SWC-Net and Multi-Phase Heterogeneous FDTD Model for Void Detection Underneath Airport Pavement Slab

Void underneath a slab is a common damage type in airport cement concrete pavement. Existing methods mainly use ground penetrating radar (GPR) images and deep learning to detect the voids but face two problems: inherent limitation of image-wise deep learning and imperfect GPR datasets. This study has proposed a signal-wise cascade deep network (SWC-net) to detect voids underneath slabs, which was trained by a dataset with real-world and simulation GPR signals. In the study, a multi-phase heterogeneous pavement model was first built, in which each pavement layer was reconstructed based on the real-world morphologies and distributions of aggregates and other materials. The model was then used for the finite-difference time-domain (FDTD) simulation of GPR signal propagation. The simulation signals were merged with real-world GPR signals to generate a void signal dataset covering comprehensive pavement conditions. Finally, the proposed SWC-net was trained by the merged dataset to detect voids. The inputs of the network were the GPR signals, while the outputs were the signal classes and abnormal intervals of voids to perform signal-wise void detection. The experiments on Nanjing Dajiaochang and Xuzhou Guanyin airports showed that the FDTD results generalized the signal dataset by simulating different real-world conditions, while the proposed network outperformed the image-wise deep networks on void detection.