A deep mineral exploration method based on multi-source information fusion and overlay

Abstract

This invention discloses a deep mineral exploration method based on multi-source information fusion and overlay, comprising: acquiring various trace element contents, fluid inclusion physicochemical parameters, and infrared spectral data from several sampling points in the exploration area; extracting geochemical indicators based on trace element contents, extracting spectral parameter indicators based on infrared spectral data and generating model prediction indicators, and selecting physical environment indicators from the fluid inclusion physicochemical parameters; interpolating non-sampling point grids in the exploration area using the indicator data from the sampling points for each indicator; obtaining a mineralization potential index for each grid point by weighting the indicators, and correcting the mineralization potential index based on the physical environment indicators; and determining the deep mineral exploration target area in the exploration area based on the geological background and the corrected mineralization potential. This invention establishes a dynamic mineral exploration model by comprehensively utilizing multi-source geoscientific information, and corrects and optimizes the model using multi-source monitoring data, thereby achieving accurate prediction of deep concealed ore bodies.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, specifically to a deep mineral exploration method that integrates and superimposes multi-source information. Background Technology

[0002] Identification and prospecting prediction of deep concealed ore bodies is an important research direction in the field of geological prospecting. Although traditional drilling and geophysical methods can provide local information, they suffer from high costs, low efficiency, and limited spatial resolution, making it difficult to accurately characterize deep mineralization systems. Existing single technologies (such as geochemistry, fluid inclusions, and infrared spectroscopy) can provide some mineralization information, but the information is isolated and has multiple interpretations, making it difficult to systematically characterize the spatial distribution of deep hydrothermal channels and ore bodies.

[0003] In recent years, short-wave infrared spectroscopy (SWIR, 300–2500 nm) and thermal infrared spectroscopy (TIR, 8000–14000 nm) have been widely used in alteration mineral identification, mineralization zoning analysis, and hydrothermal activity identification due to their advantages such as speed, non-destructive nature, and in-situ measurement capabilities. However, single spectral information cannot comprehensively reveal the characteristics of mineralization systems. Trace element geochemistry and fluid inclusion analysis can constrain mineralization conditions such as hydrothermal source, temperature, salinity, and pH, but lack spatial continuity. Therefore, there is an urgent need for an integrated prospecting method that can integrate multi-source information such as spectral, geochemical, and fluid inclusion data to achieve intelligent prediction of deep deposits and target area optimization. Although existing studies have revealed the single-factor spectral and geochemical characteristics of ore deposits, these characteristics exhibit ambiguity and isolation in geospatial contexts. Summary of the Invention

[0004] This invention provides a deep mineral exploration method that integrates and superimposes multi-source information. By comprehensively utilizing multi-source geoscience information, a dynamic mineral exploration model is established, and the model is corrected and optimized using multi-source monitoring data, thereby achieving accurate prediction of deep concealed ore bodies.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0006] A deep mineral exploration method based on multi-source information fusion and overlay includes:

[0007] Obtain the contents of various trace elements, physicochemical parameters of fluid inclusions, and infrared spectral data from several sampling points in the exploration area;

[0008] Geochemical indicators were extracted based on trace element content, and model prediction indicators and spectral parameter indicators were extracted based on infrared spectral data. Physical environment indicators were selected from the physicochemical parameters of fluid inclusions. For each indicator, the indicator data of the sampling points were used to interpolate the non-sampling point grid of the exploration area.

[0009] For each grid point, a mineralization potential index is obtained by weighting geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators, and the mineralization potential index is then corrected based on the physical environment indicators.

[0010] Based on the revised mineralization potential index, the deep mineral exploration target area of ​​the exploration area is determined.

[0011] Furthermore, the physicochemical parameters of the fluid inclusions include temperature, pH value and / or pressure.

[0012] Furthermore, spectral parameters are extracted based on infrared spectral data, including illite crystallinity based on short-wave infrared and / or quartz crystallinity based on thermal infrared.

[0013] Furthermore, the extraction of spectral parameters based on infrared spectral data includes:

[0014] The spectral data of all wavelengths in the short infrared or thermal infrared are reconstructed into a two-dimensional spectral image, with each pixel representing the spectral data of one wavelength.

[0015] Sensitive wavelength bands can be screened by using orthogonal partial least squares discriminant analysis or by performing Grad-CAM thermal analysis on a convolutional neural network that predicts mineralization probability based on spectral data matrix.

[0016] Spectral parameters are calculated using spectral data in the sensitive wavelength range.

