A mineral resource prediction method and system using digital twinning technology

Abstract

The present application relates to the technical field of resource prediction, and particularly relates to a mineral resource prediction method and system using digital twin technology.A mineral resource prediction system using digital twin technology comprises a mineral digital twin construction module, a simulation data filling module, a mineral resource prediction module and an updating module.A large number of geological model samples are generated by Monte Carlo modeling, a backtracking and forward geological evolution mechanism is introduced, and the optimal target sample is screened from the samples meeting the geological process constraints to fill the unexplored voxels.Through the above method, the digital twin not only has a static three-dimensional attribute mapping capability, but also has a state deduction capability adjusted with the geological evolution process, thereby significantly enhancing the expression accuracy and stability of the subsequent AI prediction model on the occurrence probability of the ore body, and effectively improving the generalization capability of the mineral resource prediction system in the sparse control, complex structure and high uncertainty scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of resource prediction, in particular to a mineral resource prediction method and system using digital twin technology. BACKGROUND

[0002] At present, mineral resource prediction mainly relies on geological models and spatial interpolation methods based on drilling survey data. However, due to the complex geological environment of the mine site and the high cost of drilling, the number of available control points is extremely limited, and the control points are sparsely and unevenly distributed in space, making it difficult to fully reflect the occurrence characteristics of underground ore bodies. Under this background, traditional prediction methods often use Kriging interpolation, inverse distance weighting and other algorithms to fill the unexplored area, but these methods are mainly based on statistical approximation, lack of modeling of geological genetic mechanism, and are difficult to guarantee the geological rationality and spatio-temporal evolution consistency of the generated model, thereby limiting the accuracy and reliability of the prediction results. SUMMARY

[0003] The present application constructs a three-dimensional mineral digital twin model based on historical drilling, geological layers and remote sensing geophysical data in the mining area, introduces survey feature data based on the voxel corresponding to the control anchor, and generates a fault dip angle label by combining the spatial projection of the fault surface, thereby realizing explicit modeling of structural constraints; further, a large number of geological model samples are generated by Monte Carlo modeling, the backtracking and forward geological evolution mechanism is introduced, the optimal target sample is selected from the samples that meet the geological process constraints, and the unexplored voxel is filled with the optimal target sample, realizing high consistency and high rationality of spatial attribute expansion; through the above method, the digital twin not only has the ability of static three-dimensional attribute mapping, but also has the ability of state deduction adjusted with the geological evolution process, thereby significantly enhancing the expression accuracy and stability of the subsequent AI prediction model for the occurrence probability of ore bodies, and effectively improving the generalization ability of the mineral resource prediction system in the sparse control, complex structure and high uncertainty scenarios.

[0004] The present application provides a mineral resource prediction method using digital twin technology, comprising:

[0005] Constructing a mineral digital twin based on historical survey data in the mining area, the mineral digital twin comprising a plurality of voxels, and a survey feature data associated with the voxel corresponding to the control anchor;

[0006] The simulation data filling is performed on the voxels of the mineral digital twin that are not associated with the survey feature data, and the specific content is as follows: a plurality of groups of simulation geological model samples are determined through Monte Carlo sampling modeling, the simulation geological model samples include voxels corresponding to control anchor points in the mineral digital twin and the voxels are filled based on probability distribution sampling, the backtracking evolution operation and the forward evolution operation are performed on each simulation geological model sample to obtain an evolved geological model sample, the target geological model sample is determined based on the difference between the voxels corresponding to the control anchor points in the evolved geological model sample and the voxels corresponding to the control anchor points in the mineral digital twin, the voxels corresponding to the control anchor points in the target geological model sample are replaced by the voxels corresponding to the control anchor points in the mineral digital twin to obtain a complete mineral digital twin;

[0007] The mineral resource prediction operation is performed based on the complete mineral digital twin, and the specific content is as follows: the corresponding mine geological map structure is constructed based on the complete mineral digital twin, and then the mine geological map structure is sent to a mineral resource prediction network for processing to obtain mineral resource label information corresponding to each voxel, the mineral resource label information includes a mineral resource label and an occurrence probability value corresponding to the voxel.

[0008] Preferably, a plurality of groups of simulation geological model samples are determined through Monte Carlo sampling modeling, and the specific steps include the following: the voxels of the mineral digital twin that are not associated with the survey feature data are recorded as empty voxels, a simulation feature data is set for each empty voxel, the simulation feature data is consistent with the data items of the survey feature data, and the simulation feature data is initially empty, for each data item of the simulation feature data, a corresponding probability distribution function is selected, and sampling is performed based on the corresponding probability distribution function to fill the data item of the simulation feature data with the obtained sampling value.

