Research on Rockburst Classification Prediction Based on BP-SVM Model

Rockburst is a complex destabilization phenomenon which is a combination of multiple factors, the study of rockburst for classification prediction can help prevent and control engineering geological hazards, reduce casualties and property damage. To achieve efficient and accurate rockburst classification prediction and solve the problem of rockburst propensity assessment, six evaluation factors are selected as the rock explosion prediction and evaluation system: tangential stress <inline-formula> <tex-math notation="LaTeX">$\sigma _{\theta }$ </tex-math></inline-formula>, uniaxial compressive strength <inline-formula> <tex-math notation="LaTeX">$\sigma _{\mathrm {c}}$ </tex-math></inline-formula>, uniaxial tensile strength <inline-formula> <tex-math notation="LaTeX">$\sigma _{\mathrm {t}}$ </tex-math></inline-formula>, tangential stress to uniaxial compressive strength ratio <inline-formula> <tex-math notation="LaTeX">$\sigma _{\mathrm {\theta }}/\sigma _{\mathrm {c}}$ </tex-math></inline-formula> (BCF), uniaxial compressive strength to tensile strength ratio <inline-formula> <tex-math notation="LaTeX">$\mathrm {\sigma }_{\mathrm {c}}/\sigma _{\mathrm {t}}$ </tex-math></inline-formula> (SCF), and elastic deformation energy index <inline-formula> <tex-math notation="LaTeX">$\mathrm {W}_{\mathrm {et}}$ </tex-math></inline-formula> in this study. Widely collected domestic and international groups of rock explosion evaluation data, and 420 sets of valid samples were obtained by data processing. Establish rockburst grading prediction evaluation models based on BP neural networks and support vector machines respectively, then establish BP-SVM prediction models based on arithmetic mean weights and standard deviation weights, analyzing and comparing the prediction rating results of 120 groups of samples among them. Accuracy, Precision, Recall, Specificity, and F1 Score metrics are selected to evaluate the performance of different models, the results show that several models can obtain effective prediction results, among which the standard deviation weight combination BP-SVM model proposed in this paper has the best prediction accuracy and the best effect, which is better than the traditional single machine learning method.


I. INTRODUCTION
Rock explosion is a sudden geological disaster caused by the accumulation of elastic strain energy after the rapid release, it is in a state of high stress or limit equilibrium of the rock or geological structure in a very short period of time, the violent release of internal stored strain energy, resulting in the dynamic destabilization of the rock around the excavation space, which is a complex phenomenon of multi-factor synthesis. It is generally accepted that the necessary conditions for the rock explosion to occur are: (a) The rock is in a high stress state, and the rock has a high strength with its principal stress greater than or close to the strength of the rock. (b)The elasticity and brittleness of the rock is large, when the rock is damaged by the force, there will be ejected, flying burst The associate editor coordinating the review of this manuscript and approving it for publication was Li Zhang . in the form of strong shock waves instantly releasing a lot of kinetic energy, resulting in rockburst phenomenon [1]. With the development of deeper mining, the flake spalling of the surrounding rock in the empty area caused by rock explosion, flake gang, loud noise and rock ejection, causing a great threat to the life and property safety of staff, has also become an important disaster of underground space engineering. Because of its sudden, complex, catastrophic, difficult to control and other characteristics, rock explosion classification prediction study, rock explosion propensity study, rock explosion risk evaluation and other issues have attracted the attention of a large number of experts and scholars. Scientific and reasonable prediction, assessment and prevention of rock blast problems for engineering geological aspects have great significance.
Rockburst predictions are usually divided into short-term and long-term predictions. Short-term predictions are usually VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ made during the rock excavation phase, with corresponding models built from field monitoring data to predict the risk of near-term rockbursts. Long-term prediction is usually carried out at the beginning of the project design or excavation, using rock mechanics parameters and the nature of the original rock to establish predictive models to determine the propensity of rockbursts to occur at different locations, but it is not possible to derive the evolutionary trend of rockbursts over time.
Commonly used rockburst monitoring techniques include electromagnetic radiation method [2], drill chip method [3], infrared thermal imaging method [4], micro-gravity method [5], micro-seismic monitoring method [6], acoustic emission method [7], etc. Short-term rockburst risk prediction studies include:time-series precursor characteristics method [8], fractal theory [9], probabilistic method [10], empirical method [11], machine learning method [12], etc. Long-term rockburst risk prediction studies include: empirical evidence method [13], multi-indicator assembly method [12], uncertainty theory (such as fuzzy integrated evaluation method [14], cloud Model [15], object element topology model [16], set-pair analysis method [17]and so on), comprehensive ranking method [18], numerical simulation method [19], mutation theory [20], machine learning method [21], etc. The rock explosion prediction classification is shown in Figure 1. Different forecasting methods' advantages and disadvantages are as follows: 1) In time-series precursor characteristics method, the precursor information obtained based on the source parameters in different engineering geological conditions is difficult to apply universally, and the prediction thresholds for different rockburst risks are difficult to determine precisely; 2) Although the fractal theory method can deeply explore the spatio-temporal fractal characteristics of energy, it is difficult to determine the critical value of the energy fractal dimension corresponding to different rockburst risks.
3) Probabilistic method can predict the degree of risk of rock blast and the probability of occurrence at the same time, but it requires the establishment of a large database of cases. Besides, there is subjectivity in the scoring of each evaluation index, and the experience derived from the method may not be universal.

