A geological comprehensive information analysis system and method for gold mine exploration

Abstract

The application discloses a geological comprehensive information analysis system and method for gold mine exploration, and relates to the technical field of cross of mineral resources exploration and artificial intelligence. The three-dimensional geological probability digital twin of the present application is obtained by converting multi-source exploration data into a three-dimensional geological probability digital twin carrying the metallogenic favorable degree and the uncertainty variance through a digital twin model construction module. The voxel posterior distribution is dynamically corrected by adopting a memory attenuation Bayesian updating rule based on geological constraints through a real-time data synchronization and updating module. The geological feasibility factor is introduced in the node selection through a Monte Carlo tree search and deep reinforcement learning fusion deduction engine to simulate the expected discounted net income of different engineering deployment sequences. The optimal engineering sequence, resource quantity upgrading probability curve and uncertainty risk map are generated through a quantitative analysis and decision output module. The re-planning is triggered after each step of engineering through a closed-loop self-correcting scheduling controller to realize data-driven quantitative decision and adaptive optimization in the whole exploration process.

Description

Technical Field

[0001] The present invention relates to the cross technical field of mineral resource exploration and artificial intelligence, and specifically relates to a geological comprehensive information analysis system and method for gold deposit exploration. Background Art

[0002] In the actual production of gold exploration enterprises, the prospecting work in deep and covered areas faces prominent problems of complex data types, high interpretation difficulty and long decision-making cycle. A mining area under exploration usually has geological mapping results, ground magnetic survey and gravity data collected in different years, multiple batches of rock geochemical and soil geochemical analysis results, several induced polarization sounding or controlled source audio magnetotelluric sounding profiles, and logging data of dozens of ore-bearing and non-ore-bearing boreholes. Technical personnel need to unify these multi-source information with significant differences in spatial resolution, detection depth and reliability into the same three-dimensional space, delineate favorable metallogenic sections through manual interpretation, and then the chief engineer judges the position, depth and construction sequence of the next batch of boreholes based on personal experience. However, the heterogeneity of the underground metallogenic system and the complexity of ore-controlling structures often lead to differences in the interpretation of the same set of data by different experts, and the manual interpretation process is difficult to quantitatively express the spatial coupling relationship between geophysical and geochemical anomalies and lithology, alteration and structure, resulting in the prediction of concealed ore bodies remaining in the qualitative or semi-quantitative stage for a long time.

[0003] The deficiencies of traditional technical means are particularly obvious in the gradual advancement of exploration engineering. First, the existing comprehensive information analysis mostly adopts static fusion methods such as layer weighted superposition or weights of evidence, and the weight of each data layer is not adjusted after being set at the initial stage of analysis, which cannot reflect the dynamic change of data reliability in different exploration stages. When the high-precision borehole data added in the detailed exploration stage conflicts with the low-resolution geophysical data in the reconnaissance stage, the model lacks an effective conflict resolution mechanism. Second, engineering deployment decisions are highly dependent on the personal experience of the technical person in charge, and there is a lack of quantitative comparison tools for probability simulation of resource upgrading paths under different drilling sequences and input-output ratios. Under the condition of limited budget, it is difficult to answer sequential decision-making problems such as whether to construct deep verification holes first or use geophysical prospecting to encrypt and control structures before arranging drilling, which easily leads to premature consumption of funds in low-probability target areas. Third, when the construction result deviates from the pre-judgment, it often takes several weeks to re-interpret and model to revise the understanding and adjust the subsequent engineering scheme. The delayed response prolongs the entire exploration cycle, during which equipment leasing and personnel costs continue to occur, and the exploration cost increases invisibly.

[0004] The aforementioned problems indicate that the traditional analysis and decision-making model, which relies primarily on human experience and supplemented by static models, is no longer sufficient to meet the stringent requirements of current gold exploration companies for prospecting efficiency and return on investment. There is a need for a closed-loop intelligent analysis system that can dynamically integrate multi-source heterogeneous data, explicitly express underground uncertainties, quantitatively extrapolate the expected returns of exploration engineering sequences, and automatically update the model and replan subsequent deployments after each step of construction. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a comprehensive geological information analysis system and method for gold exploration. This system can transform multi-source exploration data into a three-dimensional geological probabilistic digital twin carrying mineralization favorability and uncertainty variance through a digital twin model construction module. A real-time data synchronization and update module uses a memory decay Bayesian update rule based on geological constraints to dynamically correct the posterior distribution of voxels. A Monte Carlo tree search and deep reinforcement learning fusion inference engine introduces geological feasibility factors during node selection to simulate the expected discounted net benefit of different engineering deployment sequences. A quantitative analysis and decision output module generates the optimal engineering sequence, resource upgrade probability curve, and uncertainty risk map. Finally, a closed-loop self-correcting scheduling controller triggers replanning after each engineering step, achieving data-driven quantitative decision-making and adaptive optimization throughout the entire exploration process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a comprehensive geological information analysis system for gold mine exploration, comprising the following components: Digital twin model construction module: used to receive multi-source exploration data and generate a three-dimensional geological probabilistic digital twin with voxels as units and carrying mineralization favorability and uncertainty variance; Real-time data synchronization and update module: used to receive newly added engineering measured data and dynamically correct the posterior distribution of parameters of each voxel in the digital twin; A Monte Carlo tree search and deep reinforcement learning fusion simulation engine: taking the current digital twin state vector and remaining exploration budget as input, it evaluates the state value and action prior probability through a deep neural network, and drives a Monte Carlo tree search that introduces geological feasibility factors for action selection when selecting nodes, to simulate resource discovery and economic benefits of different engineering deployment sequences. Quantitative analysis and decision output module: used to generate the optimal exploration project sequence, resource quantity upgrade probability curve and uncertainty risk map based on the search statistics returned by the inference engine.

