Mineral prospective prediction method based on attention mechanism driven graph convolutional network

Abstract

The invention relates to the technical field of mineral prospective prediction, and discloses an attention mechanism driven graph convolutional network mineral prospective prediction method, which comprises the following steps of: after data processing, introducing an attention mechanism on the basis of a traditional graph neural network so as to capture the most important node features. As a mineral prospective prediction method, the complex spatial relationship between remote geological features can be effectively captured, the non-Euclidean spatial characteristics of geological information are fully considered, and the relevance between geological data can be analyzed more accurately. Compared with a traditional method, the method has higher accuracy, reliability and robustness in the aspect of identifying the potential mineral product distribution area, and scientific basis and technical support are provided for subsequent mineral product exploration and resource evaluation work.

Description

Technical Field

[0001] The present invention relates to the field of mineral prospect prediction, and in particular to a graph convolutional network mineral prospect prediction method driven by an attention mechanism. Background Art

[0002] Mineral Prospectivity Mapping (MPM) is an important cornerstone in the field of geosciences, which aims to identify and delineate areas with potential mineral deposits. This complex work integrates multiple data sets, including geological, geochemical, geophysical and remote sensing information, to build predictive models to facilitate mineral exploration and resource assessment. Over time, the evolution of various models has significantly advanced the MPM paradigm, from traditional knowledge-driven methods to more complex data-driven and hybrid models.

[0003] Early MPM models were based on expert knowledge-driven techniques, which prioritized the interpretation of geological information and expert insights to identify potential areas for mineral exploration. Techniques such as weight of evidence provide intuitive ways to integrate exploration information. These methods have limitations in capturing complex relationships and patterns associated with mineralization. Among traditional data-driven methods, although Random Forest (RF) can handle complex geological features and evaluate the importance of features, when dealing with high-dimensional sparse data, each tree tends to fully fit its own subsample, making it difficult to generalize, resulting in poor model performance. Convolutional Neural Network (CNN) can capture patterns related to mineralization through the spatial characteristics of the input image, but due to its fixed convolution kernel size, it may encounter difficulties in dealing with the anisotropy of mineral space, and considering that the mineral space is a non-Euclidean space, it further limits its ability to fully grasp the spatial relationship between geological features.

[0004] Graph Neural Networks (GNNs) have attracted attention mainly due to their adaptability to handle non-Euclidean spatial data. GNNs overcome the limitations of traditional CNNs by capturing complex nonlinear spatial relationships between multiple nodes and their edges. However, current GNN-based MPM methods mainly rely on direct geographic connections to construct graph structures, which simplifies the capture of spatial patterns and correlations in geological data and ignores the connections between geological features that are not geographically adjacent.

[0005] The application number is 202210874998.2, and the invention name is a method for predicting the travel time of open-pit mine trucks based on attention-based spatiotemporal graph convolution. It belongs to the field of mine scheduling. It includes firstly obtaining the original trajectory data and road network data of the open-pit mine truck, combining the external factor attribute information, and obtaining the open-pit mine truck trajectory data and road network data with attribute information; secondly, extracting the factor attribute information features in the trajectory; thirdly, capturing the traffic flow information features in the time and space dimensions through the spatiotemporal graph convolution network based on the attention mechanism; then, capturing the features of the adjacent road sections through the convolution network; then, connecting the above three parts of feature information, designing the prediction model and loss function, and training the model; finally, according to the trajectory to be predicted, using the trained model to predict the travel time. Advantages: Make full use of the trajectory, road network and external information features to achieve accurate prediction of the travel time of open-pit mine trucks.

[0006] The invention mainly uses attention to calculate the weights of the trajectory, road network and external information features within the node, and uses graph convolution to aggregate adjacent node features. However, the invention performs attention calculation on the features within the node, especially in the graph convolution stage, it aggregates the features of adjacent node features. Summary of the invention

[0007] The purpose of the present invention is to solve the defects of the above-mentioned prior art and provide a method for predicting mineral prospects based on a graph convolutional network driven by an attention mechanism. On the basis of the traditional graph neural network, the connection between geological features of non-adjacent geographical locations is calculated through the attention mechanism, which overcomes the inability of CNN and other methods to handle non-Euclidean spatial features, while being able to capture the complex spatial features between more distant geological features and more accurately understand the relationship between geological data.