4)
The empirical method is the easiest to operate and understand, but the method is the most subjective, requires a large number of experts with accumulated experience, and is not applicable to large and complex systems. 5) Multi-indicator assembly method integrates a variety of factors, simple and convenient operation, but the threshold of rockburst classification there is subjectivity. 6) Uncertainty theory can deal with both qualitative and quantitative information, taking into account the combined factors and the uncertainty of the rockburst prediction process, but the method needs to determine the weight of each indicator, and the corresponding affiliation function is difficult to determine precisely. 7) Comprehensive ranking method is often combined with fuzzy theory to combine the effects of doing factors, but it is more difficult to judge the standard sample, and the method still needs to determine the weights of each indicator. 8) Numerical simulation method can simulate the aggregation, transfer and release of energy processes, determine the location of the rock explosion, the depth of the crater, but the construction of the model and rock parameters are difficult to determine, and the model is affected by the parameters.
9) The mutation theory approach is highly theoretical and the derivation of the formula is rigorous, but the method is only applicable to a few simple engineering environments.
Since different methods have their unavoidable shortcomings, rockburst prediction grading research is evolving and innovating, from traditional rockburst prediction methods to cross-fertilization in the field of deep learning. Machine learning methods have powerful ability to handle nonlinear complex problems and have been widely used in several fields with good results. Machine learning methods are applicable to both long-term and short-term rockburst prediction, however, the methods are mostly focused on shallow simulation and prediction at present. Commonly used machine learning methods include neural networks, decision trees, plain Bayes, random forests, support vector machines, etc. For instance, Rui et al. [22] validated a Dropout-based improved deep neural network model using sample data from engineering examples(DA-DNN); Li et al. [23] optimized extreme learning machine hyperparameters using genetic algorithms. Wu et al. [24] supervised rockburst learning using COPULA theory with support vector machine models; Xuebin et al. [25] used the XGBoost algorithm for rockburst prediction and assigned weights to the sample data using the CRITIC algorithm.;Lin et al. [26] compared the Cloud model, Bayesian model and random forest methods respectively to analyze their rockburst classification prediction performance;Pu et al. [27] analyzed the decision strategies through support vector machine models and Gaussian processes respectively. Gao [28] improved the computational accuracy of the traditional ant colony algorithm by using an abstract ant colony clustering algorithm for rockburst hierarchy prediction. Pu et al. [29] predicted the rockburst risk using support vector machine method by recalibrating the original data through K-means clustering method. In view of this, this paper proposes a combined modeling approach using BP neural networks and support vector machines for hierarchical prediction of rockburst samples.
Since BP neural networks are prone to local minimal value problems and SVM methods may have dimensional errors, this paper proposes a combined BP-SVM model. For the multi-factorial nature of rock blasts, uncertainty, and the advantages and disadvantages of different prediction methods, according to existing research [21], [30]- [32], the rock tangential stress σ θ , rock uniaxial compressive strength σ c , rock uniaxial tensile strength σ t , tangential stress to rock uniaxial compressive strength ratio σ θ /σ c (BCF), rock uniaxial compressive strength to tensile strength ratio σ c /σ t (SCF), elastic deformation energy index W et six evaluation factors as a rock burst prediction evaluation system. Extensively collected domestic and international groups of rock explosion evaluation data, processed to obtain 420 sets of valid samples. Establish rockburst grading prediction evaluation models based on BP neural networks and support vector machines respectively, then establish BP-SVM prediction models based on arithmetic mean weights and standard deviation weights, analyze and compare the prediction rating results of 120 groups of these samples. Accuracy, Precision, Recall, Specificity, and F1 Score metrics are selected to evaluate the performance of different models, which provide some theoretical support and empirical guidance for practical engineering applications.