[0007] Furthermore, the digital twin model construction module employs a multi-source data conflict adaptive fusion algorithm to calculate the mineralization favorability for each voxel. This algorithm is expressed as follows: ,in It is a voxel The mineralization favorableness index, with a value between 0 and 1; It represents the total number of exploration data types involved in the fusion; It is the first Class data in voxels The anomaly intensity value is normalized to the interval [0, 1] after being processed by the chromaticity normalization process; It is a voxel Center to the The Euclidean spatial distance of the nearest known control point in the data is in meters. The known control point is the control point within the exploration grid to which the voxel belongs that has been drilled and has a core recovery rate of more than 85%. It is the distance attenuation coefficient, which is the ratio of the nugget value to the sill value obtained by fitting the experimental semi-variogram of the main ore-forming elements in the mining area. It is the first The initial credibility weights of the data source are given by domain experts based on the data exploration depth and resolution. They are updated by minimizing the binary cross-entropy loss between the fusion result and the mineralization label after each new borehole reveals a mineralization mark through online gradient descent. The learning rate of gradient descent is the preset initial learning rate divided by the sum of the total number of historical verifications of the data source and 1. The total number of historical verifications is the cumulative number of times the data source has been verified by subsequent boreholes with or without mineralization. The final mineralization favorability of the voxel and the variance calculated by at least 100 implementations of sequential Gaussian simulation are stored together in the digital twin.

[0008] Furthermore, the digital twin model construction module is configured to: digitally extract geological plans, exploration line profiles, and borehole columnar sections of the exploration area, and construct a three-dimensional triangular mesh surface model of faults, rock masses, and strata; perform constrained tetrahedral subdivision within the triangular mesh, and convert the subdivision results into a hexahedral voxel grid with a side length of 1 / 4 of the average borehole spacing in the mining area; assign lithological values ​​to voxels using known borehole lithology columns, and generate at least 10 equally probable lithological realizations using truncated Gaussian simulation for voxels without borehole control; for each lithological realization, statistically fit a three-parameter log-normal distribution to the gold grade distribution of known samples within each lithological unit, and generate gold grade realization values ​​for each voxel for each lithological realization through sequential Gaussian simulation and record the variance between realizations; and perform diurnal variation analysis on ground magnetic and gravity data respectively. After correction, normal field correction, extension, and polarization processing, three-dimensional physical property inversion is used to obtain magnetic susceptibility and density voxel models, which are then aligned to the geological voxel grid using bilinear interpolation. For rock geochemical and soil geochemical data, contrast values ​​are calculated after iterative removal of outliers using twice the standard deviation, and factor scores are assigned to the voxels corresponding to the first three principal factors. Two-dimensional inversion is performed on electrical resistivity data, and resistivity sections are interpolated to three-dimensional voxels. Within a range of twice the voxel width on both sides of the known ore-controlling fault confirmed by surface mapping or geophysical interpretation, the grid is densified to half the original resolution and attributes are redistributed. Finally, for each voxel, lithology probability vector, gold grade expectation and variance, magnetic susceptibility expectation, density expectation, resistivity expectation, geochemical factor score, and mineralization favorability and variance are stored, forming a digital twin with uncertainty measures.

[0009] Furthermore, the real-time data synchronization and update module adopts a memory decay Bayesian sequential update rule based on geological constraints. When a new measured grade value or physical property value of a spatial point is acquired, the update will be performed. After that, the voxel at that location The posterior probability density function is based on Update; among them Voxel representation The actual mineralization parameters to be estimated; The prior probability density function of the voxel before the update is taken from the distribution currently stored in the digital twin; The standard deviation is the Gaussian likelihood function, and its standard deviation is the estimated value of the standard deviation of repeated measurements of the measuring instrument under similar conditions. When there is no reference estimate, the instrument's nominal repeatability index is taken as the standard deviation. It is the memory decay factor, which is... calculate, The minimum spatial Euclidean distance between the new sampling point and all existing sampling points within the voxel, expressed in meters. The experimental semivariogram range of the main ore-forming elements in this mining area is given in meters. This is the normalized influence coefficient of the distance from the voxel to the nearest known active fault. Normalization is achieved by first calculating the ratio of the planar distance from the voxel to the nearest active fault to the maximum value of this distance for all voxels in the mining area, thus obtaining the normalized distance. Then through Mapped to the interval [0.2, 3.0]; when Approaching 0 Approaching 1 preserves the prior knowledge, when Far exceeding hour Approaching 0 causes historical information to be quickly forgotten; the updated posterior distribution replaces the original distribution, and the neighborhood voxel weights with a correlation coefficient greater than 0.5 are simultaneously corrected and updated.

[0010] Furthermore, the real-time data synchronization and update module is also configured to: listen for data interfaces on a preset message queue, triggering a callback when new measurement records are pushed by the measurement-while-drilling equipment, core scanning equipment, or downhole geophysical exploration equipment; verify the timestamps and spatial coordinates of the records, removing duplicate data frames and data frames whose coordinate drift exceeds twice the nominal positioning accuracy of the measurement equipment at this exploration stage; convert the WGS84 coordinates to mine-area independent grid coordinates and retrieve the voxel indexes that fall into the data; if the voxel has not been directly sampled before, mark it as controlled and use the measured value as hard data input; if it is indirect geophysical data... First, a small-range constraint inversion is performed to obtain the updated property values. Then, the memory decay Bayesian update formula is called to correct the posterior distribution of parameters of the voxel and its neighboring voxels with a spatial correlation coefficient > 0.5. At the same time, the uncertainty scalar field of the corresponding region in the digital twin is updated, and the number of voxels involved in this update and the maximum posterior value change range are recorded. When the maximum posterior value change range exceeds the coefficient of variation of the sequence of the five most recent posterior maximum values ​​of the voxel, a replanning request signal is set and sent to the inference engine through the message queue. Finally, the original measurement data and the parameter snapshots before and after the update are written to the audit log to complete the synchronization.

