A three-dimensional geological model construction and resource evaluation method for precise prospecting

Abstract

The application discloses a kind of three-dimensional geological model construction and resource evaluation methods for accurate prospecting, belong to mine three-dimensional geological modeling and resource reserve estimation technical field. Including: obtaining multi-source data to construct basic information database;With geological interpretation line as prior constraint, construct initial implicit surface, with new data as dynamic control point, drive surface local deformation by elastic deformation, realize ore body model update reconstruction;With sliding window division space, extract experimental variogram, after optimal parameter output by depth parameter prediction network, carry out kriging evaluation and parallel comparison verification;Based on roadway point cloud, generate global feature descriptor to reconstruct roadway model, with reconstructed wall surface reverse constraint ore body boundary to carry out mutual feedback correction;With new data timestamp trigger incremental update, superimpose and fuse three kinds of models to generate comprehensive three-dimensional model.The application realizes the substantial change from stage display to continuous production guidance tool of three-dimensional geological achievement.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional geological modeling and resource reserve estimation technology for mines, and more specifically, to a method for constructing three-dimensional geological models and assessing resources for precise mineral exploration. Background Technology

[0002] With the gradual depletion of shallow mineral resources, concealed ore bodies in deep and complex geological conditions have become a key focus of geological prospecting. However, deep exploration faces challenges such as difficulty in obtaining geological information, high engineering costs, and complex structural analysis. The traditional working mode, which mainly relies on two-dimensional maps and manual estimation, can no longer meet the actual needs of modern mines for precise exploration and efficient development.

[0003] Three-dimensional geological modeling technology integrates multi-source data from drilling, geophysical exploration, geochemical exploration, surveying, and mining engineering to construct three-dimensional digital models of ore bodies, structures, and tunnels in a computer environment, providing a scientific basis for mineralization prediction, resource estimation, and mining design. Currently, international commercial software such as Leapfrog, Surpac, and Datamine are widely used in the global mining industry. Surface modeling, block models, and geostatistical kriging based on TIN have become the mainstream techniques for ore body morphology reconstruction and resource reserve estimation. The latter is recognized as an international standard method due to its unbiased optimal interpolation characteristics. Despite the increasing maturity of the technology system, the following prominent problems still exist in its practical application in domestic mines: First, the diverse sources, formats, and coordinates of data from drilling, geophysical exploration, geochemical exploration, and point cloud measurement, coupled with weak data fusion and standardized preprocessing, restrict modeling efficiency and reliability. Second, domestic mines heavily rely on imported software, resulting in high costs, limited services, and dependence on foreign suppliers for upgrades. More critically, there are compatibility bottlenecks between imported software and existing domestic geological database formats and industry standards, making data interoperability difficult and hindering the achievement of fully domestically produced and controllable processes. Third, after completing the initial modeling, most mines lack dynamic supplementation and model correction mechanisms for information on new boreholes, new roadways, and newly discovered ore bodies, leading to a gradual disconnect between the model and actual production. The results remain at the level of mere "display," failing to fully realize their practical value. Fourth, resource reserve estimation still relies mainly on traditional methods such as the cross-section method and the geological block method, with insufficient systematic introduction of the kriging method and a lack of comparative verification with traditional methods, affecting the scientific validity and international comparability of the estimation results. How to overcome the bottleneck of multi-source data fusion, realize the dynamic updating and fine reconstruction of existing models, establish a standardized resource estimation process with the Kriging method as the core, and promote the substantial transformation of 3D geological results from "emphasizing display" to "emphasizing practicality" are the urgent problems to be solved in the field of precision mining. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration. Through multi-source data fusion, adaptive surface reconstruction, dynamic kriging estimation, intelligent point cloud modeling, and incremental update mechanisms, an intelligent three-dimensional geological model system that dynamically evolves with production is constructed.

[0005] This invention discloses a method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration, comprising the following steps: Data from drilling, geophysical exploration, geochemical exploration, laser point cloud, and mining logging are acquired, and after unified coordinate normalization, gross errors are eliminated through adaptive filtering to construct a basic information database. Using existing geological interpretation lines as prior constraints, an initial implicit surface is constructed through radial basis function interpolation. Newly acquired borehole ore-bearing points and anomalous boundaries are used as dynamic control points. Local deformation of the surface is driven by elastic deformation to update the existing ore body model and reconstruct the model of the newly discovered ore body. Based on the ore body model, borehole grade data is extracted, and the experimental variogram is calculated in three dimensions. The space is divided by a sliding window, and the experimental variogram curve and statistical characteristics within the window are input into a pre-trained depth parameter prediction network to output the optimal variogram parameters for the window. After determining the optimal search ellipsoid based on the optimal variogram parameters of the window, ordinary Kriging interpolation is performed, and the nearest neighbor method is used in parallel for comparison and verification. If the deviation between the two exceeds the threshold, the variogram is backfitted and the converged Kriging estimation result is used as the basis for the reserves. Global feature descriptors are generated based on laser point clouds of underground roadways. The cross-sectional contour and centerline are predicted and the roadway model is reconstructed based on the global feature descriptors. The reconstructed wall is used as the real boundary to constrain the ore body boundary convergence and perform geological-engineering mutual feedback correction. Local incremental updates are triggered by new data timestamps, and the updated geological entities, block attributes, and tunnel models are superimposed and merged with a unified control network as anchor points to generate a comprehensive three-dimensional visualization model.

