Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1139 https://internationalpubls.com Deep Learning for Real-Time Crack Detection in Concrete Using Electromagnetic Imaging Isha Katariya Assistant Professor, Civil Engineering, Deogiri Institute Of Engineering And Management Studies, Ch.Sambhaji Nagar Maharashtra Email Id: Ishakatariya@Dietms.Org Vivek Ashok Ballal Assistant Professor, Electronics & Telecommunication Engineering, Deogiri Institute Of Engineering And Management Studies Ch.Sambhaji Nagar, Maharashtra Email Id: Vivekballal@Dietms.Org Megha R. Pardeshi Assistant Professor, Computer Science And Engineering, Artificial Intelligence & Machine Learning Deogiri Institute Of Engineering And Management Studies, Ch.Sambhaji Nagar, Maharashtra Email Id: Meghapardeshi@Dietms.Org Yogita Lakade Assistant Professor, Cse. Aiml, Deogiri Institute Of Engineering And Management Studies, Ch.Sambhaji Nagar, Maharashtra Email Id: Yogitalakade@Dietms.Org Amruta D.Agharde Assistant Professor, Civil Engineering, Deogiri Institute Of Engineering And Management Studies, Ch.Sambhaji Nagar Maharashtra Email Id: Amrutaagharde@Dietms.Org Article History: Received: 01/10/2024 Revised: 06/11/2024 Accepted: 10/12/2024 Abstract: In total, crack detection has been very important in the durability and safety of concrete infrastructure and must be undertaken correctly and on time. The conventional forms of inspection are time-consuming, surface-restricted and not real-time. This study attempts to offer deep learning-based solution by implementing electromagnetic (EM) imaging so that real-time crack detection on surface and sub surface of concrete structures could be possible. To train the 4 deep learning approaches, including Convolutional Neural Network (CNN), Residual Neural Network (ResNet-50), U-Net and MobileNetV2, 5,000 EM images were utilized. Within this, the one that succeeded in classification with mailto:ishakatariya@dietms.org mailto:vivekballal@dietms.org mailto:meghapardeshi@dietms.org mailto:yogitalakade@dietms.org Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1140 https://internationalpubls.com the best accuracy at 96.2% was ResNet-50 which gained precision of 95.6%, F1-score of 95.3%; U-Net exhibited better segmentation at an IoU of 89.2%. MobileNetV2 had the lowest inference time of 12 ms per image, implying that it is good to be used in real- time applications. The performance of the results is better compared to the current methodology employing the use of conventional GPR or RGB imaging and shows the promise of deep learning-based EM systems in non-destructive concrete inspection. This paper identifies a scalable and automated infrastructure monitoring tool that will help manage assets safely and more efficiently. Keywords: Crack Detection, Deep Learning, Electromagnetic Imaging, Concrete Structures, Structural Health Monitoring I. INTRODUCTION The stability and strength of concrete buildings is of essence in making civil infrastructure, such as bridge, tunnels, and buildings as well as highways, safe and long lasting. Concrete is prone to different kinds of deteriorations with some surface and internal cracks constituting some of the worst and most frequent concrete defects. These cracks need to be identified as early as possible so as to facilitate proactive maintenance, cost efficient repairs as well as avoiding disastrous structural collapse [1]. The common crack detection processes include visual examination and tribimeters, schematic evaluation by which methods are tedious, subjective, and may also be not very accurate. In addition to this, they cannot provide a real- time inspection and they can miss underground anomalies [2]. Recent developments in non- destructive testing (NDT) brought about a series of promising techniques involved in visualizing internal structural defects, with no physical intrusions, namely electromagnetic (EM) imaging. EM imaging, to detect changes in dielectric properties of concrete, also can show the cracks and empty spaces that are hidden internally and not visible on the surface [3]. The interpretation of ED imaging data is however complex and time consuming and