Detection of subsurface cracks in marble: application of artificial intelligence based on ultrasonic technique

In this study, a novel artificial intelligence-based measurement system using ultrasonic data was developed to detect the existence and orientation of artificially induced cracks in marble without axial loading. Ultrasonic sensors, acting as detectors, were moved across the marble surface by a PLC-controlled mechanical system. During position-controlled scanning, 268 data points—including average longitudinal wave velocity, Young’s modulus, Poisson’s ratio, and velocity variations—were recorded with a V-Meter Mark IV™ ultrasonic testing device and saved in CSV format. Feature selection was applied to identify crack-related indicators. Since crack data were underrepresented, the Smote method was used for balancing. As no single dominant feature emerged, all data except 256 measurements containing longitudinal wave variations were retained for classification. Decision Tree, Random Forest, and Logistic Regression models were first tested, but due to limited success, deep learning methods were applied. Multiple neural network structures were evaluated, achieving a maximum accuracy of 0.825. The estimated and actual discontinuity positions were then visualised, showing that the orientation of cracks within marble blocks relative to the scanning surface can be determined, thereby reducing waste in the cutting process.