[0008] The present invention adopts the following technical solution:

[0009] A method for predicting mineral prospects based on graph convolutional networks driven by an attention mechanism, including:

[0010] Step 1. Geoscience component data processing

[0011] Logarithmic ratio transformation of component data, the central logarithmic ratio transformation can map the variable x belonging to the D-dimensional simplex space to the variable y in the D-dimensional vector space in a one-to-one manner, that is:

[0012]

[0013] Among them, y1, ..., y D Represents the mapped variables in the D-dimensional simplex space, x1…x drepresents the variable before mapping in the D-dimensional simplex space, x and y both represent element composition data, g(x) is the geometric mean with x as the variable, clr is the central logarithmic ratio transformation, and the formula is as follows:

[0014]

[0015] Step 2. Perform robust principal component analysis on the data after logarithmic ratio transformation to eliminate the noise between the data. By calculating the principal component factors, it is ensured that each principal component can maximize the intrinsic characteristics of the data. The loss is as follows:

[0016]

[0017] Where I is the identity matrix, V is the orthogonal matrix, V T is the transpose of V;

[0018] Step 3. The m chemical data obtained after processing are interpolated using the inverse distance weighted method in ArcGIS, namely:

[0019]

[0020] Among them, w i Represents the weight of the known point to the point to be interpolated, d i Represents the distance between the known point and the interpolation point

[0021]

[0022] Among them, h(x, y) represents the interpolation of the kth channel of the interpolation point. There are N known points in total, z i is the function value at a known point.

[0023] All known point chemical elements are interpolated to obtain the characteristic matrix H required by the present invention. The components of the characteristic matrix are pixel points, that is, nodes in the graph, the node set is V0, the edge set is E0, and the graph G0 = (V0, E0) is formed. All nodes are connected, that is:

[0024] E0={(ν i ,ν j )|ν i ,ν j ∈V,i≠j}

[0025] Where, ν i ,ν j Represents a node;

[0026] Step 4. Build a graph network neural model

[0027] Using the attention mechanism, the attention coefficient e between node i and node j is calculated according to the characteristics of node i and node j in the graph ij , each node has m+q channels, represented by the feature vector h i and h j , attention score e ij Right now

[0028] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0029] Among them, a is the learnable attention parameter vector, W is the weight matrix, and LeakReLU is the activation function.

[0030] Get the attention score e ij Finally, the attention scores of each node are summed up, and the k nodes with the highest attention scores are selected and retained.

[0031]

[0032] S knodes ={i∈{1,2,…,N}|e i is among the k highest values}

[0033] where e i is the sum of the attention scores of each node, S knodes Select the k nodes with the highest attention coefficients; for these k nodes, select the edges whose attention scores are greater than the threshold θ (that is, as long as the attention score is greater than the average value, it is considered to be an important edge):

[0034] E1={(i,j)|i∈S knodes， j∈{1,2,…,N},e ij >θ}

[0035] E1 is the updated edge set, G1 is the updated graph set, V0 is the node set that has not been updated, and finally a new graph is formed:

[0036] G1=(V0,E1)

[0037] Step 5. Aggregate information from neighboring nodes and update the feature representation of each node. 0 represents the initial feature matrix, H l Represents the feature matrix of the l-th layer node, and uses graph convolution to iteratively update the feature representation of the node:

[0038]

[0039] in To add self-connection to fuse the adjacency matrix of node features, for The degree matrix, W l is the learnable weight matrix of the lth layer, σ is the activation function, and ReLU is used as the activation function here.

[0040] The graph convolution operation aggregates information from neighboring nodes by multiplying the feature matrix with the normalized adjacency matrix. By iteratively applying the graph convolution operation, the feature representation of the model nodes is continuously optimized, deep spatial patterns and correlations are captured, and the potential characteristics of mineral prospects are effectively learned.