II. BASIC THEORY AND PREDICTION ALGORTITHM
A. BP NEURAL NETWORK BP neural network (Back Propagation), a feed-forward neural multilayer network trained according to error back propagation, is also the core of forward neural network. The basic BP algorithm consists of two processes: signal forward propagation and error back propagation.(a) In forward propagation, the signal passes from the input layer through each neuron in the implicit layer in turn and then out of the output layer, outputting the predicted value of each node.(b)When backward propagation, if there is an error between the predicted value and the actual value in the output layer, the error is calculated recursively layer by layer and the connection weights of the nodes in each layer are corrected and re-entered into forward propagation, and it will be repeated until the best mapping is achieved [33]- [35]. The typical BP neural network feedforward topology model is shown in Figure 2. Let the number of nodes in the input layer be M, the mth neuron be xm, the input signal vector X = (x1, x2, . . . , xm)T; the number of nodes in the hidden layer is L and the lth neuron is kl. The number of nodes in the output layer is J, the jth neuron is yj, and the output signal vector Y = (y1, y2, . . . , yj)T. The weight is w, the threshold is b, the expected output is d, the number of iterations is n, and the expected value is d. The implied layer transfer function is a Sigmoid function and the output layer transfer function is a Purelin linear function, as shown in Eqs. (1) and (2), respectively.
From Figure 2,(a) the input to the implicit layer is the weighted sum of the input layer signal with the connection weights wmi and threshold bi of each node.(b)The input of the output layer is the weighted sum of the output of the implicit layer and the connection weights wki and threshold bki. The output models of nodes in the implicit layer and output layer are shown in Eqs. (3) and (4), respectively.
The output result is compared with the expected value: the calculation is finished if the accuracy meets the requirement, and if the accuracy does not meet the requirement, the weights of each neuron are corrected according to the gradient descent method and transferred to the back propagation stage until the requirement is met. The model errors and weight corrections are shown in Eqs. (5) and (6), respectively.
where, η denotes the learning rate, δ denotes the local gradient, and v denotes the upper layer signal. This signal propagation and feedback method is the basic principle of machine learning of BP neural network.

