Mineral-forming data expansion method based on generative adversarial network

Abstract

The invention provides a generative adversarial network-based metallogenic data expansion method, and belongs to the technical field of metallogenic prediction, and the method comprises the following steps: S1, collecting and preprocessing geochemical remote data of known ore deposits and non-ore deposits, and constructing a metallogenic prediction data set; s2, classifying samples in the metallogenic prediction data set by adopting a hierarchical clustering algorithm, and performing category division; s3, determining a corresponding category weight by calculating the frequency distribution of each category, and carrying out standardization processing on the category weight; s4, constructing a selectable parameter containing a standardized weight in the generative adversarial network model, wherein the selectable parameter is used for transmitting weight information of each category; s5, according to the optional parameters of the standardized weight, generating synthetic data by adopting a weighted sampling strategy in combination with a preset specified condition; and S6, performing quality evaluation on the synthetic data. According to the method, high-quality and diversified metallogenic samples can be generated, and the performance of a mineral resource prediction model is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mineralization prediction, and in particular to a mineralization data expansion method based on a generative adversarial network. Background Art

[0002] Mineralization prediction is an important part of geological exploration. Traditional exploration methods rely on a lot of field work, which is often time-consuming, labor-intensive and costly. The acquisition of a large amount of high-quality mineral sample data is often limited by factors such as geographical location, economic conditions and time. Therefore, mineral resource exploration data often have an imbalanced data set. Specifically, the number of mineralized samples is far less than that of non-mineralized samples. This imbalance has a significant negative impact on the training and prediction accuracy of machine learning models.

[0003] In order to deal with the problem of data imbalance, especially the expansion of mineralization data, many researchers have proposed a series of methods, such as undersampling, oversampling, synthetic minority oversampling technology SMOTE, and adaptive synthetic sampling ADASYN and other integrated learning methods. Although the various methods proposed by predecessors have made some progress, there are still certain challenges in the high complexity and scarcity of mineralization data in mineral exploration. Summary of the invention

[0004] The purpose of the present invention is to provide a mineralization data expansion method based on a generative adversarial network, which can generate high-quality and diversified mineralization samples and significantly improve the performance of the mineral resource prediction model. It is of great significance in the field of mineralization prediction and has broad application prospects.

[0005] To achieve the above object, the present invention provides a method for mineralization data expansion based on a generative adversarial network, comprising the following steps:

[0006] S1. Collect geophysical and chemical remote data of known mineral points and non-mineral points, pre-process the collected geophysical and chemical remote data, and construct a mineralization prediction data set;

[0007] S2. Use a hierarchical clustering algorithm to classify the samples in the mineralization prediction data set and divide the samples into categories;

[0008] S3. Determine the corresponding category weight by calculating the frequency distribution of each category, and standardize the category weight to obtain a standardized weight;

[0009] S4. Construct an optional parameter containing standardized weights in the generative adversarial network model to pass the weight information of each category;

[0010] S5. Generate synthetic data by adopting a weighted sampling strategy based on the optional parameters of the standardized weights and combining preset specified conditions, such as the number of generated data, specific category values, etc.;

[0011] S6. Perform quality assessment on the synthetic data.

[0012] Preferably, in step S1, the geophysical remote sensing data includes regional geological data, remote sensing data, soil geochemical data and geophysical data.

[0013] Preferably, in step S1, preprocessing includes missing value processing and reclassification.

[0014] Preferably, in step S3, the frequency distribution of each category is as follows:

[0015]

[0016] Among them, f i is the frequency distribution of category i; N i is the number of samples for each category i; N total is the total sample size;

[0017] The class weights are set to the frequency distribution of each class:

[0018] w i =f i (2)

[0019] Among them, w i is the category weight;

[0020] The formula for normalizing the category weights is:

[0021]

[0022] Among them, w' i is the normalized class weight; is the sum of all class weights; k is the total number of classes.

[0023] Preferably, in step S4, an optional parameter containing standardized weights is constructed in the generative adversarial network model to transmit weight information of each category, and the specific operation is as follows:

[0024] First, add an optional parameter weights to the constructor of the generative adversarial network model. The optional parameter weights adjusts the generation strategy of the generator and the discrimination strategy of the discriminator according to the standardized weight of each category during the training process.

