Detection and diagnosis of tunnel lining apparent defects: a deep graph clustering approach

There have been rapid advancements in detecting defects using convolutional neural networks (CNNs) in tunnelling; however, the application of these detection results for diagnosing apparent defects remains underexplored. This paper proposes a method to diagnose apparent tunnel defects using graph clustering. Initially, label images segmented by CNNs serve as the foundational data. Key parameters such as length, maximum width, number of fractal dimensions, and area are separately calculated for four prevalent types of apparent defects: cracks, leakage, segment damage, and falling off of fireproofing coating. Subsequently, within a fixed-length tunnel as the diagnostic scale, the cumulative values of these defect parameters are utilised to construct the graph data. Defects are clustered using a hybrid HSAN graph neural network (GNN) and K-means algorithm. A Tunnel Risk Indicator for Apparent Defects (TRI-A) is then developed via regression analysis. The results demonstrate that this method effectively diagnoses various apparent tunnel defects, significantly aiding efficient tunnel operation and maintenance.