Metallization prediction method for synthesizing rare data based on mineralization space structure

Abstract

The invention relates to a mineralization prediction method for synthesizing rare data based on a mineralization space structure. The method comprises the following steps: collecting and preprocessing geological data based on a prediction demand; based on the geological background of the to-be-studied area, constructing a metallogenic prediction model, determining a prediction variable and creating a prediction layer; performing mineralization space structure analysis, synthesizing a preset number of prediction variables of rare data by combining the prediction variables of the to-be-studied area, and further generating a plurality of training data sets; adopting a plurality of training data sets to respectively train the constructed metallogenic prediction model so as to predict the metallogenic probability of the to-be-researched area; and performing performance evaluation on the metallogenic prediction model, and delineating a prospecting target region based on a prediction result of the metallogenic prediction model with relatively good performance. According to the metallogenic prediction method disclosed by the invention, the nuclear density analysis is carried out on the space distribution of the known ore deposit, so that a certain number of positive samples can be synthesized under the condition of considering the geological connotation, the training of the metallogenic prediction model is facilitated, and the prospecting prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the field of geological mineral resources, and in particular relates to a mineralization prediction method based on synthesizing rare data of mineralization spatial structure. Background Art

[0002] Machine learning / deep learning algorithms are widely used in mineralization prediction because they can better handle geological variables with complex nonlinear relationships and mine deep prospecting information. Among them, supervised machine learning / deep learning algorithms are guided by similarity analogy theory and predict undiscovered deposits by mining the correlation between geological data and the locations of discovered deposits.

[0003] In the process of mineralization prediction, the study area is usually divided into a series of grids of the same size. Grid cells containing minerals are often selected as positive samples, while grid cells without minerals are often selected as negative samples. In this way, mineralization prediction can be regarded as a binary classification problem. However, since the mineralization process is a rare event, the number of known ore deposits (i.e., mineralization units) in the study area is usually small. Therefore, when using supervised machine learning / deep learning to carry out mineralization prediction, there is a challenge of insufficient number of training samples (especially positive samples available for training), which leads to insufficient model learning ability and poor generalization ability, which in turn affects the prediction accuracy.

[0004] At present, in supervised machine learning / deep learning, "synthesizing" a certain number of positive samples is one of the main ways to make up for the lack of positive samples. In the synthetic positive sample technology, SMOTE (Synthetic Minority Over-sampling Technique), image affine transformation (rotation, scaling, displacement), image processing (such as contrast transformation) and adding random noise are more commonly used. However, mineralization is a singular process in which a large number of mineralizing materials are enriched in a relatively small space, and mineralization itself has spatial heterogeneity. At the same time, since the formation of ore deposits is often controlled by specific geological factors (such as strata, structures, etc.), mineralization often has a certain directionality. The spatial heterogeneity and directionality of mineralization indicate that it has a specific spatial structure. Although the SMOTE method can synthesize a certain number of positive samples in the feature space, it only considers the properties of the ore-bearing unit itself, but ignores the spatial relationship and attribute characteristics between the ore-bearing unit and its neighboring units, that is, it does not consider the spatial structure of mineralization. Image radiometric transformation, image processing and other methods take into account the spatial structure of mineralization because the selected training samples are images (i.e., the combination of mineral-bearing units and their adjacent units). However, these transformations may cause the geological connotation of the original data to change, thus affecting the prediction results. At the same time, although the method of adding noise to the training images can synthesize a large number of positive samples to a certain extent and take into account the spatial structure of mineralization, it will also increase uncertainty due to the addition of noise. Therefore, how to synthesize rare positive samples while taking into account the geological connotation of the mineralized structure has become a key technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of the defects existing in the prior art, the purpose of the present invention is to provide a mineralization prediction method based on synthesizing rare data based on the spatial structure of mineralization, which can synthesize a preset number of positive samples while taking into account the geological connotation, thereby improving the generalization ability and prediction accuracy of the mineralization prediction model, thereby improving the mineralization prediction accuracy and reducing exploration risks and costs.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a mineralization prediction method based on synthesizing rare data of mineralization spatial structure, comprising the following steps:

[0007] S1. Collect and preprocess geological data based on prediction needs;

[0008] S2. Based on the geological background of the area to be studied, a mineralization prediction model is constructed, prediction variables are determined, and prediction layers are created;

[0009] S3, perform mineralization spatial structure analysis, combine the prediction variables of the area to be studied to synthesize a preset number of rare data prediction variables, and then generate several training data sets;

[0010] S4. Use several training data sets to train the constructed mineralization prediction models respectively, predict the mineralization probability of the study area based on the created prediction layers, evaluate the performance of the mineralization prediction models, and delineate the prospecting target areas based on the prediction results of the mineralization prediction models with better performance.

