A method for prospecting granite type tin deposit based on multi-source spatial scale

Abstract

The present application belongs to the technical field of geological mineral exploration. A granite tin ore prospecting method based on multi-source spatial scale is provided, comprising: based on the detection data obtained by using the gravity and magnetic method and the electromagnetic method in the target area, a first three-dimensional prospecting model and a second three-dimensional prospecting model are respectively constructed by using an inversion algorithm, and a plurality of difference areas are obtained by comparison; based on auxiliary detection data and joint inversion algorithm, each difference area is modified to obtain a difference area model; the difference area model is fused with the first three-dimensional prospecting model or the second three-dimensional prospecting model to obtain a final three-dimensional prospecting model of the granite tin ore. Through the collaborative inversion and model fusion of multi-source spatial scale data, the machine learning joint inversion correction is introduced for the gravity-magnetic-electromagnetic model difference area, the multi-solution of a single method is effectively eliminated, the three-dimensional prospecting model formed provides a precise target area for deep tin ore exploration, and the prospecting efficiency and accuracy can be significantly improved.

Description

Technical Field

[0001] This invention relates to the field of geological and mineral exploration technology, and more specifically, to a method for prospecting granite-type tin deposits based on multi-source spatial scales. Background Technology

[0002] With the continuous increase in demand for tin resources from modern industry, surface and shallow tin resources are gradually being depleted, making the exploration of deep granite-type tin deposits a key direction for ensuring resource supply. However, deep tin deposits are characterized by their great depth and strong concealment of ore bodies, posing many challenges to conventional prospecting techniques.

[0003] Currently, gravity and magnetic methods and electromagnetic methods are commonly used geophysical techniques for deep tin ore exploration. Gravity and magnetic methods, by detecting differences in rock density and magnetic properties, can infer the distribution and structural morphology of deep granite bodies; electromagnetic methods utilize differences in the electrical properties of underground media to identify potential mineralization anomalies. However, each method has significant limitations: gravity and magnetic methods do not respond significantly to weakly magnetic or low-density ore bodies at depth, and the small differences in the physical properties of geological bodies easily lead to multiple interpretations of anomalies; electromagnetic methods are affected by complex underground electrical structures, resulting in uncertainties in the interpretation of low-resistivity bodies and making it difficult to accurately distinguish between mineralized and non-mineralized geological bodies.

[0004] To improve exploration accuracy, some studies have attempted to construct three-dimensional mineral exploration models. However, existing technologies mostly rely on data from a single method to build models, and models constructed using different methods lack effective comparison and integration. Even when both gravity and magnetic methods are used to construct models, the differences in their detection principles and data characteristics result in significant discrepancies in the generated three-dimensional models, making it difficult to determine the reliability of the model results. This leads to large prediction errors for deep ore bodies, high levels of uncertainty in drilling deployment, and a dramatic increase in exploration costs and inefficiency.

[0005] Therefore, there is an urgent need for a new method that can integrate multi-source spatial scale data, eliminate differences in models of different detection methods, and improve the accuracy of deep granite-type tin deposit exploration, so as to improve the efficiency and accuracy of deep tin deposit exploration and reduce exploration costs. Summary of the Invention

[0006] To address the technical problems existing in the background art, the present invention provides a granite-type tin ore prospecting method based on multi-source spatial scale, electronic equipment, computer storage medium, and computer program product.

[0007] This invention provides a method for prospecting granite-type tin deposits based on multi-source spatial scales, comprising the following steps:

[0008] Based on the detection data obtained by using gravity and magnetic methods and electromagnetic methods in the target area, an inversion algorithm was used to construct a first three-dimensional mineral exploration model and a second three-dimensional mineral exploration model, and a comparison was made between the first three-dimensional mineral exploration model and the second three-dimensional mineral exploration model to identify several areas of difference.

[0009] Auxiliary detection data corresponding to another detection method different from gravity and magnetic methods and electromagnetic methods is obtained. Based on the auxiliary detection data and the joint inversion algorithm driven by machine learning, the difference regions are corrected to obtain the difference region model.

[0010] The models of the different regions are fused with the first three-dimensional prospecting model or the second three-dimensional prospecting model to obtain the final three-dimensional prospecting model of granite-type tin deposits.

