关键词: 声发射, 裂纹分类, 深度学习, 实时识别

Abstract: Accurate identification of crack types during rock fracturing is of great significance for revealing rock mass failure mechanisms and preventing engineering disasters. Shear tests and Brazilian splitting acoustic emission (AE) tests were conducted on sandstone. The Pearson correlation coefficient and random forest methods were employed to optimize AE parameters. The SMOTE (synthetic minority over-sampling technique) method was used to balance the tensile-shear crack dataset, thereby constructing an integrated AE dataset that incorporates tensile-shear crack characteristics. By considering various machine learning model performance evaluation metrics, the convolutional neural network (CNN) was determined as the foundational architecture for the identification model. Subsequently, three deep learning models were constructed: CNN-LSTM (long short term memory)-multi head attention, CNN-GRU (gated recurrent unit), and BO (Bayesian optimization)-CNN-LSTM. It was found that the BO-CNN-LSTM model is most suitable for the real-time identification of rock crack types. The eight optimal identification parameters are average frequency, duration, initial frequency, rise time, peak frequency, centroid frequency, amplitude, and root mean square (RMS) voltage, achieving a comprehensive accuracy of 97.89%. Based on sandstone uniaxial compression tests, the generalization performance of the three deep learning models was compared. The difference in crack classification results between the optimal real-time identification model BO-CNN-LSTM and the conventional Gaussian mixture model-support vector machine (GMM-SVM) clustering model was less than 5%. The newly developed real-time crack type identification model possesses high recognition accuracy and generalization capability, providing a reliable basis for the real-time identification of rock crack types.

Key words: acoustic emission, crack classification, deep learning, real-time identification