Deposit grade prediction method, device, equipment and medium

Abstract

The application relates to the technical field of mineral exploration, and discloses a deposit grade prediction method, device, equipment and medium. The method comprises the following steps: performing a pretreatment operation on deposit exploration data, and constructing a feature data set according to the processed deposit exploration data; training an XGBoost main regression model based on the feature data set, and constructing a hybrid residual model; generating an intermediate prediction result by adopting a two-level fusion mechanism, and establishing a mapping relationship between a spatial structure, a variation function and a prediction error; adaptively selecting and dynamically adjusting variation function parameters according to the mapping relationship, so as to obtain an optimized variation function; performing Kriging interpolation by using the optimized variation function, so as to obtain a preliminary grade prediction result with spatial continuity; and fusing the intermediate prediction result and the preliminary grade prediction result, so as to obtain a final grade prediction result. By means of the technical scheme, the accuracy and stability of deposit grade prediction are significantly improved.

Description

Technical Field

[0001] This application relates to the field of mineral exploration technology, and in particular to a method, apparatus, equipment and medium for predicting the grade of mineral deposits. Background Technology

[0002] Ore grade prediction is a core and critical task in the mining exploration and development process. Its accuracy directly determines the accuracy of mineral resource reserve assessment, the scientific nature of mining planning, and the economic and safety benefits of mineral resource development. Accurate ore grade prediction provides mining companies with reliable resource distribution data, helping them to rationally allocate mining equipment, plan mining sequences, and control mining costs. It also provides crucial technical support for the rational development and protection of mineral resources, serving as a vital bridge connecting geological exploration and mining production.

[0003] Among existing mineral deposit grade prediction technologies, Kriging interpolation is one of the most widely used traditional spatial interpolation techniques. Based on the principle of spatial autocorrelation of geological variables, this method constructs a variogram to describe the spatial distribution structure of grade data, effectively addressing the spatial continuity problem of grade data. The resulting predictions exhibit good spatial smoothness and geological rationality, playing a crucial role in traditional mineral exploration. However, as mineral resource exploration progresses to deeper and more complex geological conditions, the nonlinear characteristics of mineral deposit grade distribution become increasingly significant. Simultaneously, the geological data accumulated during exploration exhibits multivariable and high-dimensional characteristics, gradually highlighting the limitations of traditional Kriging interpolation. Since Kriging interpolation is essentially a linear interpolation method, its ability to model complex nonlinear geological patterns is weak, making it difficult to accurately fit the distribution characteristics of high-dimensional, strongly nonlinear grade data. This leads to a significant decrease in prediction accuracy under complex geological conditions, failing to meet the needs of modern refined mine development.

[0004] In recent years, with the rapid development of artificial intelligence technology, machine learning methods (such as Extreme Gradient Boosting (XGBoost), deep learning, and random forests) have demonstrated powerful advantages in the field of data modeling, especially in processing complex nonlinear data and mining deep correlation features, where they possess capabilities unmatched by traditional methods. Therefore, some studies have attempted to apply machine learning methods to the field of mineral deposit grade prediction, constructing data-driven prediction models to uncover hidden nonlinear grade distribution patterns in exploration data, thereby improving prediction accuracy to some extent.

[0005] However, machine learning methods, when applied alone to predict ore grade, still have significant technical limitations. Most machine learning models focus on the nonlinear mapping between data features and grade values ​​during the modeling process, easily neglecting the inherent spatial correlation characteristics of ore grade data. This leads to predictions that often lack reasonable spatial continuity, potentially resulting in abrupt grade changes in adjacent areas and spatial distributions that contradict geological patterns—in other words, insufficient spatial consistency. While such predictions may have smaller fitting errors at local sample points, they fail to accurately reflect the actual spatial distribution characteristics of the ore deposit, making them difficult to directly apply to actual production scenarios such as mine planning.

[0006] In summary, while traditional kriging interpolation has the advantage of handling spatial continuity, it is difficult to deal with nonlinear and high-dimensional complex data. While machine learning methods alone can fit nonlinear patterns, they suffer from insufficient spatial consistency. Neither of these methods can simultaneously meet the dual requirements of high prediction accuracy and good spatial consistency for ore grade prediction. Summary of the Invention

[0007] This application provides a method, apparatus, equipment, and medium for predicting ore deposit grade, aiming to solve many technical problems in related technologies, such as the difficulty in dealing with nonlinear and high-dimensional complex data, and the lack of spatial consistency of individual machine learning methods, although they can fit nonlinear laws.

[0008] In a first aspect, embodiments of this application provide a method for predicting the grade of a mineral deposit, the method comprising: Preprocess the mineral deposit exploration data and construct a feature dataset based on the processed mineral deposit exploration data; An XGBoost master regression model is trained based on the aforementioned feature dataset, and a hybrid residual model is constructed. A two-stage fusion mechanism is used to generate intermediate prediction results, while establishing a mapping relationship between spatial structure, variogram, and prediction error. Based on the mapping relationship, the mutation function parameters are adaptively selected and dynamically adjusted to obtain the optimized mutation function; Kriging interpolation is performed using the optimized variogram to obtain preliminary grade prediction results with spatial continuity; The intermediate prediction results and the preliminary grade prediction results are fused to obtain the final grade prediction results.

[0009] In one embodiment, optionally, the preprocessing operations include: outlier removal, missing value imputation, feature construction, dimensionality reduction, standardization, and high-grade sample enhancement. Specifically, the preprocessing operations on the mineral deposit exploration data include: The standard score constraint strategy and the interquartile range constraint strategy are used to remove outliers from the mineral deposit exploration data and fill in missing values ​​to obtain the preliminary processed mineral deposit exploration data. Based on the pre-processed mineral deposit exploration data, spatial geometric features with normalized radial distance and depth are constructed, and combined with the inherent features of the mineral deposit to form an original feature set. Principal component analysis is used to reduce the dimensionality of the original feature set to obtain the dimensionality-reduced feature data. The dimensionality-reduced feature data is transformed into a preset scale by standard score standardization to form a standard dataset. A small perturbation conforming to a Gaussian distribution is applied to the high-quality samples in the standard dataset to enhance the high-quality samples, thereby obtaining the feature dataset.

[0010] In one embodiment, optionally, training an XGBoost master regression model based on the feature dataset includes: The feature dataset is divided into a training set, a validation set, and a test set according to a preset ratio; The hyperparameters of the XGBoost principal regression model are optimized using the Bayesian optimization algorithm with the root mean square error of the five-fold cross-validation as the objective function. The XGBoost master regression model is trained based on optimized hyperparameters, and the nonlinear relationship between the ore deposit grade is fitted by minimizing the loss function. Cross-validation was used to evaluate the XGBoost principal regression model.

[0011] In one embodiment, optionally, the construction process of the hybrid residual model includes: Temporary kriging interpolation is performed using the initial default mutation function; Based on the three-dimensional spatial coordinates of all sample points, and utilizing the spatial continuity modeling capability of Kriging interpolation, a preliminary estimate of the overall grade distribution of the ore deposit is made to obtain a temporary predicted value for the corresponding position of each sample point. The sample points are discrete data points containing three-dimensional spatial coordinates and corresponding grade values, which are formed by associating ore samples collected through borehole exploration. Based on the provisional predicted value, the difference between the actual grade value and the provisional predicted value of each sample point is calculated to obtain the provisional Kriging interpolation prediction residual, and nonlinear geological laws that cannot be modeled by provisional Kriging interpolation are extracted. The K-means algorithm is used to cluster the sample points according to their spatial location, dividing them into K spatially similar regions. For each region cluster, the feature data of the sample points within the cluster and the corresponding temporary kriging interpolation prediction residuals are separately divided to construct the target training dataset for that region cluster. Based on the target training dataset for each regional cluster, independent local XGBoost master regression models are trained to accurately fit the nonlinear geological patterns of that regional cluster. The residual prediction results output by the local XGBoost master regression model are superimposed with the temporary Kriging interpolation prediction residuals to obtain the prediction results of the hybrid residual model.

