Improved gold ore prospecting prediction method based on deep residual network

Abstract

The invention discloses an improved deep residual network gold ore prospecting prediction method, and relates to the field of gold ore prospecting prediction. The method comprises the following steps: acquiring gold ore related data in a research area, and selecting various element data as first sample data; acquiring four load factors by adopting an adaptive kernel density estimation method and a factor analysis method; taking content data of various geological elements containing four load factors in the research area as second sample data; constructing a regular grid for the second sample data by adopting an inverse distance weighted interpolation method, and fusing the data in the regular grid to form a multi-dimensional data sample; processing the multi-dimensional data sample through a step translation method and a random discarding method to obtain a sample data subset; constructing an improved deep residual network model obtained by integrating three deep residual network sub-models; and respectively predicting the sample data subsets by adopting an improved deep residual network gold mine prospecting prediction model to obtain a potential distribution area of the gold mine.

Description

Technical Field

[0001] The invention relates to the technical field of mineral exploration, and in particular to an improved deep residual network gold mine prospecting prediction method. Background Art

[0002] At present, with the development of emerging technologies, machine learning technology is being used more and more widely in the geological industry. The introduction of machine learning algorithms makes big data intelligent prospecting and prediction possible. The core issue of prospecting and prediction is to analyze and judge the nonlinear relationship between multi-source data such as geology, minerals, remote sensing, geophysics, geochemistry, and the distribution of mineral resources. By extracting the characteristic information of a large amount of data and integrating it comprehensively, machine learning methods can effectively explore the relationship between multi-source information and mineral deposits, deeply reveal the internal logic of mineralization, and effectively improve the accuracy of prospecting. Machine learning, especially deep learning methods, have unique advantages in geological big data feature extraction, information fusion, prospecting information mining, classification, and prediction.

[0003] In the existing technology, deep learning, as a kind of machine learning method, has gradually become the main method of mineral prospecting prediction with its strong complex image processing and spatial relationship mining capabilities. Convolutional neural network (CNN), as one of the deep learning models, has been widely used in mineral prospecting prediction. Many scholars have used convolutional neural network to build mineral prospecting prediction models, proving that convolutional neural network (CNN) is an effective mineral prospecting prediction method.

[0004] However, in the field of prospecting prediction, the prospecting prediction method based on convolutional neural network used in the prior art cannot guarantee the learning and generalization ability of convolutional neural network for different data features due to the essential characteristics of small amount of geological data samples and low data quality in the study area, and thus cannot learn the complex features in geological data in a deeper and more comprehensive way. In addition, the prospecting prediction model used in the prior art is relatively single, and different models have different judgments on different data samples, so it is impossible to avoid biased dependence on different models, which makes it difficult to improve the accuracy of the model.

[0005] Therefore, how to enhance the model's ability to recognize and capture diverse features on the basis of solving the problems of small sample data volume and low sample data quality of geological elements in mineral exploration prediction, so as to further improve the prediction accuracy is a problem to be solved by existing technologies. Summary of the invention

[0006] Based on this, it is necessary to provide an improved deep residual network gold mine prospecting prediction method to address the above technical problems.

[0007] The present invention adopts the following technical solutions: The present invention provides a prospecting prediction method, comprising: Obtaining content data of multiple geological elements in soil sediment samples in the study area, and using the content data of multiple geological elements as first sample data; Adaptive kernel density estimation method and factor analysis method were used to statistically analyze the first sample data to obtain four loading factors; and the content data of multiple geological elements containing four loading factors in the study area were used as the second sample data; By using an inverse distance weighted interpolation method, a regular grid is constructed for the second sample data, and the second sample data in the regular grid is fused to form multi-dimensional sample data; Through the step-size shift method and random discard method, the multi-dimensional sample data is expanded, enhanced and divided to obtain the sample data subset; By integrating three deep residual network models, an improved deep residual network prospecting prediction model is obtained; wherein the three sub-models of the improved deep residual network prospecting prediction model include: two deep residual network models containing 18 residual units and one deep residual network model containing 34 residual units; The improved deep residual network gold prospecting prediction model is used to predict the sample data subset and obtain the final prediction value to characterize the potential distribution area of ​​gold mines.

