Predicting the Building Envelope in BIM Models Using Graph Convolutional Neural Networks
Kuikka, Toni (2022)
Kuikka, Toni
2022
Master's Programme in Computational Big Data Analytics
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2022-05-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202204273944
https://urn.fi/URN:NBN:fi:tuni-202204273944
Tiivistelmä
In recent years, Building Information Modeling (BIM) has become one of the leading techniques to maintain data from the lifespan of the building in the Architecture, Engineering and Construction industry. Compared to the traditional modeling methods, BIM requires less human contribution, making it an advantageous approach to represent the characteristics of buildings digitally. In addition, one interesting innovation regarding BIM is semantic enrichment in which an existing model is used to obtain new features and include them in the data entity. Despite the increasing use of BIM, the literature concerning the subject remains limited and thus, the full potential of BIM-based applications appears not to be achieved yet.
In this thesis, the objective is to review the potential to employ Machine Learning solutions for BIM-related supervised prediction tasks. Beginning with the BIM dataset, different approaches to formulate 3D data are considered and based on the selected format, the experiments are made by predicting the envelope for each building using a supervised Machine Learning algorithm. Eventually, the buildings are decided to be formatted as graphs and the chosen algorithm is a Graph Convolutional Neural Network with varying architectures that emphasizes the relationships between elements in different ways. This type of graph-based approach for BIM-related classification problems is an area that has not been much examined previously.
Comparing three different neural network models, the classification is performed in two different scenarios. In the first scenario, data utilized for the training and testing are from the same building whereas in the second one, both the training and testing data comprise distinct, complete buildings. In the first scenario, the Graph Convolutional Neural Network is observed to improve classification performance especially for the minor classes compared to the traditional neural network. Also in the second scenario, the accuracy is higher when employing the graph models, although this type of classification task turns out to be more challenging compared to the one in the first scenario.
The results illustrate a great potential for solving BIM-related classification problems using Machine Learning algorithms. For the first prediction task, a potential application area could be dealing with missing data that occur in BIM models frequently. The second scenario, in turn, has an even higher potential to produce useful tools for semantic enrichment. These types of investigations play an important role in developing new methods to process BIM models.
In this thesis, the objective is to review the potential to employ Machine Learning solutions for BIM-related supervised prediction tasks. Beginning with the BIM dataset, different approaches to formulate 3D data are considered and based on the selected format, the experiments are made by predicting the envelope for each building using a supervised Machine Learning algorithm. Eventually, the buildings are decided to be formatted as graphs and the chosen algorithm is a Graph Convolutional Neural Network with varying architectures that emphasizes the relationships between elements in different ways. This type of graph-based approach for BIM-related classification problems is an area that has not been much examined previously.
Comparing three different neural network models, the classification is performed in two different scenarios. In the first scenario, data utilized for the training and testing are from the same building whereas in the second one, both the training and testing data comprise distinct, complete buildings. In the first scenario, the Graph Convolutional Neural Network is observed to improve classification performance especially for the minor classes compared to the traditional neural network. Also in the second scenario, the accuracy is higher when employing the graph models, although this type of classification task turns out to be more challenging compared to the one in the first scenario.
The results illustrate a great potential for solving BIM-related classification problems using Machine Learning algorithms. For the first prediction task, a potential application area could be dealing with missing data that occur in BIM models frequently. The second scenario, in turn, has an even higher potential to produce useful tools for semantic enrichment. These types of investigations play an important role in developing new methods to process BIM models.