[0011] Furthermore, the Monte Carlo tree search and deep reinforcement learning fusion inference engine executes a PUCT action selection strategy with geological feasibility correction during the node selection phase of the tree search, namely... ,in The state vector is composed of the mean mineralization favorability of all voxels in the digital twin, the arithmetic mean of the variance of mineralization favorability within the control voxel range of the constructed project, the spatial mask of the constructed project, and the remaining budget, after batch normalization. For optional engineering actions, the action space is discretized into vertical boreholes on grid nodes, geophysical survey lines along grid lines, and trenching points, with additional depth or survey line length parameters. The expected discounted net return of the state-action pair output by the deep value network is calculated with a discount rate of 0.08. The return is the economic value of the discovered gold quantity minus the total engineering cost. In-state output by deep policy network Select action The prior probability; For state The total number of times it is visited in the search tree For state Next action The number of times it was selected; The coefficient is dynamically explored, with an initial value of 1.5. After each round of actual project completion, the standard deviation of the most recent 10 actual returns is calculated, along with the standard deviation of the previous 20 actual returns. If the former is greater than the latter, then... Multiply by 1.2, otherwise multiply by 0.8, and limit the range to [0.5, 2.5]. It is a geological feasibility factor, calculated by first considering the actions... The normalized value is obtained by dividing the vertical distance from the corresponding engineering point to the nearest known ore-controlling fault surface by the median of the alteration width on both sides of the exposed fault zone in the mining area. The normalized value is then multiplied by the membership degree of the alteration buffer zone. The alteration buffer zone is formed by expanding the boundary of the identified alteration mineral assemblage by 1 voxel width. The membership function is a triangular membership function with 1 for the center line and 0 for the edge of the buffer zone, so that the search focuses on the ore-guiding and ore-accommodating spaces.

[0012] Furthermore, the training sample construction process for the deep value network and deep strategy network in the inference engine is as follows: A three-dimensional digital twin state snapshot of each exploration decision moment is extracted from the historical gold mine exploration case library. In addition to voxel attributes, the state snapshot also includes spatial mask channels for constructed engineering works and mineralization marker channels. Its shape is a four-dimensional tensor composed of the number of voxel grids in the X, Y, and Z directions of the current digital twin and nine channels. The channels include mineralization favorability, variance, expected gold grade, expected magnetic susceptibility, expected density, expected resistivity, geochemical principal factor score, engineering mask, and mineralization marker. Action samples... The encoding is a structured vector containing project type, grid coordinates, and depth parameters; the reward value is defined as the final discovered gold metal quantity multiplied by the average monthly settlement price of the Shanghai Gold Exchange Au99.99 contract in the month the training case report is submitted, minus the present value of the total project investment cost, with a discount rate of 0.08; the Monte Carlo reward for each step is calculated backwards from the end of the case, constructing a state-action-reward triplet and storing it in the experience pool; the value network uses a 3D residual convolutional encoder to extract spatial features and concatenates them with the action encoding, outputting a single value through a fully connected layer, trained using mean squared error loss and a weight decay coefficient of 1×10. -4The Adam optimizer is used; the policy network shares the same encoder and outputs the action probability distribution using softmax, with cross-entropy as the training loss; after pre-training, it is deployed to the inference engine, where online decision-making and final reward samples are collected during subsequent actual exploration, with the learning rate being 1 / 10 of the final learning rate in the pre-training stage and not less than 1×10. -5 The learning rate is incrementally fine-tuned, and the network parameters are exponentially averaged every 100 iterations, with checkpoints reserved for rollback.

[0013] Furthermore, the quantitative analysis and decision output module is configured to: extract the number of visits and average state value of each action branch under the root node of the Monte Carlo tree search; collect samples of the mineralized metal quantity from all leaf node returns along each action's branch, fit a log-normal distribution, and calculate the expected value and the lower quantile at 95% confidence level as the lower limit of resource upgrade potential; calculate the return on investment by comparing the expected net present value of each action with the corresponding engineering implementation cost, and sort them in descending order; select the action with the highest return on investment and a cost not exceeding 80% of the remaining budget as the first recommended project, and continue to extract subsequent condition recommendations under its branch according to the same criteria to form a decision tree. The maximum depth of the decision tree is the remaining budget divided by the average cost of a single project, rounded up, and not exceeding 6; Each node on the decision tree is labeled with its engineering coordinates, type, expected cost, and mineralization probability. The mineralization probability is taken as the proportion of mineralization marker samples in the leaf node of that branch, and rendered with different line types. All voxels in the digital twin are traversed, and the standard deviation of the posterior mineralization favorability of each voxel is taken. After Min-Max standardization, a thermal color spectrum is generated according to the rule that standard deviation <0.15 is mapped to blue, 0.15-0.35 is mapped to yellow, and >0.35 is mapped to red. This spectrum is then overlaid on the surface of the 3D geological model of the mining area to form an uncertainty risk map. At the same time, high-risk areas with a standard deviation exceeding 0.35 are highlighted in the 3D model, and textual annotations are used to suggest supplementing the type and scope of geophysical exploration work. Finally, the decision tree, risk map, and resource probability curve are packaged into a PDF report and output to the display terminal.

