Clustering and diagnosis of crack images of tunnel linings via graph neural networks

Abstract

ABSTRACT Intelligent detection of the apparent defects of tunnel linings on the basis of deep learning has become a mainstream trend, and how to objectively diagnose the hazard level of tunnel defects via semantic segmentation images is the key to intelligent control of tunnel health. In this study, a tunnel lining defect image diagnosis method that uses graph neural networks is proposed. First, the binary images obtained by semantic segmentation are used as the data set, and quantitative parameters, such as crack length, maximum width, box dimension, and area density, are calculated according to the crack orientation and employed as graph node attributes. Second, the construction of graph data is based on the similarity of defect attributes. Third, clustering is completed using graph neural networks and K-means, and the number of clusters and danger levels of crack defects are reasonably determined. Last, a tunnel crack risk index is developed via partial least squares regression analysis, and the grading criteria for defect levels are established. Results show that the method is effective in extracting the key attributes of crack images and clustering them into different hazard levels, which is important for maintaining tunnel health.