B. SUPPORT VECTOR MACHINE
Support Vector Machine (SVM) is one of the machine learning algorithms for supervised learning, which can be used for data classification, regression fitting and outlier detection based on supervised learning principles [36]. The main principle of its action is to find the optimal plane to separate two or more classes of data according to the maximum interval of two or more classes of variables in the feature space, so that the data can be classified. It is characterized by using a subset of the training set as a support vector to represent the decision boundary, which is the maximum edge hyperplane, and finally finding the optimal hyperplane that maximizes the distance from a point in space to itself [37]. The schematic diagram of the support vector machine structure is shown in Figure 3. SVM methods can be divided into three categories: linearly divisible, linearly indivisible, and nonlinearly divisible. The rockburst grading prediction study in this paper falls into the category of nonlinearly divisible, so the principles of nonlinearly divisible support vector machines are described in detail here. For nonlinear support vector machines, the following principles exist [38], [39] If the properties of the data (i.e., the original space) are finite, there must exist spaces of higher or infinite dimensions such that the variables become linearly separable after mapping. The nonlinear regression model can be represented by Eq. (7) and the model optimization function is Eq. (8).
where, y,ω T (x) +b ≥ 1 − ξ, ξ ≥ 0, i = 1, 2, . . . , t.ω is the normal vector in the decision surface, (x) is the mapping function, b is the displacement term, C is a non-negative penalty parameter, ξ i is a non-negative relaxation factor. According to the Lagrange multiplier method, the pairwise problem is obtained as in Eq. (9): where, α i is non-negative Lagrangian multipliers. From the equation above, it can be seen that the mapping function is not directly involved in finding the parameters, so the kernel function is introduced to facilitate the next step of calculation. The commonly used kernel functions are linear kernel function, polynomial kernel function, Gaussian radial kernel function (RBF), etc. In this paper, the Gaussian radial kernel function (RBF) is mainly selected as the kernel function of the model for regression prediction. With the introduction of the kernel function, the nonlinear support vector machine model can be represented by Eq. (10). where, The advantage over traditional statistical methods is that the method does not require a large amount of sample data, which enables learning, classification, prediction and regression fitting of small samples. The advantages over neural networks include easier training, good learning and generalization performance, absence of local minima, and no need to determine the network topology in advance. In the field of regression prediction, the support vector machine model exerts better results and achieves better structural risk minimization.

C. BP-SVM COMBINED MODEL
Since BP neural networks are prone to the problem of local minima and SVM methods may have dimensional errors, the combined BP-SVM model proposed in this paper combines the two methods to take advantage of the strengths of both models in order to complement each other's strengths and strive to get the best results and more accurate prediction values. Compared to traditional single-method forecasting, portfolio forecasting is a combination of two or more methods in some way, with more emphasis on the absorption of information derived from different methods. The key to the combination method is to find the right weights, and only when the weights are right, the results will appear close to the real situation. The commonly used combination forecasting methods include arithmetic mean method and standard deviation combination method, etc. In this paper, the arithmetic mean method and standard deviation combination method are used to assign weights for weighted combination, respectively. The  weights are calculated as shown in Eq. (11) and (12) [40]: where, σ i is the standard deviation of the prediction error in the ith prediction model. σ is the sum of standard deviations of a single model, n is the number of prediction models. The process of the model prediction method is shown in Figure 4.

A. DETERMINATION OF EVALUATION INDICATORS
Rockburst is a highly complex and nonlinear dynamic instability phenomenon influenced by many factors. There is no clear definition of the mechanism of the occurrence of rockbursts, and has not formed a unified standard, it is generally believed that the necessary conditions for the occurrence of rockbursts are: (a) The rock is in a high stress state and the rock has a high principal stress strength which is greater than or close to the rock strength. (b)The elasticity and brittleness of the rock is large, when the rock is damaged by the force, will be ejected, flying burst in the form of a strong shock wave instantly release a lot of kinetic energy, resulting in rock burst phenomenon [1]. Therefore, the rock explosion factors and rock stress state, rock properties, rock energy storage is highly correlated. Combined with the occurrence of rockburst conditions and rockburst impact characteristics, according to existing research [21], [30]- [32], the rock tangential stress σ θ , rock uniaxial compressive strength σ c , rock uniaxial tensile strength σ t , tangential stress to rock uniaxial compressive strength ratio σ θ /σ c (BCF), rock uniaxial compressive strength to tensile strength ratio σ c /σ t (SCF), elastic deformation energy index W et six evaluation factors as a rock burst prediction evaluation system. Rockburst intensity classification is not yet standardized, with reference to the research results of Wang Yuanhan and other scholars [14], [41]- [43], taking into account the main factors affecting the degree of rock explosion intensity, in this paper, the rock explosion classification is classified as: None, Light, Moderate, Strong [22], [44]. The specific values and indicators are shown in Table 1.