[0012] In one embodiment, optionally, the two-stage fusion mechanism includes a first-stage ridge regression linear fusion and a second-stage meta-limit gradient boosting nonlinear fusion. The generation of intermediate prediction results using the two-stage fusion mechanism includes: Using the first prediction result of the XGBoost principal regression model and the second prediction result of the hybrid residual model as input, a base model prediction matrix is ​​constructed, and the optimal weights are solved by minimizing the objective function. At the same time, non-negativity constraints and normalization constraints are applied to obtain the linear fusion result. Using the linear fusion result, the first prediction result, and the second prediction result as inputs, an XGBoost principal regression model is trained to learn the nonlinear complementary relationship between different models and generate the intermediate prediction result. The establishment of the mapping relationship between spatial structure, variogram, and prediction error includes: Extract the spatial structure features of the ore deposit to form a spatial structure feature vector, wherein the spatial structure features include the spatial distribution features of sample points, geological trend features, and grade distribution features; Generate candidate sets of variogram parameter combinations for spherical, exponential, and Gaussian models; Calculate the prediction error for each parameter combination and construct a training dataset for mapping relationships.

[0013] In one embodiment, optionally, the method further includes: The final grade prediction result is evaluated using multi-dimensional evaluation indicators, and a decision is made based on the evaluation results as to whether to retrain the model. The multi-dimensional evaluation indicators include: coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error.

[0014] In one embodiment, optionally, the method further includes: Based on incremental learning technology, when new mineral deposit exploration data is acquired, it is preprocessed and then added to the feature dataset to update the feature dataset. Based on the updated feature dataset, the hyperparameters of the XGBoost master regression model, the parameters of the mixed residual model, and the variogram parameters were re-optimized.

[0015] Secondly, embodiments of this application provide a mineral deposit grade prediction device, comprising: The preprocessing module is used to preprocess mineral exploration data and construct a feature dataset based on the processed mineral exploration data. The module is used to train the XGBoost master regression model based on the feature dataset, construct a hybrid residual model, generate intermediate prediction results using a two-stage fusion mechanism, and establish a mapping relationship between spatial structure, variogram and prediction error. An adjustment module is used to adaptively select and dynamically adjust the mutation function parameters according to the mapping relationship to obtain an optimized mutation function; The interpolation module is used to perform Kriging interpolation using the optimized mutation function to obtain preliminary grade prediction results with spatial continuity. The fusion module is used to fuse the intermediate prediction results and the preliminary grade prediction results to obtain the final grade prediction results.

[0016] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for predicting the grade of mineral deposits.

[0017] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described method for predicting the grade of a mineral deposit.

[0018] In the above-described scheme for predicting ore grade, the following steps are taken: Preprocessing of ore exploration data is performed, and a feature dataset is constructed based on the processed data. An XGBoost master regression model is trained on this feature dataset, and a hybrid residual model is constructed. A two-stage fusion mechanism is used to generate intermediate prediction results, while simultaneously establishing a mapping relationship between spatial structure, variogram, and prediction error. Based on this mapping relationship, the variogram parameters are adaptively selected and dynamically adjusted to obtain an optimized variogram. Kriging interpolation is performed using the optimized variogram to obtain a preliminary grade prediction result with spatial continuity. The intermediate prediction results and the preliminary grade prediction results are then fused to obtain the final grade prediction result. The above-mentioned technical solution of the present invention combines the XGBoost principal regression model with the hybrid residual model and adopts a two-stage fusion mechanism to generate intermediate prediction results. At the same time, it establishes a mapping relationship between spatial structure, variogram, and prediction error to achieve adaptive optimization of the variogram. Finally, it fuses the intermediate prediction results with the preliminary prediction results obtained by optimizing Kriging interpolation. This fully leverages the accurate modeling capability of machine learning for nonlinear complex data and the excellent handling advantage of Kriging interpolation for spatial continuity, effectively making up for the shortcomings of single methods. It not only significantly improves the accuracy and stability of ore grade prediction, but also ensures the spatial consistency and geological rationality of the prediction results. This provides more reliable technical support for mine resource assessment and mining planning, and can effectively promote the improvement of the level of refined mine development. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic flowchart of a mineral deposit grade prediction method according to an embodiment of this application is shown.

[0021] Figure 2 A block diagram of a mineral deposit grade prediction device according to an embodiment of this application is shown.

[0022] Figure 3 A block diagram of a computer device according to one embodiment of this application is shown. Detailed Implementation

[0023] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0024] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0025] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0026] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0027] Please see Figure 1 , Figure 1 A schematic flowchart of a mineral deposit grade prediction method according to an embodiment of this application is shown.

[0028] like Figure 1 As shown, the methods for predicting ore deposit grades include: Step S101: Perform preprocessing operations on the mineral deposit exploration data and construct a feature dataset based on the processed mineral deposit exploration data.

[0029] Mineral deposit exploration data preprocessing: refers to a series of operations such as targeted cleaning, correction, feature derivation and standardization of raw data collected during drilling exploration, geological surveying, etc., with the aim of improving data quality and providing reliable input for subsequent model training; Feature dataset construction: based on the preprocessed data, integrate the original features and derived features to form a structured dataset that is complete, effective and adaptable, to support subsequent machine learning model training and prediction.

[0030] In one embodiment, optionally, the preprocessing operations include: outlier removal, missing value imputation, feature construction, dimensionality reduction, standardization, and high-grade sample enhancement. Specifically, the preprocessing operations on the mineral deposit exploration data include: The standard score constraint strategy and the interquartile range constraint strategy are used to remove outliers from the mineral deposit exploration data and fill in missing values ​​to obtain the preliminary processed mineral deposit exploration data. A dual screening mechanism is employed, combining a standard score constraint strategy and a quartile range constraint strategy. The standard score constraint is determined by calculating the multiple of the standard deviation of the sample from the mean (|X|). i -μ x | / σ xThe system removes outlier data by eliminating values ​​exceeding a preset threshold. The interquartile range constraint, based on the data's quartiles (Q1 being the first quartile and Q3 the third quartile), removes extreme values ​​falling outside the range [Q1 - 1.5IQR, Q3 + 1.5IQR] (IQR = Q3 - Q1 being the interquartile range). For missing values, the appropriate method is selected based on the data type: continuous grade data is imputed using the mean / median, while discrete geological attribute data is imputed using the mode, ensuring data integrity.

[0031] Based on the pre-processed mineral deposit exploration data, spatial geometric features with normalized radial distance and depth are constructed, and combined with the inherent features of the mineral deposit to form an original feature set. Based on the pre-processed exploration data, on the one hand, spatial geometric features (planar radial distance) are constructed. Used to characterize the planar position of a sample point relative to the center of the deposit; depth normalization On the one hand, samples at different depths are unified to the [0,1] scale to eliminate the influence of depth span; on the other hand, the inherent characteristics of the deposit (such as ore layer depth, geological lithology coding, borehole number, ore layer thickness, etc.) are integrated to finally form the original feature set.

[0032] Principal component analysis is used to reduce the dimensionality of the original feature set to obtain the dimensionality-reduced feature data.

[0033] Principal Component Analysis (PCA) is used to reduce the dimensionality of the original feature set. By calculating the covariance matrix of the features, the top k principal components with the largest variance contribution are extracted (usually k≤5, which can be adjusted according to the variance explanation rate), redundant features are removed, multicollinearity is eliminated, and the computational complexity of subsequent model training is reduced.

[0034] The dimensionality-reduced feature data is transformed into a preset scale by standardizing with standard scores to form a standard dataset.

[0035] Standardization using Z-scores transforms the dimensionality-reduced feature data into a uniform, pre-defined scale. The calculation formula is as follows: (where μ) j Let σ be the mean of the j-th feature. j The standard deviation is set to 0, and the mean and variance of all features are set to 1. This avoids excessive bias of the model towards features with larger values ​​due to differences in feature units (such as depth in meters and quality in percentages).

[0036] A small perturbation conforming to a Gaussian distribution is applied to the high-quality samples in the standard dataset to enhance the high-quality samples, thereby obtaining the feature dataset.

[0037] Apply a small Gaussian perturbation to the high-quality samples in the standardized dataset (typically samples with quality values ​​above the industry threshold or the 90th quantile of the dataset). Specifically... (where Z is the original high-grade value, ε~N(0, σ) Z 0.1), σ Z (where the standard deviation of the grade value is used). By introducing small random perturbations, the number and diversity of samples are increased while retaining the core characteristics of high-grade samples, thus alleviating the underfitting problem of the model caused by the scarcity of high-grade samples.