requires expert opinion. The combination of deep learning and the EM imaging can become a groundbreaking approach to real-time and automated identification of cracks. Convolutional neural network (CNN) Deep learning models have seen massive successes in image classification, segmentation, anomaly detection tasks in multiple spheres. These models used on EM imaging data have the ability to learn complex patterns with cracks and give accurate and quick judgments. This study addresses the methods and technicalities of the development and application of deep learning algorithm to improve crack detection in concrete building using EM imaging. The project seeks to develop a real time automated smart system which is able to detect and locate cracks in a precise manner. The proposed method allows improving the depth sensitivity of electromagnetic imaging and pattern recognition capabilities of deep learning which can be used to overcome weaknesses of conventional structural health monitoring and lead to a novel category of smart, AI-based structural health monitoring of target structures. II. RELATED WORKS Current innovations in the field of structural health monitoring (SHM) regarding monitoring of concrete structures, in particular, have become more interested in combining non-destructive testing (NDT) methods and artificial intelligence (AI)-based diagnosis. In harsh conditions Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1141 https://internationalpubls.com (marine or underground) concrete buildings suffer an escalated rate of deterioration, and thus scientists refer to more durable inspection and detection techniques. Lee et al. [15], recommended a general review on surface coatings and inspection technologies that are specific to a marine exposed concrete structure. They gave an emphasis in early detection of defects using automated inspection systems to ensure that life of a service is longer and maintenance cost reduced. Important electromagnetic methods are ground penetrating radar (GPR) that has potential in subsurface crack detection. Li et al. [16] proposed a multi- domained feature extraction method on the GPR-aliased signals representing a scenario where signal processing may be used to improve crack detection in noisy environments. Convolutional neural networks (CNNs) have grown to be an effective way of automating the process of detecting the crack with the application of deep learning models. Liu et al. [17] applied deep CNNs to capture concrete surface cracks simultaneously using binocular video to obtain quantification on the crack at a pixel level automatically. Their system recorded a high spatial accuracy and this means that it can be used to evaluate damaged sites in detail. On the same note, Liu et al. [18] used GPR imaging in tandem with CNN based classification models to classify internal damage in concrete beams, prompting the importance of integrating electromagnetic imaging with deep learning towards more interpretable models. Going beyond detection to prediction, Luo and Matsumoto [19] introduced a deep learning method of tensile strength of ultra-high-performance fiber-reinforced concrete (UHPFRC) being estimated as a factor of predicted locations of cracking. This is what can be seen as a transformation of the passive inspection into predictive AF with the use of AI. Sensing technologies have continued to be a leader in SHM. Mardanshahi et al. [20] overviewed contemporary sensing approaches, such as EM and acoustic-based sensing, and explained how they can be compatible with the AI algorithms. The intersection is especially useful when monitoring concrete tunnel and bridges in real-time. As an instance, Mengwen et al. [21] used U-Net segmentation on electrical resistivity tomography data under contamination localization, demonstrating the efficiency of the semantic segmentation models in the subsurface mapping exercise. Patrizia et al. [22] assessed the use of electromagnetic methods in cultural heritage diagnostics, citing the issues regarding the technique, as well as opportunities in the future perspective of using EM imaging to deal with poorly-held or old structures. Their