Command prediction based on early 3D modeling design logs by deep neural networks

Abstract Command prediction based on BIM logs is an important computer-aided design (CAD) method to help avoiding design errors especially on early design stages in architecture, engineering and construction (AEC). On this issue, methods for data preprocessing, predictive model training and model evaluation are three primary questions to answer. In this study, a data augmentation method is developed to prepare high quality data input for 3D modeling event logs. A standard Transformer model is trained with the augmented data as input. Six predictive models on three different input data are compared for evaluating the method. In this case, the most accurate of the six models is Transformer with 94% accuracy on top one command prediction. As results, the data augmentation method proposed in this study improves the accuracy of the predictive models up to 1.75 times. Intelligent CAD tool for command prediction with high accuracy during 3D modeling design can be further developed.