[0009] Preferably, the backtracking evolution operation and the forward evolution operation are performed on each simulation geological model sample to obtain an evolved geological model sample, and the specific functions include the following:

[0010] The position information of each voxel is added at the end of the survey feature data or the simulation feature data corresponding to each voxel in the simulation geological model sample, the corresponding simulation node feature vector is constructed, all simulation node feature vectors are spliced into a first simulation geological feature map from top to bottom in the order of the voxel corresponding number, and a simulation feature edge is constructed between two voxels with a distance less than a distance threshold, a simulation geological adjacency matrix is constructed based on the simulation feature edge, the first simulation geological feature map and the simulation geological adjacency matrix are sent to a backtracking evolution network for processing, and a corresponding second simulation geological feature map is output, and the backtracking evolution network is established based on a graph convolutional neural network.

[0011] The second simulated geological feature map and the simulated geological adjacency matrix are sent to a forward evolution network for processing, and a corresponding evolution geological model sample is output.

[0012] Preferably, the target geological model sample is determined based on the difference between the voxel corresponding to the control anchor point in the evolution geological model sample and the voxel corresponding to the control anchor point in the mineral digital twin, specifically including the following contents: the similarity between the simulated feature data of each control anchor point in the evolution geological model sample and the survey feature data corresponding to the control anchor point in the mineral digital twin is calculated, denoted as a target similarity, and then the average value of all target similarities is calculated, denoted as a comprehensive similarity, and if the comprehensive similarity is higher than a similarity threshold, the evolution geological model sample is recorded as the target geological model sample.

[0013] Preferably, a corresponding mine geological map structure is constructed, specifically by adding position information corresponding to the voxel at the end of the survey feature data or the simulated feature data corresponding to each voxel in the complete mineral digital twin, constructing a corresponding analysis node feature vector, and splicing all analysis node feature vectors into an analysis geological feature map from top to bottom in the order of voxel corresponding numbers, and constructing an analysis feature edge between two voxels with a distance less than a distance threshold, constructing an analysis geological adjacency matrix based on the analysis feature edge, and the analysis geological feature map and the analysis geological adjacency matrix constitute the mine geological map structure.

[0014] Preferably, a mineral digital twin update time point is also set, at which the mineral digital twin is updated according to the record of the actual mining operation of the mine, and the mineral resource prediction network is reinforced learning operation according to the record of the actual mining operation of the mine, and the simulation data filling and mineral resource prediction operation are also performed based on the updated mineral digital twin and the reinforced learning mineral resource prediction network.

[0015] Preferably, the mineral resource prediction network is reinforced learning operation according to the record of the actual mining operation of the mine, specifically including the following operations:

[0016] The actual mineral resource label information corresponding to the voxel is determined according to the record of the actual mining operation of the mine, the actual mineral resource label information includes the actual mineral resource label and the actual occurrence probability value, and the voxel with the determined actual mineral resource label and actual occurrence probability value is recorded as a labeled voxel, the similarity between the actual mineral resource label information corresponding to the labeled voxel and the mineral resource label information output by the labeled voxel at the last mineral digital twin update time point is calculated, denoted as a difference similarity, and then the average value of all difference similarities is calculated, denoted as a comprehensive difference similarity, and the parameters in the mineral resource prediction network are adjusted in the direction of maximizing the comprehensive difference similarity through the gradient ascent method, completing the reinforcement learning operation of the mineral resource prediction network.

[0017] The application also provides a mineral resource prediction system using a digital twin technology, comprising:

[0018] A mineral digital twin construction module is configured to construct a mineral digital twin based on historical survey data of a mining area, the mineral digital twin comprising a plurality of voxels, and a survey feature data is associated with a voxel corresponding to a control anchor point;

[0019] A simulation data filling module is configured to fill simulation data in voxels in the mineral digital twin that are not associated with survey feature data, and the specific content is as follows: a plurality of groups of simulation geological model samples are determined through Monte Carlo sampling modeling, the simulation geological model samples comprising voxels corresponding to control anchor points in the mineral digital twin and the voxels being filled based on probability distribution sampling, backtracking evolution operation and forward evolution operation are performed on each simulation geological model sample to obtain an evolved geological model sample, a target geological model sample is determined based on a difference between voxels corresponding to control anchor points in the evolved geological model sample and the voxels corresponding to the control anchor points in the mineral digital twin, the voxels corresponding to the control anchor points in the target geological model sample are replaced by the voxels corresponding to the control anchor points in the mineral digital twin to obtain a complete mineral digital twin;