[0014] Furthermore, the system also includes a closed-loop self-correcting scheduling controller, which is connected to the four modules respectively and configured to: upon receiving a replanning request signal from the real-time data synchronization and update module, or after accumulating at least one set of logging and sampling data from a borehole since the last simulation, retrieve the latest voxel database from the digital twin model construction module and extract the state vector; input the state vector and the current remaining exploration budget into the simulation engine, initiate a Monte Carlo tree search, with the number of search iterations set to 50 times the number of uncontrolled voxels in the digital twin, and at least 2000 times; and output the latest voxel database from the digital twin model construction module and extract the state vector; input the state vector and the current remaining exploration budget into the simulation engine, initiate a Monte Carlo tree search, with the number of search iterations set to 50 times the number of uncontrolled voxels in the digital twin, and at least 2000 times; and output the latest voxel database from the search results. The optimal first-step project and condition sequence are pushed to the quantitative analysis and decision output module to generate visualization instructions; the horizontal offset distance of the recommended first-step project spatial position before and after this replanning is calculated. If the offset distance is less than the voxel grid side length, the cumulative convergence counter is incremented by 1; otherwise, the counter is cleared to zero; when the convergence counter reaches 3 consecutively, and the horizontal offset distance of the recommended first-step project in the last 3 replannings is less than 20% of the design borehole spacing in the current exploration stage, an exploration scheme convergence prompt is generated and it is recommended to terminate the dynamic replanning; the timestamp, triggering reason, recommended project change amount and convergence counter status of each replanning are written to the scheduling log database.

[0015] On the other hand, a comprehensive geological information analysis method for gold mine exploration is characterized by the following specific steps: Step S1: Obtain multi-source data of the exploration area and generate a three-dimensional geological probability digital twin with mineralization favorableness and uncertainty by the digital twin model construction module; Step S2: Take a version snapshot of the current digital twin and generate a state vector. Input the state vector and the remaining budget into the simulation engine and execute a Monte Carlo tree search that incorporates geological feasibility factors. During the search, use a deep value network and a policy network to evaluate the expected discounted net benefit of each engineering action and output the search statistics. Step S3: The quantitative analysis and decision output module extracts and visualizes the optimal first-step engineering action and conditional programming tree from the search statistics; Step S4: Conduct on-site drilling, geophysical exploration, or trenching according to the optimal initial engineering action, and collect measured geological data; Step S5: Input the measured data into the real-time data synchronization and update module, and update and correct the digital twin through memory decay Bayes based on geological constraints; Step S6: Determine whether the termination condition has been met. The termination condition is that the ore section has been delineated after being controlled by at least 3 boreholes and the product of the gold grade and the true thickness of a single project reaches the boundary value of m·g / t in the current gold mine geological exploration specification, or the exploration budget has been consumed to a level lower than the preset single project cost threshold. If the conditions are not met, return to step S2 to realize closed-loop adaptive exploration decision-making.

[0016] Compared with existing technologies, this comprehensive geological information analysis system and method for gold exploration has the following advantages: I. This invention constructs a three-dimensional geological probabilistic digital twin with voxels as units and carrying mineralization favorability and uncertainty variance. Combined with a memory decay Bayesian sequential update rule based on geological constraints, the digital twin can adaptively adjust the balance between prior information and new data based on the spatial distance of sampling points and the influence coefficient of tectonic activity when new borehole or geophysical data is acquired during the exploration process. It can quickly correct the posterior distribution of voxels and simultaneously refresh neighboring voxels. This solves the problem that traditional static models cannot dynamically depict the deepening of exploration understanding and the reduction of local uncertainty, and makes the model always close to the latest revealed real geological features.

[0017] Second, this invention utilizes a Monte Carlo tree search and deep reinforcement learning fusion engine to introduce a geological feasibility factor at the node selection stage. This factor is composed of the normalized distance from the engineering point to the ore-controlling fault surface and the membership degree of the alteration buffer zone. This drives the search to prioritize the ore-guiding and ore-containing space. In conjunction with a closed-loop self-correcting scheduling controller, the optimal engineering sequence is re-deduced based on the updated digital twin after each engineering step. The quantified resource upgrade probability, expected net present value, and return on investment assist decision-making. This transforms exploration deployment from subjective judgment based on personal experience to data-driven closed-loop optimization that is calculable, comparable, and traceable. This significantly reduces the financial risks caused by blind construction and shortens the discovery cycle of hidden ore bodies.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is an overall flowchart of a comprehensive geological information analysis system for gold mine exploration. Figure 2 A flowchart of the replanning process for a closed-loop self-correcting scheduling controller of a comprehensive geological information analysis system for gold exploration. Figure 3 This is a flowchart of a geological comprehensive information analysis method for gold mine exploration. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] Example 1 This embodiment uses a prospecting scenario in the deep and peripheral areas of a production mine as an example to illustrate the application of the system of the present invention in the exploration decision-making of the controlled ore body extending into the deeper areas.

[0023] A gold mine has been in operation since the 1990s. The main ore body has a controlled length of approximately 1200 meters along its strike, and its dip depth has been controlled to the mid-400-meter elevation. The mine processes approximately 300,000 tons of ore annually. Recent mining practices have shown that the main ore body still has a significant tendency to extend into deeper areas. However, due to insufficient control by existing exploration projects, the morphology, occurrence, and grade variation patterns of the deep ore body are still unclear. Continuing to use traditional experience-based drilling methods carries significant exploration risks. The mine plans to conduct deep exploration within the elevation range of -400 to -800 meters, with a budgeted drilling depth of approximately 8000 meters.