[0038] All features processed through the above 5 steps are integrated with the corresponding taste labels to form the final feature dataset. This dataset needs to be divided into a training set, a validation set, and a test set (e.g., a ratio of 7:1:2) for model training, hyperparameter optimization, and performance evaluation, respectively.

[0039] Through the above technical solutions, the dual outlier removal mechanism, compared to a single method, can more accurately identify and remove extreme outlier data caused by geological measurement errors, instrument malfunctions, etc., while reasonably filling in missing values ​​to avoid interference from missing data on model training and ensure the reliability of input data. The construction of spatial geometric features enables the model to capture the spatial distribution patterns of ore deposit grades. Combined with inherent geological features, it achieves multi-dimensional representation of geological attributes and spatial location, improving the model's adaptability to complex geological conditions. PCA dimensionality reduction eliminates multicollinearity between features, reducing the computational load of model training and shortening training time. Standardization ensures the stability of model training and avoids training bias caused by differences in units. High-grade sample augmentation effectively solves the problem of high-grade sample scarcity, enabling the model to fully learn the distribution patterns of high-grade areas and improve the prediction accuracy of high-grade ore deposits. High-grade areas directly affect the economic benefits of mining, making this step of significant value to actual production. The final constructed feature dataset balances completeness, simplicity, and relevance, accurately matching the training needs of subsequent machine learning models such as XGBoost, laying the foundation for high-performance models.

[0040] In practical applications, the parameters of the preprocessing step can be dynamically adjusted according to the type of deposit. For example, for deep deposits, due to the more complex geological conditions, the outlier threshold can be appropriately relaxed (e.g., the standard score constraint threshold can be adjusted from 3.0 to 4.0). For deposits with extremely scarce high-grade samples, the intensity of Gaussian perturbation can be appropriately increased, or other sample augmentation algorithms such as SMOTE can be combined to further enhance sample diversity. In addition, the feature construction stage can add derived features (such as ore dip angle, rock density, etc.) according to the specific geological characteristics of the deposit to further enhance the characterization ability of the features.

[0041] The above technical solution will be described in detail below with a specific embodiment.

[0042] First, the borehole profile data is discretized into continuous sampling points. For the i-th borehole, its depth interval is... Corresponding spatial coordinates ( ), grade This constitutes the set of sampling points: (1) in,( () represents the drilling plane position. For depth, Let TFe be the depth at which the borehole samples are located, and D be the set of all discrete borehole samples. The number of sampling points for the i-th borehole.

[0043] This formula describes the smallest unit of 3D modeling, namely the spatial sampling point along the depth of the borehole. By discretizing the borehole section data into point samples, continuous 3D modeling becomes possible, thus making up for the problem of large-scale and sparse data in drilling.

[0044] To ensure data quality, variables Apply dual constraints of Z-score and IQR:

[0045] The above formula is for outlier removal, a common statistical strategy used to prevent outliers from biasing the fit of the variogram and machine learning. Z-score is used to filter extreme values ​​that are too far from the mean, and IQR is used to provide a safety margin. To address the issue of erroneous deletions caused by skewed distributions in the interquartile range, the intervention of both methods avoids bias in fitting the variogram. For the i-th sample of any variable, Let X be the mean. Let X be the standard deviation of X, and Q1 and Q3 be the first and third quartiles, respectively.

[0046] Then, spatial geometric features are constructed: (3) Where r represents the radial distance in the plane, which helps to reflect the overall spatial trend of the ore body; For depth normalization, it reflects the nonlinear relationship of vertical variation.

[0047] Formula 3 incorporates geometric and depth information through feature extension to capture geological heterogeneity and spatial anisotropy.

[0048] To compensate for regression bias caused by the scarcity of high-quality samples, a lightweight perturbation is introduced for enhancement: (4) Formula 4 is a random perturbation enhancement formula, which can regress the regression bias caused by the scarcity of high-grade samples and improve the model's sensitivity to extreme values. Since the number of high-grade samples is scarce, the model is prone to underfitting in these areas. Therefore, adding a small amount of Gaussian noise to some extreme value samples can enhance the distribution stability of high-grade samples, improve sample diversity while preserving geological trends, and enhance the learning ability of high-grade areas.

[0049] Z represents the original grade, indicating the enhanced grade. This is Gaussian perturbation noise.

[0050] Next, principal component analysis (PCA) is performed on the feature matrix X: (5) X is the original feature matrix, and W is the PCA principal direction matrix. (Preserving the original feature matrix...) Principal components can eliminate multicollinearity and stabilize subsequent model learning. Z is the principal component matrix, d is the original feature tree, and k is the number of principal components.

[0051] Then, standardization is performed: (6) Formulas 5 and 6 aim to reduce dimensionality and eliminate collinearity, preventing scale bias in tree models and Kriging. Principal component analysis (PCA) is used to extract the most important address structure features, reducing feature variance, stabilizing the variogram fit, and reducing noise. Z-score standardization is used to reduce the influence of dimensions, accelerate the optimization process, ensure consistent feature scale, and avoid high variance dominating the model. For the j-th feature, These are the standardized features.

[0052] Step S102: Train the XGBoost master regression model based on the feature dataset, construct a hybrid residual model, generate intermediate prediction results using a two-stage fusion mechanism, and establish the mapping relationship between spatial structure, variogram and prediction error.

[0053] XGBoost Master Regression Model: The core regression model built on the Extreme Gradient Boosting (XGBoost) algorithm is used to explore the nonlinear mapping relationship between feature data and ore grade, and undertakes the main task of grade prediction. Hybrid residual model: A composite model that integrates temporary kriging interpolation and local XGBoost residual fitting, used to capture local nonlinear geological features not covered by the main regression model and the initial kriging interpolation; Two-level fusion mechanism: Through a two-level strategy of linear fusion and nonlinear fusion, the prediction results of different models are integrated to generate intermediate prediction results that combine stability and accuracy; Spatial structure-variogram-prediction error mapping relationship: a quantitative correspondence model used to describe the spatial distribution characteristics of ore deposits, the correlation between Kriging interpolation variogram parameters and prediction errors, and to provide a basis for subsequent variogram optimization.

[0054] In one embodiment, optionally, training an XGBoost master regression model based on the feature dataset includes: The feature dataset is divided into a training set, a validation set, and a test set according to a preset ratio; The hyperparameters of the XGBoost principal regression model are optimized using the Bayesian optimization algorithm with the root mean square error of the five-fold cross-validation as the objective function. In this step, a Bayesian optimization algorithm is employed, using the root mean square error (RMSE) of five-fold cross-validation as the objective function to optimize the key hyperparameters of the XGBoost model. Core hyperparameters include tree depth (typically 3-10, controlling model complexity), learning rate (typically 0.01-0.3, controlling the contribution weight of each tree), regularization terms (λ, γ, used to prevent overfitting), and the minimum number of samples per leaf node. Bayesian optimization predicts the performance of hyperparameter combinations using a Gaussian process model, finding the optimal hyperparameter combination more efficiently than grid search or random search, thus improving optimization efficiency.

[0055] Specifically, Bayesian optimization of Kriging: (7) in, Location to be estimated Given sample points, is the spatial weight, and n is the number of samples.

[0056] To ensure unbiased prediction, the weights must satisfy the following constraints: (8) The weights are obtained from the Kriging linear system: (9) in, It is the semivariance matrix; Let be the semivariance vector between the point to be estimated and the sample points. These are Lagrange multipliers used to satisfy constraints.

[0057] Equation 7 represents the ordinary kriging method. Equation 8 represents the unbiased constraint, which ensures that the predicted mean matches the true value, fulfilling the unbiased condition of ordinary kriging. Equation 9 represents the kriging linear system. By solving the system of equations, we can obtain the Lagrange multipliers corresponding to the weights and constraints, and then substitute them back into Equation 7 for calculation, thereby achieving spatial interpolation with minimum variance and unbiasedness. (10) Explanation of the three parameters: : nugget represents microscale perturbation, and c is the structural variance. To determine the extent of influence, the attenuation scale of spatial correlation is determined.