[0011] The present invention also provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any of the preceding claims.

[0012] The present invention also provides a computer storage medium storing a computer program that can be executed by a processor to implement the method as described in any of the preceding claims.

[0013] The present invention also provides a computer program product comprising a computer program that can be executed by a processor to implement the method as described in any of the preceding claims.

[0014] The beneficial effects of this invention are as follows: By collaborative inversion and model fusion of multi-source spatial scale data, and by introducing machine learning joint inversion corrections for discrepancies in gravity-magnetic-electromagnetic models, this invention effectively eliminates the ambiguity of single methods. Simultaneously, by utilizing the fusion algorithm of generative adversarial networks and U-Net, it achieves full-chain data integration from regional tectonics to orebody scale. The resulting three-dimensional prospecting model provides precise target areas for deep tin ore exploration, significantly improving prospecting efficiency and accuracy. Attached Figure Description

[0015] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0016] Figure 1 This is a schematic flowchart of a granite-type tin ore prospecting method based on multi-source spatial scale disclosed in an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention. Detailed Implementation

[0018] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0019] like Figure 1 As shown in the figure, this invention discloses a method for prospecting granite-type tin deposits based on multi-source spatial scales, comprising the following steps:

[0020] S10, based on the detection data obtained by using gravity and magnetic methods and electromagnetic methods on the target area, the first three-dimensional prospecting model and the second three-dimensional prospecting model are constructed by using the inversion algorithm, and several difference areas in the first three-dimensional prospecting model and the second three-dimensional prospecting model are compared and identified.

[0021] This step involves constructing three-dimensional mineral exploration models using detection data obtained from gravity and magnetic methods, and extracting areas of difference between the models. The specific implementation is as follows:

[0022] Density and magnetic data of the target area were acquired using high-precision gravity measurements (e.g., 1:50,000 scale) and airborne magnetic surveys. Gravity measurements collected Bouguer gravity anomaly values, and magnetic surveys acquired total field magnetic anomaly data. The data were processed using a regularized inversion algorithm to construct a first three-dimensional prospecting model reflecting the boundaries and structural morphology of the deep granite body. In this model, low-density anomaly zones correspond to the distribution of the granite body, and magnetic anomaly gradient zones indicate the locations of fault structures.

[0023] Electrical data of the subsurface medium were collected using methods such as controlled-source audio-frequency magnetotellurics (CSAMT) at depths of 1000-2000 meters and sampling points spaced 50 meters × 100 meters. A second three-dimensional prospecting model was constructed using a finite element inversion algorithm, in which low-resistivity anomalies were preliminarily inferred to be ore-bearing hydrothermal channels or tin ore bodies.

[0024] The constructed first and second three-dimensional prospecting models were imported into three-dimensional geological modeling software (such as Petrel) and compared by overlay based on spatial coordinates. By calculating the root mean square error of the depth of the top interface of the rock mass and the spatial location of the low resistivity anomaly in the two three-dimensional prospecting models, several areas with significant differences can be identified. For example: (1) The first three-dimensional prospecting model (gravity and magnetic model) shows that the granite body is buried at a depth of 800 meters, and there is no low resistivity anomaly at the corresponding location in the electromagnetic model; (2) The low resistivity body at a depth of 1200 meters in the second three-dimensional prospecting model (electromagnetic model) does not show density anomaly in the gravity and magnetic model; (3) There is a 15° deviation between the two models in the extension direction of the NE-trending fault zone.

[0025] S20: Obtain auxiliary detection data corresponding to another detection method different from the gravity and magnetic methods and the electromagnetic methods. Based on the auxiliary detection data and the joint inversion algorithm driven by machine learning, correct each of the difference regions to obtain the difference region model.

[0026] This step employs a machine learning-driven joint inversion algorithm, which integrates a different detection method than gravity and electromagnetic methods to correct the difference region, thereby obtaining a difference region model that is closer to the actual situation.

[0027] Another detection method is the combination of hyperspectral remote sensing and borehole geochemical measurement. The joint inversion algorithm driven by machine learning is, for example, a joint inversion model based on generative adversarial networks (GANs), which will be explained in detail later and will not be repeated here.