[0020] A mineral resource prediction module is configured to perform a mineral resource prediction operation based on the complete mineral digital twin, and the specific content is as follows: a corresponding mine field geological map structure is constructed based on the complete mineral digital twin, and the mine field geological map structure is sent to a mineral resource prediction network for processing to obtain mineral resource label information corresponding to each voxel, the mineral resource label information comprising a mineral resource label and an occurrence probability value corresponding to the voxel;

[0021] An updating module is configured to set a mineral digital twin updating time point, update the mineral digital twin according to a record of actual mining operations of a mine field at the mineral digital twin updating time point, perform reinforcement learning operation on the mineral resource prediction network according to the record of actual mining operations of the mine field, and perform simulation data filling and mineral resource prediction operation based on the updated mineral digital twin and the reinforcement-learned mineral resource prediction network.

[0022] The application has the following advantages:

[0023] The present application is based on the construction of a three-dimensional mineral digital twin model from multi-source data such as historical drilling in the mining area, geological layers, and remote sensing geophysical prospecting. The method introduces survey feature data based on the control of anchor point corresponding voxels, and generates fault dip angle markers by combining fault plane spatial projection, thereby realizing explicit modeling of structural constraints. Further, a large number of geological model samples are generated by Monte Carlo modeling, and the backtracking and forward geological evolution mechanism is introduced to filter the optimal target sample in the sample that meets the geological process constraints, and fill the un-surveyed voxels with the optimal target sample, thereby realizing high-consistency and high-rationality spatial attribute expansion. Through the above method, the digital twin not only has the ability of static three-dimensional attribute mapping, but also has the ability of state deduction adjusted with the geological evolution process, thereby significantly enhancing the expression accuracy and stability of the subsequent AI prediction model for the occurrence probability of ore bodies, and effectively improving the generalization ability of the mineral resource prediction system in the sparse control, complex structure and high uncertainty scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The structural schematic diagram of the mineral resource prediction system using digital twin technology adopted by the embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to enable personnel in the technical field to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0026] Embodiment 1, a mineral resource prediction method using digital twin technology, comprising:

[0027] A mineral digital twin is constructed based on historical survey data of the mining area. The mineral digital twin is stored in a three-dimensional model structure, which includes a plurality of voxels. The voxel corresponding to the control anchor point is associated with a survey feature data. The survey feature data includes a lithology label (such as sandstone and diabase, etc.), a grade value (the content of the target mineral), a density, a porosity, a mineral composition data, and a fault marker information. The mineral composition data refers to the proportion of each mineral obtained by analyzing the sample taken at the position corresponding to the voxel by XRD (X-ray diffraction), MLA (mineral automatic analysis), etc. For example, the proportion of each item such as plagioclase, hornblende, quartz, and calcite. The fault marker information includes the fault dip angle. The lithology label, the grade value, the density, the porosity, and the mineral composition data are obtained by drilling experiments. The fault plane marker is determined according to regional geological map legends and route geological survey data.

[0028] The drilling experiment refers to selecting control anchor points in a mine to be tested, drilling rock and soil samples at positions corresponding to the control anchor points, and recording lithology labels. The rock and soil samples are sent for inspection to determine grade value, density, porosity, and mineral composition data. It should be noted that the selected control anchor points are mapped to voxels, and only a small part of the voxels are occupied.

[0029] The regional geological map legend is compiled by geological survey institutions (such as the Geological Survey Bureau and the Provincial Geological and Mineral Bureau). Route geological survey refers to geological personnel observing, measuring, and recording structural information in the field along a specific route. Both of them can determine the surface lithology in the mine and the fault line of the known exposed mineralization zone, and record the strike, trend, and dip angle of the fault line. The strike, trend, and dip angle of the fault line are mapped into the mineral digital twin, and the fault line is extended in the vertical plane, generally by 1000 meters. Then, based on all the extended fault lines, a curved surface fitting is performed to construct a fault plane network. Then, all the control anchor points corresponding to the voxels in the mineral digital twin are traversed. For each control anchor point corresponding to the voxel, if it is crossed by the fault plane network, the corresponding fault plane dip angle is recorded. It should be noted that the specific operation of the fault line is generally performed in the LeapfrogGeo software.