[0041] Step 6. Train the model

[0042] The feature matrix after the full connection operation of the features obtained from the graph convolution is H, which is input into a fully connected layer for classification. The output of the fully connected layer is recorded as Y FC :

[0043] Y FC =W fc H+b fc

[0044] Where W fc and b fc They are the weights and biases of the fully connected layer. Next, we use the cross entropy loss function to train the model:

[0045]

[0046] Among them, y i is the true label, is the probability distribution predicted by the model, obtained through the softmax function:

[0047]

[0048] Among them, C is the number of categories. The final loss function is:

[0049]

[0050] Step 7. Use the trained model to predict mineral production

[0051] The trained model is used to predict mineral prospects in areas where it is unknown whether there are mines.

[0052] Furthermore, step 4 also includes: obtaining the attention score e ij Finally, only the spatial information of the most important edges of the most important k nodes is retained, and the edge set must be symmetric before updating the feature representation of the node.

[0053] Furthermore, step 5 also includes: after each node is updated, in addition to saving the current iteration result H l and W l In addition, historical information needs to be stored for the next update, that is, the current attention score and the spatial variables and parameters used in the calculation process are saved so that the model can be quickly updated when new nodes are added.

[0054] Beneficial effects of the present invention:

[0055] The present invention proposes a method for mineral prospect prediction based on a graph convolutional network driven by an attention mechanism, which significantly improves the prediction accuracy compared with traditional convolutional networks, graph attention networks and graph convolutional network methods. In the mineral prospect prediction area, this method can effectively capture the complex spatial features between more distant geological features, breaking through the limitations of traditional methods. Experiments have shown that this method achieves an AUC of 91.7% and a recall rate of 83.6, showing its significant advantage in prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 This is a comparison chart of ROC curves;

[0058] Figure 3 These are empirical diagrams of four types of mineral prediction results: (a) is the prediction diagram of the traditional CNN method; (b) is the prediction diagram of the GCN method; (c) is the prediction diagram of the GAT method; and (d) is the prediction diagram of the method of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] like Figure 1 As shown, a graph convolutional network mineral prospect prediction method based on an attention mechanism driven by the present invention comprises:

[0061] Step 1. Geoscience component data processing

[0062] Logarithmic ratio transformation of component data, the central logarithmic ratio transformation can map the variable x belonging to the D-dimensional simplex space to the variable y in the D-dimensional vector space in a one-to-one manner, that is:

[0063]

[0064] Among them, y1, ..., y D Represents the mapped variables in the D simplex space, x1…x d represents the variable before mapping in the D-dimensional simplex space, x and y both represent element composition data, g(x) is the geometric mean with x as the variable, clr is the central logarithmic ratio transformation, and the formula is as follows:

[0065]

[0066] Step 2. Perform robust principal component analysis on the data after logarithmic ratio transformation to eliminate the noise between the data. By calculating the principal component factors, it is ensured that each principal component can maximize the intrinsic characteristics of the data. The loss is as follows:

[0067]

[0068] Where I is the identity matrix, V is the orthogonal matrix, V T is the transpose of V;

[0069] Step 3. The 39 chemical data obtained after processing are interpolated in ArcGIS using the inverse distance weighted method, that is, the 39 data are interpolated into 28×28×39 raster data using the inverse distance weighted method in ArcGIS, that is:

[0070]

[0071] Among them, w i Represents the weight of the known point to the point to be interpolated, d i Represents the distance between a known point and an interpolated point:

[0072]

[0073] Among them, h(x, y) represents the interpolation of the 39th channel of the interpolation point. There are N known points in total, z i is the function value of the known point. The obtained data is enhanced using a sliding window to enhance each 28×28×39 grid data into 16 24×24×39 grid data.

[0074] The above method is applied to all raster data to obtain the feature matrix H required by the present invention. The components of the feature matrix are pixels, that is, nodes in the graph. The node set is V0, and the edge set is E0. They are combined into a graph G0 = (V0, E0). All nodes are connected, ν i ,ν j Represents a node, namely:

[0075] E0={(ν i,ν j )|ν i ,ν j ∈V,i≠j}

[0076] Step 4. Build a graph network neural model

[0077] Using the attention mechanism, the attention coefficient α between node i and node j is calculated according to the characteristics of node i and node j in the graph ij , each node has 39 channels, represented by the feature vector h i and h j , attention score e ij Right now

[0078] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0079] Among them, a is the learnable attention parameter vector, W is the weight matrix, and LeakReLU is the activation function.

