A method and system for intelligent prediction of blasting vibration PPV of jointed rock mass

Abstract

The present invention belongs to the technical field of rock blasting prediction, and discloses a method and system for intelligent prediction of blasting vibration PPV of jointed rock mass. The method collects blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions; different geology includes different lithology and joint parameters; the collected blasting vibration PPV test data is used to divide the training set and the test set according to the training set random sampling method based on the working condition grouping, and on this basis, a jointed rock mass blasting vibration PPV prediction model based on the gradient boosting algorithm is established; the jointed rock mass blasting vibration PPV intelligent prediction model is verified and applied. The present invention can more quickly and accurately predict the PPV of the jointed rock mass at different distances from the blasting surface, realize the optimization of blasting parameters and the stability evaluation of the surrounding rock of the jointed rock mass, reduce the impact of blasting vibration on the surrounding environment, protect the ecological environment, and meet the concept of green construction.

Description

Technical Field

[0001] The invention belongs to the technical field of rock mass blasting prediction, and in particular relates to a jointed rock mass blasting vibration PPV intelligent prediction method and system. Background Art

[0002] The drilling and blasting method is widely used in the excavation and construction of rock mass projects such as tunnels and mines due to its wide adaptability to geological conditions and flexible cross-section excavation forms. However, the blasting vibration generated by blasting operations is an inevitable negative problem in the process of rock mass excavation. The propagation of blasting vibration will have different degrees of impact on the surrounding environment and structures. Strong blasting vibration may cause the rock layer to loosen, crack, or even collapse, which will destroy the stability of the surrounding rock of tunnels and mines, and often cause major engineering disasters. Therefore, it is very important to monitor, predict and control the rock blasting vibration.

[0003] At present, the intensity of blasting vibration is often characterized by the particle peak velocity PPV, and empirical equations are often used to predict the blasting vibration PPV. It has been verified that the empirical equation has good accuracy in predicting the blasting vibration PPV of intact rock mass, but the accuracy is low when predicting the blasting vibration PPV of jointed rock mass. With the continuous development of underground engineering, the geological environment of the surrounding rock of underground engineering is becoming more and more complex. The traditional empirical prediction equation of blasting vibration PPV is difficult to meet the needs of engineering. There is an urgent need for a high-precision and high-efficiency prediction method for blasting vibration PPV of jointed rock mass. With the advancement of artificial intelligence big data machine learning methods and the reduction of the cost of related intelligent model training, it has become possible to establish a high-performance intelligent diagnostic model based on ensemble learning methods using a large amount of blasting vibration PPV data monitored by field tests. This technology can timely and accurately predict the blasting vibration PPV, thereby realizing the safe, efficient and green construction of rock mass engineering excavation by drilling and blasting.

[0004] Through the above analysis, the problems and defects of the existing technology are: the traditional empirical prediction method has low accuracy in predicting the blasting vibration PPV of jointed rock mass and cannot meet the requirements of fast and high-precision prediction on site. Summary of the invention

[0005] In order to overcome the problems existing in the related art, the disclosed embodiment of the present invention provides a method and system for intelligent prediction of blasting vibration PPV of jointed rock mass. The method is based on the collected blasting vibration PPV data under different geological conditions (different lithology and joint parameters), different explosive equivalents, and different monitoring distances, and uses an integrated learning method to establish an intelligent jointed rock mass blasting vibration PPV prediction model, which overcomes the problem that the traditional empirical prediction method has low accuracy in predicting blasting vibration PPV of jointed rock mass and cannot meet the requirements of fast and high-precision prediction on site.

[0006] The technical solution is as follows: A jointed rock mass blasting vibration PPV intelligent prediction method, comprising:

[0007] S1: Collect blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions; different geology includes different lithology and joint parameters;

[0008] S2: Establishment of intelligent prediction model of blasting vibration PPV of jointed rock mass. The collected blasting vibration PPV test data are used to divide the training set and the test set according to the random sampling method of the training set based on the working condition grouping. On this basis, a PPV prediction model of blasting vibration of jointed rock mass based on the gradient boosting algorithm is established.

[0009] S3: Validation and application of PPV intelligent prediction model for blasting vibration in jointed rock mass.

[0010] In step S1, blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions are collected, including:

[0011] S1.1: For tunnels constructed by drilling and blasting, a monitoring test working face is selected, and core samples are taken along the sides, top and bottom of the tunnel. Physical and mechanical parameters such as rock density, elastic modulus, uniaxial compressive strength and Poisson's ratio are measured by indoor test methods. The number, length, inclination and dip angle of joints within 50m of the blasting excavation face are statistically analyzed.

[0012] S1.2: Installation of explosives with different blast equivalents and layout of blasting vibration PPV monitoring points;

[0013] S1.3: At different blasting sites, select different geological conditions and repeat the above steps to collect blasting vibration PPV data under different geological conditions.

[0014] In step S2, the jointed rock mass blasting vibration PPV intelligent prediction model is established, including:

[0015] S2.1: Collect sample data under different working conditions to divide the data into training set and test set. Use the random sampling method of training set based on working condition grouping to make the extracted training set data contain different geological working conditions and blasting parameter combinations; the training set accounts for 80% of the total data and the test set accounts for 20%;

[0016] S2.2: Establishment of LightGBM gradient lifting jointed rock mass blasting vibration PPV intelligent prediction model;

[0017] S2.3: Hyperparameter optimization of the PPV intelligent prediction model for blasting vibration in jointed rock mass.

[0018] In step S2.1, a training set random sampling method based on working condition grouping is adopted, including:

[0019] (1) Select important features from the training data set to construct a grouping list. The factors in the grouping list include two types of factors: on-site geological conditions and blasting conditions. The on-site geological conditions refer to the geological parameters considered in the international rock mass classification standard RMR, including lithology, uniaxial compressive strength, joint spacing, and joint direction parameters; the blasting conditions include the amount of explosives;

[0020] (2) Refine the grouping of training set data according to the important features in the selected list;

[0021] (3) Obtained For each group, 80% of the data is randomly selected as the training set and 20% is randomly selected as the test set.

[0022] (4) Combine the training sets in all groups into a sample training set, and combine the test sets in all groups into a sample test set.

[0023] In step (2), the training set data is refined and grouped according to the important features in the selected list, including:

[0024] (2.1) Group the monitoring data according to the first feature;

[0025] (2.2) Group each group of data after grouping according to the first feature according to the second factor in the grouping list;

[0026] (2.3) Repeat the above steps. Each feature in the list has m groups, and there are n features in total, resulting in m to the power of n groups. For the two features of lithology and joint spacing, each feature has two groups, resulting in lithology 1 joint spacing 1, lithology 1 joint spacing 2, lithology 2 joint spacing 1, and lithology 2 joint spacing 2, for a total of four working conditions.

