Solid mineral prospecting prediction method and system based on geological big data

Abstract

The application relates to a solid mineral prospecting prediction method and system based on geological big data. The method comprises the following steps: performing three-dimensional space discretization and attribute assignment processing on multi-source geological big data to obtain a multi-dimensional feature vector of each three-dimensional voxel unit, and calculating a geological rule prior score of each three-dimensional voxel unit based on a formalized knowledge model of metallogenic geology theory; based on a multi-objective optimization algorithm, training and solving a Pareto frontier of a deep neural network prediction model through the multi-dimensional feature vector and the geological rule prior score to obtain a prospecting prediction model set; performing parallel prospecting target area prediction on a target exploration area based on all the Pareto optimal models in the prospecting prediction model set to obtain a plurality of three-dimensional metallogenic probability bodies, and fusing the three-dimensional metallogenic probability bodies to obtain a prospecting prediction pedigree; and based on a dynamic threshold, delineating a prospecting target area in the target exploration area through the prospecting prediction pedigree to obtain a final prospecting prediction scheme. The method can provide a robust prospecting prediction result.

Description

Technical Field

[0001] This invention belongs to the field of mineral development technology, and in particular relates to a method and system for predicting solid mineral exploration based on geological big data. Background Technology

[0002] With the continuous accumulation of geological big data in the field of solid mineral exploration and prediction, data-driven mineral exploration and prediction methods have emerged. These methods can autonomously mine potential correlations from massive amounts of data and efficiently process complex, high-dimensional, and heterogeneous data, providing a brand-new technical path for mineral exploration and prediction.

[0003] However, while data-driven mineral exploration prediction methods such as deep learning compensate for the difficulty of traditional methods in fully mining the complex relationships hidden in geological big data, resulting in insufficient comprehensiveness and accuracy of predictions, the nonlinear relationships they uncover often deviate from basic geological principles, easily generating a large number of geologically inexplicable anomalies and causing ineffective waste of exploration costs. Existing solutions that attempt to integrate expert knowledge with data-driven approaches only achieve coupling through rule constraints or feature splicing, failing to resolve the inherent conflict between the deterministic laws of expert knowledge and the exploration of unknown relationships by data-driven approaches. Furthermore, single-objective model training frameworks cannot quantify and balance this conflict, still requiring a large amount of manual trial and error, and are difficult to provide multi-dimensional decision support. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and system for solid mineral exploration prediction based on geological big data that can make explicit and quantify the conflict between a priori geological laws and data-driven exploration, and achieve an adaptive, interpretable, and selectable balance during model training, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for predicting solid mineral deposits based on geological big data, including:

[0006] The multi-source geological big data of the target exploration area is discretized in three dimensions and attribute assigned to obtain the multi-dimensional feature vector of each three-dimensional voxel unit. Based on the formal knowledge model of metallogenic geology theory, the prior score of geological law of each three-dimensional voxel unit is calculated.

[0007] Based on a multi-objective optimization algorithm, a set of mineral exploration prediction models is obtained by training a deep neural network prediction model and solving the Pareto front using multi-dimensional feature vectors and geological law prior scores. The multi-objective optimization algorithm aims to minimize the geological law conformity loss and maximize the information entropy loss.

[0008] Based on all Pareto optimal models in the mineral exploration prediction model set, parallel mineral exploration target area prediction is performed in the target exploration area to obtain multiple three-dimensional mineralization probability volumes. All three-dimensional mineralization probability volumes are then fused to obtain a mineral exploration prediction spectrum. The mineral exploration prediction spectrum includes a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume.

[0009] Based on dynamic thresholds, the target exploration area is delineated using a mineral exploration prediction spectrum to obtain the final mineral exploration prediction scheme.

[0010] In one embodiment, based on a multi-objective optimization algorithm, a deep neural network prediction model is trained and its Pareto front is solved using multi-dimensional feature vectors and prior scores of geological laws to obtain a set of mineral exploration prediction models, including:

[0011] Based on a deep neural network prediction model, the prospecting target area is predicted for each three-dimensional voxel unit by using multi-dimensional feature vectors, and the prediction probability of each three-dimensional voxel unit is obtained.

[0012] Based on the prior score of geological laws and the predicted probability, the KL divergence of the probability distribution is calculated to obtain the geological law conformity loss.

[0013] Spatial conditional entropy is calculated based on the prediction probability of each three-dimensional voxel unit, and information entropy loss is obtained.

[0014] With the optimization objectives of minimizing the loss of geological regularity conformity and minimizing the negative value of information entropy loss, the model population is iteratively evolved using a non-dominated sorting genetic algorithm. When the iterative evolution process reaches the target termination condition, all deep neural network prediction models in the first non-dominated layer of the final model population are output to obtain the mineral exploration prediction model set. The model population includes multiple deep neural network prediction models with different parameter sets. The iterative evolution corresponds to evaluating the optimization target value of each deep neural network prediction model in each generation, and performing non-dominated sorting and crowding calculation based on Pareto dominance relationship, so as to obtain a new generation of population through selection, crossover and mutation of individuals.

[0015] In one embodiment, the KL divergence of the probability distribution is calculated based on the prior score of geological laws and the predicted probability to obtain the geological law conformity loss, including:

[0016] The continuous range of geological law prior scores is evenly divided into multiple intervals, and the frequency of geological law prior scores of each three-dimensional voxel unit falling into each interval is counted to obtain the geological prior probability distribution.

[0017] The model's predicted probability distribution is obtained based on the frequency of predicted probabilities falling into each interval.

[0018] Based on the prior geological probability distribution and the model prediction probability distribution, the geological law conformity loss is obtained; the formula for calculating the geological law conformity loss is as follows: ,in, The predicted probability distribution in the model is at the th... The frequency of each interval The geological prior score distribution in the geological prior probability distribution is the th The frequency of each interval The total number of intervals, This is the parameter set for a deep neural network prediction model.

[0019] In one embodiment, the spatial conditional entropy is calculated based on the predicted probability of each three-dimensional voxel unit to obtain the information entropy loss, including:

[0020] The fundamental component term is obtained by calculating the negative of the sum of pointwise binary cross-entropy based on the predicted probabilities of all 3D voxel units; the expression for the fundamental component term is: ,in, It is a three-dimensional voxel unit. To predict probabilities;

[0021] Based on the total variation of the predicted probabilities of all 3D voxel units on the 3D spatial mesh, a spatial smoothness regularization term is generated; the expression for the spatial smoothness regularization term is: ;

[0022] The information entropy loss is obtained by fusing the basic component terms and the spatial smoothness regularization term according to the weight coefficients.

[0023] In one embodiment, parallel mineral exploration target area prediction is performed on the target exploration area based on all Pareto optimal models in the mineral exploration prediction model set, resulting in multiple three-dimensional mineralization probability volumes. These three-dimensional mineralization probability volumes are then fused to obtain a mineral exploration prediction spectrum, including:

[0024] The average of the predicted probability values ​​of each three-dimensional mineralization probability volume on the same three-dimensional voxel unit is calculated to obtain the three-dimensional consensus probability volume.

[0025] The standard deviation of the predicted probability values ​​of each three-dimensional mineralization probability volume on the same three-dimensional voxel unit is calculated to obtain the three-dimensional uncertainty quantification volume.

[0026] From the set of mineral exploration prediction models, the three-dimensional mineralization probability volume corresponding to the Pareto optimal model with the smallest loss of geological law conformity is determined as the conservative prediction volume, and the three-dimensional mineralization probability volume corresponding to the Pareto optimal model with the smallest loss of information entropy is determined as the aggressive prediction volume.

[0027] The three-dimensional consensus probability body, the three-dimensional uncertainty quantification body, the conservative prediction body, and the radical prediction body are integrated into a mineral exploration prediction system.

[0028] In one embodiment, based on a dynamic threshold, a mineral exploration target area is delineated using a mineral exploration prediction spectrum to obtain a final mineral exploration prediction scheme, including:

[0029] The three-dimensional consensus probability volume and the three-dimensional uncertainty quantification volume in the mineral exploration prediction spectrum are loaded into the three-dimensional visualization interactive platform, and spatial coordinate registration and overlay rendering are performed with the pre-stored basic geological layer data to obtain a three-dimensional comprehensive prediction view.

[0030] In response to the acquisition of the dynamic probability threshold parameter, the set of voxel space coordinates of all three-dimensional voxel units in the three-dimensional consensus probability volume that satisfy the average value ≥ the dynamic probability threshold parameter is extracted, and three-dimensional spatial clustering analysis is performed based on the connectivity of the voxel space coordinate set to obtain the candidate abnormal three-dimensional entity model.

[0031] In response to the acquisition of the dynamic uncertainty threshold parameter, all three-dimensional voxel units in the three-dimensional uncertainty quantization volume that satisfy the standard deviation > the dynamic uncertainty threshold parameter are marked as high uncertainty exploration areas, and the high uncertainty exploration areas are superimposed and displayed in the three-dimensional comprehensive prediction view using a preset rendering method.

