A Method for Predicting Blasting Effects in Open-Pit Mines by Integrating Feature Filtering and XGBoost

By integrating feature filtering and XGBoost, a physical coupling derived parameter is constructed based on the blasting physics mechanism, and an asymmetric misjudgment cost matrix is ​​built. This solves the shortcomings of existing technologies in predicting blasting effects in open-pit mines, and achieves high-precision and safe prediction of blasting effects, providing a basis for blasting design optimization and safety management.

CN123116014APending Publication Date: 2026-11-17FUJIAN PROVINCIAL INSTITUTE OF PRODUCTION SAFETY SCIENCE (FUJIAN PROVINCIAL OCCUPATIONAL HEALTH TESTING & INSPECTION CENTER)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611352045.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for predicting blasting effects in open-pit mines have shortcomings in feature construction, feature selection, characterization of misjudgment costs, and model training and evaluation. These shortcomings make it difficult to meet the actual needs of blasting design optimization and safety management, resulting in insufficient prediction accuracy and safety.

Method used

We employ a method that combines feature selection with XGBoost. By constructing physical coupling derived parameters based on the blasting physics mechanism, we select features using a fusion algorithm of mutual information and random forest based on physical coupling perception, construct an asymmetric misjudgment cost matrix, and train a modified XGBoost model to achieve accurate prediction of blasting effects.

Benefits of technology

It improves the accuracy and safety of blasting effect prediction, provides quantitative decision-making basis before blasting, reduces safety hazards, and enhances the model's generalization ability and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116014A_ABST
    Figure CN123116014A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of mining blasting engineering and artificial intelligence data processing technology, and particularly to a method for predicting the blasting effect in open-pit mines by integrating feature selection and XGBoost. The method involves acquiring historical blasting samples, extracting original working condition parameters, constructing physically coupled derived parameters based on the blasting physics mechanism, and forming an initial feature set. A fusion algorithm of physically coupled mutual information and random forest is used, combined with the feature coupling correlation matrix to calculate the physical coupling compensation score. Redundancy is adaptively removed based on the Pearson correlation coefficient and the coupling correlation matrix to obtain the optimal feature subset. Blasting effect levels are classified according to multi-dimensional quantitative scores, and an asymmetric misjudgment cost matrix is ​​constructed. A modified XGBoost model with an asymmetric cost-sensitive loss function is constructed, and after hierarchical cross-validation training, the predicted blasting effect level is output. This invention balances the effect of parameter physical coupling with the cost of safety misjudgment, improving prediction accuracy and the ability to identify substandard blasting.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of mining blasting engineering and artificial intelligence data processing technology, specifically a method for predicting the blasting effect in open-pit mines that integrates feature screening and XGBoost. Background Technology

[0002] Open-pit mine blasting is a crucial process in ore extraction and rock stripping, and its effectiveness directly impacts subsequent loading, transportation efficiency, and production safety. The quality of blasting is influenced by a combination of factors, including bench height, borehole parameters, charge structure, and explosive ratio, exhibiting strong nonlinearity and uncertainty. For a long time, the effectiveness of open-pit mine blasting has relied primarily on the experience of on-site technicians for post-blast evaluation. This has made it difficult to quantitatively predict the blasting effect of the design scheme before implementation, thus failing to provide a basis for optimizing blasting parameters and ensuring safety control.

[0003] With the widespread application of machine learning technology in mining engineering, some existing technologies have emerged that use historical blasting data to train predictive models to predict blasting effects. However, these methods still have the following shortcomings in practical applications: Existing methods often directly use raw working condition parameters such as step height, number of holes, and explosive dosage as model inputs, without characterizing the implicit physical coupling relationship between parameters from the perspective of blasting rock dynamics and charge energy distribution mechanism. The input features are not good at representing the blasting effect, which limits the further improvement of prediction accuracy. Existing feature selection methods mostly rely on single statistical indicators such as mutual information or feature importance to determine feature selection. When removing redundancy, they simply eliminate highly correlated features based on the statistical correlation between features, without considering the objective physical synergy between the bursting parameters. This can easily lead to the misjudgment and elimination of features with strong physical synergy that should be retained as statistical redundancy, resulting in the loss of key information and affecting the rationality of feature subsets. The cost of misjudging the blasting effect level is significantly asymmetric. Misjudging a substandard blasting with safety hazards as good or excellent has far more serious consequences than underestimating a blasting with excellent effects as good. Existing prediction models generally use symmetric loss functions, treating misjudgments in different directions equally. Furthermore, under class imbalance conditions where the number of substandard samples is small, the model tends to converge towards the majority class, resulting in weak identification ability for substandard blasting and the risk of missing safety hazards. Existing model training often uses ordinary random partitioning and is evaluated with a single accuracy metric. This makes it difficult to ensure a reasonable distribution of each blasting effect level in the training and validation sets, and also makes it difficult to balance the model's balanced prediction performance across levels and cross-subset stability, thus limiting the generalization ability and field usability of the prediction model.

