Training method of geological mineral resources prediction model based on artificial intelligence

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses a geological mineral resource prediction model training method based on artificial intelligence, including: using a data expansion model to expand geological mineral resource data to obtain a training data set; constructing a prediction model through an extreme learning machine based on conjugate gradient, inputting the geological mineral resource data in the training data set into the prediction model for training; randomly initializing the hidden layer parameters of the extreme learning machine, calculating the hidden layer output matrix; calculating the output layer weight of the extreme learning machine through the conjugate gradient method; and updating the hidden layer output matrix. The present invention adopts a belief propagation mechanism, which can simulate the inherent correlation of geological mineral resource data, and adjusts the generated samples through the covariance matrix to be more in line with the distribution characteristics of actual geological mineral resources; adopts a dynamic constraint game strategy, and dynamically adjusts the strategies of the generator and the discriminator to enhance the adaptability to complex data.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a geological mineral resources prediction model training method based on artificial intelligence. Background Art

[0002] With the increasing global demand for mineral resources, the exploration and development of geological mineral resources has become increasingly important. However, traditional geological exploration methods face challenges such as difficult data collection, high cost, and low efficiency. Especially in complex geological environments, it is very difficult to obtain enough high-quality data samples. In addition, geological mineral resource data are usually highly nonlinear and complex, and existing machine learning models are difficult to effectively process these complex data, resulting in poor generalization ability of prediction models and low prediction accuracy. In this context, how to expand and optimize geological mineral resource data through intelligent technology and improve the accuracy and efficiency of prediction models has become an important direction of current technological development.

[0003] The Chinese invention patent with publication number CN118709523A proposes a method for predicting geomechanical parameters based on deep learning, which involves the field of geological survey. The method includes: obtaining a learning set of the area to be tested; the learning set of the area to be tested includes the physical parameters of each position in the area to be tested; according to the learning set of the area to be tested, a mechanical parameter prediction model is used to predict the mechanical parameters of each position in the area to be tested, and a prediction set of the area to be tested is obtained; the mechanical parameter prediction model is obtained by pre-training the machine learning network with a training sample set; the training sample set includes multiple groups of known physical parameters and corresponding mechanical parameters; according to the learning set and prediction set of the area to be tested, a three-dimensional spatial parameter distribution model of the area to be tested is established. This patent constructs a mechanical parameter prediction model based on machine learning, which is suitable for the prediction of mechanical parameters of complex shale formations and can finely characterize the spatial distribution characteristics of three-dimensional mechanical parameters.

[0004] The Chinese invention patent with publication number CN116523136A proposes a method and device for intelligent spatial prediction of mineral resources based on multi-model integrated learning. The method includes: obtaining a feature data set of a geological space to be predicted, wherein the feature data set of the geological space to be predicted includes: a feature sub-data set based on a knowledge perspective and a feature sub-data set based on a data science perspective; using a trained convolutional neural network to extract features from the feature data set to obtain a new feature data set; predicting mineral resources based on the new feature data set and the trained Stacking multi-model integrated learning spatial intelligent modeling to output the final mineral resource prediction value.

[0005] The Chinese invention patent with the publication number CN117951551A proposes a method for predicting mineral resources based on singularity index and self-organizing neural network, which belongs to the field of mineral exploration technology. The method includes the following steps: obtaining the spatial distribution of the catchment basins in the study area; calculating the singularity index of each catchment basin; extracting the geological ore-controlling elements in the study area; obtaining the metallogenic favorableness of each geological ore-controlling element to the catchment basin; classifying the catchment basins through self-organizing neural network; identifying the metallogenic favorable area in the study area. This patent is based on the method for predicting mineral resources based on singularity index and self-organizing neural network. The mineralization information of key metal mineral resources is weak and difficult to identify. Based on the sampling principle of water system sediments, the sampling unit is used as the basic unit for extracting geochemical anomalies, and the geological ore-controlling factors in the catchment basin unit are introduced as constraints to identify geochemical anomalies and provide reference for the next step of mineral exploration.

[0006] The prior art including the above patents has the following problems:

[0007] 1. Traditional data augmentation methods, such as simple oversampling or augmentation based on linear interpolation, are often unable to effectively cope with complex geological and mineral resource data. The fidelity of the augmented samples is low and they cannot truly reflect the distribution of geological data, resulting in poor generalization ability of the prediction model.

[0008] 2. When generating data, traditional generative adversarial networks often lack effective control over the quality of generated samples. The distribution of generated samples may be overly concentrated or discrete, resulting in the generated data being unable to well represent the diversity and complexity of actual data.