findings can be used in civil infrastructure, where non-invasive procedures are of essence. In case of the tunnel inspections, Poncetti et al. [23] examined the new normative practice and advanced methods of NDT to perform inspection of concrete-covered wall of a tunnel. Their work confirms on the applicability of AI-powered EM inspection in restricted or risky areas. Further elaboration of this was undertaken by Prakash et al. [24], who provided a narrative review of current practices in artificial intelligence used to monitor the health of concrete bridge structures, which confirmed that deep learning is increasingly becoming a dominant method of large-scale infrastructure inspection. Although the bulk of investigations lie in civil applications, advances in adjacent fields can contribute new ideas into inter-disciplinary practice. Qiu and Lau [25] devised a hybrid deep learning technique based on UAV images to evaluate the structural defects of trees and Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1142 https://internationalpubls.com proposed that the same techniques could be applied with tall or remote concrete buildings. Finally, Rakin et al. [26] developed the fault diagnosis model base on the YOLOv11 network that is tailored to particular infrastructure faults in the urban environment. The real-time potential of their system is rather similar to the objectives of this study. On the whole, the literature illustrates an increased convergence between the electromagnetic inspection technologies and deep learning models. This paper expounds on these premises as the study combines EM imaging with optimized neural networks to make crack inspection in concrete structures a real-time possibility. III. METHODS AND MATERIALS 1. Materials and Methods This segment represents dataset description, data cleaning methods, and four deep learning algorithms to detect cracks in real time on the basis of electromagnetic (EM) imaging data. To offer a reasonable and standard comparison, each algorithm was executed and tested against the same data set. The general methodology incorporates the acquisition of the data, preprocessing, training of the models, evaluating the results and analysing the results. 1.1 Data Collection and Preprocessing The data was produced by synthetic and real electromagnetic imaging scan of concrete slabs with artificially generated cracks of depth, length and orientation varied. Each scan possesses a 2D gray scale (plot) depiction on the dielectric normal part possession on the material [4]. The 5 000 images were gathered 3500 intended for training, 750 validation and 750 testing. The following steps were used in preprocessing the data: ● Normalization: Pixel values of the image were normalized within the range 0 and 1. ● Resizing: All the images had been made to take 128128 pixels. ● Labeling: Every single image was labeled as either Cracked or Non-Cracked. ● Augmentation: In order to make the same robust we used effects of horizontal flipping, rotation (45 degrees plus or minus), zoom-in. 1.2 Selected Deep Learning Algorithms To compare four deep learning algorithms, Convolutional Neural Network (CNN), Residual Neural Network (ResNet-50), U-Net, and MobileNetV2 have been selected. These models have been selected as being most appropriate to image classification and segmentation. 1.2.1 Convolutional Neural Network (CNN) A CNN is a deep learning network that has found particular efficiency at image-related tasks. It works by implementing convolutional filters to detect local features which include edges, lines and textures. In our paper, we have adopted a pre-configured CNN consisting of three layers, each Conv layer, ReLU, and max pooling. The network is terminated at two fully connected layers and softmax classifier [5]. