Blasting fragmentation prediction method based on machine learning and extreme learning machine hybrid model

Abstract

The present application relates to a blasting block size prediction method based on a machine learning and extreme learning machine hybrid model, comprising the following steps: A) obtaining original data of blasting, determining variable parameters thereof, and calculating skewness values of the variable parameters; B) classifying the variable parameters according to the skewness values, transforming the variable parameters of each class based on a set transformation method to generate new variable parameters, and processing the new variable parameters based on at least one machine learning method to generate a new feature data set; C) determining the number of hidden layer neurons of an extreme learning machine model according to the new feature data unit, obtaining the best extreme learning machine model network structure accordingly, and saving the corresponding weight parameter values in the best extreme learning machine model network structure; and D) predicting the average block size of blasting based on the new feature data set, the best extreme learning machine model network structure and the weight parameter values. The blasting block size prediction method based on the machine learning and extreme learning machine hybrid model can reduce the amount of calculation and has high precision.

Description

TECHNICAL FIELD

[0001] The present application relates to a blasting prediction method, in particular to a blasting fragmentation prediction method based on a machine learning and extreme learning machine hybrid model. BACKGROUND

[0002] Blasting is a widely used rock breaking method in underground space engineering, mining engineering and tunnel engineering. Its main purpose is to divide the intact rock mass or large stone into small pieces.

[0003] The size of the rock fragmentation after blasting directly affects the productivity of downstream operations and the economic benefits of the entire project. Large rock blocks can greatly reduce the full load rate during the loading stage, and usually require additional secondary blasting to further break up larger rock blocks, which not only increases costs, but also poses new safety problems for equipment and workers. On the contrary, too fine a blasting fragmentation will result in increased explosives consumption and costs, ore loss during transportation, and ultimately recovery, so accurately predicting the size distribution of rock fragmentation after each blasting is an effective way to analyze and optimize blasting design.

[0004] In the prior art, fragmentation size can be measured by two methods: direct method and indirect method. The most commonly used direct method is the sieve method, which sieves through sieves of different aperture, and then weighs the weight of the fragments passing through different sieves to obtain the overall fragmentation distribution characteristics. Although more accurate results can be obtained in this way, the time and labor costs are huge. The indirect method includes observation method, experience method and image photography method, and further combines digital analysis processing to obtain the distribution of fragmentation size. The existing digital processing algorithm of fragmentation size distribution is too redundant and complex in details, resulting in low prediction efficiency and poor generalization ability. Often a formula can only adapt to one scene.

[0005] Therefore, it is necessary to design a blasting fragmentation prediction method based on a machine learning and extreme learning machine hybrid model. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a blasting fragmentation prediction method based on a machine learning and extreme learning machine hybrid model.

[0007] In order to solve the above technical problems, the present application provides a blasting fragmentation prediction method based on a machine learning and extreme learning machine hybrid model, comprising the following steps:

[0008] A) Obtain the original data of blasting, determine the variable parameters of the original data, and calculate the skewness value of each variable parameter of the original data;

[0009] B) classifying the variable parameters according to the skewness values of the variable parameters, transforming the variable parameters of each class based on a set transformation method to generate new variable parameters, and processing the new variable parameters based on at least one machine learning method to generate a new feature data set;

[0010] C) determining the number of hidden layer neurons of an extreme learning machine model based on the new feature data unit, obtaining the best extreme learning machine model network structure accordingly, and saving the corresponding weight parameter values in the best extreme learning machine model network structure;

[0011] D) predicting the average block size of blasting based on the new feature data set, the best extreme learning machine model network structure, and the weight parameter values.

[0012] Specifically, the variable parameters include blasting design parameters: S / B, H / B, B / D, T / B, explosive parameters Pf, and rock mechanical property parameters: , E, where S represents the hole spacing, B represents the row spacing, H represents the bench height, D represents the hole diameter, T represents the stemming length, Pf represents the explosive unit consumption, X B represents the field rock block size, and E represents the Young's modulus of elasticity of the rock.

