A Machine Learning Prediction Model for Rockburst Based on Oversampling Algorithm and Bayesian-XGBoost

Rockburst is a key source of risk in engineering construction of deep-buried tunnels, which is induced by high in-situ stress and strong dynamic disturbance. Due to the highly complex relation between rockburst and the impact factors, traditional mechanism-based prediction methods have some limitations. Extreme gradient boosting (XGBoost) is herein introduced to predict the rockburst problem, where six hyperparameters are optimized by Bayesian optimization algorithms. This study collected 384 data sets based on real rockburst cases. Rockburst prediction can be divided into four categories: no rockburst, light rockburst, moderate rockburst and high rockburst. The occurrence probability of these four categories is different in actual engineering, which leads to imbalanced samples of the four rockburst categories. The synthetic minority oversampling technique (SMOTE) algorithm is utilized to process the collected data sets as the sample imbalance can affect the prediction accuracy of XGBoost model. The SMOTE-Bayesian-XGBoost model proposed in this study does not rely on the internal mechanism of rockburst. It provides a simple but effective model for rockburst prediction, which is of practical significance to reduce the risk during deep-buried tunnel construction.

Keywords: Rockburst prediction, extreme gradient boosting, Bayesian optimization algorithms, synthetic minority oversampling technique.