Enhanced convolutional neural network-based prediction and source mechanism interpretation of high-energy microseismic events in coal mines

Rockburst is a common geological hazard in mining operations, and the occurrence of high-energy microseismic events often precedes rockbursts. As coal mining depths increase, existing prediction methods for these hazards are often inefficient and inaccurate. To address these issues, this study proposes an enhanced rockburst prediction method that integrates microseismic sensitive feature indicators with an improved model combining convolutional neural network and gated recurrent unit. First, the b-value is calculated using the Gutenberg–Richter theory, and trends in daily energy, frequency, and b-value leading up to high-energy events are analyzed. Fourier transform is then used to extract the dominant frequency and amplitude, revealing their correlation with the daily maximum energy. Based on these features, the model is applied to predict the daily maximum microseismic energy, with the convolutional neural network responsible for feature extraction and the gated recurrent unit for energy prediction. The model has been validated with actual mine data and applied in field conditions, showing superior predictive performance for high-energy events compared to traditional neural network models, achieving an R2 of 0.916 05, a mean absolute percentage error of 0.985 38, a mean absolute error of 0.114 57 × 104, and a root mean square error of 0.1612 × 104. By integrating the model's predictions with source mechanism analysis, the method enables targeted preventive measures in high-risk areas, such as implementing preemptive drilling and blasting pressure relief, effectively reducing the risk of rockburst. These results suggest that the model demonstrates strong theoretical predictive capability and practical reliability in real mining environments.