Robotic subsurface defect inspection system and an unsupervised deep neural network-based abnormal traces reconstruction method for ground penetrating radar data autonomous collection

Abstract

Abstract Ground-penetrating radar (GPR) has been widely utilized for inner structural defect inspection of infrastructure. However, automatic GPR control for inspecting the infrastructure and collecting high-quality GPR data remain challenging. To address such limitations, we developed a GPR-based robotic inspection system inner defect inspection and an unsupervised deep learning-based abnormal trace reconstruction network tailored for data postprocessing. A GPR status feedback module and an adaptive GPR scanning control method were developed, and these were assembled on the robotic inspection system to manipulate the GPR to scan over complex infrastructure surfaces and avoid obstructions. Furthermore, a cycle-consistent adversarial network with a distance attention mechanism was devised to process the abnormal traces induced by the automatic GPR manipulation such that the high-quality GPR data were gained. Both indoor and outdoor experimental results demonstrate the adaptability of the GPR-based robotic inspection system. Additionally, the superiority of the reconstruction network was validated through comparative experiments.