[0011] Furthermore, the prediction requirements in step S1 include mineralization prediction accuracy, predicted mineral species, and required ore deposit type.

[0012] Furthermore, the geological data collected in step S1 include geological maps of the area to be studied, mineral distribution maps, geochemical data, geophysical data, etc.

[0013] Furthermore, step S2 includes analyzing the geological background of the area to be studied, summarizing the mineralization geological conditions, conducting research on typical mineral deposits, combining the analysis of the mineralization laws of the study area by predecessors, sorting out the key geological factors that control the formation of mineral deposits and important prospecting signs, and thus constructing a mineralization prediction model.

[0014] Furthermore, in step S2, based on the preprocessed geological data, prediction factors that can indicate key geological elements and important prospecting signs are used as prediction variables.

[0015] Further, the prediction variables in step S2 include Carboniferous-Permian strata, NE and NW-trending structures, Yanshanian intrusive bodies, aeromagnetic anomalies and geochemical anomalies related to the prediction elements. Step S2 includes dividing the area to be studied into a series of grid units of the same size according to the prediction accuracy, performing distance analysis on the ore-controlling strata, ore-controlling structures and ore-controlling rock bodies in the study area, and obtaining corresponding prediction variable layers respectively; performing factor analysis on geochemical data, extracting factor scores of geochemical element combinations related to mineralization, and performing local singularity analysis to obtain prediction variable layers of geochemical anomalies related to mineralization; performing local singularity analysis on aeromagnetic data to obtain prediction variable layers indicating aeromagnetic anomalies, and storing information of each layer in the grid unit.

[0016] Further, step S3 includes the following sub-steps:

[0017] S31, determining the number of synthetic training samples and the corresponding analysis window size according to the needs;

[0018] S32, performing a kernel density analysis on the first known mineral deposit in the study area to obtain a kernel density of the first known mineral deposit;

[0019] S33, taking the location of the unit where the first known mineral deposit is located as the center, performing weighted summation of the normalized weights on the original first prediction variable value of the prediction variable layer in the analysis window of a preset size and the kernel density of the first known mineral deposit, and taking the weighted summation result as the final first prediction variable at the location of the first known mineral deposit, that is, the first prediction variable of the synthetic rare data, replacing the first prediction variable with the Nth prediction variable, and performing weighted summation in the same way until all the prediction variables of the synthetic rare data are obtained;

[0020] S34, replacing the first known mineral deposit in steps S32-S33 with the Nth known mineral deposit, repeating steps S32-S33 to calculate the prediction variables of the rare data at the locations of other known mineral deposits, and combining the obtained prediction variables of the rare data of each known mineral deposit to form a positive sample of the expected multiple;

[0021] S35. Randomly select the same number of units as positive samples in areas with poor mineralization conditions as negative samples, and obtain a training data set of a preset sample size based on the generated positive and negative samples.

[0022] Further, in step S33, the prediction variables of rare data are synthesized according to the following formula:

[0023]

[0024] Among them, i is the analysis window size, i is an odd number, Y l u is the lth predictor variable Y at position u in the analysis window l The value of K u is the ore deposit core density K value at position u in the analysis window.

[0025] Furthermore, in step S35, negative samples are randomly selected a preset number of times from the negative samples having the same number as the positive samples, and the selected negative samples are respectively combined with the positive samples to obtain different training data sets.

[0026] Furthermore, in step S4, the prediction results of the mineralization prediction models with better performance are averaged, which is the mineralization probability map of the target mineral species in the study area. The mineralization probability is classified using the natural break point method, and the high-value area of ​​mineralization probability is circled as the prospecting target area.

