Deep Semantic Fusion and Overlap-Aware Transformer Registration for Intelligent Recognition and 3D Modeling of Power Grid Equipment

Abstract

Artificial intelligence, computer vision, and digital twin technologies have increased the demand for intelligent recognition and dynamic modeling of power grid equipment. Existing methods remain limited by noisy multi-source documents, weak semantic-to-geometric mapping, and unstable registration under low overlap and occlusion. To address these issues, this paper proposes a multi-source semantic fusion framework following the chain of semantic information extraction, parameter-to-model mapping, and dynamic point cloud registration. First, PCNN-ATTRA-RL is designed as a distantly supervised relation extraction model. It uses PCNN encoding, relation-aware sentence-level attention, and reinforcement-learning-based noise filtering to extract equipment objects, attributes, and relations from heterogeneous texts. Second, a semantic-driven parametric modeling method maps equipment type, dimensions, ports, and topology to reusable geometric templates, so 3D models can be generated and assembled automatically. Third, an overlap-aware Transformer registration network combines engineering-rule weights and sliding-window optimization to update the generated model with field point clouds. Experiments show that the proposed framework improves knowledge acquisition, semantic consistency of generated models, and point cloud alignment accuracy.