[0017] Furthermore, orthogonal partial least squares discriminant analysis is used to screen sensitive wavelength bands, including:

[0018] The orthogonal partial least squares discriminant analysis method is used to decompose the spectral data matrix X into a spectral feature part that can distinguish the mineralization level and a noise part that is unrelated to the mineralization level.

[0019] Based on mineralization-related weight matrix Calculate the projection importance index (VIP) for each wavelength point, that is, calculate the contribution of each wavelength point to mineralization classification. After projection, each wavelength point obtains a VIP value.

[0020] If the VIP value of a certain band exceeds the preset threshold, the band is identified as a sensitive band.

[0021] Alternatively, Grad-CAM heatmap analysis can be performed on a convolutional neural network that predicts mineralization probability based on a spectral data matrix to screen for sensitive wavelength bands, including:

[0022] Using a two-dimensional spectral image with known mineralization levels as input to a trained convolutional neural network model, the gradient of the target category score with respect to the last convolutional feature map is calculated, and global average pooling is performed to obtain the importance score of each feature channel. :

[0023]

[0024] In the formula, Representing the The feature channel (i.e. the first feature channel) (wavelength bands) for high-potential mineralization categories Importance weights; The output score for the category belonging to high-potential mineralization; For the last convolutional feature map The spatial location of each channel The activation response value at the location; and The height and width of the two-dimensional spectral image are input into the model, respectively.

[0025] Furthermore, the extraction of geochemical indicators based on trace element content specifically involves: using factor analysis algorithms to extract multiple element combinations and calculating the score of each element combination; determining the element combination most relevant to mineralization among the multiple element combinations based on prior knowledge, and using its score as the geochemical indicator.

[0026] Furthermore, based on infrared spectral data, model prediction indicators are extracted. Specifically, a neural network model trained on spectral data is used to predict the probability that a sampling point belongs to a high-potential mineralization, which serves as the model prediction indicator.

[0027] Furthermore, the weighted calculation of all indicators includes:

[0028] First, the objective weights of geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators are determined using the entropy weight method.

[0029] Then, prior knowledge is introduced to correct the objective weights, resulting in the coupled weights of geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators.

[0030] Finally, the mineralization potential index is obtained by weighting geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators based on coupling weights.

[0031] Furthermore, the fluid homogenization temperature of fluid inclusions is selected as a physical environment indicator to correct the mineralization potential index, specifically including:

[0032] Establish a space constraint function based on fluid cooling precipitation logic When the fluid homogenization temperature of the fluid inclusion When within the ore-forming temperature range, The value is 1; when the fluid homogenization temperature of the fluid inclusion is... When the temperature is less than or greater than the ore-forming temperature range, It exhibits Gaussian decay with temperature difference;

[0033] Furthermore, for geochemical indicators indicating mineralization anomalies within the study area, if the spatial temperature gradient modulus calculated based on the fluid temperature field model for a certain grid point is greater than the preset multiple of the cooling gradient at the non-altered rock of the deposit, it indicates that the grid point is located on an effective hydrothermal migration path. The value is 1.

[0034] Furthermore, based on the revised mineralization potential, the deep mineral exploration target areas of the exploration region are identified, including:

[0035] If the mineralization potential of a grid after correction is above the 90th percentile and spatially overlaps with the anomaly center of the near-ore halo zone, then the grid is a high-potential mineralization area.

[0036] If the mineralization potential of a certain grid after correction is at the 75-90th percentile, then the grid is a mineralization potential zone;

[0037] If the corrected mineralization potential of a grid is below the 75th percentile, then the grid is considered a low-potential mineralization zone.

[0038] Compared with the prior art, the invention employing the above technical solution has the following advantages:

[0039] High information integration: By comprehensively utilizing multi-source geoscience information such as elemental geochemistry, infrared spectroscopy and fluid inclusions, a comprehensive deep mineral exploration index system has been established, overcoming the coupling and limitations of single technical means;

[0040] High level of intelligence: By introducing machine learning algorithms such as convolutional neural networks and OPLS-DA, the automatic identification and extraction of mineralization parameter features are realized, reducing the subjectivity of human interpretation;

[0041] High prediction accuracy: Through multi-source data fusion and spatial overlay analysis, the accuracy of locating deep ore bodies is significantly improved;

[0042] Low exploration cost: Compared with traditional drilling methods, the cost is reduced to a very low level, and some steps achieve non-destructive testing;

[0043] Wide range of applications: It is not only applicable to antimony deposits, but can also be extended to deep mineral exploration in other types of hydrothermal deposits. Attached Figure Description

[0044] Figure 1This is a schematic flowchart of the method of the present invention;

[0045] Figure 2 The classification and feature selection diagram for OPLS-DA;

[0046] Figure 3 This is a diagram showing the CNN training process and feature extraction.

