An intelligent evaluation method for mineral resources integrating GAT and GCN

Abstract

The present invention provides a method for intelligent evaluation of mineral resources integrating GAT and GCN. First, the original data set is trained by using the GAT model to obtain a training data set; then, positive and negative sample points with high confidence are selected from the training data set to cover the corresponding sample points in the original data set to form a new training data set to achieve correction of the original data; finally, the GCN model is trained using the new data set to establish a deep learning network model coupling GCN and GAT to achieve intelligent evaluation of mineral resources. An intelligent evaluation device and storage device for mineral resources integrating GCN and GAT are provided to achieve an intelligent evaluation method for mineral resources integrating GAT and GCN. The beneficial effects of the present invention are: intelligent evaluation of mineral resources is achieved to provide guidance for finding the most mineralized areas.

Description

Technical Field

[0001] The present invention relates to the field of mineral exploration, and in particular to an intelligent mineral resource evaluation method integrating GAT and GCN. Background Art

[0002] The quantitative prediction and evaluation of mineral resources requires consideration of a large amount of geological, geochemical, geophysical and remote sensing data, and these geological big data need to be processed comprehensively. In the field of digital and information-based earth sciences, the acquisition of geological big data is more real-time, updated and faster. Geological big data is of great value, but due to its huge data volume, multi-source, multi-category and multi-dimensional characteristics, it has the problems of low value density and difficulty in obtaining useful value. How to accurately and efficiently mine the valuable prospecting information hidden behind the data and provide decision-making services for mineral prediction and prospecting exploration has become a technical challenge for the application of geological big data. Using advanced mathematical geological theories and methods, the multi-source geological spatial information is integrated to draw mineral distribution prediction maps, estimate the predicted resource volume, and then screen the target area, narrow the optimal prospecting area and reduce the exploration risk, which has become an important way to find the most mineralized areas.

[0003] The formation of mineral deposits is controlled by multiple complex mineralization processes, such as the source of mineralizing materials, the migration channels of mineralizing fluids, and the places where mineralizing materials gather, which derive a series of mineralization-controlling variables. These mineralization-controlling variables form complex spatial correlations and nonlinear coupling relationships with mineral deposits. Traditional machine learning models, such as random forests and support vector machines, have limitations in capturing complex nonlinear relationships and processing unstructured data, which makes the evaluation of mineral resources inaccurate. Summary of the invention

[0004] In order to solve the above problems, the present invention provides an intelligent evaluation method for mineral resources that integrates a graph convolutional neural network (GCN) and a graph attention neural network (GAT), which can represent the graph structure between mining-controlling variables in a graphical way, take into account the connection information between nodes, and simultaneously perform information propagation and parameter update between nodes and neighborhoods, so that the intelligent evaluation method for mineral resources can more effectively capture the nonlinear relationship between mining-controlling variables and realize deep-level mining information identification and mining.

[0005] An intelligent mineral resource evaluation method integrating GAT and GCN mainly includes the following steps:

[0006] S1: Use GIS spatial analysis technology to extract mining control variables as node features; set corresponding labels according to the location of the nodes as the basis for classification; thus obtain the initial data set D0 = {N, F, L}, where N represents the node set, F represents the node feature set, L represents the node label set, and N = {n1, n2, ..., n k}, F={f1,f2,...,f k}, L={l1,l2,...,l k},n k represents the kth node, f k represents the feature of the kth node, l k represents the label of the kth node. When the jth node is a mining point, it is used as a positive sample point. j =1, when the jth node is a non-mining point, it is used as a negative sample point, l j = 0, when the label of the jth node is unknown, l j =unkown, j = 1, 2, ..., k, k represents the total number of nodes;

[0007] S2: Use half of the data in D0 as training set D2, and the other half as validation set D3;

[0008] S3: Construct the initial GAT model. According to the training set D2, randomly select equal number of positive and negative sample points to train the initial GAT model. At the same time, evaluate the performance of the trained GAT model according to the validation set D3.