[0027] In step S2.2, the LightGBM gradient boosting jointed rock blasting vibration PPV intelligent prediction model is established, including: using the blasting vibration training set data to train the gradient boosting decision tree model, first discretizing different input features to obtain histogram bins, and using the histogram statistical method to obtain the best leaf splitting node to construct a decision tree, where the best leaf splitting node is the splitting node with the largest split gain value; based on the current model residual, the next level of decision tree is constructed until the iterative gradient boosting model achieves the minimum prediction error of blasting vibration PPV; the objective function of the gradient boosting tree is composed of the mean square error and the L2 regularization term, and the expression is:

[0028] ;

[0029] In the formula, To reduce the loss function of the error between the model and the true value, are model parameters, are the true sample value and the predicted value respectively, is the sample size, is the L2 regularization strength parameter, is the penalty term of the L2 norm of the complexity of the model difference parameters, is the total number of model parameters, For the model parameters;

[0030] The method of using the blasting vibration training set data to perform gradient boosting decision tree model training includes:

[0031] (1) Gradient boosting decision tree uses decision tree as the basic weak learner unit, and constructs a strong learner to predict the burst PPV by accumulating and iteratively optimizing the weak learner;

[0032] (2) In each round of iteration, a new decision tree is constructed to fit the residuals of the previous round, and the error of the previous decision tree is corrected by calculating the negative gradient direction of the residuals;

[0033] The method of obtaining the best leaf splitting node by using the histogram statistical method to construct a decision tree includes:

[0034] (i) Discretize each eigenvalue under the node into k integers and construct a histogram;

[0035] (ii) Traverse the samples and add the gradient value of each sample to the corresponding histogram interval. Each feature histogram bin interval obtains the corresponding gradient sum;

[0036] (iii) Perform traversal and segmentation calculations on the bins under each feature to obtain the optimal split gain of the feature; compare the split gain values ​​of different features to obtain the optimal feature and bin split value;

[0037] The next level of decision tree is constructed based on the current model residual until the iterative gradient boosting model achieves the minimum prediction error of the blasting vibration PPV, including:

[0038] (a) Make predictions based on the initial decision tree model and calculate the residual of the current initial decision tree model;

[0039] (b) construct the prediction residual of the next level decision tree, and superimpose the next level decision tree and the current initial decision tree model according to the weight; calculate the negative gradient value of the superimposed strong prediction model to update the superimposed model;

[0040] (c) Repeat the above steps until the set number of iterations is reached.

[0041] Furthermore, the blasting vibration training set data is used to train the gradient boosting decision tree model, including:

[0042] S2.2.1, construct the initial decision tree model;

[0043] S2.2.11, feature discretization, for a feature in the training data , discretize its value into bins, each bin width is:

[0044] ;

[0045] In the formula, is the bin width value, Features The maximum value in Features The minimum value in ;

[0046] S2.2.12, build a histogram; for each feature, count the samples according to the bins, and the current node contains samples, calculate the sum of the gradient information in each bin , the expression is:

[0047] ;

[0048] In the formula, For the binning The partial derivative of the loss function for each sample in is For binning The partial derivative of each prediction sample in ;

[0049] S2.2.13, select the optimal splitting point for the current feature, calculate the post-splitting gain for each splitting point between different bins, and for a given splitting point , the left side of the split point contains the 1st to Boxes, total samples, the right side contains To Boxes, total samples, the expression is:

[0050] ;

[0051] In the formula, At a given split point Split gain value at position;

[0052] Traverse all possible bin splitting points of the current feature and select the splitting point that maximizes the gain for splitting. The optimal splitting point selection formula is as follows:

[0053] ;

[0054] In the formula, To select the maximum value of splitting gain at different bin positions in the current feature;

[0055] Select the optimal segmentation points for all features; obtain the optimal segmentation point set under each feature parameter for different feature parameters ,in Represents the number of features;

[0056] ;

[0057] In the formula, To represent traversal of all The split node corresponding to the maximum split gain value of the feature;

[0058] S2.2.14, split the original data into a left node and a right node according to the optimal split point, and repeat the selection of the optimal split node for the two nodes according to the above step S2.2.13 until the maximum depth specified by the decision tree is reached to complete the construction of the initial decision tree;

[0059] S2.2.2, residual calculation and model updating;

[0060] Based on the decision tree model obtained in step S2.2.1, the training set data is predicted to obtain the predicted value , according to the parameters of the extreme model of predicted value and true value, the formula is as follows:

[0061] ;

[0062] In the formula, For the The true value in the iteration is The residuals between rounds of predictions;

[0063] Train a new weak learner to fit the residual and obtain a decision tree model that can fit the minimum residual error ;

[0064] ;

[0065] In the formula, To select the decision tree with the smallest fitting error, For all decision trees, Based on the input features The trained decision tree;

[0066] Add the new weak learner model to the current model and obtain the optimal coefficient by calculating the negative gradient between the iterative model and the training data. , the expression is:

[0067] ;

[0068] In the formula, To select the weight coefficient corresponding to the minimum loss, To calculate the loss between the iterative model and the true value, For the The decision tree obtained in turn, is the weight coefficient to be selected, For the The decision tree with the smallest fitting residual is obtained through round training;

[0069] After correcting the residual of the previous model, we get the current model:

[0070] ;

[0071] In the formula, For the t The decision tree after rounds of corrected residuals, For the The decision tree obtained in turn, For the The weight coefficient of the wheel, For the The decision tree with the smallest fitting residual is obtained;

[0072] S2.2.3, repeat the previous step S2.2.2 until the predetermined number of iterations is reached;

[0073] Repeat step S2.2.2 for residual calculation and model update. After T rounds of iterations, the final prediction model is obtained as follows: The accumulation of:

[0074] ;

[0075] In the formula, is the integrated decision tree after T rounds of iterations, is the initial decision tree before iteration, For the Iterations, is the total number of iterations.

[0076] In step S2.3, the hyperparameters of the jointed rock mass blasting vibration PPV intelligent prediction model are optimized, including:

[0077] S2.3.1, Parameter definition of the hyperparameter combination space to be optimized; Parameter search of the hyperparameter setting in the integrated gradient learning method in continuous or discrete space;

[0078] S2.3.2, traverse all hyperparameters and perform random sampling of parameters from each parameter definition space according to certain sampling criteria;

[0079] S2.3.3, perform gradient boosting tree model training and objective function evaluation according to a set of randomly obtained hyperparameters;

[0080] S2.3.4, repeat the above steps until the specified number of iterations is reached and the stopping condition is obtained to obtain the optimal hyperparameter combination.

[0081] In step S3, the jointed rock mass blasting vibration PPV intelligent prediction model verification and application includes:

[0082] S3.1. Based on the test set data, the performance of the established PPV intelligent prediction model for blasting vibration of jointed rock mass is verified to evaluate its prediction error and model robustness under different geological conditions. The root mean square index is used to compare the error between the actual PPV value and the model prediction value under different explosive equivalents and joint numbers under blasting.

[0083] S3.2. Engineering application of PPV intelligent prediction model for blasting vibration of jointed rock mass;

[0084] S3.2.3, repeat the above steps to obtain the maximum explosive equivalent and PPV values ​​at different distances to ensure tunnel blasting safety;

[0085] In step S3.2, the engineering application of the jointed rock mass blasting vibration PPV intelligent prediction model includes:

[0086] S3.2.1, when using the drilling and blasting method to make a real-time and rapid prediction of the blasting vibration PPV of the jointed rock mass during the construction process, it is necessary to obtain the physical and mechanical parameters of the rock mass within 50m of the construction blasting excavation location, the parameters mainly include rock sample density, elastic modulus, uniaxial compressive strength, Poisson's ratio physical and mechanical parameters; count the number, length, inclination and dip parameters of the joints within 50m of the on-site blast source; collect the above parameters as input parameters for analyzing the current blasting vibration safety analysis;

[0087] S3.2.2, based on engineering experience, assume the minimum explosive equivalent required for blasting and excavation of the current section , with a unit increment of 2kg, assuming that the explosive equivalent of the blasting surface is ; Each assumed explosive equivalent and the measured rock physical and mechanical parameters and joint parameters are respectively input into the jointed rock blasting vibration PPV intelligent prediction model to obtain the blasting vibration PPV values ​​at different assumed explosive equivalents and different distances from the blasting source; By analyzing the PPV at different blasting source distances from the blasting source combined with safety criteria, the safety status under the current explosive equivalent is obtained.

