Real-Time Identification of Rock Failure Stages Using Deep Learning: A Case Study of Acoustic Emission Analysis in Rock Engineering

The accurate and timely determination of rock failure processes is crucial for various rock engineering applications, especially for preventing dynamic disasters such as rock bursts and roof failures. The primary aim of this study was to determine the current rock failure stage using a single acoustic emission (AE) event signal recorded during the failure process. To achieve this, we proposed a deep learning model that employs advanced convolutional modules and a soft-threshold technique to extract the full waveform features of AE events from four different stages of rock failure in a laboratory uniaxial compressive strength (UCS) test. Once fully trained, our model can instantaneously determine the current rock failure stage from a raw waveform of a single recorded AE event, achieving a classification accuracy of 84%. Subsequently, the trained model was applied to on-site microseismic data analysis at a coal mine working face. Compared to traditional methods of microseismic data analysis that consider large-energy events, our model can identify the rock failure stage at the time of a specific microseismic event. Furthermore, by analyzing microseismic events triggered by post-peak rock fracturing, we identified potential hazard areas for rock bursts in the working face, and the results closely matched the site's burst prevention logs. This study successfully developed a real-time method for determining rock failure stages using deep learning, which can be effectively applied to microseismic data analysis in engineering sites to provide more precise early warnings of rock dynamic disasters.