[0080] Get the attention score e ij Finally, the attention scores of each node are summed up, and the k nodes with the highest attention scores are selected and retained.

[0081]

[0082] S knodes ={i∈{1,2,…,N}|e i is among the k highest values}

[0083] where e i is the sum of the attention scores of each node, S knodes Select the k nodes with the highest attention coefficients; for these k nodes, select the edges whose attention scores are greater than the threshold θ (that is, as long as the attention score is greater than the average value, it is considered to be an important edge):

[0084] E1={(i,j)|i∈S knodes ,j∈{1,2,…,N},e ij >θ}

[0085] E1 is the updated edge set, G1 is the updated graph set, V0 is the node set that has not been updated, and finally a new graph is formed:

[0086] G1=(V0,E1)

[0087] Get the attention score e ijFinally, only the spatial information of the most important edges of the most important k nodes is retained, and the edge set must be symmetric before updating the feature representation of the node.

[0088] Step 5. Aggregate information from neighboring nodes and update the feature representation of each node. 0 represents the initial feature matrix, H l Represents the feature matrix of the l-th layer node, and uses graph convolution to iteratively update the feature representation of the node:

[0089]

[0090] in To add self-connection to fuse the adjacency matrix of node features, for The degree matrix, W l is the learnable weight matrix of the lth layer, σ is the activation function, and ReLU is used as the activation function here.

[0091] The graph convolution operation aggregates information from neighboring nodes by multiplying the feature matrix with the normalized adjacency matrix. By iteratively applying the graph convolution operation, the feature representation of the model nodes is continuously optimized, deep spatial patterns and correlations are captured, and the potential characteristics of mineral prospects are effectively learned.

[0092] After each node is updated, in addition to saving the current iteration result H l and W l In addition, historical information needs to be stored for the next update, that is, the current attention score and the spatial variables and parameters used in the calculation process are saved so that the model can be quickly updated when new nodes are added.

[0093] Step 6. Train the model

[0094] The feature matrix after the full connection operation of the features obtained from the graph convolution is H, which is input into a fully connected layer for classification. The output of the fully connected layer is recorded as Y FC :

[0095] Y FC =W fc H+b fc

[0096] Where W fc and b fc They are the weights and biases of the fully connected layer. Next, we use the cross entropy loss function to train the model:

[0097]

[0098] Among them, y i is the true label, is the probability distribution predicted by the model, obtained through the softmax function:

[0099]

[0100] Among them, C is the number of categories. The final loss function is:

[0101]

[0102] Step 7. Use the trained model to predict mineral production

[0103] The trained model is used to predict mineral prospects in areas where it is unknown whether there are mines.

[0104] In the embodiment of the present invention, the final mineral prospect prediction result is obtained by Figure 2 and Figure 3 A detailed presentation was given. Figure 2 The AUC curve comparison of different methods in mineral prospect prediction is shown, including traditional CNN, GCN, GAT methods and the graph convolution network method driven by the attention mechanism of the present invention. Figure 2 It can be clearly seen from the curve in that the method of the present invention shows a higher value in the AUC index than other methods, indicating that it has better accuracy and reliability in mineral prospect prediction.

