Regional geochemical exploration extension method and device

Abstract

The invention provides a regional geochemical exploration extension method and device.The regional geochemical exploration extension method comprises the steps that a first geochemical database and a second geochemical database are provided, element types and contents in a first element group and a second element group in the first geochemical database are known quantities, and element types and contents in the first element group and the second element group in the second geochemical database are known quantities; the content in the second element group in the second geochemical database is an unknown quantity; constructing a deep neural network taking a full connection layer as a main body framework, and configuring parameter information of the deep neural network; training the deep neural network by using the first geochemical database to generate a prediction model; and based on the prediction model, by taking the first element group in the second geochemical database as an input quantity, determining the element content in a second element group in the second geochemical database so as to generate a geochemical anomaly map under the second sampling density. By adopting the technical scheme, the content value of the second element group can be quickly and effectively estimated.

Description

Technical Field

[0001] The invention belongs to the technical field of mineral resource exploration, and in particular relates to a regional geochemical exploration extension method and device. Background Art

[0002] Eleven minerals, including iron, aluminum, potassium salt, boron, beryllium, lithium, manganese, niobium, titanium, uranium and zirconium, are strategically scarce minerals in my country. Increasing their reserves through exploration is the key to solving the contradiction between supply and demand. Among them, regional geochemical exploration is one of the most effective means. By measuring the contents of 11 elements or oxides, such as Fe2O3, Al2O3 / SiO2, K2O, B, Be, Li, Mn, Nb, Ti, U and Zr, direct prospecting signs for the corresponding minerals can be obtained.

[0003] However, in important mineralization belts or small-scale regional geochemical anomalies, most of the 1:50,000 geochemical surveys carried out for the purpose of geological prospecting only analyzed and tested the 16 mandatory elements required by the "Specifications for Geochemical Surveys (1:50,000)", including Au, Ag, As, Bi, Cd, Cr, Co, Cu, Hg, Mo, Ni, Pb, Sb, Sn, W, and Zn. The above 11 elements or oxides are rarely involved, which poses a great challenge to the exploration of the above-mentioned strategic scarce minerals.

[0004] Although data can be obtained by reanalyzing and testing duplicate samples, this will consume a lot of manpower, material resources, financial resources and time. Therefore, for the above-mentioned strategic scarce minerals, there is an urgent need for the emergence of efficient and low-cost 1:50000 regional geochemical exploration extension methods.

[0005] At present, it is not uncommon to use artificial intelligence methods to estimate element content, and some results have been achieved, but these methods also have obvious defects. Although various interpolation methods (such as inverse distance weighting, kriging, spline function, trend surface, etc.) can estimate the content of target elements or oxides at unknown locations by internal interpolation of sparse data to achieve data encryption, they are basically unable to generate new abnormal information with practical value; the disadvantage of various machine learning methods (such as linear regression, random forest and shallow neural network, etc.) is that they cannot stimulate the advantages of exploration geochemical big data and cannot achieve more accurate estimation of the content of target elements or oxides. In summary, it is often difficult to systematically and comprehensively discover the complex and hidden mapping relationship between massive elements or oxide data using the above existing methods, which is not enough for accurately predicting the content value of target elements or oxides.

[0006] Therefore, how to estimate the content of target elements or oxides more quickly and effectively becomes a difficult point in geochemical exploration. Summary of the invention

[0007] In view of this, an embodiment of the present invention provides a method and device for extending regional geochemical exploration, which can quickly and effectively estimate the content value of the second element group.

[0008] An embodiment of the present invention provides a regional geochemical exploration extension method, comprising:

[0009] Providing a first geochemical database obtained based on a first sampling density and a second geochemical database obtained based on a second sampling density, wherein the types and contents of elements in the first element group and the second element group in the first geochemical database are known, and the contents of elements in the second element group in the second geochemical database are unknown;

[0010] Construct a deep neural network with a fully connected layer as the main architecture, and configure parameter information of the deep neural network;

[0011] Using the first geochemical database, training the deep neural network to generate a prediction model;

[0012] Based on the prediction model, the first element group in the second geochemical database is used as input to determine the element content in the second element group in the second geochemical database to generate a geochemical anomaly map under the second sampling density.

[0013] Optionally, the constructing a deep neural network with a fully connected layer as the main architecture and configuring parameter information of the deep neural network includes:

[0014] Determining a network architecture with the fully connected layer as the main body, including: determining an input layer of the network architecture, the input layer is composed of a plurality of first neurons, and the number of the first neurons is equal to the number of elements in the first element group; determining an output layer of the network architecture, the output layer is composed of a plurality of second neurons, and the number of the second neurons is equal to the number of elements in the second element group; determining a plurality of hidden layers for connecting the input layer and the output layer, and each hidden layer is connected to each other;

[0015] The deep neural network is configured to perform data standardization processing on each hidden layer, wherein the standardization processing formula is:

[0016]

[0017] in, represents the value after batch normalization, γ and β represent the values ​​of scaling and translating the normalized data, respectively, and x i represents the i-th input value in the hidden layer, x meanrepresents the average value of the input value in the hidden layer, σ represents the standard deviation of the input value in the hidden layer, and ξ is a very small constant;

[0018] Define the activation function in each hidden layer of the deep neural network, where the activation function is:

[0019] PReLU(x)=max(0,x)+ηmin(0,x)

[0020] Where η represents the learning parameter;

[0021] And determine the number of neurons that are randomly inactivated in the hidden layer during each training process of the deep neural network.