[0009] S4: Use the trained GAT model to train nodes with unknown labels N1 = {n1, n2, ..., n i |l i = unkown} to predict labels and obtain the probability P = {p1, p2, ..., p i |l i =1};

[0010] S5: For positive sample points with a confidence level greater than or equal to a preset confidence level, correspond them to the initial data set D0, and change the labels of the nodes at the mining point locations to 1; correspond the remaining samples, i.e., negative sample points with a high confidence level, to the initial data set D0, and change the labels of the nodes at non-mining point locations to 0; through the above steps, the initial data set is corrected to obtain a new data set D1 = {N', F', L'}, and half of the data is used as the training set D2', and the other half of the data is used as the validation set D3';

[0011] S6: Based on the optimized GAT model, a GCN model is constructed, and an equal number of positive and negative sample points are randomly selected from the training set D2' to train the GCN model until convergence; at the same time, the GCN model is verified using the validation set D3';

[0012] S7: Based on the trained GCN model, the mineral probability is calculated, and the mineral potential prediction map is drawn based on the mineral probability, and the ROC curve is used to evaluate the performance of the GCN model.

[0013] Furthermore, in step S3, the GAT model contains two graph attention layers. The first layer is a multi-head graph attention layer with an input dimension of H (0) ∈R k×29 , the output dimension is H (1) ∈R k×(18×3) , k is the number of nodes; the second layer is a single-head graph attention layer with an input dimension of H (1) ∈R k×(18×3) , the output dimension is P0∈R k×2 , the activation function is ReLu(·).

[0014] Furthermore, in step S3, when the performance of the trained GAT model is evaluated using the validation set D3, the preset accuracy is 70%.

[0015] Furthermore, in step S5, the confidence is preset to 0.99, and nodes with a probability P>=0.99 of the node label being 1 are positive sample points with high confidence, and the nodes of the remaining samples are nodes with a probability P<=0.01 of the node label being 1.

[0016] Furthermore, in step S6, the GCN model includes two graph convolutional layers, and the input dimension of the first graph convolutional layer is H (0) ∈R k×29 , the output dimension is H (1) ∈R k×18 , k is the number of nodes; the input dimension of the second graph convolution layer is H (1) ∈R k×18 , the output dimension is P0∈R k×2 , the activation function is ReLu(·).

[0017] Furthermore, the GCN model uses the Adam optimizer to optimize parameters, where the parameters are the trainable weight matrix W (0) and W (1) , used to transform data dimensions.

[0018] Furthermore, the GCN model is finally normalized and converted into a probability distribution using the softmax(·) function:

[0019]

[0020] Among them, P1 is the probability of mineral production, ReLu(·) is the activation function, and H (0) is the input dimension of the first graph convolutional layer, W (0) and W (1) are all trainable weight matrices, is the adjacency matrix, is the degree matrix.

[0021] A storage device stores instructions and data for implementing an intelligent mineral resource evaluation method integrating GAT and GCN.

[0022] A mineral resource intelligent evaluation device integrating GCN and GAT comprises: a processor and the storage device; the processor loads and executes instructions and data in the storage device to implement a mineral resource intelligent evaluation method integrating GAT and GCN.

[0023] The technical solution provided by the present invention has the following beneficial effects: the present invention uses the GAT model to train the original data to obtain the training set data. Then, positive and negative sample points are respectively selected from the training set data to cover the corresponding sample points in the original data set to form a new data set, thereby achieving the purpose of correcting the original data. Finally, the GCN model is trained using the corrected data set to obtain the prediction results of the training set, and the performance of the GCN model is evaluated using the validation set. The GCN model assigns different weights to different neighbor nodes, which can better capture the relationship between a node and its neighbor nodes, thereby obtaining a more comprehensive and accurate global distribution law of mining points. At the same time, the training convergence speed of the GCN model is faster than that of the GAT model, which can improve the operational efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0025] Figure 1 It is a flow chart of a method for intelligent evaluation of mineral resources integrating GAT and GCN in an embodiment of the present invention.

[0026] Figure 2 Schematic diagram of a graph attention network model in an embodiment of the present invention.

[0027] Figure 3 Schematic diagram of a graph convolutional neural network model in an embodiment of the present invention.

[0028] Figure 4 In the figure, (a) is the loss rate curve and accuracy curve of the training set based on the fusion model GAT-GCN, as well as the accuracy curve of the validation set, and (b) is a schematic diagram of the superposition of the mineralization prediction map and the mineral points based on the fusion model GAT-GCN.