[0032] By comparing the conservative predictive volume and the radical probabilistic volume on a three-dimensional voxel basis, the difference probabilistic volume is obtained;

[0033] Threshold segmentation and spatial clustering are performed on the differential probability volumes, and high-value clustering regions unique to the radical probability volumes are marked as prospective exploration candidate regions;

[0034] Based on the candidate anomaly 3D entity model, 3D comprehensive prediction view, and prospective exploration candidate area, spatial overlay analysis is performed in conjunction with preset target area fusion rules to generate a final mineral exploration prediction scheme. The final mineral exploration prediction scheme includes the category, 3D spatial coordinate range, confidence level, and associated geological attribute information of each mineral exploration target area.

[0035] In one embodiment, the method further includes:

[0036] Positive sample labels are assigned to mineral-bearing locations verified in the field engineering of the final mineral exploration prediction scheme, and negative sample labels are assigned to non-mineral-bearing locations. A new verification sample set is constructed by combining the spatial coordinates of the three-dimensional voxel units corresponding to each label location and the multi-dimensional feature vector. The new verification sample set is used to train the deep neural network prediction model.

[0037] The systematic deviation is calculated based on the newly added verification sample set and the prior score of geological laws. If the prior score of geological laws exceeds the preset threshold in multiple consecutive verification periods, the corresponding three-dimensional voxel unit is a positive sample label, triggering an update of the formal knowledge model of metallogenic geology. The update of the formal knowledge model of metallogenic geology corresponds to the correction of the rule confidence, the conditional probability table of the probabilistic graphical model, and the quantization function parameters of the prediction elements in the formal knowledge model of metallogenic geology.

[0038] Secondly, this application also provides a solid mineral exploration prediction system based on geological big data, including:

[0039] The prior module is used to perform three-dimensional spatial discretization and attribute assignment processing on multi-source geological big data of the target exploration area to obtain the multi-dimensional feature vector of each three-dimensional voxel unit, and calculate the geological law prior score of each three-dimensional voxel unit based on the formal knowledge model of metallogenic geology theory.

[0040] The balancing module is used to train and solve the Pareto front of a deep neural network prediction model based on a multi-objective optimization algorithm, using multi-dimensional feature vectors and prior scores of geological laws, to obtain a set of mineral exploration prediction models. The multi-objective optimization algorithm aims to minimize the geological law conformity loss and maximize the information entropy loss.

[0041] The prediction module is used to perform parallel mineral exploration target area prediction based on all Pareto optimal models in the mineral exploration prediction model set, obtain multiple three-dimensional mineralization probability volumes, and fuse all three-dimensional mineralization probability volumes to obtain a mineral exploration prediction spectrum; the mineral exploration prediction spectrum includes a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume.

[0042] The targeting module is used to delineate the target exploration area based on dynamic thresholds and mineral exploration prediction spectrum, and obtain the final mineral exploration prediction scheme.

[0043] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned solid mineral exploration and prediction methods based on geological big data.

[0044] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described solid mineral exploration and prediction methods based on geological big data.

[0045] The aforementioned method and system for solid mineral exploration prediction based on geological big data discretizes multi-source geological big data into three-dimensional voxel units and integrates mineralization theories to construct a priori scores of geological laws as knowledge constraints. Then, a multi-objective optimization algorithm drives a deep neural network prediction model to simultaneously minimize the loss of geological law conformity and maximize the loss of information entropy during the training process. This results in a set of Pareto optimal models representing different trade-off strategies between knowledge adherence and data exploration. Based on this set of Pareto optimal models, parallel predictions are performed on all voxels to generate multiple three-dimensional mineralization probability volumes, which are then integrated into a prediction spectrum that includes consensus probability and uncertainty quantification. By applying dynamic thresholds to the prediction spectrum, the precise delineation of mineral exploration target areas is achieved. This allows for the proactive mining of unconventional mineralization information hidden in the data while strictly adhering to the basic framework of geological mineralization laws, and provides robust and interpretable mineral exploration prediction results by quantifying uncertainty. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying 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.

[0047] Figure 1 This is a flowchart illustrating the solid mineral exploration and prediction method based on geological big data of the present invention.

[0048] Figure 2 This is a structural diagram of the solid mineral exploration and prediction system based on geological big data of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] In one embodiment, such as Figure 1 As shown, a method for predicting solid mineral deposits based on geological big data is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S101. Perform three-dimensional spatial discretization and attribute assignment processing on the multi-source geological big data of the target exploration area to obtain the multi-dimensional feature vector of each three-dimensional voxel unit, and calculate the geological law prior score of each three-dimensional voxel unit based on the formal knowledge model of metallogenic geology theory.

[0052] Indicatively, based on a clearly defined spatial boundary of the target exploration area, which can be delineated according to the latitude, longitude, and elevation ranges specified in the exploration task statement, the continuous three-dimensional geological space is divided into a regular three-dimensional voxel grid. The resolution of the voxel grid can be adjusted according to the exploration accuracy requirements of the target exploration area, and each three-dimensional voxel unit has unique spatial coordinates. ,in , , These represent the grid indices of voxels along the x, y, and z axes, respectively. The multi-source geological big data encompasses remote sensing data, geophysical data, geochemical data, and geological survey data. For each three-dimensional voxel unit, various data attribute values ​​corresponding to its spatial location are extracted. For example, these data attribute values ​​include multi-band reflectance from remote sensing data, gravity and magnetic anomalies from geophysical data, ore-forming element content from geochemical data, and lithology and structural type codes from geological survey data. Optionally, for voxel units with missing data in some spatial locations, kriging interpolation or neighborhood voxel attribute mean filling is used to complete the attribute values. All completed attribute values ​​are then integrated according to a preset dimension to form a multi-dimensional feature vector for that voxel unit. ,in, This represents the total number of feature dimensions of multi-source geological data.

[0053] Based on a formal knowledge model of metallogenic geology, the prior scores of geological laws for each three-dimensional voxel unit are calculated. This formal knowledge model transforms classic metallogenic geology theories such as igneous rock mineralization, tectonic control of ore formation, and stratigraphic control of ore formation into a quantifiable mathematical model. The model includes quantitative rules for various favorable conditions for mineralization. For example, the formula for calculating the prior scores of geological laws is as follows: ,in The number of categories of favorable conditions for mineralization. For the first The weighting coefficients of favorable mineralization conditions, all weighting coefficients satisfy... The weighting coefficients can be determined using the analytic hierarchy process (AHP) or a scoring method by geological experts. For the first A quantification function for favorable mineralization conditions, which uses the multidimensional eigenvectors of voxel units. The input is a score of a single type of favorable mineralization condition, ranging from 0 to 1. Taking the tectonic control type of favorable mineralization condition as an example, its quantification function can be expressed as follows: ,in This represents the fault density value of the voxel unit, which is the fault length per unit area within the spatial range of the voxel. This represents the fault intersection degree of the voxel unit, i.e., the number of fault intersection points per unit area within the spatial range of the voxel. The weighting coefficients are and satisfy the following conditions: The final calculated The value is normalized to the range of 0 to 1. The higher the value, the more the voxel unit conforms to the geological laws of mineralization.

[0054] S102. Based on a multi-objective optimization algorithm, the deep neural network prediction model is trained and the Pareto front is solved by multi-dimensional feature vectors and geological law prior scores to obtain a set of mineral exploration prediction models. The multi-objective optimization algorithm takes minimizing the geological law conformity loss and maximizing the information entropy loss as its optimization objectives.

[0055] Specifically, the input layer dimension and multidimensional feature vectors of deep neural network prediction models The dimension n remains consistent. The hidden layer contains network layers such as convolutional layers, pooling layers, and fully connected layers. The convolutional layer is used to extract spatial correlation features from the feature vector. The pooling layer is used to reduce dimensionality and retain key features. The fully connected layer is used to integrate features and output the predicted value of mineralization probability. The output layer is a single neuron, and the output value is the mineralization probability value in the range of 0 to 1.