[0022] Optionally, the number of third neurons in each hidden layer is greater than the number of the first neurons;

[0023] During each training process of the deep neural network, the number of randomly inactivated neurons in the hidden layer is 20% of the number of all neurons.

[0024] Optionally, the using the first geochemical database to train the deep neural network to generate a prediction model includes:

[0025] Dividing the first element group and the second element group in the first geochemical database into a training data set and a verification data set according to a preset ratio;

[0026] The training data set is sequentially input into the deep neural network and forward propagated to determine the cost function of the prediction result of the deep neural network, wherein the cost function is:

[0027]

[0028] Among them, m is the number of training samples; y gt i is the true value of the i-th training sample; y pre i is the predicted value of the i-th training sample; L(y gt i ,y pre i ) is the loss function of training sample i; δ is a threshold; λ is the penalty coefficient; R(w) is the penalty term, where w is the weight of the network;

[0029]

[0030] Among them, n is the number of network weights, w i is the i-th weight;

[0031] Using the gradient descent method for back propagation to update the weight parameters in the cost function to obtain an initial prediction model;

[0032] Input all validation data sets into the initial prediction model and forward propagate to determine one or more of the mean square error and absolute error of the prediction results as validation evaluation indicators;

[0033] Repeat the iterative training and validation steps until the cost function value and validation evaluation index meet the requirements.

[0034] Optionally, in the process of determining the cost function of the deep neural network prediction result, it also includes:

[0035] Determine an optimization function that is compatible with the deep neural network, and optimize the weight parameters in the cost function during the training process, wherein the optimization function includes:

[0036]

[0037] Where w is the weight parameter, := means that the value on its left is updated or replaced by the value on its right, α is the learning rate, β1 is the first momentum, β2 is the second momentum, dw is the derivative of the cost function with respect to the weight parameter, v dw is the momentum exponential weighted average of dw, s dw for dw 2 Momentum exponential weighted average, ε is a constant, and t represents the number of iterations;

[0038] And initialize the weight parameters.

[0039] Optionally, before dividing the first element group and the second element group in the first geochemical database into a training data set and a verification data set according to a preset ratio, the method further includes:

[0040] The first element group and the second element group in the first geochemical database, and the first element group in the second geochemical database are preprocessed, and the preprocessing includes: replacing the missing elements in the first element group and the second element group with any value between zero and the element detection limit, wherein the detection limit is the lowest value of the corresponding element detected; and, standardizing the data in the first element group and the second element group after the missing processing, and the parameters of the standardization processing in the first geochemical database and the second geochemical database are the same.

[0041] Optionally, the first geochemical database and the second geochemical database further include geographical landscape nominal scale variables, and the geographical landscape nominal scale variables are used to characterize different types of geographical landscapes;

[0042] The deep neural network is trained by using the first geochemical database to generate a prediction model. The training process also includes: using the geographical landscape nominal scale variables as a data set to train the deep neural network.

[0043] Optionally, the step of determining the element content in the second element group in the second geochemical database based on the prediction model and taking the first element group in the second geochemical database as input to generate a geochemical anomaly map under the second sampling density includes:

[0044] Taking the first element group in the second geochemical database as input, and inputting it into the prediction model, generating geochemical standard values ​​corresponding to each element in the second element group, and restoring them to original values ​​of each element in the second element group;

[0045] interpolating the original values ​​of each element in the second element group to generate a geochemical map corresponding to the second element group;

[0046] The geochemical map is processed by taking the mean of the original values ​​of each element in the second element group and the standard deviation of the original values ​​of each element as the anomaly lower limit to generate a geochemical anomaly map.

[0047] Optionally, the regional geochemical exploration extension method further includes:

[0048] Based on the geochemical anomaly map and multivariate information data, the prospecting target area for the target mineral species is determined.

[0049] Accordingly, an embodiment of the present invention further provides a regional geochemical exploration extension device, comprising:

[0050] a geochemical database providing unit, adapted to provide a first geochemical database obtained based on a first sampling density and a second geochemical database obtained based on a second sampling density, wherein the types and contents of elements in the first element group and the second element group in the first geochemical database are known quantities, and the contents of elements in the second element group in the second geochemical database are unknown quantities;

[0051] A neural network construction unit, adapted to construct a deep neural network with a fully connected layer as the main architecture, and configure parameter information of the deep neural network;

[0052] The processing unit is suitable for training the deep neural network using the first geochemical database to generate a prediction model; and based on the prediction model, taking the first element group in the second geochemical database as input, determining the element content in the second element group in the second geochemical database to generate a geochemical anomaly map under the second sampling density.