[0004] In summary, existing methods for predicting blasting effects in open-pit mines still have significant shortcomings in feature construction, feature selection, characterization of misjudgment costs, and model training and evaluation, making it difficult to meet the actual needs of blasting design optimization and safety management in open-pit mines.

[0005] Therefore, a method for predicting blasting effects in open-pit mines that integrates feature selection and XGBoost is proposed to address the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting blasting effects in open-pit mines that integrates feature filtering and XGBoost, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting blasting effects in open-pit mines by integrating feature filtering and XGBoost includes the following steps: S1: Obtain historical sample data of blasting operations in open-pit non-coal mines, extract original working condition parameters, and construct physical coupling derived parameters by feature engineering based on the blasting physics mechanism. Combine the original working condition parameters and physical coupling derived parameters to form an initial feature set. S2: The initial feature set is filtered using a physically coupled mutual information-random forest fusion algorithm to obtain the optimal feature subset. The physically coupled mutual information-random forest fusion algorithm includes: calculating the mutual information ranking and random forest feature importance ranking of each feature in the initial feature set; constructing a feature coupling correlation matrix based on prior knowledge of blasting engineering and calculating the physical coupling compensation score of each feature; calculating the comprehensive priority score of each feature using a weighted model that combines the inverse ranking and physical coupling compensation; and performing adaptive redundancy removal based on the Pearson correlation coefficient and the feature coupling correlation matrix to obtain the optimal feature subset. S3: Obtain multi-dimensional quantitative scores of historical blasting operations, classify blasting effect levels as target labels, construct a training dataset by combining the optimal feature subset, and construct an asymmetric misjudgment cost matrix between different blasting effect levels based on the safety and economic constraints of blasting engineering. S4: Construct a modified XGBoost prediction model. The modified XGBoost prediction model is equipped with an asymmetric cost-sensitive loss function that combines the asymmetric misjudgment cost matrix. The modified XGBoost prediction model is trained using the training dataset using the hierarchical cross-validation method to obtain the trained prediction model for blasting effects in open-pit non-coal mines. S5: Obtain the target feature data of the blasting operation to be predicted in the open-pit non-coal mine, input it into the trained open-pit non-coal mine blasting effect prediction model, and output the corresponding blasting effect prediction level.

[0008] As can be seen from the technical solution provided by the present invention above, the beneficial effects of the open-pit mine blasting effect prediction method that integrates feature screening and XGBoost provided by the present invention are: This invention performs feature engineering on the original working condition parameters based on the blasting physics mechanism, and constructs physical coupling derived parameters covering multiple dimensions such as geometry and structure, charge and energy, explosive composition and ratio, and design and process control. Together with the original working condition parameters, these parameters form an initial feature set, which characterizes the implicit coupling relationship between blasting parameters from the perspective of physical mechanism. This makes up for the lack of characterization ability when only the original working condition parameters are used, and provides a feature basis with sufficient information for accurately predicting blasting effects. This invention employs a fusion screening method combining physically coupled mutual information and random forest. It integrates mutual information ranking, random forest feature importance ranking, and physical coupling compensation score obtained based on the feature coupling correlation matrix to calculate a comprehensive priority. This ensures that feature importance evaluation has both statistical basis and prior physical support, resulting in a more objective and reasonable evaluation. In the redundancy removal stage, it distinguishes between physically synergistic features and purely statistically redundant features based on the Pearson correlation coefficient and the feature coupling correlation matrix. Features with physical synergistic effects are packaged into coupled feature groups and retained together, while purely statistically redundant features are removed. This reduces feature dimensions and model complexity while avoiding the accidental deletion of key coupling information, ensuring that the optimal feature subset is both concise and physically reasonable. This invention constructs an asymmetric misjudgment cost matrix based on the safety and economic constraints of blasting engineering. It assigns the highest cost to the misjudgment of overestimation of the level of unqualified blasting. Furthermore, it introduces a cost-sensitive factor and a focus penalty modulation factor into the loss function of the XGBoost model. This enables the model to actively optimize decision tree splitting in the direction of reducing the cost of high-risk misjudgment during training. At the same time, it reduces the weight of easily classified samples and increases the weight of difficult-to-classify boundary samples. This enhances the ability to identify unqualified blasting under class imbalance conditions, reduces the risk of missing safety hazards, and meets the actual requirement of prioritizing safety in blasting operations. This invention employs a hierarchical cross-validation method to train the model, ensuring that the proportion of each blasting effect level in each subset is consistent with the overall proportion. The macro-average F1 value is used as the main evaluation index, and the model parameters are selected by combining the accuracy and standard deviation. This approach balances the model's balanced predictive performance across all levels and its stability across subsets, thereby improving the generalization ability and reliability of the prediction model. Before blasting operations are carried out, this invention only requires collecting target working condition parameters based on the blasting design plan to output the predicted blasting effect level and the probability distribution of each level. This realizes the transformation of blasting effect from post-event evaluation to pre-event quantitative prediction. When the prediction is unqualified or at the level boundary, it can prompt on-site personnel to review the blasting parameters. This provides a quantitative decision-making basis for the optimization of blasting design and safety management in open-pit mines and has good engineering application value. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the steps in the method for predicting blasting effects in open-pit mines that integrates feature screening and XGBoost, according to the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0011] To better understand the above technical solutions, the following will provide a detailed description of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0012] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for predicting the blasting effect in open-pit mines by integrating feature filtering and XGBoost, including the following steps: S1: Obtain historical sample data of blasting operations in open-pit non-coal mines, extract original working condition parameters, and construct physical coupling derived parameters by feature engineering based on the blasting physics mechanism. Combine the original working condition parameters and physical coupling derived parameters to form an initial feature set. S2: The initial feature set is filtered using a physically coupled mutual information-random forest fusion algorithm to obtain the optimal feature subset. The physically coupled mutual information-random forest fusion algorithm includes: calculating the mutual information ranking and random forest feature importance ranking of each feature in the initial feature set; constructing a feature coupling correlation matrix based on prior knowledge of blasting engineering and calculating the physical coupling compensation score of each feature; calculating the comprehensive priority score of each feature using a weighted model that combines the inverse ranking and physical coupling compensation; and performing adaptive redundancy removal based on the Pearson correlation coefficient and the feature coupling correlation matrix to obtain the optimal feature subset. S3: Obtain multi-dimensional quantitative scores of historical blasting operations, classify blasting effect levels as target labels, construct a training dataset by combining the optimal feature subset, and construct an asymmetric misjudgment cost matrix between different blasting effect levels based on the safety and economic constraints of blasting engineering. S4: Construct a modified XGBoost prediction model. The modified XGBoost prediction model is equipped with an asymmetric cost-sensitive loss function that combines the asymmetric misjudgment cost matrix. The modified XGBoost prediction model is trained using the training dataset using the hierarchical cross-validation method to obtain the trained prediction model for blasting effects in open-pit non-coal mines. S5: Obtain the target feature data of the blasting operation to be predicted in the open-pit non-coal mine, input it into the trained open-pit non-coal mine blasting effect prediction model, and output the corresponding blasting effect prediction level.

