Abstract: The intelligent identification of complex rock-mass discontinuities in tunnels faces multiple challenges, including multi-scale feature loss and low segmentation accuracy. To address these issues, an enhanced geometric segmentation network (GeoSegNet++) that integrates geometric features is proposed to enable intelligent segmentation and extraction of occurrence parameters of rock discontinuities. First, the model employs a deformable geometric convolutional network (DGCv2) to dynamically adjust the neighborhood sampling range, thereby effectively capturing local geometric features. Second, a multi-modal interaction pyramid is constructed that utilizes a cross-level attention mechanism to adaptively fuse multi-scale geometric features, which significantly improves boundary segmentation accuracy for complex discontinuities. Finally, an adaptive density clustering algorithm is introduced to automatically identify dominant sets of discontinuities based on normal vector angular distance constraints and local density peak detection. The proposed method is validated using two types of data: regular geometric bodies and tunnel-scale point cloud data. The experimental results demonstrate the following: (1) For regular geometric bodies, the difference in dip direction of the dominant discontinuities extracted by the proposed method is less than 3° compared with those identified using a discontinuity set extractor. (2) For tunnel-scale point cloud data, the discontinuity surface parameters extracted by the proposed method exhibit an orientation error of up to 3.21° relative to the dominant discontinuity surface group in the actual data, with dip angle errors below 1°.

Key words: tunnel, surrounding rock discontinuity surface, intelligent identification, deep learning, dominant grouping, cluster analysis