[0058] Formula 10 represents three major variogram models, employing the spherical model, exponential model, and Gaussian model. These three models can cover most of the spatial structure of ore bodies, and the optimal model is automatically selected based on BO.

[0059] Then, Bayesian optimization was used to automatically select the optimal mutation function type and parameters, with the RMSE of five-fold cross-validation as the objective function: (11) This formula employs Bayesian hyperparameter optimization, using the RMSE of five-fold cross-validation as the objective to automatically select the optimal spatial variation structure. It treats the variogram type and parameters as optimization variables, evaluating their performance through the average RMSE of K-fold cross-validation. Ultimately, it selects the Kriging model that performs most stably across different spatial subsets, ensuring that subsequent interpolation both closely matches the data and possesses good spatial generalization ability. Here, K represents the number of k-folds. Let m be the number of samples in the k-th fold, and m be the variogram model type. For Kriging parameters, For predicted values, This is the actual value.

[0060] The XGBoost master regression model is trained based on optimized hyperparameters, and the nonlinear relationship between the ore deposit grade is fitted by minimizing the loss function. XGBoost regression model: (12) (13) (14) Each of them It is a CART regression tree, where T is the number of trees, Equation 13 is the loss function, and Equation 14 is the round-by-round improvement formula. It is the learning rate, used to control the update magnitude. For XGBoost predictions, For regularization parameters, Let be the complexity of the tree.

[0061] Formulas 12-14 describe the XGBoost learning mechanism: XGBoost is a linear function-multi-tree stacking-second-order gradient boosting. XGBoost compensates for the non-linear properties that Kriging cannot fit. Formula 12 explains that XGBoost approximates the grade function using the additive ensemble of multiple regression trees. XGBoost approximates the true grade function as the additive superposition of T regression trees, with each tree learning a portion of the structure. All trees combined form the final non-linear mapping f. Formula 13 balances accuracy and generalization ability through the goal of "fitting error and complexity penalty," which is the overall goal that XGBoost aims to minimize. In the fitting error term, the goal is to make the predicted value as close as possible to the true grade; the smaller the error, the better. The second term, model complexity penalty, applies more strongly to complex trees, resulting in a more stable model. This controls model complexity and prevents overfitting on local data. Formula 14 describes the small-step correction boosting process using the residuals of new trees in each round. Essentially, it's "old prediction" + "small correction" = "new prediction." Each tree compensates for the shortcomings of the previous model, and the learning rate is used... The scale is incorporated into the existing model to achieve a gradual approximation of complex nonlinear geological relationships.

[0062] Cross-validation was used to evaluate the XGBoost principal regression model.

[0063] The trained model is evaluated using a cross-validation method (five-fold cross-validation). The model's fit and generalization ability are judged by calculating the RMSE and coefficient of determination (R²) of the validation set. If overfitting is found, the hyperparameters such as the regularization term need to be readjusted until the model performance meets the requirements.

[0064] In one embodiment, optionally, the construction process of the hybrid residual model includes: Temporary kriging interpolation is performed using the initial default mutation function; Based on the three-dimensional spatial coordinates of all sample points, and utilizing the spatial continuity modeling capability of Kriging interpolation, a preliminary estimate of the overall grade distribution of the ore deposit is made to obtain a temporary predicted value for the corresponding position of each sample point. The sample points are discrete data points containing three-dimensional spatial coordinates and corresponding grade values, which are formed by associating ore samples collected through borehole exploration. Based on the provisional predicted value, the difference between the actual grade value and the provisional predicted value of each sample point is calculated to obtain the provisional Kriging interpolation prediction residual, and nonlinear geological laws that cannot be modeled by provisional Kriging interpolation are extracted. The K-means algorithm is used to cluster the sample points according to their spatial location, dividing them into K spatially similar regions. For each region cluster, the feature data of the sample points within the cluster and the corresponding temporary kriging interpolation prediction residuals are separately divided to construct the target training dataset for that region cluster. Based on the target training dataset for each regional cluster, independent local XGBoost master regression models are trained to accurately fit the nonlinear geological patterns of that regional cluster. The residual prediction results output by the local XGBoost master regression model are superimposed with the temporary Kriging interpolation prediction residuals to obtain the prediction results of the hybrid residual model.

[0065] In this scheme, the Kriging residual is calculated first: (15) K-means divides the sample points into K regions based on their spatial location: (16) A separate local XGBoost model fk is trained within each cluster to fit the residuals: (17) The final Hybrid prediction is: (18) in, For the k-th spatial cluster, The coordinates of the cluster center, This is the k-th XGB model.

[0066] Equations 15-18 describe spatial residual learning. First, Kriging is used to establish a spatial baseline to ensure smoothness. Then, residual K-means is used for spatial clustering to maintain geological contiguity. Next, local XGBoost is used to fit nonlinear differences. Finally, residual addition is used to compensate for missing Kriging patterns. This local learning approach avoids overfitting and spatial mismatch. Equation 15 calculates the Kriging residual for each sampling point to explicitly extract complex nonlinearities or mineralization patterns that Kriging cannot model. Equation 16 uses K-means to cluster the sample point coordinates, dividing them into K spatially similar clusters, ensuring geological continuity and isomorphism in residual modeling. Equation 17 trains an independent XGBoost model within each spatial cluster to predict residuals, more accurately compensating for interpolation errors. Equation 18 shows that the final predicted value of the rockfall location is composed of the spatial structure-controlled Kriging and the residual model fitted by local nonlinearity.

[0067] In one embodiment, optionally, the two-stage fusion mechanism includes a first-stage ridge regression linear fusion and a second-stage meta-limit gradient boosting nonlinear fusion. The generation of intermediate prediction results using the two-stage fusion mechanism includes: Using the first prediction result of the XGBoost principal regression model and the second prediction result of the hybrid residual model as input, a base model prediction matrix is ​​constructed, and the optimal weights are solved by minimizing the objective function. At the same time, non-negativity constraints and normalization constraints are applied to obtain the linear fusion result. Using the linear fusion result, the first prediction result, and the second prediction result as inputs, an XGBoost principal regression model is trained to learn the nonlinear complementary relationship between different models and generate the intermediate prediction result.

[0068] Level 1: Ridge regression linear fusion. Based on the first prediction result of the XGBoost principal regression model ( ) and the second prediction result of the mixed residual model ( Using as input, construct the base model prediction matrix P. Solve for the optimal fusion weights by minimizing the objective function with L2 regularization. (Where y is the true grade value, w is the fusion weight, b1 is the bias term, and α is the regularization coefficient). Simultaneously, nonnegativity constraints (w←max(w,0) to ensure the physical meaning of the weights is reasonable) and normalization constraints are applied. (ensuring the sum of weights is 1), ultimately yielding a linear fusion result. .

[0069] Second stage: Meta-XGBoost nonlinear fusion. This combines the linear fusion results from the first stage (…). ), the first prediction result of the XGBoost principal regression model ( ), the second prediction result of the mixed residual model ( Using these as inputs, the Meta-XGBoost model is trained. The core function of this model is to learn the nonlinear complementary relationship between the prediction results of different models, correct the biases that were not captured during the linear fusion process, and finally generate intermediate prediction results.

[0070] The establishment of the mapping relationship between spatial structure, variogram, and prediction error includes: Extract the spatial structure features of the ore deposit to form a spatial structure feature vector, wherein the spatial structure features include the spatial distribution features of sample points, geological trend features, and grade distribution features.

[0071] Based on the preprocessed feature dataset, core features characterizing the spatial structure of the ore deposit are extracted from the sample points to form a spatial structure feature vector S. These core features include: ① Spatial distribution characteristics (planar density, depth-direction density, and cluster compactness of the sample points, obtained through spatial density calculation and cluster profile coefficient evaluation); ② Geological trend characteristics (quantitative indicators of the strike and dip angle of the ore layer, obtained by fitting the borehole location coordinates to the depth, such as strike angle and dip angle values); ③ Grade distribution characteristics (spatial autocorrelation of grade values ​​and spatial clustering range of high-grade areas, obtained through initial calculation of the variogram and spatial hotspot analysis).