[0029] Figure 5 is a ROC curve diagram of the fusion model GAT-GCN in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0031] Please refer to Figure 1 , Figure 1 It is a flow chart of a method for intelligent evaluation of mineral resources integrating GAT and GCN in an embodiment of the present invention, which specifically includes:

[0032] S1: Use GIS spatial analysis technology to extract mining control variables as node features; set corresponding labels according to the location of the nodes as the basis for classification; thus obtain the initial data set D0 = {N, F, L}, where N represents the node set, F represents the node feature set, L represents the node label set, and N = {n1, n2, ..., n k}, F={f1,f2,...,f k}, L={l1,l2,...,l k},n k represents the kth node, f k represents the feature of the kth node, l k represents the label of the kth node. When the jth node is a mining point, it is used as a positive sample point. j =1, when the jth node is a non-mining point, it is used as a negative sample point, l j = 0, when the label of the jth node is unknown, l j =unkown, j = 1, 2, ..., k, k represents the total number of nodes;

[0033] S2: Use half of the data in D0 as training set D2, and the other half as validation set D3;

[0034] S3: Construct the initial GAT model. According to the training set D2, randomly select equal number of positive and negative sample points to train the initial GAT model. At the same time, evaluate the performance of the trained GAT model according to the validation set D3.

[0035] S4: Use the trained GAT model to train nodes with unknown labels N1 = {N1, N2, ..., N i |l i = unkown} to predict labels and obtain the probability P = {p1, p2, ..., p i |l i =1};

[0036] S5: For positive sample points with high confidence, correspond them to the initial data set D0, and change the labels of nodes at the mining point position to 1; correspond the remaining samples, that is, negative sample points with high confidence, to the initial data set D0, and change the labels of nodes at non-mining point positions to 0; through the above steps, the initial data set is corrected to obtain a new data set D1 = {N', F', L'}, and half of the data is used as the training set D2', and the other half of the data is used as the verification set D3';

[0037] S6: Based on the optimized GAT model, a GCN model is constructed, and an equal number of positive and negative sample points are randomly selected from the training set D2' to train the GCN model until convergence; at the same time, the GCN model is verified using the validation set D3';

[0038] S7: Based on the trained GCN model, the mineral probability is calculated, and the mineral potential prediction map is drawn based on the mineral probability, and the ROC curve is used to evaluate the performance of the GCN model.

[0039] This example uses the Nanling tungsten polymetallic mineralization belt as the research area, and selects a total of 29 variables as graph node features. This method takes into account the spatial correlation between multiple ore-controlling factors, fully utilizes the advantages of the graph topology, and can improve the accuracy of prediction.

[0040] (1) Set grid nodes covering the entire study area and extract the eigenvalues ​​of 29 variables to the grid nodes to obtain the initial data set D0.

[0041] (2) The study area is divided into two equal parts, the lower half as the training set D2, and the upper half as the validation set D3. 80% of the positive sample points (mineral points) and an equal number of negative sample points (non-mineral points) are randomly selected from the two sub-areas.

[0042] (3) Construct Figure 2 The GAT model shown in the figure contains two graph attention layers. The first layer is a multi-head (the number of attention heads M = 3) graph attention layer with an input dimension of H (0) ∈R k×29 (k is the number of nodes), the output dimension is H (1) ∈R k×(18×3) ; The second layer is a single-head (number of attention heads M = 1) graph attention layer, with an input dimension of H (1) ∈R k×(18×3) , the output dimension is P0∈R k ×2, the activation function is ReLu(·), the learning rate is 0.003, and the cross-entropy loss function CrossEntropyLoss(·) is selected to calculate the training loss. To prevent overfitting, the Dropout rate of the graph attention layer is set to p = 0.5, and the e2 regularization coefficient is 0.002. The training round is set to 200 times, and the Adam optimizer is used for parameter Optimization, where and is the attention vector, which is used to give different attention weights to neighbor nodes. (0) and W (1) They are all trainable weight matrices, which will be updated with the training rounds and used to transform the data dimension. Finally, softmax(·) is used to normalize and convert them into probability distribution.

[0043] (4) The GAT model is trained for 200 rounds and the performance of the GAT model is evaluated using the validation set D3. If the final accuracy on the validation set is lower than 70%, the parameters of the GAT model are adjusted and retrained until the validation accuracy of the GAT model is equal to or greater than 70%, otherwise, step (5) is executed.