[0013] In this embodiment, step S1 extracts original working condition parameters from historical sample data of blasting operations in open-pit non-coal mines, and constructs physically coupled derived parameters based on the blasting physics mechanism through feature engineering. The original working condition parameters and physically coupled derived parameters are combined to form an initial feature set, laying the data foundation for subsequent feature selection and prediction model training. The detailed steps are as follows: Step S1-1: Collect historical blasting operation sample data from the blasting design and construction records, blasting acceptance reports, and mine production management system of open-pit non-coal mines. Each sample corresponds to one complete blasting operation. Perform quality checks on the collected samples: remove samples with missing key fields, inconsistent dimensions, or values ​​exceeding the physically reasonable range. For example, step height, average hole depth, and charge length must not be negative, charge length must not exceed the average hole depth, detonator usage and number of holes must be positive integers, and the total charge amount must not exceed the blasting amount. After the quality check is completed, a historical sample dataset is formed, and the total number of samples is recorded for subsequent feature calculation and model training. Step S1-2: For each sample in the historical sample dataset, extract nine categories of original working condition parameters, namely, step height, number of holes, detonator usage, emulsion explosive usage, main explosive type, main explosive usage, blasting volume, average hole depth, and charge length; among them, the main explosive type is a categorical parameter, which is converted into a numerical variable using a categorical coding method; the other eight categories of parameters are all numerical parameters, maintaining their original dimensions; Step S1-3: Based on the dynamics of blasting rock mass, charge structure and energy distribution mechanism, feature engineering is performed on the original working condition parameters to construct four types of physical coupling derived parameters, namely geometric and structural parameters, charge and energy characteristics, explosive composition and proportion, and design and process control. The geometric and structural parameters include relative hole depth, packing length, packing rationality, charge-to-hole depth ratio, and charge-to-step ratio. The calculation formulas are as follows: ,in, For relative hole depth, The height of the step. The average hole depth; ,in, For the filling length, This refers to the length of the propellant charge; ,in, To ensure the rationality of the filling; ,in, The ratio of explosive charge to hole depth; ,in, The ratio of the charge to the step; Charge and energy characteristics include single-hole energy intensity, linear charge density, volume breaking capacity per unit charge, step-to-energy coupling index, and hole depth-to-energy coupling index, calculated as follows: ,in, For single-pore energy intensity, This is the dosage per orifice; ,in, Density of the linear charge; ,in, This refers to the dosage per unit of medicine; ,in, The coupling index between step and unit consumption. This refers to the unit consumption of explosives, which is the ratio of the total charge to the explosive yield. ,in, The coupling index between hole depth and unit consumption; the explosive composition and proportion category includes the proportion of emulsion explosive and the proportion of main explosive, and the calculation formulas are as follows: ,in, The percentage of emulsion explosives, This refers to the dosage of emulsion explosives; ,in, The proportion of main explosives, The main explosive usage; design and process control includes detonator utilization rate and single-hole controlled blasting volume, the calculation formulas are as follows: ,in, To improve detonator utilization, For detonator usage, Number of holes; ,in, To control the blasting volume in a single hole; Step S1-4: Combine the nine types of original working condition parameters with the above fourteen types of physical coupling derived parameters sample by sample to form an initial feature set containing twenty-three feature dimensions, and simultaneously establish a mapping description of feature name, feature meaning, calculation formula and value range as input for feature screening in step S2.

