Mineral reserve estimation method based on artificial intelligence

Abstract

The invention relates to the technical field of mineral resource estimation, and discloses a mineral reserve estimation method based on artificial intelligence, comprising the following steps: step 1, collecting and preprocessing data, acquiring multi-source data, and performing data standardization and feature extraction; 2, carrying out ore body modeling through a 3D convolutional neural network, constructing the 3D convolutional neural network, and carrying out ore body prediction through three-dimensional convolution calculation; step 3, optimizing ore body space prediction in combination with Kriging interpolation, preliminarily predicting ore body space distribution through Kriging interpolation, and further optimizing by using a deep learning model; and step 4, estimating reserves based on Bayesian optimization. The technical scheme of combining multi-source data fusion, a 3D convolutional neural network and Kriging interpolation is adopted, the effect of accurately capturing ore body space distribution under the complex geological condition is achieved, and compared with an existing scheme depending on a traditional geological statistical method, the problem that the fault zone and heterogeneous region modeling capacity is insufficient is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of mineral resource estimation, and in particular to a mineral reserve estimation method based on artificial intelligence. Background Art

[0002] This paper provides an artificial intelligence-based mineral reserve estimation method designed to address the current challenges of insufficient ore body modeling capabilities and insufficient uncertainty analysis in reserve estimation in complex geological environments. Conventional mineral reserve estimation methods typically rely on geostatistical methods, such as kriging interpolation, which are effective in some situations but cannot accurately capture the spatial distribution of ore bodies in fault zones and heterogeneous regions. Consequently, these methods often suffer from prediction errors when dealing with complex geological environments, particularly when dealing with concealed ore bodies and fault zones, where accuracy is significantly reduced.

[0003] Furthermore, existing reserve estimation techniques often rely on static Monte Carlo simulations, which are unable to dynamically adjust to changes in data during the estimation process. Consequently, errors and uncertainties in reserve estimates cannot be corrected in real time, resulting in reduced credibility and insufficient error quantification. In actual mineral resource development, this leads to estimation bias, which in turn affects the scientific nature of decision-making and the economic benefits of mines.

[0004] Furthermore, existing techniques often overlook sampling bias and model errors when estimating reserves in complex geological environments. This can lead to inaccurate estimates, particularly when conducting uncertainty analysis. Because traditional methods rely on simplified assumptions and models, they fail to fully account for orebody uncertainty, leading to either optimistic or pessimistic estimates and compromising risk assessments for mine development.

[0005] Therefore, those skilled in the art provide a mineral reserve estimation method based on artificial intelligence to solve the above-mentioned problems. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the present invention provides a mineral reserve estimation method based on artificial intelligence to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a mineral reserve estimation method based on artificial intelligence, comprising the following steps:

[0008] Step 1: Data collection and preprocessing: acquiring multi-source data and performing data standardization and feature extraction;

[0009] Step 2: 3D convolutional neural network is used to model the ore body, construct a 3D convolutional neural network and perform ore body prediction through three-dimensional convolution calculation;

[0010] Step 3: Combine Kriging interpolation to optimize the spatial prediction of ore bodies. Kriging interpolation is used to preliminarily predict the spatial distribution of ore bodies and further optimize it using a deep learning model.

[0011] Step 4: Reserve estimation based on Bayesian optimization, optimizing model hyperparameters through Gaussian processes and maximizing the expected improvement acquisition function;

[0012] Step 5: quantify uncertainty based on the Markov Chain Monte Carlo method and calculate the confidence interval of reserves.

[0013] Preferably, in step 1, the multi-source data includes collected drilling data, geophysical data and remote sensing data, and the acquired data are spatially aligned and standardized.

