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

54 4

2.2. Artificial intelligence model Fig. 1 shows flowchart of this study. As shown in Fig. 1a, the process begins by loading scalogram and FFT (or STFT) images derived from GPR and IE signals. A random selection of images ensures balanced dataset representation. Each selected image pair (Wavelet + FFT or Wavelet + STFT) is flattened into numerical feature vectors and labeled as either sound (CLS) or defective (CLD). Two separate preprocessed datasets are created for model training—one for each image combination (Fig. 1b). Two dataset groups were considered in this study, as illustrated in Fig. 1b.

b)

a)

c)

Fig. 1. Flowchart of study with main steps, a) flowchart, b) input, c) feature section.

Each sample is represented as a 3D tensor of size 224×224×4, where the first two dimensions define spatial resolution and the four channels represent two wavelet scalogram features and two FFT features (magnitude and phase). This structure captures both spatial and spectral characteristics of radar-derived images, forming a rich representation for classification (Fig. 1c). To ensure balanced learning, equal numbers of samples from each class are selected. Features are standardized using Z-score normalization (zero mean and unit variance), improving numerical stability and ensuring all features contribute equally during model training. Each decision tree uses high-dimensional pixel-based features to distinguish between CLS and CLD classes. During training, the model automatically identifies optimal feature–threshold pairs that minimize Gini impurity, forming supervised boundaries that maximize class separation. Each node represents a learned decision rule based on the intensity distribution of specific image pixels. Multiple decision trees are combined using bagging (bootstrap aggregation). Each tree is trained on a random subset of data, and their outputs are aggregated through majority voting. This ensemble approach reduces overfitting, increases robustness, and enhances generalization by leveraging diverse decision boundaries. Model performance was assessed using cross-validated accuracy, confusion matrix, and key metrics such as True Positive Rate (TPR) and True Negative Rate (TNR), ensuring reliable and reproducible defect classification in bridge materials. This classification framework was applied to both Impact Echo (IE) and Ground Penetrating Radar (GPR) datasets, allowing the model to capture complementary frequency–time and subsurface reflection patterns. By jointly analyzing features derived