[0008] Preferably, the multiple elements in the soil sediment samples of the study area include: Au, Ag, As, Bi, Cu, Mo, Pb, Sb, W and Zn.

[0009] Preferably, the acquisition of the four load factors specifically includes: Adopting the adaptive kernel density estimation method, the bandwidth of the kernel function is adjusted to adapt to the local density of the first sample data, the sparse area and dense area of ​​the first sample data are estimated, and the element distribution of the first sample data in the study area is statistically analyzed to determine the metallogenic conditions of the elements in the study area; Using factor analysis method, we extracted the main factor with eigenvalue greater than 1 and non-single factor factors to obtain the loading factor matrix; The load factor values ​​in the load factor matrix represent the correlation between the factors in the first sample data and the potential load factors, including: When the loading factor of the factor in the first sample data and the latent factor is close to 1, it means that the two are highly correlated; when the loading value is close to 0, it means that there is no significant correlation between the two; when the loading value is close to -1, it means that there is a negative correlation between the two; Set the factor load threshold, divide the factors of the first sample data according to the threshold, and obtain 4 groups of load factors, including: F1 load factor, including Ag, Bi, Cu and Pb elements; F2 load factor, including Au, Mo and W elements; F3 load factor, including As and Sb elements; The F4 load factor is a load factor constructed by dividing the high temperature element of the F3 load factor by the high temperature element of the F1 load factor, and the F4 load factor is used to enhance the degree of change of the F3 low temperature element combination.

[0010] Preferably, the second sample data includes: 10 types of soil sediment geochemical exploration data, digital elevation model data, structural data and four load factors, a total of 16 types of data.

[0011] Preferably, obtaining the three sample data subsets comprises the following steps: The inverse distance weighted interpolation method is used to construct a regular grid for the second sample data. The construction formula is as follows: ; in, Indicates at point The estimated value of is the measured value at this point, is the number of known points used for interpolation, is the distance between the estimated value and the known point, is the weight index; Fusing the second sample data in the regular grid to obtain a multidimensional data set; The step-size translation method is adopted, and a 16×16 window is selected as the sample clipping size of the multidimensional data set to form 256 grid units. By translating the window, the mineral-bearing points are sequentially placed in each grid unit and clipped. Each mineral-bearing point can obtain 256 sample data, and n mineral-bearing points can obtain 256×n sample data, thus obtaining an expanded sample data set. The random discard method is adopted. While ensuring that the geological meaning remains unchanged, some data points in the expanded sample data set are randomly selected as label samples and their values ​​are replaced with 0. Noise is added to the label samples, and a 16×16 window is selected to crop the label samples. The specified number of sample enhancement data is obtained repeatedly, and the obtained sample enhancement data is added to the expanded sample data set. Sampling with replacement is used to obtain three sample data subsets with the same number.

[0012] Preferably, the step of constructing a regular grid for the second sample data using an inverse distance weighted interpolation method specifically includes: The regular grid size is set to 20m×20m, and the total number of grid cells is 103×133=13699; The digital elevation model data in the second sample data is resampled using data with a resolution of 12.5 meters to conform to the grid size.

[0013] Preferably, the method of translating the window so that the mineral-bearing points fall into each grid unit in turn and perform cropping specifically includes: during the window translation process, it is stipulated that there is only one mineral-bearing point in a sample, and cropping is skipped for the case where more than one mineral-bearing point is included.

[0014] Preferably, the improved deep residual network prospecting prediction model is obtained by comprising the following steps: Define the residual unit: ; in, Indicates l The output of the residual unit, Indicates l The input of the residual unit is Indicates l+ The output of 1 residual unit, F () is the residual function, represents the weight coefficient of the residual function, f () is the activation function, represents the target mapping function; The residual units are accumulated to obtain the residual block group, and the formula is as follows: ; in, Indicates l The output of shallow units and the cumulative sum of a series of residual functions, Represents the weight coefficient of the residual function; Construct a deep residual network including an initial convolutional layer, a residual block group, a global average pooling layer, and a fully connected layer, and obtain a deep residual network prediction model by training it. By integrating three deep residual network prediction models through voting, an improved deep residual network prospecting prediction model is obtained; The voting method is specifically: averaging the output results of the three deep residual network prediction models.