[0072] Generate candidate sets of variogram parameter combinations for spherical, exponential, and Gaussian models; For three types of variograms—spherical, exponential, and Gaussian—reasonable parameter candidate intervals are pre-defined based on industry experience and deposit type: microscale perturbation c0∈[0, σ Z / 5](σ Z (Standard deviation of grade value), structural variance c∈[σ] Z / 5,2σ Z The influence range a∈[average spacing of sample points, 3×maximum spacing of sample points]. Multiple sets of parameter combinations are generated through grid sampling (50-100 sets of parameters are generated for each model), forming a candidate set of variogram parameter combinations P={(type m, parameter θ=(c0, c, a))}.

[0073] Calculate the prediction error for each parameter combination and construct a training dataset for mapping relationships.

[0074] For each parameter combination (m, θ) in the candidate set P, perform the following operations: ① Use the variogram of that parameter combination to perform kriging interpolation to obtain the grade prediction value under that parameter combination. P ② Calculate the predicted value P The error index (RMSE using five-fold cross-validation) between the sample point and the true grade value Z is denoted as E (i.e., the prediction error corresponding to the spatial structure and parameter combination); ③ Establish sample pairs of “spatial structure feature vector S + parameter combination (m, θ) → error E”, and summarize all sample pairs to form a mapping relationship training dataset T={(S, m, θ, E)}.

[0075] The XGBoost algorithm is used to train a mapping model M, with (S, m, θ) from the training dataset T as input features (encoding the variogram type m as a numerical feature), and the corresponding error E as the output label. Bayesian optimization is used to adjust the model's hyperparameters, minimizing the deviation between the model's prediction error and the true error, ultimately yielding a stable mapping model M. This model can achieve the function of "inputting any spatial structure feature S and a combination of variogram parameters (m, θ) → outputting the corresponding estimated prediction error E".

[0076] In the above scheme, Bayesian optimization of hyperparameters and regularized loss functions are used to accurately mine complex nonlinear grade distribution patterns. Compared with traditional linear models, it can make fuller use of multi-dimensional feature data and improve the accuracy of basic predictions. The cross-validation evaluation mechanism effectively avoids model overfitting, ensures the model's generalization ability, and can adapt to the grade prediction needs of different types of deposits. Through the hybrid design of "temporary Kriging interpolation + local XGBoost residual fitting", it retains the advantage of Kriging interpolation in handling spatial continuity, and accurately captures nonlinear geological patterns that Kriging cannot cover through local models, making up for the shortcomings of single models. The strategy of spatial clustering to divide regional clusters ensures that each local model focuses only on the residual features of its own region, avoiding fitting bias caused by cross-regional geological differences, and further improving the accuracy of residual prediction. The first-level ridge regression linear fusion, through regularization constraints and weight normalization, ensures the stability of the fusion result and avoids extreme prediction biases of a single model. The second-level Meta-XGBoost nonlinear fusion further explores the complementary relationships between different models, correcting the shortcomings of linear fusion. The resulting intermediate predictions possess both high accuracy and stability, laying the foundation for subsequent optimization of the final prediction results. A quantitative correlation between spatial structure, variogram, and prediction error is established, replacing the traditional Kriging interpolation method that relies on manual experience to select variogram parameters. This provides a scientific and accurate decision-making basis for the subsequent adaptive optimization of the variogram. Simultaneously, this mapping model can quickly predict the errors of different parameter combinations, avoiding the need to repeatedly perform complete interpolation calculations during subsequent optimization processes, significantly improving optimization efficiency.

[0077] Step S103: Based on the mapping relationship, adaptively select and dynamically adjust the mutation function parameters to obtain the optimized mutation function.

[0078] Adaptive selection of variogram parameters: Based on the established spatial structure-variogram-prediction error mapping model, the system automatically searches for and selects the variogram type and parameter combination that minimizes the prediction error, replacing the traditional manual experience-based selection method; Dynamic adjustment: According to the current spatial structure characteristics of the deposit, the system optimizes the key parameters of the variogram (microscale perturbation c0, structural variance c, and influence range a) in real time to ensure that the variogram accurately matches the actual spatial distribution characteristics of the deposit; Optimized variogram: Refers to the variogram (including the optimal type and optimal parameter combination) that best suits the current spatial structure of the deposit and minimizes the prediction error, which is used for subsequent formal kriging interpolation.

[0079] Specifically, the extracted spatial structure feature vector S of the current ore deposit is input into the mapping model M. A heuristic search algorithm (such as genetic algorithm or particle swarm optimization algorithm) is used to traverse the generated candidate set of variogram parameter combinations P. For each parameter combination (m, θ), the corresponding prediction error E is predicted through the mapping model M. The optimal variogram type m is selected with the goal of minimizing the prediction error E. (Choose from spherical, exponential, and Gaussian models) and the corresponding optimal parameter combination θ =(c0 c a If multiple parameter combinations have prediction errors close to the minimum, they can be further verified using the actual interpolation errors on the validation set to ultimately determine the unique optimal combination. The optimal parameter combination (m) θ Substitute the values ​​into the Kriging interpolation algorithm and perform interpolation calculations on the validation set. Validate the actual performance of the variogram by calculating metrics such as RMSE and MAE on the validation set. If the validation error deviates significantly from the prediction error of model M (e.g., the deviation exceeds 10%), make small adjustments within the neighborhood of the optimal parameter combination (e.g., adjust the influence range a by ±5%), and re-validate until the error meets the target, ultimately obtaining the optimized variogram.

[0080] By employing precise prediction and heuristic search through a mapping relationship model, intelligent selection of variogram type and parameters is achieved. This ensures that the optimized variogram accurately matches the spatial structural characteristics of the current deposit, avoiding subjective bias and improving the adaptability of the variogram compared to traditional manual selection methods. The optimized variogram more accurately describes the spatial autocorrelation of deposit grade data, providing a reliable spatial structural basis for subsequent kriging interpolation, effectively reducing interpolation prediction errors and improving the accuracy of spatial prediction. Leveraging the mapping relationship model's ability to quickly predict parameter combination errors avoids the repetitive iterative process of "trying different parameter combinations - performing complete interpolation - calculating errors" found in traditional methods, significantly shortening the variogram optimization time and improving the efficiency of the entire prediction process. This adaptive optimization strategy is independent of specific deposit types and can automatically adjust the variogram based on the spatial structural characteristics of different deposits, enabling subsequent kriging interpolation to adapt to deposits with different geological conditions, enhancing the versatility and applicability of the entire prediction method.

[0081] Step S104: Kriging interpolation is performed using the optimized variogram to obtain preliminary grade prediction results with spatial continuity; Kriging interpolation of the optimized variogram: Substitute the optimal variogram (including type and parameters) obtained in step S103 above into the Kriging interpolation algorithm, and make spatial predictions on the grade distribution of the entire deposit based on the principle of spatial autocorrelation of sample points. Preliminary grade prediction results: These are grade prediction results with good spatial continuity obtained through optimized Kriging interpolation. This result is one of the core components of the final prediction results, focusing on ensuring the rationality of spatial distribution. It needs to be integrated with intermediate prediction results to improve accuracy.

[0082] Specifically, the three-dimensional spatial coordinates (x, y, z) of all sample points and their corresponding true grade values ​​are compiled as the basic data for interpolation. Simultaneously, the predicted area of ​​the deposit is determined (e.g., extending a certain boundary based on the borehole coverage area), and the predicted area is discretized into uniform grid nodes (the grid size is adjusted according to the required prediction accuracy, such as a 10m × 10m × 5m three-dimensional grid). Each grid node represents the location where the grade needs to be predicted. The resulting optimized variogram (with its type and parameter c0 clearly defined) is then used to refine the variogram. c a Substituting the core formula of Kriging interpolation, we calculate the spatial variability (i.e., the spatial correlation strength of grade data) between any two sample points and between a sample point and a prediction node. For each prediction node, based on its spatial coordinates and the spatial coordinates of surrounding sample points, we construct a system of Kriging linear equations. By solving this system of linear equations, we obtain the optimal weight coefficient for each sample point to the prediction node. Based on the obtained optimal weight coefficients, we calculate the grade prediction value for the prediction node. We traverse all grid nodes within the prediction area, repeating steps S103-S104 to obtain the grade prediction value for each node. Finally, we integrate these results to form a preliminary grade prediction result that covers the entire prediction area and has spatial continuity.