[0025] Then, adjust the generator and discriminator;

[0026] The generator generates samples based on the input random noise and conditional information. During the training process, the generator uses the optional parameter weights for weighted sampling to generate samples that meet the weight requirements;

[0027] The loss function of the generator is as follows:

[0028]

[0029] Among them, G(z,c i ) is the sample generated by the generator, D is the output of the discriminator, c i is the category condition; L G is the loss function of the generator; is the expected value operator, which represents the average of all possible values ​​of the random noise vector z; z is the random noise vector; p z (z) is the probability distribution of the random noise vector z;

[0030] The discriminator distinguishes between real samples and generated samples;

[0031] The loss function of the discriminator is as follows:

[0032]

[0033] Among them, p data (x) is the true sample distribution; D(x,c i ) is the predicted probability of the discriminator for the real sample x; G(z,c i ) is the sample generated by the generator; L D is the loss function of the discriminator; is the expected value operator for the real sample x, which represents the average of all possible values ​​of the real sample; x is the real sample; p data (x) is the probability distribution of the real data.

[0034] Preferably, in step S6, the quality of the synthesized data is evaluated, and the evaluation indicators used include mean square error, AUC and F1 score.

[0035] Therefore, the present invention adopts the above-mentioned mineralization data expansion method based on generative adversarial network, which can generate high-quality and diversified mineralization samples and significantly improve the performance of mineral resource prediction model. The present invention provides an effective new method for solving the data imbalance problem in mineral resource exploration, which is of great significance in the field of mineralization prediction and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention is a flow chart of a method for mineralization data expansion based on a generative adversarial network. DETAILED DESCRIPTION

[0037] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0038] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0039] Embodiment 1

[0040] In this embodiment, a manganese mining area is taken as the research area, and the weights of the ore-controlling factors of mineralization prediction are determined by using the collected multi-source data such as geophysical, chemical, and remote sensing.

[0041] like Figure 1 As shown, it is a flow chart of a method for mineralization data expansion based on a generative adversarial network of the present invention, which specifically includes the following steps:

[0042] Step S1, collect geophysical and remote sensing data of known mineral points and non-mineral points to construct a model data set. According to the prospecting signs, the regional geological data, remote sensing data, soil geochemical data, and geophysical data are preprocessed by missing value processing, reclassification, etc. Finally, the mineralization element information of the four types of data is extracted into numerical form as features and included in the data set, and a mineralization prediction data set is established based on the mineralization situation of the drilling data. The data is stored in a tabular form, with each row representing a set of geological data and each column representing a feature variable. Among them, a single piece of data has 27-dimensional features, including physical features (XY coordinates, upward extension 2km, etc.), chemical composition (Mn, Cu, etc. and oxides Al2O3, Fe2O3, etc.), geological features (strata, faults, etc.) and remote sensing features. Some of the data are shown in Table 1:

[0043] Table 1 Partial data display

[0044] id X coordinate Y coordinate 2km Mn Cu <![CDATA[Fe2O3]]> <![CDATA[Al2O3]]> … Strata Fault 1 887122 3120015 6 4 4 4 7 … 3 10 2 887452 3121005 5 5 4 4 6 … 11 7 3 887572 3120015 6 4 5 5 7 … 3 1 … … … … … … … … … … … 647 866032 3130005 4 8 3 3 6 … 8 9

[0045] Step S2, using a hierarchical clustering algorithm to classify the existing mineralization samples, and divide the samples into multiple categories. This embodiment adopts agglomerative hierarchical clustering. Each mineralization data point is regarded as an independent clustering object, the distance between clusters is calculated, and the most similar clusters are gradually merged to finally form a single cluster containing all data points. The distance metric used in this embodiment is Euclidean distance. For two points p=(p1, p2, ..., p n ) and q=(q1,q2,…,q n ), the Euclidean distance calculation formula is as follows:

[0046]

[0047] Among them, qi is the coordinate of point q in the i-th dimension; p i is the coordinate of point p in the i-th dimension; d Euclidean Represents the Euclidean distance between two points; i is an index, indicating the dimension currently being calculated.

[0048] Next, two clusters with the smallest distance are selected to be merged to form a new cluster. The merging strategy adopted in this embodiment includes but is not limited to the Ward chain method. Specifically, the core formula of the Ward chain method is as follows:

[0049]

[0050] Among them, x i and x j is a data point in cluster C; |x i -x j | represents the Euclidean distance between the two points. After merging, the distance matrix is ​​updated to reflect the distance of the new cluster. This process is repeated until all mineralization data objects reach the predetermined number of clusters.

