Learning from Class-Imbalanced Bridge and Weather Data for Supporting Bridge Deterioration Prediction

Abstract

In recent years, Building Information Modelling (BIM) has been pushing forwards the digitalization of global AEC industries continuously. Developing countries with low computerized level, such as China, still use CAD drawings as delivery in building project. It is significant to transform architecture drawings to BIM models in a cost-efficient way. Current researches on 3D reconstruction of architecture drawings are restricted by accuracy and automation level. In this paper, an automated layer classification method is proposed as pretreatment in transforming CAD to BIM models. It analyses the content in each layer of a drawing and classifies the layer into a specific category. Detailed methods to find out grid text layer, dimension layer, window & door layer and wall layer are presented in the paper. The approach is tested using 70 sample drawings. The average accuracy degree of classification is around 95%. Based on layer classification, the existed recognition algorithms could have better performance since obstructs are removed, and the detection method of section drawings can be optimized.