[0088] Another object of the present invention is to provide a jointed rock mass blasting vibration PPV intelligent prediction system, the system implements the jointed rock mass blasting vibration PPV intelligent prediction method, the system comprises:

[0089] The data collection module is used to collect blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions; different geology includes different lithology and joint parameters;

[0090] The module for establishing the intelligent prediction model of blasting vibration PPV of jointed rock mass is used to use the collected blasting vibration PPV test data to divide the training set and the test set according to the training set random sampling method based on the working condition grouping, and on this basis, establish the blasting vibration PPV prediction model of jointed rock mass based on the gradient boosting algorithm;

[0091] Verification and application module, used for verification and application of PPV intelligent prediction model for blasting vibration in jointed rock mass.

[0092] Combined with all the above technical solutions, the beneficial effects of the present invention are as follows: the present invention collects blasting vibration PPV monitoring data under various geological conditions (different lithology and joint parameters) and different monitoring distances under explosive equivalent; processes the acquired data according to different working conditions, and obtains training set and test set data randomly sampled by group; uses the training set data to train the jointed rock blasting vibration PPV intelligent prediction model and uses random space search to optimize the model parameters; uses the test set data to verify the generalization prediction performance of the model. The present invention constructs a jointed rock blasting vibration PPV intelligent prediction model based on ensemble learning based on blasting vibration PPV data under different geological conditions (different lithology and joint parameters), which can more quickly and accurately predict the PPV of the jointed rock at different distances from the blasting surface, realize the optimization of blasting parameters and the stability evaluation of the surrounding rock of the jointed rock, reduce the impact of blasting vibration on the surrounding environment, protect the ecological environment, and conform to the concept of green construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0094] Figure 1 It is a flow chart of the PPV intelligent prediction method for blasting vibration of jointed rock mass provided by an embodiment of the present invention;

[0095] Figure 2 Schematic diagram of PPV field monitoring of tunnel surrounding rock blasting vibration provided by an embodiment of the present invention;

[0096] Figure 3 It is the national blasting vibration GB677-2014 standard diagram provided by the embodiment of the present invention;

[0097] Figure 4 It is a block diagram of the PPV intelligent prediction method for blasting vibration of jointed rock mass provided by an embodiment of the present invention;

[0098] Figure 5 It is a specific implementation principle diagram of the jointed rock mass blasting vibration PPV intelligent prediction method provided by an embodiment of the present invention;

[0099] Figure 6 This is a graph of the actual value and predicted value prediction results in the training set;

[0100] Figure 7 This is a graph of the actual values ​​and predicted values ​​in the test set;

[0101] Figure 8 This is the PPV monitoring value and predicted value result diagram at different distances from the explosion source under blasting vibration conditions. DETAILED DESCRIPTION

[0102] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0103] The innovation of the present invention is that: the present invention first constructs the original engineering database based on the field monitoring data of rock mass parameters and vibration velocity under different geological conditions of blasting construction; then uses the training set random sampling method based on working condition grouping to segment the data so that the training set covers different geological conditions and blasting parameter combinations; uses gradient boosting ensemble learning and random parameter search to train the optimal prediction model and perform on-site blasting vibration PPV prediction and explosive equivalent selection application in jointed rock mass engineering. The technical solution established by the present invention can reduce the on-site monitoring cost, improve the prediction accuracy of jointed rock mass blasting vibration PPV, and has a wide range of on-site application prospects.

[0104] Embodiment 1, as Figure 1 As shown, the jointed rock mass blasting vibration PPV intelligent prediction method provided by the embodiment of the present invention includes:

[0105] S1: Collect blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions; different geology includes different lithology and joint parameters;

[0106] S1.1: In the construction tunnel excavated by drilling and blasting, a monitoring test working face is selected, and core samples are drilled along both sides, above and below the tunnel; the physical and mechanical parameters of rock sample density, elastic modulus, uniaxial compressive strength and Poisson's ratio are measured by indoor test methods; the number, length, inclination and dip angle of joints within 50m of the blasting excavation face are statistically analyzed;

[0107] S1.2: Installation of explosives with different blasting equivalents and arrangement of blasting vibration PPV monitoring points. Install explosives of different equivalents on the blasting surface; then arrange vibration PPV monitoring points at a certain interval within a range of 5m to 30m from the blasting surface, and measure the blasting wave vibration PPV under each working condition through blasting vibration tests; the schematic diagram of the blasting source position, rock joints and vibration PPV monitoring point arrangement is shown in the figure. Figure 2 shown.

[0108] S1.3: At different blasting sites, select different geological conditions (different lithology and joint parameters) and repeat the above steps S1.1 and S1.2 to collect blasting vibration PPV data under different geological conditions (different lithology and joint parameters).

[0109] S2: Establishment of intelligent prediction model of blasting vibration PPV of jointed rock mass. The collected blasting vibration PPV test data are used to divide the training set and the test set according to the random sampling method of the training set based on the working condition grouping. On this basis, a PPV prediction model of blasting vibration of jointed rock mass based on the gradient boosting algorithm is established.

[0110] Among them, the factors in the grouping list constructed by the training data set include two types of factors: on-site geological conditions and blasting construction conditions. The on-site geological conditions refer to the geological parameter grouping considered in the international rock mass classification standard RMR, including lithology, uniaxial compressive strength, joint spacing, and joint direction parameters; the blasting conditions include the amount of explosives. Specifically, the relevant geological parameter grouping is based on the classification standard of geological parameters considered in the international rock mass classification standard (RMR). The above-mentioned conditions are closely related to the prediction of blasting vibration. The relevant geological grouping is based on the international rock mass classification standard, which is not easy for non-industry-related personnel to think of.

[0111] Specifically include:

[0112] S2.1: Sample data preprocessing operation. Collect sample data under different working conditions to divide the data into training set and test set. Use the training set random sampling method based on working condition grouping to make the extracted training set data contain different geological working conditions and blasting parameter combinations; the training set accounts for 80% of the total data and the test set accounts for 20%;

[0113] The traditional method of dividing the training set and test set randomly divides the relevant data without considering the impact of the grouping of different geological conditions and blasting parameter combinations in the data set, resulting in the prediction model being unable to fully reflect the influence of jointed rock mass characteristics and lacking widespread engineering use.

[0114] The present invention adopts a training set random sampling method based on working condition grouping, including:

[0115] (1) Select important features from the training data set to construct a grouping list. The factors in the grouping list include two types of factors: on-site geological conditions and blasting conditions. The geological conditions refer to the geological parameters considered in the International Rock Mass Classification (RMR), including lithology, uniaxial compressive strength, joint spacing, and joint direction parameters; the blasting conditions include the amount of explosives.

[0116] (2) Refine the grouping of the training set data according to the important features in the selected list. The steps are as follows:

[0117] (2.1) First, the monitoring data are grouped according to the first feature;

[0118] (2.2) Group each group of data after grouping according to the first feature according to the second factor in the grouping list;

[0119] (2.3) Repeat the above steps, assuming that each feature in the list has m groups, and there are n features in total, and get m to the power of n groups. The grouping of relevant geological parameters in the working condition grouping is divided according to the standard of the international rock mass classification standard RMR; taking the two characteristics of lithology and joint spacing, and two groups for each feature as an example, we can get four working conditions, such as lithology 1 joint spacing 1, lithology 1 joint spacing 2, lithology 2 joint spacing 1, and lithology 2 joint spacing 2.

[0120] (3) The obtained m n For each group, 80% of the data is randomly selected as the training set and 20% as the test set.

[0121] (4) Combine the training sets in all groups into a sample training set, and combine the test sets in all groups into a sample test set.