[0044] (5) Use the trained GAT model to train the unknown label nodes N1 = {n1, n2, ..., n i |l i = unkown} to predict the label and obtain the probability P = {p1, p2, ..., p i |l i =1}. For nodes with probability P>=0.99, i.e., positive sample points with high confidence, they are mapped to the initial data set D0, and the node label in the initial data set is changed to 1; for the remaining samples, i.e., negative sample points with high confidence of P<=0.01, similar operations are performed, and the node labels in the corresponding initial data set are changed to 0 (the mining point does not exist).

[0045] (6) Through the above step (5), the correction operation of the initial data set D0 is implemented to obtain a new data set D1 = {N', F', L'}, and half of the data is used as the training set D2', and the other half of the data is used as the verification set D3'.

[0046] (7) Construct Figure 3 The GCN model shown in the figure contains two graph convolutional layers. The input dimension of the first graph convolutional layer is H. (0) ∈R k×29 (k is the number of nodes), the output dimension is H (1) ∈R k×18; The input dimension of the second graph convolution layer is H (1) ∈R k×18 , the output dimension is P0∈R k×2 The activation function is ReLu(·), the learning rate is 0.003, and the cross-entropy loss function CrossEntropyLoss(·) is selected to calculate the training loss. To prevent overfitting, the dropout rate of the Dropout layer is set to p = 0.5, and the l2 regularization coefficient is 0.002. The training round is set to 1000 times, and the Adam optimizer is used for parameter (W (0) , W (1) ) optimization, W (0) and W (1) They are all trainable weight matrices, which will be updated with the training rounds and used to transform the data dimension. Finally, softmax(·) is used to normalize and convert them into probability distribution. The whole process can be expressed by the following formula:

[0047]

[0048] Among them, P1 is the probability of mineral production, ReLu(·) is the activation function, and H (i) is the input data of the (i+1)th graph convolutional layer, H (0) is the input data of the first graph convolution layer, is the adjacency matrix, is the degree matrix.

[0049] (8) The GCN model is formally trained for 1000 rounds using the training set D2' until the training results converge. The trained GCN model is then verified for accuracy using the validation set D3'. Figure 4 The training loss and accuracy as well as the prospect map of mineralization are shown. Figure 5 As shown in the figure, the AUC value (0.862) of the ROC curve shows that the GCN model has good performance in the intelligent evaluation of mineral resources.

[0050] In order to avoid accidental errors in the experiment, the GAT-GCN model used 5 random seeds to conduct 5 identical experiments, and finally took the average of the 5 experiments as the experimental result of the GAT model.

[0051] In order to improve the accuracy of the model, two attempts are needed when dividing the training set and the validation set, namely: a. Use the upper half of the image as the training set and the lower half as the validation set; b. Use the upper half of the image as the validation set and the lower half as the training set. Select the one with higher accuracy from the two.

[0052] A storage device stores instructions and data for implementing an intelligent mineral resource evaluation method integrating GAT and GCN.

[0053] A mineral resource intelligent evaluation device integrating GCN and GAT comprises: a processor and the storage device; the processor loads and executes instructions and data in the storage device to implement a mineral resource intelligent evaluation method integrating GAT and GCN.

