Multi-domain transfer generation of cavity defect data in asphalt pavements using 3D GPR and 3D forward modeling

Abstract

The paucity of GPR data pertaining to cavity defects significantly impedes the advancement of intelligent nondestructive testing methods in pavement engineering. This paper illustrates that heterogeneous forward models of cavity defects, constructed using pseudo-random generation algorithms, exhibit remarkable accuracy in mimicking the electromagnetic responses within asphalt pavement structures. A unified multi-domain transfer learning framework, employing StarGAN, facilitates the cross-domain generation of data representing cavity defects in asphalt pavements. The model effectively suppresses clutter interference, thereby preserving cavity defect characteristics in heterogeneous forward images, while adeptly synthesizing signals conforming to heterogeneous structural properties in homogeneous forward images. Quantitative assessments reveal an exceptionally high degree of similarity between the synthetically generated data and actual samples (LPIPS≈0). The measured cavity defect features generated by StarGAN exhibit high physical regularity and morphological diversity compared to real samples (LPIPS<0.1). This paper introduces a novel approach to data augmentation for GPR applications in asphalt roads.