Identifying different classes of geoacoustic events using machine learning

Abstract

Microseismic monitoring is widely used to detect instability hazards. Specifically, the quick and effective identification of microseismic events, such as noise, explosions, and drilling activities, is critical for determining mine stability and safety. This study employed features including the mean and standard deviation of frequencies extracted by short-time Fourier transform (STFT), short-term energy, total energy, and waveform length as multivariate parameters for event classification. These features were then standardized and integrated, and techniques such as KMeansSMOTE and OneSidedSelection were used to balance the dataset distribution through oversampling and undersampling. K-fold cross-validation combined with an advanced retention network deep neural network architecture was used to generate a comprehensive geoacoustic event classification model (GSEC) to classify different types of geoacoustic events, such as microseisms, blasts, drilling, and noise. In-depth comparisons were then performed using common deep-learning models, such as a convolutional neural network, long short-term memory network, residual network, dense network, and transformer. The proposed GSEC model outperformed the baseline models in terms of key performance metrics such as accuracy, precision, recall, and the F1 score. Thus, the developed GSEC model represents a new and effective tool for improving mine safety management.