[0009] 3. Many existing machine learning models often have low computational efficiency when processing large-scale geological data, especially when processing high-dimensional data and sparse matrices. The training of prediction models takes a long time and is prone to overfitting problems.

[0010] 4. Traditional prediction models often show weak generalization ability when faced with complex nonlinear geological data and are unable to adapt to complex changes in data distribution. Especially when training samples are scarce, the prediction model is prone to overfitting, affecting the prediction accuracy. Summary of the invention

[0011] In order to solve the above technical problems, the present invention provides a geological and mineral resources prediction model training method based on artificial intelligence. By constructing a generative adversarial network based on belief propagation, combined with the energy conservation mechanism and the dynamic constraint game strategy, the problems of difficult geological and mineral resource data collection, insufficient samples and difficult modeling of complex data distribution are solved, thereby improving the quality and diversity of data expansion, optimizing the reasonable distribution of generated samples, and improving the computational efficiency and generalization ability of the extreme learning machine on large-scale geological data sets by adopting the conjugate gradient method and sparse matrix decomposition technology, ultimately achieving more accurate and efficient geological and mineral resource prediction.

[0012] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0013] A method for training a geological mineral resource prediction model based on artificial intelligence, which predicts the category of geological mineral resource data by completing the trained prediction model; the training method of the prediction model includes the following steps:

[0014] Step 1: training a data augmentation model, using the trained data augmentation model to augment the geological and mineral resource data, and obtaining a training data set consisting of real geological and mineral resource data and generated geological and mineral resource data; the data augmentation model adopts a generative adversarial network based on belief propagation, including a generator and a discriminator; at the output end of the generator, a belief propagation mechanism is used to optimize the quality of the generated geological and mineral resource data, and the training process of the generator and the discriminator is optimized through a dynamic constraint game strategy;

[0015] Step 2: construct a prediction model through an extreme learning machine based on conjugate gradient, input the geological mineral resource data in the training data set into the prediction model for training, which specifically includes the following steps:

[0016] S21, randomly initialize the hidden layer parameters of the extreme learning machine and calculate the hidden layer output matrix of the extreme learning machine ;

[0017] S22, through the conjugate gradient method and the hidden layer output matrix Calculate the output layer weights of the extreme learning machine;

[0018] S23, using sparse matrix decomposition technology, the hidden layer output matrix Make updates;

[0019] S24, repeating iterative steps S21 to S23 until a preset stop iteration condition is met.

[0020] Furthermore, step S21 specifically includes:

[0021] Randomly select features for each input geological mineral resource data and combine them to obtain feature combinations : , is the random combination generating function;

[0022] Calculate the dynamic weight of the jth feature combination : ; is the weight parameter of the feature combination; It is a function that evaluates the contribution of the j-th feature combination in the k-th hidden layer of the extreme learning machine; is the Sigmoid activation function; represents the total number of hidden layers of the extreme learning machine;

[0023] Initialize the parameters of the extreme learning machine hidden layer. The parameters of the extreme learning machine hidden layer include the hidden layer weights of the extreme learning machine and the hidden layer bias is ; Hidden layer output of extreme learning machine The input geological mineral resource data is converted through a nonlinear activation function:

[0024] ;

[0025] In the formula, Represents the input geological and mineral resource data; Represents the element-wise product operation; It is the scaling coefficient vector of the eigenvalue vector corresponding to all geological and mineral resource data input to the extreme learning machine; is the dynamic weight of the feature combination.

[0026] Furthermore, the scaling coefficient vector corresponding to all features is calculated using an adaptive feature scaling strategy, specifically including: each element in the scaling coefficient vector corresponding to all features , calculated by analyzing the correlation between features and target classification:

[0027] ;

[0028] In the formula, Indicates The correlation coefficient between the feature and the target classification, is a hyperparameter that adjusts the steepness of the response curve.

[0029] Further, in step S22, the conjugate gradient method and the hidden layer output matrix Calculate the output layer weights of the extreme learning machine, including:

[0030] ;

[0031] In the formula, is the output layer weight of the extreme learning machine, Represents a conjugate gradient solver, used to solve linear equations ,in is a symmetric positive definite matrix, the intermediate variable , is the target output matrix, for The transpose of .

[0032] Further, step S23 specifically includes:

[0033] Extract the hidden layer output matrix through singular value decomposition Before The feature corresponding to the largest singular value is the output matrix of the hidden layer. To update:

[0034] ;

[0035] In the formula, Indicates parameter update operation; represents singular value decomposition.