[0013] Further, the skewness value is obtained based on the standard deviation of each variable parameter:

[0014]

[0015]

[0016] wherein, is the variable parameter, n is the sample size of the variable parameter, S represents the standard deviation of the variable parameter, is the skewness value.

[0017] Further, the transformation process of the variable parameters includes: for the variable parameters with a skewness value greater than 0, using the root mean square method to transform; for the variable parameters with a skewness value less than 0, using the method of taking the opposite number to obtain new variable parameters :

[0018] .

[0019] Further, the machine learning method includes a random forest regression algorithm, an adaptive boosting regression algorithm, a gradient boosting regression algorithm, and an extreme tree regression algorithm. Each of the new variable parameters The random forest regression algorithm, the adaptive boosting regression algorithm, the gradient boosting regression algorithm, and the extreme tree regression algorithm are used for machine learning, and each of the new variable parameters The new feature data set obtained after machine learning becomes the new variable parameter .

[0020] Further, the new variable parameter The step of processing by using the random forest regression algorithm comprises:

[0021] A1) For the new variable parameter The best split variable j and split point s are selected based on a split selection formula, and the split selection formula is as follows:

[0022]

[0023] Wherein: is the sample output mean of the D1 data set, c2 is the sample output mean of the D2 data set, the split variable j is traversed, for the fixed split variable j, the split point s is traversed, and the split pair (j, s) that makes the split selection formula reach the minimum value is found;

[0024] A2) For the selected split pair (j, s), the region is divided and the corresponding output value is determined:

[0025]

[0026]

[0027] Wherein, is the value of the variable parameter in the data set corresponding to the variable parameter to be split, is the value of the variable parameter in the data set corresponding to the variable parameter to be split, , is the selected split variable , is or the number of samples in the region, is the divided and , is and the corresponding output value X 50 , representing and the mean of the output;

[0028] A3) dividing the original data into M sub-regions , generating a final decision tree:

[0029]

[0030] wherein I represents an identity matrix.

[0031] Further, the new variable parameters The step of processing by using the adaptive boosting regression algorithm comprises:

[0032] B1) initializing the weight of the variable parameters, defining the weight of each sample of the new variable parameters initialized as 1 / N:

[0033]

[0034] wherein, represents the number of loop iterations;

[0035] B2) loop iteration, on the basis of the sample distribution of the variable parameters , training weak classifiers by using the training set, calculating the maximum error of the weak classifier on the training set:

[0036]

[0037] wherein, represents the prediction result of the weak classifier on the new variable parameters , and represents the target value of the new variable parameters ;

[0038] B3) according to the maximum error , calculating the relative error of each new variable parameter of the weak classifier :

[0039]

[0040] and according to the relative error , calculating the error rate of the weak classifier :

[0041]

[0042] wherein, represents the weight value corresponding to each of the new variable parameters; based on the error rate updating the weights of the weak classifiers :

[0043]

[0044]

[0045]

[0046] wherein, represents the weight coefficient of the weak classifier , represents the weight of the last updated sample, represents the weight of the current updated sample, represents the normalization factor;

[0047] B4) ending the iteration, based on the error rate obtaining the strong regressor :

[0048]

[0049] wherein, is the weight of all the weak classifiers , the median of the results, and .

[0050] Further, the new variable parameters processed by the gradient boosting regression algorithm include:

[0051] C1) initializing the base learner:

[0052]

[0053]

[0054] wherein, represents the mean square error loss function, is a constant, representing the mean value of the target values of all new variable parameters .

[0055] C2) calculating the negative gradient of the variable parameters after t iterations:

[0056] ​

[0057] where denotes the partial derivative of the loss function, and ) , the first regression tree corresponding to the leaf node region is :

[0058]

[0059] where is the number of leaf nodes, and the best fitting value is calculated for the leaf node region :

[0060] ;

[0061] C3) Update the strong learner based on the leaf node region :

[0062] .