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1143 https://internationalpubls.com CNNs are useful to binary classification tasks and have few parameters as opposed to fully connected networks that are slower to train. Nevertheless, they can be difficult to learn long- range dependencies or minute patterns unless they have more architecture or an ability to make residual connections [6]. “Initialize weights and biases For each training epoch: For each input image: Apply convolutional layer + ReLU Apply max pooling Flatten the output Apply dense layers Compute loss using cross-entropy Backpropagate errors and update weights Return trained model” 1.2.2 Residual Neural Network (ResNet-50) ResNet-50 is a 50-layer deep convolutional neural network, which was able to overcome the vanishing gradient problem with the help of a residual learning strategy. Rather than the whole output, it learns the discrepancy between the input and output (residual), which allows training extremely deep networks more easily. This architecture has identity short cut connections that by pass one or more layers [7]. ResNet-50 in our implementation had been pretrained on ImageNet and fine tuned on the EM imaging dataset. It performed well in capturing intricacy features concerning internal crack patterns and more precise compared to shallow networks. “Input image For each ResNet block: Apply convolutional layers Add shortcut connection Apply batch normalization and Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1144 https://internationalpubls.com ReLU Flatten and pass through fully connected layers Output classification result” 1.2.3 U-Net U-Net is a ConvNet, developed to perform semantic segmentation. It is composed by a contracting path (encoder) and an expansive path (decoder). This model especially applies to localization of cracks in an image as the skip connection acts by maintaining the spatial resolution [8]. U-Net used in this research work was trained to produce the map of crack regions pixel-wise. During structural health diagnostics, it assisted in imaging where the exact defects are which is essential in diagnosis. “Input image Encoder: Apply conv + ReLU + maxpool (downsampling) Decoder: Upsample Concatenate with corresponding encoder features (skip connections) Apply conv + ReLU Final layer: Apply 1×1 conv to produce segmentation map Return pixel-wise classification” 1.2.4 MobileNetV2 MobileNetV2 is an efficient architecture that is narrow and needs to perform mobile and real- time applications. It employs depthwise separable convolutions and inverted residual blocks to keep accuracy, but at a heavily reduced computational cost. This also qualifies it to be well deployed on-site and where hardware resources are scarce [9]. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1145 https://internationalpubls.com We applied the MobileNetV2 pretrained model on the ImageNet and fine-tuned to the binary classification. It detected cracks with close to real time accuracy and competitive accuracy. “Input image Apply depthwise separable convolution blocks Use inverted residuals with linear bottlenecks Global average pooling Fully connected layer for binary output Return predicted label” IV. RESULTS AND ANALYSIS This section provides the experimental setup, implementation details, evaluation processes, and performance results obtained from the proposed deep learning models for detecting cracks in concrete in real-time through electromagnetic (EM) imaging. The results presented below are compared across four different architectures—CNN, ResNet-50, U-Net, and MobileNetV2— and evaluated based on standard evaluation metrics [10]. It also benchmarks the proposed models to several select related studies to see the potential relative advancements. Figure 1: “Detection of Surface Cracks in Concrete Structures using Deep Learning” 4.1 Experimental Setup All experiments were conducted in Python 3.10 using TensorFlow and Keras frameworks. The experiments were run on a system with an NVIDIA RTX 3080 GPU, 32GB of RAM, and running Ubuntu OS. The same training environment was used for all four models to ensure comparisons remain uniform; that being said, we used EM images of the same scale—both in size and count—to provide consistent training inputs. In our dataset, we had 5,000 EM images Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1146 https://internationalpubls.com that we split into a training 70%, validation 15%, and testing 15% sets, as described in Section 3. Each of the models was allowed to be trained for 50 epochs, wherein we used the early stopping technique based on validation loss to thwart overfitting [11]. We used the Adam optimizer with a learning rate of 0.001 and binary cross-entropy as the loss function for the models. For segmentation (U-Net), we used Dice loss to make a more meaningful pixel-by-pixel evaluation process. 