Real-time identification of acoustic emission signals of rock tension-shear fracture based on machine learning and study on precursory characteristics

The identification of rock tensile-shear cracks and the study of precursory characteristics related to micro-crack types are of significant importance for monitoring and early warning of rock fracture states. The direct shear acoustic emission monitoring test of granite was carried out, and the PSO-GMM-SVM model is constructed to classify RA and AF data. The feature importance of acoustic emission parameters is analyzed using the Maximum Information Coefficient (MIC) and Random Forest (RF). Three models are then built: CNN-LSTM-Multi Head Attention, CNN-SVM, DT-AdaBoost. By comparing the overall performance of three different crack type identification models, the best model and optimal parameters are selected. Finally, based on the crack classification results, a parameter representing the dynamic proportion of the source types, namely the TSR-value, was constructed and used to study the precursor characteristics of rock fracture . The results show that the DT-AdaBoost model is most suitable for the real-time identification of rock tensile-shear cracks. The seven parameters determined to be most suitable for rock tensile and shear crack identification are: Average Frequency, Duration Time, Initial Frequency, Rise Time, AE Count, Amplitude, and Peak Count. The acoustic emission b-value and the TSR-value followed the same trend, both showing a general downward trend. When the TSR-value continues to decrease and the magnitude of the decrease reaches 61.37% of the historical maximum value, it is considered a precursor characteristic of rock instability and fracture. The research results provide a basis for the classification method of tension-shear cracks based on machine learning and acoustic emission technology, offering a new index for the early warning of rock instability and fracturing.