[0036] Furthermore, in step 1, the training data expansion model specifically includes the following steps:

[0037] S11. Initialize the generator Network parameters , Discriminator Network parameters ;

[0038] S12. Enter the adversarial training process: The goal of the generator is to maximize the probability of misjudgment of the discriminator, while the discriminator attempts to distinguish between real geological and mineral resource data and generated geological and mineral resource data. The constraints of the adversarial training process are:

[0039] ;

[0040] In the formula, Represents real geological and mineral resource data, Data distribution representing real geological and mineral resource data; represents the noise vector sampled from a random noise distribution, used to generate simulated geological mineral resource data; represents random noise distribution; express expectations;

[0041] S13. At the output end of the generator, the quality of the generated geological and mineral resource data is optimized by simulating the inherent correlation of the geological and mineral resource data using the belief propagation mechanism, and the generated geological and mineral resource data is adjusted:

[0042] ;

[0043] In the formula, represents the generated geological mineral resource data after confidence propagation adjustment; is the belief propagation algorithm function; is the adjustment coefficient of the belief propagation algorithm;

[0044] S14. Use dynamic constraint game strategy to adjust the strategies of generator and discriminator according to the characteristics of geological and mineral resource data: By adjusting the objective functions of generator and discriminator, the adaptability of data augmentation model to the distribution of geological and mineral resource data is enhanced, and the loss function of data augmentation model is for:

[0045] ;

[0046] In the formula, is the reconstruction error loss between the geological and mineral resource data generated by the generator and the real geological and mineral resource data; represents the additional loss term adopted by the dynamic constraint game strategy, which is used to adjust the strategy and constraints during training. is the first generative adversarial network loss weight, is the loss weight of the second generative adversarial network;

[0047] S15. Repeat iterative steps S12 to S14 until a preset stop iteration condition is met, thereby completing the training of the data expansion model.

[0048] Furthermore, the belief propagation algorithm function is:

[0049] ;

[0050] In the formula, The first one represents the geological mineral resource data generated by the generator Features, is the dimension of geological mineral resource data, is the adjustment coefficient obtained from the covariance matrix of geological and mineral resource data, It is the average value of geological and mineral resource data.

[0051] Furthermore, the additional loss term adopted by the dynamic constraint game strategy is for:

[0052] ;

[0053] In the formula, It represents the entropy of generating geological and mineral resource data and is used to measure the diversity of geological and mineral resource data; It represents the variance of the generated geological and mineral resource data and is used to measure the concentration of geological and mineral resource data; is the first dynamic constraint coefficient, is the second dynamic constraint coefficient.

[0054] Furthermore, it also includes:

[0055] During the generator training process, the energy state of the generated geological and mineral resource data is adjusted by simulating the principle of energy conservation: in each generation process, the generated geological and mineral resource data adjusted by confidence propagation is The energy state is defined as The norm of :

[0056] ;

[0057] In the formula, represents the L2 norm; for energy state;

[0058] Regulate by setting target energy The target energy is the mean value of the energy distribution based on the real geological mineral resource data. and standard deviation The calculation results are:

[0059] ;

[0060] In the formula, is the target energy, is a random variable from a standard normal distribution;

[0061] The loss function of the data augmentation model Replace with:

[0062] ;

[0063] In the formula, is the energy conservation loss term.

[0064] Furthermore, the energy conservation loss term Used to generate geological mineral resource data after confidence propagation adjustment The energy state is close to the target energy to adjust Energy status:

[0065] .

[0066] Compared with the prior art, the beneficial technical effects of the present invention are:

[0067] 1. Generative Adversarial Network Based on Belief Propagation:

[0068] The traditional generative adversarial network algorithm mainly relies on the adversarial training of the generator and the discriminator. The present invention adopts the belief propagation mechanism, which can simulate the intrinsic correlation of geological and mineral resource data, and adjust the generated samples through the covariance matrix to make them more consistent with the distribution characteristics of actual geological and mineral resources. The present invention enhances the fidelity and consistency of the generated samples through belief propagation, making the generated virtual samples closer to the real data, thereby improving the quality of data expansion and the accuracy of the prediction model.

[0069] 2. Energy conservation mechanism is used to generate sample quality control:

[0070] The present invention uses a mechanism based on the energy conservation principle in the generative adversarial network to control the energy state of the generated samples. By defining the energy of the generated samples and controlling it to be consistent with the energy distribution of the real samples during training, it is ensured that the generated samples maintain a reasonable distribution in the feature space. This mechanism effectively prevents the excessive concentration or dispersion of the generated samples, ensures the diversity and reasonable distribution of the generated samples, and thus enhances the generalization ability of the model in a complex data environment.