B. ROCKBURST PREDICTION SAMPLE DATA PROCESSING
Using the literature survey method, extensive collection of domestic and international rock explosion sample data [21], [24], [45]- [47], and the sample data were pre-processed. The pre-processing consists of three steps as follows. (a)Sieve out samples with missing information in the data, eliminate duplicate data, and clean up unwanted data. (b)The abnormal data sets in the sample are identified by drawing box plots, and replacement values are randomly generated from the interval of values consisting of the mean and median to bring in the abnormal data sets. (c)The mapminmax function is called to normalize the data to between [0,1] and to denormalize the output data at the end of training. Normalization can avoid errors in sample data due to different orders of magnitude, yet accurately retain the relationships between the original data sets without introducing new biases. The normalized formula is shown in Eq. (13).
where, y is the normalized data, x is the original data, min x is the minimum value in x, and max x is the maximum value in x.
In the sample data set, the four rockburst classes were resampled separately, thus avoiding the problem of sample imbalance. A new sample set is formed by sampling 1:1:1:1 for None, Light, Moderate and Strong respectively. Draw the first 70% as the training and validation sets for the training samples. The latter 30% of the data were used as the test set, separately using the trained BP neural network model, the support vector machine model, and the improved SVM-BP model for the graded prediction of rock bursts to compare the accuracy of different methods.

C. ALGORITHM MODEL PREDICTION
In this paper, MatlabR2016a is used for training and prediction of the sample data. BP neural network, support vector machine model, and improved BP-SVM combined model were used to build a training set for the first 70% of the sample data respectively to obtain a model with better fit, and then the last 30% of the sample data were brought into the training model for prediction regression fitting to obtain the prediction results of the three algorithmic models. In addition, to prevent overfitting due to complex models, a cross-validation approach is introduced here, i.e., repeated training and validation using a randomly generated subset of samples. Crossvalidation, also known as round robin estimation, requires the following two conditions to be met: (a)The proportion of the training set is greater than 50%. (b)The training set and test set samples are to be generated by uniform sampling. This study used 10-fold cross-validation [21], [48].

D. COMPARATIVE STUDY OF PREDICTION RESULTS
In classification prediction, the first criterion for the degree of model prediction of interest is accuracy (also known as precision), which is the ratio of the number of accurately predicted samples to the overall sample size. Similarly, the ratio of the number of incorrectly predicted samples to the total number of samples is the error rate (also known as error). In the machine learning process, there are two types of errors, one is the training error that occurs during the training process and the other is the generalization error that occurs during the testing process. The smaller the training error, the higher the accuracy of model building, but it tends to lead to overfitting; while the generalization error is often unpredictable, so other metrics need to be introduced to judge the accuracy of the prediction model.   Table 2. The commonly used model evaluation metrics are Accuracy, Precision, Recall, Specificity, Falsepositive Rate, F1 Score, etc. The specific meanings and calculation formulas are shown in Table 3. In this paper, Accuracy, Precision, Recall, Specificity, and F1 Score metrics are chosen to evaluate the evaluation performance of different models. Accuracy represents the ratio of the number of accurate samples to the total number of samples; Precision represents the accuracy of the predicted results in positive cases; Recall represents the recall rate, which represents the completeness of the search for positive cases in the predicted results; usually, the accuracy is negatively correlated with the recall rate [49].

