Intelligent rockburst prediction model with sample category balance using feedforward neural network and Bayesian optimization

The rockburst prediction becomes more and more challenging due to the development of deep underground projects and constructions. Increasing numbers of intelligent algorithms are used to predict and prevent rockburst. This paper investigated the drawbacks of neural networks in rockburst prediction, and aimed at these shortcomings, Bayesian optimization and the synthetic minority oversampling technique + Tomek Link (SMOTETomek) were applied to efficiently develop the feedforward neural network (FNN) model for rockburst prediction. In this regard, 314 real rockburst cases were collected to establish a database for modeling. The database was divided into a training set (80%) and a test set (20%). The maximum tangential stress, uniaxial compressive strength, tensile strength, stress ratio, brittleness ratio, and elastic strain energy were selected as input parameters. Bayesian optimization was implemented to find the optimal hyperparameters in FNN. To eliminate the effects of imbalanced category, SMOTETomek was adopted to process the training set to obtain a balanced training set. The FNN developed by the balanced training set received 90.48% accuracy in the test set, and the accuracy improved 12.7% compared to the imbalanced training set. For interpreting the FNN model, the permutation importance algorithm was introduced to analyze the relative importance of input variables. The elastic strain energy was the most essential variable, and some measures were proposed to prevent rockburst. To validate the practicability, the FNN developed by the balanced training set was utilized to predict rockburst in Sanshandao Gold Mine, China, and it had outstanding performance (accuracy 100%).