[0027] The effect of the present invention is that compared with the positive sample synthesis method in the prior art, the mineralization prediction method disclosed in the present invention that synthesizes rare data based on the mineralization spatial structure can synthesize a certain number of positive samples while taking into account the mineralization spatial structure, thereby improving the prediction accuracy of the mineralization prediction model, improving the generalization ability of the mineralization prediction model, thereby reducing exploration risks, and providing a new method and idea for mineralization prediction of rare data. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The regional geological map of the Taqian-Zhuxi area described in Example 1 of the present invention;

[0029] Figure 2 This is a schematic diagram of kernel density estimation shown in a method for mineralization prediction based on synthesizing rare data of mineralization spatial structure according to the first embodiment of the present invention;

[0030] Figure 3 The neighborhood map (3×3) of the mineral deposit kernel density analysis based on pixel units shown in the method for mineralization prediction based on synthesizing rare data of mineralization spatial structure described in the first embodiment of the present invention;

[0031] Figure 4 A general ore deposit kernel density analysis neighborhood map (3×3) shown in a method for mineralization prediction based on synthesizing rare data of mineralization spatial structure according to the first embodiment of the present invention;

[0032] Figure 5 A general schematic diagram of synthetic rare data prediction variables shown in a mineralization prediction method based on synthetic rare data of mineralization spatial structure according to the first embodiment of the present invention;

[0033] Figure 6 This is a schematic diagram of the process of synthesizing rare data prediction variables shown in a method for mineralization prediction based on synthesizing rare data of mineralization spatial structure according to the first embodiment of the present invention;

[0034] Figure 7 The AUC values ​​of the prediction results of each model after the mineralization prediction models are trained using different training data sets in the mineralization prediction method based on synthesizing rare data of mineralization spatial structure described in Example 1 of the present invention;

[0035] Figure 8 The mineralization probability map of W-Cu polymetallic ore in the Taqian-Zhuxi area is generated by a mineralization prediction method based on synthesizing rare data of mineralization spatial structure as described in Example 1 of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0037] Embodiment 1

[0038] In this embodiment, the W-Cu polymetallic ore in the Taqian-Zhuxi area of ​​Jiangxi Province is taken as an example to illustrate a mineralization prediction method based on synthesizing rare data of mineralization spatial structure disclosed in the present invention.

[0039] The geotectonic structure of the Taqian-Zhuxi area is located in the Jiangxi section of the Jiangnan orogenic belt at the junction of the Yangtze Plate and the Cathaysia Plate, and on the northwest side of the deep fault in northeast Jiangxi. It is one of the most important Cu-Au-Pb-Zn-Ag polymetallic mineralization areas in South China. The strata in the area are relatively well exposed, with a typical binary structure, namely the basement composed of the Neoproterozoic Wannian Group and the Shuangqiaoshan Group and the sedimentary cover composed of the Devonian, Carboniferous, Permian, Triassic, Jurassic, Cretaceous, Paleogene and Quaternary. Affected by the NE-trending regional thrust nappe structure, the Neoproterozoic Wannian Group is superimposed on the Upper Triassic strata. The structures in the area mainly develop NE, NW and nearly EW-trending structures. The concealed rock bodies revealed by the drilling are mainly granite and granite porphyry, followed by granodiorite, diorite porphyry and lamprophyre. The Zhuxi super-large W-Cu deposit, the Taqian porphyry tungsten-molybdenum deposit, the moon-shaped vein Cu-Pb-Zn deposit, etc. have been discovered in the area, and the geological conditions for mineralization are very favorable. Therefore, it is of great significance to systematically carry out mineralization prediction and use machine learning algorithms to mine deep prospecting information for mineral exploration in the area.

[0040] This embodiment provides a method for mineralization prediction based on synthesizing rare data of mineralization spatial structure, the method comprising the following steps:

[0041] S1. Collect and preprocess geological data based on prediction needs

[0042] According to the requirements of mineralization prediction accuracy and the needs of actual prediction of mineral species and deposit types, systematically collect geological data in the geology, geochemistry, geophysics and remote sensing of the area to be studied. The scale of collecting geological data can be 1:50,000, 1:200,000, etc. Geological data preprocessing includes unifying the data coordinate system, sorting and cleaning the data, eliminating missing values ​​and replacing data below the detection limit.