[0006] This plan involves constructing a basic information database, specifically including: Acquire five types of data: drilling, geophysical exploration, geochemical exploration, laser point cloud, and mining logging. Transform unstructured data into structured data tables and establish original index ledgers according to type. Using the downhole high-precision control network as a benchmark, the coordinates of each data source are uniformly normalized to the control network coordinate system, and the length is compared and verified using the actual measured length downhole as the verification benchmark. When the deviation exceeds the limit, geometric correction is applied. An adaptive robust filter based on median absolute deviation is used to remove outliers. The median of the data sequence is used as the center and the median absolute deviation is used as the discrete measure. The cutoff range is adaptively determined according to the thickness of the tail of the data distribution. Outliers outside the effective interval are removed, and gaps are filled by local weighted addition of adjacent effective data. A unique identifier code is generated for each data item, embedding spatial grid location information in the code, establishing spatial proximity associations, and constructing a stratigraphic sequence constraint network and shaft connectivity topology to generate a basic information database.

[0007] In this scheme, existing geological interpretation lines are used as prior constraints, and an initial implicit surface is constructed through radial basis function interpolation, specifically including: Existing geological profile interpretation lines are extracted and discretized into a three-dimensional spatial point sequence. The credibility is graded and weighted according to the degree of agreement with borehole control points. Using quadratic radial basis functions as kernel functions, three-dimensional implicit functions are constructed by taking the discrete points of the interpretation line as constraint points. The geological interface is defined as the zero isosurface with F(x)=0. Anisotropic distance scaling factors are set along the strike, dip and thickness directions of the ore body, and the weight coefficients are solved to construct the initial implicit surface.

[0008] In this scheme, the newly acquired borehole ore-bearing points and anomalous boundaries are used as dynamic control points. Local surface deformation is driven by elastic deformation to update the existing orebody model and reconstruct the model of the newly discovered orebody. Specifically, this includes: The newly acquired borehole mineralization points and geophysical and geochemical anomaly interpretation boundaries are used as dynamic control points and assigned confidence levels, and then fused with prior constraints with confidence levels. A tight-support radial basis function is constructed with each dynamic control point as the center and a finite influence radius is set. Anisotropic propagation radii are defined along the strike, dip and thickness directions of the ore body. The weighted distance of any point in space relative to each dynamic control point is calculated. The weighted distance is substituted into the tight-support radial basis function to obtain the deformation influence coefficient of each control point on the selected point. The influence coefficients of all control points are superimposed to obtain the total deformation displacement at the selected point. The total deformation displacement at each point is used as the adjustment amount of the implicit function value to perform point-by-point displacement correction on the initial implicit surface, thereby realizing local adaptive correction of the existing ore body model. For newly discovered ore bodies without a prior model, a cross-sectional sequence is generated with the ore-bearing point group and the anomaly center line as the backbone. The ore-bearing points falling within the search zone of each cross section are projected onto the cross-sectional plane to generate an initial profile. The contour range is constrained by the geochemical anomaly contour lines as the external geometric boundary. Corresponding points are matched for adjacent cross-sectional profiles and transition surfaces between cross sections are generated by cubic spline interpolation. The first and last cross sections are closed to form a three-dimensional solid model of the newly discovered ore body.

[0009] In this scheme, obtaining the optimal mutation function parameters for the window includes: The updated orebody model is discretized into regular block units. Drill sample data falling within the orebody area are extracted and sample length is standardized. The three-dimensional experimental variogram curves are calculated along the strike, dip and thickness directions of the orebody. The ore body space is uniformly divided into overlapping three-dimensional sliding windows, and the discrete value sequence of experimental variogram curves, sample statistical features and spatial location codes in each window are extracted as input feature vectors. A deep parameter prediction network consisting of a one-dimensional convolutional encoder, a multi-head attention layer, and a parameter regression output layer is constructed, and supervised pre-training is performed using historical window data and the corresponding standard variogram parameters as training samples. The feature vectors of each window are sequentially input into the pre-trained network. The curve shape features are extracted by a one-dimensional convolutional encoder, the three-directional variogram structure correlation is obtained by a multi-head attention layer, and the optimal variogram parameters of each window are predicted and output by a parameter regression output layer.

[0010] In this scheme, after obtaining the optimal mutation function parameters, the following steps are also included: An anisotropic search ellipsoid is constructed using a three-dimensional variable range. Neighborhood samples are searched for each block element within the ellipsoid and a set of ordinary Kriging equations is constructed. After solving the sample weights, the grade estimate and Kriging estimate variance of each block are calculated. Parallel execution of nearest neighbor interpolation based on inverse spatial distance weighting yields independent comparative estimates; The relative deviations of the two methods for each block are calculated and a spatial deviation sequence is formed. When the deviation sequence statistic exceeds a preset threshold, parameter prediction and kriging estimation are re-executed in the deviation threshold region. The iteration continues until the deviation sequence converges, and the converged kriging estimation result is used as the basis for resource reserves.