[0071] 3. Dynamic constraint game strategy:

[0072] Traditional generative adversarial networks often have difficulty adapting to complex geological data distributions in adversarial training between generators and discriminators. The present invention adopts a dynamic constraint game strategy to dynamically adjust the strategies of the generator and discriminator during the training process to enhance adaptability to complex data. By constraining the diversity and quality of generated samples, the dynamic constraint game strategy improves the robustness of the model, avoids the problem of overly consistent or unbalanced distribution of generated samples, and effectively improves the generalization ability and accuracy of the model.

[0073] 4. Extreme Learning Machine Based on Conjugate Gradient:

[0074] As a single-layer feedforward neural network, the extreme learning machine can quickly process a large amount of data. The present invention optimizes the output layer weight calculation of the extreme learning machine by adopting the conjugate gradient method, thereby improving the training efficiency and prediction accuracy of the extreme learning machine. The application of the conjugate gradient method greatly reduces the calculation time and improves the model's adaptability to large-scale geological and mineral resource data, which is suitable for complex and high-dimensional data sets.

[0075] 5. Sparse matrix decomposition technology:

[0076] Traditional neural networks often encounter computational bottlenecks when processing large-scale data. The present invention adopts sparse matrix decomposition technology to reduce feature dimensions and reduce model computational complexity. Sparse matrix decomposition helps the model remain efficient and accurate when processing large-scale data, and reduces the demand for storage and computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 Schematic diagram of the flow of the prediction model training method in an embodiment of the present invention. DETAILED DESCRIPTION

[0078] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0079] The present invention proposes a geological mineral resource prediction model training method based on artificial intelligence. The training of the prediction model is realized through training data, and the training data includes real geological mineral resource data and generated geological mineral resource data. The real geological mineral resource data is obtained by combining remote sensing technology with actual ground exploration, and specifically includes satellite images, ground sample point analysis results and historical geological survey data. The data are collected through data acquisition equipment and software tools, and transmitted to a central data processing center through a unified data interface.

[0080] In a preferred embodiment, the attributes of geological mineral resource data include:

[0081] Rock type Ra, indicating the rock composition of the site;

[0082] Density Da, which represents the density information of geological samples;

[0083] Metal content Aa, indicating the content of a specific metal;

[0084] Saturation Sa, represents the saturation of groundwater;

[0085] Magnetic Ma, which indicates the magnetic characteristics of geological samples;

[0086] Conductivity Ca, which indicates the conductivity of the geological sample;

[0087] Pressure Pa, indicating underground pressure data;

[0088] Temperature Ta, indicating the temperature of the measurement point;

[0089] Biomarker Ba, represents the biomarker in geological samples;

[0090] Seismicity Ea represents the seismicity data of the region.

[0091] It should be noted that the above-mentioned attributes of geological and mineral resource data are only used to illustrate one data format and type of the present invention. In practical applications, the attributes of geological and mineral resource data are usually more than 10, and can reach dozens or even hundreds.

[0092] Furthermore, the categories of the collected geological and mineral resource data are marked. The marking method of the present invention is manual marking. In a preferred embodiment, the manual marking includes three categories:

[0093] Category 1: Metal minerals, such as gold, silver, copper, etc.;

[0094] Category 2: Non-metallic minerals, such as limestone, stone coal, etc.;

[0095] Category three: Energy-type minerals, such as coal, oil, etc.

[0096] It is understandable that in the task of geological and mineral resource category prediction, the acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples easily lead to poor generalization ability of the prediction model, while affecting the accuracy of the prediction model. The present invention adopts a generative adversarial network based on belief propagation for sample generation, thereby realizing geological and mineral resource data expansion. The traditional generative adversarial network consists of a generator and a discriminator, wherein the generator is responsible for generating sufficiently realistic geological and mineral resource data, and the discriminator attempts to distinguish between the generated geological and mineral resource data and the real geological and mineral resource data. The present invention adopts a belief propagation mechanism so that the generated geological and mineral resource data better reflects the intrinsic distribution characteristics of geological and mineral resource data. In addition, in order to cope with the high nonlinearity and complexity in geological and mineral resource data, the present invention also uses a dynamic constraint game strategy to optimize the training process of the generator and the discriminator, thereby enhancing the geological and mineral resource data expansion capability and fidelity of the model.