[0047] Figure 4 Contour plot of factor scores and shortwave infrared parameters;

[0048] Figure 5 This is a diagram showing the fluid temperature field distribution.

[0049] Figure 6 A graph for overlaying and analyzing multi-source information;

[0050] Figure 7 Metallogenic model diagram of the Banxi deposit. Detailed Implementation

[0051] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0052] This embodiment provides a deep mineral exploration method that integrates and overlays multi-source information, taking a certain antimony deposit as the research object, such as... Figure 1 As shown, follow these steps:

[0053] 1. Sampling design and data acquisition.

[0054] Step 101: Construct a multi-source data acquisition system.

[0055] 110 sampling points were established at the intersection of the exploration line and five vertical sections (sections 17, 19, 21, 22, and 23) in the mining area. Sampling was denser along the exploration line in structurally developed areas, with a spacing of 10-20 meters, and sparser along the exploration line in non-structural areas, with a spacing of 50 meters. Rock samples were collected at each sampling point, and 3-5 nearby samples were collected from each point and mixed together, with a sample weight of 1-2 kg to ensure sample representativeness. This invention is also applicable to horizontal surface and intermediate section sampling.

[0056] Step 102: Multi-technology collaborative testing and data acquisition.

[0057] Trace element analysis: The contents of 19 trace elements in 87 samples were analyzed by ICP-MS / OES. The analyzed elements included Au, As, Sb, Ag, Cd, Cu, Pb, Zn, Co, Cr, Fe, Ni, W, Sn, and Bi.

[0058] Fluid inclusion testing: 29 representative samples were selected, and temperature data of 29 fluid inclusions were obtained using the Linkam THMS600 hot and cold stage temperature measurement system.

[0059] Infrared spectroscopy scanning: 91 samples were scanned using an ASD TerraSpec 4 high-resolution mineral spectrometer (SWIR, 300-2500 nm) and an Agilent 4300 handheld FTIR spectrometer (TIR, 8000-14000 nm), acquiring 2300 spectral data.

[0060] Step 103: Data transmission and preprocessing, i.e., performing max-min standardization on each type of data collected in step 102.

[0061] 2. Extract multiple indicators from multi-source information.

[0062] Step 201: Geochemical Indicators. This involves using factor analysis to analyze the content of all trace elements, extracting principal factors (elemental combinations), and identifying the zonal characteristics of each element combination.

[0063] First, calculate the correlation coefficient matrix R of all trace elements based on their elemental content, and then solve for the eigenvalues ​​of the coefficient matrix R. The eigenvector U is used to output the KMO and Bartlett's sphericity test results to determine the effectiveness of factor analysis.

[0064] Based on factor analysis to extract principal factor combinations, the factor loading matrix is ​​first calculated. :

[0065]

[0066] In the formula, U is the eigenvector matrix of the microelement correlation coefficient matrix R. It is a diagonal matrix composed of eigenvalues;

[0067] Then, the factor score coefficient matrix B and the geochemical factor score F for each sampling point on each principal factor are calculated:

[0068]

[0069] In the formula, R is the trace element correlation coefficient matrix, X is the trace element content matrix of the sampling points after the maximum and minimum standardization process, and F is the geochemical factor score of each sampling point on each principal factor.

[0070] For example, four principal factors were extracted, and the cumulative variance contribution rate (71.32%) was calculated: Among them, factor F1 (Ag, Cd, Cu, Pb, Zn): the element combination of the leading edge halo, which should be located in the shallow part and periphery of the ore body; factor F2 (Co, Cr, Fe, Ni, W): the element combination of the tail halo, which should be located in the deep part of the ore body; factor F3 (Au, As, Sb): the element combination of the near-ore halo, which matches the spatial location of the ore body; factor F4 (Bi, Sn): the high-temperature element combination, which should be located in the fluid center.

[0071] Furthermore, factor rotation, specifically orthogonal rotation using the variance maximization method, can be used to polarize factor loadings towards 0 or 1, making the factor loadings of all elements easier to interpret. This yields clear elemental combinations (e.g., F1: Ag-Cd-Cu-Pb-Zn; F2: Co-Cr-Fe-Ni-W; F3: Au-As-Sb), and the score for each sampling point based on the trace element content on each principal factor can be calculated.