[0014] Preferably, in step 1, the data standardization and feature extraction include:

[0015] Step 1.1, use principal component analysis to reduce data dimensionality;

[0016] For high-dimensional data, principal component analysis is used to reduce the dimensionality. The mathematical formula of principal component analysis is as follows: Let the original data matrix be X∈R m×n ,

[0017] Among them, m is the number of samples and n is the feature dimension;

[0018] a. Calculate the data covariance matrix:

[0019] b. Calculate the eigenvalues ​​and eigenvectors of the covariance matrix C:

[0020] Cv i =λ i v i , i=1,2,...,n, where C is the covariance matrix, v i is the eigenvector, λ i is the eigenvalue;

[0021] c. Select the eigenvectors corresponding to the first k largest eigenvalues ​​and construct the dimensionality reduction transformation matrix:

[0022] W:W=[v1,v2,...,v k ],

[0023] d. Calculate the data after dimensionality reduction: X′=XW, where X′ is the data after dimensionality reduction, X is the original data, and W is the transformation matrix;

[0024] Step 1.2, use the semivariogram to calculate spatial correlation and optimize data interpolation;

[0025] In order to optimize the spatial interpolation of data, it is necessary to calculate the semivariogram to measure the trend of variable changes at different spatial locations. The semivariogram calculation formula is:

[0026]

[0027] Among them, h is the spatial distance, Z(s i ) is the position s i The observation value at , N(h) is the number of sampling logarithms with a spacing of h, Z(s i +h) is the position s i The observation value at +h, γ(h) is the value of the semivariogram;

[0028] After the calculation is completed, the exponential model or Gaussian model is used to fit the semivariogram:

[0029] Exponential model: γ(h)=C0+C(1-e -h / a )

[0030] Gaussian model:

[0031] Among them, C0 is the base value, C is the reference value, a is the range, γ(h) is the value of the semivariogram, h is the spatial distance, and e is a constant;

[0032] After the semivariogram is calculated, it is used for Kriging interpolation to optimize the ore body distribution prediction.

[0033] Preferably, in step 2, the 3D convolutional neural network outputs the ore body prediction result through three-dimensional convolution operation, pooling layer dimensionality reduction and full connection layer, and uses the cross entropy loss function to optimize model training.

[0034] Preferably, the three-dimensional convolution operation formula is:

[0035]

[0036] in, represents the input ore body data tensor, represents the weight of the three-dimensional convolution kernel, ζ represents the bias parameter, φ(·) represents the activation function, i, j, k represent the corresponding spatial position index in the output tensor, and p max ,q max 、r max Indicates the size of the convolution kernel in the X, Y, and Z directions. is the eigenvalue obtained at position (i, j, k) after the convolution operation;

[0037] The pooling layer dimensionality reduction formula:

[0038]

[0039] in, represents the pooled output tensor, α max , β max , γ max represents the size of the pooling window in the X, Y, and Z directions, α, β, and γ represent the local offset along the X, Y, and Z directions within the pooling window. is the convolution output input to the pooling layer.

[0040] Preferably, the fully connected layer output formula is:

[0041] in, is the mth element of the flattened output of the pooling layer, is the mth input and the mth The weight parameter between outputs, m max is the length of the flattened vector, is the output node in the fully connected layer The bias parameter of , softmax(·) is the softmax function, is the output of the fully connected layer The predicted probability of the class;

[0042] The cross entropy loss function formula:

[0043] Among them, Y e is the one-hot encoding of the true label element, is the output of the fully connected layer The predicted probability of the class, is the total number of output categories, ln(·) is the natural logarithm function, and L is the cross entropy loss value.

[0044] Preferably, in step 3, the Kriging interpolation performs spatial distribution interpolation of the ore body by calculating the semivariogram and minimizing the estimated variance.

[0045] Preferably, the Kriging interpolation uses a semivariogram to calculate the influence weight of each sampling point on the point to be estimated; assuming the point to be estimated is s0, the Kriging weight is calculated by minimizing the estimation variance; for the known sampling points S = {s1, s2, ..., s N}, estimated value The Kriging interpolation expression is:

[0046]

[0047] Among them, λ i are the kriging weights, which are calculated by solving the following system of equations;

[0048] Z(si ) is the known sampling point s i reserves value.