[0015] Preferably, the deep residual network prediction model specifically includes: The initial convolutional layer receives the sample subset data, performs convolution operations on the data and performs batch normalization, and outputs the basic features in the data through the activation function; The residual block group receives the basic features output by the initial convolutional layer, obtains the residual features by performing residual learning on the basic features, and adds the basic features to the residual features by using a skip link to obtain the target mapping features; The global average pooling layer receives the target mapping features output by the residual block group, performs a simple average calculation on the target mapping features, and obtains the pooling features; The fully connected layer receives the pooled features and outputs the model prediction value through feature integration and complex mapping learning.

[0016] Preferably, obtaining the final predicted value comprises the following steps: The three sample subset data are input into the three sub-models of the improved deep residual network gold prospecting prediction model respectively; The three sub-models predict the three sample subset data respectively to obtain the predicted value of each sub-model; The predicted values ​​of all sub-models are added together and averaged to obtain the final predicted value, which is used to characterize the potential distribution area of ​​gold mines.

[0017] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects: The present invention first obtains four main load factors based on the original geological element data through adaptive kernel estimation function and factor analysis method, and obtains sample data through integration. Then, the sample data is expanded by combining step-size shift method and random discarding method, and the features of the sample data are enhanced. A sample data subset with equal sample data volume is obtained through sampling with replacement. This not only solves the problems of insufficient geological data samples and low data quality in mineral prospecting prediction, but also improves the diversity of sample data while ensuring that each sample subset contains the overall characteristics of the study area.

[0018] Secondly, the improved deep residual network prediction method formed by integrating multiple deep residual network models in the present invention solves the problems of sample data quantity and data quality, so that the model provided by the method of the present invention can learn different data features, thereby improving the model's recognition ability and feature capture ability for diversified data.

[0019] In summary, the present invention uses the step-size shifting method and the random discarding method to expand and enhance the sample data, and combines the improved deep residual network gold prospecting prediction model to perform gold prospecting prediction on the sample data. It not only solves the problems of small sample size and low quality of geological element data, but also the improved deep residual network gold prospecting prediction model integrates multiple deep residual network sub-models. While avoiding the biased dependence of different models, it can also comprehensively learn and capture the diversity characteristics of geological element data in the study area, improve the model's recognition ability for diversified data, enhance the model's learning and generalization capabilities for diversified data, and improve the prediction accuracy of gold prospecting prediction, thereby providing scientific guidance for the practical application of prospecting in the study area. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 A schematic diagram of a flow chart of a mineral prospecting prediction method provided by the present invention; Figure 2 An adaptive kernel density estimation diagram of a mineral prospecting prediction method provided by the present invention; Figure 3 A multi-dimensional sample data point density map of a mineral prospecting prediction method provided by the present invention; Figure 4 A schematic diagram of step-length translation of a mineral prospecting prediction method provided by the present invention; Figure 5 A schematic diagram of random discarding of a mineral prospecting prediction method provided by the present invention; Figure 6 A schematic diagram of the structure of an improved deep residual network model of a mineral prospecting prediction method provided by the present invention; Figure 7 A comparison chart of the loss function and training accuracy of a CNN model, a ResNet 18 model, and a ResNet 34 model of a prospecting prediction method provided by the present invention; Figure 8 This is a comparison chart of the accuracy verification of all models of a mineral prospecting prediction method provided by the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0023] Figure 1 The present invention is a schematic diagram of an improved deep residual network gold ore prospecting prediction method, which specifically includes the following steps: S101: Obtaining content data of multiple geological elements in soil sediment samples in a study area, and using the content data of multiple geological elements as first sample data; The study area is the Kongquegou area, which is located in the Elbin Mountain-Kumish residual basin of the Southern Tianshan-Luoyachu Mountain arc basin system, and belongs to the Elbin mineralization sub-belt of the northern margin of the Tarim plate. The geological structure in the area is complex. Affected by the northwest-trending Kawabulak regional fault, the gold deposits are mainly distributed in the northwest, and its secondary faults control the spatial distribution of the ore body; The source of the first sample data is the data of 10 elements including Au, Ag, As, Bi, Cu, Mo, Pb, Sb, W and Zn from 474 soil sediment samples in the study area; S102: using an adaptive kernel density estimation method and a factor analysis method to statistically analyze the first sample data elements to obtain four load factors; and using the content data of multiple geological elements containing the four load factors in the study area as the second sample data; An adaptive kernel density estimation method is used to perform skewness analysis on the first sample data, specifically including: adjusting the bandwidth of the kernel function to adapt to the local density of the data, accurately estimating the sparse area and the dense area of ​​the data, thereby accurately reflecting the characteristics of the sample data; The first sample data is analyzed for skewness by adaptive kernel density estimation. For the adaptive kernel density estimation graph, see Figure 2 Specific analysis results include: Affected by the detection limit, the distance between the mean and the peak of Au and Sb elements is large; the distribution of As, W, and Zn elements is relatively uniform, and the distance between the peak and the average value is small; the distribution range of Au, Ag, Bi, Cu, Mo, Pb, and Sb elements is wide and uneven, and the peaks are all located on the left side of the average value, indicating that the elements are dispersed and conducive to mineralization; In addition, from the comparison of the average element content of soil sediments in the study area with the element abundance value of the continental lithosphere and the average value of soil geochemical background in China, it can be seen that the content of Au, As, Mo, and W elements is much higher than the element abundance value of the continental lithosphere, the content of Ag and Cu elements is basically the same, and the content of other elements is relatively low; the content of Au, Ag, As, Bi, Cu, Mo, and W elements is much higher than the average value of soil geochemical background in China, and the content of other elements is relatively low, indicating that the element mineralization conditions in the study area are good; The statistical analysis of the first sample data by using the factor analysis method specifically includes: SPSS software was used to perform factor analysis on the first sample data. The analysis results showed that its KMO value was 0.667, and the Bartlett test result was passed, indicating that the data set was statistically suitable for factor analysis.

