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

55 5

from these two nondestructive testing methods, the approach enhances defection detection reliability and enables cross-validation of subsurface anomaly predictions across sensing modalities. 3. Result Table 2 presents the performance of different signal processing methods—STFT and Wavelet versus FFT and Wavelet—applied to GPR signals for classifying defects versus sound areas using 2D image scalograms. The performance is evaluated based on True Positive Rate (TPR), True Negative Rate (TNR), and Accuracy (Acc) across various training set sizes: 237, 500, 1000, 2000, and 3000 signals per class. The K-fold cross-validation results show that both Accuracy and TPR improve notably when the training set increases to 1000 signals per class. Beyond this point, the performance metrics stabilize, indicating minimal gains with additional data. However, increasing the dataset size beyond 1000 signals still contributes to a more generalized model compared to models trained with smaller datasets.

Table 2. Primary output Performance matrix based GPR 2D image scalogram.

Defect/Sound Signals

F1-score (%)

Method

TPR (%)

TNR (%)

Accuracy (%)

Precision (%)

STFT + Wavelet FFT + Wavelet STFT + Wavelet FFT + Wavelet STFT + Wavelet FFT + Wavelet STFT + Wavelet FFT + Wavelet STFT + Wavelet FFT + Wavelet

72 74 71 70 62 72 68 70 68 72

72 60 80 88 93 92 90 92 92 92

72 67 75 80 78 82 79 81 80 82

72

72

237 / 237

68.1

70.9 74.3 77.3 73.2 80.5 77.2 78.7

78

500 / 500

86.4 89.9 91.1 89.5 90.2 90.6 90.6

1000 / 1000

2000 / 2000

78

3500 / 3500

80.3

The comparison between GPR and Impact Echo signals for defect detection reveals that both signal types perform well when processed using FFT combined with Wavelet transforms; however, Impact Echo signals achieve comparable or even better performance with significantly fewer training samples. Specifically, Impact Echo achieves an accuracy of 84% and a TPR of 87% using only 237 signals per class (Table 3), matching the F1-score (80.5%) of GPR at 1000 signals per class. While GPR slightly outperforms in precision (91.1% vs. 82%) when more data is available, the performance gains plateau beyond 1000 samples, suggesting diminishing returns. Therefore, based on Table 3, Impact Echo signals offer more efficient training with strong generalization, whereas GPR requires a larger dataset to achieve similar performance.

Table 3. Trained model-based impact echo signals.

Defect/Sound Signals

Precision (%)

Input

TPR (%)

TNR (%)

Acc (%)

F1-score (%)

STFT + Wavelet FFT + Wavelet STFT + Wavelet FFT + Wavelet STFT + Wavelet FFT + Wavelet

80

79

79

78.78 82.63 76.64

79.39 85.15 79.23

237/237

87.84

81.53

84.68

82 88 66 75

75 76 72 82

78.5

100/100

82 69 79

79

83 68

70.21

50/50

78.4

76.5