[0053] Compared with the prior art, the technical solution of the embodiment of the present invention has the following advantages:

[0054] By using the regional geochemical exploration extension method provided by the embodiment of the present invention, the element types and contents in the first element group and the second element group in the first geochemical database are all known quantities, indicating that the first geochemical database is reliable. In this way, the first geochemical database is used to train the deep neural network, so that the prediction model can better learn and more comprehensively and objectively reflect the relationship between the first element group and the second element group in the first geochemical database. In this way, based on the prediction model, the element content in the second element group in the second geochemical database can be quickly and effectively ascertained, and then a geochemical anomaly map for guiding mineral exploration work can be generated. In other words, the above process uses the first geochemical database to train the deep neural network, and through transfer learning, it is used to predict the element content of the second geochemical database. The whole process is simpler, so the content value of the second element group can be quickly and effectively estimated. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0056] Figure 1 A flow chart of a regional geochemical exploration extension method according to an embodiment of the present invention is shown;

[0057] Figure 2 A flowchart for determining a deep neural network in an embodiment of the present invention is shown;

[0058] Figure 3 A schematic diagram of the architecture of a deep neural network in an embodiment of the present invention is shown;

[0059] Figure 4 A schematic diagram of the deactivation principle of a deep neural network in an embodiment of the present invention is shown;

[0060] Figure 5A flowchart showing a process of generating a prediction model in an embodiment of the present invention is shown;

[0061] Figure 6 A schematic structural diagram of a regional geochemical exploration extension device in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0062] As described in the background technology, the retesting and analysis of elements or oxides such as Fe2O3, Al2O3 / SiO2, K2O, B, Be, Li, Mn, Nb, Ti, U, Zr, etc. in the 1:50000 regional geochemical exploration method will consume a lot of manpower, material resources, financial resources and time.

[0063] The inventors found in actual application that China's 1:200000 regional geochemical survey has collected a large amount of data (for example, more than 1.47 million data points, with a total data volume of about 63.21 million) on 39 elements or oxides, including 16 elements (namely: Au, Ag, As, Bi, Cd, Cr, Co, Cu, Hg, Mo, Ni, Pb, Sb, Sn, W, and Zn) measured by the 1:50000 regional geochemical survey. These elements or oxides are SiO2, Al2O3, K2O, Na2O, CaO, MgO, Fe2O3, Ag, As, Au, B, Ba, Be, Bi, Cd, Co, Cr, Cu, F, Hg, La, Li, Mn, Mo, Nb, Ni, P, Pb, Sb, Sn, Sr, Th, Ti, U, V, W, Y, Zn, and Zr. Moreover, as far as 1:200000 and 1:50000 regional geochemical surveys are concerned, in similar geographical landscapes, although the sampling densities are obviously different, the sampling media and analysis methods are basically the same, so the correlation between the different elements reflected by the two should also be similar.

[0064] Based on this, an embodiment of the present invention provides a regional geochemical exploration extension method, in which the element types and contents in the first element group and the second element group in the first geochemical database are all known quantities, indicating that the first geochemical database is reliable. In this way, the first geochemical database is used to train the deep neural network, so that the prediction model can better learn and more comprehensively and objectively reflect the relationship between the first element group and the second element group in the first geochemical database. In this way, based on the prediction model, the element content in the second element group in the second geochemical database can be quickly and effectively ascertained, and then a geochemical anomaly map for guiding mineral exploration work can be generated. In other words, the above process uses the first geochemical database to train the deep neural network, and through transfer learning, it is used to predict the element content of the second geochemical database. The whole process is simpler, so the content value of the second element group can be quickly and effectively estimated.

[0065] In order to enable those skilled in the art to better understand the inventive concept, working principle and advantages of the embodiments of the present invention, the regional geochemical exploration extension method in the embodiments of the present invention is described in detail below.

[0066] See also Figure 1 A flow chart of a regional geochemical exploration extension method in an embodiment of the present invention is shown in FIG. Figure 1 As shown, the following steps can be performed:

[0067] S11, providing a first geochemical database obtained based on a first sampling density and a second geochemical database obtained based on a second sampling density.

[0068] Specifically, the first geochemical database and the second geochemical database can characterize the distribution information of elements under different geographical conditions, and the first geochemical database and the second geochemical database are acquired under different sampling densities.

[0069] In this embodiment, the first sampling density may refer to 1:200,000 exploration geochemistry, and the second sampling density may refer to 1:50,000 exploration geochemistry.

[0070] It should be noted that the numerical values ​​listed in the above examples are only for illustration and are used to represent different sampling densities, and should not be construed as limiting the present invention. In some other embodiments, the first sampling density and the second sampling density may also be other values, as long as they are different.

[0071] In this embodiment, the element types and contents in the first element group and the second element group in the first geochemical database are known quantities. As a non-limiting example, the first element group may include 16 elements (namely: Au, Ag, As, Bi, Cd, Cr, Co, Cu, Hg, Mo, Ni, Pb, Sb, Sn, W, Zn), and the second element group may include 11 elements (namely: Fe2O3, Al2O3 / SiO2, K2O, B, Be, Li, Mn, Nb, Ti, U, Zr elements or oxides).

[0072] The element content in the second element group in the second geochemical database is an unknown quantity, or in other words, the second element group in the second geochemical database is a null value, which is the quantity that needs to be predicted in the present invention.