[0024] Extract the main factor eigenvalue greater than 1 and non-single factor factors to obtain the factor loading matrix, as shown in the following table:

[0025] Taking 0.5 as the factor loading standard, 4 main loading factors are obtained, including: F1 load factor (Ag-Bi-Cu-Pb element combination), F2 load factor (Au-Mo-W element combination), F3 load factor (As-Sb element combination); In addition, in order to strengthen the degree of variation of the low-temperature element combination of F3, this factor is combined with the medium- and high-temperature elements of F1 to construct a new main load factor F4 (F3 / F1).

[0026] The content data of various geological elements including four loading factors in the study area are used as the second sample data; The second sample data includes: 10 kinds of soil sediment geochemical exploration data, digital elevation model data, structural data and four main load factor variables F1, F2, F3 and F4 in the study area, a total of 16 kinds of data; S103: constructing a regular grid for the second sample data by using an inverse distance weighted interpolation method, superimposing and fusing the 16 types of element data therein to form multidimensional sample data, and expanding the data volume, enhancing the samples, and dividing the multidimensional sample data by using a step-size shift method and a random discarding method to improve the diversity of the sample data and obtain a sample data subset; The second sample data is constructed into a regular grid by using the inverse distance weighted interpolation method, and the construction formula is as follows: ; in, Indicates at point The estimated value of is the measured value at this point, is the number of known points used for interpolation, is the distance between the estimated value and the known point, is the weight index; The 16 types of data of the second sample data are superimposed and fused using the constructed regular grid to obtain multidimensional sample data. The multidimensional sample data point density map can be found in Figure 3 .

[0027] S104: expanding the data volume, enhancing the samples, and dividing the multidimensional sample data by a step-size shift method and a random discard method to obtain a sample data subset; Using the step-shift method, see Figure 4 , select a 16×16 window as the sample clipping size to form 256 grid units. By translating the window, the mineral-bearing points are sequentially placed in each grid unit and clipped. Each mineral-bearing point can obtain 256 sample data, and n mineral-bearing points can obtain 256×n sample data to obtain the expanded sample data set; Using random dropout method, see Figure 5 , while ensuring that the geological meaning remains unchanged, some data points in the expanded sample data set are randomly selected as label samples and their values ​​are replaced with 0, noise is added to the label samples, and a 16×16 window size is selected to crop the label samples. The required number of sample data is obtained by repeated multiple times, and the obtained sample data are added to the expanded sample data set. Sampling with replacement is used to obtain three sample data subsets with the same number.