[0063] Further, the new variable parameter is processed using the extreme tree regression algorithm, including the following steps:

[0064] D1) For a random split pair (j, s), divide the region and determine the corresponding output value:

[0065]

[0066] ;

[0067] D2) Divide the original data into M sub-regions , and generate the final decision tree:

[0068] .

[0069] Further, the steps for determining the optimal extreme learning machine model network structure include calculating the root mean square error and R-square between the average block value predicted by the extreme learning machine model and the true block value, determining the number of neurons of the extreme learning machine model corresponding to the smallest root mean square error and the largest R-square, thereby obtaining the optimal extreme learning machine network structure, and saving the corresponding weight parameter value in the network structure:

[0070]

[0071] where, is the activation function, is the weight value of the i-th neuron in the hidden layer, is the weight value of the i-th neuron output layer, bias value of the i-th hidden layer neuron, number of neurons in the hidden layer, output result value, taking the root mean square error and R square as the evaluation criteria, taking the root mean square error minimum and R square maximum, the corresponding number of neurons is the optimal value:

[0072]

[0073]

[0074] wherein, new variable parameter corresponding target value, new variable parameter corresponding predicted value, all new variable parameters mean value of the target value.

[0075] Through the above technical scheme, in the blasting block size prediction method based on the machine learning and extreme learning machine hybrid model provided by the application, after the variable parameters of the data are determined, the skewness value of the variable parameters is calculated first, so that the variable parameters can be classified according to the skewness value, and then each type of variable parameter is converted based on the corresponding conversion method to obtain new variable parameters. According to the skewness information corresponding to different blasting variable parameters, different conversion methods are used, which can make the entire data set more conform to the normal distribution, and can ensure the accuracy of the prediction. The new variable parameters are processed based on at least one machine learning method, so that each new variable parameter can obtain a new feature data set with multiple new feature data after being processed by at least one machine learning method. The new feature data set obtained by processing by the machine learning method is a low-dimensional feature parameter, which can avoid the data complexity and modeling difficulty caused by directly using high-dimensional parameters with multiple influences as input variables, reduce the modeling difficulty of prediction, and reduce the calculation amount of the prediction process. Then, the number of hidden layer neurons of the extreme learning machine model is determined according to the new feature data unit, and the optimal extreme learning machine model network structure is obtained accordingly, and the corresponding weight parameter value in the optimal extreme learning machine model network structure is saved, so that the blasting average block size can be predicted based on the new feature data set, the optimal extreme learning machine model network structure and the weight parameter value. Using the extreme learning machine to predict the blasting block size can reduce the number of model training optimization iterations, quickly obtain the prediction model, and fully guarantee the accuracy, model robustness and reliability of the blasting average block size prediction.

[0076] Other advantages of the application and technical effects of the preferred embodiments will be further described in the specific embodiments below. Attached Figure Description

[0077] Figure 1 This is a flowchart of the blasting block degree prediction method based on a hybrid model of machine learning and extreme learning machine according to the present invention;

[0078] Figure 2 This is a flowchart of the new feature generation process in the explosive block degree prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0079] Figure 3 This is a bar chart of the variable parameter S / B in the explosive block size prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0080] Figure 4 This is a bar chart of the variable parameter H / B in the explosive block size prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0081] Figure 5 This is a bar chart of the variable parameter B / D in the explosive block size prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0082] Figure 6 This is a bar chart of the variable parameter T / B in the explosive block size prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0083] Figure 7 This is a bar chart of the variable parameter Pf in the explosive block size prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0084] Figure 8 This is a bar chart of the variable parameter XB in the explosive block size prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0085] Figure 9 This is a bar chart of the variable parameter E in the blasting block size prediction method based on the hybrid model of machine learning and extreme learning machine in this invention;

