In order to accurately evaluate the safety of existing structures, it is desirable to obtain not only surface information but also detailed internal cracks information in concrete structures. This paper presents a fundamental study of a method for visualizing internal cracks in concrete structures using GPR and GAN. Specifically, radar measurements are conducted on concrete specimens containing artificial defects at different locations and sizes. From this experiments, a number of training data sets are acquired. After that, we validated the accuracy of crack visualization of the network model obtained by training the acquired data sets. In addition, a FDTD is used to reproduce the experiments, and the possibility of using the simulation-generated training data sets is also verified. The results of the study show that the proposed method is able to approximately estimate the location and size of artificial defects. It is also found that the FDTD simulation has the potential to generate a large amount of significant training data. However, it is found that there are currently challenges in modeling the antenna and non-homogeneity.

In the future social infrastructure diagnosis for concrete structures such as bridges, tunnels and buildings, it is expected that the digital hammering inspection, which is objective, quantitative, and recordable to the conventional hammering inspection, will play a significant role. However, since there are a wide variety of defects due to poor construction and age-related deteriorations in the concrete structures, they have not been quantitatively identified enough by the digital hammering inspection in terms of their scales (size, depth from the surface, etc.). Also, establishing the data base of defects and deteriorations by mockup testing is not practical. As an application of machine leaning, an alternative model is proposed in this study that complements the experimental data of digital hammering inspections with the FEM analysis model and further expands the data by machine learning.