[0097] Specifically, the training process of the generative adversarial network based on belief propagation adopted by the data augmentation model is as follows:

[0098] 1. Initialize the network parameters of the generator and discriminator. Let the generator be , the discriminator is , the network parameters of the generator are , the network parameters of the discriminator are In a preferred embodiment, the initialization method of the generator and the discriminator is expressed as:

[0099] ;

[0100] ;

[0101] In the formula, represents a normal distribution, It means the mean is 0 and the variance is Normal distribution of Indicates that it obeys a specific distribution.

[0102] 2. In the adversarial training stage, the generator and the discriminator enter a game process. The generator tries to generate more and more realistic geological and mineral resource data, while the discriminator strives to distinguish which are real geological and mineral resource data (real samples) and which are generated geological and mineral resource data (generated samples). That is, the goal of the generator is to maximize the probability of the discriminator's misjudgment, while the discriminator tries to accurately distinguish between real samples and generated samples. The constraints of this process are expressed as:

[0103] ;

[0104] In the formula, represents the real sample, Represents the data distribution of real samples; represents the noise vector sampled from a random noise distribution, used to generate simulated samples; represents random noise distribution; represents the discriminator, Represents a generator; Express expectations.

[0105] 3. In order to make the generated geological and mineral resource data more realistic and statistically consistent, the present invention uses a belief propagation mechanism at the output end of the generator to simulate the inherent correlation of the geological and mineral resource data to optimize the quality of the generated geological and mineral resource data, so as to optimize the consistency and fidelity of the generated samples. The way of adjusting the generated samples is expressed as:

[0106] ;

[0107] In the formula, represents the generated samples after belief propagation adjustment; It is a belief propagation algorithm function, which is used to enhance the intrinsic connection and fidelity of the generated samples; is the adjustment coefficient of the belief propagation algorithm. Preferably, Set to 0.3.

[0108] In a preferred embodiment, the belief propagation algorithm function reflects the deviation of each sample point from the average sample point, and adjusts this deviation through the covariance matrix to enhance the consistency of geological mineral resource data. As the input of the function, the implementation of the belief propagation algorithm function is expressed as:

[0109] ;

[0110] In the formula, is the dimension of geological mineral resource data (sample), is the adjustment coefficient obtained from the covariance matrix of geological and mineral resource data, is the sample mean.

[0111] 4. In the training process of the generator of the generative adversarial network, a quality control mechanism for generated samples based on energy conservation is adopted. By simulating the energy conservation principle in physics, the energy state of the generated samples is adjusted to control the quality and diversity of the samples and ensure that the generated samples maintain a reasonable distribution in the feature space. Specifically, in each generation process, the energy of the generated sample is defined as the norm of the generated sample adjusted by belief propagation, which reflects the activity and complexity of the sample and is expressed as:

[0112] ;

[0113] In the formula, represents the L2 norm; is the energy function for generating samples.

[0114] Furthermore, the energy state of the generated samples is adjusted by setting the target energy. The target energy is dynamically calculated based on the mean and standard deviation of the energy distribution of the real samples to keep the activity and complexity of the generated samples matching the real samples. The calculation method is:

[0115] ;

[0116] In the formula, is the target energy; and are the mean and standard deviation of the true sample energy distribution, is a random variable from a standard normal distribution. Preferably, Set to 0.5.

[0117] 5. Use dynamic constraint game strategy to adjust the strategies of generator and discriminator according to the characteristics of geological and mineral resource data; specifically, by adjusting the objective functions of generator and discriminator, enhance the adaptability of the model to the distribution of complex geological and mineral resource data, and generate the loss function of adversarial network. for:

[0118] ;

[0119] In a preferred embodiment, the loss function of the generative adversarial network is for:

[0120] ;

[0121] In the formula, is the reconstruction error loss between the geological and mineral resource data generated by the generator and the real geological and mineral resource data; represents the additional loss term adopted by the dynamic constraint game strategy, which is used to adjust the strategy and constraints during training. is the first generative adversarial network loss weight, is the loss weight of the second generative adversarial network; is the energy conservation loss term. Preferably, Set to 0.1, Set to 0.4.

[0122] In a preferred embodiment, the additional loss term used in the dynamic constraint game strategy includes two parts, one part emphasizes sample diversity, and the other part emphasizes sample quality. It not only considers the diversity of generated samples, but also emphasizes that samples should not be too concentrated or consistent, thereby improving the adaptability and generalization ability of the model. The calculation method is:

[0123] ;

[0124] In the formula, Represents the entropy of generated samples, which is used to measure sample diversity; Represents the variance of the generated samples, which is used to measure the concentration of the samples; is the first dynamic constraint coefficient, is the second dynamic constraint coefficient. Preferably, Set to 0.35, Set to 0.1.