[0086] Figure 10 This is a bar chart of the variable parameter X50 in the explosive block size prediction method based on the hybrid model of machine learning and extreme learning machine of this invention. Detailed Implementation

[0087] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0088] like Figure 1As shown, as an embodiment of the blasting block degree prediction method based on a hybrid model of machine learning and extreme learning machine provided by the present invention, the method includes the following steps:

[0089] A) Obtain the original data of the blasting. At the blasting site, the blasting design parameters can be measured by equipment such as laser rangefinders and measuring tapes, and the rock mechanical property parameters can be obtained by indoor rock mechanics experiments. Determine the variable parameters of the original data and calculate the skewness value of each variable parameter of the original data.

[0090] B) Classify the variable parameters according to their skewness values, transform the variable parameters of each class according to the set transformation method to generate new variable parameters, and process the new variable parameters according to at least one machine learning method to generate a new feature dataset.

[0091] C) Determine the number of hidden layer neurons in the Extreme Learning Machine model based on the new feature data units, obtain the optimal Extreme Learning Machine model network structure accordingly, and save the corresponding weight parameter values ​​in the optimal Extreme Learning Machine model network structure.

[0092] D) Predict the average block size of blasting based on the new feature dataset, the optimal extreme learning machine model network structure, and the weight parameter values.

[0093] Specifically, the variable parameters include blasting design parameters: S / B, H / B, B / D, T / B, explosive parameter Pf, and rock mechanical property parameters: E, where S represents the borehole spacing, B represents the row spacing, H represents the step height, D represents the borehole diameter, T represents the tamping length, Pf represents the explosive consumption per unit volume, and X represents the tamping length. B The scale of the rock block is represented by E, and the Young's modulus of elasticity of the rock can be obtained by means of... Figures 3 to 10 The bar chart shown above illustrates the distribution of the seven variable parameters.

[0094] Furthermore, the skewness values ​​of the above seven variable parameters are calculated. skewness value It is obtained based on the standard deviation of each variable parameter:

[0095]

[0096]

[0097] in, Here, n is the number of samples for the variable parameter, and S represents the standard deviation of the variable parameter. When the skewness is <0, it is called negative skewness, and the data has a long tail on the left. When the skewness is >0, it is called positive skewness, and the data has a long tail on the right. Based on this, the above 7 types of variable parameters are classified so that different types of variable parameters can be transformed in different ways.

[0098] Furthermore, the transformation process of the variable parameters includes: transforming variable parameters with skewness values ​​greater than 0 using the root mean square method; and transforming variable parameters with skewness values ​​less than 0 using the inverse method to obtain new variable parameters. (No conversion is needed if the skewness value is 0):

[0099]

[0100] By using different transformation methods based on the skewness information corresponding to different explosion variable parameters, the entire dataset can be made to better conform to a normal distribution and be closer to the actual situation.

[0101] Furthermore, machine learning methods include Random Forest Regression (RFR), Adaptive Boosting Regression (ABR), Gradient Boosting Regression (GBR), and Extreme Tree Regression (ETR), which are applied to the parameters of each new variable. Machine learning algorithms such as Random Forest Regression, Adaptive Boosting Regression, Gradient Boosting Regression, and Limit Tree Regression are used to generate new features, such as... Figure 2 As shown, the original dataset D={( X 1, Y1), ( X 2, Y2), ... ( X N-1 Y N-1 ), ( X N Y N )}, where each set of data X i Includes S / B, H / B, B / D, T / B, Pf, X B E has 7 parameter variables. Y i This represents the corresponding target block size value, for each set of parameters. X i Machine learning algorithms such as Random Forest Regression, Adaptive Boosting Regression, Gradient Boosting Regression, and Limit Tree Regression all yield four new feature values. F 1. F 2. F 3. F 4. And set the parameters of each new variable. New feature data obtained after machine learning ( F 1. F 2. F 3. F 4) The set becomes the parameter of the new variable. The new feature data set. Therefore, the original blasting data dimension is N x 7, and the new feature dimension generated after conversion by the machine learning model is N x 4. The new feature parameters are generated by the random forest regression algorithm, the adaptive boosting regression algorithm, the gradient boosting regression algorithm, and the extreme tree regression algorithm. On the one hand, all factors that can possibly affect the blasting fragmentation are fully considered. On the other hand, by generating new low-dimensional feature parameters, the problem of data complexity and modeling difficulty caused by directly using multiple impact (high-dimensional) parameters as input variables is avoided.