[0056] The multi-objective optimization algorithm selected is NSGA-III or MOEA / D, which are suitable for solving the Pareto front. The optimization objective is set as minimizing the geological regularity conformity loss and maximizing the information entropy loss. For example, the formula for calculating the geological regularity conformity loss function is as follows: ,in To determine the total number of three-dimensional voxel units in the training set, For deep neural networks to the first Predicted mineralization probability values ​​for individual element units. For the first The prior geological score of individual element units is used as a loss function to quantify the deviation between the model's prediction results and the prior score of metallogenic geological laws. A smaller value indicates that the prediction results are closer to the metallogenic geological theory. The formula for calculating the information entropy loss function is as follows: This function quantifies the information entropy of the model's prediction results. A higher value indicates greater uncertainty in the prediction results, enabling the uncovering of hidden correlations in multi-source geological data that are not covered by traditional metallogenic geological patterns. The objective function for multi-objective optimization is... By minimizing The optimization objective is to maximize the information entropy loss. During model training, the multi-objective optimization algorithm iteratively optimizes the parameters such as weights and biases of the deep neural network. In each iteration, the geological law conformity loss value and the information entropy loss value are calculated. Non-dominated solutions are selected based on the Pareto dominance relationship. After iterating to the convergence state, the Pareto front is obtained. Each solution on the Pareto front corresponds to a set of deep neural network parameters, that is, a Pareto optimal mineral exploration prediction model. All Pareto optimal models together constitute a set of mineral exploration prediction models.

[0057] S103. Based on all Pareto optimal models in the mineral exploration prediction model set, perform parallel mineral exploration target area prediction for the target exploration area to obtain multiple three-dimensional mineralization probability volumes, and fuse all three-dimensional mineralization probability volumes to obtain a mineral exploration prediction spectrum; the mineral exploration prediction spectrum includes a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume.

[0058] Furthermore, the TensorFlow distributed training framework or PyTorch's DataParallel parallel computing mechanism can be used to simultaneously input the multidimensional feature vectors of all three-dimensional voxel units in the target exploration area into each Pareto optimal model. Each model independently completes the mineralization probability prediction, and the output of each model is a tensor with the same dimension as the three-dimensional voxel grid of the target exploration area. This tensor is the three-dimensional mineralization probability volume, and each element in the tensor corresponds to the mineralization probability value of a voxel unit. All three-dimensional mineralization probability volumes are fused to obtain a mineralization prediction spectrum, which includes two parts: a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume. For example, suppose there are M Pareto optimal models in the mineralization prediction model set, and the voxel output of the m-th model... The probability value of mineralization is Then the consensus probability value of this voxel in the three-dimensional consensus probability volume. This value reflects the degree of consensus among multiple Pareto optimal models regarding the mineralization probability of the voxel unit; a higher value indicates greater consistency in the judgments of different models regarding the mineralization probability of the voxel. The three-dimensional uncertainty quantification uses variance as the uncertainty quantification index, and the voxel... The formula for calculating the uncertainty value is as follows: This value reflects the degree of dispersion of the mineralization probability prediction results of multiple Pareto optimal models for this voxel unit. The higher the value, the greater the difference in prediction between models, and the higher the uncertainty of the mineralization prediction in the corresponding region.

[0059] S104. Based on dynamic thresholds, the target exploration area is delineated using the mineral exploration prediction spectrum to obtain the final mineral exploration prediction scheme.

[0060] Optionally, the dynamic threshold is adaptively adjusted based on the statistical characteristics of the three-dimensional consensus probability volume in the mineral exploration prediction spectrum. First, the mean of the consensus probability values ​​of all voxel units in the three-dimensional consensus probability volume is calculated. with standard deviation Dynamic threshold The calculation formula is: ,in This is an adjustment coefficient, which can be adjusted according to the prospecting stage of the target exploration area, such as the general exploration stage. Take 0.5, detailed investigation stage The value of the adjustment coefficient is set to 1.0. The specific value needs to be determined comprehensively based on the geological background of the exploration area, exploration cost constraints, and mineral exploration objectives. The consensus probability value is then calculated by traversing each voxel unit in the three-dimensional consensus probability volume. Greater than or equal to dynamic threshold Voxel units are labeled as ore-forming favorable voxels; simultaneously, a secondary screening is performed using a three-dimensional uncertainty quantization volume, with a preset uncertainty threshold. For consensus probability values ​​reaching a dynamic threshold but uncertainty values... Exceed The voxel units can be selected for retention or removal based on the actual exploration scenario requirements. For example, voxels can be removed in high-risk exploration scenarios and retained in exploratory exploration scenarios. All voxel units marked as favorable for mineralization are integrated into a continuous three-dimensional spatial region according to their spatial coordinates. This region is the initially delineated mineral exploration target area. Spatial integration and parameter statistics are performed on the initially delineated target area to output key parameters such as the spatial coordinate range of the target area, the number of voxel units, the average consensus probability value, and the average uncertainty value, forming a final mineral exploration prediction scheme that includes the spatial location of the target area, mineralization probability characteristics, and uncertainty characteristics.

[0061] The aforementioned method for predicting solid mineral deposits based on geological big data involves three-dimensional spatial discretization and attribute assignment of multi-source geological big data from the target exploration area to construct multi-dimensional feature vectors for each three-dimensional voxel unit. Based on a formalized knowledge model of metallogenic geology, the prior geological law scores for each voxel unit are calculated. With the optimization objectives of minimizing geological law conformity loss and maximizing information entropy loss, a multi-objective optimization algorithm is used to train a deep neural network prediction model and solve for the Pareto front, resulting in a set of mineral exploration prediction models. All Pareto optimal models are then used to predict the target exploration area. Parallel mineral exploration target area prediction is performed, generating multiple three-dimensional mineralization probability volumes and fusing them to obtain a mineral exploration prediction spectrum that includes a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume. Based on dynamic thresholds, the mineral exploration target area is delineated through this prediction spectrum, ultimately generating the final mineral exploration prediction scheme. This effectively balances the inherent conflict between geological knowledge constraints and data-driven exploration, avoids the blindness and inefficiency of manual trial and error adjustments, provides flexible and transparent decision support, improves the reliability of mineral exploration prediction and exploration success rate, and provides effective guidance for precise deployment of exploration work by quantifying the uncertainty of the prediction.

[0062] In one embodiment, based on a multi-objective optimization algorithm, a deep neural network prediction model is trained and its Pareto front is solved using multi-dimensional feature vectors and prior scores of geological laws to obtain a set of mineral exploration prediction models, including:

[0063] S11. Based on the deep neural network prediction model, the prospecting target area is predicted for each three-dimensional voxel unit through multi-dimensional feature vectors, and the prediction probability of each three-dimensional voxel unit is obtained.

[0064] Specifically, a forward propagation mechanism is used to calculate the prediction probability. Multidimensional feature vectors are first input to the model's input layer and then passed to the hidden layer via input layer neurons. In the hidden layer, convolutional layers extract spatial correlation features from the feature vectors using pre-defined convolutional kernels. The size and number of convolutional kernels are set according to the feature dimension and spatial correlation requirements. After convolution, a non-linear feature mapping is introduced through the ReLU activation function to enhance the model's ability to fit complex geological correlations. Pooling layers use max pooling or average pooling to reduce the dimensionality of the convolutional feature maps, preserving key feature information while reducing the model's computational load. The dimensionality-reduced feature vectors are input to a fully connected layer, where weighted summation of multiple neurons and activation function transformation integrate high-dimensional features into low-dimensional feature vectors. The output layer uses the Sigmoid activation function to map the integrated features, obtaining the prediction probability of each three-dimensional voxel unit, directly reflecting the possibility of an economic mineral deposit existing in the corresponding voxel unit.

[0065] S12. Based on the prior score of geological laws and the predicted probability, calculate the KL divergence of the probability distribution to obtain the geological law conformity loss.

[0066] Schematic, the geological law conformity loss is achieved by calculating the KL divergence between the predicted probability distribution and the prior score distribution of the geological law. The KL divergence is used to quantify the degree of difference between the two probability distributions, i.e. ,in, This represents the loss value in terms of conformity to geological laws. Here is the KL divergence calculation function. Discrete probability distribution for predicting probabilities for all three-dimensional voxel units. The discrete probability distribution of the prior scores for geological regularities for all three-dimensional voxel units. This represents the total number of discretization intervals. To predict the probability falling on the th The proportion of voxel units within each discrete interval. The a priori score for geological laws falls at the first The proportion of voxel units within a discrete interval. During discretization, the prediction probability and the prior score of geological patterns are divided into... The results were obtained by counting the number of voxel units contained in each equidistant interval. and The smaller the KL divergence value, the closer the distribution of the predicted probability is to the distribution of the prior score of geological laws, and the more the model prediction results conform to the deterministic laws of metallogenic geology.

[0067] S13. Calculate the spatial conditional entropy based on the prediction probability of each three-dimensional voxel unit to obtain the information entropy loss.

[0068] Indicatively, the information entropy loss is obtained by calculating the spatial conditional entropy, which comprehensively considers the pointwise information entropy of the prediction probability and the spatial smoothness constraint, i.e. ,in, This represents the information entropy loss value. The total number of three-dimensional voxel units in the target exploration area. For the first The predicted probability of a three-dimensional voxel unit. These are the total variation regularization coefficients, used to balance the weights of pointwise information entropy and spatial smoothness. The total variation of the prediction probability is used to quantify the drastic change of the prediction probability in three-dimensional space, i.e. ,in , , The first The predicted probability of a voxel being adjacent to another voxel in the x, y, and z axes. The negative value of the pointwise binary cross-entropy is encouraged by maximizing this term, which can encourage the model to output predictions with clear discrimination and uncover novel associations hidden in the data. It can punish drastic fluctuations in the predicted probability in space, ensuring that the prediction results conform to the geological reality of continuous distribution of mineralized bodies.