[0030] The voxels in the mineral digital twin that are not associated with the survey feature data are filled with simulated data. The specific content is as follows: a plurality of groups of simulated geological model samples are determined by Monte Carlo sampling modeling. The simulated geological model samples include the voxels corresponding to the control anchor points in the mineral digital twin and the voxels filled based on probability distribution sampling. The backtracking evolution operation and the forward evolution operation are performed on each simulated geological model sample to obtain an evolved geological model sample. The target geological model sample is determined based on the difference between the voxels corresponding to the control anchor points in the evolved geological model sample and the voxels corresponding to the control anchor points in the mineral digital twin. The lithology label, grade value, density, porosity, mineral composition data, and fault marker information corresponding to each voxel in the target geological model sample conform to the geological evolution rule. The target geological model sample is used to fill the voxels in the mineral digital twin that are not associated with the survey feature data, which has high accuracy and can improve the accuracy of subsequent mineral resource prediction. The voxels corresponding to the control anchor points in the target geological model sample are replaced by the voxels corresponding to the control anchor points in the mineral digital twin to obtain a complete mineral digital twin. The forward evolution operation here refers to the simulated evolution operation on the simulated geological model sample based on the geological evolution rule.

[0031] The mineral resource prediction operation is performed based on the complete mineral digital twin, and the specific content is that a corresponding mine geological map structure is constructed based on the complete mineral digital twin, and then the mine geological map structure is sent to a mineral resource prediction network for processing to obtain mineral resource label information corresponding to each voxel, the mineral resource label information including a mineral resource label and an occurrence probability value corresponding to the voxel, and the mineral resource prediction network is established based on a graph convolutional neural network, and the connection between the voxels can be better analyzed through the graph convolutional neural network;

[0032] The corresponding mine geological map structure is constructed, specifically, the position information corresponding to the voxel is added at the end of the survey feature data or the simulation feature data corresponding to each voxel in the complete mineral digital twin, the analysis node feature vector is constructed, all analysis node feature vectors are spliced into an analysis geological feature map from top to bottom according to the voxel corresponding number sequence, and the analysis feature edge is constructed between two voxels with a distance less than a distance threshold, the distance between the two voxels is calculated based on the position information, and the Euclidean distance algorithm can be used for calculation, the distance threshold is set by an operator, the analysis geological adjacency matrix is constructed based on the analysis feature edge, if the (i, j) element in the analysis geological adjacency matrix is 0, it indicates that there is no analysis feature edge between the i-th voxel and the j-th voxel, and if the (i, j) element in the analysis geological adjacency matrix is 1, it indicates that there is an analysis feature edge between the i-th voxel and the j-th voxel; the analysis geological feature map and the analysis geological adjacency matrix constitute the mine geological map structure;

[0033] The mineral digital twin update time point is set, at the mineral digital twin update time point, the mineral digital twin is updated according to the record of the actual mining operation of the mine, and the mineral resource prediction network is subjected to reinforcement learning operation according to the record of the actual mining operation of the mine, and the simulation data filling and the mineral resource prediction operation are performed based on the updated mineral digital twin and the reinforcement learning mineral resource prediction network; it should be noted that in the actual mining process, the lithology label, grade value, density, porosity, mineral composition data and fault label information corresponding to the voxel can also be determined according to the mined rock and soil samples, and the mineral resource corresponding to the voxel can also be determined;

[0034] The present application is based on the construction of a three-dimensional mineral digital twin model from multi-source data such as historical drilling in the mining area, geological layers, and remote sensing geophysical prospecting, introduces survey feature data based on the control of anchor point corresponding voxels, and generates fault dip angle markers by combining fault plane spatial projection, thereby realizing explicit modeling of structural constraints; further, a large number of geological model samples are generated by Monte Carlo modeling, a backtracking and forward geological evolution mechanism is introduced, the optimal target sample is selected from the samples that meet the geological process constraints, and the unexplored voxels are filled with the optimal target sample, realizing high consistency and high rationality of spatial attribute expansion; through the above method, the digital twin not only has the ability of static three-dimensional attribute mapping, but also has the ability of state deduction adjusted with the geological evolution process, thereby significantly enhancing the expression accuracy and stability of the subsequent AI prediction model for the occurrence probability of ore bodies, and effectively improving the generalization ability of the mineral resource prediction system in the sparse control, complex structure and high uncertainty scenarios.

