Heterogeneous Graph Neural Networks-Based Paper Drawings Automatic Layering Method

Building Information Modelling (BIM) can significantly improve aging buildings’ smart operation and maintenance (O&M) efficiency, and support advanced equipment maintenance and emergency response. However, many aging buildings lack BIM due to the era in which they were built, and the manual reconstruction process requires experts with specialized knowledge and consumes considerable modeling time. Many existing studies utilize deep learning methods based on image features to extract component information from paper drawings for BIM reconstruction, however, the paper drawing layering pre-processing, which is essential for improving the information extraction accuracy, still requires inefficient manual drafting and line classification. Although existing studies have proposed methods for drawings layering by line classification, these methods perform poorly on construction engineering drawings due to the complexity and line dense in drawings. To fill this gap, we propose a heterogeneous graph neural networks (GNNs) based method that predicts the category of line elements on paper drawings to achieve automatic layering with three modules: 1) paper drawing vectorization by line extraction and duplicate elements merging to detect the line elements representing component contour and annotation; 2) graph structure construction by considering the different topological relationships among lines to represent the line properties; 3) heterogeneous graph nodal classification model to predict the line category and realize automatic layering. The proposed method was tested on an actual engineering drawing dataset, and the results show that the method has an overall F1 score more than 0.74 and exceeds the baseline model by over 0.1. This research improves paper drawing pre-processing efficiency and provides a new solution for information extraction in ageing buildings BIM reconstruction.