BIM log mining: learning and predicting design commands

• Main Authors: Pan, Yue, Zhang, Limao
• Other Authors: School of Civil and Environmental Engineering
• Format: Article
• Language: English
• Published: 2022
• Subjects: Engineering::Civil engineering; Log Data Mining; Design Command Prediction
• Online Access: https://hdl.handle.net/10356/161070
• Institution: Nanyang Technological University

This paper develops a framework to learn and predict design commands based upon building information modeling (BIM) event log data stored in Autodesk Revit journal files, which has the potential to improve the modeling efficiency. BIM design logs, which automatically keep detailed records on the modeling process, are the basis of data acquisition and data mining. Long Short-Term Memory Neural Network (LSTM NN), as a probabilistic deep learning model for learning sequential data with varying lengths from logs, is established to provide designers with predictions about the possible design command class in the next step. To demonstrate the feasibility of this method, a case study runs at large design logs over 4 GB from an international design firm for command class prediction. To begin with, useful data retrieved from logs is cleaned and saved in a 320 MB Comma Separated Values (CSV) file with totally 352,056 lines of commands over 289 projects. Subsequently, various design commands are categorized into 14 classes according to their effects and given numerical labels, which are then fed into LSTM NN for training and testing. As a result, the overall accuracy of this particular case study can reach 70.5% in the test set, which outperforms some classical machine learning methods, like k nearest neighbor, random forest and support vector machine. This research contributes to applying a probabilistic LSTM NN with optimal parameters to learn features from designers' subjective behaviors effectively and predict the next possible design command class intelligently towards automation of the design process. Moreover, the three most possible command classes will be offered as the recommendations under the assumption that the correct class tends to appear owning the top three highest probabilities, which can possibly enhance the reliability of predictions.