[0035] A plurality of sets of simulated geological model samples are determined by Monte Carlo sampling modeling, specifically including the following steps:

[0036] The voxels in the mineral digital twin that are not associated with survey feature data are recorded as empty voxels, and a simulated feature data is set for each empty voxel. The data items of the simulated feature data and the survey feature data are consistent, including lithology labels, grade values, densities, porosities, mineral composition data, and fault marker information. The simulated feature data is initially empty. For each data item of the simulated feature data, a corresponding probability distribution function is selected, and sampling is performed based on the corresponding probability distribution function to fill the data item of the simulated feature data with the obtained sampling value.

[0037] The probability distribution function here is generally fitted by normal distribution. First, select all data values of the corresponding data item in the historical record, and then fit all data values of the corresponding data item in the historical record by normal distribution to construct the probability distribution function of the corresponding data item.

[0038] The backtracking evolution operation and the forward evolution operation are performed on each simulated geological model sample to obtain an evolved geological model sample, specifically including the following functions:

[0039] Adding position information corresponding to the voxel at the end of the survey feature data or the simulation feature data corresponding to each voxel in the simulation geological model sample, the position information is generally stored in the form of (x, y, z) here, the corresponding simulation node feature vector is constructed, and all simulation node feature vectors are spliced into a first simulation geological feature map from top to bottom according to the voxel corresponding number sequence, and a simulation feature edge is constructed between two voxels with a distance less than a distance threshold, the distance between the two voxels is calculated based on the position information, and the calculation method can adopt the Euclidean distance algorithm, the distance threshold is set by the operator, a simulation geological adjacency matrix is constructed based on the simulation feature edge, if the ith row and jth column in the simulation geological adjacency matrix is 0, it means that there is no simulation feature edge between the ith voxel and the jth voxel, if the ith row and jth column in the simulation geological adjacency matrix is 1, it means that there is a simulation feature edge between the ith voxel and the jth voxel, the first simulation geological feature map and the simulation geological adjacency matrix are sent into the backtracking evolution network for processing, and the corresponding second simulation geological feature map is output, the backtracking evolution network is established based on the graph convolutional neural network;

[0040] The second simulation geological feature map and the simulation geological adjacency matrix are sent into the forward evolution network for processing, and the corresponding evolution geological model sample is output; the forward evolution network is established based on the graph convolutional neural network.

[0041] Based on the difference between the voxel corresponding to the control anchor point in the evolution geological model sample and the voxel corresponding to the control anchor point in the mineral digital twin, a target geological model sample is determined, which includes the following contents:

[0042] The similarity between the simulation feature data corresponding to each control anchor point in the evolution geological model sample and the survey feature data corresponding to the control anchor point in the mineral digital twin is calculated, the calculation method adopts cosine similarity calculation method, and is recorded as target similarity, then the average value of all target similarities is calculated, which is recorded as comprehensive similarity, if the comprehensive similarity is higher than the similarity threshold, the similarity threshold is determined by the operator, the evolution geological model sample is recorded as the target geological model sample.

[0043] According to the record of the actual mining operation of the mine, the mineral resource prediction network is reinforced learning, which includes the following operations:

[0044] Based on the actual mining operation records of the mine, the actual mineral resource labeling information corresponding to the voxels is determined. The actual mineral resource labeling information includes the actual mineral resource tag and the actual occurrence probability value. The voxels with the determined actual mineral resource tag and actual occurrence probability value are recorded as labeled voxels. The actual occurrence probability value here is generally 1. The similarity between the actual mineral resource labeling information corresponding to the labeled voxel and the mineral resource labeling information output by the labeled voxel at the previous mineral digital twin update time point is calculated and recorded as the difference similarity. Then, the average value of all difference similarities is calculated and recorded as the comprehensive difference similarity. In order to maximize the direction of the comprehensive difference similarity, the parameters in the mineral resource prediction network are adjusted by the gradient ascent method to complete the reinforcement learning operation of the mineral resource prediction network.

[0045] Training the backtracking evolution network and the forward evolution network involves the following steps:

[0046] Several evolutionary training samples were obtained, including pre-geological feature maps and post-geological feature maps. The post-geological feature maps are... Figure 1 Generally, geological feature maps are constructed from survey data obtained by geologists at the current time scale, while pre-geological feature maps are constructed by geologists using professional knowledge and professional geological evolution software to retrospectively analyze evolution; all evolution training samples are combined into an evolution training set;

[0047] The backtracking evolution network is trained using an evolution training set. During training, the post-geological feature map in the evolution training samples is used as input, and the pre-geological feature map in the evolution training samples is used as the training target. It is determined whether the training conditions are met. The training conditions are generally to meet a certain number of training times. If the training conditions are met, the trained backtracking evolution network is output; otherwise, the backtracking evolution network is trained again using the evolution training set.