[0011] In this plan, the reconstruction of the tunnel model specifically includes: Laser point clouds of underground roadways are acquired, preprocessed by coordinate stitching, statistical filtering for noise reduction and voxel downsampling, and the point cloud coordinates are normalized. For each point in the normalized point cloud, search the neighborhood and aggregate local geometric features. Calculate the importance score of each point feature through a gating network, retain key points according to a preset ratio based on the score, and output a sparse scene feature tensor. The sparse feature tensor is input in parallel into the self-attention branch and the relative position encoding branch. The self-attention branch captures the global semantic dependencies between points, and the relative position encoding branch aggregates local geometric relationships. The outputs of the two branches are fused through channel concatenation to obtain the enhanced scene features. The enhanced scene features are compressed using NetVLAD soft-assignment residual aggregation and a fully connected layer to output a global feature descriptor; The global feature descriptor is used as input parameters to predict the cross-sectional morphology type, cross-sectional size, and centerline trajectory points of the regression head, and then a three-dimensional solid model of the tunnel is reconstructed through generalized sweeping.

[0012] In this scheme, after the tunnel model reconstruction is completed, geological-engineering feedback correction is also included: The reconstructed tunnel wall is spatially superimposed with the ore body model. The boundary of the adjacent ore body is traced along the normal direction of the wall. The wall points within the tolerance range are marked as ore body boundary control points and assigned high confidence weights. The updated control point set drives the elastic deformation model to make the ore body boundary converge towards the tunnel wall, thus completing the geological-engineering mutual feedback correction.

[0013] In this scheme, a local incremental update is triggered by the new data timestamp, specifically including: The newly added data is filtered by type and the data with the following characteristics are marked as valid updated data: the borehole mineralization points are located outside the existing ore body boundary, the point cloud of the roadway wall is located within the existing roadway space, and the geophysical and geochemical anomaly boundary has a spatial deviation from the existing boundary. Define spatial influence domains for the valid updated data points, retrieve model units falling within each spatial influence domain based on the three-dimensional spatial grid index, and after merging all influence domains and smoothly expanding their boundaries, determine the spatial update area for this incremental update.

[0014] In this scheme, the updated geological entities, block attributes, and tunnel models are overlaid and fused using a unified control network as anchor points to generate a comprehensive 3D visualization model, specifically including: At the boundary of the updated region, the constrained surface is consistent with the original model. The internal surface is updated by elastic deformation driven by the fusion of old and new control points and then spliced ​​with the surface outside the boundary. After determining the weight of blocks within the updated area and expanding the sample search range, the grade is recalculated, and the tunnel model is locally corrected. Using the underground control network as the anchor point, the updated three types of model spaces are superimposed. The roadway model is used to perform Boolean difference set on the ore body model to generate openings. After Boolean intersection of the block model and the ore body model, the volume of the external block is set to zero. After intersection of the block and the roadway, the volume of the block falling into the roadway is set to zero, and the partial cutting is calculated proportionally. After local fine-tuning and eliminating conflicts between models, the three types of models are integrated with a unified coordinate scene to generate a comprehensive 3D visualization model.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes a multi-source heterogeneous data standardization and fusion, along with a dynamic incremental update mechanism, to treat 3D geological models as a continuous production tool. An adaptive surface reconstruction technique based on prior constraints and dynamic boundary coupling allows the orebody model to evolve locally with new exploration data rather than being fully reconstructed, balancing the continuity of historical geological understanding with timely response to new data. A kriging estimation method based on sliding window cross-validation and dynamic inversion adaptively matches the variogram parameters to local grade fluctuations. This is further enhanced by a deep learning parameter prediction network replacing manual experience fitting and a parallel comparison and backtracking mechanism, significantly improving the accuracy, efficiency, and reliability of resource reserve estimation. A point cloud intelligent modeling technique combining LSA sparse coding and dual spatial feature enhancement significantly reduces computational overhead while maintaining both global layout and local details in tunnel model reconstruction. The reverse constraint of the orebody boundary on the reconstructed wall enables dynamic feedback between mining operations and the geological model. Incremental updates based on timestamps and multi-model spatial coupling integration ensure that model maintenance costs are proportional to the update scope rather than the entire mine scale, completely resolving the bottlenecks of update lag and computational redundancy in traditional methods. Ultimately, using the unified underground control network as an anchor point, the three types of models—geological entities, block attributes, and tunnel engineering—are integrated and combined into a unified three-dimensional visualization scene through spatial overlay and conflict elimination. This scene evolves synchronously with actual mine production, serving mining design, resource management, and precise exploration decision-making. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the accompanying drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the process of constructing a three-dimensional geological model and assessing resources for precision mineral exploration is shown. Figure 2 A flowchart for constructing the ore body model is shown; Figure 3 The flowchart for obtaining the optimal mutation function parameters for the window is shown; Figure 4 A flowchart of the reconstructed tunnel model is shown. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] like Figure 1 As shown in the figure, this embodiment provides a method for constructing a three-dimensional geological model and assessing resources for precise mineral exploration, including: Data from drilling, geophysical exploration, geochemical exploration, laser point cloud, and mining logging are acquired, and after unified coordinate normalization, gross errors are eliminated through adaptive filtering to construct a basic information database. Using existing geological interpretation lines as prior constraints, an initial implicit surface is constructed through radial basis function interpolation. Newly acquired borehole ore-bearing points and anomalous boundaries are used as dynamic control points. Local deformation of the surface is driven by elastic deformation to update the existing ore body model and reconstruct the model of the newly discovered ore body. Based on the ore body model, borehole grade data is extracted, and the experimental variogram is calculated in three dimensions. The space is divided by a sliding window, and the experimental variogram curve and statistical characteristics within the window are input into a pre-trained depth parameter prediction network to output the optimal variogram parameters for the window. After determining the optimal search ellipsoid based on the optimal variogram parameters of the window, ordinary Kriging interpolation is performed, and the nearest neighbor method is used in parallel for comparison and verification. If the deviation between the two exceeds the threshold, the variogram is backfitted and the converged Kriging estimation result is used as the basis for the reserves. Global feature descriptors are generated based on laser point clouds of underground roadways. The cross-sectional contour and centerline are predicted and the roadway model is reconstructed based on the global feature descriptors. The reconstructed wall is used as the real boundary to constrain the ore body boundary convergence and perform geological-engineering mutual feedback correction. Local incremental updates are triggered by new data timestamps, and the updated geological entities, block attributes, and tunnel models are superimposed and merged with a unified control network as anchor points to generate a comprehensive three-dimensional visualization model.