[0122] S2.2: LightGBM gradient boosting intelligent prediction model for blasting vibration PPV of jointed rock mass is established. The blasting vibration training set data is used to train the gradient boosting decision tree model. First, the different input features are discretized to obtain histogram bins, and the best leaf splitting nodes are obtained through the histogram statistical method to construct the decision tree. Among them, the best leaf splitting node is the splitting node for the maximum splitting gain value pair. The next level of decision tree is constructed based on the current model residual until the iterative gradient boosting model achieves the minimum prediction error of blasting vibration PPV. The objective function of the gradient boosting tree is composed of the mean square error and the L2 regularization term:

[0123] ;

[0124] In the formula, To reduce the loss function of the error between the model and the true value, are model parameters, are the true sample value and the predicted value respectively, is the sample size, is the L2 regularization strength parameter, is the penalty term of the L2 norm of the complexity of the model difference parameters, is the total number of model parameters, For the j model parameters;

[0125] The objective function technology is used to prevent the model from overfitting, reduce the complexity of the model, and limit the model to use fewer hyperparameters; improve the generalization ability and robustness of the model so that it can be used in unseen data sets.

[0126] Exemplarily, in the gradient boosting decision tree model, the gradient boosting decision tree uses the decision tree as the basic weak learner unit, and constructs a strong learner to predict the maximum vibration velocity of the blasting by cumulatively iteratively optimizing the weak learner.

[0127] Exemplarily, using the blasting vibration training set data to perform gradient boosting decision tree model training includes:

[0128] (1) Gradient boosting decision tree uses decision tree as the basic weak learner unit, and constructs a strong learner to predict the burst PPV by accumulating and iteratively optimizing the weak learner;

[0129] (2) In each round of iteration, a new decision tree is constructed to fit the residuals of the previous round, and the error of the previous decision tree is corrected by calculating the negative gradient direction of the residuals;

[0130] In another exemplary embodiment, the steps of constructing the gradient boosting decision tree model include:

[0131] S2.2.1, construct an initial decision tree model.

[0132] S2.2.11, feature discretization. For a feature in the training data x , discretize its value into k bins, each bin width is:

[0133] ;

[0134] In the formula, is the bin width value, Features The maximum value in Features The minimum value in ;

[0135] S2.2.12, construct a histogram. For each feature, count the samples according to the bins. Assume that the current node contains samples, calculate the sum of the gradient information in each bin :

[0136] ;

[0137] In the formula, For the binning The partial derivative of the loss function for each sample in is For binning The partial derivative of each prediction sample in ;

[0138] S2.2.13, select the optimal split point for the current feature. Calculate the post-splitting gain for each possible split point between different bins. For a given split point , the left side of the split point contains the 1st to Boxes, total samples, the right side contains To Boxes, total samples, the expression is:

[0139] ;

[0140] In the formula, At a given split point Split gain value at position;

[0141] Traverse all possible bin splitting points of the current feature and select the splitting point that maximizes the gain for splitting. The optimal splitting point selection formula is as follows:

[0142] ;

[0143] In the formula, To select the maximum value of splitting gain at different bin positions in the current feature;

[0144] Select the optimal segmentation points for all features; obtain the optimal segmentation point set under each feature parameter for different feature parameters ,in Represents the number of features;

[0145] ;

[0146] In the formula, To represent traversal of all The split node corresponding to the maximum split gain value of the feature;

[0147] S2.2.14, based on the upper part, split the original data into a left node and a right node according to the optimal splitting point, and repeat the selection of the optimal splitting node for the two nodes according to the above step S2.2.13 until the maximum depth specified by the decision tree is reached to complete the construction of the initial decision tree.

[0148] S2.2.2, residual calculation and model updating;

[0149] Based on the decision tree model obtained in step S2.2.1, the training set data is predicted to obtain the predicted value , according to the parameters of the extreme model of predicted value and true value, the formula is as follows:

[0150] ;

[0151] In the formula, For the The true value in the iteration is The residuals between rounds of predictions;

[0152] Train a new weak learner to fit the residual and obtain a decision tree model that can fit the minimum residual error ;

[0153] ;

[0154] In the formula, To select the decision tree with the smallest fitting error, For all decision trees, Based on the input features The trained decision tree;

[0155] Add the new weak learner model to the current model and obtain the optimal coefficient by calculating the negative gradient between the iterative model and the training data. , the expression is:

[0156] ;

[0157] In the formula, To select the weight coefficient corresponding to the minimum loss, To calculate the loss between the iterative model and the true value, For the The decision tree obtained in turn, is the weight coefficient to be selected, For the The decision tree with the smallest fitting residual is obtained through round training;

[0158] After correcting the residual of the previous model, we get the current model:

[0159] ;

[0160] In the formula, For the t The decision tree after rounds of corrected residuals, For the The decision tree obtained in turn, For the The weight coefficient of the wheel, For the The decision tree with the smallest fitting residual is obtained;

[0161] S2.2.3, repeat the previous step S2.2.2 until the predetermined number of iterations is reached;

[0162] Repeat step S2.2.2 for residual calculation and model update. After T rounds of iterations, the final prediction model is obtained as follows: The accumulation of:

[0163] ;

[0164] In the formula, the integrated decision tree after T rounds of iterations is is the initial decision tree before iteration, For the Iterations, is the total number of iterations.

[0165] In each round of iteration, a new decision tree is constructed for the residuals of the previous round to fit the residuals. The error of the previous decision tree is corrected by calculating the negative gradient direction of the residuals, thereby improving the prediction performance of the model.

[0166] Exemplarily, obtaining the best leaf splitting node by a histogram statistical method to construct a decision tree includes:

[0167] The best splitting node needs to be determined during the decision tree generation process, so the splitting gain of each leaf node needs to be compared. For any leaf node in the decision tree, the splitting gain is calculated using the histogram statistical method:

[0168] (i) Discretize each eigenvalue under the node into k integers and construct a histogram.

[0169] (ii) Traverse the samples and add the gradient value of each sample to the corresponding histogram interval. Each feature histogram bin interval obtains the corresponding gradient sum.

[0170] (iii) Perform traversal and segmentation calculations on the bins under each feature to obtain the optimal split gain of the feature; compare the split gain values ​​of different features to obtain the optimal feature and bin split value.

[0171] For example, the next level of decision trees are constructed until the gradient boosting model after iteration achieves the minimum prediction error of jointed rock blasting vibration PPV. The principle of the gradient boosting method based on ensemble learning is to construct a strong prediction model by combining several cumulative decision trees. Subsequent decision trees are used to continuously correct the residual of the previous predictor, thereby improving the prediction performance. The steps are as follows:

[0172] (a) Make predictions based on the initial decision tree model and calculate the residual of the current initial decision tree model.

[0173] Among them, the initial decision tree model is the first generation decision tree PPV prediction model trained based on the training data. As shown in the construction steps of the gradient boosting decision tree model, it gives the training process of the initial decision tree model. It is mainly used as the first training of weak learners, and then on its basis, more decision trees are superimposed to form strong learners;

[0174] (b) Construct the prediction residual of the next level decision tree, and superimpose the next level decision tree with the current initial decision tree model according to the weight. Calculate the negative gradient value of the superimposed strong prediction model to update the superimposed model.

[0175] Among them, the steps of building the next level decision tree are the same as the initial decision tree steps, the difference is that the prediction target of the decision tree is the residual of the previous step:

[0176] ;

[0177] The next level decision tree is superimposed on the current initial decision tree model according to the weights, such as the residual calculation and model update steps. ;

[0178] The strong prediction model is a combination of several weak prediction models, which improves the prediction performance of the model by continuously eliminating residuals. Its formula is as follows: ;

[0179] The negative gradient value of the superimposed strong prediction model is calculated to update the superimposed model, such as: ;

[0180] (c) Repeat steps (a) to (b) above until the set number of iterations is reached.

[0181] S2.3: Hyperparameter optimization of the PPV intelligent prediction model for blasting vibration of jointed rock mass. Based on the above objective function, the random space search method is used to optimize the hyperparameters of the gradient boosting tree model, including the learning rate, maximum depth, number of leaf nodes and L2 regularization strength parameters. The hyperparameter optimization method uses random space search to randomly combine the parameters in the parameter combination. Based on the objective function as an indicator for evaluating the accuracy of the PPV intelligent prediction model for blasting vibration of jointed rock mass, the hyperparameter combination with the smallest prediction error is selected as the optimal parameter of the PPV intelligent prediction model for blasting vibration of jointed rock mass.