4.2 Model Evaluation Metrics We used the following evaluation metrics: ● Accuracy: The percentage of correct predictions. . ● Precision: The percentage of true positives in all predicted positives. ● Recall: The percentage of true positives in all actual positives. ● F1-Score: The harmonic mean of precision and recall. ● IoU (Intersection-over-Union): For segmentation, to check the overlap between predicted and ground truth masks,. ● Inference Time: The average time per prediction in milliseconds, there is significance in time for real-time application. 4.3 Model Performance Comparison The performance of each model on the test data-set is shown below: Table 1: Model Performance Metrics Model Accuracy (%) Precision (%) Recall (%) F1-Score (%) IoU (%) Inference Time (ms) CNN 91.4 90.1 89.7 89.9 77.3 18 ResNet-50 96.2 95.6 95.1 95.3 86.4 42 U-Net 95.7 94.9 95.3 95.1 89.2 65 MobileNetV2 94.1 93.2 92.7 92.9 84.6 12 Chas a brief look at the dimensions of these three models you can see in Table 1, ResNet-50 had the best overall classification performance - while U-Net had the best localization performance (IoU). Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1147 https://internationalpubls.com MobileNetV2 was a little less accurate, but it had a phenomenal inference time, making it the optimal model to implement in real time on an embedded system [12]. Figure 2: “Deep Learning-Based Automated Detection of Cracks in Historical Masonry Structures” 4.4 Confusion Matrix Analysis The following confusion matrices summarize the prediction accuracy of each model across the classes: Table 2: Confusion Matrix (Sample of 200 Test Images) Model True Positive False Positive True Negative False Negative CNN 88 7 92 13 ResNet-50 94 3 95 8 U-Net 95 4 93 8 MobileNetV2 92 6 91 11 ResNet-50 and U-Net exhibited favorable trade-offs in terms of precision and recall. CNN had slightly more false negatives, but MobileNetV2 still achieved some respectable trade-offs, despite being the lightest of all models [13]. 4.5 Visual Analysis of crack detection To visualize predictions made from a trained model, specifically a pixel-wise evaluation of true vs. predicted crack locations using U-Net, we constructed a table that presents IoU values for 5 sample test images. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1148 https://internationalpubls.com Table 3: IoU Evaluation on Segmentation Test Samples (U-Net) Image ID Ground Truth Mask Area (px) Predicted Mask Area (px) IoU (%) IMG_0101 1082 1120 90.5 IMG_0133 943 910 87.8 IMG_0210 1002 978 88.6 IMG_0245 1354 1401 89.9 IMG_0287 820 792 85.2 The U-Net model was able to track the true position of the cracks very closely, supporting its usefulness for detailed structural health diagnostics [14]. Figure 3: “Deep Learning for Concrete Crack Detection and Measurement” 4.6 Convergence of models To ensure that correct model convergence was achieved, training and validation accuracy and loss curves were tracked to be able to detect overfitting. Table 4: Final Training Epoch Performance (After 50 Epochs) Model Train Accuracy (%) Validation Accuracy (%) Train Loss Validation Loss CNN 92.1 90.3 0.19 0.24 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1149 https://internationalpubls.com ResNet-50 97.8 95.8 0.08 0.14 U-Net 96.3 94.6 0.10 0.16 MobileNetV 2 94.5 93.1 0.14 0.18 The findings confirm that ResNet-50 and U-Net generalize better than the other models with lesser degrees of validation loss which indicates there was less overfitting. The CNN showed signs of overfitting at the outset [27]. 