[0043] like Figure 1 As shown, in this embodiment, a 1:50,000 regional geological map (including strata, structures and igneous rocks) of the Taqian-Zhuxi area, a mineral distribution map, a 1:50,000 stream sediment geochemical data (including 17 elements such as Au, Ag, Cu, Pb, Zn, W, Sn, Mo, Bi, Cd, Hg, Cr, Co, Ni, Sb, As, Ba) and a 1:50,000 aeromagnetic data were collected.

[0044] S2. Based on the geological background of the area to be studied, a mineralization prediction model and prediction variables are constructed, and a prediction variable layer is created.

[0045] Analyze the geological background of the area to be studied, systematically summarize the mineralization geological conditions, carry out research on typical mineral deposits, combine previous analysis of the mineralization laws of the area, sort out the key geological factors that control the formation of mineral deposits and important prospecting signs, and thus construct a mineralization prediction model.

[0046] According to the prediction accuracy, the study area is divided into a series of grid units of the same size. On this basis, according to the mineralization prediction model, the prediction variables are constructed based on the existing geological data. That is, based on the preprocessed geological data, the prediction factors that can indicate key geological elements and important prospecting signs are used as prediction variables.

[0047] The prediction variable layer is the display form of the prediction variable, which is expressed in the form of a layer. When performing model calculations, the variables need to be converted into variable layers that can be used for GIS calculations.

[0048] Distance analysis is used to analyze the ore-controlling strata, ore-controlling structures and ore-controlling rock bodies in the study area, and the corresponding prediction variable layers are formed respectively; factor analysis is performed on geochemical data to extract the factor scores of geochemical element combinations related to mineralization, and local singularity analysis is performed to obtain the prediction variable layer of geochemical anomalies related to mineralization; local singularity analysis is performed on aeromagnetic data to obtain the prediction variable layer indicating aeromagnetic anomalies. The information of each layer is stored in the grid unit.

[0049] Continuing from the above example, according to the analysis of the geological conditions and mineralization laws of the Taqian-Zhuxi area by predecessors, the main mineralization strata in the area are mainly carbonate rocks of the Carboniferous-Permian system. The strata contain high contents of mineralization elements such as W, Cu, Pb, and Zn, which provide an important material basis for mineralization. The main structure in the area is the NE-trending thrust-nappe structure. The thrust-nappe structure can cause interlayer thrusting and slippage of the mineralization strata, thus forming a structural weak zone, which is an important space for magma emplacement and ore body occurrence. The rock bodies closely related to mineralization in the area are mainly peraluminous granites, and the mineralization age is mainly concentrated around 150Ma. S and Pb isotope analysis shows that the mineralization materials mainly come from crustal remelting granites, surrounding rocks, etc. According to the exploration experience in the area, the magnetic rocks in the area are mainly various types of skarns. The positive and negative magnetic anomalies distributed in the carbonaceous tectonic rock slices or their edges in the late Paleozoic are favorable locations for prospecting skarn-type deposits and are one of the important signs for prospecting. At the same time, surface W, Cu, Pb, Zn and other anomalies have a good indication of concealed ore bodies and are important geochemical prospecting signs. As shown in Table 1, on this basis, a mineralization prediction model and prediction variables related to W-Cu polymetallic mineralization were preliminarily constructed for the region.

[0050] Table 1 Ore-forming prediction model related to W-Cu polymetallic mineralization in the Taqian-Zhuxi area

[0051]

[0052] According to Table 1, the mineralization prediction model has seven prediction variables, namely, Carboniferous-Permian strata Y 1 , NE, NW-oriented structure Y 2 , Yanshanian intrusive body Y 3 、Aeromagnetic anomaly Y 4 , geochemical anomaly Y related to W 5 , Geochemical anomaly Y related to Pb-Zn 6 and geochemical anomalies Y related to Cu 7 .