[0073] S12, constructing a deep neural network with a fully connected layer as the main architecture, and configuring parameter information of the deep neural network.

[0074] Specifically, the deep neural network with a fully connected layer as the main architecture can better learn and more comprehensively and objectively reflect the relationship between the first element group and the second element group in the first geochemical database, and by configuring the parameter information of the deep neural network, it can better process the first element group and the second element group in the first geochemical database.

[0075] S13: Using the first geochemical database, the deep neural network is trained to generate a prediction model.

[0076] Specifically, the element types and contents in the first element group and the second element group in the first geochemical database are all known quantities, indicating that the first geochemical database is reliable. In this way, when the deep neural network is trained using the first geochemical database, the prediction model can better learn and more comprehensively and objectively reflect the relationship between the first element group and the second element group in the first geochemical database, so that the prediction model has better generalization ability.

[0077] S14, based on the prediction model, taking the first element group in the second geochemical database as input, determining the element content in the second element group in the second geochemical database to generate a geochemical anomaly map under the second sampling density.

[0078] Specifically, the prediction model is based on the first geochemical database, so that the prediction model can quickly and effectively find out the element content in the second element group in the second geochemical database, and then generate a geochemical anomaly map for guiding mineral exploration work.

[0079] In other words, the above process uses the first geochemical database to train a deep neural network, and uses it for predicting the element content of the second geochemical database through transfer learning. The whole process is simpler, and thus the content value of the second element group can be estimated quickly and effectively.

[0080] In this embodiment, combined with Figure 2 and Figure 3 ,in, Figure 2 is a flowchart of determining a deep neural network in an embodiment of the present invention. Figure 3 Schematic diagram of a deep neural network architecture in an embodiment of the present invention. Figure 2 and Figure 3 As shown, the following steps can be performed:

[0081] S21, determining a network architecture with the fully connected layer as the main body.

[0082] Specifically, the fully connected layer can integrate the extracted features so that the output of the deep neural network can more realistically reflect the relationship between the first element group (which can be used as an independent variable) and the second element group (which can be used as a dependent variable) in the first geochemical database.

[0083] In this embodiment, the following steps may be performed:

[0084] The input layer of the network architecture is determined.

[0085] Specifically, the input layer is used to process the first element group in the first geochemical database input into the network architecture so that the deep neural network can correctly identify the first element group.

[0086] In this embodiment, the input layer is composed of a plurality of first neurons, and the number of the first neurons is equal to the number of elements in the first element group, so that the first neurons can completely correspond to the elements in the first element group.

[0087] As a specific embodiment, when the number of elements in the first element group is 17, the input layer can be composed of 17 first neurons, so that the first neurons numbered 1 to 16 in the input layer correspond to Au, Ag, As, Bi, Cd, Cr, Co, Cu, Hg, Mo, Ni, Pb, Sb, Sn, W, and Zn in the first element group, respectively, and the first neuron numbered 17 is a geographical landscape.

[0088] An output layer of the network architecture is determined.

[0089] Specifically, the output layer is used to process the second element group in the first geochemical database input into the network architecture so that the deep neural network can correctly identify the second element group.

[0090] In this embodiment, the output layer is composed of a plurality of second neurons, and the number of the second neurons is equal to the number of elements in the second element group, so that the second neurons can completely correspond to the elements in the second element group.

[0091] As a specific embodiment, when the number of elements in the second element group is 11, the output layer can be composed of 11 second neurons, so that the second neurons numbered 1 to 11 in the output layer correspond to Fe2O3, Al2O3 / SiO2, K2O, B, Be, Li, Mn, Nb, Ti, U, and Zr in the second element group, respectively.

[0092] A plurality of hidden layers for connecting the input layer and the output layer are determined, and each hidden layer is connected to each other.

[0093] In this embodiment, the setting of the hidden layer includes the setting of the number of hidden layers and the number of neurons in each hidden layer. The more the number of hidden layers and the number of third neurons, the more features will be extracted, but at the same time, the larger the number of parameters will be, and the risk of overfitting will also increase.

[0094] In this case, the number of third neurons in each hidden layer is greater than the number of first neurons, which can avoid missing important features.

[0095] Furthermore, the number of hidden layers should be deep enough to extract more advanced features.

[0096] In a specific embodiment, the number of hidden layers may be set to 20, and the number of third neurons in each hidden layer may be set to 32.

[0097] S22, configuring the deep neural network to perform data standardization processing on each hidden layer.

[0098] Specifically, by performing data normalization, the search for hyperparameters becomes simple, making deep neural networks more robust and easier to train.

[0099] In this embodiment, the normalization processing formula is:

[0100]

[0101] in, represents the value after batch normalization, γ and β represent the values ​​of scaling and translating the normalized data, respectively, and x i represents the i-th input value in the hidden layer, x mean represents the average value of the input in the hidden layer, σ represents the standard deviation of the input in the hidden layer, and ξ is a very small constant to prevent the denominator from being 0.

[0102] S23, defining an activation function in each hidden layer in the deep neural network.