[0102] Specifically, the new variable parameters The steps of processing by the random forest regression algorithm include:

[0103] A1) For the new variable parameters The best split variable j and split point s are selected based on the split selection formula as follows:

[0104]

[0105] Wherein: is the sample output mean of the D1 data set, c2 is the sample output mean of the D2 data set, and the split variable j is traversed. For a fixed split variable j, the split point s is traversed, and the split pair (j, s) that makes the split selection formula reach the minimum value is found. Specifically, taking Table 1 as an example:

[0106]

[0107] Table 1

[0108] Suppose the split variable j is the first variable S / B among all variable parameters, and the split point s is set to 1.2, that is, the variable S / B whose value is less than or equal to 1.2 is classified into the D1 data set, and the variable S / B whose value is greater than 1.2 is classified into the D2 data set. Then the D1 data set and the D2 data set are shown in Table 2 and Table 3, respectively:

[0109]

[0110] Table 2

[0111]

[0112] Table 3

[0113] Then the newly divided D1 data set and D2 data set are further divided according to the above method, and the process is repeated;

[0114] A2) For the selected split pair (j, s), the region is divided and the corresponding output value is determined:

[0115]

[0116]

[0117] wherein, is the corresponding variable parameter to be split in the data set of all variable parameters , is the region where the value of the corresponding variable parameter to be split in the data set of all variable parameters is greater than the split point s, is the corresponding variable parameter to be split in the data set of all variable parameters , is the selected variable parameter to split, , is or the number of samples in the region, is the divided and , is and the corresponding output value X 50 , denotes and the mean value of the output, taking the example in step A1), assuming that our split variable j is the first variable S / B among all parameters, and our split point s is 1.2, that is, (j, s) is (S / B, 1.2), then the parameter meaning in x respectively: x represents all characteristic variables S / B, H / B, B / D, T / B), explosive parameters (Pf) and rock mechanics property parameters ( , E); denotes the characteristic we selected to split this time (for example, S / B); that s represents the split point value (for example, 1.2); so denotes the region in all data sets corresponding to S / B value less than or equal to s (i.e., Table 2); similarly is to represent the region in all data sets corresponding to S / B value greater than s (i.e., Table 3); denotes or the number of samples in the region; is the divided two regions and ; denotes and the corresponding output value X 50 ; denotes and ​The average value of the output;

[0118] A3) Divide the original data into M sub-regions. Generate the final decision tree:

[0119]

[0120] Where I represents the identity matrix.

[0121] Specifically, for the new variable parameters The steps involved in processing using the adaptive boosting regression algorithm include:

[0122] B1) Initialize the weights of the variable parameters and define each new variable parameter. The weights of all samples are initialized to 1 / N:

[0123]

[0124] in, Indicates the number of loop iterations;

[0125] B2) Iterate in a loop, with the sample distribution of the variable parameters as follows: Based on this, a weak classifier is trained using the training set. Calculate the weak classifier Maximum error on the training set :

[0126]

[0127] in, Representing a weak classifier For new variable parameters The prediction results Indicates new variable parameters Target value;

[0128] B3) Based on the maximum error Calculate the weak classifier For each new variable parameter relative error :

[0129]

[0130] And based on relative error Calculate the weak classifier error rate :

[0131]

[0132] in, denotes the weight value corresponding to each new variable parameter; based on error rate updating the weight of the weak classifier :

[0133]

[0134]

[0135]

[0136] wherein, denotes the weight coefficient of the weak classifier , denotes the weight of the last updated sample, denotes the weight of the current updated sample, denotes the normalization factor; B4) ending the Tthiteration, based on error rate

[0137] obtaining the strong regressor :

[0138]

[0139] wherein, is the weight of all weak classifiers , the median of the output result, and .