[0069] S14. With the optimization objectives of minimizing the geological regularity conformity loss and minimizing the negative information entropy loss, the model population is iteratively evolved using a non-dominated sorting genetic algorithm. When the iterative evolution process reaches the target termination condition, all deep neural network prediction models in the first non-dominated layer of the final model population are output to obtain the mineral exploration prediction model set. The model population includes multiple deep neural network prediction models with different parameter sets. The iterative evolution corresponds to evaluating the optimization objective value of each deep neural network prediction model in each generation, and performing non-dominated sorting and crowding calculation based on Pareto dominance relationship, so as to obtain a new generation of population through selection, crossover and mutation of individuals.

[0070] Specifically, in the model population initialization phase, multiple deep neural network prediction models with different parameter sets are randomly generated according to a preset population size. Each model's parameters include convolutional layer weights, pooling layer parameters, fully connected layer weights, and bias terms. The weight parameters are generated using the Xavier initialization method, and the bias terms are initialized to 0 to ensure sufficient diversity in the initial population. The iterative evolution process aims to minimize the geological regularity conformity loss and the negative value of the information entropy loss. The optimization objective function is expressed as: ,in, This is the parameter set for a deep neural network prediction model. This represents the loss value in terms of geological regularity corresponding to this parameter set. This is the negative value of the information entropy loss corresponding to this parameter set.

[0071] For example, in each generation of iterative evolution, the optimization target value is evaluated for all models in the population, and the corresponding value for each model is calculated. and Then, non-dominated ranking is performed, that is, the superiority of models is determined according to Pareto dominance. If both optimization objective values ​​of model A are not inferior to model B, and at least one optimization objective value is superior to model B, then model A dominates model B. The population is divided into multiple non-dominated layers according to the dominance relationship, and the first non-dominated layer is the set of optimal models in the population. Within the same non-dominated layer, the crowding degree of each model is calculated. The crowding degree is obtained by calculating the sum of the distances of the model to its neighboring models in the objective space. The larger the distance, the higher the crowding degree. Models with high crowding degree are better able to maintain the diversity of the population.

[0072] Furthermore, the selection operation can employ a binary tournament selection method, randomly selecting two models from the population for comparison, prioritizing the model with a higher non-dominated layer level. If both models are in the same non-dominated layer, the model with higher crowding is selected as the parent individual, and this process is repeated until a predetermined number of parent individuals are selected. The crossover operation can use a simulated binary crossover algorithm, performing crossover operations on the parameter vectors of the selected parent individuals. A crossover probability is set to control the frequency of crossover operations, and a randomly generated crossover factor is used to linearly combine the parent parameters to generate the offspring parameter vector. The mutation operation employs a multinomial mutation algorithm, setting a mutation probability to control the frequency of mutation operations. Each element in the parameter vector is randomly perturbed according to a preset mutation distribution to achieve small changes in parameters and avoid the population getting trapped in local optima.

[0073] The parent and offspring populations are merged, and non-dominated sorting and crowding calculations are performed again. Models with the optimal preset population size are selected to form a new generation population. This iterative evolution process continues until a target termination condition is reached. The target termination condition can be set as the maximum number of iterations or no significant change in the optimization objective value of the models in the first non-dominated layer within a consecutive preset number of generations. When the iteration terminates, all models in the first non-dominated layer of the final population are extracted. These models are all Pareto optimal solutions and together constitute a mineral exploration prediction model set.

[0074] In one embodiment, the KL divergence of the probability distribution is calculated based on the prior score of geological laws and the predicted probability to obtain the geological law conformity loss, including:

[0075] S21. Divide the continuous range of geological law prior scores into multiple intervals evenly, and count the frequency of geological law prior scores of each three-dimensional voxel unit falling into each interval to obtain the geological prior probability distribution.

[0076] Indicatively, the continuous range of the geological law prior score is the span of the geological law prior scores for all three-dimensional voxel units. When uniformly dividing the intervals, the starting and ending values ​​of this continuous range can be determined first. The starting value is the minimum value among all prior scores, and the ending value is the maximum value among all prior scores. The range is then divided equally according to the preset total number of intervals, with the width of each interval being the ratio of the continuous range to the total number of intervals. Optionally, when calculating the frequency of the geological law prior score of each three-dimensional voxel unit falling into each interval, the interval to which the prior score of each voxel unit belongs is determined one by one. Then, the number of voxel units contained in each interval is accumulated, and the number of voxel units in each interval is divided by the total number of three-dimensional voxel units in the target exploration area to obtain the frequency corresponding to each interval. The frequencies of all intervals together constitute the geological prior probability distribution, which fully reflects the distribution characteristics of the geological law prior score in each numerical interval.

[0077] S22. Based on the frequency of the predicted probability falling into each interval, obtain the model's predicted probability distribution.

[0078] Similarly, the interval division criteria for model prediction probabilities are completely consistent with the interval division of geological law prior scores. When calculating the frequency of prediction probabilities falling into each interval, the interval assignment of the prediction probability of each three-dimensional voxel unit is first determined, the number of voxel units in each interval is recorded, and then the number of voxel units in each interval is divided by the total number of three-dimensional voxel units to obtain the frequency corresponding to each interval.

[0079] S23. Based on the prior geological probability distribution and the model prediction probability distribution, the geological law conformity loss is obtained; the formula for calculating the geological law conformity loss is: ,in, The predicted probability distribution in the model is at the th... The frequency of each interval The geological prior score distribution in the geological prior probability distribution is the th The frequency of each interval The total number of intervals, This is the parameter set for a deep neural network prediction model.

[0080] Specifically, This is the geological law conformity loss value, used to quantify the degree of difference between the model's predicted probability distribution and the geological prior probability distribution; This is the parameter set of a deep neural network prediction model. The parameter set contains all trainable parameters such as weights and biases of each layer of the model. Different parameter sets correspond to different model prediction results. The total number of intervals is the number of intervals divided by the prior score of geological laws and the predicted probability. For the model to predict the probability distribution of the th The frequency of each interval reflects the prediction probability in the th interval. The distribution percentage of each interval; The first in the geological prior probability distribution The frequency of each interval reflects the prior score of geological laws in the th interval. The distribution percentage of each interval.

[0081] In one embodiment, the spatial conditional entropy is calculated based on the predicted probability of each three-dimensional voxel unit to obtain the information entropy loss, including:

[0082] S31. Calculate the negative of the sum of pointwise binary cross-entropies based on the predicted probabilities of all three-dimensional voxel units to obtain the fundamental component terms; the expression for the fundamental component terms is: ,in, It is a three-dimensional voxel unit. To predict probabilities.

[0083] Specifically, It is a basic component term used to quantify the overall discriminative power of the prediction probabilities of all three-dimensional voxel units; The total number of three-dimensional voxel units in the target exploration area, and ; , , These represent the number of grid cells in the x-axis, y-axis, and z-axis directions of the 3D spatial grid, respectively. The coordinate index is Three-dimensional voxel units, , , These correspond to the position indices along the x-axis, y-axis, and z-axis in three-dimensional space, respectively. For deep neural network prediction models on parameter sets Below, for three-dimensional voxel units The output predicted probability; The binary cross-entropy component, which is a natural logarithm operation, measures the degree of deviation of a single predicted probability from the extreme value, reflecting the clarity of the model's judgment on the mineralization probability of the voxel unit.

[0084] S32. Based on the total variation of the predicted probabilities of all 3D voxel units on the 3D spatial mesh, generate a spatial smoothness regularization term; the expression for the spatial smoothness regularization term is: .

[0085] Specifically, This is a spatial smoothness regularization term used to quantify the degree of drastic change in the predicted probability in three-dimensional space; Three-dimensional voxel unit Predicted probability of adjacent voxel units in the positive x-axis direction The coordinate index is , and when At that time, the prediction probability of adjacent voxel units is assumed to be the same as... Consistent; for Predicted probability of adjacent voxel units in the positive y-axis direction The coordinate index is , and when At that time, the prediction probability of adjacent voxel units is assumed to be the same as... Consistent; for Predicted probability of adjacent voxel units in the positive z-axis direction The coordinate index is , and when At that time, the prediction probability of adjacent voxel units is assumed to be the same as... Consistency; the absolute value operation is used to calculate the difference in predicted probabilities between adjacent voxel units. The sum of the three difference values ​​is the spatial variation contribution of a single voxel unit. The total variation is obtained by summing the contributions of all voxel units. The larger the value, the more drastic the spatial fluctuation of the predicted probability.