[0103] Specifically, by introducing the activation function, deep neural networks have the ability to solve nonlinear problems. At present, the ReLU activation function is widely used in hidden layers. It can solve the gradient vanishing problem caused by the saturated activation function and greatly speed up the learning speed. However, the disadvantage of this function is that when a very large gradient passes through the back propagation process, the back propagation update may cause the center of the weight distribution to be less than 0, so that the derivative at this point is always 0, and the back propagation cannot update the weight, that is, it enters the inactivated state.

[0104] Therefore, unlike the ReLU activation function in the existing solution, the activation function in the embodiment of the present invention is:

[0105] PReLU(x)=max(0,x)+ηmin(0,x) (2)

[0106] Where η represents the learning parameter.

[0107] By using formula (2), the parameter η can be adaptively learned from the data, which not only solves the above problem but also has the characteristics of fast convergence speed and low error rate.

[0108] It should be noted that data normalization has been performed before defining the activation function.

[0109] S24, determining the number of randomly inactivated neurons in the hidden layer during each training process of the deep neural network.

[0110] Specifically, the most common problem encountered by deep neural networks during training is overfitting. In order to alleviate overfitting and improve the generalization ability of deep neural networks, random dropout is used in the model. The so-called random dropout means that in each training process, for each hidden layer of neurons, according to a certain probability, some neurons are randomly excluded from model training.

[0111] For example, see Figure 4 A schematic diagram of the deactivation principle of a deep neural network in an embodiment of the present disclosure is shown in FIG. Figure 4 As shown, in each training process, the probability of random inactivation can be made 20%, that is, in each training process of the deep neural network, the number of randomly inactivated neurons in the hidden layer (for example Figure 4 The number of neurons (shown by the dotted line) is 20% of the total number of neurons in the hidden layer, which can effectively prevent overfitting.

[0112] Therefore, by adopting a network architecture with a fully connected layer as the main body, supplemented by data standardization processing, activation functions and random inactivation methods, the prediction structure of the deep neural network is made more accurate and reliable.

[0113] In this embodiment, after the parameter information of the deep neural network is configured, the deep neural network can be trained to generate a prediction model.

[0114] See also Figure 5 FIG. 1 is a flowchart of a process for generating a prediction model in an embodiment of the present invention, as shown in FIG. Figure 5 As shown, the following generation steps can be performed:

[0115] S31, dividing the first element group and the second element group in the first geochemical database into a training data set and a verification data set according to a preset ratio.

[0116] Specifically, the first element group and the second element group in the first geochemical database can be randomly shuffled to ensure that the relationship between the samples in the first geochemical database is random and has no logic in order, thereby preventing similar samples from being too concentrated in the training set or the validation set, resulting in the low generalization ability of the trained deep learning model; then, the first geochemical database with a randomly shuffled order is divided into corresponding training set and validation set folders according to a certain ratio, for example, 7:3 or 8:2, 9:1, to form training data sets and validation data sets for the training and validation process of the deep neural network.

[0117] In this embodiment, considering that the contents of some elements in the first element group and the second element group are low, for example, lower than the detection limit, or the formats of some elements in the first element group and the second element group are inconsistent, before dividing the first element group and the second element group in the first geochemical database into a training data set and a verification data set according to a preset ratio, the regional geochemical exploration extension method may further include:

[0118] The first element group and the second element group in the first geochemical database, and the first element group in the second geochemical database are preprocessed.

[0119] By performing the preprocessing operation, the integrity of the first element group and the second element group in the first geochemical database and the elements in the first element group in the second geochemical database can be improved, and the consistency of the data format is improved, which is conducive to reducing the difficulty of training.

[0120] In this embodiment, the preprocessing may include:

[0121] Any value between zero and a detection limit is used to replace the missing elements in the first element group and the second element group, wherein the detection limit is the lowest value at which the corresponding element is detected.

[0122] Specifically, in order to make up for the defect that some elements cannot be detected, this scheme replaces the missing values ​​with random values ​​between zero and the detection limit, so that there are no missing elements in the first element group and the second element group, and can prevent the data analysis results from being biased or misleading.

[0123] Furthermore, as shown in formula (3), the random value obeys uniform distribution, and the reason for adopting uniform distribution is that the probability of the random value appearing in any value in this interval is equal.

[0124] X~U(0,ρ)(3)

[0125] Where X represents a random value, U represents a uniform distribution, and ρ represents the detection limit of the element.

[0126] Next, the data in the first element group and the second element group after the missing data processing are standardized, so as to eliminate the impact of the dimensions between different variables and make them comparable.

[0127] In addition, in order to improve the convergence speed and stability, the standardization processing method in this scheme can be seen in formula (4):

[0128]

[0129] Among them, x i is the i-th value, x mean is the mean value of the training set, σ is the standard deviation of the training set, ξ is a very small constant to prevent the denominator from being 0, x std is the normalized value.

[0130] Furthermore, the parameters of the standardization process in the standardization process of the first geochemical database and the second geochemical database are the same.

[0131] It should be noted that the so-called same parameters here refer to the mean and standard deviation used in the standardization processing of the first element group and the second element group are the mean and standard deviation of the training data set. The purpose is to ensure that the model evaluation and model prediction are consistent with the training set in terms of measurement, thereby ensuring the reliability of the prediction effect.