[0182] Exemplarily, the method of using random space search to optimize the hyperparameters of the gradient boosting tree model includes:

[0183] S2.3.1, definition of the spatial parameters of the hyperparameter combination to be optimized. The hyperparameters to be optimized in the LightGBM gradient boosting blasting vibration PPV intelligent prediction model are [a, b, c, …, h] (total Hyperparameters ) Set the parameter search range for each hyperparameter in .

[0184] The continuous search space of hyperparameters is usually in the form of an interval, indicating that all hyperparameters within the interval are possible candidates. For example, if the learning rate is set to [0.0001, 0.1], it means that the continuous space between 0.0001 and 0.1 is the parameter selection range.

[0185] The discrete space representation uses parameters in the set of braking parameter values ​​as candidates, and the number of parameters is usually limited. For example, the maximum depth of the tree can be selected from [3, 5, 7, 9].

[0186] S2.3.2, traverse all hyperparameters, perform random sampling of parameters from each parameter definition space according to the uniform sampling criterion, and obtain a set of random hyperparameter combinations [a1, b1, c1, …, h1].

[0187] Among them, the uniform sampling criterion in performing random sampling of parameters according to the uniform sampling criterion from each parameter definition space adopts an equal probability uniform sampling method.

[0188] S2.3.3, train the PPV intelligent prediction model for blasting vibration of jointed rock mass according to a set of randomly obtained hyperparameters [a1, b1, c1, …, h1], and record its prediction error results on the objective function.

[0189] The training of the jointed rock blasting vibration PPV intelligent prediction model adopts step S2.2.1 in the step of building a gradient boosting decision tree model to build an initial decision tree model for training;

[0190] S2.3.4, repeat the above steps S2.3.2-step S2.3.3, and continue to randomly generate a set of hyperparameter combinations until the specified number of iterations is reached and the stopping condition is obtained. The set of hyperparameters with the smallest objective function error among all the parameter combinations tried is the optimal hyperparameter combination.

[0191] Among them, the optimal hyperparameter combination is the one with the smallest objective function error among all parameter combinations tried during the hyperparameter optimization process.

[0192] S3: Verification and application of PPV intelligent prediction model for blasting vibration of jointed rock mass, including:

[0193] S3.1. Based on the test set data, the performance of the PPV intelligent prediction model for blasting vibration of jointed rock mass established in step S2 is verified to evaluate its prediction error and model robustness under different geological conditions. The root mean square index is used to compare the error between the actual PPV value and the model prediction value under different explosive equivalents and numbers of joints under blasting.

[0194] S3.2 Engineering application of PPV intelligent prediction model for blasting vibration of jointed rock mass, including:

[0195] S3.2.1, when using the drilling and blasting method to make a real-time and rapid prediction of the blasting vibration PPV of the jointed rock mass during the construction process, it is necessary to obtain the physical and mechanical parameters of the rock mass within 50m of the construction blasting excavation location, which mainly include rock sample density, elastic modulus, uniaxial compressive strength, Poisson's ratio physical and mechanical parameters; and to count the number, length, inclination, and inclination of joints within 50m of the on-site blast source. The above parameters are collected as input parameters for analyzing the current blasting vibration safety analysis.

[0196] For example, in the real-time rapid prediction of blasting vibration PPV of jointed rock mass during the drilling and blasting construction process, the integrated gradient learning method based on the gradient boosting algorithm proposed in the present invention is adopted. The improved technical effects include: avoiding the use of vibration sensing equipment for on-site test monitoring and reducing the prediction cost.

[0197] S3.2.2, based on engineering experience, assume the minimum explosive equivalent required for blasting and excavation of the current section xkg, and then in unit increments of 2 kg, the present invention innovatively proposes that, assuming that the explosive equivalent of the blasting surface is [ x+ 2, x+ 4. x+ 6. x+ 8. x+ 10]; Each assumed explosive equivalent and the measured rock physical and mechanical parameters and joint parameters are respectively input into the jointed rock mass blasting vibration PPV intelligent prediction model to obtain the blasting vibration PPV at different distances from the blasting source under different assumed explosive equivalents; by analyzing the PPV at different blasting source distances from the blasting source combined with the safety criterion, the safety status under the current explosive equivalent is obtained.

[0198] It can be understood that the explosive equivalent is [ x+ 2, x+ 4. x+ 6. x+ 8. x+ 10] formula is used to provide different explosive equivalent numerical inputs to the integrated gradient learning intelligent prediction method for prediction.

[0199] For example, the intelligent prediction model of blasting vibration PPV of jointed rock mass is an integrated gradient learning prediction machine learning model obtained by training and optimizing the LightGBM algorithm based on the on-site blasting vibration monitoring data. The model can predict the blasting vibration PPV by inputting the rock physical and mechanical parameters, joint parameters and blasting parameters during blasting construction.

[0200] For example, in step S3.2.2, the minimum blasting explosive equivalent is 4 kg. At this time, the rock density of the medium-weathered granite section is 2.5 g / cm 3 , elastic modulus 3.6GPa, uniaxial compressive strength 30MPa, Poisson's ratio 0.27, number of joints 8, the input characteristics of the predicted blasting vibration wave velocity at a distance of 10m from the blast source are [4,2.5,3.6,30,0.27,8,10]. The above matrix is ​​input into the LightGBM algorithm PPV intelligent prediction model to obtain the maximum blasting vibration velocity at the current position of 0.13m / s. At the same time, it is assumed that the explosive equivalent of the blasting surface is [ x+ 2, x+ 4. x+ 6. x+ 8. x+10]=[6,8,10,12,14] is combined with rock parameters to obtain [6,2.5,3.6,30,0.27,8,10], [8,2.5,3.6,30,0.27,8,10], [10,2.5,3.6,30,0.27,8,10], [12,2.5,3.6,30,0.27,8,10]. They are respectively input into the LightGBM algorithm PPV intelligent prediction model to obtain the blasting vibration PPV under different explosive equivalents.

[0201] For example, by analyzing the PPV at different distances from the explosion source and combining it with the safety criterion, the safety state under the current explosive equivalent is obtained. Based on step S3.2.2, the PPV values ​​at 10m from the explosion source under different explosive equivalents [4, 6, 8, 10, 12, 14] are 0.07m / s, 0.09m / s, 0.14m / s, 0.16m / s, 0.18m / s, and 0.21m / s, respectively. The predicted PPV value (frequency>50Hz) of the present invention is compared with the GB677-2014 standard (such as Figure 3 (as shown) By comparison, we can see that (1) for hydraulic tunnels, when the blasting equivalent is 6kg, the vibration velocity does not exceed the standard value of 0.1m / s, so 6kg is selected to guide on-site construction; (2) for traffic tunnels, when the blasting equivalent is 8kg, the vibration velocity does not exceed the standard value of 0.15m / s, so 8kg is selected to guide on-site construction; (3) for traffic tunnels, when the blasting equivalent is 12kg, the vibration velocity does not exceed the standard value of 0.2m / s, so 12kg is selected to guide on-site construction.

[0202] S3.2.3, repeat the above steps S3.2.1-S3.2.2 to obtain the maximum explosive equivalent and PPV at different distances to ensure the safety of jointed rock blasting, provide guidance for the selection of on-site explosive quantity, and provide support for ensuring the safe and efficient on-site blasting operations.