[0125] In this embodiment, the energy conservation loss term ensures that the energy of the generated sample is close to the target energy, so as to adjust the energy state of the generated sample to make it tend to be stable and reasonably distributed. The calculation method is:

[0126] ;

[0127] The generated sample quality control mechanism based on energy conservation can effectively control the quality and diversity of generated samples and prevent over-concentrated or over-discrete sample distribution.

[0128] 6. Repeat the above steps until the preset stop iteration condition is met, indicating that the data augmentation model training is completed. In a preferred embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0129] After the data expansion model training is completed, the trained data expansion model is used to increase the number of geological mineral resource data. In a preferred embodiment, assuming that the original collected geological mineral resource data is 800, and the data expansion model is expanded to generate 200 samples, the expanded training data set contains 1000 samples.

[0130] Furthermore, the expanded training data set is input into the prediction model for geological mineral resource prediction. The prediction model adopts an extreme learning machine based on conjugate gradient. The extreme learning machine is an efficient single-layer feedforward neural network, which is characterized by randomly initializing the hidden layer weights without adjustment, and only calculating the output layer weights. Based on the traditional extreme learning machine, the present invention optimizes the output layer weight calculation of the extreme learning machine by adopting the conjugate gradient method, which not only improves the training speed of the algorithm, but also enhances the generalization ability of the prediction model. In addition, the present invention combines sparse matrix decomposition technology to process large-scale geological mineral resource data sets, ensuring that the classifier can still maintain high efficiency and accuracy when facing a large number of features and samples. Sparse matrix decomposition helps reduce the complexity of the model and reduces the demand for storage and computing resources.

[0131] Specifically, Figure 1 As shown, the training process of the prediction model using the extreme learning machine based on conjugate gradient includes the following steps:

[0132] S1. Initialize the parameters of the hidden layer of the extreme learning machine. Set the hidden layer weight of the extreme learning machine to , the hidden layer bias of the extreme learning machine is , the hidden layer output matrix of the extreme learning machine is , the hidden layer output matrix of the extreme learning machine is obtained by transforming the input geological mineral resources data through a nonlinear activation function, which is expressed as:

[0133] ;

[0134] In the formula, represents the geological mineral resource data input to the extreme learning machine, is the Sigmoid activation function; represents the element-wise product operation, It is the scaling coefficient vector of the eigenvalue vector corresponding to all geological and mineral resource data input to the extreme learning machine; is the dynamic weight of the feature.

[0135] In a preferred embodiment, the scaling coefficient vector corresponding to all features is calculated using an adaptive feature scaling strategy, and the scaling ratio of each feature is dynamically adjusted to maximize the response sensitivity and classification accuracy of the classifier. Specifically, each element of the scaling coefficient vector corresponding to all features is , calculated by analyzing the correlation between the features and the target classification, expressed as:

[0136] ;

[0137] In the formula, Indicates The correlation coefficient between the feature and the target classification, is a hyperparameter that adjusts the steepness of the response curve.

[0138] Among them, the target classification refers to the category corresponding to the geological and mineral resource data to be classified.

[0139] In this embodiment, a dynamic weight adjustment mechanism based on random feature combination is adopted to enhance the performance and generalization ability of the model when processing complex geological mineral resource data by randomly combining input features and dynamically adjusting their weights in the model. Specifically, features are randomly selected and combined for each input sample, and each feature combination generates a new feature vector through a combination function, which is expressed as:

[0140] ;

[0141] In the formula, It is a random combination generation function, which can be implemented as a weighted sum of features, polynomial or other nonlinear functions.

[0142] Furthermore, the corresponding dynamic weight is calculated for each feature combination, and the weight is dynamically adjusted based on the contribution of the feature combination to the output of the prediction model, which is expressed as:

[0143] ;

[0144] In the formula, is the feature combination weight parameter, which is a training parameter, that is, obtained through training; It is a function that evaluates the contribution of the j-th feature combination in the k-th hidden layer of the extreme learning machine, which is calculated through cross-validation.

[0145] S2. Calculate the output layer weight of the extreme learning machine. The output layer weight of the extreme learning machine is achieved by minimizing the output error. In the traditional extreme learning machine, the Moore-Penrose pseudo-inverse is used to solve the output layer weight of the extreme learning machine. The present invention uses the conjugate gradient method to solve the output layer weight of the extreme learning machine, which improves the solution efficiency and stability, and is expressed as:

[0146] ;

[0147] In the formula, Indicates that the extreme learning machine obtains the output layer weight; Represents a conjugate gradient solver, used to solve linear equations ,in is a symmetric positive definite matrix, , is the target output matrix, for The transpose of .