[0053] According to the prediction accuracy requirements, the study area is divided into a series of grid cells of the same size as the benchmark for the range size and grid size of each prediction layer. According to the above-mentioned mineralization prediction model, the Carboniferous-Permian salt rock strata are screened out in the 1:50,000 geological map, and the prediction variable Carboniferous-Permian Y is constructed using the inverse distance. 1 Distance prediction layer. Screen out the prediction variables NE and NW structural Y in the 1:50,000 geological map 2 , using the inverse distance, a prediction layer of distance to the fracture is constructed. Similarly, the prediction variable Yanshanian intermediate acidic intrusive rock Y is selected from the 1:50,000 geological map. 3, using the inverse ratio of distance, a prediction layer of the distance to the Yanshanian acidic intrusive rock mass is constructed. The local singularity analysis of the 1:50,000 aeromagnetic pole is performed to obtain the predicted variable aeromagnetic anomaly Y 4 The layer where it is located. Factor analysis was performed on 17 geochemical elements to screen factors related to W-Cu polymetallic mineralization: F1 (W-Mo-As), F2 (Pb-Zn-Cd) and F3 (Cu-Co-Ba), which respectively represent the geochemical element combinations related to W, Pb-Zn and Cu mineralization. Local singularity analysis was performed on F1, F2 and F3 respectively, and the geochemical anomaly Y related to W was obtained. 5 , Geochemical anomaly Y related to Pb-Zn 6 and geochemical anomalies Y related to Cu 7 The prediction layer.

[0054] S3. Perform mineralization spatial structure analysis, combine the prediction variables of the area to be studied to synthesize a preset number of prediction variables of rare data, and then generate several training data sets.

[0055] Since ore deposits are formed randomly by mineralization, mineralization is often the result of multi-factor coupling and critical transformation, so they have chaotic effects, and their spatial distribution has both randomness and statistical regularity. Therefore, the spatial distribution pattern of ore deposits can be studied as a realization of random processes using spatial point pattern analysis. Kernel density analysis is an effective method for analyzing the spatial heterogeneity of spatial point patterns. Figure 2 As shown, kernel density analysis estimates the local intensity of the spatial point process at each position in space under a given bandwidth (i.e., search radius).

[0056] Commonly used kernel functions include Gaussian kernel function, binary quadratic kernel function, etc. Taking the binary quadratic kernel function as an example, the kernel function can be expressed as:

[0057]

[0058] Where K(u) is the kernel density estimate at u, ||ux i || is u to x i The distance at which the kernel density is calculated is h, and h is the bandwidth (or search radius). The choice of bandwidth h has a great influence on the estimation of kernel density. When the bandwidth h is large, the kernel function is smoother. At this time, the kernel function can better reflect the changes in the region and ignore the variability of the local scale; when the bandwidth h is small, the kernel function can better reflect the variability of the local scale. The Silverman rule of thumb is often used to determine the bandwidth. According to this rule, the optimal bandwidth can be expressed by the following formula:

[0059]

[0060] Where IQR is the interquartile range, σ is the standard deviation of the analysis sample, and N is the total number of spatial point processes.

[0061] The kernel density estimation of the spatial distribution of known mineral deposits can reflect the spatial structure of mineralization (such as the spatial heterogeneity and directionality of mineralization) to a certain extent: on the one hand, the closer the location of the known mineral deposit is, the higher the kernel density and the stronger the mineralization; on the other hand, since the distribution of mineral deposits is usually controlled by specific geological factors, their spatial distribution has a certain directionality. The kernel density of the mineral deposit also reflects the directionality of the distribution of the mineral deposit to a certain extent, and also reflects the directionality of the spatial distribution of mineralization. Figure 3 As shown in the figure, the mineralization prediction based on pixel units can be regarded as a special case where the kernel density of the ore-bearing unit position is 1, while the kernel density of other adjacent units is 0, that is, a special case where the bandwidth infinitely approaches 0. The kernel density of the general case is as follows: Figure 4 shown.

[0062] According to the number of positive samples to be synthesized, the size of the analysis window to be used for synthesizing positive samples is determined with the mineral-bearing unit as the center. For example, if you need to synthesize 1 times the number of positive samples, select an analysis window of 3×3; if you need to synthesize 2 times the number of positive samples, select analysis windows of 3×3 and 5×5 respectively; if you need to synthesize 3 times the number of positive samples, select analysis windows of 3×3, 5×5 and 7×7 respectively. That is, when you need to synthesize n times the number of positive samples, the number of analysis windows required for synthesizing positive samples is also n, and the sizes of analysis windows required for synthesizing positive samples are 3×3, 5×5, ···, (2m+1)×(2m+1), ···, (2n+1)×(2n+1) (m is a natural number from 1 to n).

