U-Net learning for the automatic identification of the sandstone crack tip position to determine mixed-mode stress intensity factors utilizing digital image correlation method

Abstract

Tracking the crack tip position in crack growth is a key and difficult problem in calculating mixed-mode I–II dynamic stress intensity factors based on displacement field obtained by digital image correlation method (DIC). U-Net in deep learning is used to perform pixel-level semantic segmentation on the experimental images of mixed-mode I–II cracks propagation under the three-point bending of brittle sandstone to identify crack features and locate the crack tip position. On this basis, the digital image correlation displacement data are substituted into the Williams displacement field equation to calculate the dynamic stress intensity factors at different loading stages. The results show that the U-Net network model can achieve image segmentation and process experimental pictures in batches, and accurately detects cracks in images with speckles which has been demonstrated potential in identifying crack propagation paths and tip positions. When the position of the crack tip is known, the solution for the stress intensity factor converts the non-linear equation system to a linear one and solves it with the linear least squares method. This significantly improves the solution accuracy and speed of the mixed-mode I–II dynamic stress intensity factors. Simultaneously, according to the position of the crack tip, the crack propagation path during the loading process can be drawn, which provides a reference for research on the evolution in crack growth.