[0140] Further, the new variable parameter is processed by the gradient boosting regression algorithm, and the steps include:

[0141] C1) initializing the base learner:

[0142]

[0143]

[0144] wherein, denotes the mean square error loss function, is a constant, representing the mean value of the target value of all new variable parameters ;

[0145] C2) calculating the negative gradient of the variable parameter after t iterations:

[0146]

[0147] wherein denotes the partial derivative of the loss function, using ( ) , obtaining the​​​ The leaf node region corresponding to the regression tree is :

[0148]

[0149] wherein is the number of leaf nodes, and the leaf node region , the best fitting value is calculated:

[0150] ;

[0151] C3) Based on the leaf node region update the strong learner:

[0152] .

[0153] Further, the new variable parameter is processed by the extreme tree regression algorithm. The steps are similar to the steps of the random forest regression algorithm, except that the extreme tree regression uses all training samples to obtain each decision tree. At the same time, the extreme tree regression is completely random to obtain the split value. The other regression process is exactly the same as the random forest regression algorithm. Specifically, the new variable parameter is processed by the extreme tree regression algorithm, and the steps include:

[0154] D1) For a random split pair (j, s), divide the region and determine the corresponding output value:

[0155]

[0156] ;

[0157] D2) Divide the original data into M sub-regions , and generate the final decision tree:

[0158] .

[0159] Further, the steps for determining the best extreme learning machine model network structure include: calculating the root mean square error and R square between the average block value predicted by the extreme learning machine model and the true block value, determining the number of neurons of the extreme learning machine model corresponding to the smallest root mean square error and the largest R square, thereby obtaining the best extreme learning machine network structure, and saving the corresponding weight parameter value in the network structure:

[0160]

[0161] wherein, represents the activation function, represents the weight value of the i-th neuron in the hidden layer, denotes the weight value of the i-th neuron output layer, denotes the bias value of the i-th hidden layer neuron, denotes the number of neurons in the hidden layer, denotes the output result value (i.e. the prediction result of the blasting fragmentation), and the root mean square error and R square are taken as the judgment criteria, the root mean square error is taken as the minimum, the R square is taken as the maximum, and the number of neurons corresponding to the minimum root mean square error and the maximum R square is taken as the best value

[0162]

[0163]

[0164] wherein, denotes the new variable parameter corresponding target value, denotes the new variable parameter corresponding predicted value, denotes all new variable parameters the mean value of the target value.

[0165] Through the above technical scheme, in the blasting fragmentation prediction method based on the machine learning and extreme learning machine hybrid model provided by the application, after the variable parameters of the data are determined, the skewness value of the variable parameters is calculated first, so that the variable parameters can be classified according to the skewness value, and then each type of variable parameter is converted based on the corresponding conversion method to obtain a new variable parameter. According to the skewness information corresponding to different blasting variable parameters, different conversion methods are used, which can make the entire data set more conform to the normal distribution, and can ensure the accuracy of the prediction. The new variable parameters are processed based on at least one machine learning method, so that each new variable parameter can obtain a new feature data set with multiple new feature data after being processed by at least one machine learning method. The new feature data set obtained by processing by the machine learning method is a low-dimensional feature parameter, which can avoid the problems of data complexity and modeling difficulty caused by directly using high-dimensional parameters with multiple influences as input variables, reduce the modeling difficulty of prediction, and reduce the calculation amount of the prediction process. Then, the number of hidden layer neurons of the extreme learning machine model is determined according to the new feature data unit, and the best extreme learning machine model network structure is obtained, and the corresponding weight parameter value in the best extreme learning machine model network structure is saved, so that the blasting average fragmentation can be predicted based on the new feature data set, the best extreme learning machine model network structure and the weight parameter value. Using the extreme learning machine to predict the blasting fragmentation can reduce the number of model training optimization iterations, quickly obtain the prediction model, and fully guarantee the accuracy, model robustness and reliability of the blasting average fragmentation prediction.

