Abstract
There is substantial potential for future underground construction operations to be performed by autonomous multi-functional robots. However, a core challenge is accurate digital mapping of the constructed structure, which is essential for optimising robot operations. This study proposes a deep learning (DL)-based framework for 360-degree digital reconstruction of underground structures using in-pipe rotating ground penetrating radar (GPR). Unlike traditional ground-level GPR applications, this study positions the GPR within underground pipes, shortening the distance to the target to reduce signal attenuation, and thus improving imaging accuracy. To alleviate data scarcity for DL training, this paper proposes a high- fidelity in-pipe GPR generator that combines calibrated synthetic data with real-world pipe reflections, clutter, and random noises. First, a ‘stochastic-ellipse-union’ method is introduced to mathematically model the outcomes of robot-constructed structures for a realistic, yet diverse, dataset. Then, a new 2D digital antenna model is proposed and optimized to reflect real-world measurements using a genetic algorithm. Notably, the new antenna reduces radargram generation time by 99.2% compared to traditional 3D models. Lastly, benchmark predictions of the structure maps from the present dataset using seven popular DL networks are also provided. The best performance was achieved using U-Net enhanced by ResNet101, with an intersection-over-union score of 0.937 demonstrating the feasibility of in-pipe GPR for digital reconstruction of underground structures. To the authors’ knowledge, this is the first general framework for the 360-degree reconstruction of underground structures. Moreover, this framework can be extended to other pipe-related applications such as pipe defect identification and pipe joint rehabilitation.