[0083] Kriging interpolation is essentially optimal linear unbiased interpolation based on spatial autocorrelation. Combined with an optimized variogram, it can accurately capture the spatial distribution structure of ore grade, resulting in preliminary predictions with good spatial smoothness and geological rationality. This avoids problems such as abrupt grade changes in adjacent areas and discrepancies between spatial distribution and geological laws, conforming to the objective laws of natural ore grade distribution. The optimized variogram can more accurately describe the spatial variability of grade data, making the weighting coefficients obtained from solving the Kriging equations more reasonable, thereby improving the grade prediction accuracy of each prediction node. Compared with traditional Kriging interpolation using default parameters or manually selected parameters, the prediction error is significantly reduced. By discretizing the prediction area into a grid and performing global interpolation, grade distribution data covering the entire ore prediction area is obtained, rather than being limited to sample point locations. This provides comprehensive spatial distribution data for subsequent mine resource assessment and mining planning, meeting the needs of modern refined mine development for comprehensive grade data. The preliminary grade prediction results prioritize spatial continuity and complement the intermediate prediction results obtained in step S102 (which prioritize prediction accuracy). This provides high-quality input for the subsequent fusion of the two, ensuring that the final prediction results can meet the dual requirements of "high accuracy" and "good spatial continuity".

[0084] Step S105: The intermediate prediction results and the preliminary grade prediction results are fused to obtain the final grade prediction results.

[0085] The specific fusion method can employ weighted fusion (weights are dynamically determined based on model performance), with the following steps: Based on the validation set data, calculate the prediction errors (RMSE) between the intermediate prediction results and the preliminary grade prediction results. Let the error of the intermediate prediction results be RMSE. mid The error of the preliminary grade prediction result is RMSE. Kriging The fusion weights are calculated using the "inverse error weighting method," which means the weights w of the intermediate prediction results are... mid =1 / RMSE mid The weight w of the preliminary grade prediction results Kriging=1 / RMSE Kriging Then, the weights are normalized: w mid '=w mid / (w mid +w Kriging ), w Kriging '=w Kriging / (w mid +w Kriging To ensure the weights sum to 1, and that results with smaller errors have larger weights, the fusion is made more reasonable. For each grid node within the prediction area, its intermediate predicted value and preliminary predicted value are weighted and summed according to the above weights to obtain the final grade prediction value for that node. final =w mid '× mid + w Kriging '× Kriging The algorithm iterates through all grid nodes, integrates the final grade prediction values ​​of each node, and forms a final grade prediction result covering the entire prediction area. This result can be output as 3D raster data, contour maps, heat maps, etc., facilitating subsequent visualization analysis and practical applications.

[0086] Optionally, if further improvement in fusion effect is required, a nonlinear fusion strategy can be adopted (such as constructing a simple XGBoost fusion model), using intermediate and preliminary prediction values ​​as input features and the true quality value of the validation set as the label, to train the fusion model, and the model automatically learns the optimal fusion rule to generate the final prediction result.

[0087] By employing a fusion strategy that uses the reciprocal of the error weighting, intermediate predictions with higher accuracy are given greater weight, while preserving the spatial continuity advantage of the initial predictions. The resulting predictions address both the insufficient spatial consistency of single machine learning models and the weak nonlinear fitting ability of traditional Kriging interpolation, achieving the dual goals of "high prediction accuracy" and "good spatial continuity." This fusion strategy integrates the predictions of two different types of models, reducing prediction bias caused by the limitations of a single model and making the final predictions more stable and resilient to interference. For example, when one model has a large prediction bias in a local area, the advantages of the other model can effectively compensate for this bias, improving the reliability of the overall prediction results. The final predictions combine accuracy and spatial rationality, accurately reflecting the actual grade distribution characteristics of the ore deposit. They can be directly used in actual production processes such as calculating mineral resource reserves, dividing mining areas, and planning mining sequences, providing reliable technical support for the refined development of mining enterprises and helping to improve mining efficiency and reduce mining costs.

[0088] In one embodiment, optionally, the method further includes: The final grade prediction result is evaluated using multi-dimensional evaluation indicators, and a decision is made based on the evaluation results as to whether to retrain the model. The multi-dimensional evaluation indicators include: coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error.

[0089] Multi-dimensional evaluation indicators: A set of indicators that quantify the accuracy of the final grade prediction results from different perspectives, avoiding the limitations of a single indicator; Evaluation Result Decision: Based on the achievement of evaluation indicators, determine whether the final prediction result meets the actual application requirements, and then decide whether it is necessary to re-optimize the model parameters and retrain to ensure the reliability of the prediction result.

[0090] Evaluation Indicators Selection: Four core multi-dimensional evaluation indicators were selected to comprehensively cover key dimensions such as prediction accuracy, goodness of fit, and error distribution. Specifically, these include: ① Coefficient of Determination (R²): measures the linear correlation between the predicted value and the true value, with a value range of [0,1]. The closer to 1, the higher the goodness of fit; ② Root Mean Square Error (RMSE): reflects the overall magnitude of the prediction error. The smaller the value, the higher the accuracy; ③ Mean Absolute Error (MAE): reflects the average deviation between the predicted value and the true value. It is not sensitive to extreme values ​​and better reflects the level of normal error; ④ Mean Absolute Percentage Error (MAPE): reflects the prediction error in percentage form, making it easy to intuitively understand the proportion of error to the true value. The smaller the value, the smaller the relative error.

[0091] Evaluation index calculation: Substitute the actual quality values ​​of the test set and the corresponding final predicted values ​​into the above formula to calculate the specific values ​​of the four evaluation indicators.

[0092] Evaluation criteria are set: based on actual needs such as deposit type and exploration accuracy requirements, threshold values ​​for each evaluation indicator are established. For example, for metallic deposits, threshold values ​​are typically set as R² ≥ 0.85 and RMSE ≤ σ. Z / 3(σ Z (Standard deviation of the true grade values ​​in the test set), MAE≤σ Z 5. MAPE ≤ 10%.

[0093] Decision-making: ① If all evaluation indicators meet the threshold, the final prediction result meets the requirements, and there is no need to retrain the model. The result can be directly output for practical application. ② If at least one indicator fails to meet the standard, it is necessary to combine indicator characteristics and visualization analysis (such as scatter plots of predicted and true values, error histograms, residual analysis plots) to locate the root cause of the problem (such as large prediction deviation in high-grade areas, insufficient local spatial continuity, etc.). Then, adjust the model parameters (such as adjusting XGBoost hyperparameters, the number of clusters in the hybrid residual model, fusion weights, etc.) to address the problem, and re-execute the model training, prediction, and fusion steps until the evaluation indicators meet the standard.

[0094] Multi-dimensional evaluation metrics quantify prediction results from different perspectives, such as goodness of fit, absolute error, and relative error. Compared to single metrics, they provide a more comprehensive and objective reflection of the final prediction performance, avoiding misjudgments of prediction accuracy due to the limitations of a single metric. Through evaluation results and visualization analysis, model problems (such as overfitting, insufficient prediction of high-grade areas, and poor spatial continuity) can be accurately identified, providing clear direction for model parameter adjustments, avoiding blind optimization, and improving the efficiency and effectiveness of iterative model optimization. By constraining the evaluation to meet thresholds, it ensures that the final prediction output meets the accuracy requirements of practical applications, preventing the use of inaccurate predictions in mine resource assessment and mining planning, reducing enterprise production risks, and ensuring the scientific nature of production decisions.

[0095] In one embodiment, optionally, the method further includes: Based on incremental learning technology, when new mineral deposit exploration data is acquired, it is preprocessed and then added to the feature dataset to update the feature dataset. Based on the updated feature dataset, the hyperparameters of the XGBoost master regression model, the parameters of the mixed residual model, and the variogram parameters were re-optimized.

[0096] Incremental learning technology: a machine learning technique that updates the model without retraining the entire model, by simply fine-tuning the model parameters with new data. It is suitable for scenarios with dynamically growing data. Feature dataset update: new exploration data is preprocessed and then integrated into the original feature dataset to make the dataset more complete and representative.