IV. EXAMPLE ANALYSIS A. SAMPLE DATA PROCESSING
According to part III-A of the article, the rock tangential stress σ θ , rock uniaxial compressive strength σ c , rock uniaxial tensile strength σ t , tangential stress to rock uniaxial compressive strength ratio σ θ /σ c (BCF), rock uniaxial compressive strength to tensile strength ratio σ c /σ t (SCF), elastic deformation energy index W et six evaluation factors as a rock burst prediction evaluation system.the rock explosion classification is classified as: None, Light, Moderate, Strong.
Using the literature survey method, a wide range of domestic and international rock explosion sample data were collected, after eliminating duplicate data and missing values, a total of 275 sets of sample data with complete impact factors and classification results were obtained. The ratio of  be seen that there are no outliers in the newly formed sample data, and the pre-processed sample data meet the training requirements.

B. BP NEURAL NETWORK AND SVM MODLE PREDICTION
After pre-processing the data, there are 105 groups of each category (N,L,M,S), 30 groups are randomly selected from each category as the test set and 75 groups as the training and validation sets, where the data order of the training and validation sets are randomly assigned by matlab. In total, 120 sets of prediction samples and 300 sets of training samples are used. To facilitate the prediction results, each category (N,L,M,S) is assigned the values (1,2,3,4) respectively. Using BP neural network with support vector machine model, the predicted values of the test samples are obtained after repeated training and taking the best fit results.
The comparison of the prediction results of the two models is shown in Figure 6. It can be visualized from the figure that both the BP neural network and support vector machine models can produce more accurate prediction results, but the SVM model is overall closer to the real results with less curve fluctuations. The standard deviation of the prediction errors of BP neural network and support vector machine are 23.4192 and 15.5132 respectively, which are more accurate overall for SVM model.
From part II-C of this paper, the arithmetic mean combination model and the standard deviation weight combination model are obtained. Let y n be the predicted value of the BP-SVM combined prediction model, into from Eq. (11), the arithmetic mean combined model equation is: From Eq. (12), the combined weights of the BP and SVM models are derived as: w1 = 0.3985, w2 = 0.6015. Therefore, the BP-SVM standard deviation combination prediction model equation is: The BP-SVM model prediction results are further calculated from the combined model, and the histogram of the VOLUME 10, 2022  predicted data of the combined BP-SVM model with two weights is shown in Figure 7. The standard deviation of the prediction error of the arithmetic mean combination model is 9.4783, and the standard deviation of the prediction error of the standard deviation weight combination model is 9.0875. The combination model has substantially lower errors and more accurate predictions than even the original model, but the standard deviation weight combination model is better than the arithmetic mean combination model. See Appendix 2 for specific values.

C. MODEL EVALUATION
After the above analysis, it can be seen that there is a certain difference between the correctness of the rockburst prediction results and the degree of model optimization under the four methods, but it is not obvious from the result graph alone. The following section evaluates the performance of different models in terms of Accuracy, Precision, Recall, Specificity, and F1 Score metrics. For the convenience of representation, BP-SVM-1 is used to represent the arithmetic mean combined BP-SVM model, and BP-SVM-2 is used to represent    the standard deviation weight combined BP-SVM model. The index statistics are shown in Table 4.
It can be seen that the prediction accuracy of all four models is high, among which the standard deviation weight combination BP-SVM model has the largest accuracy of 97.5%, which is better than the BP model (86.67%), SVM model (90.83%) and arithmetic mean combination BP-SVM model (94.17%). The F1 Score values of the four models are obtained for the overall evaluation of Precision and Recall on the basis of Precision and Recall. Among the four models, the F1 Score values are at a high level, among which the standard deviation weight combination BP-SVM model has the highest accuracy rate of 0.975, which is better than the BP model (0.875), SVM model (0.912) and arithmetic mean combination BP-SVM model (0.941). It indicates that the combined model has better classification prediction performance than the single model and can compensate the shortcomings of the single model to some extent. Moreover, the standard deviation weight combination method can start from the error of the prediction results, and the improvement effect is more optimized than the arithmetic average combination. As can be seen from Figure 8, the evaluation indicators of the four models fluctuate consistently, and the dotted line graph clearly illustrates the optimality ranking of the models as: standard deviation weight combination BP-SVM model, arithmetic mean combination BP-SVM model, SVM model, BP model.
The standard deviation weight combination BP-SVM model has achieved good results in this rockburst grading prediction study, but there are still shortcomings, such as: (a) Rockburst model impact indicators are relatively small. Input indicators were selected only six, taking into account the complexity of rock blast generating factors, there is a need to introduce more technical indicators, geological factors, mechanical indicators, etc.(b)The sample size is relatively small. Although the sample size was expanded to 420 data sets during the data processing phase, there is still room for development of an accurate training set. The next step will be to expand the sample database capacity and make further adjustments to the model training.

