Influence of acoustic emission sequence length on intelligent identification accuracy of 3-D loaded rock’s fracture stage

Abstract

Accurate prediction of impending disasters in underground projects is crucial and requires the identification of rock fracture stages. Currently, rock fractures are commonly analyzed using microseismic parameter statistics or individual waveform features for engineering disaster early warning systems. However, rock fracturing is a continuous process, and waveform sequences contain a wealth of information on fractures, which is often overlooked by existing research that generally neglects the information within continuous waveforms. In this study, we leverage acoustic emission (AE) data and employ a transfer learning approach with a convolutional neural network (CNN) to identify rock fracture stages under three-dimensional (3-D) stress paths induced by true triaxial compression tests. Failure experiments were performed on seven sandstone specimens under various 3-D stress paths. To fully utilize the characteristics of the crack rupture sequence, we introduce the concept of AE waveform sequence length. This concept integrates the discrete features of AE time–frequency images, thereby improving the CNN model’s performance. Utilizing waveforms of six different lengths (1, 2, 3, 5, 10, and 15) to train the neural networks, our findings reveal that a sequence length of 10 enables the CNN to effectively identify rock fracture stages under 3-D stresses with an accuracy rate of up to 90.4%. This demonstrates that appropriately increasing the sequence length to process the discrete features of AE waveforms structurally is a viable strategy to enhance CNN identification accuracy. Our results underscore that rock fracturing is a sequential process with significant inter-sequence correlations, which critically influence the CNN model’s ability to accurately identify rock fracture stages. These insights offer valuable theoretical contributions to the automatic monitoring of rock fracture stages in deep engineering projects.