[0132] S32, inputting the training data set into the deep neural network in sequence, and forward propagating, to determine the cost function of the prediction result of the deep neural network.

[0133] Specifically, the cost function is an indicator of the "badness" of the deep learning model, which is used to measure the effect of the weight parameter w on the training set. When the cost function value converges to a suitable accuracy, it indicates that the deep neural network has reached the optimal level during the training stage.

[0134] By inputting the training data set into the deep neural network in sequence, the prediction result of the deep neural network under the current state of the training data set can be determined, and further, the cost function can be determined.

[0135] The cost function is:

[0136]

[0137] Among them, m is the number of training samples; y gt i is the true value of the i-th training sample; y pre i is the predicted value of the i-th training sample; L(y gt i ,y prei ) is the loss function of training sample i; δ is a threshold (e.g. 1); λ is the penalty coefficient (e.g. 0.01); R(w) is the penalty term, where w is the weight of the network;

[0138]

[0139] Among them, n is the number of network weights, w i is the ith weight.

[0140] In this embodiment, in the process of determining the cost function of the deep neural network prediction result, it also includes: determining an optimization function compatible with the deep neural network, and optimizing the weight parameters in the cost function.

[0141] The optimization function is used to optimize the weight parameters. Among them, Adam combines the advantages of the two methods of Momentum and RMSprop, realizes efficient search of parameter space, and is widely used. This deep learning model optimization adopts this method.

[0142] In this embodiment, the optimization function is:

[0143]

[0144] Where w is the weight parameter, := means that the value on its left is updated or replaced by the value on its right, α is the learning rate, β1 is the first momentum, β2 is the second momentum, dw is the derivative of the cost function with respect to the weight parameter, v dw is the momentum exponential weighted average of dw, s dw for dw 2 Momentum exponential weighted average, ε is a constant, and t represents the number of iterations.

[0145] Next, the weight parameters are initialized.

[0146] Specifically, the weight parameter initialization is to prevent the so-called "symmetry" problem, that is, to prevent all neurons in the same hidden layer from having the same function, and not allowing all weight parameters to be initialized to 0, so random initialization is required.

[0147] Specifically, in order to meet the two conditions that the initialization parameters are relatively small and the changes are relatively stable, this deep learning method uses a random value that conforms to the standard normal distribution as the initialization parameter, that is, the formula:

[0148] W~N(0,1) / 100(8)

[0149] Among them, W is the weight parameter and N(0,1) is the standard normal distribution.

[0150] S33, using the gradient descent method to perform back propagation to update the weight parameters in the cost function to obtain an initial prediction model.

[0151] Specifically, through steps S31 and S32, the cost function of the deep neural network prediction results can be determined, and then through the Adam gradient descent method, the weight parameters in the cost function can be further optimized, so that the prediction structure of the deep neural network is closer to the actual result, thereby obtaining the corresponding initial prediction model.

[0152] S34, input all the verification data sets into the initial prediction model, and forward propagate to determine one or more of the mean square error and absolute error of the prediction results as verification evaluation indicators.

[0153] Specifically, by executing step S34, inputting all verification data sets into the initial prediction model and forward propagating, the verification evaluation indicators of the prediction results can be determined, for example, one or more of the mean square error and the absolute error can be used as the verification evaluation indicators, so as to evaluate the quality of the initial prediction model.

[0154] Specifically, the mean square error MES and absolute error MAE can be determined by the following formula:

[0155]

[0156] Among them, m is the number of training samples; y gt i is the true value of the i-th training sample; y pre i is the predicted value of the i-th training sample.

[0157] S35, repeat the iterative training and verification steps until the cost function value and the verification evaluation index meet the requirements.

[0158] For example, the above steps are repeated multiple times (for example, 1000 times) until the cost function and the verification evaluation index reach the required accuracy, which is mainly manifested in that the cost function value no longer decreases significantly and the verification evaluation index no longer increases significantly, that is, the optimal deep learning model (prediction model) is obtained.

[0159] Therefore, the above training method, on the one hand, by dividing the training and validation data sets, lays a sample foundation for the deep learning model based on big data; on the other hand, the loss function is improved by introducing penalty coefficients and penalty terms, which alleviates the problem of sample imbalance to a certain extent, thereby improving the generalization ability of the prediction model.

[0160] In addition, for the loss function, the mean square error and absolute error loss are fused, which can prevent the negative impact of outliers in exploration geochemical data on model training and improve the generalization ability of the model.

[0161] In this embodiment, the first geochemical database and the second geochemical database also include geographical landscape nominal scale variables, which are used to characterize different types of geographical landscapes. By adding the geographical landscape nominal scale independent variable, the trained model can adapt to different geographical landscapes.

[0162] In a specific embodiment, different symbols can be used to represent different types of geographical landscapes, for example, 0—medium and low mountains, hills; 1—forest swamps; 2—arid and semi-arid deserts; 3—alpine mountainous areas; 4—alpine lake deserts; 5—karst; 6—tropical rainforest; 7—Loess Plateau; 8—alluvial plain.