[0049] Calculate the kriging weights λ by minimizing the estimated variance i , we get the following equation:

[0050]

[0051] Among them, γ(h ij ) is the semivariogram value between the i-th and j-th points, γ(h0) is the semivariogram value between the point to be estimated and the known point, and N is the total number of known sampling points involved in the interpolation calculation.

[0052] Preferably, in step 3, the deep learning model further optimizes the Kriging interpolation result and fuses it through deep neural network and Bayesian model averaging.

[0053] Preferably, the method of further optimizing the Kriging interpolation result by using the deep learning model is specifically as follows:

[0054] Suppose the output of the deep neural network is: y D =g(u;Θ),

[0055] Where u represents the input feature vector, Θ represents the parameters of the deep neural network, g(·) is the forward propagation mapping function of the deep neural network, and y D is the output of the deep neural network model;

[0056] Assume that the Kriging interpolation result is: z K =I Kriging (s),

[0057] Among them, z K is the Kriging interpolation result, s represents the geographic location, I Kriging (·) represents the Kriging interpolation calculation function;

[0058] Using the Bayesian model averaging idea, the deep neural network prediction value and the Kriging result are weighted and fused to obtain the final output estimate: Z F =ω D y D +ω K z K ,

[0059] Among them, Z F is the final ore reserve estimate, y D is the output of the deep neural network model, z K is the Kriging interpolation result, ω D Represents the contribution ratio of the deep neural network model, ω KIt represents the contribution ratio of the Kriging interpolation result, and the weight satisfies ω D +ω K =1;

[0060] The weight is determined by the error value of each model on the validation set. The specific calculation formula is:

[0061]

[0062] Among them, E D Represents the estimation error of the deep neural network model, and the calculation formula is:

[0063]

[0064] Among them, N D is the number of deep neural network validation samples, y true,i is the true ore body parameter of the i-th sample, y D,i is the ore body parameter predicted by the deep neural network model for the i-th sample;

[0065] E K It represents the estimation error of the Kriging model and is calculated as follows:

[0066]

[0067] Among them, N K is the number of Kriging validation samples, z K,j It represents the predicted value calculated by Kriging interpolation for the j-th validation sample.

[0068] The present invention provides a mineral reserve estimation method based on artificial intelligence. It has the following beneficial effects:

[0069] 1. This invention adopts a technical solution that combines multi-source data fusion, 3D convolutional neural network and Kriging interpolation to accurately capture the spatial distribution of ore bodies under complex geological conditions. Compared with existing solutions that rely on traditional geostatistical methods, it solves the problem of insufficient modeling capabilities for fault zones and heterogeneous regions.

[0070] 2. The present invention uses Bayesian optimization in conjunction with a deep neural network to dynamically adjust model hyperparameters, achieving the effect of real-time correction of reserve estimation errors and narrowing the uncertainty range. Compared with the static Monte Carlo simulation solution in the prior art, it solves the problems of low reserve estimation credibility and insufficient error quantification.

[0071] 3. The present invention adopts Markov chain Monte Carlo sampling and integrates Bayesian model averaging to achieve the effect of clearly constructing reserve confidence intervals and clarifying the prediction model error. Compared with the existing solution that ignores sampling bias, it solves the problem of inaccurate reserve estimation in complex geological environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0073] To help those skilled in the art understand the present invention, the following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0074] The present invention is described in detail below with reference to the accompanying drawings:

[0075] Example:

[0076] Please see the attached Figure 1 The embodiment of the present invention provides a mineral reserve estimation method based on artificial intelligence, which is characterized by comprising the following steps:

[0077] Step 1: Data collection and preprocessing: Acquire multi-source data and perform data standardization and feature extraction. Multi-source data includes borehole data, geophysical data, and remote sensing data. The acquired data is spatially aligned and standardized. Data standardization and feature extraction include:

[0078] Step 1.1, use principal component analysis to reduce data dimensionality;

[0079] For high-dimensional data, principal component analysis is used to reduce the dimensionality. The mathematical formula of principal component analysis is as follows: Let the original data matrix be X∈R m×n ,

[0080] Among them, m is the number of samples and n is the feature dimension;

[0081] a. Calculate the data covariance matrix:

[0082] b. Calculate the eigenvalues ​​and eigenvectors of the covariance matrix C:

[0083] Cv i =λ i v i , i=1,2,...,n, where C is the covariance matrix, v i is the eigenvector, λ i is the eigenvalue;

[0084] c. Select the eigenvectors corresponding to the first k largest eigenvalues ​​and construct the dimensionality reduction transformation matrix:

[0085] W:W=[v1,v2,...,v k ],

[0086] d. Calculate the data after dimensionality reduction: X′=XW, where X′ is the data after dimensionality reduction, X is the original data, and W is the transformation matrix;

[0087] Step 1.2, use the semivariogram to calculate spatial correlation and optimize data interpolation;

[0088] In order to optimize the spatial interpolation of data, it is necessary to calculate the semivariogram to measure the trend of variable changes at different spatial locations. The semivariogram calculation formula is:

[0089]

[0090] Among them, h is the spatial distance, Z(s i ) is the position s i The observation value at , N(h) is the number of sampling logarithms with a spacing of h, Z(s i +h) is the position s i The observation value at +h, γ(h) is the value of the semivariogram;

[0091] After the calculation is completed, the exponential model or Gaussian model is used to fit the semivariogram:

[0092] Exponential model: γ(h)=C0+C(1-e -h / a )

[0093] Gaussian model:

[0094] Among them, C0 is the base value, C is the reference value, a is the range, γ(h) is the value of the semivariogram, h is the spatial distance, and e is a constant;

[0095] After the semivariogram is calculated, it is used for Kriging interpolation to optimize the ore body distribution prediction;

[0096] Step 2: 3D convolutional neural network is used to model the ore body. A 3D convolutional neural network is constructed and ore body prediction is performed through three-dimensional convolution calculation. The 3D convolutional neural network outputs the ore body prediction results through three-dimensional convolution operation, pooling layer dimensionality reduction and full connection layer, and uses the cross entropy loss function to optimize model training. The three-dimensional convolution operation formula is:

[0097]

[0098] Where χ represents the input ore body data tensor, represents the weight of the three-dimensional convolution kernel, ζ represents the bias parameter, φ(·) represents the activation function, i, j, k represent the corresponding spatial position index in the output tensor, and p max ,q max 、rmax Indicates the size of the convolution kernel in the X, Y, and Z directions. is the eigenvalue obtained at position (i, j, k) after the convolution operation;

[0099] Pooling layer dimensionality reduction formula:

[0100]

[0101] in, represents the pooled output tensor, α max , β max , γ max represents the size of the pooling window in the X, Y, and Z directions, α, β, and γ represent the local offset along the X, Y, and Z directions within the pooling window. is the convolution output input to the pooling layer;

[0102] Fully connected layer output formula:

[0103] in, is the mth element of the flattened output of the pooling layer, is the mth input and the mth The weight parameter between outputs, m max is the length of the flattened vector, is the output node in the fully connected layer The bias parameter of , softmax(·) is the softmax function, is the output of the fully connected layer The predicted probability of the class;

[0104] Cross entropy loss function formula:

[0105] in, is the one-hot encoding of the true label element, is the output of the fully connected layer The predicted probability of the class, is the total number of output categories, ln(·) is the natural logarithm function, and L is the cross entropy loss value;

[0106] Step 3: Combine Kriging interpolation to optimize the spatial prediction of ore bodies. Kriging interpolation is used to preliminarily predict the spatial distribution of ore bodies and further optimize it using a deep learning model. Kriging interpolation interpolates the spatial distribution of ore bodies by calculating the semivariogram and minimizing the estimated variance.

[0107] Kriging interpolation uses the semivariogram function to calculate the influence weight of each sampling point on the estimated point; let the estimated point be s0, and the Kriging weight is calculated by minimizing the estimated variance; for the known sampling points S = {s1, s2, ..., s N}, estimated value The Kriging interpolation expression is:

[0108]

[0109] Among them, λ i are the kriging weights, which are calculated by solving the following system of equations;

[0110] Z(s i ) is the known sampling point s i reserves value.