4.7 Related Work Comparison We presented a comparison of our models to select recent work that has applied AI and specifically image data or electromagnetic techniques for crack detection in concrete. Table 5: Comparison with Related Work Study Method Used Accuracy (%) IoU (%) Real-Time Capability Li et al. (2021) SVM on EM data 83.6 — No Kaur et al. (2022) CNN on RGB 89.2 72.3 Moderate Park et al. (2023) Hybrid DL+EM 92.5 80.1 Partial Our Work (ResNet-50 + EM) ResNet-50 96.2 86.4 Limited Our Work (MobileNetV2 + EM) MobileNetV2 94.1 84.6 Yes (Fastest) Our Work (U-Net + EM) U-Net 95.7 89.2 Moderate In contrast to prior studies: ● Models were applicable here much better than in past EM-only classifiers (see SVM- based models) Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1150 https://internationalpubls.com ● ResNet-50 had over a 6 percent jump in accuracy compared to the hybrid methods in Park, et al. (2023). ● U-Net was able to provide better segmentation than image-only CNNs in Kaur et al. (2022) and increased the IoU by 16.9%. ● MobileNetV2 had the best performance-latency balance and presented a way to complete clack detection in real-time on site. Figure 4: “Automated Vision-Based Detection of Cracks on Concrete Surfaces Using a Deep Learning Technique” 4.8 Key Observations and Insights 1. Algorithmic Superiority: Deep CNNs including ResNet-50 enabled increased accuracy through depth and residual learning properties to accommodate slightly visible EM crack features. 2. Segmentation Capability: U-Net provided pixel level crack localization would be important for structural engineers and the U-Net with skip connections would maintain resolution [28]. 3. Real-Time Capability: MobileNetV2 was able to achieve results quite consistent with larger models, but at a lower cost with less latency, which would be suitable for embedded systems. 4. Generalizability: The inclusion of data augmentation, early stopping and dropout help minimize overfitting, which from the still-strong test set performance reflects good generalizability. 5. Electromagnetic Imaging Advantage: EM images have surface and internal defect qualities. When processing EM images through a deep learning protocol, detection rates performed better when comparing image alone. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1151 https://internationalpubls.com 4.9 Limitations ● In order to undertake EM imaging, specialised tools are often required, and may not always be accessible on-the-scene. ● The model was tested in noise-free EM environments, controlled and managed, but it is important to note that noise from the environment or inhomogeneity in the material may alter accuracy [29]. ● Model inference times mapped on GPU, and would vary, if inference was conducted on mobile or edge devices. 4.10 Future Improvements ● Studies have shown that performance from a model will increase if data from multiple modalities is used (EM + visual images). ● Model compression (pruning, quantization, etc...) can be employed to help lower latency without losing accuracy. ● The data can be made more robust and versatile by enhancing the size of the dataset further, by adding cracks from different variables, which would help with logistical perspectives, as well as, multiple physical variables [30]. V. CONCLUSION The study provides the detailed investigation of combining the deep learning algorithms with electromagnetic (EM) imaging into the real-time crack detection in concrete structures. The project fills a much needed niche in structural health monitoring in that EM imaging has the ability to scan below the surface due to the fact that the EM field travels much further into the ground than most other methods of surface scanning and the advanced neural networks have pattern recognition capabilities than individual nodes and groupings. By applying and comparing the accuracy of the classification methods and performance in segmentation of four deep learning models, i.e., CNN, ResNet-50, U-Net and MobileNetV2, we were able to go through the test to find difference in classification accuracies and capabilities of segmenting an image and using the platform in real-time. The best classification result was achieved by ResNet-50, which has outperformed the other tested models, and U-Net showed the best pixel- level crack localization. Although not as precise, MobileNetV2 was most efficient in