[0028] S105: constructing a deep residual network model, and obtaining an improved deep residual network gold prospecting prediction model by integrating three deep residual network models, including the following steps: Define the residual unit: ; in, Indicates l The output of the residual unit, Indicates l The input of the residual unit is Indicates l+ The output of 1 residual unit, F () is the residual function, Indicates l The weight coefficients of the residual function, f () is the activation function, h ( ) represents the target mapping; The residual units are accumulated to obtain the residual block group, and the formula is: ; in, Indicates l The output of shallow units and the cumulative sum of a series of residual functions, Indicates l The weight coefficient of the residual function; Construct a deep residual network including an initial convolutional layer, a residual block group, a global average pooling layer, and a fully connected layer, and obtain a deep residual network prediction model by training it. See also Figure 6 , three deep residual network prediction models were integrated by voting to obtain an improved deep residual network prospecting prediction model (ResNet 18-18-34); The voting method is: averaging the output results of the three deep residual network prediction models; Wherein, each of the deep residual network models is a sub-model of the improved deep residual network gold prospecting prediction model; the three sub-models of the improved deep residual network gold prospecting prediction model are specifically: two deep residual network models (ResNet 18) containing 18 residual units and one deep residual network model (ResNet 34) containing 34 residual units; S106: Use the improved deep residual network gold prospecting prediction model to predict the sample data subset to obtain the potential distribution area of ​​gold mines, including: The three sample subset data are input into the three sub-models of the improved deep residual network gold prospecting prediction model respectively; The three sub-models predict the three sample subset data respectively to obtain the predicted value of each sub-model; Add up the predicted values ​​of all sub-models and take the average value to obtain the potential distribution area of ​​the final gold mine; In order to verify the beneficial effect of the improved deep residual network gold prospecting prediction method provided by the present invention, this example uses the ResNet 18-18-34 model provided by the method of the present invention to compare with three single models of CNN model, ResNet 18 model and ResNet 34 model and the integrated ResNet 18-18-18 model, specifically including: First, set the model parameters, including: in the CNN model, set the convolution kernel size to 3×3 window, the number of convolution kernels in the first layer to 32, the step size to 1, select the Adam optimization algorithm, and the learning rate to 0.001; in the ResNet18 and ResNet34 models, set the convolution kernel size to 3×3 window, the number of convolution kernels in the initial convolution layer to 64, the step size to 1, select the Adam optimization algorithm, and the learning rate to 0.001; Then, for three single models: CNN model, ResNet 18 model and ResNet 34 model, the center point cropping and step size translation methods are used to crop the control sample sets according to the fixed size. After cropping, each sample has 16 channels at the same time. The first 80% of the data is intercepted as the training set, and the last 20% of the data is taken as the validation set as the input data of the CNN model, ResNet 18 model and ResNet 34 model. The effects of different sample expansion methods on the model accuracy are verified through experiments, as shown in the following table:

[0029] The experimental results of the CNN model, ResNet 18 model, and ResNet 34 model show that the overall verification accuracy of the sample set model generated based on step-shift is higher than that of the sample set generated without step-shift. The accuracy of the model without step-shift is lower than 90%, indicating that the step-shift method effectively expands the number of samples, enhances the diversity of samples, and improves the verification accuracy of the model. Next, for the integrated models: ResNet 18-18-18 model and ResNet 18-18-34 model, based on the data set expanded by step-size shift, the random dropout method is used to enhance the sample data and expand it to the sample data set generated by the step-size shift method. In addition, combined with the boost idea, three sample data subsets with the same number are formed by sampling with replacement as the input data of the two integrated models. The above five prediction models are used to predict the sample data and the model accuracy is obtained as shown in the following table;