[0072] Among them, the elemental assemblage associated with ore-forming elements is the elemental assemblage near the ore body. The score of the elemental assemblage near the ore body is used as a geochemical index for subsequent weighted calculation of the ore-forming potential index. Based on prior knowledge, other elemental assemblages are compared with the elemental assemblage near the ore body. The elemental assemblage with a lower temperature than the elemental assemblage near the ore body is the elemental assemblage at the leading edge of the ore body, located in the periphery and shallow part of the ore body; the elemental assemblage with a higher temperature than the elemental assemblage near the ore body is the elemental assemblage at the tail edge of the ore body, located in the deep part of the ore body and the hydrothermal center.

[0073] Step 202: Extract spectral parameter indicators.

[0074] First, the short-wave infrared spectral data (wavelength range 300-2500 nm, totaling 2150 nm after noise removal) of rock samples collected from each sampling point in the mining area were integrated into a spectral data matrix. Simultaneously, based on the geological logging results (mineralization intensity, alteration degree, and distance from the ore body), the samples were divided into three categories: high potential, medium potential, and low potential, thus constructing a mineralization category matrix. For the independent variable matrix Centralization is performed to eliminate the influence of different spectral band dimensions on model stability. The quantitative discrimination criteria for high, medium, and low mineralization need to be determined in conjunction with the actual deposit conditions. The aim is to extract spectral parameters related to high mineralization levels using OPLS-DA and CNN, and to train a CNN-based model to predict probabilities directly related to hydrothermal activity in the main mineralization stage during actual mineral exploration. For example, high-potential samples must meet at least two of the following three conditions: high grade, strong alteration, and location within or at a certain distance from a vein.

[0075] (1) Wavelength band selection based on OPLS-DA. That is, orthogonal partial least squares discriminant analysis (OPLS-DA) is used to select spectral parameters with VIP>1 from the SWIR spectral data.

[0076] First, the spectral data matrix X is decomposed into two parts: a spectral feature component that distinguishes mineralization grades and a noise component that is unrelated to mineralization grades. Specifically, by maximizing the covariance between the spectral data matrix X and the mineralization category matrix Y, the component in X that is related to the mineralization category matrix is ​​extracted. The components with the highest correlation are used as predicted components, and a mapping relationship between spectral features and mineralization grades is established:

[0077]

[0078] in, This represents the spectral characteristic intensity matrix related to the wavelength and mineralization. for The corresponding weight matrix, This is a noise matrix that is independent of mineralization at the wavelength point. for The corresponding weight matrix, E is the residual matrix, U Q The response score matrix (which connects the response variables with the predicted scores by weights, establishing a quantitative relationship between spectral characteristics and mineralization grade) is given by F, where F is the response residual.

[0079] Then, based on the weight matrix related to mineralization The projected importance index (VIP) for each wavelength point is calculated, which is the contribution of each wavelength point to mineralization classification. After projection, each wavelength point yields a VIP value. The formula for calculating the VIP value is as follows:

[0080]

[0081] In the formula, For the first VIP value at each wavelength point; This represents the total number of wavelength points. The spectral characteristic intensity matrix obtained by decomposing X Number of principal components; : The variance explained by the a-th component; For the first The weight of each wavelength point in the a-th component.

[0082] Screen key spectral parameters and set This serves as a screening threshold for sensitive wavelength bands. For example, if the VIP value continuously exceeds 1.0 within a certain wavelength range (such as 2180-2220 nm), then that wavelength range is determined to be a sensitive wavelength band (actually the core alteration indicator band of the deposit, such as the Al-OH absorption band of sericite).

[0083] (2) Train a CNN-based mineralization classification model.

[0084] The 2150 wavelength points of SWIR (300 nm-2500 nm, excluding the first 50 noise points) were reconstructed into a 50×43 two-dimensional matrix, where each element represents the spectral data of one wavelength point. A two-layer convolutional neural network (Conv2D(32)-MaxPooling2D-Conv2D(64)-MaxPooling2D-Dense(128)-Output) was constructed.

[0085] Input: Dimension The original spectrum was a one-dimensional spectrum with 2150 feature points. It was then reshaped into... The purpose of this two-dimensional matrix is ​​to capture the "local texture" features of the spectral curve (such as the slope of the absorption peak and the shape of the shoulder) using the convolution kernel. This matrix allows the convolution kernel to take into account the correlation information between the center and the wings of the peak during scanning, and is more capable of identifying the fine fingerprints of altered minerals than pure one-dimensional feature extraction.

[0086] Convolution kernel: size Small-sized convolution kernels are chosen to facilitate deep feature overlay. When processing spectral data at 2nm resolution, Convolution kernels can precisely lock the morphological changes of alteration characteristic peaks (usually spanning 5-15 wavelength points), avoiding the feature smoothing caused by large-sized convolution kernels, thus preserving the slight spectral shifts caused by mineralization.