[0111] Calculate the kriging weights λ by minimizing the estimated variance i , we get the following equation:

[0112]

[0113] Among them, γ(h ij ) is the semivariogram value between the i-th and j-th points, γ(h0) is the semivariogram value between the point to be estimated and the known point, and N is the total number of known sampling points involved in the interpolation calculation;

[0114] The deep learning model further optimizes the Kriging interpolation results by integrating deep neural networks and Bayesian model averaging;

[0115] The method of using deep learning models to further optimize Kriging interpolation results is as follows:

[0116] Suppose the output of the deep neural network is: y D =g(u;Θ),

[0117] Where u represents the input feature vector, Θ represents the parameters of the deep neural network, g(·) is the forward propagation mapping function of the deep neural network, and y D is the output of the deep neural network model;

[0118] Assume that the Kriging interpolation result is: z K =I Kriging (s),

[0119] Among them, z K is the Kriging interpolation result, s represents the geographic location, I Kriging (·) represents the Kriging interpolation calculation function;

[0120] Using the Bayesian model averaging idea, the deep neural network prediction value and the Kriging result are weighted and fused to obtain the final output estimate: Z F =ω D y D +ω K z K ,

[0121] Among them, Z F is the final ore reserve estimate, y D is the output of the deep neural network model, z K is the Kriging interpolation result, ω D Represents the contribution ratio of the deep neural network model, ω K It represents the contribution ratio of the Kriging interpolation result, and the weight satisfies ω D +ω K =1;

[0122] The weight is determined by the error value of each model on the validation set. The specific calculation formula is:

[0123]

[0124] Among them, E D Represents the estimation error of the deep neural network model, and the calculation formula is:

[0125]

[0126] Among them, N D is the number of deep neural network validation samples, y true,i is the true ore body parameter of the i-th sample, y D,i is the ore body parameter predicted by the deep neural network model for the i-th sample;

[0127] E K It represents the estimation error of the Kriging model and is calculated as follows:

[0128]

[0129] Among them, N K is the number of Kriging validation samples, z K,j Is the predicted value calculated by Kriging interpolation for the j-th validation sample

[0130] Step 4: Reserve estimation based on Bayesian optimization, optimizing model hyperparameters through Gaussian processes and maximizing the expected improvement acquisition function;

[0131] Step 5: quantify uncertainty based on the Markov Chain Monte Carlo method and calculate the confidence interval of reserves.

[0132] This step integrates multi-source data, combining borehole, geophysical, and remote sensing data for spatial alignment and standardization, effectively integrating diverse resource information. Principal component analysis is used to reduce data dimensionality, optimizing the processing of high-dimensional data and significantly improving data accuracy and efficiency. Furthermore, the semivariogram is used to calculate spatial correlation, optimizing the kriging interpolation process and ensuring the quality of the spatial interpolation.

[0133] This step uses a 3D convolutional neural network to model the ore body. It employs three-dimensional convolution operations and pooling layers for dimensionality reduction, accurately capturing the spatial distribution characteristics of the ore body. Model training is optimized using fully connected layers and a cross-entropy loss function to ensure accurate prediction of the ore body's spatial distribution. Compared to traditional modeling methods, deep learning models are self-optimizing and can handle complex geological environments.

[0134] Combining kriging interpolation with deep learning models, this step significantly improves the accuracy of orebody spatial distribution predictions by optimizing kriging interpolation results. The weighted fusion of deep neural networks and Bayesian models further enhances the reliability of predictions, overcoming the limitations of traditional kriging interpolation methods in complex geological conditions. This is particularly effective when dealing with concealed orebodies and heterogeneous regions.

[0135] By dynamically adjusting model hyperparameters through Bayesian optimization, this step can correct reserve estimation errors in real time, flexibly adjust the estimation range, and improve the accuracy and credibility of reserve estimates. Uncertainty quantification using Markov Chain Monte Carlo methods, constructing confidence intervals for reserves, effectively quantifies model uncertainty. Compared to traditional static Monte Carlo simulation methods, this method can accurately reflect the uncertainty in ore body estimation, improving the scientific nature of decision support.