V. CONCLUSION
(1) The rock tangential stress σ θ , rock uniaxial compressive strength σ c , rock uniaxial tensile strength σ t , tangential stress to rock uniaxial compressive strength ratio σ θ /σ c (BCF), rock uniaxial compressive strength to tensile strength ratio σ c /σ t (SCF), elastic deformation energy index W et six evaluation factors as a rock burst prediction evaluation system.the rock explosion classification is classified as: None, Light, Moderate, Strong.
(2) BP neural network and support vector machine were used to develop hierarchical prediction of rockburst sample database, and BP-SVM combined model was established by arithmetic mean combined weights and standard deviation combined weights, to analyze the prediction results of four rockburst grades under different algorithm models.
(3) The prediction accuracy of all four models was high, and the standard deviation weight combination BP-SVM model had the highest accuracy of 97.5%, which was better than the BP model (86.67%), SVM model (90.83%) and arithmetic mean combination BP-SVM model (94.17%). The F1 Score values are all at a high level, and the standard deviation weight combination BP-SVM model has the highest accuracy of 0.975, which is better than the BP model (0.875), the SVM model (0.912) and the arithmetic mean combination BP-SVM model (0.941). The model optimality ranking is: standard deviation weight combination BP-SVM model, arithmetic mean combination BP-SVM model, SVM model, and BP model.
(4)The standard deviation weight combination BP-SVM model has achieved good results in this rockburst classification prediction study, but there are still deficiencies in the rockburst model in terms of fewer impact indicators and relatively small sample size. The next step will be to expand the sample database capacity and make further adjustments to the model training.
JIANG GUO is currently a Professor of Central South University, a Ph.D. Supervisor, and the Director of the Hunan Institute of Rock Mechanics. He is also a Visiting Scholar at The University of Tokyo. He also involved in metal mining and mine filling teaching and research work. He has presided over or undertaken more than ten projects, including the National Science and Technology Support Program, the National Natural Science Foundation Project, the Provincial Natural Science Foundation, and the Nonferrous Metals Fund. He is the Director of the Hunan Blasting Society and the Hunan Rock Mechanics Society. He is a member of the Mine Information Intelligence Professional Committee of China Nonferrous Metals Society.
JINGWEN GUO was born in Shandong, in 1998. She is currently pursuing the master's degree in safety engineering with the School of Resource and Safety Engineering, Central South University. Her research interests include safety evaluation, risk assessment, and safety economics.
QINLI ZHANG is currently a Professor of Central South University and a Ph.D. Supervisor. He is a well-known expert in the field of filling mining in China. He has published more than 200 professional papers related to the research of this project. His research interests include mining methods for complex and difficult mining bodies, research and design of filling materials and filling systems, hydraulic calculation of filling pipeline conveying, filling management, and residual ore recovery in mining areas. He has been awarded two first-class, three second-class, and three third-class prizes for scientific and technological progress at provincial and ministerial level, two invention patents, and five monographs.
MINGJIAN HUANG was born in Changsha, Hunan, China. He received the Ph.D. degree in engineering. He is currently the Deputy General Manager of Hongda Blasting Engineering Group Ltd., the General Manager of Angang Mining and Blasting Company Ltd., and a Senior Engineer. His research interests include theoretical research, scheme design, operation and safety management related to mine shaft construction, and mining and engineering blasting. VOLUME 10, 2022