[0021] In this embodiment, the standardized collection of multi-source heterogeneous data and database construction are performed first.

[0022] Acquire five types of raw data from the target mine: drilling core logging data, geophysical survey data, geochemical analysis data, laser point cloud data, and mining logging data. For drilling core logging data, paper ledgers or unstructured electronic documents are converted into structured data tables, and borehole coordinates, inclinometer trajectories, geological stratification boundaries, and sample grade values ​​are extracted. For geophysical survey data, discrete measurement point data are converted into regular grid data through grid interpolation. For geochemical data, sampling point coordinates, indicator element content, and detection limit labels are organized, and maximum likelihood estimation is used to assign reasonable values ​​to truncated data. For laser point cloud data, the scan point sets under the local coordinate system of each station are stitched together to a unified coordinate system through target control points. For mining logging data, the mining engineering plan and profile are vectorized to extract the centerline coordinates, cross-sectional specifications, and tunneling direction information of the roadway.

[0023] Using the downhole high-precision measurement control network as the spatial reference, the original coordinates of each data source are uniformly normalized to the control network coordinate system. For data with known coordinate transformation parameters, a Bursa seven-parameter model is used for rigid spatial rectangular coordinate transformation. For data with unknown transformation parameters but corresponding control points, at least three uniformly distributed corresponding points are selected, and the similarity transformation parameters are solved by least squares adjustment before transformation. For data with historically accumulated non-uniform deformation, an error control grid is established, and thin-plate spline interpolation is used for local nonlinear fine correction. After coordinate normalization, the measured length downhole is used as the check criterion for length comparison verification. Multiple check sections are selected in important survey line directions, and the length calculated from the transformed coordinates is compared with the measured length downhole. When the deviation exceeds the limit, geometric correction is applied to ensure that the coordinate transformation accuracy meets the requirements of subsequent modeling.

[0024] An adaptive robust median absolute deviation filter is used to remove outliers. For numerical attribute fields in various types of data after coordinate normalization, the median of each attribute is calculated. and median absolute deviation The truncation range is adaptively determined based on the tail thickness of the data distribution, and the filtering decision criterion is as follows: ,in To achieve an adaptive truncation coefficient, a truncation coefficient of 2.5 to 3.0 is used for approximately normally distributed data. For skewed data, the truncation coefficient is adjusted based on the skewness coefficient. Data points whose attribute values ​​fall outside the truncation interval are identified as outliers and removed; gaps are filled with locally weighted adjacent valid data. A unique identifier code embedded with a spatial grid index is generated for each data item, embedding spatial grid location information to establish spatial proximity associations. Based on this, a stratigraphic sequence constraint network and a mine tunnel connectivity topology are constructed, forming a stratigraphic sequence constraint and mine tunnel spatial connectivity topology structure for the mining area. A credibility weight coefficient is added to each data item, with initial values ​​assigned according to the data source type, embedding the data acquisition timestamp and database entry timestamp, and establishing a data version management mechanism. Updated data at the same spatial location is stored as append versions without overwriting the original data, ultimately generating a basic information database with spatial topology awareness, dynamic weight evaluation, and time dimension traceability functions.

[0025] In this embodiment, the construction and updating of the ore body entity model adopts an adaptive strategy driven by the coupling of prior constraints and dynamic boundaries.

[0026] like Figure 2 As shown, existing geological profile interpretation lines are extracted and discretized into a three-dimensional spatial point sequence. Confidence levels are assigned based on the degree of agreement with borehole control points, with high-confidence segments assigned high-confidence weights and low-confidence segments or segments without borehole control points assigned medium-to-low-confidence weights. Multiple quadratic radial basis functions are used. For kernel function, This represents the distance between the input point and the center point. Let F(x) be the shape parameter. A three-dimensional implicit function is constructed using discrete points of the interpretation line as constraint points, and the geological interface is defined as the zero isosurface of the implicit function F(x) = 0. The three-dimensional implicit function F(x) is expressed as: , in For the first Spatial coordinates of the constraint points For radial basis functions, Let be the weight coefficients to be solved. It is a linear polynomial trend term.

[0027] Substituting the coordinates of all discrete constraint points on the prior interpretation line and their corresponding implicit function values, a system of linear equations is formed. The conjugate gradient method is then used to iteratively solve for all coefficients. During the solution process, anisotropic distance scaling factors are set along the strike, dip, and thickness directions of the ore body to ensure that the interpolation surface primarily conforms to the main extension direction of the ore body, avoiding distortions that violate the regional structural framework, thus completing the initial implicit surface construction. After the initial implicit surface construction is completed, newly acquired borehole ore-bearing points and geophysical and geochemical anomaly interpretation boundaries are used as dynamic control points. Confidence values ​​are assigned to each dynamic control point based on data accuracy and interpretation reliability. The dynamic control points are then fused with the prior constraints. When the spatial positions of the dynamic control points and prior constraints are consistent, the weighted average position is used as the updated target position. When the dynamic control points are located in areas with low confidence or no prior constraints, the position of the dynamic control points takes precedence. When there is a conflict, a compromise position is allocated based on the deviation between the predicted prior position and the dynamic control point position according to the confidence ratio.