[0014] In this embodiment, step S2 uses a physically coupled, mutually informative, random forest fusion algorithm to filter the initial feature set and obtain the optimal feature subset. This algorithm integrates two types of statistical importance evidence—mutual information and random forest feature importance—with prior knowledge of blasting engineering, and performs adaptive redundancy removal based on Pearson correlation coefficient and feature coupling correlation matrix. The detailed steps are as follows: Step S2-1: Taking each feature in the initial feature set as the object and the explosion effect level as the target variable, calculate the mutual information value between the feature and the target variable; for continuous features, first perform equal-frequency bin discretization processing; the mutual information calculation formula is: ,in, For a certain feature variable in the initial feature set, The target variable is the explosion effect level. Discretization of the feature variables To assign rank values ​​to the target variable, For taking values With value The joint probability of simultaneous occurrence and These are the corresponding marginal probabilities; Sort all features from highest to lowest mutual information value. The higher the mutual information value, the lower the rank number. mutual information ranking of features ,in, For the first The mutual information ranking of each feature, the smaller the value, the stronger the mutual information between the feature and the explosion effect level; Step S2-2: Using the initial feature set as input and the explosion effect level as the target variable, train a random forest model. Calculate the random forest feature importance of each feature based on the Gini impurity reduction. The calculation formula is as follows: ,in, For the first The importance of random forest features for each feature. The number of decision trees in the random forest. For the first In the decision tree, the first The set of nodes with each feature serves as a splitting feature. For nodes The decrease in Gini impurity caused by splitting; Sort all features from highest to lowest importance value to obtain the [number]. Ranking of the importance of features in a random forest ,in, For the first Ranking of the importance of each feature in a random forest; the smaller the value, the higher the importance of that feature in the random forest. Step S2-3: Based on the dynamics of blasted rock mass and the stress wave propagation mechanism, determine whether there is a physical synergy between any two features in the initial feature set. If a physical synergy exists, assign the corresponding matrix element as the first correlation value (1); if no physical synergy exists, assign the corresponding matrix element as the second correlation value (0). This constructs a feature coupling correlation matrix, with the diagonal elements taking the second correlation value. For the first feature in the initial feature set... For each feature, the number of other features in the statistical matrix that have a first correlation value with that feature is multiplied by a preset coupling benchmark coefficient to obtain the physical coupling compensation score for that feature. ,in, For the first Physical coupling compensation score for each feature The preset coupling reference coefficients, The total number of features in the initial feature set. For the characteristic coupling correlation matrix, the first Line 1 Column elements; The physical coupling compensation score is normalized and calculated using the following formula: ,in, For the first Physical coupling compensation score after feature normalization This is the maximum value among all feature physical coupling compensation scores; the physical coupling compensation score is used to increase the priority of features with strong physical synergy during fusion weighting. Step S2-4: Calculate the comprehensive priority score for each feature using a weighted model that combines the inverse ranking with physical coupling compensation. The calculation formula is as follows: ,in, For the first The overall priority score of each feature , , These are the fusion weight coefficients of the reciprocal mutual information ranking, the reciprocal random forest importance ranking, and the normalized physical coupling compensation score, respectively, satisfying... The default configuration is , , It can be adjusted based on practical experience in blasting engineering. All features are sorted from high to low based on their comprehensive priority scores. If there are identical scores, the feature with higher mutual information ranking is given priority, forming a feature priority sequence for adaptive redundancy removal. Step S2-5: Add features to the candidate feature subset in descending order of comprehensive priority score; when adding the current feature, calculate the Pearson correlation coefficient between the current feature and existing features in the candidate feature subset; let the two features to be compared be the first feature in the first position. The values ​​on each historical sample are respectively and After removing the mean, they are respectively denoted as and ,in, and Let be the means of the two features across all historical samples, then the Pearson correlation coefficient is calculated as follows: ,in, The Pearson correlation coefficient between the two features. This represents the number of historical samples. If the absolute value of the Pearson correlation coefficient is greater than the preset correlation threshold Then, query the feature coupling correlation matrix, where, The preset correlation threshold is 0.85 by default and can be adjusted according to the data distribution characteristics. If the corresponding element of the current feature and the existing feature in the feature coupling correlation matrix is ​​the first correlation value, it is determined that there is a physical synergy between the two, and the current feature and the existing feature are packaged into a coupled feature group and kept together in the candidate feature subset. If the corresponding element of the current feature and the existing feature in the feature coupling correlation matrix is ​​the second correlation value, it is determined that the two are purely statistical redundancy, the current feature is removed, and the existing feature is retained. If the absolute value of the Pearson correlation coefficient is not greater than the correlation threshold, the current feature is directly retained in the candidate feature subset.

