PSI - Issue 64

Mercedes Solla et al. / Procedia Structural Integrity 64 (2024) 293–300 M. Solla et al. / Structural Integrity Procedia 00 (2019) 000–000

299

7

Fig. 8. Preliminary results of using the GPR Insights software tool for AI-assisted hyperbola detection of the first layer of rebars.

Another experiment is applied using a custom YOLOv8 with a model trained with rebar samples from the DECKGPRH1.0 dataset (Asadi et al. 2019) and the SDNET2021 dataset (Ichi and Dorafshan 2021). The dataset has a variety of rebar samples from two different pulsed radar systems (a GSSI SIR System 3000 with 1.6 GHz and 2.6 GHz antennas, and a MALA Geoscience ProEx System with a 2.3 GHz antenna) and was trained at various scale levels. The model was applied to the SFCW data produced by Specimen III. Observing Figure 9, the preliminary results have shown comparable results in detectability than those obtained using the GPR Insights software tool (see Figure 8).

Fig. 9. An example of the prediction results using a custom YOLOv8 model trained with data from pulsed radar systems.

5. Conclusions In this study, the comparison between the time domain and frequency domain systems revealed similar behavior in detection capabilities, as both systems showed consistent findings and limitations. It is noteworthy that the SFCW system provided more detailed imaging and higher reproducibility of the signatures which was somehow expected due to its larger bandwidth. However, it should be remarked that overlapping arrangement of reinforcement bars may exceed the resolution capabilities of both GPR systems, making difficult to detect and characterize individual bars. In this context, future studies may explore strategies to mitigate the challenges posed by overlapping features, such as optimizing survey parameters, implementing advanced data processing techniques, or utilizing alternative GPR system configurations. Further works will include collecting data with both parallel and perpendicular polarizations to gain a more comprehensive understanding of the subsurface structures and properties, with the purpose of optimizing the detection, as recommended for similar applications (Rasol et al. 2022b). Additionally, manual data collection will be considered to avoid inaccuracies with the survey wheel. Later, further research will be focused on the development of a DL detection model trained with both pulse radar and SFCW GPR systems to gain more accurate rate in rebar detection.