[0028] S30, the models of the different regions are fused with the first three-dimensional prospecting model or the second three-dimensional prospecting model to obtain the final three-dimensional prospecting model of granite-type tin deposits.

[0029] This step integrates the corrected difference region model with the original 3D mineral exploration model using spatial coordinate registration and machine learning fusion algorithms to generate the final 3D mineral exploration model. It can be understood that the original 3D mineral exploration model can be either a first 3D mineral exploration model or a second 3D mineral exploration model. The specific implementation is as follows:

[0030] First, a Geographic Information System (GIS) platform is used to perform a unified coordinate transformation on the data. The specific process is as follows:

[0031] The first 3D mineral exploration model, the second 3D mineral exploration model, and the difference region model generated in step S20 were all converted to the WGS84 coordinate system, with the local mean sea level as the vertical reference surface. Kriging interpolation was performed on the mineralization probability field (resolution 10m×10m×5m) of the difference region model and the attribute field (density, resistivity) of the original model to ensure consistent mesh size (uniformly 20m×20m×10m).

[0032] The pre-built U-Net semantic segmentation network is invoked, with the input being the mineralization probability map of the differential region, density outliers from the gravity and magnetic model, and resistivity values ​​from the electromagnetic model. The output is a fused comprehensive attribute label (divided into three categories: "ore body," "rock body," and "structure"). It is understandable that the U-Net semantic segmentation network can be trained using the cross-entropy loss function, with lithological data from known boreholes (several training samples and several validation samples) as supervision signals, iteratively training until the model accuracy reaches the target.

[0033] The specific process of model fusion is as follows:

[0034] The "potential orebody area" (mineralization probability > 0.6) modified by S20 is treated as an independent attribute body and embedded into the original 3D model through Boolean operations. For example, in the 1200-meter depth difference area, the mineralization probability body is spatially intersected with the granite body boundary of the gravity and magnetic model and the low resistivity body of the electromagnetic model, and "orebody = low resistivity body ∩ mineralization probability body ∩ granite contact zone" is defined.

[0035] Gaussian filtering (standard deviation 1.5) is used to smooth the fused density-resistivity-mineralization probability field, eliminating discontinuities at the model boundary and generating a continuous three-dimensional attribute distribution.

[0036] The following comprehensive mineral exploration indicators are used to delineate target areas: mineralization probability > 0.7; density anomaly Δρ ∈ [-0.2, -0.1] g / cm³. 3 (Characteristics of granite bodies); resistivity lpgρ∈[1.5,2.0]Ω·m (characteristics of mineralized hydrothermal channels). Areas meeting the above conditions are automatically marked as Class I prospecting target areas, and areas with a probability of 0.5-0.7 and satisfying a single physical property anomaly are marked as Class II target areas.

[0037] This invention utilizes collaborative inversion and model fusion of multi-source spatial scale data. For discrepancies between gravity / magnetic and electromagnetic models, it introduces machine learning-based joint inversion correction, effectively eliminating the ambiguity of single-method approaches. Simultaneously, by employing a fusion algorithm combining generative adversarial networks and U-Net, it achieves end-to-end data integration from regional tectonics to orebody scale. The resulting three-dimensional prospecting model provides precise target areas for deep tin ore exploration, significantly improving prospecting efficiency and accuracy.

[0038] In some embodiments, the comparison to identify several difference regions between the first three-dimensional mineral exploration model and the second three-dimensional mineral exploration model includes:

[0039] The spatial data of the first three-dimensional mineral exploration model and the second three-dimensional mineral exploration model are registered, and the difference in physical property parameters at the corresponding locations is calculated by the density-resistivity joint difference index. Areas exceeding the preset threshold are selected as first-level difference areas.

[0040] Morphological analysis was performed on the primary differential regions to extract continuously distributed significant differential bodies. Based on the contradiction type of gravity-magnetic-electromagnetic response combined with geological prior knowledge, several differential regions were finally identified.

[0041] In this embodiment, the density anomaly field in the first three-dimensional mineral exploration model and the resistivity anomaly field in the second three-dimensional mineral exploration model are spatially registered to establish a unified coordinate system. Then, the difference in physical property parameters at corresponding spatial locations in the two models is calculated.

