Time series prediction of microseismic energy level based on feature extraction of one-dimensional convolutional neural network

With the gradual transition of coal mining to deep mining, the number and intensity of rock burst events in the deep mining process are gradually increasing. Thus, it is of great significance to study the change of rock burst precursor signal for the prediction of rock burst. Microseismic signal monitoring plays an important role in rock burst prediction. The microseismic energy level changes with time, a good corresponding relationship exists between the high-energy microseismic events and rock burst. To advance the time node of rock burst prediction and provide more time guarantee for rock burst prevention and control, a time series prediction model of mine microseismic energy based on the one-dimensional convolutional neural network (CNN) was established to predict the temporal variation of mine microseismic energy. Through model training, the energy level of the previous 10 microseismic events can be used as input to predict the energy level of the next microseismic event. Due to the imbalance of the microseismic sample data, the microseismic events of the 10⁶-energy level were all judged as 10⁵-energy level microseismic events in the model test. To improve the prediction accuracy of the model for the 10⁶-energy level microseismic events, a hybrid sampling method was used to train the improved model. Using the microseismic energy level data of 250202 working face in Yanbei coal mine, the overall test accuracy of the improved model reaches 98.4% and the test accuracy of the 10⁶-energy level microseismic events increased to 99%. The improved prediction model of the microseismic energy level time series based on the one-dimensional convolution neural network was applied to 250202 working face of Yanbei coal mine to predict the microseismic energy level time series. The overall prediction accuracy of the model is 93.5%, and the prediction accuracy of high-energy microseismic events is close to 100%.