Research and Application of Operational Tunnel Structure Disease Prediction Based on Graph Attention Network

With the rapid development of underground engineering in China, more and more subway tunnels have been built, the operation mileage of subway tunnels has been on the rise, and the corresponding tunnel structure diseases have become increasingly prominent. At present, tunnel structure diseases are mainly treated by manual inspection and identification, and the research on disease prediction is still inadequate. Because of the complexity of factors leading to tunnel structure diseases, it is difficult to analyze the causes and development trends of diseases comprehensively via manual analysis. In addition, over recent years, deep learning has achieved great success in extracting data features (for classification and prediction), and applying deep learning algorithms to graph data is one of the most popular research directions currently. In this paper, graph attention network is introduced into tunnel structure disease prediction for the first time. Based on the characteristics of tunnel structure safety data, a spatio-temporal network is constructed, and the graph attention network is trained to predict the development trend of tunnel structure disease, which provides a new idea for tunnel disease prevention and control. In engineering practice, abundant tunnel structure safety data (68055 ring tunnel segments) were taken as the research object, both the prediction precision and recall rate of the trained model are over 80%, and the prediction results can help with auxiliary decision-making of tunnel maintenance departments and relevant government supervision departments to prevent and control tunnel structure diseases, which focus on tunnel sections where serious diseases may occur, thus further clarifying the development trend of tunnel diseases.