Rockburst prediction based on data preprocessing and hyperband‐RNN‐DNN

Abstract A data preprocessing workflow is proposed to address key challenges in rockburst data analysis, including dimensionality differences among various sample features, variations in data values within the same feature, missing data, poor data consistency, and sample class imbalance. The workflow is divided into four steps. Each step introduces multiple algorithms, which are sequentially combined according to the order of the four steps. Then, these coupled algorithms are utilized to preprocess the rockburst data set. The rockburst data set contains 459 samples, and the maximum tangential stress ( MTS ), the uniaxial compressive strength ( UCS ), the uniaxial tensile strength ( UTS ), the elastic strain energy index ( WET ), the rock stress factor ( SCF ), and the rock brittleness coefficient ( B ) are selected as the feature parameters. Subsequently, three architectures, Deep Neural Network (DNN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN), are used to evaluate the data sets processed by different coupled algorithms. The hyperband algorithm is introduced to optimize the hyperparameters of the RNN model, and the prediction accuracy of different architectures is compared between the RNN model with dense layers and without dense layers. Finally, a rockburst prediction model based on data preprocessing and the Hyperband‐DNN model is developed. The prediction results show that data preprocessing can significantly improve the model prediction accuracy; the model architecture with the highest prediction accuracy can be found quickly using the hyperband algorithm; and adding the dense layer can improve the stability and prediction accuracy of the model.