[0030] For the comparison of loss function and training accuracy of CNN model, ResNet 18 model and ResNet 34 model, see Figure 7 , the ResNet model has a significant improvement over the CNN model. From the downward trend of the loss function, it can be clearly seen that it has an effect on the vanishing gradient of the link; the accuracy of the ResNet34 model is improved compared with the ResNet18 model, but the improvement is not obvious; For all five prediction models, see the accuracy verification comparison Figure 8 The accuracy of the integrated model ResNet 18-18-18, which is an integration of three identical ResNet18 models, is 0.08% higher than that of a single ResNet18 model, and is 0.03% lower than that of a single ResNet34 model. The reason is that although the integrated model b uses three models for integration, each model has the same structure and does not learn deeper or more diverse features, resulting in a small overall improvement in the model. The integrated model ResNet 18-18-34, which is an integration of two ResNet18 models and one ResNet34 model, has a significant improvement in accuracy compared to the CNN model, ResNet18 model, and ResNet34 model. Through the analysis of the experimental results of this embodiment, the step-size shift method combined with the random discard method effectively expands and enhances the diversity of samples; In addition, since models with different structures have different judgments on different samples, the present invention integrates multiple models to complement the advantages of different models, comprehensively learns and captures the diverse characteristics of the data, and improves the model accuracy.

[0031] Guided by an improved deep residual network gold prospecting prediction method provided by the present invention, combined with the location and gold grade of known gold mineralization (Au>0.1 gt), the geological background of the study area and the gold ore-controlling factors, five prospecting target areas (A1, B1, C1, C2 and C3) were successfully identified.

[0032] In summary, the present invention uses the step-size shifting method and the random discarding method to expand and enhance the sample data, and combines the improved deep residual network gold prospecting prediction model to perform gold prospecting prediction on the sample data. It not only solves the problems of small sample size and low quality of geological element data, but also the improved deep residual network gold prospecting prediction model integrates multiple deep residual network sub-models. While avoiding the biased dependence of different models, it can also comprehensively learn and capture the diversity characteristics of geological element data in the study area, improve the model's recognition ability for diversified data, enhance the model's learning ability and generalization ability for diversified data, and improve the prediction accuracy of gold prospecting prediction, thereby providing scientific guidance for the practical application of prospecting in the study area.

[0033] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. All technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A method for prospecting prediction, characterized in that: The prediction method comprises the following steps: Obtaining content data of multiple geological elements in soil sediment samples in the study area, and using the content data of multiple geological elements as first sample data; Adaptive kernel density estimation method and factor analysis method were used to statistically analyze the first sample data to obtain four loading factors; and the content data of multiple geological elements containing four loading factors in the study area were used as the second sample data; By using an inverse distance weighted interpolation method, a regular grid is constructed for the second sample data, and the second sample data in the regular grid is fused to form multi-dimensional sample data; Through the step-size shift method and random discard method, the multi-dimensional sample data is expanded, enhanced and divided to obtain the sample data subset; By integrating three deep residual network models, an improved deep residual network prospecting prediction model is obtained; wherein the three sub-models of the improved deep residual network prospecting prediction model include: two deep residual network models containing 18 residual units and one deep residual network model containing 34 residual units; The improved deep residual network gold prospecting prediction model is used to predict the sample data subset and obtain the final prediction value to characterize the potential distribution area of ​​gold mines.

2. A method for prospecting prediction according to claim 1, characterized in that: The various elements in the soil sediment samples of the study area include: Au, Ag, As, Bi, Cu, Mo, Pb, Sb, W and Zn.

3. A method for prospecting prediction according to claim 1, characterized in that: The acquisition of the four load factors specifically includes: Adopting the adaptive kernel density estimation method, the bandwidth of the kernel function is adjusted to adapt to the local density of the first sample data, the sparse area and dense area of ​​the first sample data are estimated, and the element distribution of the first sample data in the study area is statistically analyzed to determine the metallogenic conditions of the elements in the study area; Using factor analysis method, we extracted the main factor with eigenvalue greater than 1 and non-single factor factors to obtain the loading factor matrix; The load factor values ​​in the load factor matrix represent the degree of correlation between the factors in the first sample data and the potential load factors, including: When the loading factor of the factor in the first sample data and the latent factor is close to 1, it means that the two are highly correlated; when the loading value is close to 0, it means that there is no significant correlation between the two; when the loading value is close to -1, it means that there is a negative correlation between the two; Set the factor load threshold, divide the factors of the first sample data according to the threshold, and obtain 4 groups of load factors, including: F1 load factor, including Ag, Bi, Cu and Pb elements; F2 load factor, including Au, Mo and W elements; F3 load factor, including As and Sb elements; The F4 load factor is a load factor constructed by dividing the high temperature element of the F3 load factor by the high temperature element of the F1 load factor, and the F4 load factor is used to enhance the degree of change of the F3 low temperature element combination.

