Man Ke, Wu Liwen, Liu Xiaoli, Song Zhifei, Li Kena
Rockburst is a common dynamic geological hazard in underground engineering. In order to improve the prediction effect of the rockburst grade and the generalization of the prediction model, the Gate Recurrent Unit (GRU) neural network model with excellent memory and strong temporal timing was proposed to predict the rockburst grade. Firstly, based on the characteristics of rockburst genesis and grey correlation analysis, reasonable rockburst prediction indicators were screened, and the data of 137 sets of rockburst samples with existing research results were used as input and output data sets. Then, the optimal hyperparameter and the best prediction effect were through hyperparameter tuning of the model, and the rockburst grade prediction effect of the GRU model was compared with the rockburst grade prediction effect of the RF model, SVM model, CNN model, LSTM model, BP model, Russenes criterion, Wang Yuanhan criterion, Guan Baoshu criterion, brittleness coefficient criterion and elastic energy index criterion to verify the effectiveness of the GRU model, and verify the generalization of the GRU model according to the random sampling analysis of different models. Finally, two engineering examples were used to verify the practicality of the GRU model. The results show that the reasonable prediction indicators of rockburst are tangential stress of surrounding rock σθ, stress coefficient σθ/σc, the ratio of uniaxial compressive strength to uniaxial tensile strength σc/σt and elastic energy index Wet. According to the result analysis of different prediction methods, the effect of the GRU model in predicting rockburst grade is significantly higher than the above other prediction methods. Moreover, according to the random sampling analysis results of different models, the generalization of the GRU model in predicting rockburst grade is significantly strong. The prediction results of rockburst grade in the two engineering examples are in line with the actual rockburst situation, and the GRU model proposed in this paper is practical.