[0087] Dropout rate: 0.5. Geological sampling is limited by downhole engineering, and the sample size is relatively small compared to the needs of deep learning. Setting a random dropout rate of 0.5 forces the network to learn more robust feature combinations, effectively reducing the model's dependence on specific sampling segments and preventing "overfitting" from causing deep prediction failure.

[0088] The activation function is ReLU, which has a constant gradient in the positive interval. It can effectively overcome the gradient vanishing problem in deep computing of geoscience big data, enabling the model to converge quickly and identify nonlinear alteration enhancement signals in the spectrum.

[0089] Convolution operation formula:

[0090]

[0091] in The convolution kernel weight parameters are used to automatically learn and identify mineralization-related spectral patterns through training. This is to input the absorption intensity values ​​for a specific wavelength range in the spectral data; This is a bias term that enhances the model's robustness to background noise. The ReLU activation function is used to introduce nonlinearity and highlight significant mineralization signals; To iterate through the convolution kernels, a 3×3 convolution kernel is summed at local positions on a 50×43 spectral data matrix; the final output is... The extracted spectral feature wavelengths are located at the following positions. The response intensity at that location.

[0092] Training Optimization: The Adam optimizer (learning rate 0.001) was used, with the reconstructed spectral 2D matrix and corresponding mineralization category from the training samples as input and output, respectively. Five-fold cross-validation was employed, and training was conducted for 1000 epochs with a batch size of 32. The classification cross-entropy loss was minimized to obtain a photomineralization classification model suitable for SWIR. The classification cross-entropy loss is:

[0093]

[0094] In the formula, This represents the total error loss predicted by the model, where N is the total number of training samples and C is the number of mineralization categories. and These are the sample and category indexes, It is a sample The actual category labels (high, medium, low). The sample predicted by the model Category The probability is obtained through the Softmax output. By minimizing this loss value, the model can continuously adjust its parameters to make the predicted probability distribution as close as possible to the actual geological classification, thereby effectively distinguishing different mineralization potential levels.

[0095] Similarly, by applying the above training method to TIR spectral data (different wavelength ranges require different reconstruction matrices), a photomineralization classification model suitable for TIR can be obtained.

[0096] (3) Screening of sensitive bands based on Grad-CAM thermal map analysis:

[0097] Using the mineralization classification models for SWIR and TIR obtained through the above training, a Gradient Class Activation Mapping (Grad-CAM) algorithm (handwritten) is used to calculate the gradient of the classification probability with respect to the feature map, quantifying the importance of each local feature region and generating a heatmap. Specifically, this includes constructing a gradient propagation model from the input to the output of the last convolutional layer (conv2 layer in this example) and the final Softmax classification output; forward propagation using known test set sample images to obtain the class scores. and the feature map response of the last convolutional layer The gradient partial derivative of the category score with respect to the feature map is calculated using automatic differentiation (Gradient Tape) technique, and global average pooling is performed on the spatial dimension to obtain the feature channel importance score. :

[0098]

[0099] In the formula, Representing the The feature channel (i.e. the first feature channel) (wavelength bands) for high-potential mineralization categories Importance weights; The output score for the category belonging to high-potential mineralization; For the last convolutional feature map The spatial location of each channel The activation response value at the location; and These represent the height and width of the feature map, respectively. Then, the activation value of each channel is correlated with its corresponding channel importance score. The products are multiplied together, averaged along the channel axis, and finally filtered out for negative contributions using the ReLU activation function. The result is then normalized to obtain a two-dimensional heatmap. :

[0100]

[0101] By mapping the two-dimensional grid nodes of the heatmap back to the original one-dimensional spectral band coordinates (excluding the first 50 nm noise bands, 351-2500 nm, a total of 2150 points), sensitive band intervals with extremely strong mineralization responses are automatically selected. This evaluation method quantifies the contribution of spectral data corresponding to different bands to mineralization identification by calculating the weighted feature response value of gradient-based activation mapping, thereby automatically selecting mineralization-sensitive bands.

[0102] Key findings include: the CNN model identified characteristic wavelength ranges [1920, 2017] nm and [2214, 2311] nm associated with high mineralization, corresponding to the ~1900 nm H2O absorption characteristics and the ~2200 nm Al-OH absorption characteristics of sericite, respectively.

[0103] (4) Extract spectral parameters.

[0104] Key mineralization-sensitive bands identified by OPLS-DA and CNN were combined, and key spectral parameters were extracted based on existing research, such as: Pos2200, IC value (Dep2200 / Dep1900), Pos2250, Pos2350, R8200 / R7000, Wid9300, and Dep9800. In this embodiment, illite crystallinity (SWIR-IC) and quartz crystallinity (Wid9300) in the short infrared spectrum were selected as spectral parameter indicators for subsequent weighted calculation of the mineralization potential index.