[0097] Model and parameter re-optimization: Based on the updated feature dataset, the hyperparameters of the XGBoost master regression model, the parameters of the mixed residual model, and the parameters of the variogram were fine-tuned to ensure that the model always adapts to the latest exploration data and deposit characteristics.

[0098] New data preprocessing: After acquiring new mineral deposit exploration data (such as sample data from new boreholes), the new data is processed according to the preprocessing process in step S101 (outlier removal, missing value imputation, feature construction, dimensionality reduction, standardization, and high-grade sample enhancement) to ensure that the format and scale of the new data are consistent with the original feature dataset.

[0099] Feature dataset update: The preprocessed new data is added to the original feature dataset to form the updated feature dataset. At the same time, the training set, validation set, and test set are re-split (maintaining the original proportions) to ensure the representativeness of the new dataset.

[0100] Model and parameter re-optimization: XGBoost principal regression model: Based on the updated training set, Bayesian optimization is used to fine-tune the model hyperparameters (rather than completely re-optimizing) to adapt to the features of the new data; incremental learning is used to use the original model parameters as initial values ​​and update the model weights through only a small number of iterations to improve training efficiency.

[0101] Hybrid residual model: Recalculate the temporary Kriging interpolation residuals of the newly added samples to update the target training dataset for each region cluster; perform minor fine-tuning on each local XGBoost model; if the newly added samples cause a significant change in the spatial structure of a region cluster, K-means clustering can be re-executed to adjust the region cluster partitioning.

[0102] The variogram parameters are based on the updated spatial structure feature vector (incorporating the spatial structure features of the newly added sample points). The optimal combination of variogram parameters is re-searched through the mapping relationship model M, and the variogram parameters are slightly adjusted to ensure that they are adapted to the updated spatial structure of the ore deposit.

[0103] Updated model validation and prediction: Apply the re-optimized model to the updated test set and verify the model performance through multi-dimensional evaluation indicators; if the performance meets the standards, use the updated model to regenerate the final grade prediction results for subsequent practical applications; if the performance does not meet the standards, repeat the parameter optimization steps until the standards are met.

[0104] Incremental learning technology enables models to quickly adapt to new exploration data, promptly capturing dynamic changes in deposit characteristics (such as local geological patterns revealed by new boreholes), avoiding performance degradation due to data lag, and ensuring that prediction results are always synchronized with the latest exploration findings. Incremental learning does not require retraining the entire model; only minor parameter adjustments are needed, significantly reducing the computational load and time cost of model updates. Compared to the traditional "retraining" mode, efficiency is significantly improved, making it suitable for long-term, continuous exploration and development processes in mines.

[0105] With the continuous accumulation of exploration data, the updated feature dataset is more complete and better reflects the true characteristics of the deposit. Combined with targeted optimization of parameters, the prediction accuracy of the model remains stable in long-term applications, avoiding accuracy fluctuations caused by incomplete data, and providing continuous and reliable technical support for long-term mining planning.

[0106] Figure 2 A block diagram of a mineral deposit grade prediction device according to an embodiment of this application is shown.

[0107] like Figure 2 As shown, in a second aspect, embodiments of this application provide a mineral deposit grade prediction device 20, comprising: The preprocessing module 21 is used to perform preprocessing operations on the mineral deposit exploration data and construct a feature dataset based on the processed mineral deposit exploration data; Module 22 is used to train an XGBoost master regression model based on the feature dataset, construct a hybrid residual model, generate intermediate prediction results using a two-stage fusion mechanism, and establish a mapping relationship between spatial structure, variogram and prediction error. The adjustment module 23 is used to adaptively select and dynamically adjust the mutation function parameters according to the mapping relationship to obtain the optimized mutation function; Interpolation module 24 is used to perform Kriging interpolation using the optimized mutation function to obtain preliminary grade prediction results with spatial continuity; The fusion module 25 is used to fuse the intermediate prediction results and the preliminary grade prediction results to obtain the final grade prediction results.

[0108] In one embodiment, optionally, the preprocessing operations include: outlier removal, missing value imputation, feature construction, dimensionality reduction, standardization, and high-quality sample enhancement. The preprocessing module specifically includes: The processing unit is used to remove outliers from the mineral deposit exploration data and fill in missing values ​​using a standard fraction constraint strategy and an interquartile range constraint strategy to obtain the mineral deposit exploration data after preliminary processing. The construction unit is used to construct spatial geometric features with normalized radial distance and depth based on the pre-processed mineral deposit exploration data, and combine them with the inherent features of the mineral deposit to form an original feature set. The dimensionality reduction unit is used to reduce the dimensionality of the original feature set using principal component analysis to obtain the dimensionality-reduced feature data. The transformation unit is used to transform the dimensionality-reduced feature data into a preset scale by standardizing it with standard scores, thereby forming a standard dataset. An enhancement unit is used to apply a small perturbation conforming to a Gaussian distribution to the high-quality samples in the standard dataset to enhance the high-quality samples and obtain the feature dataset.

[0109] In one embodiment, optionally, the building module includes: A partitioning unit is used to divide the feature dataset into a training set, a validation set, and a test set according to a preset ratio; The optimization unit is used to optimize the hyperparameters of the XGBoost principal regression model using the Bayesian optimization algorithm, with the root mean square error of the five-fold cross-validation as the objective function. The fitting unit is used to train the XGBoost master regression model based on the optimized hyperparameters, and fits the nonlinear relationship of the ore grade by minimizing the loss function. The evaluation unit is used to evaluate the XGBoost principal regression model using cross-validation.

[0110] In one embodiment, optionally, the building module further includes: The execution unit is used to perform temporary kriging interpolation using the initial default mutation function; The estimation unit is used to make a preliminary estimate of the overall grade distribution of the ore deposit based on the three-dimensional spatial coordinates of all sample points and the spatial continuity modeling capability of Kriging interpolation, so as to obtain a temporary predicted value for the corresponding position of each sample point. The sample points are discrete data points containing three-dimensional spatial coordinates and corresponding grade values ​​formed by associating ore samples collected by borehole exploration. The first extraction unit is used to calculate the difference between the actual grade value and the temporary predicted value of each sample point based on the temporary predicted value, to obtain the temporary kriging interpolation prediction residual, and to extract nonlinear geological laws that cannot be modeled by temporary kriging interpolation. Clustering units are used to cluster sample points according to their spatial location using the K-means algorithm, dividing them into K spatially similar regions. The data construction unit is used to separately divide the feature data of the sample points within each region cluster and the corresponding temporary kriging interpolation prediction residuals to construct the target training dataset for that region cluster. The training unit is used to train independent local XGBoost master regression models based on the target training dataset for each region cluster, so as to accurately fit the nonlinear geological laws of that region cluster. The overlay unit is used to overlay the residual prediction results output by the local XGBoost master regression model with the temporary kriging interpolation prediction residuals to obtain the prediction results of the hybrid residual model.

[0111] In one embodiment, optionally, the second-level fusion mechanism includes first-level ridge regression linear fusion and second-level meta-limit gradient boosting nonlinear fusion, and the building unit further includes: The solution unit is used to construct the base model prediction matrix with the first prediction result of the XGBoost principal regression model and the second prediction result of the hybrid residual model as input, and solve the optimal weight by minimizing the objective function, while applying non-negativity constraints and normalization constraints to obtain the linear fusion result; The first generation unit is used to take the linear fusion result, the first prediction result and the second prediction result as inputs, train the XGBoost principal regression model to learn the nonlinear complementary relationship between different models, and generate the intermediate prediction result. The second extraction unit is used to extract the spatial structure features of the ore deposit and form a spatial structure feature vector, wherein the spatial structure features include the spatial distribution features of sample points, geological trend features, and grade distribution features. The second generation unit is used to generate candidate sets of variogram function parameter combinations for spherical models, exponential models, and Gaussian models. The computational unit is used to calculate the prediction error corresponding to each set of parameters and to construct a training dataset with mapping relationships.

[0112] In one embodiment, optionally, the apparatus further includes: The evaluation module is used to evaluate the final grade prediction result through multi-dimensional evaluation indicators and determine whether to retrain the model based on the evaluation results. The multi-dimensional evaluation indicators include: coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error.