[0028] The local deformation of the curved surface is realized by an elastic deformation model driven by tightly supported radial basis functions. A finite influence radius is set with each dynamic control point as the center. Constructing compactly supported radial basis functions : , in Let be the distance from any point in space to the dynamic control point.

[0029] Anisotropic propagation radii are defined along the strike, dip, and thickness directions of the ore body, respectively, and the weighted distances of any point in space relative to each dynamic control point are calculated. : , in For the radius of travel, For the radius of inclination, Where is the thickness radius.

[0030] Substituting the weighted distance into the radial basis function of the compact support, we obtain the deformation influence coefficient of each control point on the selected point. By summing the influence coefficients of all control points, we obtain the total deformation displacement at the selected point. With the goal of minimizing deformation energy, the deformation energy objective function is: ,in For data fitting, this penalizes the deviation of the updated surface from the target position at the dynamic control point. For deformation smoothing term, The equilibrium coefficient is used. Under the premise of satisfying the dynamic control point constraints, the total deformation displacement of each point is used as the adjustment amount of the implicit function value to perform point-by-point displacement correction on the initial implicit surface, so as to realize the local adaptive correction of the existing ore body model. For newly discovered ore bodies without a prior model, a multi-constraint section lofting strategy is adopted to construct the initial entity. The cross-sectional sequence is generated with the borehole ore-bearing point group and the anomaly center line as the backbone. The ore-bearing points falling within the search zone of each section are projected onto the section plane to generate the initial contour. The contour range is constrained by the geochemical anomaly contour line as the external geometric boundary. Corresponding points of adjacent section contours are matched and transition surfaces between sections are generated by cubic spline interpolation. The first and last sections are closed to form the three-dimensional entity model of the newly discovered ore body.

[0031] In this embodiment, the resource reserve estimation adopts the Kriging adaptive estimation method based on sliding window cross-validation dynamic inversion.

[0032] like Figure 3 As shown, the updated orebody model is discretized into regular block units, and the center coordinates, three-dimensional dimensions, and effective volume of each block are recorded. Borehole sample data falling within the spatial range of the orebody entity model are extracted from the basic information database, and samples of unequal lengths are standardized and combined. Lag distance step sizes and angle tolerances are set along the strike, dip, and thickness directions of the orebody, respectively. Sample pairs are searched within the tolerance range of each direction, and the average variogram value is statistically calculated by grouping by lag distance. The formula for calculating the experimental variogram is: , in The spatial distance falls within the lag distance vector The number of sample pairs within the nearby tolerance range. In spatial location Sample grade value at the location. In spatial location The sample grade value at that location.

[0033] The ore body space is uniformly divided into overlapping three-dimensional sliding windows, with adjacent windows maintaining a preset overlap rate. The discrete value sequence of the experimental variogram curve, sample statistical features (mean, variance, skewness, kurtosis), and spatial location code are extracted from each window as input feature vectors. A deep parameter prediction network consisting of a one-dimensional convolutional encoder, a multi-head attention layer, and a parameter regression output layer is constructed. Supervised pre-training is performed using historical window data and corresponding standard variogram parameters as training samples. The input feature vectors of each window are sequentially input into the pre-trained network. The one-dimensional convolutional encoder extracts curve morphology features, the multi-head attention layer obtains the three-directional variogram structure correlation, and the parameter regression output layer predicts and outputs the optimal variogram parameters for each window. The theoretical variogram model adopts a spherical model, expressed as: , in Value of a nugget. For sill values, For variable range.

[0034] An anisotropic search ellipsoid is constructed using a three-dimensional variable range. Neighborhood samples are searched within the ellipsoid for each block element, and a system of ordinary Kriging equations is constructed. After solving for the sample weights, the grade estimates and Kriging estimate variances for each block are calculated. The block grade estimates are: ,in The spatial coordinates of the center point of the block to be estimated are: This represents the total number of neighborhood samples that fall within the search ellipsoid and participate in the estimation. For the first The Kriging weight coefficients of each neighboring sample are calculated. Parallel nearest-neighbor interpolation based on inverse spatial distance weighting is performed to obtain independent contrastive estimates, denoted as... ,in The total number of neighboring samples participating in the nearest neighbor interpolation. For the first From the neighborhood sample to the center point of the block to be estimated Spatial distance. Calculate the relative deviation between the two methods for each block and form a spatial deviation sequence. When the deviation sequence statistic exceeds a preset threshold, re-execute parameter prediction and kriging estimation in the deviation-over-threshold region. Iterate until the deviation sequence converges. After convergence, use the kriging estimation result of the final round as the grade value of each block. Multiply the grade value of each block by the effective volume of the block and the weight of the ore, and sum them to obtain the resource reserves of the entire ore body.

[0035] In this embodiment, a three-dimensional model of the underground tunnel is constructed using intelligent modeling based on deep point cloud feature extraction.