[0086] S33. The basic component terms and the spatial smoothness regularization terms are fused according to the weight coefficients to obtain the information entropy loss.

[0087] For example, the information entropy loss is obtained by fusing the basic component terms and the spatial smoothness regularization term with preset weighting coefficients, i.e. ,in, This represents the information entropy loss value. Basic component terms; is a weighting coefficient, a non-negative real number, used to balance the contribution ratio of the basic component terms and the spatial smoothness regularization term. Its value needs to be determined in combination with the geological background and mineralization characteristics of the target exploration area to ensure that the model maintains the spatial rationality of the prediction results while mining the hidden correlations in the data. For spatial smoothness regularization. After fusion, the optimization objective of the information entropy loss value is to maximize it. The basic component terms encourage the model to output prediction results with high discrimination, and to explore novel correlation information not covered by traditional metallogenic knowledge. At the same time, the spatial smoothness regularization term constrains the drastic spatial fluctuations of the prediction probability, so that the predicted metallogenic probability distribution conforms to the geological reality of continuous mineralization.

[0088] In one embodiment, parallel mineral exploration target area prediction is performed on the target exploration area based on all Pareto optimal models in the mineral exploration prediction model set, resulting in multiple three-dimensional mineralization probability volumes. These three-dimensional mineralization probability volumes are then fused to obtain a mineral exploration prediction spectrum, including:

[0089] S41. Calculate the average of the predicted probability values ​​of each three-dimensional mineralization probability volume on the same three-dimensional voxel unit to obtain the three-dimensional consensus probability volume.

[0090] Schematic, the three-dimensional consensus probability volume is obtained by calculating the arithmetic mean of the predicted probability values ​​of all three-dimensional mineralization probability volumes on the same three-dimensional voxel unit, i.e. ,in, The coordinate index in the three-dimensional consensus probability volume is Three-dimensional voxel units The corresponding consensus probability value; The total number of Pareto optimal models in the mineral exploration prediction model set, with each model corresponding to an independent three-dimensional mineralization probability volume; This is the index of the Pareto optimal model, with a value range of [value missing]. to ; For the first In the three-dimensional mineralization probability volume output by a Pareto optimal model, the voxel unit The predicted probability value; the calculation logic of the arithmetic mean can integrate the prediction results of all Pareto optimal models, weaken the local bias of a single model, highlight the consistent judgment of different models on the mineralization possibility of the same voxel unit. The higher the consensus probability value, the higher the degree of recognition of the mineralization potential of the voxel unit by most models.

[0091] S42. Calculate the standard deviation of the predicted probability values ​​of each three-dimensional mineralization probability volume on the same three-dimensional voxel unit to obtain the three-dimensional uncertainty quantification volume.

[0092] Schematic, the three-dimensional uncertainty quantification volume is obtained by calculating the standard deviation of the predicted probability values ​​of all three-dimensional mineralization probability volumes on the same three-dimensional voxel unit, i.e. ,in, Voxel units in a three-dimensional uncertain quantization volume The corresponding uncertainty value; This represents the total number of Pareto optimal models. For the first voxel units in a three-dimensional mineralization probability volume The predicted probability value; The standard deviation is the consensus probability value of the voxel unit. The calculation of the standard deviation can quantify the degree of dispersion between the predicted probability of each model and the consensus probability. The larger the uncertainty value, the greater the disagreement between different Pareto optimal models in judging the mineralization potential of the voxel unit. There may be gaps in mineralization knowledge or data contradictions in this area, which has higher exploration risks and potential value. Conversely, the smaller the uncertainty value, the more concentrated the prediction results among the models are, and the higher the reliability of the judgment of mineralization potential.

[0093] S43. From the set of mineral exploration prediction models, determine the three-dimensional mineralization probability body corresponding to the Pareto optimal model with the smallest loss of geological law conformity as the conservative prediction body, and determine the three-dimensional mineralization probability body corresponding to the Pareto optimal model with the smallest loss of information entropy as the aggressive prediction body.

[0094] Specifically, the conservative predictor is determined by iterating through all Pareto optimal models in the mineral exploration prediction model set, extracting the geological law conformity loss value corresponding to each model during training, and comparing all loss values ​​to select the model with the smallest geological law conformity loss. The core characteristic of this model is that its prediction results have the highest degree of conformity with the deterministic laws of metallogenic geology, making it suitable for scenarios with low tolerance for exploration risks and a tendency to rely on known metallogenic laws for exploration. The aggressive predictor is determined by extracting the information entropy loss value corresponding to each Pareto optimal model and selecting the model with the smallest information entropy loss. Since the smallest information entropy loss is equivalent to the model with the strongest information exploration capability, it can uncover hidden correlations in the data that are not covered by traditional metallogenic knowledge, making it suitable for scenarios that seek to break through traditional cognition and are willing to take higher exploration risks to explore new mineralized areas.

[0095] S44. Integrate the three-dimensional consensus probability body, the three-dimensional uncertainty quantification body, the conservative prediction body, and the radical prediction body into a mineral exploration prediction system.

[0096] Among them, the three-dimensional consensus probability volume provides a basic reference for mineralization potential that is consistently recognized by multiple models, the three-dimensional uncertainty quantification volume supplements the risk dimension information of the prediction results, and the conservative and aggressive prediction volumes cover the prediction needs under different risk preferences. The integrated mineral exploration prediction spectrum forms a multi-level and multi-dimensional prediction result system, which includes both core prediction information reflecting general consensus and extended information reflecting risk differences and exploration tendencies. It can provide comprehensive and flexible decision support for subsequent mineral exploration target area delineation and adapt to the mineral exploration needs of different exploration stages and different risk tolerance levels.

[0097] In one embodiment, based on a dynamic threshold, a mineral exploration target area is delineated using a mineral exploration prediction spectrum to obtain a final mineral exploration prediction scheme, including:

[0098] S51. Load the three-dimensional consensus probability volume and the three-dimensional uncertainty quantification volume in the mineral exploration prediction spectrum into the three-dimensional visualization interactive platform, and perform spatial coordinate registration and overlay rendering with the pre-stored basic geological layer data to obtain a three-dimensional comprehensive prediction view.

[0099] Indicatively, the 3D visualization interactive platform supports loading, rendering, spatial analysis, and human-computer interaction of 3D spatial data. The platform incorporates a unified geodetic coordinate system and elevation datum, maintaining consistency with the spatial reference systems of the 3D consensus probability volume and the 3D uncertainty quantification volume. Pre-stored basic geological layer data encompasses various types of geological data, including regional geological maps, structural outline maps, stratigraphic columnar sections, geophysical anomaly maps, and geochemical anomaly maps of the target exploration area. All basic geological layer data has undergone standardization to ensure compatibility of its spatial coordinate format with the 3D probability volume. The spatial coordinate registration process extracts the spatial reference parameters of each data layer, mapping the coordinate systems of the 3D consensus probability volume, the 3D uncertainty quantification volume, and the basic geological layer data to the platform's preset spatial framework, achieving precise alignment of each data layer in 3D space. Optionally, during overlay rendering, a color gradient mapping rule is used to convert the probability value of the 3D consensus probability volume and the uncertainty value of the 3D uncertainty quantization volume into visual colors. The higher the probability value and uncertainty value, the higher the corresponding color saturation. At the same time, the geological element outlines of the basic geological layer are preserved, and finally a 3D comprehensive prediction view that integrates geological background and prediction information is formed, which supports multi-view rotation, zoom and profile cutting viewing.

[0100] S52. In response to obtaining the dynamic probability threshold parameter, extract the set of voxel space coordinates of all three-dimensional voxel units in the three-dimensional consensus probability volume that satisfy the average value ≥ dynamic probability threshold parameter, and perform three-dimensional spatial clustering analysis based on the connectivity of the voxel space coordinate set to obtain the candidate abnormal three-dimensional entity model.

[0101] Furthermore, the dynamic probability threshold parameter can be input autonomously by geological experts based on exploration objectives, risk preferences, and regional metallogenic background, or adaptively generated by the system based on the statistical characteristics of the 3D consensus probability volume. When extracting 3D voxel units that satisfy the condition that the average value is greater than or equal to the dynamic probability threshold parameter, all coordinate indices in the 3D consensus probability volume are traversed. voxel units Filter out The voxel units are collected, and their spatial coordinates are used to form a voxel spatial coordinate set. The three-dimensional spatial clustering analysis adopts a connectivity clustering algorithm, which sets voxel connectivity rules, that is, judging whether adjacent voxels simultaneously meet the probability threshold conditions and are spatially adjacent. The interconnected voxel unit clusters are divided into independent three-dimensional entities, and isolated voxel clusters with a volume smaller than the preset minimum threshold are removed. The final three-dimensional entity is a candidate abnormal three-dimensional entity model, and each model corresponds to a potential mineralization favorable area.