terms of inference time and, therefore, was the best on-site/on-board implementation solution. The application of a well selected EM imaging set and strict pre-processing, model training and testing also provided the robustness and reliability of the results. Comparison with the current related studies proved the higher superiority of our method in terms of performance and practical feasibility of implementation. As opposed to conventional surface level inspection or visual inspection, our technique would capture the interior cracks within an infrastructure, providing a more thorough and preemptive infrastructure maintenance solution. Effective combination of EM imaging and deep learning offers innovative future in terms of non- invasive, high and wide scale accuracy measurement that can be used to observe critical infrastructures like bridges, tunnels, and marine infrastructure. To sum up, the study helps in Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1152 https://internationalpubls.com advancing the domain of intelligent infrastructure inspection, suggesting a crack detection system with excellent performance and real-time capabilities. Other tasks in the future can include expanding the datasets, application to edge devices, and multiple-modal data to increase the sensitivity of detection and efficiency of operation in various environments. REFERENCE [1] Abdollahi-Mamoudan Farima, Ibarra-Castanedo Clemente & Maldague, X.P., V. 2025, "Non-Destructive Testing and Evaluation of Hybrid and Advanced Structures: A Comprehensive Review of Methods, Applications, and Emerging Trends", Sensors, vol. 25, no. 12, pp. 3635. [2] Cavadia, P., Benjumea, J.M., Begambre, O., Osorio, E. & Mantilla, M.A. 2025, "An Open Database of the Internal and Surface Temperatures of a Reinforced-Concrete Slab- on-I-Beam Section", Data, vol. 10, no. 2, pp. 21. [3] Congde, L., Senguo, C., Wang, X., Guanglai, J., Wang, S. & Wenlong, C. 2025, "OEM- HWNet: A Prior Knowledge-Guided Network for Pavement Interlayer Distress Detection Based on Computer Vision Using GPR", Remote Sensing, vol. 17, no. 9, pp. 1554. [4] Demeke, A.A., Hendlmeier, A., Dunn, M., Arablouei, R., Lomov, S.V., Pietro, A.D. & Nikzad, M. 2024, "Detecting Multi-Scale Defects in Material Extrusion Additive Manufacturing of Fiber-Reinforced Thermoplastic Composites: A Review of Challenges and Advanced Non-Destructive Testing Techniques", Polymers, vol. 16, no. 21, pp. 2986. [5] Egodawela, S., Gostar, A.K., Samith Buddika, H. A. D., Harischandra, W.A.N.I., Dhammika, A.J. & Mahmoodian, M. 2025, "Metal loss defect detection and depth estimation using multi-spectral image analysis of cooling excited steel specimen with corrosion", Scientific Reports (Nature Publisher Group), vol. 15, no. 1, pp. 23894. [6] Eyad, A. 2025, "Nondestructive Testing of Externally Bonded FRP Concrete Structures: A Comprehensive Review", Polymers, vol. 17, no. 9, pp. 1284. [7] Gabriel de, S.M., Ferreira Guedes, J.V. & Edilson de, S.B. 2024, "UAV-Embedded Sensors and Deep Learning for Pathology Identification in Building Façades: A Review", Drones, vol. 8, no. 7, pp. 341. [8] Ghattas, A., Al-Sharawi, R., Zakaria, A. & Qaddoumi, N. 2025, "Detecting Defects in Materials Using Nondestructive Microwave Testing Techniques: A Comprehensive Review", Applied Sciences, vol. 15, no. 6, pp. 3274. [9] Golovastikov, N.V., Kazanskiy, N.L. & Khonina, S.N. 2025, "Optical Fiber-Based Structural Health Monitoring: Advancements, Applications, and Integration with Artificial Intelligence for Civil and Urban Infrastructure", Photonics, vol. 12, no. 6, pp. 615. