UPVnet: A Neural Network for First-Arrival Picking in Ultrasonic Pulse Velocity Testing on Rock Samples

Abstract

Ultrasonic pulse velocity (UPV) testing is widely employed as a nondestructive technique in rock engineering to determine the essential physical properties of rocks. This study introduces UPVnet, an innovative deep learning-based neural network specifically developed to enhance the accuracy and reliability of first-arrival picking in UPV testing. By advancing multidimensional feature extraction methods and optimizing the network architecture, UPVnet achieves significantly higher picking accuracy and enhanced generalization capability compared to traditional methods. Additionally, the study validates the feasibility of UPV testing on small-sized rock samples, demonstrating that P-wave and S-wave velocities measured on 20-mm-thick samples exhibit good consistency with those from standard plug samples, despite challenges such as size effects and velocity dispersion.