[0042] For each grid cell (i,j,k), calculate the density-resistivity joint difference index DI:

[0043] DI(i,j,k)=|Δρ(i,j,k) / ρ max |×|Δlogρ(i,j,k) / logρ max |

[0044] Where Δρ is the density anomaly, ρ max The absolute value of the maximum density anomaly; Δlogρ is the logarithmic anomaly of resistivity, logρ max This represents the absolute value of the logarithmic anomaly of the maximum resistivity.

[0045] For example, setting a difference threshold DI threshold =0.25, so DI>DI threshold The region is marked as a first-level difference region.

[0046] Morphological analysis was performed on each primary region of difference: continuous volumes greater than 1000 m³ were extracted. 3 Connected regions; isolated differences scattered at the edges are removed.

[0047] Classify the regions of difference by combining prior geological knowledge, for example:

[0048] Type A: The gravity and magnetic model shows high density, but the electromagnetic model has no corresponding low-resistance anomaly;

[0049] Type B: The electromagnetic model shows low resistance, but the gravity and magnetic model shows no corresponding density anomaly;

[0050] Type C: The two models have a boundary positioning deviation of more than 50 meters for the same geological body;

[0051] Type D: The two models deviate more than 10° in judging the direction of the fracture structure.

[0052] The first-level anomaly regions that meet the above criteria are identified as the final anomaly regions.

[0053] In some embodiments, the correction of each of the difference regions based on the auxiliary detection data and the machine learning-driven joint inversion algorithm to obtain a difference region model includes:

[0054] Acquire auxiliary detection data, including the spectral characteristics of altered minerals obtained by hyperspectral remote sensing and the distribution characteristics of metal elements obtained by borehole geochemical measurements, and extract auxiliary feature vectors from the auxiliary detection data;

[0055] The difference region between the first three-dimensional mineral exploration model and the second three-dimensional mineral exploration model, along with the auxiliary feature vector, is input into a joint inversion model constructed based on a generative adversarial network. The generator of the joint inversion model generates a corrected geological attribute field, and the discriminator is used to evaluate the authenticity and mineralization correlation of the corrected geological attribute field.

[0056] When the mineralization probability value output by the generator exceeds a set threshold, the corresponding difference region is marked as a potential ore body region, and a corrected difference region model is obtained.

[0057] This embodiment integrates hyperspectral remote sensing and borehole geochemical data, and utilizes a generative adversarial network (GAN) to achieve intelligent correction of discrepancies in regions. The specific implementation is as follows:

[0058] First, auxiliary exploration data were obtained using hyperspectral remote sensing and borehole geochemical measurements:

[0059] Using a drone equipped with an AVIRIS-NG sensor, data in the 400-2500nm band with a resolution of 3.5nm was acquired. Surface alteration mineral assemblages (such as muscovite and tourmaline) were extracted using a spectral unmixing algorithm to delineate the range of alteration halos associated with deep tin deposits. Simultaneously, three verification boreholes (500-800 meters deep) were deployed around the differential areas to collect core samples for LA-ICP-MS analysis, obtaining the vertical distribution characteristics of elements such as Sn, W, and Be, and establishing a zoning model of deep mineralization elements.

[0060] Feature extraction was performed on the above-mentioned auxiliary detection data to obtain auxiliary feature vectors: for remote sensing data, spectral feature vectors (dimension 128) of altered minerals were extracted using a convolutional neural network (such as ResNet-50) and converted into spatial distribution probability maps; and for geochemical data, principal component analysis (PCA) was used to reduce dimensionality and extract the first three principal components (cumulative variance contribution rate 85%) to characterize the degree of element enrichment and mineralization correlation.

[0061] Simultaneously, a joint inversion model based on a generative adversarial network (GAN) is pre-constructed, with the following network structure:

[0062] Generator (G): The input is the gravity-magnetic-electromagnetic difference region data (density anomaly value, resistivity value) and auxiliary feature vectors of auxiliary detection data. It generates the corrected geological body attribute field (density distribution field, resistivity distribution field, mineralization probability field) through a multi-layer Transformer structure (8-layer encoder-decoder).

[0063] Discriminator (D): A dual-task design that simultaneously judges the authenticity of the generated data (compared with known borehole data) and mineralization relevance (matching degree with geochemical characteristics), and outputs a comprehensive score.

