GNN-Based Semantic Segmentation for Contact Network Insulators

With the vigorous development and wide application of electrified railroad in the world, the research of contact network has been gradually deepened. The state of contact network equipment is closely related to the normal operation of trains, and the failure and geometric parameters of components will directly affect the quality and safety of trains. Realizing the automation and intelligent inspection of each link of contact network of high-speed railroad is the development goal of future railroad inspection, and the intelligence of contact network overhaul task is imminent. As the upstream task of intelligent maintenance of contact network components, the component segmentation of contact network is very important. Due to the limited data characterization ability of traditional point cloud segmentation methods, this work proposes HC-SPG, a Superpoint Graph (SPG) algorithm based on the spatial attention mechanism. This work chooses insulators as the segmentation target, and the experimental results show that our algorithm operates well on the contact network dataset, and the Overall Accuracy (OA) improves by 1.04% and the Mean Intersection over Union (MIoU) improves by 5.21% compared with the SPG algorithm.