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

52 2

Different numerical approaches have been used to analyze data or images. For example, IE and GPR signals have been analyzed both numerically and empirically. The time-domain approach was initially selected as the primary method to interpret IE data, but it was time-consuming and suitable only for simple test geometries (Clausen et al., 2012). Researchers widely accept and use data collected from the frequency response of the IE device. For instance, Carino et al. (1986) used IE waveform data to indicate debonding and subsurface defects in bridge decks with cement and asphalt overlays. Employing the frequency response approach, the IE device was applied to four concrete bridge decks—one control sample and three asphalt decks—to monitor the structures before and after overlay installation. Chang and colleagues applied time–frequency and Morlet Wavelet transform methods to indicate crack responses and rebar locations. The peak frequencies in the FFT spectrum were analyzed to obtain the amplitude histogram at selected frequencies. The results showed that signals reflected from the concrete–steel interface were less noticeable, with weaker and shorter reflective energy at rebar locations Carino et al (1986). Predicting the P-wave velocity is difficult when selecting the P-wave from the first arrival, and both Rayleigh and P-waves are affected by the near-field zone, where the wave velocity is lower than the theoretical value. Medina and Bayón (2010) combined IE and shear wave approaches to inspect crack width and depth, demonstrating that the NDT approach can effectively assist in visual bridge inspection. Kee and Gucunski (2016) used thin-plate vibration theory and a numerical model to develop fundamental frequency equations. They also analyzed a simple numerical model to predict the flexural vibration of a concrete plate, achieving better interpretation of delamination areas than previous methods. Epp et al. (2018) implemented an Artificial Neural Network (ANN) to identify faulty placements in reinforced concrete beams, showing that the semi-automated IE approach saves time and cost. Furthermore, Jafari and Dorafshan (2021, 2022) analyzed the time–frequency domain of IE data using the Short-Time Fourier Transform (STFT), applying both supervised and unsupervised learning for delamination detection. GPR has been used both independently and in conjunction with IE to detect delamination in concrete structures. Yehia et al. (2005) compared GPR and IE in detecting simulated defects in fabricated bridge decks and validated the complementary performance of both techniques. Similarly, Gucunski et al. (2005) used IE together with GPR to produce condition assessment maps of bridge decks, correlating GPR results to determine suitable thresholds for distinguishing sound and deteriorated sections. Coleman and Schindler (2022) investigated GPR’s defect detection capabilities across eighteen specimens and found that GPR effectively detects corrosion and water-filled voids in the cover layer of concrete structures. Sun et al. (2018) employed a scanning system and GPR to detect delamination; their study extracted attenuation maps related to concrete deterioration and reinforcement corrosion, showing that combining acoustic and GPR scanning offers a comprehensive evaluation of concrete bridges. Dinh and Gucunski (2021) analyzed real GPR data and concluded that rebar proximity to delamination areas can influence model performance in detecting defects. Despite these advancements, there remains a limited body of research focusing on the comparative analysis or integration of GPR and IE signals for detecting delamination in real-world bridge structures. Although both techniques are widely employed in Nondestructive Evaluation (NDE) to distinguish defective regions from sound areas, they rely on fundamentally different physical principles and signal characteristics. While Artificial Intelligence (AI) has been increasingly applied to process these modalities, few studies have directly compared their performance or explored data fusion strategies. The complementary features extracted from GPR and IE signals have the potential to enhance detection robustness and reliability when applied to real bridge inspections. The primary objective of this study is to develop neural network– based classification models using GPR and IE data independently, evaluate their respective accuracies in delamination detection, and investigate the effectiveness of multimodal data fusion in improving the classification of sound versus defective regions. To achieve this objective, a dataset comprising GPR and IE signals collected from an actual Precast Reinforced Slab Bridge (PRSB) was used to classify sound versus defective signals. All acquired signals were transformed using Fast Fourier Transform (FFT), Short-Time Fourier Transform (STFT), and Wavelet Transform to generate corresponding 2D scalograms for both datasets. Feature extraction was performed by computing pixel-wise intensity information from these 2D representations. These features were then used as inputs to an ensemble learning framework to classify sound and defective signals independently for each dataset and for the combined feature space derived from both modalities. Given that FFT, STFT, and Wavelet transforms are widely adopted for signal-based defect classification, their application in this study ensures methodological consistency and comparability with previous literature.