4. A method for prospecting prediction according to claim 1, characterized in that: The second sample data includes: 10 types of soil sediment geochemical exploration data, digital elevation model data, structural data and four load factors, a total of 16 types of data.

5. A method for prospecting prediction according to claim 1, characterized in that: The acquisition of the three sample data subsets includes the following steps: The inverse distance weighted interpolation method is used to construct a regular grid for the second sample data. The construction formula is as follows: ； in, Indicates at point The estimated value of is the measured value at this point, is the number of known points used for interpolation, is the distance between the estimated value and the known point, is the weight index; Fusing the second sample data in the regular grid to obtain a multidimensional data set; The step-size translation method is adopted, and a 16×16 window is selected as the sample clipping size of the multidimensional data set to form 256 grid units. By translating the window, the mineral-bearing points are sequentially placed in each grid unit and clipped. Each mineral-bearing point can obtain 256 sample data, and n mineral-bearing points can obtain 256×n sample data, thus obtaining an expanded sample data set. The random discard method is adopted. While ensuring that the geological meaning remains unchanged, some data points in the expanded sample data set are randomly selected as label samples and their values ​​are replaced with 0. Noise is added to the label samples, and a 16×16 window is selected to crop the label samples. The specified number of sample enhancement data is obtained repeatedly, and the obtained sample enhancement data is added to the expanded sample data set. Sampling with replacement is used to obtain three sample data subsets with the same number.

6. A method for prospecting prediction according to claim 5, characterized in that: The method of constructing a regular grid for the second sample data using the inverse distance weighted interpolation method specifically includes: The regular grid size is set to 20m×20m, and the total number of grid cells is 103×133=13699; The digital elevation model data in the second sample data is resampled using data with a resolution of 12.5 meters to conform to the grid size.

7. A method for prospecting prediction according to claim 6, characterized in that: The method of making the mineral-bearing points fall into each grid unit in turn and perform cropping by translating the window specifically includes: during the window translation process, it is stipulated that there is only one mineral-bearing point in a sample, and the case of more than one mineral-bearing point is skipped for cropping.

8. A method for prospecting prediction according to claim 1, characterized in that: The improved deep residual network prospecting prediction model is obtained, comprising the following steps: Define the residual unit: ； in, Indicates l The output of the residual unit, Indicates l The input of the residual unit is Indicates l+ The output of 1 residual unit, F () is the residual function, represents the weight coefficient of the residual function, f () is the activation function, h ( ) represents the target mapping function; The residual units are accumulated to obtain the residual block group, and the formula is as follows: ； in, Indicates l The output of shallow units and the cumulative sum of a series of residual functions, Represents the weight coefficient of the residual function; Construct a deep residual network including an initial convolutional layer, a residual block group, a global average pooling layer, and a fully connected layer, and obtain a deep residual network prediction model by training it. By integrating three deep residual network prediction models through voting, an improved deep residual network prospecting prediction model is obtained; The voting method is specifically: averaging the output results of the three deep residual network prediction models.

9. A method for prospecting prediction according to claim 8, characterized in that: The deep residual network prediction model specifically includes: The initial convolutional layer receives the sample subset data, performs convolution operations on the data and performs batch normalization, and outputs the basic features in the data through the activation function; The residual block group receives the basic features output by the initial convolutional layer, obtains the residual features by performing residual learning on the basic features, and adds the basic features to the residual features by using a skip link to obtain the target mapping features; The global average pooling layer receives the target mapping features output by the residual block group, performs a simple average calculation on the target mapping features, and obtains the pooling features; The fully connected layer receives the pooled features and outputs the model prediction value through feature integration and complex mapping learning.

10. A method for prospecting prediction according to claim 1, characterized in that: The acquisition of the final prediction value comprises the following steps: The three sample subset data are input into the three sub-models of the improved deep residual network prospecting prediction model respectively; The three sub-models predict the three sample subset data respectively to obtain the predicted value of each sub-model; The predicted values ​​of all sub-models are added together and averaged to obtain the final predicted value, which is used to characterize the potential distribution area of ​​gold mines.

Images