[0166] ​The preferred embodiments of the present application are described in detail above with reference to the drawings, but the present application is not limited thereto. Within the technical concept of the present application, various simple modifications can be made to the technical solutions of the present application, including that various specific technical features are combined in any suitable manner. In order to avoid unnecessary repetition, the present application does not further describe various possible combination manners. However, these simple modifications and combinations should also be regarded as the disclosed content of the present application, and all fall within the protection scope of the present application.

Claims

1. A method for predicting the blasting fragmentation based on a hybrid model of machine learning and extreme learning machine, characterized in that, The method comprises the following steps: A) obtaining original data of blasting, determining variable parameters of the original data, and calculating skewness values of the variable parameters of the original data; B) classifying the variable parameters according to the skewness values of the variable parameters, transforming each type of the variable parameters based on a set transformation method to generate new variable parameters, and processing the new variable parameters based on at least one machine learning method to generate a new feature data set; C) determining the number of hidden layer neurons of an extreme learning machine model according to the new feature data set, obtaining a best extreme learning machine model network structure according to the number of hidden layer neurons, and saving corresponding weight parameter values in the best extreme learning machine model network structure; D) predicting average blockiness of blasting based on the new feature data set, the best extreme learning machine model network structure, and the weight parameter values. The skewness values are obtained based on standard deviations of the variable parameters: wherein, is the variable parameter, n is the sample size number of the variable parameter, S represents the standard deviation of the variable parameter, is the skewness value; The machine learning method includes a random forest regression algorithm, an adaptive boosting regression algorithm, a gradient boosting regression algorithm, and an extreme tree regression algorithm Machine learning is performed using the random forest regression algorithm, the adaptive boosting regression algorithm, the gradient boosting regression algorithm, and the extreme tree regression algorithm, and each of the new variable parameters The new feature data set obtained after machine learning becomes the new variable parameter of the new feature data set. 2.The blasting fragmentation prediction method based on a hybrid model of machine learning and extreme learning machine according to claim 1, wherein, The variable parameters include the blasting design parameters: S / B, H / B, B / D, T / B, the explosive parameter Pf and the rock mechanics property parameter X , E, where S represents the hole spacing, B represents the row spacing, H represents the bench height, D represents the hole diameter, T represents the stemming length, Pf represents the unit consumption of explosive, X B represents the field rock block size, and E represents the Young's modulus of the rock. 3.The blasting fragmentation prediction method based on the hybrid model of machine learning and extreme learning machine according to claim 1, wherein, The machine learning method includes a random forest regression algorithm, an adaptive boosting regression algorithm, a gradient boosting regression algorithm, and an extreme tree regression algorithm Machine learning is performed using the random forest regression algorithm, the adaptive boosting regression algorithm, the gradient boosting regression algorithm, and the extreme tree regression algorithm, and each of the new variable parameters The new feature data set obtained after machine learning becomes the new variable parameter of the new feature data set. 4.The blasting fragmentation prediction method based on the hybrid model of machine learning and extreme learning machine according to claim 3, wherein, the new variable parameter the step of processing with the random forest regression algorithm comprises A1) selecting the new variable parameter The best segmentation variable j and segmentation point s are selected based on a segmentation selection formula as follows: wherein: is the sample output mean of D1 data set, c2 is the sample output mean of D2 data set, traversing the split variable j, for the fixed split variable j, traversing the split point s, the split pair (j, s) is obtained so that the split selection formula reaches the minimum value; A2) dividing regions and determining corresponding output values for selected split pairs (j, s): wherein, for all the variable parameters in the data set corresponding to the variable parameter to be split whose value is less than or equal to the split point s, for all the variable parameters in the data set corresponding to the variable parameter to be split whose value is greater than the split point s, x for all the variable parameters , for the variable parameter selected to split, for or the number of samples in the region, for the divided and , for and the corresponding output value X 50 , denotes the mean of the outputs in and . A3) dividing the original data into M sub-regions , generating a final decision tree: where I represents a unit matrix. 