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

53 3

2. Method The Impact Echo (IE) and Ground Penetrating Radar (GPR) signals were collected from the Park River Southbound Bridge located in Grand Forks, North Dakota. For this bridge, four regions, labeled A, B, C, and D, were selected for IE data acquisition. Each region was gridded using an 11 × 11 layout, with a uniform spacing of 0.3 meters between adjacent grid lines in both longitudinal and transverse directions. A chain dragging method was initially employed to qualitatively inspect the bridge deck and to identify potential delaminated areas. GPR data were collected across the entire bridge deck. The GPR acquisition parameters included a vertical time scale of 12 nanoseconds and 512 samples per scan. Approximately 16,383 amplitude values were recorded along the length of the bridge, with an additional 1,225 amplitude points collected across the width. However, for the purpose of this study, only the GPR data corresponding to the four selected regions (A, B, C, and D) were utilized for delamination detection and analysis. The GPR and IE signals were annotated using a ground truth map derived from field inspections and chain dragging results. These annotations enabled the differentiation between sound and delaminated areas within the regions of interest. The Park River Southbound Bridge was selected for this investigation due to its relatively high prevalence of delamination compared to other surveyed bridges, resulting in a more balanced dataset between defected and sound signals. This balance was critical for training and evaluating machine learning models effectively. From both datasets, random signals were collected from both groups (defect (CLD) and sound (CLS)) to create a dataset for delamination detection. It notably mentioned that location defect area has been determined based on chain dragging output. Table 1 shows final dataset of this paper for all datasets. 2.1. Filtering Approach Impact Echo (IE) signals were preprocessed using the filtering method described in Carino et al. (1986) where all signal points with amplitudes below 10% of the peak value were removed. This step reduced noise and enhanced the localization of meaningful transient features in the time–frequency domain, resulting in sharper and more discriminative wavelet scalograms for model training. In contrast, no filtering was applied to signals used for FFT analysis to preserve the true frequency content and peak locations essential for identifying resonant behavior related to delamination. Thus, distinct preprocessing strategies were applied to optimize feature extraction for each transformation. Continuous Wavelet Transform (CWT) was employed to generate time–frequency representations (scalograms) for non-stationary signals such as GPR data. Using the Morlet wavelet—previously shown effective for delamination detection—the CWT computed correlations between the signal and scaled versions of the mother wavelet to extract localized frequency information. The resulting scalograms visualize how frequency content evolves over time and will later be compared with other wavelet type (Clausen et al ,.2012). Fast Fourier Transform (FFT) converted the entire unfiltered signal from the time domain to the frequency domain, capturing dominant frequencies that characterize the material condition. This global frequency representation was preserved without preprocessing to maintain accurate identification of peak frequencies critical for signal classification. Short-Time Fourier Transform (STFT) provided a joint time–frequency representation, allowing analysis of how frequency components change over time. Like FFT, no filtering was applied prior to transformation to retain the true location of local peaks. Together, the CWT, FFT, and STFT transformations enabled comprehensive representation of both temporal and spectral features, enhancing defect classification performance. Table 1. Dataset for Artificial intelligence model in this study. Origin Original A Original B Original C Original D Total signals Class D S D S D S D S D S IE signal’s Number 11 110 94 27 94 27 38 83 237 253 GPR signal’s Number 138 1635 1418 355 1338 436 615 1159 3509 3585