[0051] Step S3, by calculating the frequency distribution of each category, determine the corresponding category weight, and standardize the weight. The following explains this process in detail:

[0052] First, we need to count the number of samples in each category in the statistical data set. Suppose there are k categories and the number of samples in each category i is N i , the total number of samples is N total The frequency distribution formula for each category i is:

[0053]

[0054] Among them, f i is the frequency distribution of category i; N i is the number of samples for each category i; N total is the total sample size.

[0055] Secondly, the category weight is determined. In this embodiment, the number of samples of each category in the generated data set is consistent with the category distribution in the actual data set. Category weight w i Can be set to the frequency distribution for each category:

[0056] w i =f i (4)

[0057] Among them, w i is the category weight;

[0058] This way the weight of each class is proportional to its frequency in the original dataset.

[0059] Finally, the weights are normalized. The normalized weights ensure that the sum of the weights of all categories is 1, so that the generative model can sample according to the specified ratio when generating synthetic samples. The formula for normalizing the weights is:

[0060]

[0061] Among them, w' i is the normalized class weight; is the sum of all class weights; k is the total number of classes.

[0062] Step S4, constructing an optional parameter containing standardized weights in the generative adversarial network model to transmit the weight information of each category.

[0063] The first step is to construct the optional parameter weights. In order to introduce standardized weights, an optional parameter weights needs to be added to the constructor of the generative adversarial network model to pass the weight information of each category. This parameter allows the generation strategy of the generator and the discrimination strategy of the discriminator to be adjusted according to the weight of each category during training.

[0064] Then comes the adjustment of the generator and the discriminator. The generator generates samples based on the input random noise and conditional information. During the training process, the generator uses standardized weights for weighted sampling to generate samples that meet the category weight requirements. The loss function of the generator may take into account the category weights to increase the emphasis on the generation of minority category samples. The loss function of the generator can be expressed as:

[0065]

[0066] Among them, G(z,c i ) is the sample generated by the generator; D is the output of the discriminator; c i is the category condition; L G is the loss function of the generator; is the expected value operator, which represents the average of all possible values ​​of the random noise vector z; z is the random noise vector; p z (z) is the probability distribution of the random noise vector z.

[0067] The goal of the discriminator is to distinguish between real samples and generated samples. When considering the class weight, the discriminator's loss function will also include the class weight factor to balance the ability to distinguish samples of different categories. The discriminator's loss function calculation formula is:

[0068]

[0069] Among them, p data (x) is the true sample distribution; D(x, c i ) is the predicted probability of the discriminator for the true sample x; G(z, c i ) is the sample generated by the generator; L D is the loss function of the discriminator; is the expected value operator for the true sample x, representing the average of all possible values of the true sample; x is the true sample data; p data (x) is the probability distribution of the true data.

[0070] Step S5: According to the standardized class weights, adopt a weighted sampling strategy to generate synthetic data with higher balance. The weighted sampling strategy means that when generating synthetic samples, samples are sampled according to the pre-set class weights. Classes with larger weights will be sampled more, while classes with smaller weights will be sampled less. This strategy ensures that the number of samples in each class in the generated synthetic dataset is more in line with the expected ratio, avoiding the situation where some classes in the generated dataset are too many or too few.

[0071] Suppose there are k classes in the dataset, and the standardized weight of each class is w i (i from 1 to k), and the sum of the standardized weights is W: The sampling probability p i of class i is calculated as follows:

[0072]

[0073] Here, p i represents the probability that class i is selected when generating samples.

[0074] When generating synthetic samples, use the calculated sampling probability p i to determine the class of the generated samples. Using the method of random sampling, select the class according to the sampling probability p i . Generate a random number r in the interval [0, 1], and then determine the class according to the standardized weights: if 0 ≤ r < p1, select class 1; if p1 ≤ r < p1 + p2, select class 2; if p1 + p2 + … + p n-1 ≤ r < p1 + p2 + … + p n-1 + p n , then select class n; and so on until a class is selected.

[0075] Step S6: Evaluate the quality of the synthetic data.

