Research on rock fracture pattern recognition based on convolutional neural networks and digital image correlation

Rock failure modes are a key factor for assessing the stability of geological structures and predicting crack propagation, and are critical to ensuring the safety of rock engineering. To address the identification of rock fracture modes, an automatic recognition method for rock mass fracture modes integrating Digital Image Correlation (DIC) with deep learning is proposed. Original image data were collected through laboratory uniaxial compression, biaxial compression and Brazilian splitting tests, and deformation field cloud map data of fractured rock masses were obtained via the DIC method. Fracture segmentation datasets were constructed using the original images of fractured rock masses, and a large-scale image segmentation method for fracture identification trained on the U-Net model was developed, which can accurately capture the positional and morphological characteristics of fractures. For the fracture segmentation task, 1,350 image pairs were divided into a training set (810 pairs), a validation set (270 pairs) and a test set (270 pairs) at a ratio of 6:2:2. On this basis, a dataset of local deformation field maps was established for fracture mode classification. Five convolutional neural network (CNN) models, namely GoogLeNet, Xception, ResNet50, InceptionV3 and MobileNetV2, were selected for classification experiments and their performance was evaluated. For the fracture mode classification task, 419 DIC displacement vector cloud maps were split into a training set (294 maps) and a test set (125 maps) at a ratio of 7:3. All five CNN models exhibited favorable performance in image-based recognition of rock fracture modes, among which the ResNet50 model achieved the highest classification accuracy, with recognition accuracies of 97.89%, 98.33% and 99.36% for shear fractures, shear-tension mixed fractures and tensile fractures, respectively, and its overall performance outperformed the other networks. The research findings provide an automatic and efficient method for identifying and analyzing rock failure modes in geotechnical mechanics experiments, and are of great significance for predicting fracture development and evaluating engineering stability.