Prediction of rockburst intensity grade based on convolutional neural network

Rockburst is one of the urgent problems to be addressed in the process of deep resource extraction. In order to predict the rockburst disasters safely and efficiently, a rockburst intensity grade prediction model (MICE-CNN) based on the Multiple Imputation by Chained Equations (MICE) and Convolutional Neural Network (CNN) was proposed. Specifically, a predictive indicator system was established based on the main influencing factors and the acquisition conditions of rockburst. A total of 120 sets of raw data from rockburst cases were collected, with the outliers processed by pauta criterion. Then, the missing data were interpolated with the four interpolation models of RF, BLR, ET and KNN, which were selected using MICE. Besides, data interpolation was performed with the optimal model selected according to E_RMS, in combination with the two traditional interpolation methods (Mean and Median), resulting in a complete data set. In addition, the data were flattened into a 6×1×1 one-dimensional image data as the input layer, and the sizes of the convolutional kernel and pooling kernel were calculated to be 3×1 based on the size of the input layer. Moreover, zero-padding was applied for the feature edge processing. Batch normalization layers were added to improve the model stability and convergence speed. Thus, ReLU activation function and SGDM optimizer function were selected. Further, the CNN prediction model was trained, with accuracy rates of 100.00% for the training set and 91.67% for the validation set. Meanwhile, the RBF, SVM and PNN models were established for the comparison and verification of their test set data with that of the CNN model. Generally, the CNN model shows higher accuracy (91.67%) than the other models. By comparing the confusion matrix of the PNN model with the CNN model, it is found that the CNN model tends to overestimate the degree of rockburst compared to the actual results, indicating better safety after misjudgment. This demonstrates the feasibility of the MICE-CNN prediction model of rockburst intensity grade.