[0203] It can be seen from the above embodiments that the LightGBM algorithm provided by the present invention adopts a gradient-based optimization method with a fast convergence speed, can complete model training in a shorter time, and improve work efficiency. This method can effectively identify and filter noise in the data, reduce the impact of outliers on the prediction results, and improve the stability of the prediction. Automatic data analysis by intelligent algorithms reduces human intervention and subjective judgment, and reduces the uncertainty of the prediction results. By accurately predicting the PPV of blasting vibration in jointed rock mass, the risks that may be brought about by blasting operations can be assessed in advance, and corresponding safety measures can be taken to ensure the safety of the project. Accurately predicting the PPV of blasting vibration in jointed rock mass helps to optimize the blasting plan and reduce unnecessary blasting operations, thereby saving blasting materials and reducing project costs. By reducing the impact of blasting vibration on the surrounding environment and protecting the ecological environment, it is in line with the concept of green construction.

[0204] In summary, the intelligent prediction method of blasting vibration PPV of jointed rock mass based on LightGBM algorithm has significant advantages in improving prediction accuracy, efficiency and safety.

[0205] Furthermore, the present invention provides a method for intelligent prediction of PPV of blasting vibration in jointed rock mass: on the one hand, it can effectively reduce the personnel, equipment purchase and rental costs in the process of formulating, organizing and implementing on-site monitoring plans during tunnel on-site blasting construction, thereby reducing construction investment costs; on the other hand, after technological transformation, a blasting vibration PPV intelligent prediction system embedded with an intelligent prediction model is formed, which can attract on-site engineering applications of tunnel construction units, has broad intellectual property benefits, and has potential commercial value in obtaining patent licensing fees and technology transfer.

[0206] The technical solution of the present invention is based on blasting vibration PPV monitoring data under various geological conditions (different lithology and joint parameters) and different monitoring distances under explosive equivalents, combined with the powerful nonlinear learning ability in the field of machine learning, filling the technical gap in the interdisciplinary application of intelligent prediction of tunnel blasting construction process.

[0207] The traditional empirical prediction model of blasting vibration PPV has problems such as high on-site monitoring cost and poor prediction of blasting vibration PPV of jointed rock mass. This technical solution effectively uses the intelligent model to learn the engineering experience of on-site monitoring data, and achieves the purpose of quickly and accurately predicting the blasting vibration PPV of jointed rock mass based on on-site geological parameters.

[0208] The present invention breaks through the limitations of traditional empirical methods and on-site monitoring methods by using methods based on construction data and intelligent learning.

[0209] Example 2, taking the PPV on-site monitoring of blasting vibration of tunnel surrounding rock as an example, the present invention describes in detail a method for intelligent prediction of blasting vibration PPV of jointed rock mass based on the collected rock physical and mechanical properties, geological joint parameters and monitored blasting vibration PPV data at the blasting tunnel engineering site. The flow chart is shown in Figure 4.

[0210] The specific steps of the jointed rock mass blasting vibration PPV intelligent prediction method are as follows:

[0211] Step 1: Collect the PPV field monitoring test data of tunnel blasting vibration under different geological conditions (different lithology and joint parameters), such as Figure 5 As shown in step 1, its contents mainly include:

[0212] (1.1) The data collected from the on-site tunnel blasting test were collected in slightly weathered granite, tuff, and sandstone. At the same time, rock drilling was performed within 50m of the blasting test working surface under different lithologies and the physical and mechanical properties of the samples were measured in the laboratory, including density, elastic modulus, uniaxial compressive strength, and Poisson's ratio. At the same time, the joint parameters including the number, length, dip, and inclination of joints were investigated through geological mapping of the excavation surface. For example, the density of the rock sample within 50m of the blasting surface of the slightly weathered granite section was 2.6g / cm 3 , elastic modulus is 30Gpa, uniaxial compressive strength is 88MPa, Poisson's ratio is 0.25, there are 10 joints in this section, with an average length of 0.8m, a dip of NE30°, and a joint surface dip of 46°.

[0213] (1.2) Installation of different blasting explosive equivalents and layout of blasting vibration PPV measurement points on the on-site excavation surface. Taking the blasting vibration measurement test of slightly weathered granite as an example, blasting tests were carried out by installing explosives of different equivalents, and the explosive equivalents were set to 5kg, 10kg, 15kg, 20kg and 30kg respectively; the blasting vibration PPV measurement points were arranged within a range of 6m to 26m from the blasting surface, and a PPV monitoring point was arranged every 2m to measure the blasting vibration PPV values ​​of different monitoring points under different explosive equivalents.

[0214] (1.3) Repeat the above steps (1.1) to (1.2) to collect on-site blasting vibration PPV data under different geological conditions (different lithology and joint parameters).

[0215] Step 2: Based on the collected PPV monitoring data of tunnel blasting construction sites under different geological conditions (different lithology and joint parameters), a PPV intelligent prediction model for blasting vibration of jointed rock mass is constructed. The specific steps are as follows:

[0216] (2.1) Collect sample data preprocessing work, such as Figure 5 As shown in step 2. Three types of rock conditions, granite, tuff, and sandstone, were collected. Each rock condition contained five types of explosive equivalents. At the same time, considering the different joint numbers and monitoring point layout positions within 50m of the blasting surface, a total of 156 sets of on-site blasting vibration PPV measurement data were collected. By adopting the training set random sampling method based on working condition grouping, the extracted training set data contains different geological working conditions and blasting parameter combinations; the training set accounts for 80% of the total data, and the test set accounts for 20%.

[0217] (2.2) Based on the training set data, a gradient boosting blasting vibration PPV regression prediction model is constructed, such as Figure 5As shown in step 3. The parameters and measurement results collected by the on-site blasting vibration test are used to train the LightGBM gradient boosting model. First, the input features are discretized to obtain histogram bins. Taking the uniaxial compressive strength as an example, the strength range is divided into 12 bins within the range of 0 to 120 MPa. When performing feature selection, the segmentation point is selected according to the statistical discrete value of the bin where the feature is located. The construction of the decision tree sequence is guided by calculating the residual of the decision tree objective function. The objective function consists of the mean square error and the L2 regularization term:

[0218] ;

[0219] (2.3) LightGBM intelligent prediction model hyperparameter optimization. The random space search method is used to optimize the model's hyperparameters. The hyperparameters mainly include learning rate, number of decision trees, maximum depth of decision trees, number of leaf node samples, and L2 regularization strength parameter. The present invention sets the learning rate range to [0.01, 0.05, 0.1], the number of decision trees to [50, 100, 200, 300], the decision tree depth to [4, 6, 8, 10], the leaf stage sample size to [10, 20, 30, 40], and the L2 regularization strength parameter to [0.1, 0.2, 0.4, 0.6]. The model is trained and evaluated by randomly constructing a combination of parameter sets, and a set of hyperparameter combinations with the smallest prediction error is selected as the optimal hyperparameters.

[0220] Step 3: Verification and application of the PPV intelligent prediction model for blasting vibration of jointed rock mass, such as Figure 5 as shown in step 4.

[0221] (3.1) The test set data is used to verify the prediction performance of the jointed rock blasting vibration PPV model established in the previous step, and its prediction error and robust performance under different geological conditions are evaluated; the root mean square error index is used as the prediction error of the evaluation model under different explosive equivalents and joint numbers, and the true value and predicted value in the training set and test set are obtained, where Figure 6 This is the prediction result diagram of the true value and predicted value in the training set. Figure 7 The prediction result graph of the true value and predicted value in the test set PPV;

[0222] Figure 6 The horizontal axis of the black dot in the figure represents the true value of the PPV prediction of the training set, and the vertical axis represents the predicted value of the PPV of the training set. The closer the black dot is to the y=x diagonal line, the more accurate the prediction effect is.

