Sequential Design Command Prediction Using BIM Event Logs: Fid-Bau Portal

Sequential Design Command Prediction Using BIM Event Logs

Pan, Yue / Zhang, Limao

A method of sequential design commands prediction is developed based on the recurrent neural network (RNN), providing a unique opportunity to discover the hidden knowledge behind the building information modeling (BIM) event logs. The modeling process will generate huge amounts of data, which will be documented automatically in BIM logs in chronological order. To fully understand a series of design commands in logs, RNN with the sequential memory is constructed to learn features of extracted sequential data from logs and predict the next possible design commands. The proposed RNN-based command prediction approach is tested in a real dataset of BIM event logs in journal file “Create” containing totally 57,915 command records, which are split in 80%–20% for the training set and testing set. Acting as a multi-classification task, hundreds of design commands are categorized into six classes and labeled by number 1–6. Ultimately, the RNN with 1 hidden layer and 64 hidden neurons will be trained to reach an overall accuracy of 63.86%. Based on the performance measurements: confusion matrix, receiver operating characteristics (ROC) curve, and area under the curve (AUC), all the six design command classes can be well distinguished from each other in the RNN model, verifying its great classification ability. The novelty of this research lies in: (a) The state of knowledge by exploiting RNN in event log data mining to predict the design command sequence accurately; (b) The state of practice by providing an intelligent modeling solution for designers to improve the efficiency and quality of the design process.