, it is necessary to do relevant data analysis to find the most representative attribute. The first purpose of this paper is to find out the attributes that can be used to represent the assembly object accurately. This thesis applies a supervised machine learning method, namely Permutation feature importance, to select essential features after basic attribute filtering, data cleaning, and transformation. The selected features will be used for further classification. The classification in this paper is incremental since the assembly objects that need to be classified in the system will increase with the growth of clients. The underlying foundation of the classification is distance-based and Single-Pass algorithms. The classification system allows users to adjust the classification results based on their preference, and the system will automatically apply the user's feedback and modification to improve performance. The selected features and workflow design are assessed by simulating actual user operations and evaluation cases. The results show that critical standard features are selected from different models, and the applied classification system can be automatically improved according to user feedback.