Deep Learning–Based Point Cloud Segmentation and Relationship Graph Construction for Hydropower Dam Digital Twins

Abstract: This paper proposes a deep learning based pipeline for generating semantic meshes and relationship graphs from intensity only laser point clouds of hydropower plants. The approach segments point clouds into object-level instances using supervised models and derives spatial relationships such as proximity and containment between components. These relationships are then refined into a structured graph, providing a lightweight yet informative representation suitable for digital twin applications. By replacing traditional rule-based methods with learning-based segmentation, the pipeline achieves higher accuracy and produces more reliable relationship graphs, enabling scalable and automated modeling of complex dam infrastructure.