[0036] like Figure 4 As shown, laser point clouds of the underground roadway were acquired and transformed to the coordinate system of the unified underground measurement and control network through splicing of target control points from multiple stations. Statistical filtering was used to remove outlier noise points, and density equalization was performed through voxel grid downsampling. The denoised point cloud was then subjected to coordinate normalization and decentering. For each point in the normalized point cloud, a neighborhood search was conducted and local geometric features were aggregated. Within a preset radius Searching for the set of neighboring points relative coordinates of neighboring points Point attributes are concatenated to form local features, which are then aggregated using a multilayer perceptron and max pooling to form local aggregated features. Each point feature vector is passed through a gating network to calculate an information importance score. Key points are retained according to a preset ratio based on the score, and a sparse scene feature tensor is output. This sparse feature tensor is input in parallel into a self-attention branch and a relative position encoding branch. The self-attention branch captures global semantic dependencies between points, while the relative position encoding branch calculates relative coordinate offsets and distance mappings between points to form position codes, aggregating local geometric relationships. The outputs of the two branches are cascaded and fused to obtain enhanced scene features. These enhanced scene features are then aggregated using NetVLAD soft-assignment residuals, as shown below: , in After NetVLAD aggregation, the first The first cluster center, the first The cumulative value of each feature dimension, all After concatenation, the data is compressed through a fully connected layer to form a global feature descriptor. , For the first The, the Cluster centers, For temperature parameters, For the number of key points, The number of cluster centers. For the first The feature vectors of key points after dual-space enhancement.

[0037] The residual cumulative vectors of each cluster center are concatenated, L2 regularized, and compressed using a fully connected layer to output a global feature descriptor. This global feature descriptor is then input into a multilayer perceptron parameter regression head to predict the probability of cross-sectional morphology type, key cross-sectional dimension parameters, and coordinates of centerline trajectory points. Based on the cross-sectional morphology and dimensions, a standard cross-sectional contour point sequence is generated in the local coordinate system. A generalized sweep algorithm is used to generate a triangular mesh for the tunnel wall along the centerline trajectory. After closing the beginning and end faces, the 3D solid model of the tunnel is reconstructed. After the tunnel model reconstruction is completed, the reconstructed tunnel wall is spatially superimposed with the ore body model. The boundaries of adjacent ore bodies are traced along the wall normal direction. Wall points within the tolerance range are marked as ore body boundary control points and assigned high confidence weights. The updated control point set, obtained by merging the original control points and the tunnel wall control points, is re-input into the elastic deformation model to drive the ore body boundary to converge towards the tunnel wall, completing the reverse constraint correction of the ore body model by the mining engineering model.

[0038] In another preferred embodiment, a dynamic refinement step of the orebody boundary driven by mining engineering data is also included. This involves acquiring on-site measured data of the orebody roof and floor boundaries during the mining process, and injecting the three-dimensional coordinates of the measured boundary lines as high-confidence control points of the orebody boundary into the elastic deformation model. Ore grade data for each ore extraction route or mining area is also acquired. When the deviation between the ore grade and the grade estimate of the corresponding block in the block model exceeds a preset threshold, the ore grade is used as the true grade constraint for that spatial location. The grade value of the block at that location is then corrected in reverse, and the corrected block grade is used as a guide point to drive a local recalculation of the grade estimate of surrounding blocks. The borehole halo sampling results are spatially located and sample lengths are returned to their positions according to the borehole trajectory. After merging the densified samples of the ore-bearing section according to the stratigraphic level, a sequence of the elemental content of the halo in each borehole is formed. The borehole trajectory is used as the spatial skeleton to perform three-dimensional spatial interpolation on the original halo data of each borehole, and the geochemical anomaly contour lines are redrawn. The updated anomaly contour lines serve as dynamic boundary constraints for subsequent model updates.

[0039] In this embodiment, a timestamp-triggered incremental update mechanism is used for dynamic model maintenance. The database embeds an acquisition timestamp and an entry timestamp for each data record and maintains a global version number. The database update status is monitored, and the timestamps of new data records are scanned. When a new record with a timestamp later than the latest timestamp of the data used by the current model is found, and the newly added data is spatially concentrated, the incremental update mode is triggered. New data is filtered by type based on relevance, and only data where borehole encounters are located outside the existing orebody boundary, roadway wall point clouds are within the existing roadway space, and geophysical and geochemical anomaly boundaries deviate spatially from existing boundaries are marked as valid update data. Spatial influence domains are defined for the valid update data points. Model units falling within each spatial influence domain are retrieved based on a 3D spatial grid index. After merging all influence domains and smoothly expanding their boundaries, the spatial update area for this incremental update is determined.

[0040] At the boundary of the updated region, the surface position is forced to be consistent with the original model. Within the region, the old and new control points are merged, and the local surface is updated through elastic deformation and then stitched with the surface outside the boundary. For the blocks within the updated region, the ore body / outside determination and effective volume calculation are performed. The search range of neighboring samples is expanded beyond the boundary of the updated region before re-performing kriging estimation. For the tunnel model, the cross-sectional contour is locally corrected in the corresponding segment of the new point cloud data, or topological extension and reconstruction are performed at the end.

[0041] Using the underground high-precision measurement and control network as a unified spatial anchor point, the updated orebody model, block attribute model, and tunnel engineering model are spatially superimposed and coupled for integration. Boolean difference operations are performed on the orebody model using the tunnel model, and Boolean intersection operations are performed on the block model and orebody model, setting the effective volume of blocks falling outside the orebody to zero. After spatial intersection with the tunnel, the volume of blocks falling into the tunnel is set to zero, and partial cuts are proportionally calculated. Conflicts such as mesh penetration and gaps between models are eliminated through local fine-tuning using the minimum energy deformation method. After spatial coupling and conflict elimination, the three types of models are integrated using a unified coordinate scene to generate a comprehensive 3D visualization scene. The system returns to data monitoring status, continuously waiting for a new round of timestamps to trigger a new round of incremental updates.