[0024] The specific implementation process for making decisions on deep exploration using the system described in this invention is as follows.

[0025] The first step was to construct a three-dimensional geological probabilistic digital twin of the exploration area. This involved collecting accumulated exploration data from the mine over the years, including: eight 1:2000 geological plans, 27 exploration line profiles, 186 borehole columnar sections, approximately 3400 meters of tunnel sketches, high-precision surface magnetic survey data, induced polarization gradient profile data, rock geochemical sampling data, and borehole primary halo data. These geological maps were then digitized, and a constrained Delaunay triangulation algorithm was used to construct a three-dimensional triangular mesh surface model of faults, rock masses, and strata. Constrained tetrahedral subdivisions were applied within the triangular mesh to form a voxel grid with sides of 10 meters, generating approximately 2.86 million voxels in total. For voxels without direct borehole control, truncated Gaussian simulation was used to generate 15 equally probable lithological realizations. For each lithological realization, the gold grade distribution of known samples within each lithological unit was statistically analyzed, fitted with a three-parameter log-normal distribution, and sequential Gaussian simulation was used to generate gold grade realization values ​​for each voxel for each lithological realization, recording the variance between realizations. Diurnal variation correction, normal field correction, and polarization processing were applied to the ground magnetic survey data, and a magnetic susceptibility voxel model was obtained using three-dimensional property inversion. For the rock geochemical data, contrast values ​​were calculated after iterative elimination of outliers using twice the standard deviation, and factor scores were assigned to the voxels corresponding to the first three principal factors. Finally, for each voxel, the following data were stored: lithological probability vector, gold grade expectation and variance, magnetic susceptibility expectation, density expectation, resistivity expectation, geochemical factor score, and mineralization favorability and variance. The mineralization favorability was calculated using a multi-source data conflict adaptive fusion algorithm, with the number of data types M participating in the fusion set to 6, the distance attenuation coefficient η set to 0.015, and the initial confidence weight... Based on the historical prediction accuracy of various data sources, the weights for lithology and grade were set to 0.25, physical property data to 0.15, and geochemical data to 0.10.

[0026] The second step involves taking a version snapshot of the current digital twin and generating a state vector. The state vector contains the expected mineralization favorability, expected gold grade, lithological probability, and uncertainty variance information for all voxels in the digital twin. After 3D convolutional encoding, it forms a fixed-length feature vector. The state vector, along with the remaining exploration budget, is then input into a Monte Carlo tree search and deep reinforcement learning fusion inference engine. The deep value network and deep policy network in the inference engine use a 3D residual convolutional encoder to extract spatial features, and have already been pre-trained based on historical exploration case libraries of this and neighboring mining areas. The number of search iterations is set to 50 times the number of uncontrolled voxels in the digital twin, approximately 98.5 million iterations. During the node selection phase of the tree search, a PUCT action selection strategy for geological feasibility correction is executed, dynamically exploring coefficients. The value is 1.2, which is the geological feasibility factor. The value is a continuous value between 0 and 1, calculated based on factors such as whether the construction location of the selected project meets the minimum hole depth requirement, whether it avoids known goaf areas, and whether it is located in a workable terrain.

[0027] The third step involves the quantitative analysis and decision output module extracting the optimal initial engineering action from the search statistics. The inference engine returns the number of visits and average state value of each action branch under the root node. For each action branch, it collects samples of the mineralized metal quantity from all leaf node reports, fits a log-normal distribution, and calculates the expected value and the lower quantile at a 95% confidence level. The calculation shows that the recommended initial engineering action is to construct a slant borehole at an elevation of -520 meters on exploration line 7. The expected net present value of this action branch is 28.6 million yuan, the engineering implementation cost is approximately 980,000 yuan, the return on investment is approximately 29.2%, and the probability of mineralization is 72.6%. Following the same criteria, subsequent condition recommendations are extracted under this branch, forming a decision tree containing 6 boreholes, with a total budget of approximately 5.72 million yuan and an expected cumulative mineralized metal quantity of approximately 3.8 tons.

[0028] The fourth step involves conducting on-site drilling operations according to the optimal procedures outlined in the initial engineering steps. An XY-6B core drilling rig, equipped with NQ-caliber wireline coring tools, was used, and drilling was conducted strictly according to the designed hole depth, azimuth, and zenith angle. Core logging data was collected every 5 meters during drilling, and a basic analytical sample was collected every 10 meters. The samples were sent to a CMA-certified laboratory for gold grade determination using foam adsorption-atomic absorption spectrometry.

[0029] The fifth step involves transmitting the measured data to the real-time data synchronization and update module. After drilling is completed, the measurement-while-drilling (MWD) equipment automatically pushes out inclination data, the core scanning equipment pushes out hyperspectral core image data, and the laboratory pushes out gold grade analysis results via a data interface. Upon receiving new data in the preset message queue, the system triggers a callback to verify the recorded timestamp and spatial coordinates. After confirming the data's validity, it calls the memory decay Bayesian update formula to correct the posterior distribution of parameters for the voxel and its neighboring voxels with a spatial correlation coefficient greater than 0.5. (Memory decay factor) The value was set to 0.85. After the update, the standard deviation of voxel mineralization favorability on the borehole control path decreased from 0.31 to 0.12, significantly reducing uncertainty.

[0030] Step 6: Determine if the termination condition has been met. Following control by the first borehole, a mineralized section was encountered at an elevation of -505 meters to -538 meters. The product of the gold grade and true thickness for a single drilling operation was 2.8 m·g / t, which did not yet reach the boundary m·g / t value in the current "Specification for Geological Exploration of Gold Deposits" (DZ / T0205-2020), thus the termination condition was not met. The system returns to step 2, initiating a new round of simulation based on the updated digital twin to achieve closed-loop adaptive exploration decision-making.