[0048] The forward evolutionary network is trained using an evolutionary training set. During training, the pre-geological feature map in the evolutionary training sample is used as input, and the post-geological feature map in the evolutionary training sample is used as the training target. It is determined whether the training conditions are met. The training conditions are generally to meet a certain number of training times. If the training conditions are met, the trained forward evolutionary network is output; otherwise, the forward evolutionary network is trained again using the evolutionary training set.

[0049] Training a mineral resource prediction network involves the following steps:

[0050] The mineral resource prediction training sample is obtained, the mineral resource prediction training sample includes a mine geological map structure, the mine geological map structure is constructed by an operator according to actual mining records, and the mineral resource prediction training sample is labeled through mineral resource label information, the mineral resource label information is artificially labeled, all the labeled mineral resource prediction training samples are combined to form a mineral resource prediction training set, the mineral resource prediction network is trained through the mineral resource prediction training set, whether the training condition is met is determined, the training condition is generally that the accuracy of the mineral resource prediction network meets the expectation, if the training condition is met, the trained mineral resource prediction network is output; otherwise, the mineral resource prediction network is continuously trained through the mineral resource prediction training set.

[0051] In embodiment 2, a mineral resource prediction system using a digital twin technology is provided, as shown in Figure 1 , comprising:

[0052] The mineral digital twin construction module is configured to construct a mineral digital twin based on historical survey data of a mining area, the mineral digital twin is stored in a three-dimensional model structure, and includes a plurality of voxels, and a survey feature data is associated with a voxel corresponding to a control anchor point, the survey feature data includes a lithology label (such as sandstone and diabase), a grade value (a content of a target mineral), a density, a porosity, a mineral composition data, and fault marking information, wherein the mineral composition data refers to a proportion of each mineral obtained by analyzing a sample taken at a position corresponding to the voxel through XRD (X-ray diffraction), MLA (mineral automatic analysis), and other experimental means, such as proportions of plagioclase, hornblende, quartz, and calcite, and the fault marking information includes a fault dip angle, the lithology label, the grade value, the density, the porosity, and the mineral composition data are obtained through a drilling experiment, and the fault surface marking is determined according to regional geological map legends and route geological survey data;

[0053] The simulation data filling module is configured to fill simulation data for voxels in the mineral digital twin that are not associated with survey feature data, and specifically includes the following: a plurality of simulation geological model samples are determined through Monte Carlo sampling modeling, the simulation geological model samples include voxels corresponding to control anchor points in the mineral digital twin and the voxels are filled based on probability distribution sampling, backtracking evolution operation and forward evolution operation are performed on each simulation geological model sample to obtain an evolved geological model sample, a target geological model sample is determined based on differences between voxels corresponding to control anchor points in the evolved geological model sample and voxels corresponding to control anchor points in the mineral digital twin, each voxel in the target geological model sample corresponds to a lithology label, a grade value, a density, a porosity, a mineral composition data and a fault marker information that conform to a geological evolution rule, the target geological model sample is used to fill voxels in the mineral digital twin that are not associated with survey feature data, has high accuracy and can improve the accuracy of subsequent mineral resource prediction, and voxels corresponding to control anchor points in the target geological model sample are replaced by voxels corresponding to control anchor points in the mineral digital twin to obtain a complete mineral digital twin; the forward evolution operation refers to simulation evolution operation on the simulation geological model sample based on a geological evolution rule;

[0054] The mineral resource prediction module is configured to perform mineral resource prediction operation based on the complete mineral digital twin, and specifically includes the following: a corresponding mine geological map structure is constructed based on the complete mineral digital twin, and the mine geological map structure is sent to a mineral resource prediction network for processing to obtain mineral resource marker information corresponding to each voxel, the mineral resource marker information includes a mineral resource label and an occurrence probability value corresponding to the voxel, and the mineral resource prediction network is established based on a graph convolutional neural network, and the graph convolutional neural network can better analyze the relationship between voxels;

[0055] The update module is configured to update the mineral digital twin according to records of actual mining operations of the mine at a mineral digital twin update time point, and perform reinforcement learning operation on the mineral resource prediction network according to the records of actual mining operations of the mine.

