Time Series Prediction of Microseismic Multi-parameter Related to Rockburst Based on Deep Learning

Time series prediction refers to the learning of existing observed data of a parameter and predicting its future evolution. Based on the application of machine/deep learning methods in the field of engineering geology, it is desirable to predict the time series evolution of microseismic parameters in the process of rockburst development. Our study explores key microseismic indices that help describe the development process of rockbursts based on abundant rockburst data obtained from deep underground engineering construction. The integrated process of dynamic moving-window method and improved convolutional neural network (CNN) realizes the evolution prediction of multiple microseismic parameters or their different combinations, and the modified model structures of a univariate, multivariate input and a single-step, multi-step output are established. Various models of the multiple microseismic parameters for the CNN-based time series prediction are innovated, including a univariate prediction model, a multiple parallel series model, a multiple input series model, and a multivariate multi-step prediction model. Model training, testing, and interpretation of the rockburst risk and comparative analyses of the different models are performed for the complete process of multiple rockbursts. The results show that the proposed models can well predict the evolution trends in the various key characteristics during rockbursts. The predicted trend of multiple microseismic parameters provides time labels for rockburst prediction and risk judgement, which is conducive to rockburst early warning. This study provides a new research idea for the prediction and early warning of rockbursts in the field of deep underground and mining engineering.