[0031] Through iterative processes described above, the system recommended and implemented a total of 7 boreholes. The 4th borehole controlled a thick mineralized section at an elevation of -620 meters, with a single-project gold grade multiplied by the true thickness reaching 14.6 m·g / t, thus delineating the industrial ore body. After the completion of the 7th borehole, ore sections controlled by more than 3 boreholes and meeting the boundary m·g / t values ​​were delineated. The system generated a convergence prompt for the exploration plan and suggested terminating dynamic replanning. Compared with traditional experience-based borehole layout schemes, the system of this invention reduced drilling workload by approximately 50% and increased the mineralization rate from approximately 38% using traditional methods to 71.4%.

[0032] In summary, this embodiment verifies that the system of the present invention can effectively reduce exploration risks, increase mineral discovery rate, and save exploration costs in the deep mineral exploration scenario of a production mine by fusing multi-source geological information through digital twins, optimizing engineering deployment sequences through the fusion of Monte Carlo tree search and deep reinforcement learning, and achieving closed-loop adaptive decision-making through real-time data updates.

[0033] Example 2 This embodiment uses a gold mine preliminary exploration scenario in a new district as an example to illustrate the application of the system of the present invention in the initial exploration decision-making under the condition of only surface geological and geophysical exploration data.

[0034] The exploration area, covering approximately 15.6 square kilometers, is located within a gold metallogenic belt on the southern margin of the North China Craton. The exposed strata are primarily Neoarchean Taihua Group metamorphic rocks. Several gold-bearing quartz vein-type mineralization clues have been discovered within the area, but no drilling has yet commenced, making it a typical new area in the preliminary exploration / general survey stage. Traditional preliminary exploration work typically involves deploying a small number of verification boreholes in favorable locations based on surface mineralization clues and experience. However, given the low level of geological understanding and high exploration risks in this new area, scientifically determining the location and sequence of the first batch of verification boreholes is a crucial issue that urgently needs to be addressed.

[0035] The specific implementation process for conducting new area exploration and decision-making using the system described in this invention is as follows.

[0036] The first step is to construct a three-dimensional geological probabilistic digital twin of the exploration area. This involves collecting multi-source exploration data available in the area, including: a 1:50,000 regional geological map and a 1:10,000 mining area geological map, a 1:2,000 topographic and geological sketch map, high-precision ground magnetic survey data, induced polarization gradient surface data, soil geochemical measurement data, and surface rock sample analysis data. Since there is currently no borehole control in the area, the construction of the digital twin faces the challenge of data sparsity. The system adopts the following strategy: after digitizing and extracting the strata and fault boundaries from the geological plan, an initial three-dimensional framework model is constructed. Constrained tetrahedral subdivision is applied to the triangular mesh to form a voxel grid with a side length of 20 meters, generating approximately 1.95 million voxels. Utilizing regional geological patterns and surface outcrop information, 20 equally probabilistic lithological realizations are generated using truncated Gaussian simulation, with each realization using surface geological boundaries as hard constraints. Due to the lack of borehole grade data, the initial assignment of gold grade was primarily based on the analysis of surface rock samples and inferences from soil geochemical anomalies. For each voxel, based on its lithological unit and tectonic location, and referencing the grade statistical parameters of similar deposits in the region, a sequential Gaussian simulation was used to generate the realized gold grade value, with a large initial variance assigned to reflect the high degree of uncertainty at depth. Conventional processing was performed on surface magnetic and induced polarization data to obtain voxel models of magnetic susceptibility and resistivity. Soil geochemical data, after iterative removal of outliers by two standard deviations, was used to calculate contrast values, and factor scores were assigned to the voxels corresponding to the first three principal factors. In the mineralization favorability fusion algorithm, the types of data involved in the fusion were... The value is 5, which is the distance attenuation coefficient. The initial confidence weight is set to 0.008 based on the predictive effectiveness of each data source in the region. The geochemical factor has the highest weight of 0.30, the lithology weight is 0.25, the surface mineralization clue weight is 0.20, and the physical property data weight is 0.125 each.

[0037] The second step involves inputting the state vector and the remaining exploration budget into the inference engine. The number of search iterations is set to 50 times the number of uncontrolled voxels, approximately 97.5 million iterations. Since there are no historical exploration cases available for pre-training in this area, the deep value network and policy network are pre-trained using historical exploration data from neighboring mining areas in the region through transfer learning. During the inference process, the learning rate is 1 / 10 of the final learning rate from the pre-training stage (i.e., 1 × 10⁻⁶). -5 Perform online incremental fine-tuning.

[0038] The third step involves the quantitative analysis and decision-making output module extracting the optimal initial engineering action. Based on calculations, the recommended initial engineering action is to construct a vertical borehole approximately 120 meters northeast of the intersection of the F3 and F5 faults. This location is situated in the overlapping area of ​​soil geochemical Au anomalies and induced polarization gradient high polarizability anomalies, and is located on the footwall of the fault intersection. The geological feasibility factor is [not specified]. The calculated value is 0.92. The expected net present value of this action branch is 4.2 million yuan, the project implementation cost is approximately 560,000 yuan, the return on investment is approximately 7.5 times, and the probability of encountering minerals is 43.7%. Further recommendations for subsequent conditions are extracted under this branch, forming a decision tree containing 5 boreholes, with a total budget of approximately 2.75 million yuan.