[0163] It should be noted that, first, the nominal scale variables listed in the above examples, i.e. 0 to 8, are integers, but there is no difference in size or order between them. The different numerical values ​​can only represent different categories of geographical landscapes; second, the geographical landscapes listed in the above examples are also illustrative examples, which are used to represent different types of geographical landscapes.

[0164] In this case, the deep neural network is trained using the first geochemical database to generate a prediction model, and the training process also includes: using the geographic landscape nominal scale variables as a data set to train the deep neural network.

[0165] As a specific example, see Figure 3 , the first neuron numbered 17 in the input layer corresponds to the nominal scale variable of the geographic landscape.

[0166] In summary, the deep neural network is trained by using the first geochemical database to generate a prediction model, and then the prediction model can be used to predict the element content in the second element group in the second geochemical database.

[0167] More specifically, the following steps may be performed:

[0168] The first element group in the second geochemical database is used as an input and input into the prediction model to generate geochemical standard values ​​corresponding to each element in the second element group and restore them to the original values ​​of each element in the second element group.

[0169] Specifically, 17 elements (i.e., the first element group) can be loaded into the trained optimal deep learning model to carry out prospecting prediction, estimate the geochemical standardized values ​​of 11 elements (i.e., the second element group), and then obtain the original value through formula (11):

[0170] x i =x mean +x std ×σ (11)

[0171] Among them, x i is the original value, x mean is the mean of the training set, x std is the estimated geochemical normalized value mentioned above, and σ is the standard deviation of the training set.

[0172] Next, the original values ​​of the elements in the second element group are interpolated to generate a geochemical map corresponding to the second element group.

[0173] Finally, the geochemical map is processed by taking the mean of the original values ​​of each element in the second element group and the standard deviation of the original values ​​of each element as the anomaly lower limit to generate a geochemical anomaly map.

[0174] For example, the mean of the original geochemical value + 2 times the standard deviation is taken as the lower limit of the anomaly.

[0175] Furthermore, based on the generated geochemical anomaly map, the prospecting target areas of scarce minerals such as iron, aluminum, potassium salt, boron, beryllium, lithium, manganese, niobium, titanium, uranium and zirconium can be determined based on the geochemical anomaly map and multi-information data (such as geology, geophysical exploration and remote sensing).

[0176] Finally, field verification is carried out on the prospecting target area of ​​the target mineral species delineated on the surface. Based on its exposure, degree of mineralization, surface occurrence and other information, a certain degree of geological mapping, trenching and drilling (including chemical sample analysis) is used to determine the shape, scale, spatial distribution, grade, co-existing elements and resource quantity of the ore body in the target area.

[0177] The embodiment of the present invention also provides a device corresponding to the regional geochemical exploration extension method, such as Figure 6 The structural diagram of a regional geochemical exploration extension device in an embodiment of the present invention is shown in FIG. Figure 6 As shown, the regional geochemical exploration extension device 100 may include:

[0178] The geochemical database providing unit 110 is adapted to provide a first geochemical database obtained based on the first sampling density and a second geochemical database obtained based on the second sampling density, wherein the element types and contents in the first element group and the second element group in the first geochemical database are known, and the element contents in the second element group in the second geochemical database are unknown;

[0179] A neural network construction unit 120, adapted to construct a deep neural network with a fully connected layer as the main architecture, and configure parameter information of the deep neural network;

[0180] The processing unit 130 is suitable for using the first geochemical database to train the deep neural network to generate a prediction model; and based on the prediction model, taking the first element group in the second geochemical database as input, determining the element content in the second element group in the second geochemical database to generate a geochemical anomaly map under the second sampling density.

[0181] Among them, the specific working process and principle of the geochemical database providing unit 110, the neural network building unit 120 and the processing unit 130 can refer to the relevant description of the above examples.

[0182] It is understandable that the division of the above units is only a division of logical functions, and in actual implementation, they can be fully or partially integrated into one physical entity, or physically separated. In addition, the above units can be implemented in the form of a processor calling software.

[0183] Although the embodiments of the present invention are disclosed above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A regional geochemical exploration extension method, characterized in that: include: Providing a first geochemical database obtained based on a first sampling density and a second geochemical database obtained based on a second sampling density, wherein the types and contents of elements in the first element group and the second element group in the first geochemical database are known, and the contents of elements in the second element group in the second geochemical database are unknown; Construct a deep neural network with a fully connected layer as the main architecture, and configure parameter information of the deep neural network; Using the first geochemical database, training the deep neural network to generate a prediction model; Based on the prediction model, the first element group in the second geochemical database is used as input to determine the element content in the second element group in the second geochemical database to generate a geochemical anomaly map under the second sampling density.

