Application of Machine Learning for Automated Code Compliance

In current practice, qualified engineers check building designs for code compliance manually, a costly and time-consuming process. Building Information Modelling (BIM) opens the possibility for such checks to be automated and... [ view full abstract ]

In current practice, qualified engineers check building designs for code compliance manually, a costly and time-consuming process. Building Information Modelling (BIM) opens the possibility for such checks to be automated and to rapidly provide reliable results even in the early stages of design. All existing efforts to automate compliance checking require translation of building regulations into predefined rule sets with logical clauses, and painstaking preparation of building models for the checking process. We observe that qualified engineers can interpret extensive and complex requirements from regulations, and they can infer design information even if it is not explicitly given in the model. Moreover, human experts rely on previous experience and learn from it. Therefore, we propose the use of machine learning algorithms for code compliance to facilitate the work of interpreting and encoding the compliance checking rules. As for the challenge of preparation of BIM models, previous research has shown that machine learning algorithms can be beneficial for this task as well.

In general, we aim to classify BIM models as compliant/non-compliant with a given set of regulations by using a classifier trained on a given database of models with known outcomes. Since such databases do not exist today, we created an artificial database of models to check against the Israel code for security rooms in residential buildings. 10,000 models were generated using a parametric random model generator, with half of the models compliant and half not. A decision-tree algorithm classifier was trained using 7,000 models from this dataset to predict whether a new model complies with the design code. The accuracy of prediction of the created classifier, validated using the remaining 3,000 models, was 99.9%. In a rigorous test it classified 18 out of 20 marginal cases correctly.

This work demonstrates the feasibility of using machine learning for code compliance checking. The lack of a dataset of checked models required generation of an artificial dataset, which in turn required compilation of the rule sets based on the code for security rooms. However, once datasets of model review results produced by human code-compliance checkers become available, the task of rule set compilation will reduce to a simple task of feature identification. Since earlier work has already shown that the values of relevant features can be extracted automatically from BIM models following an automated process of semantic enrichment, the current work represents a potential breakthrough in terms of future automation of code compliance checking using BIM models.