[0015] In this embodiment, step S3 obtains multi-dimensional quantitative scores of historical blasting operations, classifies blasting effect levels as target labels, constructs a training dataset by combining the optimal feature subset, and constructs an asymmetric misjudgment cost matrix between different blasting effect levels based on the safety and economic constraints of blasting engineering. This provides a label system and cost basis for the cost-sensitive training of the modified XGBoost prediction model. The detailed steps are as follows: Step S3-1: Based on six core evaluation indicators—misfire, flyrock, perforation, large surface fragments, blast pile concentration, and base—a multi-dimensional comprehensive quantitative score is performed on historical blasting operations. The individual score for each indicator is determined based on on-site measured data and blasting acceptance records, using a 100-point scale. A higher score indicates a better blasting effect. The formula for calculating the comprehensive score is: ,in, This is the overall score. The evaluation index numbers correspond to the six core evaluation indicators in sequence: misfire, flyrock, perforation, large surface fragments, blast pile concentration, and bottom foundation. For the first Individual scores for each indicator For the first The weight coefficients of each indicator satisfy the following: Safety-related indicators such as anti-explosive, flying rock, and perforation are given higher weights than economic indicators, reflecting the principle of prioritizing safety. Step S3-2: Based on the score range of the comprehensive score, the blasting effect is divided into three levels: excellent, good, and unsatisfactory; the level division rule is as follows: when When it is classified as excellent, Time is classified as good, when The time frame is classified as unqualified; the score interval threshold can be adjusted according to the actual management standards of the mine; the grade classification result serves as the target label for each historical sample. Step S3-3: Using the features in the optimal feature subset as the input feature column and the explosion effect level as the target label column, align them one by one according to the sample number to construct the training dataset; statistically distribute the number of samples in the training dataset for the three levels of excellent, good and unqualified. If the number of samples in a certain level is too small, the impact of class imbalance will be suppressed by stratified sampling and cost weighting during training in step S4. Step S3-4: Based on the safety and economic constraints of blasting engineering, construct an asymmetric misjudgment cost matrix between different blasting effect levels; let... The true labels in the matrix are And the predicted label is The value of misjudgment at that time, among which, For real labels, For the predicted labels, the values ​​are all excellent, good, or unsatisfactory; The value selection rule is: the cost of misjudging a true label as unqualified but predicted as good or excellent is set as the maximum cost. The cost of misjudging a case where the actual label is good but the prediction is unsatisfactory is set as a moderate cost. Set the minimum cost of misjudging a true label as excellent but a prediction as good. The cost of any other correctly predicted combinations is set to zero; for any other unlisted incorrect combinations, the cost is set between the lowest and medium cost based on the safety and economic constraints of blasting engineering; the three types of cost values ​​satisfy... Typical configuration is , , The matrix as a whole has an asymmetric structure. When the true label of an unqualified sample is underestimated, the cost is much higher than that of other misclassification directions. This forces the model to prioritize avoiding high-risk misclassifications in the safety category.

