PSI - Issue 82

Faezeh Jafari et al. / Procedia Structural Integrity 82 (2026) 51–57 F. Jafari and S. Dorafshan / Structural Integrity Procedia 00 (2026) 000–000

56 6

3.1. Combination of impact echo with GPR Given that the combination of FFT and Wavelet transforms consistently produced better classification results in both GPR and Impact Echo (IE) datasets, an experiment was conducted using 237 randomly selected samples per class from each modality. The results reveal that although the overall accuracy rate did not significantly improve through data fusion (remaining around 83%), the combination of both datasets contributed to a more balanced performance across classes (Table 4). Specifically, the True Negative Rate (TNR) for detecting defective areas increased from 81% (using GPR alone) to 85% when combining both modalities. This improvement can be attributed to the higher sensitivity of GPR to defect-related signal patterns. Conversely, Impact Echo exhibited superior performance in identifying sound (non-defective) regions, achieving a TNR of 87% compared to 72% in GPR-only signals. By integrating both datasets, the model achieved an intermediate TNR of 82% for sound classification, suggesting that the complementary strengths of each modality can be effectively harnessed through ensemble learning. Importantly, while accuracy remains relatively constant, the fusion strategy reduces the gap between TPR and TNR, resulting in a model with more balanced sensitivity and specificity. This indicates improved generalization and robustness, especially in real-world applications where both defect detection and false positive avoidance are critical.

Table 4. Trained model-based impact echo signals.

Dataset

Input

TPR 72%

TNR 92%

Accuracy

Precision

F1-Score 80.50% 85.20% 82.20%

GPR only

FFT + Wavelet FFT + Wavelet FFT + Wavelet

82%

91.10% 82.60% 84.40%

IE only

87.80%

81.50%

84.70%

GPR + IE

82%

85%

83%

4. Conclusion This study presented a comparative study of Ground Penetrating Radar (GPR) and Impact Echo (IE) signals for the detection of delamination in reinforced concrete bridge decks. Time-frequency transformations including FFT, STFT, and Wavelet were applied to convert raw signals into 2D scalogram representations, which served as inputs to an ensemble learning model. Results showed that: • The combination of FFT and Wavelet transformations consistently outperformed STFT-based inputs. • The Impact Echo (IE) signals achieved a classification accuracy of 85%, demonstrating their effectiveness in detecting delaminations in the reinforced concrete bridge deck. This high accuracy indicates that the IE signals, which are typically sensitive to changes in the structural integrity of concrete, were able to correctly classify a large proportion of both sound and delaminated regions. • In comparison, the Ground Penetrating Radar (GPR) signals achieved a classification accuracy of 82%, which is slightly lower than that of the IE signals. While GPR is also a widely used technique for NDE, its slightly lower performance in this study could be attributed to factors such as the frequency of the radar waves, material heterogeneity, or the complexity of the delamination patterns within the concrete. • IE signals better identified sound areas, while GPR was more sensitive to defects. Combining both modalities increased model robustness and improved specificity by 5% without reducing overall accuracy. References Carino, N.J., Sansalone, M., Hsu, N.N., 1986. Flaw detection in concrete by frequency spectrum analysis of impact-echo waveforms. International Advances in Nondestructive Testing 12, 117–146. Clausen, J.S., Zoidis, N., Knudsen, A., 2012. Onsite measurements of concrete structures using impact-echo and impulse response. In: Emerging Technologies in Non-Destructive Testing V. CRC Press, pp. 117–122. Coleman, Z.W., Schindler, A.K., 2022. Investigation of ground-penetrating radar defect detection capabilities, influence of moisture content, and optimal data collection orientation in condition assessments of concrete bridge decks. Journal of Applied Geophysics 202, 104655.