[0105] Figure 3 It further demonstrates the actual effects of different methods in mineral prospect prediction. Figure 3 (a) is the prediction graph of the traditional CNN method. Figure 3 (b) is the prediction graph of the GCN method. Figure 3 (d) is the prediction diagram of the GAT method. It can be observed that these traditional methods have certain limitations in the identification and positioning of mineral prospects. Figure 3 (d) shows the prediction diagram of the method of the present invention. It can be clearly seen that the method of the present invention can more accurately identify the potential areas of mineral resources, and has a wider coverage and more precise details. This proves that the present invention effectively improves the ability of mineral prospect prediction and has significant advantages.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A graph convolutional network mineral prospect prediction method based on attention mechanism, characterized in that: include: Step 1. Geoscience component data processing The logarithmic ratio transformation of the component data, the central logarithmic ratio transformation maps the variable x belonging to the D-dimensional simplex space to the variable y in the D-dimensional vector space in a one-to-one manner, that is: Among them, y1, ..., y D Represents the mapped variables in the D-dimensional simplex space, x1…x d represents the variable before mapping in the D-dimensional simplex space, x and y both represent element composition data, g(x) is the geometric mean with x as the variable, clr is the central logarithmic ratio transformation, and the formula is as follows: Step 2. Perform robust principal component analysis on the data after logarithmic ratio transformation to eliminate the noise between the data. By calculating the principal component factors, it is ensured that each principal component can maximize the intrinsic characteristics of the data. The loss is as follows: Where I is the identity matrix, V is the orthogonal matrix, V T is the transpose of V; Step 3. The m chemical data obtained after processing are interpolated using the inverse distance weighted method in ArcGIS, namely: Among them, w i represents the weight of the known point to the point to be interpolated, d i Represents the distance between the known point and the interpolation point Among them, h(x, y) represents the interpolation of the kth channel of the interpolation point. There are N known points in total, z i is the function value at a known point; Interpolate all known chemical elements to get the required feature matrix H. The components of the feature matrix are pixels, that is, nodes in the graph. The node set is V0, and the edge set is E0. They are combined into a graph G0 = (V0, E0). Connect all nodes, ν i ,ν j Represents a node, namely: E0={(ν i ,n j )|n i ,n j ∈V,i≠j} Step 4. Build a graph network neural model Using the attention mechanism, the attention coefficient e between node i and node j is calculated according to the characteristics of node i and node j in the graph ij , each node has m+q channels, represented by the feature vector h i and h j , attention score e ij Right now e ij =LeakyReLU(a T [Wh i ||Wh j ]) Among them, a is the learnable attention parameter vector, W is the weight matrix, and LeakReLU is the activation function; Get the attention score e ij Finally, the attention scores of each node are summed up, and the k nodes with the highest attention scores are selected and retained; S knodes ={i∈{1,2,…,N}|e i is among the k highest values} where e i is the sum of the attention scores of each node, S knodes is to select the k nodes with the highest attention coefficients; for these k nodes, select the edges whose attention scores are greater than the threshold θ: E1={(i,j)|i∈S knodes ,j∈{1,2,…,N},e ij >θ} E1 is the updated edge set, G1 is the updated graph set, V0 is the node set that has not been updated, and finally a new graph is formed: G1=(V0,E1) Step 5. Aggregate information from neighboring nodes and update the feature representation of each node, H 0 represents the initial feature matrix, H 1 Represents the feature matrix of the l-th layer node, and uses graph convolution to iteratively update the feature representation of the node: in To add self-connection to fuse the adjacency matrix of node features, for The degree matrix, W l is the learnable weight matrix of the lth layer, σ is the activation function, and ReLU is used as the activation function here; The graph convolution operation aggregates information from neighboring nodes by multiplying the feature matrix with the normalized adjacency matrix. By iteratively applying the graph convolution operation, the feature representation of the model nodes is continuously optimized, deep spatial patterns and correlations are captured, and the potential characteristics of mineral prospects are effectively learned; Step 6. Train the model The feature matrix after the full connection operation of the features obtained from the graph convolution is H, which is input into a fully connected layer for classification. The output of the fully connected layer is recorded as Y FC : Y FC =W fc H+b fc Where W fc and b fc They are the weights and biases of the fully connected layer, and the cross entropy loss function is used to train the model: Among them, y i is the true label, is the probability distribution predicted by the model, obtained through the softmax function: Among them, C is the number of categories, and the final loss function is: Step 7. Use the trained model to predict mineral production The trained model is used to predict mineral prospects in areas where it is unknown whether there are mines.

2. The method for predicting mineral prospects based on graph convolutional networks driven by an attention mechanism as claimed in claim 1, characterized in that: Step 4 also includes: getting the attention score e ij Finally, only the spatial information of the most important edges of the most important k nodes is retained, and the edge set must be symmetric before updating the feature representation of the node.

3. The method for predicting mineral prospects based on graph convolutional networks driven by an attention mechanism as claimed in claim 1, characterized in that: Step 5 also includes: After each node is updated, in addition to saving the current iteration result H 1 and W l In addition, historical information needs to be stored for the next update, that is, the current attention score and the spatial variables and parameters used in the calculation process are saved so that the model can be quickly updated when new nodes are added.

Images