Identifying the Types of Loading Mode for Rock Fracture via Convolutional Neural Networks

For decades, monitoring and identification of the dynamic stress states of rock masses in reservoirs have been challenging tasks owing to the lack of effective observation and analysis methods. In this study, we used deep learning methods to identify whether acoustic emission (AE) signals of rock fractures are created by different loading conditions. We performed Brazilian split and uniaxial compressive experiments on six different engineering materials, and we obtained the AE waveforms. To take advantage of the powerful image processing capabilities of convolutional neural networks (CNNs), we transformed the AE waveforms into time-frequency images. We used five types of CNN to identify the time-frequency images of the AE signals created in the Brazilian split and uniaxial compressive experiments. As a result, we found that Xception model had the highest recognition accuracy. We analyzed the basis of the CNN models to recognize the AE signals using local interpretable model-agnostic explanations and found that the Xception model mainly used the patterns of the low-energy region of the time-frequency images to determine the loading modes of the rock fractures. Furthermore, the high-energy time-frequency region had little effect on recognition. The findings of this work can aid in automatically monitoring the dynamic stress states of fracture areas in reservoir formations and ensure the safety of oil and gas resource exploitation.