Identification of failure behaviors of underground structures under dynamic loading using machine learning

Abstract

Understanding the dynamic responses of hard rocks is crucial during deep mining and tunneling activities and when constructing nuclear waste repositories. However, the response of deep massive rocks with openings of different shapes and orientations to dynamic loading is not well understood. Therefore, this study investigates the dynamic responses of hard rocks of deep underground excavation activities. Split Hopkins Pressure Bar (SHPB) tests on granite with holes of different shapes (rectangle, circle, vertical ellipse (elliptical short (ES) axis parallel to the impact load direction), and horizontal ellipse (elliptical long (EL) axis parallel to the impact load direction)) were carried out. The influence of hole shape and location on the dynamic responses was analyzed to reveal the rocks' dynamic strengths and cracking characteristics. We used the ResNet18 (convolutional neural network-based) network to recognize crack types using high-speed photographs. Moreover, a prediction model for the stress-strain response of rocks with different openings was established using Deep Neural Network (DNN). The results show that the dynamic strengths of the granite with EL and ES holes are the highest and lowest, respectively. The strength-weakening coefficient decreases first and then increases with an increase of thickness-span ratio (h/L). The weakening of the granite with ES holes is the most obvious. The ResNet18 network can improve the analyzing efficiency of the cracking mechanism, and the trained model's recognition accuracy reaches 99%. Finally, the dynamic stress-strain prediction model can predict the complete stress-strain curve well, with an accuracy above 85%.