[0064] The inversion process is roughly as follows:

[0065] Data preprocessing: Density anomalies Δρ (range: -0.2~0.1 g / cm³) in regions of gravity and magnetic disparity were analyzed using the gravity and magnetic model. 3The resistivity logρ (range: 1.5~3.0Ω·m) of the electromagnetic model difference region and the auxiliary feature vectors (remote sensing alteration probability, geochemical principal components) are normalized to [-1,1].

[0066] Adversarial training: The generator G receives random noise (100 dimensions) and auxiliary feature vectors to generate a simulated geological property field; the discriminator D compares the generated data with real borehole data and calculates the Wasserstein distance as the loss function; the parameters of G and D are iteratively updated through the Adam optimizer (learning rate 1e-4, batch size 32) until the discriminator accuracy stabilizes above 75%. It is understandable that this adversarial training is not the usual pre-training of generative adversarial networks, but rather incremental fine-tuning of the pre-trained model based on differential region data and auxiliary probe data of the current area.

[0067] Difference Correction: When training converges, the geological attribute field output by generator G is the corrected difference region model, where regions with mineralization probability values ​​> 0.6 are marked as "potential ore bodies".

[0068] Taking a tin mine with a depth difference of 1200 meters as an example:

[0069] The original model is contradictory: the gravity and magnetic model shows that this area is the edge of a granite body (Δρ=-0.15g / cm). 3 The electromagnetic model exhibits a low-resistivity anomaly (logρ = 1.8 Ω·m), but traditional inversion methods cannot determine the cause of the anomaly (ore body or water-bearing structure).

[0070] Auxiliary data input: Hyperspectral remote sensing identified a tourmaline-mica alteration assemblage at this location (alteration probability 0.72), borehole geochemistry showed a Sn element content of 120 ppm (background value <20 ppm), and principal component analysis showed a positive correlation with W and Be elements (correlation coefficient >0.8).

[0071] The mineralization probability map output by the generator of the joint inversion model showed a mineralization probability of 0.81 for this area. After correction, the model classified it as a "hidden tin ore body" and predicted a thickness of 25±5 meters and a Sn grade of 0.9-1.1%. Subsequent drilling revealed a 23-meter-thick ore body at a depth of 1220-1245 meters with a Sn grade of 1.05%, which is consistent with the prediction of the joint inversion model, confirming the joint inversion model's ability to identify complex anomalies.

[0072] In some embodiments, the discriminator includes an authenticity branch and a mineralization correlation branch. The authenticity branch is used to determine the degree of matching between the generated geological attribute field and the measured borehole data, and the mineralization correlation branch is used to determine the coupling relationship between the corrected geological attribute field and the mineralization indication information in the auxiliary features.

[0073] In some embodiments, the discriminator is used to evaluate the authenticity of the corrected geological attribute field and its correlation with mineralization, including:

[0074] The branch structure parameters of the discriminator are dynamically adjusted based on the complexity of the physical property characteristics of the difference region. The complexity of the physical property characteristics is calculated by the spatial gradient of the density-resistivity joint difference index.

[0075] The authenticity branch receives the corrected geological attribute field and the measured borehole data, calculates the spatial difference between the two in multiple attribute dimensions, evaluates the degree of matching between the corrected geological attribute field in terms of numerical value and spatial structure, and obtains the authenticity score.

[0076] The mineralization correlation branch receives the corrected geological attribute field and the auxiliary feature vector, analyzes the coupling relationship between the mineralization probability distribution and hyperspectral alteration characteristics and geochemical element enrichment patterns, and obtains the correlation score.

[0077] The authenticity score and the relevance score are weighted to obtain a comprehensive discrimination score, which is used as the discrimination signal for generator training.

[0078] This embodiment achieves accurate evaluation of the corrected geological attribute field by dynamically adjusting the discriminator structure and integrating multi-dimensional scores. The specific implementation is as follows:

[0079] First, the spatial gradient of the density-resistivity joint difference index (DI) is used to characterize the geological scene complexity, i.e., the physical property complexity, of the difference region. For each difference region, the three-dimensional gradient magnitude of DI is calculated (…). The formula is:

[0080]

[0081] For example, when When a scene is identified as highly complex (e.g., at the intersection of fault zones), the number of network layers in the discriminator's realism branch is automatically increased (e.g., from 4 to 6 layers) to capture complex physical property interface features; when In cases such as homogeneous granite bodies, maintain the basic network structure (e.g., 4 layers) to improve computational efficiency.