[0056] It should be understood that those skilled in the art can make improvements or changes according to the above description, and all these improvements and changes shall fall within the protection scope of the appended claims of the present application. Parts not described in detail in the specification belong to the prior art known to those skilled in the art.

Claims

1. A method for predicting mineral resources using digital twin technology, characterized in that, include: A digital twin of mineral resources is constructed based on historical survey data of the mining area. The digital twin of mineral resources includes several voxels, and the voxel corresponding to the control anchor point is associated with a survey feature data. To fill in the voxels of the mineral digital twin with unrelated exploration feature data, the following steps are taken: Several sets of simulated geological model samples are determined through Monte Carlo sampling modeling. The simulated geological model samples include voxels corresponding to the control anchors in the mineral digital twin and voxels corresponding to the sampling and filling based on probability distribution. For each simulated geological model sample, back-evolution and forward evolution operations are performed to obtain evolved geological model samples. Based on the difference between the voxels corresponding to the control anchors in the evolved geological model samples and the voxels corresponding to the control anchors in the mineral digital twin, the target geological model sample is determined. The voxels corresponding to the control anchors in the target geological model sample are replaced with the voxels corresponding to the control anchors in the mineral digital twin to obtain the complete mineral digital twin. The mineral resource prediction operation is performed based on the complete mineral digital twin. Specifically, the corresponding geological map structure of the mine is constructed based on the complete mineral digital twin. Then, the geological map structure of the mine is sent to the mineral resource prediction network for processing to obtain the mineral resource labeling information corresponding to each voxel. The mineral resource labeling information includes the mineral resource label and the probability value of the voxel. For each simulated geological model sample, backward evolution and forward evolution operations are performed to obtain an evolutionary geological model sample, specifically including the following functions: In the simulated geological model sample, the location information of the voxel is added to the end of the survey feature data or simulated feature data corresponding to each voxel, and the corresponding simulated node feature vector is constructed. All simulated node feature vectors are then concatenated from top to bottom according to the voxel number order to form the first simulated geological feature map. Simulated feature edges are constructed between two voxels whose distance is less than the distance threshold. A simulated geological adjacency matrix is ​​constructed based on the simulated feature edges. The first simulated geological feature map and the simulated geological adjacency matrix are fed into the backtracking evolution network for processing, and the corresponding second simulated geological feature map is output. The backtracking evolution network is established based on a graph convolutional neural network. The second simulated geological feature map and the simulated geological adjacency matrix are then fed into the forward evolution network for processing, and the corresponding evolutionary geological model samples are output; the forward evolution network is built based on a graph convolutional neural network; Training the backtracking evolution network and the forward evolution network involves the following steps: Obtain several evolution training samples, including pre-geological feature maps and post-geological feature maps; combine all evolution training samples into an evolution training set; The retrogressive evolutionary network is trained using an evolutionary training set. During training, the post-geological feature map in the evolutionary training sample is used as input, and the pre-geological feature map in the evolutionary training sample is used as the training target. It is determined whether the training conditions are met. If the training conditions are met, the trained retrogressive evolutionary network is output; otherwise, the retrogressive evolutionary network is trained again using the evolutionary training set. The forward evolutionary network is trained using an evolutionary training set. During training, the pre-geological feature map in the evolutionary training sample is used as input, and the post-geological feature map in the evolutionary training sample is used as the training target. It is determined whether the training conditions are met. If the training conditions are met, the trained forward evolutionary network is output; otherwise, the forward evolutionary network is trained again using the evolutionary training set.

2. The mineral resource prediction method using digital twin technology according to claim 1, characterized in that, Several sets of simulated geological model samples were determined through Monte Carlo sampling modeling, specifically including the following steps: Voxels in the mineral digital twin that are not associated with survey feature data are recorded as empty voxels. For each empty voxel, a simulated feature data is set. The data items of the simulated feature data and the survey feature data are consistent, and the simulated feature data is initially empty. For each data item of the simulated feature data, the corresponding probability distribution function is selected, and sampling is performed based on the corresponding probability distribution function. The obtained sampled values ​​are used to fill the data items of the simulated feature data.