[0016] The core function of step S4 is to construct a modified XGBoost prediction model, configure an asymmetric cost-sensitive loss function that incorporates the asymmetric misjudgment cost matrix, and train the modified XGBoost prediction model using the training dataset with a hierarchical cross-validation method to obtain a trained prediction model for blasting effects in open-pit non-coal mines. The detailed steps are as follows: Step S4-1: Construct a prediction model based on the XGBoost algorithm. The model uses an ensemble decision tree structure to fit the residual direction between the current prediction and the true label in each round. The basic objective function consists of the training loss and the complexity regularization term, and the calculation formula is as follows: ,in, Based on the objective function, The number of training samples. To train the loss function, For the first The true labels of each training sample For the model to the first Predicted labels for each training sample. The number of decision trees in the ensemble. For the first The complexity of a decision tree is determined by a regularization term. The formula for calculating the complexity regularization term is: ,in, The penalty coefficient for the number of leaf nodes. For the first The number of leaf nodes in a decision tree. The L2 regularization coefficient for the leaf weights. For the first The predicted weights of each leaf node; Step S4-2: Introduce a cost-sensitive factor into the underlying objective function of the XGBoost algorithm. The cost-sensitive factor is determined by the cost of the true label and the predicted label in the asymmetric misclassification cost matrix, and the calculation formula is as follows: ,in, For the first The cost-sensitive factor for each training sample The true label in the asymmetric misclassification cost matrix is And the predicted label is The corresponding value; A focus penalty mechanism is introduced, which calculates a modulation factor based on the degree of deviation between the model's predicted probability of the current training sample and the true label; the formula for calculating the degree of deviation is: ,in, For the first The degree of deviation between the predicted probability of a training sample and the true label. For the model to the first The predicted probability of the true label class of each training sample; then the modulation factor is calculated using the following formula: ,in, For the first Modulation factor of each training sample To focus on the adjustment index, a preset constant is set, with a default setting of 2; the modulation factor uses the degree of bias to reduce the weight of easily classified samples and increase the weight of difficult-to-classify boundary samples. The original formula for calculating cross-entropy loss is: ,in, For the first The original cross-entropy loss for each training sample. For the number of explosion effect level categories, For the first The training sample belongs to the first... The actual pointer value of the class can be either 0 or 1. For the model to the first The training sample belongs to the first... The predicted probability of a class; Multiplying the cost-sensitive factor by the modulation factor and then weighting the original cross-entropy loss, we obtain the asymmetric cost-sensitive loss function, calculated as follows: ,in, For the first The asymmetric cost-sensitive loss for each training sample; substituting this into the basic objective function, we obtain the overall objective function for training the modified XGBoost prediction model: ,in, The overall objective function of the modified XGBoost prediction model is defined by this loss function, which forces the model to optimize decision tree splits during training in a way that reduces the cost of high-risk misjudgments. Step S4-3: Perform gradient boosting training with the goal of minimizing the overall objective function; in each iteration, perform a second-order Taylor expansion of the overall objective function at the current prediction position and fit a new decision tree, calculated as follows: ,in, For the current iteration round, For asymmetric cost-sensitive loss on the first The first-order gradient of the current predicted value of each training sample For the corresponding second-order gradient, For the first The new decision tree in the first round The output of each training sample, For the first The feature vector of the optimal feature subset of each training sample; When splitting a decision tree node, the optimal split point is selected based on the split gain. The formula for calculating the split gain is: ,in, The splitting gain for candidate split points. and These are the sum of the first-order gradients and the sum of the second-order gradients of the left child node after the split, respectively. and These are the sum of the first-order gradients and the sum of the second-order gradients of the right child node after the split, respectively. This is the sum of the first-order gradients of the parent nodes before the split. The sum of the second-order gradients of the parent nodes before the split; the larger the split gain, the more beneficial the split point is to reducing the loss sensitive to asymmetric costs, and the decision tree is grown layer by layer accordingly. Step S4-4: Divide the training dataset evenly into groups according to the sample proportions of each explosion effect level. Mutually exclusive subsets, wherein, The cross-validation fold is set to 5 by default, ensuring that the proportion of samples in each subset at the three levels of excellent, good, and unsatisfactory is consistent with the overall proportion. In each iteration of training, one subset is selected as the validation set, and the remaining subsets are used as the training set. This process is repeated until all subsets have been used as validation sets at least once. In each iteration, the macro-average F1 score is used as the main evaluation metric for model performance, combined with accuracy and the corresponding standard deviation for optimal selection; The formulas for calculating class precision, recall, and F1 score are as follows: ; ; ; in, For the first The number of samples whose class was correctly predicted. For being incorrectly predicted as the first Number of samples in the class For the first The number of samples that were incorrectly predicted as other classes. For the first The accuracy of the class For the first Recall rate of the class For the first The F1 value of the class; The formula for calculating the macro average F1 score is: ,in, This represents the macro average F1 value. The formula for calculating accuracy is: ,in, For accuracy, To verify the number of samples that were correctly predicted in the set, The total number of samples in the validation set; Let the first The macro average F1 value obtained from the fold verification is The mean of the F1 values ​​of each macro is ,remember The formula for calculating the standard deviation of the macro average F1 value is: ,in, The standard deviation of the macro-average F1 score reflects the stability of the model across different subsets; Traverse the preset model parameter combination space and select the model parameter combination with the highest macro-average F1 value and the smallest standard deviation as the final trained open-pit non-coal mine blasting effect prediction model.