[0082] This invention achieves "geological semantic perception" of discriminator branch structure parameters by quantifying the complexity of physical property features through DI gradient quantization: it strengthens local feature extraction and classification decision-making in complex areas (such as mineralized zones) and maintains model stability in simple areas (such as homogeneous surrounding rocks), ultimately resolving the contradiction of "underfitting in complex areas and overfitting in simple areas" in traditional fixed parameter models, and improving the reliability of deep geological anomaly identification.

[0083] Then, the adjusted discriminator is used to evaluate the authenticity and mineralization correlation of the generator's generated results. Specifically:

[0084] The corrected geological attribute fields (density distribution field, resistivity distribution field, mineralization probability field) and the measured borehole data (core physical property test results) were uniformly interpolated to a 20m×20m×10m grid, and the attribute mean values ​​of the 3×3×3 grid around the borehole location were extracted as a comparison sample.

[0085] (1) Multi-attribute matching calculation of authenticity score

[0086] Numerical matching degree: Calculates the root mean square error (RMSE) of density anomalies and the logarithm of resistivity, with a weighting of 60%. For example, the measured density anomaly at a borehole is Δρ = -0.12 g / cm³. 3 The predicted value in the generated results is -0.10 g / cm³. 3 RMSE = 0.02, corresponding to a numerical matching score of 0.92 (out of 1).

[0087] Spatial structure matching degree: The Structural Similarity Index (SSIM) is used to assess the spatial morphology of the mineralization probability field and the borehole mineralization section, with a weight of 40%. If the predicted deviation of the ore body dip angle is ≤5°, the SSIM ≥0.85, and the corresponding structural matching score is 0.88.

[0088] The authenticity score is calculated by weighting: 0.6 × (1 - RMSE) + 0.4 × SSIM.

[0089] (2) Coupling analysis of mineralization correlation score

[0090] Multi-source data feature fusion: The corrected mineralization probability field (3D grid data) and auxiliary feature vector (128-dimensional alteration spectral features + 3-dimensional geochemical principal components) are input into the fully connected layer to generate a 20-dimensional mineralization association feature vector.

[0091] Coupling relationship quantification:

[0092] Alteration-mineralization coupling degree: Calculate the spatial overlap rate between the high mineralization probability region (>0.7) and the tourmaline-mica alteration probability field. If the overlap rate is >60%, the score is 0.75.

[0093] Element-mineralization coupling: The correlation between mineralization probability and Sn and W content was analyzed by Pearson correlation coefficient. When the correlation coefficient was >0.7, the score was 0.82.

[0094] The correlation score was calculated by weighting as follows: 0.5 × alteration-mineralization coupling degree + 0.5 × element-mineralization coupling degree.

[0095] Finally, the weights are dynamically adjusted based on the complexity of the geological scene, especially for high-complexity scenes. The weighting for authenticity score is set to 0.7, and the weighting for relevance score is set to 0.3; the opposite applies to low-complexity scenarios. The overall score is calculated as ω1 × Authenticity Score + ω2 × Relevance Score.

[0096] In some embodiments, dynamically adjusting the branch structure parameters of the discriminator based on the complexity of the physical property characteristics of the difference region includes:

[0097] By calculating the correlation matrix between the spatial gradient feature map and the pre-trained parameters, a differentiated update strategy is implemented for the two parameter sets based on the correlation matrix; the two parameter sets are the parameters of the last few Transformer encoder layers and the classification head parameters of the discriminator branch.

[0098] First, by calculating the correlation matrix Ω between the DI spatial gradient feature map (reflecting material complexity) and the pre-trained parameters (including at least the parameters of the last few Transformer encoder layers and the classification head parameters of the discriminator branches), a basis for complexity-aware parameter updates is established. The mathematical expression for the correlation matrix Ω is:

[0099]

[0100] in, For DI spatial gradient feature map (3D tensor); W T w C These are the Transformer encoder parameters and the classification head parameters, respectively; Embed() is the embedding function that maps the parameter tensor to the feature space; W Ω The weight matrix is ​​a learnable matrix, which is optimized through backpropagation.