[0063] Assuming the analysis window size is i (i is an odd number), there are i predictor variable layers in the window. 2 Pixel units. l u is the lth predictor variable Y at position u in the analysis window l The value of K u is the value of the mineral deposit kernel density K at position u in the analysis window. Then under the condition of this window size, the predicted variable in the synthetic rare data The values ​​are:

[0064]

[0065] That is, combined with the mineral deposit kernel density, the values ​​within the preset window range in the prediction variable layer are weighted summed to obtain the synthetic prediction variable value, which is the prediction variable in the rare data. The composite value of .

[0066] like Figure 5 As shown, similarly, the synthetic values ​​of the prediction variables of other rare data under the same size window conditions can be obtained. Repeat the above steps, perform window analysis on the location of each mineral-bearing unit, and obtain the prediction variables of all rare data at the location of the mineral-bearing unit. Combine the prediction variables of rare data at all mineral-bearing units under the same size analysis window to generate 1 times the number of positive samples. Each time an analysis window of a certain size is used for analysis, 1 times the number of positive samples can be obtained. Therefore, the number of positive samples required by the user can be synthesized as needed. At the same time, by randomly selecting the same number of negative samples as the positive samples in areas with poor mineralization geological conditions, they together with the positive samples constitute a training data set.

[0067] Perform kernel density analysis on known mineral deposits to obtain the kernel density map of the mineral deposits. Centered on the location of the unit where the mineral deposit is located, select the corresponding analysis window according to the number of synthetic training samples. If you need to synthesize 1 times the number of positive samples, select a 3×3 analysis window.

[0068] like Figure 6 As shown, taking the synthesis method of the prediction variable Y as an example, the synthesis process of the sample is briefly described. The yellow grid is the prediction layer Y, and the pixel value of each small grid represents the value of the prediction variable Y at that location. The light blue grid is the kernel density of the ore deposit. The red in the layer represents the grid where the ore deposit is located, based on the high-value area of ​​the kernel density. The kernel density of the ore deposit can not only reflect the strength of the mineralization, but also the directionality of the mineralization (as shown in the figure, the NE-direction kernel density value is higher, indicating that the mineralization is distributed in the NE direction), that is, it characterizes the spatial structure of the mineralization. In the sample synthesis process, a 3×3 window is used for analysis. The new prediction variable value is the weighted sum of the original prediction variable value in the analysis window and the normalized weight of the ore deposit kernel density, that is, the prediction variable of rare data, which is expressed by the formula:

[0069]

[0070] Similarly, according to this method, the composite value of the predicted variables of other rare data at the location of the deposit under the size analysis window can be calculated. Similarly, the composite values ​​of the predicted variables of each rare data at the location of other deposits are calculated, and finally combined to form 1 times the number of positive samples.

[0071] In the embodiment of the present invention, the area to be studied contains 13 known mineral deposits, and positive samples of 1 times (13), 2 times (26) and 3 times (39) are synthesized for illustration. The original value of each prediction variable at each mineral-bearing unit constitutes 13 positive samples (Novel1); the original value and the synthesized 1 times positive sample constitute 26 positive samples (Novel2); the original value and the synthesized 2 times positive sample constitute 39 positive samples (Novel3); the original value and the synthesized 3 times positive sample constitute 52 positive samples (Novel4).