3. A mineral resource prediction method using digital twin technology according to claim 2, characterized in that, The target geological model sample is determined based on the difference between the voxels corresponding to the control anchors in the evolutionary geological model sample and the voxels corresponding to the control anchors in the mineral digital twin. Specifically, this includes the following: calculating the similarity between the simulated feature data corresponding to each control anchor in the evolutionary geological model sample and the survey feature data corresponding to it in the mineral digital twin, which is denoted as the target similarity. Then, the average value of all target similarities is calculated and denoted as the comprehensive similarity. If the comprehensive similarity is higher than the similarity threshold, the evolutionary geological model sample is denoted as the target geological model sample.

4. A mineral resource prediction method using digital twin technology according to claim 3, characterized in that, The corresponding geological map structure is constructed as follows: the location information of each voxel is added to the end of the survey feature data or simulation feature data corresponding to each voxel in the complete mineral digital twin, the corresponding analysis node feature vector is constructed, and all analysis node feature vectors are spliced ​​from top to bottom according to the voxel corresponding number order to form an analysis geological feature map. Analysis feature edges are constructed between two voxels whose distance is less than the distance threshold. An analysis geological adjacency matrix is ​​constructed based on the analysis feature edges. The analysis geological feature map and the analysis geological adjacency matrix constitute the geological map structure of the mine.

5. A mineral resource prediction method using digital twin technology according to claim 4, characterized in that, The system also sets update points for the mineral digital twin. At these update points, the mineral digital twin is updated based on the records of actual mining operations at the mine. Furthermore, reinforcement learning operations are performed on the mineral resource prediction network based on the records of actual mining operations at the mine. Finally, simulation data filling and mineral resource prediction operations are performed based on the updated mineral digital twin and the reinforced mineral resource prediction network.

6. A mineral resource prediction method using digital twin technology according to claim 5, characterized in that, The mineral resource prediction network is subjected to reinforcement learning operations based on the actual mining operation records of the mine. Specifically, the operations include the following: Based on the actual mining operation records of the mine, the actual mineral resource labeling information corresponding to the voxels is determined. The actual mineral resource labeling information includes the actual mineral resource tag and the actual occurrence probability value. The voxels with the determined actual mineral resource tag and actual occurrence probability value are recorded as labeled voxels. The similarity between the actual mineral resource labeling information corresponding to the labeled voxel and the mineral resource labeling information output by the labeled voxel at the previous mineral digital twin update time point is calculated and recorded as the difference similarity. Then, the average value of all difference similarities is calculated and recorded as the comprehensive difference similarity. In order to maximize the direction of the comprehensive difference similarity, the parameters in the mineral resource prediction network are adjusted by the gradient ascent method to complete the reinforcement learning operation of the mineral resource prediction network.

7. A mineral resource prediction system utilizing digital twin technology, characterized in that, The system applies a mineral resource prediction method using digital twin technology as described in any one of claims 1-6, comprising: The mineral digital twin construction module is used to construct a mineral digital twin based on the historical survey data of the mining area. The mineral digital twin includes several voxels, and the voxel corresponding to the control anchor point is associated with a survey feature data. The simulated data filling module is used to fill voxels of unrelated exploration feature data in the mineral digital twin. Specifically, it involves determining several sets of simulated geological model samples through Monte Carlo sampling modeling. These simulated geological model samples include voxels corresponding to control anchor points in the mineral digital twin and voxels corresponding to the sampling and filling based on probability distribution. For each simulated geological model sample, a backtracking evolution operation and a forward evolution operation are performed to obtain an evolved geological model sample. Based on the difference between the voxels corresponding to control anchor points in the evolved geological model sample and the voxels corresponding to control anchor points in the mineral digital twin, a target geological model sample is determined. The voxels corresponding to control anchor points in the target geological model sample are replaced with the voxels corresponding to control anchor points in the mineral digital twin to obtain a complete mineral digital twin. The mineral resource prediction module is used to perform mineral resource prediction operations based on a complete mineral digital twin. Specifically, it constructs a corresponding mine geological map structure based on the complete mineral digital twin, and then sends the mine geological map structure into the mineral resource prediction network for processing to obtain the mineral resource labeling information corresponding to each voxel. The mineral resource labeling information includes the mineral resource tag and the probability value assigned to the voxel. The update module is used to set the update time point of the mineral digital twin. At the update time point, the mineral digital twin is updated according to the records of the actual mining operations in the mine. Reinforcement learning operations are also performed on the mineral resource prediction network based on the records of the actual mining operations in the mine. Simulation data filling and mineral resource prediction operations are also performed based on the updated mineral digital twin and the reinforced mineral resource prediction network.

Images