[0105] Step 203: Generate model prediction metrics.

[0106] In this embodiment, the mineralization classification model trained above is used to extract the probability that the input spectral two-dimensional image belongs to high-potential mineralization, which is denoted as the model prediction index. Simultaneously, a processing path for low-confidence branch samples is established based on the classification probability distribution output by the model: a confidence threshold is set (e.g., a sample is considered low-confidence when the maximum class prediction probability is less than 0.5); for the spectra of abnormal or boundary sampling points judged as low-confidence, a spatial neighborhood smoothing strategy is adopted, that is, the average model prediction probability of other high-confidence samples within a preset radius around the grid point is used to interpolate and smooth its probability value, thereby eliminating the interference of local spectral noise on the overall spatial prediction index.

[0107] Step 204: Extract physical environment indicators.

[0108] Based on the homogenization temperature of fluid inclusions ( The normalized score of fluid inclusions within a suitable temperature range (e.g., 180-220℃) is used as a physical environment indicator for subsequent weighted calculation of the mineralization potential index. In other embodiments, the normalized pressure data of fluid inclusions within a preset pressure range can also be used as a physical environment indicator.

[0109] This embodiment constructs a fluid temperature field contour model based on fluid inclusion microthermometry data:

[0110] High-temperature zone (>250℃): High pressure, corresponding to the hydrothermal center, should spatially coincide with the high-value zone of SWIR-IC;

[0111] Mid-temperature zone (180-220℃): The main mineralization temperature range of stibnite, which should correspond to high-grade ore bodies;

[0112] Low-temperature zone (<180℃): The pressure is relatively low, which should correspond to the low-temperature alteration zone on the periphery of the deposit;

[0113] Hydrothermal channels are identified using the aforementioned fluid temperature field contour lines: based on the hydrothermal migration patterns and combined with the gradient change direction of the temperature field contour lines (i.e., the temperature / pressure decrease direction) and the spatial distribution of the structure, faults that serve as hydrothermal migration channels are identified.

[0114] Meanwhile, this fluid temperature field model provides key temperature constraints for the selection of deep target areas and realizes the visual characterization of the ore-forming fluid system.

[0115] 3. Multi-source information fusion.

[0116] Step 301: Construct an indicator system.

[0117] This embodiment uses the following indicators extracted in step 2 to construct the indicator system:

[0118] Geochemical indicators ( For example, the element combination factor score of the near-ore halo.

[0119] Model prediction metrics ( ): The predicted probability value of the "High" category output by the CNN model, representing the mineralization correlation of the alteration combination.

[0120] Spectral parameters ( ): High SWIR-IC (illite crystallinity) value, and an abnormally low Wid9300 (quartz crystallinity) value.

[0121] Physical environmental indicators ( ): Homogenization temperature of fluid inclusions ( The normalized score within a suitable temperature range (e.g., 180-220℃).

[0122] Step 302: Indicator space interpolation gridding.

[0123] Due to spatial misalignment of the data from trace element analysis (87 points), fluid inclusion testing (29 points), and infrared spectroscopy scanning (91 points) at each sampling point, this embodiment first analyzes the geochemical indicators. Model prediction indicators Spectral parameters and physical environment indicators Ordinary kriging interpolation was performed on spatial contour lines to spatially grid the indices, generating continuous grid indices covering non-sampling grid points in the exploration area: geochemical grid indices, model-predicted grid indices, spectral parameter grid indices, and physical environment grid indices. The formula for calculating the estimated values ​​of ordinary kriging interpolation is as follows:

[0124]

[0125] In the formula, The spatial coordinates of the grid points to be estimated; These are estimated values ​​for grid point indices; For drift or trend items; The number of sample points is known; Given the spatial coordinates of the sample points; These are the actual observed values ​​of known sample points; The Kriging weights are assigned to known sample points.

[0126] Step 303: Quantitative calculation of combined weights.

[0127] To avoid the arbitrariness of human-imposed weights, this embodiment first uses the entropy weight method to calculate the objective contribution of each data source.

[0128] (1) Calculate the first Entropy value of each indicator :

[0129]

[0130] in These are the normalized index observations.

[0131] (2) Calculate the objective weights Introducing prior knowledge from geological experts to adjust weights. (For example, near a known tectonic fracture zone, increase) The weights of geochemical indicators are assigned, ultimately resulting in the coupling weights of each indicator. .