[0017] In this embodiment, step S5 obtains the target feature data of the blasting operation in the open-pit non-coal mine to be predicted, inputs it into the trained open-pit non-coal mine blasting effect prediction model, and outputs the corresponding blasting effect prediction level; the detailed steps are as follows: Step S5-1: Collect nine types of original working condition parameters from the blasting design scheme of the blasting operation to be predicted in the open-pit non-coal mine, including bench height, number of holes, amount of detonators, amount of emulsion explosives, type of main explosive, amount of main explosives, blasting volume, average hole depth and charge length, and perform the same quality check as in step S1-1 to ensure data integrity and physical rationality. Step S5-2: According to the calculation formula in Step S1-3, construct all fourteen types of physical coupling derived parameters based on the collected original working condition parameters; according to the optimal feature subset obtained in Step S2, retain the features contained therein, remove the features that were not selected, and align the feature order, category encoding method and numerical precision according to the statistical caliber of the training dataset to form the target feature data. Step S5-3: Input the target feature data into the trained open-pit non-coal mine blasting effect prediction model; the model performs probability normalization on the output logical values ​​of each category to obtain the prediction probability of each category, calculated using the following formula: ,in, The model predicts that this blasting operation belongs to the [number missing] [section missing]. The probability of a class For the model to the first The logical value output by the class; the predicted level of the explosion effect is determined according to the principle of maximum probability, and the calculation formula is as follows: ,in, The predicted level of the explosion effect output by the model can be one of three levels: excellent, good, or unqualified. Step S5-4: Output the corresponding blasting effect prediction level and the prediction probability distribution of the three levels; when the prediction level is unqualified, or when the boundary sample between excellent and good appears, prompt the on-site blasting design and safety management personnel to review the blasting parameters in conjunction with the prediction probability, so as to provide a basis for blasting design optimization and safety control decisions.

[0018] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting blasting effects in open-pit mines by integrating feature filtering and XGBoost, characterized in that: Includes the following steps: S1: Obtain historical sample data of blasting operations in open-pit non-coal mines, extract original working condition parameters, and construct physical coupling derived parameters by feature engineering based on the blasting physics mechanism. Combine the original working condition parameters and physical coupling derived parameters to form an initial feature set. S2: The initial feature set is filtered using a physically coupled mutual information-random forest fusion algorithm to obtain the optimal feature subset; The physically coupled mutual information-random forest fusion algorithm includes: calculating the mutual information ranking and random forest feature importance ranking of each feature in the initial feature set respectively; constructing a feature coupling correlation matrix based on prior knowledge of blasting engineering, and calculating the physical coupling compensation score of each feature; calculating the comprehensive priority score of each feature using a weighted model that combines the inverse ranking and physical coupling compensation; and performing adaptive redundancy removal based on Pearson correlation coefficient and feature coupling correlation matrix to obtain the optimal feature subset. S3: Obtain multi-dimensional quantitative scores of historical blasting operations, classify blasting effect levels as target labels, construct a training dataset by combining the optimal feature subset, and construct an asymmetric misjudgment cost matrix between different blasting effect levels based on the safety and economic constraints of blasting engineering. S4: Construct a modified XGBoost prediction model. The modified XGBoost prediction model is equipped with an asymmetric cost-sensitive loss function that combines the asymmetric misjudgment cost matrix. The modified XGBoost prediction model is trained using the training dataset using the hierarchical cross-validation method to obtain the trained prediction model for blasting effects in open-pit non-coal mines. S5: Obtain the target feature data of the blasting operation to be predicted in the open-pit non-coal mine, input it into the trained open-pit non-coal mine blasting effect prediction model, and output the corresponding blasting effect prediction level.

2. The method for predicting open-pit mine blasting effects by fusing feature screening and XGBoost as described in claim 1, characterized in that: The original operating parameters include step height, number of holes, detonator usage, emulsion explosive usage, main explosive type, main explosive usage, blasting volume, average hole depth, and charge length. The physical coupling-derived parameters include geometric and structural parameters, charge and energy characteristics, explosive composition and proportions, and design and process control parameters. Among them, the geometric and structural parameters include relative hole depth, filling length, filling rationality, charge-to-hole depth ratio, and charge-to-step ratio; the charge and energy characteristics include single-hole energy intensity, linear charge density, unit charge volume, step-to-unit consumption coupling index, and hole depth-to-unit consumption coupling index; the explosive composition and proportioning includes the proportion of emulsion explosive and the proportion of main explosive; and the design and process control includes detonator utilization rate and single-hole controlled blasting volume. The relative hole depth is the ratio of the step height to the average hole depth; the plugging length is the difference between the average hole depth and the charge length; the plugging rationality is the ratio of the plugging length to the average hole depth; the charge to hole depth ratio is the ratio of the charge length to the average hole depth; and the charge to step ratio is the ratio of the charge length to the step height. The single-hole energy intensity is the ratio of the single-hole charge to the step height; the linear charge density is the ratio of the single-hole charge to the charge length; the unit charge volume is the ratio of the blasting volume to the total charge; the step-to-consumption coupling index is the product of the step height and the explosive consumption; and the hole depth-to-consumption coupling index is the product of the average hole depth and the explosive consumption. The proportion of emulsion explosive is the ratio of the amount of emulsion explosive used to the total charge, and the proportion of main explosive is the ratio of the amount of main explosive used to the total charge. The detonator utilization rate is the ratio of the number of detonators used to the number of holes, and the controlled blasting amount per hole is the ratio of the blasting amount to the number of holes.