[0076] After the generative adversarial network model generates synthetic data, in order to ensure that the generated data can be effectively used as input to the prediction model, its quality must be comprehensively evaluated. The essential purpose of the data augmentation method is to improve the performance of the prediction model. Therefore, the generated data is used to train or test the prediction model and evaluate the performance indicators of the model. In this implementation, MSE, AUC and F1 scores are selected as evaluation indicators to judge the impact of generated data on the prediction effect of the model. The calculation formula for Mean Squared Error (MSE) is:

[0077]

[0078] Among them, y i is the true value; is the model prediction value; m is the number of mineralization samples. The smaller the MSE, the more helpful the data generated by this method is for the model to make accurate predictions. AUC is a measure of the area under the ROC curve, which indicates the overall performance of the model in binary classification problems. The closer the AUC value is to 1, the better the classification effect of the model. The F1 score is the harmonic mean of the model's precision and recall, which is used to weigh the accuracy and coverage of the model. The higher the F1 score, the better the model's ability to identify the positive class. The F1 calculation formula is as follows:

[0079]

[0080] Among them, the precision and recall are: TP is a true positive example; FP is a false positive example; FN is a false negative example.

[0081] Compare the generated dataset and the real dataset using the above indicators. If the corresponding values ​​of the generated data are close to or higher than the real dataset, it means that the generated data has a higher quality in model prediction. If it is significantly lower than the real dataset, the GAN model or data generation process may need to be improved.

[0082] Therefore, the present invention adopts the above-mentioned mineralization data expansion method based on generative adversarial network, which can generate high-quality and diversified mineralization samples and significantly improve the performance of mineral resource prediction model. This study provides an effective new method to solve the data imbalance problem in mineral resource exploration, which is of great significance in the field of mineralization prediction and has broad application prospects.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for mineralization data expansion based on generative adversarial networks, characterized in that: The following steps are involved: S1. Collect geophysical and chemical remote data of known mineral points and non-mineral points, pre-process the collected geophysical and chemical remote data, and construct a mineralization prediction data set; S2. Use hierarchical clustering algorithm to classify samples in the mineralization prediction data set and divide them into categories; S3. Determine the corresponding category weight by calculating the frequency distribution of each category, and standardize the category weight to obtain a standardized weight; S4. Construct an optional parameter containing standardized weights in the generative adversarial network model to pass the weight information of each category; S5. Generate synthetic data by adopting a weighted sampling strategy based on the optional parameters of the standardized weights and combining the preset specified conditions; S6. Perform quality assessment on the synthetic data.

2. A method for mineralization data expansion based on generative adversarial networks according to claim 1, characterized in that: In step S1, the geophysical remote sensing data includes regional geological data, remote sensing data, soil geochemical data and geophysical data.

3. A method for mineralization data expansion based on generative adversarial networks according to claim 2, characterized in that: In step S1, preprocessing includes missing value processing and reclassification.

4. A method for mineralization data expansion based on generative adversarial networks according to claim 3, characterized in that: In step S3, the frequency distribution of each category is as follows: Among them, f i is the frequency distribution of category i; N i is the number of samples for each category i; N total is the total sample size; The class weights are set to the frequency distribution of each class: w i =f i (2) Among them, w i is the category weight; The formula for normalizing the category weights is: Among them, w' i is the normalized class weight; is the sum of all class weights; k is the total number of classes.

5. A method for mineralization data expansion based on generative adversarial networks according to claim 4, characterized in that: In step S4, an optional parameter containing standardized weights is constructed in the generative adversarial network model to transmit the weight information of each category. The specific operation is as follows: First, add an optional parameter weights to the constructor of the generative adversarial network model. The optional parameter weights adjusts the generation strategy of the generator and the discrimination strategy of the discriminator according to the standardized weight of each category during the training process. Then, adjust the generator and discriminator; The generator generates samples based on the input random noise and conditional information. During the training process, the generator uses the optional parameter weights for weighted sampling to generate samples that meet the weight requirements; The loss function of the generator is as follows: Among them, G(z,c i ) is the sample generated by the generator; D is the output of the discriminator; c i is the category condition; L G is the loss function of the generator; is the expected value operator, which represents the average of all possible values ​​of the random noise vector z; z is the random noise vector; p z (z) is the probability distribution of the random noise vector z; The discriminator distinguishes between real samples and generated samples; The loss function of the discriminator is as follows: Among them, p data (x) is the true sample distribution; D(x,c i ) is the predicted probability of the discriminator for the real sample x; G(z,c i ) is the sample generated by the generator; L D is the loss function of the discriminator; is the expected value operator for the real sample x, which represents the average of all possible values ​​of the real sample; x is the real sample; p data (x) is the probability distribution of the real data.

6. A method for mineralization data expansion based on generative adversarial networks according to claim 5, characterized in that: In step S6, the quality of the synthesized data is evaluated, and the evaluation indicators used include mean square error, AUC and F1 score.

Images