[0223] (3.2) Engineering application of PPV intelligent prediction model for blasting vibration of jointed rock mass, including:

[0224] (3.2.1) The prediction model is used to make real-time and rapid predictions of the PPV of the surrounding rock during on-site blasting construction. During the blasting construction of a certain tunnel, the indoor test data of the section and the statistical data of the joint information of the geological survey at the engineering site are used as the input parameters of the blasting vibration PPV intelligent prediction model.

[0225] (3.2.2) Based on engineering experience, it is assumed that the minimum explosive equivalent required for blasting excavation of the current section is 4kg. On this basis, it is assumed that the explosive equivalent of the blasting surface is 6kg, 8kg, 10kg, and 12kg; the assumed amount of explosives is combined with the rock physics and mechanics and geological joint parameters in the previous step to form an intelligent model input parameter vector, and input into the model to predict the PPV at different blast source distances. For example, when the explosive equivalent is 4kg, the rock density of the medium weathered granite section is 2.5g / cm 3 , elastic modulus 3.6 GPa, uniaxial compressive strength 30 MPa, Poisson's ratio 0.27, and the number of joints is 8, the input characteristics of the blasting vibration PPV at a distance of 10 m from the blast source are predicted to be [4, 2.5, 3.6, 30, 0.27, 8, 10].

[0226] (3.2.3) Change the assumed blasting explosive equivalent and repeat the above steps to obtain the PPV of different explosive equivalents and blasting distances, such as Figure 8 The PPV monitoring value and predicted value results at different distances from the blast source under blasting vibration conditions are shown. The results show that when 6kg of explosive equivalent is used, the PPV within 8m from the blast source exceeds the surrounding rock stability criterion, while the PPV beyond 8m from the blast source does not exceed the surrounding rock stability criterion. It is considered that the position within 8m from the blast source is unsafe, and it is recommended that the explosive equivalent should not exceed 6kg at the blasting site. The specific value can be further predicted and analyzed using the method of the present invention.

[0227] Embodiment 3, the embodiment of the present invention provides a jointed rock mass blasting vibration PPV intelligent prediction system, comprising:

[0228] A data collection module is used to collect blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions; the different geology includes different lithology and joint parameters;

[0229] The module for establishing the intelligent prediction model of blasting vibration PPV of jointed rock mass is used to divide the training set and the test set according to the random sampling method of the training set based on the working condition grouping by using the collected blasting vibration PPV test data, and on this basis, establish the PPV prediction model of blasting vibration of jointed rock mass based on the gradient boosting algorithm; wherein, the factors in the grouping list constructed by the training data set include two types of factors: on-site geological working condition factors and blasting construction working condition factors; the on-site geological working condition factors refer to the geological parameter grouping considered in the international rock mass classification standard RMR, including lithology, uniaxial compressive strength, joint spacing, and joint direction parameters; the blasting working condition factors include the amount of explosives;

[0230] Verification and application module, used for verification and application of PPV intelligent prediction model for blasting vibration in jointed rock mass.

[0231] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A jointed rock mass blasting vibration PPV intelligent prediction method, characterized in that: The method includes: S1: Collect blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions; different geology includes different lithology and joint parameters; S2: Establishment of intelligent prediction model of blasting vibration PPV of jointed rock mass. The collected blasting vibration PPV test data are used to divide the training set and the test set according to the random sampling method of the training set based on the working condition grouping. On this basis, a PPV prediction model of blasting vibration of jointed rock mass based on the gradient boosting algorithm is established. S3: Verification and application of PPV intelligent prediction model for blasting vibration of jointed rock mass; In step S2, the jointed rock mass blasting vibration PPV intelligent prediction model is established, including: S2.1: Collect sample data under different working conditions to divide the data into training set and test set. Use the random sampling method of training set based on working condition grouping so that the extracted training set data contains different geological working conditions and blasting parameter combinations; the training set accounts for 80% of the total data and the test set accounts for 20%; S2.2: Establishment of LightGBM gradient lifting jointed rock mass blasting vibration PPV intelligent prediction model; S2.3: Hyperparameter optimization of PPV intelligent prediction model for blasting vibration in jointed rock mass; In step S2.1, a training set random sampling method based on working condition grouping is adopted, including: (1) In the training data set, important features that affect the magnitude of the blasting vibration wave velocity are selected to construct a grouping list. The factors in the grouping list include two types of factors: on-site geological conditions and blasting construction conditions. The on-site geological conditions refer to the geological parameters considered in the international rock mass classification standard RMR, including lithology, uniaxial compressive strength, joint spacing, and joint direction parameters; the blasting conditions include the amount of explosives; (2) Refine the grouping of training set data according to the important features in the selected list; (3) The obtained m n For each group, 80% of the data is randomly selected as the training set and 20% as the test set. (4) Merge all training sets in each group into a sample training set, and merge all test sets in each group into a sample test set; In step (2), the training set data is refined and grouped according to the important features in the selected list, including: (2.1) Group the monitoring data according to the first feature; (2.2) Group each group of data after grouping according to the first feature according to the second factor in the grouping list; (2.3) Repeat the above steps. Each feature in the list has m groups, and there are n features in total, resulting in m to the power of n groups. For the two features of lithology and joint spacing, each feature has two groups, resulting in lithology 1 joint spacing 1, lithology 1 joint spacing 2, lithology 2 joint spacing 1, and lithology 2 joint spacing 2, for a total of four working conditions.

2. The PPV intelligent prediction method for blasting vibration of jointed rock mass according to claim 1 is characterized in that: In step S1, blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions are collected, including: S1.1: For tunnels constructed by drilling and blasting, a monitoring test working face is selected, and core samples are taken along the sides, top and bottom of the tunnel. Physical and mechanical parameters such as rock density, elastic modulus, uniaxial compressive strength and Poisson's ratio are measured by indoor test methods. The number, length, inclination and dip angle of joints within 50m of the blasting excavation face are statistically analyzed. S1.2: Installation of explosives with different blast equivalents and layout of blasting vibration PPV monitoring points; S1.3: At different blasting sites, select different geological conditions and repeat the above steps to collect blasting vibration PPV data under different geological conditions.

3. The method for intelligent prediction of jointed rock blasting vibration PPV according to claim 1 is characterized in that: In step S2.2, the LightGBM gradient boosting jointed rock blasting vibration PPV intelligent prediction model is established, including: using the blasting vibration training set data to train the gradient boosting decision tree model, first discretizing different input features to obtain histogram bins, and using the histogram statistical method to obtain the best leaf split node to construct a decision tree, where the best leaf split node is the split node corresponding to the maximum value of the split gain value; The next level of decision tree is constructed based on the current model residual until the iterative gradient boosting model achieves the minimum prediction error of blasting vibration PPV; the objective function of the gradient boosting tree is composed of the mean square error and the L2 regularization term, and the expression is: Where L(θ) is the loss function that reduces the error between the model and the true value, θ is the model parameter, are the true value and predicted value of the sample, N is the number of samples, λ2 is the L2 regularization strength parameter, is the penalty term of the L2 norm of the complexity of the model difference parameters, m is the total number of model parameters, θ j is the jth model parameter; The method of using the blasting vibration training set data to perform gradient boosting decision tree model training includes: (1) Gradient boosting decision tree uses decision tree as the basic weak learner unit, and constructs a strong learner to predict the burst PPV by accumulating and iteratively optimizing the weak learner; (2) In each round of iteration, a new decision tree is constructed to fit the residuals of the previous round, and the error of the previous decision tree is corrected by calculating the negative gradient direction of the residuals; The method of obtaining the best leaf splitting node by using the histogram statistical method to construct a decision tree includes: (i) Discretize each eigenvalue under the node into k integers and construct a histogram; (ii) Traverse the samples and add the gradient value of each sample to the corresponding histogram interval. Each feature histogram bin interval obtains the corresponding gradient sum; (iii) Perform traversal and segmentation calculations on the bins under each feature to obtain the optimal splitting gain of the feature; compare the splitting gain values ​​of different features to obtain the optimal feature and bin splitting value; The next level of decision tree is constructed based on the current model residual until the iterative gradient boosting model achieves the minimum prediction error of the blasting vibration PPV, including: (a) Make predictions based on the initial decision tree model and calculate the residual of the current initial decision tree model; (b) construct the prediction residual of the next level decision tree, and superimpose the next level decision tree and the current initial decision tree model according to the weight; calculate the negative gradient value of the superimposed strong prediction model to update the superimposed model; (c) Repeat the above steps until the set number of iterations is reached.