[0136] Summary: This paper proposes an innovative mineral reserve estimation method by integrating artificial intelligence and geostatistical techniques. The advantages of each step include improved modeling accuracy, optimized reserve estimation errors, and enhanced uncertainty analysis capabilities. This method effectively addresses the shortcomings of traditional methods in complex geological environments and provides accurate and reliable technical support for mineral resource development and decision-making.

[0137] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A mineral reserve estimation method based on artificial intelligence, characterized in that: The steps include: Step 1: Data collection and preprocessing: acquiring multi-source data and performing data standardization and feature extraction; Step 2: 3D convolutional neural network is used to model the ore body, construct a 3D convolutional neural network and perform ore body prediction through three-dimensional convolution calculation; Step 3: Combine Kriging interpolation to optimize the spatial prediction of ore bodies. Kriging interpolation is used to preliminarily predict the spatial distribution of ore bodies and further optimize it using a deep learning model. Step 4: Reserve estimation based on Bayesian optimization, optimizing model hyperparameters through Gaussian processes and maximizing the expected improvement acquisition function; Step 5: quantify uncertainty based on the Markov Chain Monte Carlo method and calculate the confidence interval of reserves.

2. The method for estimating mineral reserves based on artificial intelligence according to claim 1, characterized in that: In step 1, the multi-source data includes collected drilling data, geophysical data and remote sensing data, and the acquired data are spatially aligned and standardized.

3. The method for estimating mineral reserves based on artificial intelligence according to claim 1, characterized in that: In step 1, the data standardization and feature extraction include: Step 1.1, use principal component analysis to reduce data dimensionality; For high-dimensional data, principal component analysis is used to reduce the dimensionality. The mathematical formula of principal component analysis is as follows: Let the original data matrix be X∈R m×n , Among them, m is the number of samples and n is the feature dimension; a. Calculate the data covariance matrix: b. Calculate the eigenvalues ​​and eigenvectors of the covariance matrix C: Cv i =λ i v i , i=1,2,...,n, where C is the covariance matrix, v i is the eigenvector, λ i is the eigenvalue; c. Select the eigenvectors corresponding to the first k largest eigenvalues ​​and construct the dimensionality reduction transformation matrix: <h2 style=";text-align:left;direction:ltr">W:W=[v1,v2,...,v<h2 style=";text-align:left;direction:ltr"> k <h2 style=";text-align:left;direction:ltr"> ], d. Calculate the data after dimensionality reduction: X ′ =XW, where X ′ is the data after dimensionality reduction, X is the original data, and W is the transformation matrix; Step 1.2, use the semivariogram to calculate spatial correlation and optimize data interpolation; In order to optimize the spatial interpolation of data, it is necessary to calculate the semivariogram to measure the trend of variable changes at different spatial locations. The semivariogram calculation formula is: Among them, h is the spatial distance, Z(s i ) is the position s i The observation value at , N(h) is the number of sampling logarithms with a spacing of h, Z(s i +h) is the position s i The observation value at +h, γ(h) is the value of the semivariogram; After the calculation is completed, the exponential model or Gaussian model is used to fit the semivariogram: Exponential model: γ(h)=C0+C(1-e -h / a ) Gaussian model: Among them, C0 is the base value, C is the reference value, a is the range, γ(h) is the value of the semivariogram, h is the spatial distance, and e is a constant; After the semivariogram is calculated, it is used for Kriging interpolation to optimize the ore body distribution prediction.

4. The method for estimating mineral reserves based on artificial intelligence according to claim 1, characterized in that: In step 2, the 3D convolutional neural network outputs the ore body prediction result through three-dimensional convolution operation, pooling layer dimensionality reduction and full connection layer, and uses the cross entropy loss function to optimize model training.