[0148] The conjugate gradient method is an iterative solver that is suitable for optimization problems of large-scale geological and mineral resource data, especially when dealing with sparse systems. In each iteration, the method attempts to find the optimal weight update direction, thereby converging to the minimum error solution more quickly, greatly speeding up the training process. In a preferred embodiment, the iterative method of the conjugate gradient solver is expressed as:

[0149] ;

[0150] ;

[0151] ;

[0152] In the formula, It is The solution of the iteration, It is The solution of the iteration; It is The scalar step size for the iteration, It is The scalar step size of the iteration; It is The scalar direction of the iteration, It is The scalar direction of the iteration; It is The update parameter for the scalar step size of the iterations; It is The update parameter in the scalar direction of the iteration.

[0153] At the first iteration, the search direction is initially set to:

[0154] ;

[0155] In the formula, , ; is the initial weight vector, usually set to zero.

[0156] Furthermore, the scalar step size and direction update parameters are calculated as:

[0157] ;

[0158] ;

[0159] Based on this, the calculation of the output layer weights of the extreme learning machine is not only accurate, but also suitable for large-scale and high-dimensional geological and mineral resources data processing.

[0160] S3, using sparse matrix decomposition technology to simplify model calculations, update the hidden layer output matrix, the calculation method is:

[0161] ;

[0162] In the formula, Indicates parameter update operation; Represents singular value decomposition, which is used to extract the hidden layer output matrix Before The features corresponding to the largest singular values ​​are obtained, thereby reducing the computational complexity and retaining the most important geological and mineral resource data features.

[0163] S4, repeat the above steps until the preset stop iteration condition is met, indicating that the prediction model training is completed. In a preferred embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0164] After the prediction model training is completed, the trained prediction model is used to process new samples. Specifically, the newly collected geological and mineral resource data are input into the trained prediction model to obtain the predicted category, and the predicted category is the same as the manually labeled category.

[0165] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0166] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A geological mineral resources prediction model training method based on artificial intelligence, characterized in that: The category of geological mineral resource data is predicted by the trained prediction model; the training method of the prediction model includes the following steps: Step 1: training a data augmentation model, using the trained data augmentation model to augment the geological and mineral resource data, and obtaining a training data set consisting of real geological and mineral resource data and generated geological and mineral resource data; the data augmentation model adopts a generative adversarial network based on belief propagation, including a generator and a discriminator; at the output end of the generator, a belief propagation mechanism is used to optimize the quality of the generated geological and mineral resource data, and the training process of the generator and the discriminator is optimized through a dynamic constraint game strategy; Step 2: construct a prediction model through an extreme learning machine based on conjugate gradient, input the geological mineral resource data in the training data set into the prediction model for training, which specifically includes the following steps: S21, randomly initialize the hidden layer parameters of the extreme learning machine and calculate the hidden layer output matrix of the extreme learning machine ; S22, through the conjugate gradient method and the hidden layer output matrix Calculate the output layer weights of the extreme learning machine; S23, using sparse matrix decomposition technology, the hidden layer output matrix Make updates; S24, repeating iterative steps S21 to S23 until a preset stop iteration condition is met.

2. The method for training a geological mineral resource prediction model based on artificial intelligence according to claim 1, characterized in that: Step S21 specifically includes: Randomly select features for each input geological mineral resource data and combine them to obtain feature combinations : , is the random combination generating function; Calculate the dynamic weight of the jth feature combination : ; is the weight parameter of the feature combination; It is a function that evaluates the contribution of the j-th feature combination in the k-th hidden layer of the extreme learning machine; is the Sigmoid activation function; represents the total number of hidden layers of the extreme learning machine; Initialize the parameters of the extreme learning machine hidden layer. The parameters of the extreme learning machine hidden layer include the hidden layer weights of the extreme learning machine and the hidden layer bias is ; Hidden layer output of extreme learning machine The input geological mineral resource data is converted through a nonlinear activation function: ； In the formula, Represents the input geological and mineral resource data; Represents the element-wise product operation; It is the scaling coefficient vector of the eigenvalue vector corresponding to all geological and mineral resource data input to the extreme learning machine; is the dynamic weight of the feature combination.