4. The method for intelligent prediction of blasting vibration PPV of jointed rock mass according to claim 3 is characterized in that: The blasting vibration training set data is used to train the gradient boosting decision tree model, including: S2.2.1, construct the initial decision tree model; S2.2.11, feature discretization, for a feature x in the training data, its value is discretized into k bins, and the width of each bin is: In the formula, Δ is the bin width value, x max is the maximum value in feature x, x min is the minimum value in feature x; S2.2.12, build a histogram; for each feature, count the samples according to the bins, the current node contains n samples, and calculate the sum of the gradient information in each bin G k , the expression is: In the formula, is the partial differential of the loss function for each sample in bin k, is the partial differential of each predicted sample in bin k; S2.2.13, select the optimal splitting point of the current feature, and calculate the post-splitting gain for each splitting point between different bins. For a given splitting point s, the left side of the splitting point contains the 1st to the lth bins, a total of m samples, and the right side contains the l+1 to the kth bins, a total of n samples. The expression is: Where, Gain(s) is the splitting gain value at a given splitting point s; Traverse all possible bin splitting points of the current feature and select the splitting point that maximizes the gain for splitting. The optimal splitting point selection formula is as follows: s = arg max (Gain (s)) In the formula, arg max() is the maximum value of the splitting gain of different bin positions selected in the current feature; Select the optimal segmentation points of all features; for different feature parameters, obtain the optimal segmentation point set Gain(s i ), where i represents the number of features; s max =argmax(Gain(s i )) In the formula, s max represents the splitting node corresponding to the maximum splitting gain value of all i features; S2.2.14, split the original data into a left node and a right node according to the optimal split point, and repeat the selection of the optimal split node for the two nodes according to the above step S2.2.13 until the maximum depth specified by the decision tree is reached to complete the construction of the initial decision tree; S2.2.2, residual calculation and model updating; Based on the decision tree model obtained in step S2.2.1, the training set data is predicted to obtain the predicted value F(x i ) t-1 , according to the parameters of the extreme model of predicted value and true value, the formula is as follows: r i (t) =y i -F(x i ) t-1 In the formula, r i (t) is the residual between the true value in the t-th iteration and the predicted value in the t-1th iteration; Train a new weak learner to fit the residual and obtain the decision tree model h that can fit the minimum residual error t (x); In the formula, To select the decision tree with the smallest fitting error, h is all decision trees, h(x i ) is based on the input feature x i The trained decision tree; Add the new weak learner model to the current model and obtain the optimal coefficient γ by calculating the negative gradient between the iterative model and the training data. t , the expression is: In the formula, is to select the weight coefficient corresponding to the minimum loss, L() is to calculate the loss between the iterative model and the true value, and F t-1 (x i ) is the decision tree obtained in the t-1th round, γ is the weight coefficient to be selected, h t (x i ) is the decision tree with the smallest fitting residual obtained in the tth round of training; After correcting the residual of the previous model, we get the current model: F t (x)=F t-1 (x)+γ t h t (x) In the formula, F t (x) is the decision tree after correcting the residual in the tth round, F t-1 (x) is the decision tree obtained in the t-1th round, γ t is the weight coefficient of the tth round, h t (x) is the decision tree with the smallest fitting residual obtained in round t; S2.2.3, repeat the previous step S2.2.2 until the predetermined number of iterations is reached; Repeat step S2.2.2 for residual calculation and model update. After T rounds of iterations, the final prediction model is obtained as follows: t Accumulation of (x): In the formula, F t (x) is the integrated decision tree after T rounds of iterations, F0(x) is the initial decision tree before iteration, t is the tth iteration, and T is the total number of iterations.

5. The method for intelligent prediction of blasting vibration PPV of jointed rock mass according to claim 1 is characterized in that: In step S2.3, the hyperparameters of the jointed rock mass blasting vibration PPV intelligent prediction model are optimized, including: S2.3.1, Parameter definition of the hyperparameter combination space to be optimized; Parameter search of the hyperparameter setting in the integrated gradient learning method in continuous or discrete space; S2.3.2, traverse all hyperparameters and perform random sampling of parameters from each parameter definition space according to certain sampling criteria; S2.3.3, perform gradient boosting tree model training and objective function evaluation according to a set of randomly obtained hyperparameters; S2.3.4, repeat the above steps until the specified number of iterations is reached and the stopping condition is obtained to obtain the optimal hyperparameter combination.

6. The method for intelligent prediction of jointed rock blasting vibration PPV according to claim 1 is characterized in that: In step S3, the jointed rock mass blasting vibration PPV intelligent prediction model verification and application includes: S3.

1. Based on the test set data, the performance of the established PPV intelligent prediction model for blasting vibration of jointed rock mass is verified to evaluate its prediction error and model robustness under different geological conditions. The root mean square index is used to compare the error between the actual PPV value and the model prediction value under different explosive equivalents and joint numbers under blasting. S3.

2. Engineering application of PPV intelligent prediction model for blasting vibration of jointed rock mass; S3.2.3, repeat the above steps to obtain the maximum explosive equivalent and PPV values ​​at different distances to ensure tunnel blasting safety; In step S3.2, the engineering application of the jointed rock mass blasting vibration PPV intelligent prediction model includes: S3.2.1, when using the drilling and blasting method to make a real-time and rapid prediction of the blasting vibration PPV of the jointed rock mass during the construction process, it is necessary to obtain the physical and mechanical parameters of the rock mass within 50m of the construction blasting excavation location, including the rock sample density, elastic modulus, uniaxial compressive strength, and Poisson's ratio physical and mechanical parameters; the number, length, inclination, and inclination parameters of the joints within 50m of the on-site blast source are counted; the above parameters are collected as input parameters for analyzing the current blasting vibration safety analysis; S3.2.2, based on engineering experience, assume that the minimum explosive equivalent xkg required for blasting and excavation of the current section is met, with a unit increment of 2kg. Assume that the explosive equivalent of the blasting surface is [x+2, x+4, x+6, x+8, x+10]; input each assumed explosive equivalent and the measured rock physical and mechanical parameters and joint parameters into the jointed rock blasting vibration PPV intelligent prediction model respectively, and obtain the blasting vibration PPV values ​​at different distances from the blasting source under different assumed explosive equivalents; by analyzing the PPV at different distances from the blasting source combined with safety criteria, the safety status under the current explosive equivalent is obtained.

7. A jointed rock mass blasting vibration PPV intelligent prediction system, characterized in that: The system implements the PPV intelligent prediction method for blasting vibration of jointed rock mass as claimed in any one of claims 1 to 6, and the system comprises: The data collection module is used to collect blasting vibration PPV test data under different geological conditions, different blasting explosive amounts, and different monitoring positions; different geology includes different lithology and joint parameters; The module for establishing the intelligent prediction model of blasting vibration PPV of jointed rock mass is used to use the collected blasting vibration PPV test data to divide the training set and the test set according to the training set random sampling method based on the working condition grouping, and on this basis, establish the blasting vibration PPV prediction model of jointed rock mass based on the gradient boosting algorithm; Verification and application module, used for verification and application of PPV intelligent prediction model for blasting vibration in jointed rock mass.

Images