[0054] The beneficial effect of the present invention is as follows: the present invention uses the GAT model to train the original data to obtain the training set data. Then, positive and negative sample points are respectively selected from the training set data to cover the corresponding sample points in the original data set to form a new data set, thereby achieving the purpose of correcting the original data. Finally, the GCN model is trained using the corrected data set to obtain the prediction results of the training set, and the performance of the GCN model is evaluated using the validation set. The GCN model assigns different weights to different neighbor nodes, which can better capture the relationship between a node and its neighbor nodes, thereby obtaining a more comprehensive and accurate global distribution law of mining points. At the same time, the training convergence speed of the GCN model is faster than that of the GAT model, which can improve the operational efficiency of the model.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent mineral resource evaluation method integrating GAT and GCN, characterized in that: include: S1: Use GIS spatial analysis technology to extract mining control variables as node features; The corresponding labels are set according to the positions of the nodes as the basis for classification. Thus, the initial data set D0 = {N, F, L} is obtained, where N represents the node set, F represents the node feature set, L represents the node label set, and N = {n1, n2, ..., n k }, F={f1,f2,...,f k }, L={l1,l2,...,l k },n k represents the kth node, f k represents the feature of the kth node, l k represents the label of the kth node. When the jth node is a mining point, it is used as a positive sample point. j =1, when the jth node is a non-mining point, it is used as a negative sample point, l j = 0, when the label of the jth node is unknown, l j =unkown, j = 1, 2, ..., k, k represents the total number of nodes; S2: Use half of the data in D0 as training set D2, and the other half as validation set D3; S3: Construct the initial GAT model. According to the training set D2, randomly select equal number of positive and negative sample points to train the initial GAT model. At the same time, evaluate the performance of the trained GAT model according to the validation set D3. S4: Use the trained GAT model to train nodes with unknown labels N1 = {N1, N2, ..., N i |l i = unkown} to predict labels and obtain the probability P = {p1, p2, ..., p i |l i =1}; S5: For positive sample points with a confidence level greater than or equal to a preset confidence level, correspond them to the initial data set D0, and change the label of the node at the mining point location to 1; The remaining samples are mapped to the initial data set D0, and the labels of the nodes at non-mining point locations are changed to 0; through the above steps, the initial data set is corrected to obtain a new data set D1 = {N', F', L'}, and half of the data is used as the training set D2', and the other half of the data is used as the validation set D3'; S6: Based on the optimized GAT model, a GCN model is constructed. An equal number of positive and negative sample points are randomly selected from the training set D2' to train the GCN model until convergence. At the same time, the GCN model is verified using the validation set D3'. S7: According to the trained GCN model, the mineral probability is calculated, and a mineral potential prediction map is drawn based on the mineral probability to achieve intelligent evaluation of mineral resources.

2. The method for intelligent evaluation of mineral resources integrating GAT and GCN as claimed in claim 1, characterized in that: In step S3, the GAT model contains two graph attention layers. The first layer is a multi-head graph attention layer with an input dimension of H (0) ∈R k×29 , the output dimension is H (1) ∈R k×(18×3) , k is the number of nodes; the second layer is a single-head graph attention layer with an input dimension of H (1) ∈R k×(18×3) , the output dimension is P0∈R k×2 , the activation function is ReLu(·).

3. The intelligent mineral resource evaluation method integrating GAT and GCN as claimed in claim 1, characterized in that: In step S3, when the performance of the trained GAT model is evaluated using the validation set D3, the preset accuracy is 70%.

4. The method for intelligent evaluation of mineral resources integrating GAT and GCN as claimed in claim 1, characterized in that: In step S5, the confidence is preset to 0.99, and nodes with a probability P>=0.99 of the node label being 1 are positive sample points with high confidence, and the nodes of the remaining samples are nodes with a probability P<=0.01 of the node label being 1.

5. The method for intelligent evaluation of mineral resources integrating GAT and GCN as claimed in claim 1, characterized in that: In step S6, the GCN model contains two graph convolutional layers. The input dimension of the first graph convolutional layer is H (0) ∈R k×29 , the output dimension is H (1) ∈R k×18 , k is the number of nodes; the input dimension of the second graph convolution layer is H (1) ∈R k×18 , the output dimension is P0∈R k×2 , the activation function is ReLu(·).

6. The method for intelligent evaluation of mineral resources integrating GAT and GCN as claimed in claim 1, characterized in that: The GCN model uses the Adam optimizer for parameter optimization, where the parameter is a trainable weight matrix W. (0) and W (1) , used to transform data dimensions.

7. The method for intelligent evaluation of mineral resources integrating GAT and GCN as claimed in claim 1, characterized in that: The GCN model is finally normalized using the softmax(·) function and converted into a probability distribution: Among them, P1 is the probability of mineral production, ReLu(·) is the activation function, and H (0) is the input data of the first graph convolutional layer, W (0) and W (1) are all trainable weight matrices, is the adjacency matrix, is the degree matrix.

8. A storage device, characterized in that: The storage device stores instructions and data for implementing the intelligent mineral resource evaluation method integrating GAT and GCN as described in any one of claims 1 to 7.

9. An intelligent mineral resource evaluation device integrating GCN and GAT, characterized in that: include: Processor and storage device; the processor loads and executes instructions and data in the storage device to implement the intelligent mineral resource evaluation method integrating GAT and GCN as described in any one of claims 1 to 7.

Images