5.The blasting fragmentation prediction method based on the hybrid model of machine learning and extreme learning machine according to claim 4, wherein, The new variable parameter The step of processing with the adaptive boosting regression algorithm comprises: B1 ) initializing the weights of the variable parameters, defining each of the new variable parameters the weights of the samples of the new variable parameters are initialized to 1 / N: wherein denotes the number of loop iterations; B2) iteratively cycling, on the basis of a sample distribution of the variable parameters , training weak classifiers using the training set, calculating the maximum error of the weak classifiers on the training set. wherein, denotes the weak classifier a prediction result of the new variable parameter , denotes a target value of the new variable parameter . B3) calculating said maximum error , said weak classifier for each said new variable parameter relative error : and according to said relative error , the error rate of said weak classifier is calculated wherein, represents a weight value corresponding to each of the new variable parameters; based on the error rate updating the weights of the weak classifiers :​ wherein, denotes a weight coefficient of a weak classifier denotes a weight coefficient of a weak classifier denotes a weight of a last updated sample, denotes a weight of a current updated sample, denotes a normalization factor; B4) ending the iteration based on the error rate obtaining a strong regressor : wherein, is the weight of all the weak classifiers the median of the results, and .​ 6.The blasting fragmentation prediction method based on the hybrid model of machine learning and extreme learning machine according to claim 5, wherein, the new variable parameter the step of processing with the gradient boosting regression algorithm comprises C1) initializing a base learner: wherein, represents a mean squared error loss function, is a constant representing the mean of the target values of all new variable parameters . C2) calculating negative gradients of the variable parameters after t iterations: wherein denotes the partial derivation of the loss function, using (1) ) , the leaf node region corresponding to the first regression tree is :​ wherein is the number of leaf nodes, for the leaf node region , computing the best fit value: ； C3) based on the leaf node region updating the strong learner: 。 7.The blasting fragmentation prediction method based on the hybrid model of machine learning and extreme learning machine according to claim 6, wherein, The new variable parameter The step of processing with the limit tree regression algorithm comprises: D1) dividing regions and determining corresponding output values for random split pairs (j, s): ； D2) dividing the original data into M sub-regions generating a final decision tree: 。 8.The method of claim 7, wherein the method is characterized by, The step of obtaining the best extreme learning machine model network structure comprises the following steps: calculating root mean square errors and R squares between predicted average blockiness values of the extreme learning machine model and true blockiness values, determining the number of neurons of the extreme learning machine model corresponding to the minimum root mean square error and the maximum R square, thereby obtaining the best extreme learning machine network structure, and saving corresponding weight parameter values in the network structure: wherein, represents an activation function, represents a weight value of the i-th neuron in the hidden layer, represents a weight value of the i-th neuron output layer, represents a bias value of the i-th hidden layer neuron, represents the number of neurons in the hidden layer, represents an output result value, and the root mean square error and the R square are taken as the evaluation criteria, the root mean square error is the smallest, the R square is the largest, and the number of neurons corresponding to the root mean square error is the best value: wherein, denotes the new variable parameter the corresponding target value, denotes the new variable parameter the corresponding predicted value, denotes all new variable parameters the mean of the target values.

Images