2. The regional geochemical exploration extension method according to claim 1, characterized in that: The step of constructing a deep neural network with a fully connected layer as the main structure and configuring parameter information of the deep neural network includes: Determining a network architecture with the fully connected layer as the main body, including: determining an input layer of the network architecture, the input layer is composed of a plurality of first neurons, and the number of the first neurons is equal to the number of elements in the first element group; determining an output layer of the network architecture, the output layer is composed of a plurality of second neurons, and the number of the second neurons is equal to the number of elements in the second element group; determining a plurality of hidden layers for connecting the input layer and the output layer, and each hidden layer is connected to each other; The deep neural network is configured to perform data standardization processing on each hidden layer, wherein the standardization processing formula is: in, represents the value after batch normalization, γ and β represent the values ​​of scaling and translating the normalized data, respectively, and x i represents the i-th input value in the hidden layer, x mean represents the mean value of the input value in the hidden layer, and σ represents the standard deviation of the input value in the hidden layer; Define the activation function in each hidden layer of the deep neural network, where the activation function is: PReLU(x)=max(0,x)+ηmin(0,x) Where η represents the learning parameter; And determine the number of neurons that are randomly inactivated in the hidden layer during each training process of the deep neural network.

3. The regional geochemical exploration extension method according to claim 2, characterized in that: The number of third neurons in each hidden layer is greater than the number of the first neurons; During each training process of the deep neural network, the number of randomly inactivated neurons in the hidden layer is 20% of the number of all neurons.

4. The regional geochemical exploration extension method according to claim 1, characterized in that: The using the first geochemical database to train the deep neural network to generate a prediction model includes: Dividing the first element group and the second element group in the first geochemical database into a training data set and a verification data set according to a preset ratio; The training data set is sequentially input into the deep neural network and forward propagated to determine the cost function of the prediction result of the deep neural network, wherein the cost function is: Where m is the number of training samples; is the true value of the i-th training sample; is the predicted value of the i-th training sample; is the loss function of training sample i; δ is a threshold; λ is the penalty coefficient; R(w) is the penalty term, where w is the weight of the network; Among them, n is the number of network weights, w i is the i-th weight; Using the gradient descent method for back propagation to update the weight parameters in the cost function to obtain an initial prediction model; Input all validation data sets into the initial prediction model and forward propagate to determine one or more of the mean square error and absolute error of the prediction results as validation evaluation indicators; Repeat the iterative training and validation steps until the cost function value and validation evaluation index meet the requirements.

5. The regional geochemical exploration extension method according to claim 4, characterized in that: In the process of determining the cost function of the prediction result of the deep neural network, it also includes: Determine an optimization function that is compatible with the deep neural network, and optimize the weight parameters in the cost function during the training process, wherein the optimization function includes: Where w is the weight parameter, := means that the value on the left is updated or replaced by the value on the right, α is the learning rate, β1 is the first momentum, β2 is the second momentum, dw is the derivative of the cost function with respect to the weight parameter, v dw is the momentum exponential weighted average of dw, s dw for dw 2 Momentum exponential weighted average, ε is a constant, and t represents the number of iterations; And initialize the weight parameters.

6. The regional geochemical exploration extension method according to claim 4, characterized in that: Before dividing the first element group and the second element group in the first geochemical database into a training data set and a verification data set according to a preset ratio, the method further includes: The first element group and the second element group in the first geochemical database, and the first element group in the second geochemical database are preprocessed, and the preprocessing includes: replacing the missing elements in the first element group and the second element group with any value between zero and the detection limit, wherein the detection limit is the lowest value of the corresponding element detected; and, standardizing the data in the first element group and the second element group after the missing processing, and the parameters of the standardization processing in the first geochemical database and the second geochemical database are the same.

7. The regional geochemical exploration extension method according to claim 1, characterized in that: The first geochemical database and the second geochemical database further include geographical landscape nominal scale variables, and the geographical landscape nominal scale variables are used to characterize different types of geographical landscapes; The deep neural network is trained by using the first geochemical database to generate a prediction model. The training process also includes: using the geographical landscape nominal scale variables as a data set to train the deep neural network.

8. The regional geochemical exploration extension method according to claim 1, characterized in that: The method of determining the element content in the second element group in the second geochemical database based on the prediction model and taking the first element group in the second geochemical database as input to generate a geochemical anomaly map under the second sampling density includes: Taking the first element group in the second geochemical database as input, and inputting it into the prediction model, generating geochemical standard values ​​corresponding to each element in the second element group, and restoring them to original values ​​of each element in the second element group; interpolating the original values ​​of each element in the second element group to generate a geochemical map corresponding to the second element group; The geochemical map is processed by taking the mean of the original values ​​of each element in the second element group and the standard deviation of the original values ​​of each element as the anomaly lower limit to generate a geochemical anomaly map.

9. The regional geochemical exploration extension method according to claim 1 or 8, characterized in that: Also includes: Based on the geochemical anomaly map and multivariate information data, the prospecting target area for the target mineral species is determined.

10. A regional geochemical exploration extension device, characterized in that: include: a geochemical database providing unit, adapted to provide a first geochemical database obtained based on a first sampling density and a second geochemical database obtained based on a second sampling density, wherein the types and contents of elements in the first element group and the second element group in the first geochemical database are known quantities, and the contents of elements in the second element group in the second geochemical database are unknown quantities; A neural network construction unit, adapted to construct a deep neural network with a fully connected layer as the main architecture, and configure parameter information of the deep neural network; The processing unit is suitable for training the deep neural network using the first geochemical database to generate a prediction model; and based on the prediction model, taking the first element group in the second geochemical database as input, determining the element content in the second element group in the second geochemical database to generate a geochemical anomaly map under the second sampling density.

Images