A study utilizing convolutional neural networks (CNN) has been conducted to detect and classify invisible debonding-type defects in reinforced structures. Training data for these defects is collected from the finite element models of honeycomb sandwich panels and skin-stringer systems, commonly employed reinforcement structures in aerospace applications. The excitation frequency is determined based on the amplitude of the reflected wave from the defect, and the optimal sensor array is selected. The constructed two-dimensional training image, created by vertically stacking the measured responses in the time domain, exhibits high classification performance even with a shallow neural network. The neural network undergoes optimization through adjustment to the kernel parameters and initial learning rate. To assess the general performance of the training model, k-fold cross-validation is employed. The CNN-based non-destructive evaluation algorithm demonstrates high classification performance for debonding defects in honeycomb sandwich panels and skin-stringer structures. Moreover, the suggested algorithm is robust against noise, emphasizing its effectiveness in real-world applications.

Traffic accidents caused by black ice have become a serious public safety concern due to their high fatality rates and the limitations of conventional detection systems under low visibility. Millimeter-wave (mmWave) radar, capable of operating reliably in adverse weather and lighting conditions, offers a promising alternative for road surface monitoring. In this study, six representative road surface conditions—dry, wet, thin-ice, ice, snow, and sludge—were experimentally implemented on asphalt and concrete specimens using a temperature and humidity-controlled chamber. mmWave radar data were repeatedly collected to analyze the temporal variations in reflected signals. The acquired signals were transformed into range-based spectra using Range-Fast Fourier Transform (Range-FFT) and converted into statistical features and graphical representations. These features were used to train and evaluate classification models, including Extreme Gradient Boost (XGBoost), Light Gradient-Boosting Machine (LightGBM), Convolutional Neural Networks (CNN), and Vision Transformer (ViT). While machine learning models performed well under dry and wet conditions, their accuracy declined in hazardous states. Both CNN and ViT demonstrated superior performance across all conditions, with CNN showing consistent stability and ViT exhibiting competitive accuracy with enhanced global pattern-recognition capabilities. Comprehensive robustness evaluation under various noise and blur conditions revealed distinct characteristics of each model architecture. This study demonstrates the feasibility of mmWave radar for reliable road surface condition recognition and suggests potential for improvement through multimodal sensor fusion and time-series analysis.