[0039] The fourth step involves constructing the project according to the optimal initial steps. An XY-4 core drilling rig, equipped with HQ-caliber wireline coring tools, is used. The borehole curvature is strictly controlled during construction. The total core recovery rate is no less than 85%. Cores are arranged, numbered, photographed, and then boxed and stored according to their sequence.

[0040] Steps five and six are essentially the same as in Example 1. The first borehole encountered a gold-mineralized section at a depth of 320 to 358 meters. The product of the gold grade and the true thickness of the single-project borehole was 1.8 m·g / t. Although this did not reach the boundary industrial index, it confirmed the existence of a deep gold-mineralized system. After Bayesian updating of the digital twin, the distribution of deep mineralization favorability was significantly adjusted, and the area of ​​the originally highly uncertain region was reduced by about 40%. The system returned to the second step for re-analysis. The second recommended first-step project was adjusted to be carried out at a location approximately 280 meters northwest of the first borehole, as the simulation engine determined that this location was more likely to control the main mineralization zone. After iterative drilling of four boreholes, the third borehole controlled the industrial ore body at a depth of 410 to 445 meters, and the fourth borehole verified the extension of the ore body's strike, meeting the termination conditions.

[0041] In summary, this embodiment verifies that the system of the present invention can still construct a digital twin carrying uncertainty through multi-source data fusion under the conditions of no borehole control and extremely sparse data in the new area, optimize the location of the first batch of verification boreholes through Monte Carlo tree search, and gradually converge to the mineralized body through the method of exploration and updating simultaneously, effectively solving the decision problem of "where to drill the first hole" in the preliminary exploration stage of the new area.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A comprehensive geological information analysis system for gold mine exploration, characterized in that, The system includes the following components: Digital twin model construction module: used to receive multi-source exploration data and generate a three-dimensional geological probabilistic digital twin with voxels as units and carrying mineralization favorability and uncertainty variance; Real-time data synchronization and update module: used to receive newly added engineering measured data and dynamically correct the posterior distribution of parameters of each voxel in the digital twin; A Monte Carlo tree search and deep reinforcement learning fusion simulation engine: taking the current digital twin state vector and remaining exploration budget as input, it evaluates the state value and action prior probability through a deep neural network, and drives a Monte Carlo tree search that introduces geological feasibility factors for action selection when selecting nodes, to simulate resource discovery and economic benefits of different engineering deployment sequences. Quantitative analysis and decision output module: used to generate the optimal exploration project sequence, resource quantity upgrade probability curve and uncertainty risk map based on the search statistics returned by the inference engine.

2. The geological comprehensive information analysis system for gold mine exploration according to claim 1, characterized in that, The digital twin model construction module employs a multi-source data conflict adaptive fusion algorithm to calculate the mineralization favorability for each voxel. This algorithm is expressed as follows: ,in It is a voxel Mineralization favorableness index; It represents the total number of exploration data types involved in the fusion; It is the first Class data in voxels The anomaly intensity value is normalized to the interval [0, 1] after the chromaticity value is normalized. It is a voxel Center to the Euclidean spatial distance between the nearest known control points in the data set; It is the distance attenuation coefficient; It is the first Initial credibility weights for class data sources.

3. The geological comprehensive information analysis system for gold mine exploration according to claim 2, characterized in that, The digital twin model construction module is further configured to: digitally extract geological plans, exploration line profiles, and borehole columnar sections of the exploration area, and construct a three-dimensional triangular mesh surface model of faults, rock masses, and strata; perform constrained tetrahedral subdivision within the triangular mesh; assign lithological values ​​to voxels using known borehole lithology columns, and generate at least 10 equally probable lithological realizations for voxels without borehole control using truncated Gaussian simulation; for each lithological realization, statistically fit a three-parameter log-normal distribution to the gold grade distribution of known samples within each lithological unit, and generate gold grade realization values ​​for each voxel for each lithological realization through sequential Gaussian simulation and record the variance between realizations; and perform ground magnetic surveys and gravity measurements... After diurnal variation correction, normal field correction, extension and polarization processing, three-dimensional physical property inversion was used to obtain magnetic susceptibility and density voxel models, which were then aligned to the geological voxel grid using bilinear interpolation. For rock geochemical and soil geochemical data, contrast values ​​were calculated after iterative removal of outliers by 2x standard deviation, and factor scores were assigned to the voxels corresponding to the first three principal factors. Two-dimensional inversion was performed on electrical resistivity data, and resistivity sections were interpolated to three-dimensional voxels. Finally, for each voxel, lithology probability vector, gold grade expectation and variance, magnetic susceptibility expectation, density expectation, resistivity expectation, geochemical factor score, and mineralization favorability and variance were stored, forming a digital twin with uncertainty measurement.

4. The geological comprehensive information analysis system for gold mine exploration according to claim 1, characterized in that, The real-time data synchronization and update module adopts a memory decay Bayesian sequential update rule based on geological constraints. When a new measured grade value or physical property value of a spatial point is acquired, the update will be performed. After that, the voxel at that location The posterior probability density function is based on Update; among them Voxel representation The actual mineralization parameters to be estimated; This is the prior probability density function of the voxel before the update; Let Gaussian likelihood function be used. It is a memory decay factor.

5. The geological comprehensive information analysis system for gold mine exploration according to claim 4, characterized in that, The real-time data synchronization and update module is also configured to: listen to the data interface on a preset message queue, trigger a callback when the measurement while drilling equipment, core scanning equipment, or downhole geophysical exploration equipment pushes a new measurement record; verify the timestamp and spatial coordinates of the record; if the voxel has not been directly sampled before, mark it as controlled and use the measurement value as hard data input; if it is indirect geophysical data, first perform a small-range constraint inversion to obtain the updated physical property value; call the memory decay Bayes update formula to correct the posterior distribution of parameters of the voxel and its neighboring voxels with a spatial correlation coefficient > 0.5; at the same time, update the uncertainty scalar field of the corresponding region in the digital twin, and record the number of voxels involved in this update and the maximum posterior value change range.