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1153 https://internationalpubls.com [10] Guo, S., Yu, M., Xu, Z., Yue, G., Cai, W. & Tian, P. 2025, "Study on the Attribute Characteristics of Road Cracks Detected by Ground-Penetrating Radar", Sensors, vol. 25, no. 3, pp. 595. [11] Hu, D. 2025, "Development of an Open GPR Dataset for Enhanced Bridge Deck Inspection", Remote Sensing, vol. 17, no. 13, pp. 2210. [12] Huang, H., Dai, Z. & Tang, P. 2024, "Application of Microwave Imaging Combined with Electromagnetic & Ray Localization in Pipeline Deformation Detection", Journal of Physics: Conference Series, vol. 2834, no. 1, pp. 012149. [13] Lavadiya, D.N. & Dorafshan, S. 2025, "Deep learning models for analysis of non‐ destructive evaluation data to evaluate reinforced concrete bridge decks: A survey", Engineering Reports, vol. 7, no. 1. [14] Łaziński, P., Jasiński, M., Uściłowski, M., Piotrowski, D. & Ortyl, Ł. 2025, "GPR in Damage Identification of Concrete Elements—A Case Study of Diagnostics in a Prestressed Bridge", Remote Sensing, vol. 17, no. 1, pp. 35. [15] Lee, T., Kim, D., Cho, S. & Min, O.K. 2025, "Advancements in Surface Coatings and Inspection Technologies for Extending the Service Life of Concrete Structures in Marine Environments: A Critical Review", Buildings, vol. 15, no. 3, pp. 304. [16] Li, C., Wang, Y., Qibing, M. & Xiaorong, W. 2025, "Multi-Domain Feature Analysis and Application Research of GPR Aliased Signals", Sensors, vol. 25, no. 9, pp. 2741. [17] Liu, L., Shen, B., Huang, S., Liu, R., Liao, W., Wang, B. & Diao, S. 2025, "Binocular Video-Based Automatic Pixel-Level Crack Detection and Quantification Using Deep Convolutional Neural Networks for Concrete Structures", Buildings, vol. 15, no. 2, pp. 258. [18] Liu, N., Ge, Y., Bai, X., Zhang, Z., Shangguan, Y. & Li, Y. 2025, "Research on Damage Detection Methods for Concrete Beams Based on Ground Penetrating Radar and Convolutional Neural Networks", Applied Sciences, vol. 15, no. 4, pp. 1882. [19] Luo, X. & Matsumoto, T. 2025, "Tensile Strength Estimation of UHPFRC Based on Predicted Cracking Location Using Deep Learning", Materials, vol. 18, no. 10, pp. 2237. [20] Mardanshahi, A., Sreekumar, A., Yang, X., Barman, S.K. & Chronopoulos, D. 2025, "Sensing Techniques for Structural Health Monitoring: A State-of-the-Art Review on Performance Criteria and New-Generation Technologies", Sensors, vol. 25, no. 5, pp. 1424. [21] Mengwen, G., Yu, X. & Zhang, X. 2025, "Identification of NAPL Contamination Occurrence States in Low-Permeability Sites Using UNet Segmentation and Electrical Resistivity Tomography", Applied Sciences, vol. 15, no. 13, pp. 7109. [22] Patrizia, P., Rocco, C., Fabio, M. & Fabrizio, F. 2025, "Electromagnetic Techniques Applied to Cultural Heritage Diagnosis: State of the Art and Future Prospective: A Comprehensive Review", Applied Sciences, vol. 15, no. 12, pp. 6402. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 31 No. 8s (2024) 1154 https://internationalpubls.com [23] Poncetti, B.L., Ruiz, D.V., Assis Leandro, S.d., Machado, L.B., Silva Tiago, B.d., Akinlalu, A.A. & Futai, M.M. 2025, "Tunnel Inspection Review: Normative Practices and Non-Destructive Method Advancements for Tunnels with Concrete Cover", Applied Mechanics, vol. 6, no. 2, pp. 41. [24] Prakash, V., Debono, C.J., Musarat, M.A., Borg, R.P., Dylan, S., Ding, W. & Jiangpeng, S. 2025, "Structural Health Monitoring of Concrete Bridges Through Artificial Intelligence: A Narrative Review", Applied Sciences, vol. 15, no. 9, pp. 4855. [25] Qiu, Q. & Lau, D. 2024, "Assessment of Trees’ Structural Defects via Hybrid Deep Learning Methods Used in Unmanned Aerial Vehicle (UAV) Observations", Forests, vol. 15, no. 8, pp. 1374. [26] Rakin, R.Z., Mahmudur, R., Borsa, K.F., Farid, F.A., Shakila, R., Jia, U. & Karim, H.A. 2025, "Towards Safer Cities: AI-Powered Infrastructure Fault Detection Based on YOLOv11", Future Internet, vol. 17, no. 5, pp. 187. [27] Seckin, M., Demircioglu, P., Ahmet, C.S., Bogrekci, I. & Aksoy, S. 2025, "Artificial Intelligence for Non-Destructive Imaging in Composite Materials", Eng, vol. 6, no. 3, pp. 46. [28] Shayan, M., Ryan, R. & Crovella, P. 2025, "A Review of the Potential of Drone-Based Approaches for Integrated Building Envelope Assessment", Buildings, vol. 15, no. 13, pp. 2230. [29] Shrestha, P., Avci, O., Rifai, S., Abla, F., Seek, M., Barth, K. & Halabe, U. 2025, "A Review of Infrared Thermography Applications for Civil Infrastructure", Structural Durability & Health Monitoring, vol. 19, no. 2, pp. 193-231. [30] Sung-Pil, S., Sang-Yum, L. & Le Tri, H.M. 2025, "Feasibility of EfficientDet-D3 for Accurate and Efficient Void Detection in GPR Images", Infrastructures, vol. 10, no. 6, pp. 140.