[0113] In one embodiment, optionally, the apparatus further includes: The update module is used to update the feature dataset by preprocessing newly acquired mineral deposit exploration data based on incremental learning technology. The optimization module is used to re-optimize the hyperparameters of the XGBoost master regression model, the parameters of the mixed residual model, and the variogram parameters based on the updated feature dataset.

[0114] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for predicting the grade of mineral deposits.

[0115] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described method for predicting the grade of a mineral deposit.

[0116] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the above-described ore grade prediction device and its modules can be referred to the corresponding process in the aforementioned ore grade prediction method embodiments, and will not be repeated here.

[0117] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the model training device and each module described above can be referred to the corresponding processes in the aforementioned embodiments of the ore deposit grade prediction method, and will not be repeated here.

[0118] The aforementioned ore grade prediction device can be implemented as a computer program, which can, for example, Figure 3 It runs on the computer device shown.

[0119] Figure 3 A block diagram of a computer device according to one embodiment of this application is shown.

[0120] See Figure 3 The computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include storage media and internal memory.

[0121] The storage medium may store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform any of the multi-source data mineral deposit grade prediction methods provided in the embodiments of this application.

[0122] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0123] The internal memory provides an environment for the execution of computer programs stored in the storage medium. When executed by a processor, these programs enable the processor to perform any multi-source data-based mineral deposit grade prediction method. The storage medium can be non-volatile or volatile.

[0124] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0125] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0126] In addition, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions for performing the steps of the method in the first aspect embodiment.

[0127] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or electronic device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0128] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0129] It should be understood that although the terms "first," "second," etc., may be used to describe the setting units in the embodiments of this application, these setting units should not be limited to these terms. These terms are only used to distinguish the setting units from each other. For example, without departing from the scope of the embodiments of this application, the first setting unit may also be referred to as the second setting unit, and similarly, the second setting unit may also be referred to as the first setting unit.

[0130] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0131] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0132] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0133] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0134] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting the grade of a mineral deposit, characterized in that, The method includes: Preprocess the mineral deposit exploration data and construct a feature dataset based on the processed mineral deposit exploration data; An XGBoost master regression model is trained based on the aforementioned feature dataset, and a hybrid residual model is constructed. A two-stage fusion mechanism is used to generate intermediate prediction results, while establishing a mapping relationship between spatial structure, variogram, and prediction error. Based on the mapping relationship, the mutation function parameters are adaptively selected and dynamically adjusted to obtain the optimized mutation function; Kriging interpolation is performed using the optimized variogram to obtain preliminary grade prediction results with spatial continuity; The intermediate prediction results and the preliminary grade prediction results are fused to obtain the final grade prediction results.

2. The method according to claim 1, characterized in that, The preprocessing operations include: outlier removal, missing value imputation, feature construction, dimensionality reduction, standardization, and high-grade sample enhancement. Specifically, the preprocessing operations on the mineral deposit exploration data include: The standard score constraint strategy and the interquartile range constraint strategy are used to remove outliers from the mineral deposit exploration data and fill in missing values ​​to obtain the preliminary processed mineral deposit exploration data. Based on the pre-processed mineral deposit exploration data, spatial geometric features with normalized radial distance and depth are constructed, and combined with the inherent features of the mineral deposit to form an original feature set. Principal component analysis is used to reduce the dimensionality of the original feature set to obtain the dimensionality-reduced feature data. The dimensionality-reduced feature data is transformed into a preset scale by standard score standardization to form a standard dataset. A small perturbation conforming to a Gaussian distribution is applied to the high-quality samples in the standard dataset to enhance the high-quality samples, thereby obtaining the feature dataset.

3. The method according to claim 1, characterized in that, Training an XGBoost master regression model based on the aforementioned feature dataset includes: The feature dataset is divided into a training set, a validation set, and a test set according to a preset ratio; The hyperparameters of the XGBoost principal regression model are optimized using the Bayesian optimization algorithm with the root mean square error of five-fold cross-validation as the objective function. The XGBoost master regression model is trained based on optimized hyperparameters, and the nonlinear relationship between the ore deposit grade is fitted by minimizing the loss function. Cross-validation was used to evaluate the XGBoost principal regression model.

4. The method according to claim 1, characterized in that, The construction process of the hybrid residual model includes: Temporary kriging interpolation is performed using the initial default mutation function; Based on the three-dimensional spatial coordinates of all sample points, and utilizing the spatial continuity modeling capability of Kriging interpolation, a preliminary estimate of the overall grade distribution of the ore deposit is made to obtain a temporary predicted value for the corresponding position of each sample point. The sample points are discrete data points containing three-dimensional spatial coordinates and corresponding grade values, which are formed by associating ore samples collected through borehole exploration. Based on the provisional predicted value, the difference between the actual grade value and the provisional predicted value of each sample point is calculated to obtain the provisional Kriging interpolation prediction residual, and nonlinear geological laws that cannot be modeled by provisional Kriging interpolation are extracted. The K-means algorithm is used to cluster the sample points according to their spatial location, dividing them into K spatially similar regions. For each region cluster, the feature data of the sample points within the cluster and the corresponding temporary kriging interpolation prediction residuals are separately divided to construct the target training dataset for that region cluster. Based on the target training dataset for each regional cluster, independent local XGBoost master regression models are trained to accurately fit the nonlinear geological patterns of that regional cluster. The residual prediction results output by the local XGBoost master regression model are superimposed with the temporary Kriging interpolation prediction residuals to obtain the prediction results of the hybrid residual model.

5. The method according to claim 1, characterized in that, The two-stage fusion mechanism includes a first-stage ridge regression linear fusion and a second-stage meta-limit gradient boosting nonlinear fusion. The generation of intermediate prediction results using this two-stage fusion mechanism includes: Using the first prediction result of the XGBoost principal regression model and the second prediction result of the hybrid residual model as input, a base model prediction matrix is ​​constructed, and the optimal weights are solved by minimizing the objective function. At the same time, non-negativity constraints and normalization constraints are applied to obtain the linear fusion result. Using the linear fusion result, the first prediction result, and the second prediction result as inputs, an XGBoost principal regression model is trained to learn the nonlinear complementary relationship between different models and generate the intermediate prediction result. The establishment of the mapping relationship between spatial structure, variogram, and prediction error includes: Extract the spatial structure features of the ore deposit to form a spatial structure feature vector, wherein the spatial structure features include the spatial distribution features of sample points, geological trend features, and grade distribution features; Generate candidate sets of variogram parameter combinations for spherical, exponential, and Gaussian models; Calculate the prediction error for each parameter combination and construct a training dataset for mapping relationships.

6. The method according to claim 1, characterized in that, The method further includes: The final grade prediction result is evaluated using multi-dimensional evaluation indicators, and a decision is made based on the evaluation results as to whether to retrain the model. The multi-dimensional evaluation indicators include: coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error.

7. The method according to claim 1, characterized in that, The method further includes: Based on incremental learning technology, when new mineral deposit exploration data is acquired, it is preprocessed and then added to the feature dataset to update the feature dataset. Based on the updated feature dataset, the hyperparameters of the XGBoost master regression model, the parameters of the mixed residual model, and the variogram parameters were re-optimized.

8. A mineral deposit grade prediction device, characterized in that, include: The preprocessing module is used to preprocess mineral exploration data and construct a feature dataset based on the processed mineral exploration data. The module is used to train the XGBoost master regression model based on the feature dataset, construct a hybrid residual model, generate intermediate prediction results using a two-stage fusion mechanism, and establish a mapping relationship between spatial structure, variogram and prediction error. An adjustment module is used to adaptively select and dynamically adjust the mutation function parameters according to the mapping relationship to obtain an optimized mutation function; The interpolation module is used to perform Kriging interpolation using the optimized mutation function to obtain preliminary grade prediction results with spatial continuity. The fusion module is used to fuse the intermediate prediction results and the preliminary grade prediction results to obtain the final grade prediction results.

9. A computer device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, the instructions being configured to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The device stores computer-executable instructions for performing the method as described in any one of claims 1 to 7.

Images