Ultrasonic Imaging of Crack Detection in Shafts Using Convolutional Neural Networks and Grid-Wise Classification

Abstract

We present a data-driven convolutional neural network (CNN) framework for detecting internal cracks in a two-dimensional shaft cross-section using transient elastodynamic wave responses. The model is trained using a grid-wise classification approach, where each element in the domain is labeled as damaged or undamaged. Training data are generated through wave propagation simulations in ANSYS Mechanical, where elliptical cracks of varying sizes, orientations, and positions are randomly introduced. Measured wave signals at the boundary are used as input features, and corresponding damage labels for each grid element of unstructured background mesh are used as output. The CNN learns to predict the probability of damage for each grid element based on wave responses from multiple sensors. The trained model effectively reconstructs internal cracks without prior knowledge of their characteristics. Numerical results on out-of-distribution datasets show that the model can reliably detect cracks in various configurations, indicating its potential for ultrasonic structural health monitoring of complex engineering structures.