5. The method for estimating mineral reserves based on artificial intelligence according to claim 4, characterized in that: The three-dimensional convolution operation formula: in, represents the input ore body data tensor, represents the weight of the three-dimensional convolution kernel, ζ represents the bias parameter, φ(·) represents the activation function, i, j, k represent the corresponding spatial position index in the output tensor, and p max ,q max 、r max Indicates the size of the convolution kernel in the X, Y, and Z directions. is the eigenvalue obtained at position (i, j, k) after the convolution operation; The pooling layer dimensionality reduction formula: in, represents the pooled output tensor, α max , β max , γ max represents the size of the pooling window in the X, Y, and Z directions, α, β, and γ represent the local offset along the X, Y, and Z directions within the pooling window. is the convolution output input to the pooling layer.

6. The method for estimating mineral reserves based on artificial intelligence according to claim 4, characterized in that: The fully connected layer output formula: in, is the mth element of the flattened output of the pooling layer, is the mth input and the mth The weight parameter between outputs, m max is the length of the flattened vector, is the output node in the fully connected layer The bias parameter of , softmax(·) is the softmax function, is the output of the fully connected layer The predicted probability of the class; The cross entropy loss function formula: in, is the one-hot encoding of the true label element, is the output of the fully connected layer The predicted probability of the class, is the total number of output categories, ln(·) is the natural logarithm function, and L is the cross entropy loss value.

7. The method for estimating mineral reserves based on artificial intelligence according to claim 3, characterized in that: In step 3, the Kriging interpolation performs spatial distribution interpolation of the ore body by calculating the semivariogram and minimizing the estimated variance.

8. The method for estimating mineral reserves based on artificial intelligence according to claim 7, characterized in that: The Kriging interpolation uses the semivariogram function to calculate the influence weight of each sampling point on the estimated point; let the estimated point be s0, and the Kriging weight is calculated by minimizing the estimation variance; for the known sampling points S = {s1, s2, ..., s N }, estimated value The Kriging interpolation expression is: Among them, λ i are the kriging weights, which are calculated by solving the following system of equations; Z(s i ) is the known sampling point s i reserves value. Calculate the kriging weights λ by minimizing the estimated variance i , we get the following equation: Among them, γ(h ij ) is the semivariogram value between the i-th and j-th points, γ(h0) is the semivariogram value between the point to be estimated and the known point, and N is the total number of known sampling points involved in the interpolation calculation.

9. The method for estimating mineral reserves based on artificial intelligence according to claim 8, characterized in that: In step 3, the deep learning model further optimizes the Kriging interpolation result and fuses it through deep neural network and Bayesian model averaging.

10. The method for estimating mineral reserves based on artificial intelligence according to claim 9, characterized in that: The method of further optimizing the Kriging interpolation results using the deep learning model is specifically as follows: Suppose the output of the deep neural network is: y D =g(u;Θ), Where u represents the input feature vector, Θ represents the parameters of the deep neural network, g(·) is the forward propagation mapping function of the deep neural network, and y D is the output of the deep neural network model; Assume that the Kriging interpolation result is: z K =I Kriging (s), Among them, z K is the Kriging interpolation result, s represents the geographic location, I Kriging (·) represents the Kriging interpolation calculation function; Using the Bayesian model averaging idea, the deep neural network prediction value and the Kriging result are weighted and fused to obtain the final output estimate: Z F =ω D y D +ω K z K , Among them, Z F is the final ore reserve estimate, y D is the output of the deep neural network model, z K is the Kriging interpolation result, ω D Represents the contribution ratio of the deep neural network model, ω K It represents the contribution ratio of the Kriging interpolation result, and the weight satisfies ω D +ω K =1; The weight is determined by the error value of each model on the validation set. The specific calculation formula is: Among them, E D Represents the estimation error of the deep neural network model, and the calculation formula is: Among them, N D is the number of deep neural network validation samples, y true,i is the true ore body parameter of the i-th sample, y D,i is the ore body parameter predicted by the deep neural network model for the i-th sample; E K It represents the estimation error of the Kriging model and is calculated as follows: Among them, N K is the number of Kriging validation samples, z K,j It represents the predicted value calculated by Kriging interpolation for the j-th validation sample.

Images