[0072] Then, according to the principle of negative sample screening (i.e. randomly screening units with the same number of positive samples in areas with poor mineralization conditions), a corresponding number of 52 negative samples were screened. Among the 52 negative samples, 13 negative samples were randomly selected, repeated 4 times, and combined with Novel1 respectively, to obtain 4 training data sets with a total number of 26 samples (13 positive samples + 13 negative samples), namely Novel11, Novel12, Novel13 and Novel14; among the 52 negative samples, 26 negative samples were randomly selected, repeated 4 times, and combined with Novel2 respectively, to obtain 4 training data sets with a total number of 52 samples (26 positive samples + 26 negative samples), namely Novel21 , Novel22, Novel23 and Novel24; among the 52 negative samples, 39 negative samples were randomly selected, repeated 4 times, and combined with Novel3 respectively to obtain 4 training data sets with a total sample size of 78 (39 positive samples + 39 negative samples), namely Novel31, Novel32, Novel33 and Novel34; the 52 negative samples were combined with Novel4 to obtain the training data set Novel41 with a total sample size of 104 (52 positive samples + 52 negative samples).

[0073] S4. Based on the created prediction layer and the generated training data sets, mineralization prediction is performed, and the mineralization prediction model is evaluated based on the prediction results.

[0074] The supervised machine learning algorithm is used to learn the training data set to train the constructed mineralization prediction model, and the trained mineralization prediction models are used to carry out predictions for the entire study area, calculate the mineralization probability value of each grid unit in the study area, delineate the mineralization prospective area, and use the ROC curve and AUC value to evaluate the trained mineralization prediction models.

[0075] In this embodiment, supervised machine learning algorithms include decision tree, support vector machine, random forest, adaptive boosting and other machine learning algorithms.

[0076] Following the above example, in this embodiment, the random forest algorithm is used to learn Novel11, Novel12, Novel13 and Novel14 respectively. Figure 7 As shown in the figure, four mineral-bearing probability maps of the study area are obtained. The ROC curves of the four mineral-bearing probability maps are drawn respectively, and their average AUC values ​​are calculated. Repeat the above steps to calculate the average AUC values ​​of the four types of training data sets with different numbers. According to the model evaluation results, when the number of synthesized positive samples is doubled, the average AUC value is the largest, which is 0.997. Figure 8 As shown in the figure, the average of the prediction results obtained from this number of training data sets is the W-Cu mineralization probability map in the Taqian-Zhuxi area. On this basis, the mineralization probability is classified using the natural break point method, and the W-Cu polymetallic mineral prospecting target area can be delineated according to the high value area of ​​mineralization probability.

[0077] In this embodiment, only the use of the random forest algorithm to learn Novel11, Novel12, Novel13 and Novel14 is used as an example. In fact, the random forest algorithm can also be used to learn Novel21, Novel22, Novel23 and Novel24, and there is no limitation to this.

[0078] It can be seen from the above embodiments that the present invention discloses a mineralization prediction method based on synthesizing rare data from the spatial structure of mineralization. Based on the geological background of the area to be studied, a mineralization prediction model and prediction variables are constructed, and the prediction layer is preprocessed; the mineralization spatial structure analysis is performed, and the prediction variables of the rare data are synthesized in combination with the prediction variables of the area to be studied, and then a number of training data sets are generated. It is possible to synthesize a preset number of positive samples while taking into account the geological connotation, thereby improving the generalization ability and prediction accuracy of the mineralization prediction model, thereby improving the accuracy and efficiency of mineralization prediction, reducing exploration risks and exploration costs, and providing important basis and guidance for mineral exploration.

[0079] The method described in the present invention is not limited to the embodiments described in the specific implementation manner. Those skilled in the art may derive other implementation manners based on the technical solution of the present invention, which also fall within the technical innovation scope of the present invention.

Claims

1. A mineralization prediction method based on synthesizing rare data of mineralization spatial structure, comprising the following steps: S1. Collect and preprocess geological data based on prediction needs; S2. Based on the geological background of the area to be studied, a mineralization prediction model is constructed, prediction variables are determined and prediction layers are created; S3, perform mineralization spatial structure analysis, combine the prediction variables of the area to be studied to synthesize a preset number of rare data prediction variables, and then generate several training data sets; S4. Use several training data sets to train the constructed mineralization prediction models respectively. Based on the created prediction layers, the mineralization probability of the study area is predicted, and the performance of the mineralization prediction model is evaluated. The prospecting target area is delineated based on the prediction results of the mineralization prediction model with better performance.

2. A mineralization prediction method based on synthesizing rare data of mineralization spatial structure as claimed in claim 1, characterized in that: The prediction requirements in step S1 include the mineralization prediction accuracy, the predicted mineral species and the required ore deposit type.