3. The method for training a geological mineral resource prediction model based on artificial intelligence according to claim 2, characterized in that: The scaling coefficient vector corresponding to all features is calculated using an adaptive feature scaling strategy, specifically including: each element in the scaling coefficient vector corresponding to all features , calculated by analyzing the correlation between features and target classification: ； In the formula, Indicates The correlation coefficient between the feature and the target classification, is a hyperparameter that adjusts the steepness of the response curve.

4. The method for training a geological mineral resource prediction model based on artificial intelligence according to claim 1, characterized in that: In step S22, the conjugate gradient method and the hidden layer output matrix Calculate the output layer weights of the extreme learning machine, including: ； In the formula, is the output layer weight of the extreme learning machine, Represents a conjugate gradient solver, used to solve linear equations ,in is a symmetric positive definite matrix, the intermediate variable , is the target output matrix, for The transpose of .

5. The artificial intelligence-based geological mineral resource prediction model training method according to claim 1 is characterized in that: Step S23 specifically includes: Extract the hidden layer output matrix through singular value decomposition Before The feature corresponding to the largest singular value is the output matrix of the hidden layer. To update: ； In the formula, Indicates parameter update operation; represents singular value decomposition.

6. The artificial intelligence-based geological mineral resources prediction model training method according to claim 1 is characterized in that: In step 1, the training data expansion model specifically includes the following steps: S11. Initialize the generator Network parameters , Discriminator Network parameters ; S12. Enter the adversarial training process: The goal of the generator is to maximize the probability of misjudgment of the discriminator, while the discriminator attempts to distinguish between real geological and mineral resource data and generated geological and mineral resource data. The constraints of the adversarial training process are: ； In the formula, Represents real geological and mineral resource data, Data distribution representing real geological and mineral resource data; represents the noise vector sampled from a random noise distribution, used to generate simulated geological mineral resource data; represents random noise distribution; express expectations; S13. At the output end of the generator, the quality of the generated geological and mineral resource data is optimized by simulating the inherent correlation of the geological and mineral resource data using the belief propagation mechanism, and the generated geological and mineral resource data is adjusted: ； In the formula, represents the generated geological mineral resource data after confidence propagation adjustment; is the belief propagation algorithm function; is the adjustment coefficient of the belief propagation algorithm; S14. Use dynamic constraint game strategy to adjust the strategies of generator and discriminator according to the characteristics of geological and mineral resource data: By adjusting the objective functions of generator and discriminator, the adaptability of data augmentation model to the distribution of geological and mineral resource data is enhanced, and the loss function of data augmentation model is for: ； In the formula, is the reconstruction error loss between the geological and mineral resource data generated by the generator and the real geological and mineral resource data; represents the additional loss term adopted by the dynamic constraint game strategy, which is used to adjust the strategy and constraints during training. is the first generative adversarial network loss weight, is the loss weight of the second generative adversarial network; S15. Repeat iterative steps S12 to S14 until a preset stop iteration condition is met, thereby completing the training of the data expansion model.

7. The method for training a geological mineral resource prediction model based on artificial intelligence according to claim 6, characterized in that: The belief propagation algorithm function is: ； In the formula, The first one represents the geological mineral resource data generated by the generator Features, is the dimension of geological mineral resource data, is the adjustment coefficient obtained from the covariance matrix of geological and mineral resource data, It is the average value of geological and mineral resource data.

8. The method for training a geological mineral resource prediction model based on artificial intelligence according to claim 6, characterized in that: The additional loss term adopted by the dynamic constraint game strategy for: ； In the formula, It represents the entropy of generating geological and mineral resource data and is used to measure the diversity of geological and mineral resource data; It represents the variance of the generated geological and mineral resource data and is used to measure the concentration of geological and mineral resource data; is the first dynamic constraint coefficient, is the second dynamic constraint coefficient.

9. The method for training a geological mineral resource prediction model based on artificial intelligence according to claim 6, characterized in that: Also includes: During the generator training process, the energy state of the generated geological and mineral resource data is adjusted by simulating the principle of energy conservation: in each generation process, the generated geological and mineral resource data adjusted by confidence propagation is The energy state is defined as The norm of : ； In the formula, represents the L2 norm; for energy state; Regulate by setting target energy The target energy is the mean value of the energy distribution based on the real geological mineral resource data. and standard deviation The calculation results are: ； In the formula, is the target energy, is a random variable from a standard normal distribution; The loss function of the data augmentation model Replace with: ； In the formula, is the energy conservation loss term.

10. The artificial intelligence-based geological mineral resource prediction model training method according to claim 9, characterized in that: The energy conservation loss term Used to generate geological mineral resource data after confidence propagation adjustment The energy state is close to the target energy to adjust Energy status: 。

Images