Image piping and instrumentation diagrams (P&IDs) are drawings expressed in pixels that do not contain any engineering attribute information. With the advancement of the plant industry in recent years, attempts are being made to produce intelligen...
Image piping and instrumentation diagrams (P&IDs) are drawings expressed in pixels that do not contain any engineering attribute information. With the advancement of the plant industry in recent years, attempts are being made to produce intelligent P&IDs by adding attribute information to P&IDs for use in design modification and inspection, volume calculation, and maintenance. However, because drawings produced in the past often consist of image P&IDs, research is being conducted to identify ways to convert them into intelligent P&IDs.
To reconstruct intelligent P&IDs from image P&IDs, the symbols, texts, and lines in the image P&ID must be detected and their phase structures must be reconstructed. During the reconstruction of the phase structures, it is important to distinguish between piping lines and signal lines. However, it is difficult to distinguish the lines solely based on their type, thus hindering the automatic conversion of image P&IDs to intelligent P&IDs.
To this end, this study proposes an approach using graph neural networks (GNNs). First, 20 drawings were created using Microsoft Visio to generate data used for training. In addition, the application programming interface of Visio was used to generate data in XML, which were divided into 12 training sets, 4 validation sets, and 4 test sets. Then, a GNN model was defined and trained using the training data. To obtain the optimal model, hyperparameter tuning was performed with the validation data using the random search method. Finally, the performance of the final model was evaluated using the test data, resulting in an accuracy of 93.06%.