6. The geological comprehensive information analysis system for gold mine exploration according to claim 1, characterized in that, The Monte Carlo tree search and deep reinforcement learning fusion inference engine executes a PUCT action selection strategy for geological feasibility correction during the node selection phase of the tree search, namely... ,in It is a state vector; This is an optional engineering action; The expected discounted net return of state-action pairs output by the deep value network; In-state output by deep policy network Select action The prior probability; For state The total number of times it is visited in the search tree For state Next action The number of times it was selected; For dynamic exploration coefficients; It is a geological feasibility factor.

7. A comprehensive geological information analysis system for gold mine exploration according to claim 6, characterized in that, The training sample construction process of the deep value network and deep policy network in the inference engine is as follows: extract the three-dimensional digital twin state snapshot of each exploration decision moment from the historical gold mine exploration case library; calculate the Monte Carlo reward of each step from the end of the case, construct the state-action-reward triplet and store it in the experience pool. The value network extracts spatial features using a 3D residual convolutional encoder, which is then concatenated with action codes. A single value is output through a fully connected layer. It is trained using mean squared error loss and a weight decay coefficient of 1×10⁻⁶. -4 The Adam optimizer is used; the policy network shares the same encoder and outputs the action probability distribution using softmax, with cross-entropy as the training loss; after pre-training, it is deployed to the inference engine, where online decision-making and final reward samples are collected during subsequent actual exploration, with the learning rate being 1 / 10 of the final learning rate in the pre-training stage and not less than 1×10. -5 The learning rate is incrementally fine-tuned.

8. A comprehensive geological information analysis system for gold mine exploration according to claim 1, characterized in that, The quantitative analysis and decision output module is configured as follows: extract the number of visits and average state value of each action branch under the root node of the Monte Carlo tree search; for each action, collect samples of mineralized metal quantity from all leaf node returns along its branches, fit a log-normal distribution, and calculate the expected value and the lower quantile at 95% confidence level as the lower limit of resource upgrade potential; calculate the return on investment by comparing the expected net present value of each action with the corresponding engineering implementation cost, and sort them in descending order; select the action with the highest return on investment and a cost not exceeding 80% of the remaining budget as the first recommended project, and continue to extract subsequent condition recommendations under its branches according to the same criteria to form a decision tree; label each node on the decision tree with engineering coordinates, type, expected cost, and mineralization probability, with the mineralization probability taken as the proportion of mineralization marker samples in the leaf nodes of that branch, and render them with different line types; traverse all voxels of the digital twin and take the standard deviation of the posterior mineralization favorability of each voxel; finally, package the decision tree, risk map, and resource probability curve into a PDF report and output it to the display terminal.

9. A comprehensive geological information analysis system for gold mine exploration according to claim 1, characterized in that, The system also includes a closed-loop self-correcting scheduling controller, which is connected to the four modules and configured to: upon receiving a replanning request signal from the real-time data synchronization and update module, or after accumulating at least one set of logging and sampling data from a borehole since the last simulation, retrieve the latest voxel database from the digital twin model construction module and extract the state vector; input the state vector and the current remaining exploration budget into the simulation engine, initiate a Monte Carlo tree search, and set the number of search iterations to 50 times the number of uncontrolled voxels in the digital twin; push the optimal first-step project and condition sequence output by the search to the quantitative analysis and decision output module to generate visualization instructions; calculate the horizontal offset distance of the recommended first-step project before and after this replanning; if the offset distance is less than the voxel grid side length, increment the cumulative convergence counter by 1; otherwise, clear the counter; when the convergence counter reaches 3 consecutively, and the horizontal offset distance of the recommended first-step project in the last 3 replannings is less than 20% of the design borehole spacing in the current exploration stage, generate an exploration scheme convergence prompt and suggest terminating dynamic replanning.

10. A method for analyzing comprehensive geological information for gold exploration, applicable to the comprehensive geological information analysis system for gold exploration as described in any one of claims 1-9, characterized in that, The specific steps of this method are as follows: Step S1: Obtain multi-source data of the exploration area and generate a three-dimensional geological probability digital twin with mineralization favorableness and uncertainty by the digital twin model construction module; Step S2: Take a version snapshot of the current digital twin and generate a state vector. Input the state vector and the remaining budget into the simulation engine and execute a Monte Carlo tree search that incorporates geological feasibility factors. During the search, use a deep value network and a policy network to evaluate the expected discounted net benefit of each engineering action and output the search statistics. Step S3: The quantitative analysis and decision output module extracts and visualizes the optimal first-step engineering action and conditional programming tree from the search statistics; Step S4: Conduct on-site drilling, geophysical exploration, or trenching according to the optimal initial engineering action, and collect measured geological data; Step S5: Input the measured data into the real-time data synchronization and update module, and update and correct the digital twin through memory decay Bayes based on geological constraints; Step S6: Determine whether the termination condition has been met. The termination condition is that the ore section has been delineated after being controlled by at least 3 boreholes and the product of the gold grade and the true thickness of a single project reaches the boundary value of m·g / t in the current gold mine geological exploration specification, or the exploration budget has been consumed to a level lower than the preset single project cost threshold. If the conditions are not met, return to step S2 to realize closed-loop adaptive exploration decision-making.

Images