[0132] Step 304: Establish a deep mineralization potential evaluation model

[0133] Using the grid index obtained in step 302, in each non-sampling point grid cell Spatial overlay and weighted calculations are performed to generate a grid-based mineralization potential index. :

[0134]

[0135] In the formula, For grid cells The mineralization potential index For grid cells The Individual indicator values ​​(i.e.) to ); The objective and subjective coupling weights of the i-th indicator.

[0136] Step 305: Spatial logic filtering.

[0137] Establish a space constraint function based on fluid cooling precipitation logic For the mineralization potential index Make corrections:

[0138]

[0139]

[0140] In the formula, : Spatial constraint function value based on fluid cooling and sedimentation logic (value range between 0 and 1, used as a correction coefficient multiplied into the potential index); Observed values ​​of the homogenization temperature of fluid inclusions; and : The lower and upper limits of the most suitable temperature range for the main mineralization stage (for example, in this embodiment, the main mineralization temperature range of stibnite is 180-220℃). : The central reference temperature of the ore-forming temperature range (an average value can be taken; in this example, it is 220℃); Gaussian decay control variance (usually a constant greater than zero, for example, between 10 and 30, the magnitude of which determines the rate at which the potential index decays when the fluid temperature deviates from the suitable mineralization range).

[0141] When the fluid homogenization temperature of the fluid inclusion When within the mineralization temperature range (e.g., 180-220℃). The value is 1; when the fluid homogenization temperature of the fluid inclusion is... When the temperature is less than or greater than the ore-forming temperature range, It exhibits Gaussian decay with temperature difference.

[0142] Meanwhile, for geochemical indicators indicating mineralization anomalies within the study area, this embodiment utilizes the fluid temperature field model constructed in step 204 to analyze each grid cell. Calculate its two-dimensional fluid temperature gradient value. :

[0143]

[0144] Then, the modulus of the two-dimensional temperature gradient in the plane is calculated. :

[0145]

[0146] In the formula, and Grid points The partial derivatives of the spatial variation of temperature in the east-west (X-axis) and north-south (Y-axis) directions. If a certain grid point... Temperature gradient modulus at When the cooling gradient is greater than the preset multiple of the non-altered rock temperature gradient in the ore deposit, it indicates that the grid point is on an effective hydrothermal migration path. In this case, regardless of whether the homogenization temperature of the fluid at that point deviates from the suitable mineralization range, the spatial constraint function... All are forced to be 1. That is... This constitutes a piecewise conditional function: when within the ore-forming temperature range, or on an effective hydrothermal migration path, =1; the remaining areas that deviate from the mineralization conditions show a Gaussian decrease with temperature difference.

[0147] 4. Target area prediction

[0148] According to the revised mineralization potential index To determine the deep mineral exploration target area in the exploration area, including:

[0149] If the mineralization potential of a grid after correction is above the 90th percentile and spatially overlaps with the anomaly center of the near-ore halo zone, then the grid is a high-potential mineralization area.

[0150] If the mineralization potential of a certain grid after correction is at the 75-90th percentile, then the grid is a mineralization potential zone;

[0151] If the corrected mineralization potential of a grid is below the 75th percentile, then the grid is considered a low-potential mineralization zone.

[0152] Figure 2 This demonstrates the classification and feature selection of OPLS-DA in an example of the present invention. Figure 3 This demonstrates the CNN training and feature extraction of the present invention in an example; Figure 4 The figure shows the contour plot of factor scores and shortwave infrared parameters in an example of the present invention. Figure 5 The fluid temperature field distribution of the present invention is shown in an example. Figure 6 This invention demonstrates the multi-dimensional information overlay analysis in an example; Figure 7 The diagram illustrates the metallogenic model of a certain ore deposit. Example contour maps of factors, spectral parameters, and the spatial distribution of fluid temperature visually reflect the correspondence between various mineralization parameters and the spatial location of structures and ore bodies, as well as the geological metallogenic model derived from the integrated analysis of these three factors. A schematic diagram of the geochemical halo is used for geological understanding in conjunction with factor analysis.

[0153] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A deep mineral exploration method based on multi-source information fusion and overlay, characterized in that, include: Obtain the contents of various trace elements, physicochemical parameters of fluid inclusions, and infrared spectral data from several sampling points in the exploration area; Geochemical indicators were extracted based on trace element content, and model prediction indicators and spectral parameter indicators were extracted based on infrared spectral data. Physical environment indicators were selected from the physicochemical parameters of fluid inclusions. For each indicator, the indicator data of the sampling points were used to interpolate the non-sampling point grid of the exploration area. For each grid point, a mineralization potential index is obtained by weighting geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators, and the mineralization potential index is then corrected based on the physical environment indicators. Based on the revised mineralization potential index, the deep mineral exploration target area of ​​the exploration area is determined.

2. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 1, characterized in that, The physicochemical parameters of the fluid inclusions include temperature, pH value and / or pressure.

3. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 1, characterized in that, Spectral parameters are extracted based on infrared spectral data, including illite crystallinity based on short-wave infrared and / or quartz crystallinity based on thermal infrared.

4. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 3, characterized in that, Spectral parameters extracted from infrared spectral data include: The spectral data of all wavelengths in the short infrared or thermal infrared are reconstructed into a two-dimensional spectral image, with each pixel representing the spectral data of one wavelength. Sensitive wavelength bands can be screened by using orthogonal partial least squares discriminant analysis or by performing Grad-CAM thermal analysis on a convolutional neural network that predicts mineralization probability based on spectral data matrix. Spectral parameters are calculated using spectral data in the sensitive wavelength range.

5. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 4, characterized in that, Sensitive wavelength bands were screened using orthogonal partial least squares discriminant analysis, including: The orthogonal partial least squares discriminant analysis method is used to decompose the spectral data matrix X into a spectral feature part that can distinguish the mineralization level and a noise part that is unrelated to the mineralization level. Based on mineralization-related weight matrix Calculate the projection importance index (VIP) for each wavelength point, that is, calculate the contribution of each wavelength point to mineralization classification. After projection, each wavelength point obtains a VIP value. If the VIP value of a certain band exceeds the preset threshold, the band is identified as a sensitive band. Alternatively, Grad-CAM heatmap analysis can be performed on a convolutional neural network that predicts mineralization probability based on a spectral data matrix to screen for sensitive wavelength bands, including: Using a two-dimensional spectral image with known mineralization levels as input to a trained convolutional neural network model, the gradient of the target category score with respect to the last convolutional feature map is calculated, and global average pooling is performed to obtain the importance score of each feature channel. : ； In the formula, Representing the Each wavelength band corresponds to high-potential mineralization categories Importance weights; The output score for the category belonging to high-potential mineralization; For the last convolutional feature map Each wavelength band in spatial location The activation response value at the location; and The height and width of the two-dimensional spectral image are input into the model, respectively.

6. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 1, characterized in that, The model extracts predictive indicators based on infrared spectral data. Specifically, it uses a neural network-based model to predict the probability that a sampling point belongs to a high-potential mineralization, which serves as the model's predictive indicator.

7. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 1, characterized in that, The extraction of geochemical indicators based on trace element content specifically involves: using factor analysis algorithm to extract multiple element combinations and calculating the score of each element combination; determining the element combination most relevant to mineralization among the multiple element combinations based on prior knowledge, and using its score as the geochemical indicator.

8. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 1, characterized in that, The weighted calculation of all indicators includes: First, the objective weights of geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators are determined using the entropy weight method. Then, prior knowledge is introduced to correct the objective weights, resulting in the coupled weights of geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators. Finally, the mineralization potential index is obtained by weighting geochemical indicators, model prediction indicators, spectral parameter indicators, and physical environment indicators based on coupling weights.

9. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 1, characterized in that, The fluid homogenization temperature of fluid inclusions is selected as a physical environment indicator to correct the mineralization potential index, specifically including: Establish a space constraint function based on fluid cooling precipitation logic When the fluid homogenization temperature of the fluid inclusion When within the ore-forming temperature range, The value is 1; when the fluid homogenization temperature of the fluid inclusion is... When the temperature is less than or greater than the ore-forming temperature range, It exhibits Gaussian decay with temperature difference; Meanwhile, for geochemical indicators indicating mineralization anomalies within the study area, if the spatial temperature gradient modulus calculated based on the fluid temperature field model for a certain grid point is greater than the preset multiple of the cooling gradient at the non-altered rock of the deposit, it indicates that the grid point is on an effective hydrothermal migration path and is not subject to the aforementioned temperature difference attenuation limitation. The spatial constraint function value based on the fluid cooling precipitation logic... It also takes the value 1.

10. The deep mineral exploration method based on multi-source information fusion and overlay according to claim 1, characterized in that, Based on the revised mineralization potential, the deep mineral exploration target areas of the exploration region are identified, including: If the mineralization potential of a grid after correction is above the 90th percentile and spatially overlaps with the anomaly center of the near-ore halo zone, then the grid is a high-potential mineralization area. If the mineralization potential of a certain grid after correction is at the 75-90th percentile, then the grid is a mineralization potential zone; If the corrected mineralization potential of a grid is below the 75th percentile, then the grid is considered a low-potential mineralization zone.

Images