In order to realize efficient and accurate rockburst grade prediction, and prevent underground engineering disasters, this paper proposes a prediction model based on black-winged kite optimization algorithm-convolutional neural network-support vector machine (BKA-CNN-SVM). Firstly, the prediction index system was established according to six influence factors of rockburst, and 284 groups of rockburst cases at home and abroad were collected to establish a rockburst database. Secondly, Laida criterion and 1.5 times quartile difference were introduced to remove and replace the outliers in the data. The kernel principal component analysis (KPCA) was used to reduce the dimension of the data and extract the features. The extracted features were used as the model inputs. Finally, the confusion matrix was used to evaluate the model performance in terms of accuracy, precision, recall, and F₁ value. BKA-CNN-SVM model was compared with convolutional neural network (CNN) model, extreme learning machine (ELM) model, and convolutional neural network and support vector machine (CNN-SVM) integrated model. The results showed that the accuracy, precision, F₁ value, and recall of BKA-CNN-SVM model are 95.35%, 0.89, 0.92, and 0.94, respectively, which are significantly better than the other models in terms of prediction accuracy and generalization degree. In order to verify the feasibility of the BKA-CNN-SVM model, it was used to prediction the rockburst grade of the Jinping secondary hydro-power station. The prediction results have high consistency with the actual field conditions. This research can provides a new method for rockburst grade prediction.

To solve the problems of outlier samples, imbalanced samples, and local optimal of sparrow search algorithm in machine learning rockburst prediction, this paper established a rockburst prediction model from two perspectives of data preprocessing and algorithm improvement. First, based on lithology conditions and stress conditions, selected the maximum tangential stress, compressive strength, tensile strength and elastic energy index of surrounding rock as the characteristic indexes, and used three kinds of machine learning algorithms combined with 5-fold cross-validation method to construct the prediction model. In the data pre-processing stage, collected 174 groups of domestic and international rock burst cases to establish a database; for outlier samples, introduced the local outlier factor (LOF) algorithm to detect and eliminate outlier samples step by step according to the rock burst class; for sample imbalance, the adaptive synthetic sampling method (ADASYN) was introduced to increase the number of minority class samples. Three hybrid strategies were employed to improve sparrow search algorithm (ISSA) was used to optimize the parameters of three machine learning algorithms, namely limit gradient lift tree (XGBoost), random forest (RF) and multi-layer perceptron (MLP). Multiple evaluation indexes such as accuracy rate and precision rate were analyzed and discussed to verify the effectiveness of the model. The results show that the accuracy of the newly constructed optimal model, ISA-XGBoost, reaches 94.12%, indicating high prediction accuracy. In addition to the feature importance analysis of the four feature indexes, it was determined that the maximum tangential stress of the surrounding rock is the most important feature.

In order to reduce the occurrence of rock burst accidents during construction, the rock burst intensity should be assessed. In this paper, we propose a new rock burst prediction model based on the improved sandcat swam optimization-kernel based extreme learning machhine (ISCSO-KELM) algorithm. The maximum tangential stress, uniaxial compressive strength, uniaxial tensile strength and rock elastic energy index were selected as the evaluation indexes of rock burst. 105 domestic and international examples of rock burst were selected as samples for machine learning. Comparison of the relative ratios of the model presented herein with confusion matrix predicted by models including random forest (RF), K-nearest neighbor (KNN), support vector machine (SVM) and kernel based extreme learning machhine (KELM) models shows that, the ISCSO-KELM model is superior at assessing both evaluation accuracy and recall. The evaluation accuracy of the model reached 96.774 2%, indicating the superiority of ISCSO-KELM. Relevant engineering cases were used to verify the rock burst intensity. The results show that ISCSO-KELM model is more effective in capturing the connection between rock burst intensity and the indexes, thus providing a new highly applicable method for rock burst prediction.