3. The method for predicting open-pit mine blasting effects by fusing feature screening and XGBoost as described in claim 1, characterized in that: Based on prior knowledge of blasting engineering, a feature coupling correlation matrix is ​​constructed, and the physical coupling compensation score of each feature is calculated, including: Based on the dynamics of blasted rock mass and the stress wave propagation mechanism, it is determined whether there is a physical synergy between any two features in the initial feature set. If there is a physical synergy, the corresponding matrix element is assigned the first correlation value; if there is no physical synergy, the corresponding matrix element is assigned the second correlation value. Thus, a feature coupling correlation matrix is ​​constructed. For any target feature in the initial feature set, traverse the rows or columns corresponding to the target feature in the feature coupling correlation matrix, count the number of other features with the first correlation value, and use the product of the number and the preset coupling benchmark coefficient as the physical coupling compensation score of the target feature. The physical coupling compensation score is used to increase the priority of features with strong physical synergy during fusion weighting.

4. The method for predicting open-pit mine blasting effects by fusing feature screening and XGBoost as described in claim 1, characterized in that: Adaptive redundancy removal is performed based on the Pearson correlation coefficient and the feature coupling correlation matrix to obtain the optimal feature subset, including: Features are added to the candidate feature subset in descending order of their overall priority scores. When adding the current feature, calculate the Pearson correlation coefficient between the current feature and existing features in the candidate feature subset; If the absolute value of the Pearson correlation coefficient is greater than the preset correlation threshold, then query the feature coupling association matrix; If the corresponding element of the current feature and the existing feature in the feature coupling correlation matrix is ​​the first correlation value, it is determined that there is a physical synergy between the two, and the current feature and the existing feature are packaged into a coupled feature group and kept together in the candidate feature subset; If the corresponding element of the current feature and the existing feature in the feature coupling correlation matrix is ​​the second correlation value, then the two are determined to be purely statistical redundancy, the current feature is removed, and the existing feature is retained; After traversing all features, the independent features in the candidate feature subset and the features contained in the coupled feature group are combined to form the optimal feature subset.

5. The method for predicting open-pit mine blasting effects by fusing feature screening and XGBoost as described in claim 1, characterized in that: Obtain multi-dimensional quantitative scores of historical blasting operations, and classify blasting effectiveness levels as target labels, including: Based on six core evaluation indicators—misfire, flyrock, perforation, large surface blocks, blast pile concentration, and base foundation—historical blasting operations are comprehensively quantitatively scored from multiple dimensions to obtain a comprehensive score value. Based on the score range of the comprehensive score, the blasting effect level is divided into three levels: excellent, good, and unqualified, which serve as the target label; Based on the safety and economic constraints of blasting engineering, an asymmetric misjudgment cost matrix is ​​constructed between different blasting effect levels, including: The cost of misjudging a case where the actual label is unqualified but the prediction is good or excellent is set as the highest cost, the cost of misjudging a case where the actual label is good but the prediction is unqualified is set as the medium cost, the cost of misjudging a case where the actual label is excellent but the prediction is good is set as the lowest cost, and the cost of all other correct predictions is set to zero, thus constructing an asymmetric misjudgment cost matrix.

6. The method for predicting open-pit mine blasting effects by fusing feature screening and XGBoost as described in claim 1, characterized in that: The modified XGBoost prediction model is configured with an asymmetric cost-sensitive loss function that incorporates an asymmetric misjudgment cost matrix, including: A cost-sensitive factor is introduced into the underlying objective function of the XGBoost algorithm. The cost-sensitive factor is determined by the cost of the true label and the predicted label in the asymmetric misjudgment cost matrix. A focus penalty mechanism is introduced, which calculates a modulation factor based on the degree of deviation between the model's predicted probability of the current training sample and the true label. The modulation factor is used to reduce the weight of easily classified samples and increase the weight of difficult-to-classify boundary samples. By multiplying the cost-sensitive factor by the modulation factor, the original cross-entropy loss of the XGBoost algorithm is weighted and corrected to obtain an asymmetric cost-sensitive loss function, which forces the model to optimize decision tree splits in the direction of reducing the cost of high-risk misjudgments during training.

7. The method for predicting open-pit mine blasting effects by fusing feature screening and XGBoost as described in claim 1, characterized in that: The method of using hierarchical cross-validation to train the modified XGBoost prediction model using the training dataset includes: The training dataset is evenly divided into multiple mutually exclusive subsets according to the sample proportion of each explosion effect level. In each iteration of training, one subset is selected as the validation set and the remaining subsets are used as the training set. This process is repeated until all subsets have been used as validation sets at least once. In each iteration, the macro-average F1 score is used as the main evaluation index for model performance. Combined with the accuracy and the corresponding standard deviation, the model parameter combination with the highest macro-average F1 score and the smallest standard deviation is selected as the final trained open-pit non-coal mine blasting effect prediction model.