Aiming at the problem of coal–rock dynamic disasters occurring during roadway excavation in deep mines, this study employs microseismic monitoring technology to conduct real-time monitoring of microseismic events at the excavation face of Wuyang Coal Mine. Signal features are extracted using methods such as Short-Time Fourier Transform (STFT), Stockwell Transform (S-transform), and Wavelet Packet Transform (WPT). Key parameters are classified and predicted, and a microseismic event identification model is developed. The research findings indicate that microseismic events are primarily distributed above the excavation face and along the sides of the roadway, with the event frequency closely correlated with the excavation rate. Coal–rock fracture events in microseismic signals exhibit peak-like waveforms, large amplitude, and high energy. Mechanical disturbance signals exhibit periodic oscillations, their energy attenuation is rapid, and the dominant frequency of mechanical disturbance signals shows a significant lag compared to that of coal–rock fracture signals. The wavelet packet band energies of coal–rock fracture events and mechanical disturbance events comply with an overall decreasing trend. A significant residual energy signature exists in the 16th band of coal–rock fracture microseismic events. Statistical calculations revealed that 72.99% of the coal–rock fracture events in Wuyang Coal Mine were low-energy microseismic events, 15.64% were medium-energy, and 11.37% were high-energy. Using the amplitude, main frequency, and energy characteristic indexes, a coal–rock fracture event classification and identification prediction model is established, which yields prediction accuracies of 78.71%, 81.16%, and 93.48% for Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Long Short-Term Memory Network (LSTM), respectively. Comprehensive results of accuracy, precision, recall, F1-score, ROC curve, and AUC demonstrate that the LSTM model achieves optimal performance. Finally, by integrating the intrinsic correlation among multi-source monitoring indexes, such as amplitude, frequency, and energy, a comprehensive analysis method of correlated long and short-term memory network (R-LSTM) is constructed, which significantly improves the classification discrimination accuracy and prediction ability of coal-rock fracture events. The results of the study can provide important theoretical basis and technical support for the identification of microseismic events and disaster early warning in coal mines.Highlights<UnorderedList Mark="Bullet"> <ItemContent> The real-time, continuous and uninterrupted microseismic monitoring of coal mine operation is realized. </ItemContent> <ItemContent> Feature extraction of different microseismic signals. </ItemContent> <ItemContent> A machine learning algorithm is used to classify and identify microseismic events. </ItemContent> <ItemContent> The R-LSTM microseismic classification and recognition model is constructed. </ItemContent> </UnorderedList>

Identification and Classification Prediction of Microseismic Signals from Coal–Rock Fracture in Deep Coal Mines

期刊论文
发表日期：

2026年1月6日

查看全文

Tao Ma,
Yanni Zhang,
Dongji Lei,
Yajuan Wang

摘要

Aiming at the problem of coal–rock dynamic disasters occurring during roadway excavation in deep mines, this study employs microseismic monitoring technology to conduct real-time monitoring of microseismic events at the excavation face of Wuyang Coal Mine. Signal features are extracted using methods such as Short-Time Fourier Transform (STFT), Stockwell Transform (S-transform), and Wavelet Packet Transform (WPT). Key parameters are classified and predicted, and a microseismic event identification model is developed. The research findings indicate that microseismic events are primarily distributed above the excavation face and along the sides of the roadway, with the event frequency closely correlated with the excavation rate. Coal–rock fracture events in microseismic signals exhibit peak-like waveforms, large amplitude, and high energy. Mechanical disturbance signals exhibit periodic oscillations, their energy attenuation is rapid, and the dominant frequency of mechanical disturbance signals shows a significant lag compared to that of coal–rock fracture signals. The wavelet packet band energies of coal–rock fracture events and mechanical disturbance events comply with an overall decreasing trend. A significant residual energy signature exists in the 16th band of coal–rock fracture microseismic events. Statistical calculations revealed that 72.99% of the coal–rock fracture events in Wuyang Coal Mine were low-energy microseismic events, 15.64% were medium-energy, and 11.37% were high-energy. Using the amplitude, main frequency, and energy characteristic indexes, a coal–rock fracture event classification and identification prediction model is established, which yields prediction accuracies of 78.71%, 81.16%, and 93.48% for Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Long Short-Term Memory Network (LSTM), respectively. Comprehensive results of accuracy, precision, recall, F1-score, ROC curve, and AUC demonstrate that the LSTM model achieves optimal performance. Finally, by integrating the intrinsic correlation among multi-source monitoring indexes, such as amplitude, frequency, and energy, a comprehensive analysis method of correlated long and short-term memory network (R-LSTM) is constructed, which significantly improves the classification discrimination accuracy and prediction ability of coal-rock fracture events. The results of the study can provide important theoretical basis and technical support for the identification of microseismic events and disaster early warning in coal mines.

Highlights

The real-time, continuous and uninterrupted microseismic monitoring of coal mine operation is realized.

Feature extraction of different microseismic signals.

A machine learning algorithm is used to classify and identify microseismic events.

The R-LSTM microseismic classification and recognition model is constructed.