[0102] S53. In response to obtaining the dynamic uncertainty threshold parameter, mark all three-dimensional voxel units in the three-dimensional uncertainty quantization volume that satisfy the standard deviation > dynamic uncertainty threshold parameter as high uncertainty exploration areas, and overlay the high uncertainty exploration areas in the three-dimensional comprehensive prediction view using a preset rendering method.

[0103] Similarly, the dynamic uncertainty threshold parameter is set in conjunction with the overall distribution characteristics of the 3D uncertainty quantization volume. Statistical indicators such as the mean, median, and standard deviation of the uncertainty values ​​can be obtained through the platform's statistical functions, and a reasonable threshold can be determined based on these indicators. When marking high-uncertainty exploration areas, all voxel units in the 3D uncertainty quantization volume are traversed. , will satisfy The voxel units are marked as high-uncertainty exploration areas. The preset rendering method uses a color that is significantly different from the main color of the 3D comprehensive prediction view and sets a semi-transparent overlay effect to display the high-uncertainty exploration areas on top of the generated 3D comprehensive prediction view. This highlights the spatial location of the high-uncertainty areas without obscuring the details of the geological background and candidate anomaly 3D entity models below, making it easier for experts to intuitively identify high-risk and high-potential exploration areas.

[0104] S54. Compare the conservative predictive volume and the radical probabilistic volume in three dimensions to obtain the difference probabilistic volume.

[0105] For example, the difference probability volume is obtained by calculating the difference in prediction probabilities between the aggressive predictor and the conservative predictor on a per-3D voxel basis, i.e. ,in, Voxel units in the differential probability volume The corresponding probability value of difference; For radical prediction of voxel units The predicted probability value is derived from the radical predictor identified in the mineral exploration prediction spectrum; For conservative prediction of voxel units The predicted probability values ​​are derived from the conservative predictors identified in the mineral exploration prediction spectrum. A positive difference probability value indicates that the aggressive predictor's assessment of the mineralization potential of the element is higher than that of the conservative predictor; a negative difference probability value indicates that the conservative predictor's assessment is more optimistic; and a difference probability value of 0 indicates that the two assessments are consistent. The difference probability value fully reflects the divergence distribution between the two prediction models with different risk preferences.

[0106] S55. Perform threshold segmentation and spatial clustering on the differential probability volume, and mark the high-value clustering regions unique to the radical probability volume as prospective exploration candidate regions.

[0107] Optionally, when performing threshold segmentation on the difference probability volume, a positive threshold can be set. Filter out those that meet the requirements Voxel units, representing regions where the mineralization potential of aggressive predictors is significantly higher than that of conservative predictors, are further spatially clustered to form multiple independent cluster regions. By comparing the prediction probability values ​​of aggressive and conservative predictors in each cluster region, cluster regions where the prediction probability value of aggressive predictors reaches a preset high standard, while the corresponding region's prediction probability value in conservative predictors does not reach this standard, are selected as prospective exploration candidate areas. These areas represent potential mineralization areas that break through traditional understanding of mineralization patterns and have high exploration value.

[0108] S56. Based on the candidate anomaly 3D entity model, 3D comprehensive prediction view and prospective exploration candidate area, and combined with the preset target area fusion rules, perform spatial overlay analysis to generate the final mineral exploration prediction scheme; the final mineral exploration prediction scheme includes the category, 3D spatial coordinate range, confidence level and associated geological attribute information of each mineral exploration target area.

[0109] Specifically, the pre-defined target area fusion rules include three core dimensions: spatial priority determination, attribute correlation verification, and integrity screening. In spatial priority determination, candidate anomaly 3D entity models serve as the foundation of the core target area, and their spatial extent has the highest priority. Areas overlapping with candidate anomaly 3D entity models in high-uncertainty exploration zones are included in the expansion range of the core target area. Prospective exploration candidate areas are treated as independent expansion target area types. Attribute correlation verification extracts the corresponding basic geological layer data within the spatial range of each target area to obtain associated geological attribute information such as lithology, structural type, alteration intensity, and elemental anomalies, verifying the degree of matching between the target area and favorable geological conditions for mineralization. Integrity screening eliminates target areas with excessively small spatial extents, incomplete geological attribute information, or overlap with known non-mineralized areas. During the spatial overlay analysis, the core area of ​​the candidate anomaly 3D entity model, the extended part of the high uncertainty exploration area, and the prospective exploration candidate area are integrated to divide the target areas into three categories: key verification target areas, extended exploration target areas, and prospective potential target areas. The confidence level of each target area is determined based on the probability value of the 3D consensus probability volume, with higher probability values ​​indicating higher confidence levels. The final mineral exploration prediction scheme clearly includes the category definition of each target area, the maximum and minimum latitude and longitude and elevation range of the 3D spatial coordinates, the confidence level classification criteria, and the corresponding detailed information on associated geological attributes, providing a complete technical basis for subsequent exploration engineering deployment.

[0110] In one embodiment, the method further includes:

[0111] S61. Mark positive sample labels for mineral-bearing locations verified by field engineering in the final mineral exploration prediction scheme, and negative sample labels for non-mineral-bearing locations. Combine the spatial coordinates of the three-dimensional voxel units corresponding to each label location with the multi-dimensional feature vectors to construct a new verification sample set. The new verification sample set is used to train the deep neural network prediction model.

[0112] This illustrative, field engineering verification involves deploying exploration projects in various target areas defined by the final mineral exploration prediction plan. These projects include trenching, drilling, and pitting, and the actual mineralization information of the target areas is obtained through engineering exposure. The determination of mineralization locations is based on industrial mineral exploration standards. When the thickness of the mineralized body and the ore grade revealed by the engineering project meet the preset industrial index requirements, the exposed location is marked as a positive sample label, with the label value set to [value missing]. When no mineralized body is found in the area exposed by the project, or when the thickness and grade of the mineralized body do not meet the industrial standards, the corresponding location will be marked as a negative sample label, and the label value will be set to [value missing]. During the labeling process, the actual geographic coordinates of the project's exposed location are converted into spatial coordinate indices corresponding to the three-dimensional voxel grid of the target exploration area. This ensures precise association with the corresponding 3D voxel units. Constructing a new validation sample set requires integrating the spatial coordinate indices and multidimensional feature vectors of the 3D voxel units corresponding to each label location. And sample labels. A new validation sample set is added as supplementary training data for incremental training of the subsequent deep neural network prediction model. By incorporating the latest field validation information, the model prediction bias is corrected, and the model's adaptability and prediction accuracy to the mineralization regularities of the target exploration area are improved.

[0113] S62. Calculate the systematic deviation based on the newly added verification sample set and the prior score of geological laws. If the prior score of geological laws exceeds the three-dimensional voxel unit corresponding to the preset threshold in multiple consecutive verification periods, it is a positive sample label, triggering the update of the formal knowledge model of metallogenic geology. The update of the formal knowledge model of metallogenic geology corresponds to the correction of the rule confidence, the conditional probability table of the probabilistic graphical model and the quantization function parameters of the prediction elements in the formal knowledge model of metallogenic geology.

[0114] For example, systematic bias is obtained by calculating the degree of deviation between the verification results of three-dimensional voxel units whose prior geological law scores exceed a preset threshold and the theoretical expectations. ,in, This is a systematic deviation value used to quantify the degree of prediction bias in the formal knowledge model of metallogenic geological theory; The number of positive sample labels in three-dimensional voxel units where the prior score of geological law exceeds a preset threshold; The total number of three-dimensional voxel units whose a priori scores for geological patterns exceed a preset threshold. The confidence level is set based on the formal knowledge model of metallogenic geology theory, representing the theoretically expected proportion of positive samples, and reflects the probability of mineralization in areas with high prior scores as predicted by the model.

[0115] The verification cycle is measured in batches of a single field engineering verification. Multiple consecutive verification cycles refer to a predetermined number of consecutive engineering verification batches, with each batch containing verification results from multiple exploration projects in at least one target area. Within multiple consecutive verification cycles, the actual proportion of positive sample labels and the theoretically expected proportion are considered within three-dimensional voxel units where the prior geological score exceeds a preset threshold. deviation value When the deviation exceeds the preset threshold continuously, it indicates that there is a systematic deviation in the formal knowledge model of metallogenic geology that does not conform to the actual metallogenic laws, thus triggering the model update mechanism.

[0116] The updating process of the formal knowledge model of metallogenic geology revolves around the correction of core component parameters. Specifically, it involves the rule confidence in the production rule set. Based on the distribution of positive samples in the newly added validation sample set, the confidence of rules that miss positive samples is reduced, while the confidence of rules that hit positive samples and are not misclassified is increased. For the conditional probability table (CPT) of the probabilistic graphical model, the conditional probabilities between each node are recalculated using the newly added validation sample set as supplementary training data, and the values ​​in the conditional probability table are updated. For the quantization function parameters of the predicted elements, including the weight coefficients in the weighted linear combination... Constructing the ore-controlling quantification function Based on the correspondence between features and labels of the newly added validation sample set, the parameters are optimized and adjusted using the gradient descent method to make the quantization function more in line with the quantization requirements of actual mineralization conditions, ensuring that the updated knowledge model can accurately reflect the actual mineralization patterns of the target exploration area.

