关键词: 地下洞室群, 多模态感知, 图像, 点云, 点云配准, Geotransformer, 数字孪生建模

Abstract: High-precision 3D models serve as fundamental tools for the digitalization and informatization of hydraulic projects. As a high-fidelity approach, digital twin modeling enables dynamic and accurate mapping between physical entities and virtual models. However, previous modeling technologies have faced dual challenges in complicated hydraulic project scenarios－single-modal sensors suffer from blind spots and occlusions, while multimodal sensing data can compensate for these defects but often leads to poor performance in data registration and data fusion due to their large density differences and low overlap rates. This paper develops a new method for modeling digital twin underground cavern groups using image and point cloud multimodal sensing technologies. To overcome the limitation of single-modal sensors, this method adopts 3D image reconstruction and 3D laser scanning to acquire cross-source point cloud data of hydraulic underground cavern groups, fully perceiving the spatial information of the groups. We design a Dual-path Feature Fusion Attention (DPFFA) module and embed it into the Geotransformer backbone network to enhance its network capability of fusing the multi-scale features of global structures and local details, avoiding single-path attention failure caused by cross-source point cloud density differences and data missing. Then, we achieve a complete and high-accuracy digital twin model based on the accurately registered cross-source point clouds. This method is applied to the underground cavern group at the Kala hydropower station, and tested against the point cloud datasets with overlap rates of 20%, 40%, 60%, and 80%. Results show its registration recalls (RR) up to 26.9%, 72.5%, 84.4%, and 86.9% respectively, or an increase of 3.2%, 11.2%, 8.2%, and 2.5% respectively in comparison to those of the unimproved model. This demonstrates the superiority of our registration approach and validates the accuracy of the new model, thus promoting digital twin modeling improvement.

Key words: underground cavern, multimodal perception, image, point cloud, point cloud registration, Geotransformer, digital twin modeling