[0101] Then, a differentiated update strategy is implemented on the two parameter sets using the correlation matrix constructed above. The two parameter sets are the Transformer encoder parameters of the last few layers (e.g., the last 3 layers) of the discriminator and the classification head parameters of the discriminator branches, respectively. The details are as follows:

[0102] Transformer encoder parameters (W) T ): Through channel-space attention mask (M) T The multi-head attention weight is dynamically adjusted, and the calculation formula is as follows:

[0103]

[0104] in, For element-wise multiplication, GradCAM is a gradient-weighted activation map, W M σ is the learnable weight matrix; σ is the Sigmoid activation function.

[0105]

[0106] Among them, W' T These are the adjusted Transformer parameters, where α is the learning rate. For the loss function on W T The gradient.

[0107] Classification header parameters (W) C ): Design a geologically constrained attention pooling layer, using the DI gradient direction field Guided feature aggregation:

[0108]

[0109] PooledFeature is the final output pooled feature, which is the result of weighted aggregation of the original features according to rules, and is used for subsequent model calculations (such as classification and inversion).

[0110] α i,j,k Spatial attention weights reflect the feature importance of the (i,j)th position and the kth channel on the feature map. They can be calculated by attention mechanisms (such as Softmax) to allow the model to focus on key regions.

[0111] Feature i,j,k This represents the feature value at spatial location (i,j) in the k-th channel of the original feature map.

[0112] λ is the directional constraint coefficient; the larger λ is, the greater the deviation from g. center The more directional the feature, the greater the weight decay;

[0113] g i,j,k The direction vector corresponding to the position (i,j,k) on the feature map (such as prior information such as the extension direction of the geological body and the strike of the ore body);

[0114] This is the direction vector of the frame's center.

[0115] Based on the above adjustment method, first through α i,j,k Highlight key spatial features; then use exp(-λ· Add directional filters to features: and g center The more consistent the features, the higher the weight, forcing the model to focus on directions that conform to geological priors (such as the direction of mineralization zone extension) and filtering out irrelevant interference (such as the chaotic direction of the surrounding rock).

[0116] It is understandable that the input of the discriminator is the intermediate features after the encoder has performed structured abstraction on the original geological data. The geologically interpretable features (such as the gradient features of fault zones and the spatial distribution patterns of ore bodies) are the feature maps mentioned above.

[0117] PooledFeatures are geologically constrained (orientation, attention) filtered features that are input into a classification head (e.g., used to predict "mineralization probability" or "lithology category"). The classification head's parameter W... C (Such as fully connected layer weights) Based on these filtered features, predictions are made, and then the loss between the predicted results and the true labels (such as cross-entropy loss) is calculated. The classification head parameters W are then adjusted based on gradient descent. C (Such as the weights of fully connected layers), to make the next prediction more accurate. This invention indirectly adjusts the classification head parameter W based on geological constraints. C This allows it to focus more on features that conform to geological laws and avoid learning irrelevant noise.

[0118] In some embodiments, a preset color is used to mark the portions of the three-dimensional mineral exploration model corresponding to each of the different regions.

[0119] To facilitate reference for relevant personnel, the present invention further marks the parts of the three-dimensional mineral exploration model corresponding to the different regions with preset colors, so that relevant personnel can focus on the analysis and verification of the post-processed parts.

[0120] like Figure 2 As shown, embodiments of the present invention also disclose an electronic device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any of the preceding claims.

[0121] This invention also discloses a computer storage medium storing a computer program that can be executed by a processor to implement the method as described in any of the preceding claims.

[0122] This invention also discloses a computer program product comprising a computer program that can be executed by a processor to implement the method as described in any of the preceding claims.