[0117] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0118] Based on the same inventive concept, this application also provides a geological big data-based solid mineral exploration prediction system for implementing the above-mentioned geological big data-based solid mineral exploration prediction method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more geological big data-based solid mineral exploration prediction system embodiments provided below can be found in the limitations of the geological big data-based solid mineral exploration prediction method described above, and will not be repeated here.

[0119] In one exemplary embodiment, such as Figure 2 As shown, a solid mineral exploration prediction system based on geological big data is provided, including:

[0120] The prior module 201 is used to perform three-dimensional spatial discretization and attribute assignment processing on the multi-source geological big data of the target exploration area to obtain the multi-dimensional feature vector of each three-dimensional voxel unit, and calculate the geological law prior score of each three-dimensional voxel unit based on the formal knowledge model of metallogenic geology theory.

[0121] The balancing module 202 is used to train and solve the Pareto front of the deep neural network prediction model based on the multi-objective optimization algorithm through multi-dimensional feature vectors and geological law prior scores to obtain a set of mineral exploration prediction models. The multi-objective optimization algorithm takes minimizing the geological law conformity loss and maximizing the information entropy loss as its optimization objectives.

[0122] The prediction module 203 is used to perform parallel mineral exploration target area prediction on the target exploration area based on all Pareto optimal models in the mineral exploration prediction model set, to obtain multiple three-dimensional mineralization probability volumes, and to fuse all three-dimensional mineralization probability volumes to obtain a mineral exploration prediction spectrum; the mineral exploration prediction spectrum includes a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume.

[0123] The targeting module 204 is used to delineate the target exploration area based on a dynamic threshold and a mineral exploration prediction spectrum to obtain the final mineral exploration prediction scheme.

[0124] In one embodiment, the balancing module 202 is further configured to:

[0125] Based on a deep neural network prediction model, the prospecting target area is predicted for each three-dimensional voxel unit by using multi-dimensional feature vectors, and the prediction probability of each three-dimensional voxel unit is obtained.

[0126] Based on the prior score of geological laws and the predicted probability, the KL divergence of the probability distribution is calculated to obtain the geological law conformity loss.

[0127] Spatial conditional entropy is calculated based on the prediction probability of each three-dimensional voxel unit, and information entropy loss is obtained.

[0128] With the optimization objectives of minimizing the loss of geological regularity conformity and minimizing the negative value of information entropy loss, the model population is iteratively evolved using a non-dominated sorting genetic algorithm. When the iterative evolution process reaches the target termination condition, all deep neural network prediction models in the first non-dominated layer of the final model population are output to obtain the mineral exploration prediction model set. The model population includes multiple deep neural network prediction models with different parameter sets. The iterative evolution corresponds to evaluating the optimization target value of each deep neural network prediction model in each generation, and performing non-dominated sorting and crowding calculation based on Pareto dominance relationship, so as to obtain a new generation of population through selection, crossover and mutation of individuals.

[0129] In one embodiment, a geological regularity conformity loss module is also included, for:

[0130] The continuous range of geological law prior scores is evenly divided into multiple intervals, and the frequency of geological law prior scores of each three-dimensional voxel unit falling into each interval is counted to obtain the geological prior probability distribution.

[0131] The model's predicted probability distribution is obtained based on the frequency of predicted probabilities falling into each interval.

[0132] Based on the prior geological probability distribution and the model prediction probability distribution, the geological law conformity loss is obtained; the formula for calculating the geological law conformity loss is as follows: ,in, The predicted probability distribution in the model is at the th... The frequency of each interval The geological prior score distribution in the geological prior probability distribution is the th The frequency of each interval The total number of intervals, This is the parameter set for a deep neural network prediction model.

[0133] In one embodiment, an information entropy loss module is also included, for:

[0134] The fundamental component term is obtained by calculating the negative of the sum of pointwise binary cross-entropy based on the predicted probabilities of all 3D voxel units; the expression for the fundamental component term is: ,in, It is a three-dimensional voxel unit. To predict probabilities;

[0135] Based on the total variation of the predicted probabilities of all 3D voxel units on the 3D spatial mesh, a spatial smoothness regularization term is generated; the expression for the spatial smoothness regularization term is: ;

[0136] The information entropy loss is obtained by fusing the basic component terms and the spatial smoothness regularization term according to the weight coefficients.

[0137] In one embodiment, the prediction module 203 is further configured to:

[0138] The average of the predicted probability values ​​of each three-dimensional mineralization probability volume on the same three-dimensional voxel unit is calculated to obtain the three-dimensional consensus probability volume.

[0139] The standard deviation of the predicted probability values ​​of each three-dimensional mineralization probability volume on the same three-dimensional voxel unit is calculated to obtain the three-dimensional uncertainty quantification volume.

[0140] From the set of mineral exploration prediction models, the three-dimensional mineralization probability volume corresponding to the Pareto optimal model with the smallest loss of geological law conformity is determined as the conservative prediction volume, and the three-dimensional mineralization probability volume corresponding to the Pareto optimal model with the smallest loss of information entropy is determined as the aggressive prediction volume.

[0141] The three-dimensional consensus probability body, the three-dimensional uncertainty quantification body, the conservative prediction body, and the radical prediction body are integrated into a mineral exploration prediction system.

[0142] In one embodiment, the targeting module 204 is further configured to:

[0143] The three-dimensional consensus probability volume and the three-dimensional uncertainty quantification volume in the mineral exploration prediction spectrum are loaded into the three-dimensional visualization interactive platform, and spatial coordinate registration and overlay rendering are performed with the pre-stored basic geological layer data to obtain a three-dimensional comprehensive prediction view.

[0144] In response to the acquisition of the dynamic probability threshold parameter, the set of voxel space coordinates of all three-dimensional voxel units in the three-dimensional consensus probability volume that satisfy the average value ≥ the dynamic probability threshold parameter is extracted, and three-dimensional spatial clustering analysis is performed based on the connectivity of the voxel space coordinate set to obtain the candidate abnormal three-dimensional entity model.

[0145] In response to the acquisition of the dynamic uncertainty threshold parameter, all three-dimensional voxel units in the three-dimensional uncertainty quantization volume that satisfy the standard deviation > the dynamic uncertainty threshold parameter are marked as high uncertainty exploration areas, and the high uncertainty exploration areas are superimposed and displayed in the three-dimensional comprehensive prediction view using a preset rendering method.

[0146] By comparing the conservative predictive volume and the radical probabilistic volume on a three-dimensional voxel basis, the difference probabilistic volume is obtained;

[0147] Threshold segmentation and spatial clustering are performed on the differential probability volumes, and high-value clustering regions unique to the radical probability volumes are marked as prospective exploration candidate regions;

[0148] Based on the candidate anomaly 3D entity model, 3D comprehensive prediction view, and prospective exploration candidate area, spatial overlay analysis is performed in conjunction with preset target area fusion rules to generate a final mineral exploration prediction scheme. The final mineral exploration prediction scheme includes the category, 3D spatial coordinate range, confidence level, and associated geological attribute information of each mineral exploration target area.

[0149] In one embodiment, an optimization module is also included, for:

[0150] Positive sample labels are assigned to mineral-bearing locations verified in the field engineering of the final mineral exploration prediction scheme, and negative sample labels are assigned to non-mineral-bearing locations. A new verification sample set is constructed by combining the spatial coordinates of the three-dimensional voxel units corresponding to each label location and the multi-dimensional feature vector. The new verification sample set is used to train the deep neural network prediction model.

[0151] The systematic deviation is calculated based on the newly added verification sample set and the prior score of geological laws. If the prior score of geological laws exceeds the preset threshold in multiple consecutive verification periods, the corresponding three-dimensional voxel unit is a positive sample label, triggering an update of the formal knowledge model of metallogenic geology. The update of the formal knowledge model of metallogenic geology corresponds to the correction of the rule confidence, the conditional probability table of the probabilistic graphical model, and the quantization function parameters of the prediction elements in the formal knowledge model of metallogenic geology.