3. A mineralization prediction method based on synthesizing rare data of mineralization spatial structure as claimed in claim 1, characterized in that: The geological data collected in step S1 include geological maps of the area to be studied, mineral distribution maps, geochemical data, and geophysical data.

4. A mineralization prediction method based on synthesizing rare data of mineralization spatial structure as claimed in claim 1, characterized in that: Step S2 includes analyzing the geological background of the area to be studied, summarizing the mineralization geological conditions, conducting research on typical mineral deposits, combining the analysis of the mineralization laws of the study area by predecessors, sorting out the key geological factors that control the formation of mineral deposits and important prospecting signs, and thus constructing a mineralization prediction model.

5. A method for mineralization prediction based on synthesizing rare data of mineralization spatial structure as claimed in claim 1, characterized in that: In step S2, based on the preprocessed geological data, prediction factors that can indicate key geological elements and important prospecting signs are used as prediction variables.

6. A method for mineralization prediction based on synthesizing rare data of mineralization spatial structure as claimed in claim 5, characterized in that: The prediction variables in step S2 include Carboniferous-Permian strata, NE and NW-trending structures, Yanshanian intrusive bodies, aeromagnetic anomalies and geochemical anomalies related to the prediction elements. Step S2 includes dividing the area to be studied into a series of grid units of the same size according to the prediction accuracy, performing distance analysis on the ore-controlling strata, ore-controlling structures and ore-controlling rock bodies in the study area, and obtaining corresponding prediction variable layers respectively; performing factor analysis on geochemical data, extracting factor scores of geochemical element combinations related to mineralization, and performing local singularity analysis to obtain prediction variable layers of geochemical anomalies related to mineralization; performing local singularity analysis on aeromagnetic data to obtain prediction variable layers indicating aeromagnetic anomalies, and storing information of each layer in the grid unit.

7. A method for mineralization prediction based on synthesizing rare data of mineralization spatial structure as claimed in claim 1, characterized in that: Step S3 includes the following sub-steps: S31, determining a corresponding analysis window size according to the number of training samples that need to be synthesized; S32, performing a kernel density analysis on the first known mineral deposit in the study area to obtain a kernel density of the first known mineral deposit; S33, taking the location of the unit where the first known mineral deposit is located as the center, performing weighted summation of the normalized weights on the original first prediction variable value of the prediction variable layer in the analysis window of a preset size and the kernel density of the first known mineral deposit, and taking the weighted summation result as the final first prediction variable at the location of the first known mineral deposit, that is, the first prediction variable of the synthetic rare data, replacing the first prediction variable with the Nth prediction variable, and performing weighted summation in the same way until all the prediction variables of the synthetic rare data are obtained; S34, replacing the first known mineral deposit in steps S32-S33 with the Nth known mineral deposit, repeating steps S32-S33 to calculate the prediction variables of the rare data at the locations of other known mineral deposits, and combining the obtained prediction variables of the rare data of each known mineral deposit to form a positive sample of the expected multiple; S35. Randomly select the same number of units as positive samples in areas with poor mineralization conditions as negative samples, and obtain a training data set of a preset sample size based on the generated positive and negative samples.

8. A method for mineralization prediction based on synthesizing rare data of mineralization spatial structure as claimed in claim 1, characterized in that: In step S33, the prediction variables of rare data are synthesized according to the following formula: Among them, i is the analysis window size, i is an odd number, Y l u is the lth predictor variable Y at position u in the analysis window l The value of K u is the ore deposit core density K value at position u in the analysis window.

9. A method for mineralization prediction based on synthesizing rare data of mineralization spatial structure according to claim 1, characterized in that: In step S35, negative samples are randomly selected a preset number of times from the negative samples having the same number as the positive samples, and the selected negative samples are respectively combined with the positive samples to obtain different training data sets.

10. A method for mineralization prediction based on synthesizing rare data of mineralization spatial structure according to claim 1, characterized in that: In step S4, the prediction results of the mineralization prediction models with better performance are averaged, which is the mineralization probability map of the target mineral species in the study area. The mineralization probability is classified using the natural break point method, and the high-value area of ​​mineralization probability is circled as the prospecting target area.

Images