[0123] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0124] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for prospecting granite-type tin deposits based on multi-source spatial scales, characterized in that, Includes the following steps: Based on the detection data obtained by using gravity and magnetic methods and electromagnetic methods in the target area, an inversion algorithm was used to construct a first three-dimensional mineral exploration model and a second three-dimensional mineral exploration model, and a comparison was made between the first three-dimensional mineral exploration model and the second three-dimensional mineral exploration model to identify several areas of difference. Auxiliary detection data corresponding to another detection method different from gravity and magnetic methods and electromagnetic methods is obtained. Based on the auxiliary detection data and the joint inversion algorithm driven by machine learning, the difference regions are corrected to obtain the difference region model. The models of the different regions are fused with the first three-dimensional prospecting model or the second three-dimensional prospecting model to obtain the final three-dimensional prospecting model of granite-type tin deposits. The comparison revealed several areas of difference between the first 3D mineral exploration model and the second 3D mineral exploration model, including: The spatial data of the first three-dimensional mineral exploration model and the second three-dimensional mineral exploration model are registered, and the difference in physical property parameters at the corresponding locations is calculated by the density-resistivity joint difference index. Areas exceeding the preset threshold are selected as first-level difference areas. Morphological analysis was performed on the primary differential regions to extract continuously distributed significant differential bodies. Based on the contradiction type of gravity-magnetic-electromagnetic response combined with geological prior knowledge, several differential regions were finally identified. Based on the auxiliary detection data and the machine learning-driven joint inversion algorithm, the differences in each region are corrected to obtain a difference region model, including: Acquire auxiliary detection data, including the spectral characteristics of altered minerals obtained by hyperspectral remote sensing and the distribution characteristics of metal elements obtained by borehole geochemical measurements, and extract auxiliary feature vectors from the auxiliary detection data; The difference region between the first three-dimensional mineral exploration model and the second three-dimensional mineral exploration model, along with the auxiliary feature vector, is input into a joint inversion model constructed based on a generative adversarial network. The generator of the joint inversion model generates a corrected geological attribute field, and the discriminator is used to evaluate the authenticity and mineralization correlation of the corrected geological attribute field. When the mineralization probability value output by the generator exceeds a set threshold, the corresponding difference region is marked as a potential ore body region, and a corrected difference region model is obtained.

2. The method for prospecting granite-type tin deposits based on multi-source spatial scales according to claim 1, characterized in that: The discriminator includes an authenticity branch and a mineralization correlation branch. The authenticity branch is used to determine the degree of matching between the generated geological attribute field and the measured borehole data, and the mineralization correlation branch is used to determine the coupling relationship between the corrected geological attribute field and the mineralization indication information in the auxiliary features.

3. The method for prospecting granite-type tin deposits based on multi-source spatial scales according to claim 2, characterized in that: The discriminator is used to evaluate the authenticity of the corrected geological attribute field and its correlation with mineralization, including: The branch structure parameters of the discriminator are dynamically adjusted based on the complexity of the physical property characteristics of the difference region. The complexity of the physical property characteristics is calculated by the spatial gradient of the density-resistivity joint difference index. The authenticity branch receives the corrected geological attribute field and the measured borehole data, calculates the spatial difference between the two in multiple attribute dimensions, evaluates the degree of matching between the corrected geological attribute field in terms of numerical value and spatial structure, and obtains the authenticity score. The mineralization correlation branch receives the corrected geological attribute field and the auxiliary feature vector, analyzes the coupling relationship between the mineralization probability distribution and hyperspectral alteration characteristics and geochemical element enrichment patterns, and obtains the correlation score. The authenticity score and the relevance score are weighted to obtain a comprehensive discrimination score, which is used as the discrimination signal for generator training.

4. The method for prospecting granite-type tin deposits based on multi-source spatial scales according to claim 3, characterized in that: Based on the complexity of the physical properties of the differential regions, the branch structure parameters of the discriminator are dynamically adjusted, including: By calculating the correlation matrix between the spatial gradient feature map and the pre-trained parameters, a differentiated update strategy is implemented for the two parameter sets based on the correlation matrix; the two parameter sets are the parameters of the last few Transformer encoder layers and the classification head parameters of the discriminator branch.

5. The method for prospecting granite-type tin deposits based on multi-source spatial scales according to claim 1, characterized in that: The portions of the three-dimensional mineral exploration model corresponding to the different regions are marked with preset colors.

6. An electronic device, characterized in that: The electronic device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1-5.

7. A computer storage medium, characterized in that: The computer storage medium stores a computer program that can be executed by a processor to implement the method as described in any one of claims 1-5.

8. A computer program product, characterized in that: The computer program product includes a computer program that can be executed by a processor to implement the method as described in any one of claims 1-5.

Images