[0152] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0153] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0154] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0155] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting solid mineral deposits based on geological big data, characterized in that, The method includes: The multi-source geological big data of the target exploration area is discretized in three dimensions and attribute assigned to obtain the multi-dimensional feature vector of each three-dimensional voxel unit. Based on the formal knowledge model of metallogenic geology theory, the geological law prior score of each three-dimensional voxel unit is calculated. Based on a multi-objective optimization algorithm, the deep neural network prediction model is trained and Pareto front is solved using the multi-dimensional feature vector and the geological law prior score to obtain a mineral exploration prediction model set; the multi-objective optimization algorithm takes minimizing the geological law conformity loss and maximizing the information entropy loss as its optimization objectives. Based on all Pareto optimal models in the mineral exploration prediction model set, parallel mineral exploration target area prediction is performed on the target exploration area to obtain multiple three-dimensional mineralization probability volumes. All three-dimensional mineralization probability volumes are then fused to obtain a mineral exploration prediction spectrum. The mineral exploration prediction spectrum includes a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume. Based on dynamic thresholds, the target exploration area is delineated using the mineral exploration prediction spectrum to obtain the final mineral exploration prediction scheme.

2. The method according to claim 1, characterized in that, The multi-objective optimization algorithm trains the deep neural network prediction model and solves the Pareto front using the multi-dimensional feature vector and the prior score of geological laws, resulting in a set of mineral exploration prediction models, including: Based on the deep neural network prediction model, the target area for mineral exploration is predicted for each of the three-dimensional voxel units through the multi-dimensional feature vector, and the prediction probability of each of the three-dimensional voxel units is obtained. Based on the prior score of the geological law and the predicted probability, the KL divergence of the probability distribution is calculated to obtain the conformity loss of the geological law; The spatial conditional entropy is calculated based on the predicted probability of each of the three-dimensional voxel units, and the information entropy loss is obtained. The optimization objective is to minimize the loss of conformity to the geological laws and the negative value of the information entropy loss. A non-dominated sorting genetic algorithm is used to iteratively evolve the model population. When the iterative evolution reaches the target termination condition, all deep neural network prediction models in the first non-dominated layer of the final model population are output, resulting in the mineral exploration prediction model set. The model population includes multiple deep neural network prediction models with different parameter sets. The iterative evolution corresponds to evaluating the optimization objective value of each deep neural network prediction model in each generation, and performing non-dominated sorting and crowding calculation based on Pareto dominance to obtain a new generation of the population through selection, crossover, and mutation of individuals.

3. The method according to claim 2, characterized in that, The calculation of the probability distribution KL divergence based on the prior score of the geological law and the predicted probability to obtain the conformity loss of the geological law includes: The continuous range of the geological law prior score is evenly divided into multiple intervals, and the frequency of the geological law prior score of each three-dimensional voxel unit falling into each interval is counted to obtain the geological prior probability distribution. The model prediction probability distribution is obtained based on the frequency of the predicted probability falling into each of the intervals. Based on the geological prior probability distribution and the model prediction probability distribution, the geological law conformity loss is obtained; the formula for calculating the geological law conformity loss is as follows: ,in, The predicted probability distribution in the model is the probability distribution at the th... The frequency of each interval The geological prior score distribution in the geological prior probability distribution is the th... The frequency of each interval The total number of intervals, The parameter set of the deep neural network prediction model.

4. The method according to claim 2, characterized in that, The calculation of spatial conditional entropy based on the predicted probability of each of the three-dimensional voxel units to obtain the information entropy loss includes: The negative of the sum of pointwise binary cross-entropy is calculated based on the predicted probabilities of all the three-dimensional voxel units to obtain the fundamental component term; the expression for the fundamental component term is: ,in, The three-dimensional voxel unit, The predicted probability; A spatial smoothness regularization term is generated based on the total variation of the predicted probabilities of all the three-dimensional voxel units on the three-dimensional spatial mesh; the expression for the spatial smoothness regularization term is: ; The basic component term and the spatial smoothness regularization term are fused together according to weight coefficients to obtain the information entropy loss.

5. The method according to claim 1, characterized in that, The method involves performing parallel mineral exploration target area prediction on the target exploration area based on all Pareto optimal models in the mineral exploration prediction model set, obtaining multiple three-dimensional mineralization probability volumes, and fusing all the three-dimensional mineralization probability volumes to obtain a mineral exploration prediction spectrum, including: The average value of the predicted probability values ​​of each of the three-dimensional mineralization probability volumes on the same three-dimensional voxel unit is calculated to obtain the three-dimensional consensus probability volume. The standard deviation of the predicted probability values ​​of each of the three-dimensional mineralization probability volumes on the same three-dimensional voxel unit is calculated to obtain the three-dimensional uncertainty quantification volume; From the set of prospecting prediction models, the three-dimensional mineralization probability body corresponding to the Pareto optimal model with the smallest loss of geological law conformity is determined to be a conservative prediction body, and the three-dimensional mineralization probability body corresponding to the Pareto optimal model with the smallest loss of information entropy is determined to be an aggressive prediction body. The three-dimensional consensus probability body, the three-dimensional uncertainty quantification body, the conservative prediction body, and the radical prediction body are integrated together to form the mineral exploration prediction spectrum.

6. The method according to claim 5, characterized in that, The process of delineating the target exploration area based on a dynamic threshold using the mineral exploration prediction spectrum to obtain a final mineral exploration prediction scheme includes: The three-dimensional consensus probability volume and the three-dimensional uncertainty quantification volume in the mineral exploration prediction spectrum are loaded into the three-dimensional visualization interactive platform, and spatial coordinate registration and overlay rendering are performed with the pre-stored basic geological layer data to obtain a three-dimensional comprehensive prediction view. In response to obtaining the dynamic probability threshold parameter, the set of voxel space coordinates of all three-dimensional voxel units in the three-dimensional consensus probability volume that satisfy the average value ≥ the dynamic probability threshold parameter is extracted, and three-dimensional spatial clustering analysis is performed based on the connectivity of the set of voxel space coordinates to obtain candidate abnormal three-dimensional entity models. In response to obtaining the dynamic uncertainty threshold parameter, all three-dimensional voxel units in the three-dimensional uncertainty quantization volume that satisfy the standard deviation > the dynamic uncertainty threshold parameter are marked as high uncertainty exploration areas, and the high uncertainty exploration areas are superimposed and displayed in the three-dimensional comprehensive prediction view in a preset rendering mode; The conservative predictive body and the radical probabilistic body are compared on a three-dimensional voxel basis to obtain the difference probabilistic body. Threshold segmentation and spatial clustering are performed on the differential probability volume, and the high-value clustering regions unique to the radical probability volume are marked as prospect exploration candidate regions; Based on the candidate anomaly 3D entity model, the 3D comprehensive prediction view, and the prospective exploration candidate area, spatial overlay analysis is performed in conjunction with preset target area fusion rules to generate the final mineral exploration prediction scheme; the final mineral exploration prediction scheme includes the category, 3D spatial coordinate range, confidence level, and associated geological attribute information of each mineral exploration target area.

7. The method according to claim 1, characterized in that, The method further includes: Positive sample labels are marked for mineral-bearing locations verified in the field engineering of the final mineral exploration prediction scheme, and negative sample labels are marked for non-mineral-bearing locations. A new verification sample set is constructed by combining the spatial coordinates of the three-dimensional voxel unit corresponding to each label location and the multi-dimensional feature vector. The new verification sample set is used to train the deep neural network prediction model. The systematic deviation is calculated based on the newly added verification sample set and the prior score of geological laws. If the prior score of geological laws exceeds the preset threshold in multiple consecutive verification periods, the corresponding three-dimensional voxel unit is the positive sample label, triggering an update of the formal knowledge model of the metallogenic geological theory. The update of the formal knowledge model of the metallogenic geological theory corresponds to the correction of the rule confidence, the conditional probability table of the probabilistic graphical model, and the quantization function parameters of the prediction elements in the formal knowledge model of the metallogenic geological theory.

8. A solid mineral exploration and prediction system based on geological big data, characterized in that, The system includes: The prior module is used to perform three-dimensional spatial discretization and attribute assignment processing on the multi-source geological big data of the target exploration area to obtain the multi-dimensional feature vector of each three-dimensional voxel unit, and calculate the geological law prior score of each three-dimensional voxel unit based on the formal knowledge model of metallogenic geology theory. The balancing module is used to train and solve the Pareto front of the deep neural network prediction model based on the multi-dimensional feature vector and the geological law prior score, thereby obtaining a set of mineral exploration prediction models. The multi-objective optimization algorithm aims to minimize the geological law conformity loss and maximize the information entropy loss. The prediction module is used to perform parallel mineral exploration target area prediction on the target exploration area based on all Pareto optimal models in the mineral exploration prediction model set, to obtain multiple three-dimensional mineralization probability volumes, and to fuse all the three-dimensional mineralization probability volumes to obtain a mineral exploration prediction spectrum; the mineral exploration prediction spectrum includes a three-dimensional consensus probability volume and a three-dimensional uncertainty quantification volume. The targeting module is used to delineate the target exploration area based on a dynamic threshold and the mineral exploration prediction spectrum, thereby obtaining the final mineral exploration prediction scheme.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images