[0042] After the mining is completed, the actual mined ore body boundary and actual ore grade data are obtained. The spatial deviation between the actual mined ore body boundary and the ore body model boundary at the same spatial location is statistically analyzed point by point. The spatial distribution of the deviation is projected onto the strike, dip, and thickness directions of the ore body. The systematic deviation amount and direction of each direction are statistically analyzed. When the deviation in a certain direction shows a consistent trend, it is determined that there is a systematic setting deviation in the prior constraint weight or elastic deformation parameter, and parameter correction suggestions are generated. The actual ore grade and the block grade estimate at the same spatial location are statistically analyzed block by block. When the deviation shows a clustered distribution in space, it is determined that there is a systematic deviation in the sliding window variogram parameter corresponding to the cluster area, and window parameter correction suggestions are generated.

[0043] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration, characterized in that, Includes the following steps: Data from drilling, geophysical exploration, geochemical exploration, laser point cloud, and mining logging are acquired, and after unified coordinate normalization, gross errors are eliminated through adaptive filtering to construct a basic information database. Using existing geological interpretation lines as prior constraints, an initial implicit surface is constructed through radial basis function interpolation. Newly acquired borehole ore-bearing points and anomalous boundaries are used as dynamic control points. Local deformation of the surface is driven by elastic deformation to update the existing ore body model and reconstruct the model of the newly discovered ore body. Based on the ore body model, borehole grade data is extracted, and the experimental variogram is calculated in three dimensions. The space is divided by a sliding window, and the experimental variogram curve and statistical characteristics within the window are input into a pre-trained depth parameter prediction network to output the optimal variogram parameters for the window. After determining the optimal search ellipsoid based on the optimal variogram parameters of the window, ordinary Kriging interpolation is performed, and the nearest neighbor method is used in parallel for comparison and verification. If the deviation between the two exceeds the threshold, the variogram is backfitted and the converged Kriging estimation result is used as the basis for the reserves. Global feature descriptors are generated based on laser point clouds of underground roadways. The cross-sectional contour and centerline are predicted and the roadway model is reconstructed based on the global feature descriptors. The reconstructed wall is used as the real boundary to constrain the ore body boundary convergence and perform geological-engineering mutual feedback correction. Local incremental updates are triggered by new data timestamps, and the updated geological entities, block attributes, and tunnel models are superimposed and merged with a unified control network as anchor points to generate a comprehensive three-dimensional visualization model.

2. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 1, characterized in that, Constructing a basic information database, specifically including: Acquire five types of data: drilling, geophysical exploration, geochemical exploration, laser point cloud, and mining logging. Transform unstructured data into structured data tables and establish original index ledgers according to type. Using the downhole high-precision control network as a benchmark, the coordinates of each data source are uniformly normalized to the control network coordinate system, and the length is compared and verified using the actual measured length downhole as the verification benchmark. When the deviation exceeds the limit, geometric correction is applied. An adaptive robust filter based on median absolute deviation is used to remove outliers. The median of the data sequence is used as the center and the median absolute deviation is used as the discrete measure. The cutoff range is adaptively determined according to the thickness of the tail of the data distribution. Outliers outside the effective interval are removed, and gaps are filled by local weighted addition of adjacent effective data. A unique identifier code is generated for each data item, embedding spatial grid location information in the code, establishing spatial proximity associations, and constructing a stratigraphic sequence constraint network and shaft connectivity topology to generate a basic information database.

3. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 1, characterized in that, Using existing geological interpretation lines as prior constraints, an initial implicit surface is constructed through radial basis function interpolation, specifically including: Existing geological profile interpretation lines are extracted and discretized into a three-dimensional spatial point sequence. The credibility is graded and weighted according to the degree of agreement with borehole control points. Using quadratic radial basis functions as kernel functions, three-dimensional implicit functions are constructed by taking the discrete points of the interpretation line as constraint points. The geological interface is defined as the zero isosurface with F(x)=0. Anisotropic distance scaling factors are set along the strike, dip and thickness directions of the ore body, and the weight coefficients are solved to construct the initial implicit surface.

4. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 1, characterized in that, Using newly acquired borehole ore-bearing points and anomalous boundaries as dynamic control points, the existing orebody model is updated and the newly discovered orebody model is reconstructed through elastic deformation-driven local surface deformation. Specifically, this includes: The newly acquired borehole mineralization points and geophysical and geochemical anomaly interpretation boundaries are used as dynamic control points and assigned confidence levels, and then fused with prior constraints with confidence levels. A tight-support radial basis function is constructed with each dynamic control point as the center and a finite influence radius is set. Anisotropic propagation radii are defined along the strike, dip and thickness directions of the ore body. The weighted distance of any point in space relative to each dynamic control point is calculated. The weighted distance is substituted into the tight-support radial basis function to obtain the deformation influence coefficient of each control point on the selected point. The influence coefficients of all control points are superimposed to obtain the total deformation displacement at the selected point. The total deformation displacement at each point is used as the adjustment amount of the implicit function value to perform point-by-point displacement correction on the initial implicit surface, thereby realizing local adaptive correction of the existing ore body model. For newly discovered ore bodies without a prior model, a cross-sectional sequence is generated with the ore-bearing point group and the anomaly center line as the backbone. The ore-bearing points falling within the search zone of each cross section are projected onto the cross-sectional plane to generate an initial profile. The contour range is constrained by the geochemical anomaly contour lines as the external geometric boundary. Corresponding points are matched for adjacent cross-sectional profiles and transition surfaces between cross sections are generated by cubic spline interpolation. The first and last cross sections are closed to form a three-dimensional solid model of the newly discovered ore body.

5. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 1, characterized in that, Obtain the optimal mutation function parameters for the window, specifically including: The updated orebody model is discretized into regular block units. Drill sample data falling within the orebody area are extracted and sample length is standardized. The three-dimensional experimental variogram curves are calculated along the strike, dip and thickness directions of the orebody. The ore body space is uniformly divided into overlapping three-dimensional sliding windows, and the discrete value sequence of experimental variogram curves, sample statistical features and spatial location codes in each window are extracted as input feature vectors. A deep parameter prediction network consisting of a one-dimensional convolutional encoder, a multi-head attention layer, and a parameter regression output layer is constructed, and supervised pre-training is performed using historical window data and the corresponding standard variogram parameters as training samples. The feature vectors of each window are sequentially input into the pre-trained network. The curve shape features are extracted by a one-dimensional convolutional encoder, the three-directional variogram structure correlation is obtained by a multi-head attention layer, and the optimal variogram parameters of each window are predicted and output by a parameter regression output layer.

6. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 5, characterized in that, After obtaining the optimal mutation function parameters, the following is also included: An anisotropic search ellipsoid is constructed using a three-dimensional variable range. Neighborhood samples are searched for each block element within the ellipsoid and a set of ordinary Kriging equations is constructed. After solving the sample weights, the grade estimate and Kriging estimate variance of each block are calculated. Parallel execution of nearest neighbor interpolation based on inverse spatial distance weighting yields independent comparative estimates; The relative deviations of the two methods for each block are calculated and a spatial deviation sequence is formed. When the deviation sequence statistic exceeds a preset threshold, parameter prediction and kriging estimation are re-executed in the deviation threshold region. The iteration continues until the deviation sequence converges, and the converged kriging estimation result is used as the basis for resource reserves.

7. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 1, characterized in that, The reconstruction of the tunnel model specifically includes: Laser point clouds of underground roadways are acquired, preprocessed by coordinate stitching, statistical filtering for noise reduction and voxel downsampling, and the point cloud coordinates are normalized. For each point in the normalized point cloud, search the neighborhood and aggregate local geometric features. Calculate the importance score of each point feature through a gating network, retain key points according to a preset ratio based on the score, and output a sparse scene feature tensor. The sparse feature tensor is input in parallel into the self-attention branch and the relative position encoding branch. The self-attention branch captures the global semantic dependencies between points, and the relative position encoding branch aggregates local geometric relationships. The outputs of the two branches are fused through channel concatenation to obtain the enhanced scene features. The enhanced scene features are compressed using NetVLAD soft-assignment residual aggregation and a fully connected layer to output a global feature descriptor; The global feature descriptor is used as input parameters to predict the cross-sectional morphology type, cross-sectional size, and centerline trajectory points of the regression head, and then a three-dimensional solid model of the tunnel is reconstructed through generalized sweeping.

8. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 7, characterized in that, After the tunnel model reconstruction is completed, geological-engineering feedback correction is also included: The reconstructed tunnel wall is spatially superimposed with the ore body model. The boundary of the adjacent ore body is traced along the normal direction of the wall. The wall points within the tolerance range are marked as ore body boundary control points and assigned high confidence weights. The updated control point set drives the elastic deformation model to make the ore body boundary converge towards the tunnel wall, thus completing the geological-engineering mutual feedback correction.

9. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 1, characterized in that, Triggering partial incremental updates with new data timestamps specifically includes: The newly added data is filtered by type and the data with the following characteristics are marked as valid updated data: the borehole mineralization points are located outside the existing ore body boundary, the point cloud of the roadway wall is located within the existing roadway space, and the geophysical and geochemical anomaly boundary has a spatial deviation from the existing boundary. Define spatial influence domains for the valid updated data points, retrieve model units falling within each spatial influence domain based on the three-dimensional spatial grid index, and after merging all influence domains and smoothly expanding their boundaries, determine the spatial update area for this incremental update.

10. The method for constructing a three-dimensional geological model and assessing resources for precision mineral exploration according to claim 9, characterized in that, The updated geological entities, block attributes, and tunnel models are overlaid and merged using a unified control network as anchor points to generate a comprehensive 3D visualization model, specifically including: At the boundary of the updated region, the constrained surface is consistent with the original model. The internal surface is updated by elastic deformation driven by the fusion of old and new control points and then spliced ​​with the surface outside the boundary. After determining the weight of blocks within the updated area and expanding the sample search range, the grade is recalculated, and the tunnel model is locally corrected. Using the underground control network as the anchor point, the updated three types of model spaces are superimposed. The roadway model is used to perform Boolean difference set on the ore body model to generate openings. After Boolean intersection of the block model and the ore body model, the volume of the external block is set to zero. After intersection of the block and the roadway, the volume of the block falling into the roadway is set to zero, and the partial cutting is calculated proportionally. After local fine-tuning and eliminating conflicts between models, the three types of models are integrated with a unified coordinate scene to generate a comprehensive 3D visualization model.

Images