Geochemical anomaly identification methods and systems based on causal discovery and deep learning

Abstract

This invention discloses a method and system for geochemical anomaly identification based on causal discovery and deep learning, belonging to the field of geochemical anomaly identification technology. The method includes: constructing a graph structure with sampling points as nodes based on exploration geochemical data, where the node features are the elemental concentration values ​​of the sampling points, and constructing edges based on geographical proximity; performing causal analysis on the node features using a linear non-Gaussian directed acyclic model to generate a causal weight matrix; optimizing the graph structure constructed based on the geochemical data based on the causal weight matrix, and causally adjusting the node features; inputting the causally adjusted graph structure into a graph attention network, and weighting and aggregating node information through an attention mechanism; training the graph attention network model to complete the geochemical anomaly identification. This invention effectively reduces the impact of noisy data on the identification results by removing spurious correlation features and strengthening causal correlation features, exhibiting good robustness and interpretability.

Description

Technical Field

[0001] This invention relates to the field of geochemical anomaly identification technology, and in particular to geochemical anomaly identification methods and systems based on causal discovery and deep learning. Background Technology

[0002] Deep learning possesses strong nonlinear fitting and feature extraction capabilities, enabling it to deeply mine complex patterns hidden in geochemical data. It has become a cutting-edge technology for mining geochemical exploration data to identify mineralization-related geochemical patterns. However, deep learning is a data-driven model, often failing to adequately consider metallogenic theories and expert knowledge. It learns potential patterns solely from training data and primarily focuses on the correlation between elements and ore deposits, neglecting causal relationships. This leads to a series of unresolved issues regarding interpretability, generalization, and robustness. Furthermore, current deep learning models for geochemical anomaly identification are mainly based on single grid cells or pixel blocks (images), making it difficult to fully characterize the spatial pattern features of geochemistry and the spatial coupling relationship between ore deposits and geochemical elements. Graph-based deep learning algorithms input all nodes and their connections into the network, thus better capturing the relationships between training samples and effectively identifying geochemical anomalies. Graph attention networks are an important graph deep neural network method widely used in geochemical anomaly identification and extraction. Data noise may occur during the sampling and preprocessing of geochemical data in exploration. For example, due to limitations in the precision of instrumental analysis, some geochemical data cannot be effectively detected, resulting in missing values. These data interference problems caused by noise and incomplete data pose challenges to the application of deep learning in geochemical anomaly identification, especially for graph neural networks. When neighboring nodes in the graph are disturbed, the model's classification and prediction accuracy will be affected and reduced. Summary of the Invention

[0003] The purpose of this invention is to address the problem that data noise affects the accuracy of graph-based deep learning algorithms in geochemical anomaly identification, and to propose a geochemical anomaly identification method based on causal discovery and deep learning, comprising the following steps:

[0004] S1. Based on exploration geochemical data, construct a graph structure with sampling points as nodes, where the node characteristics are the element concentration values ​​of the sampling points, and construct edges based on geographical proximity.

[0005] S2. A linear non-Gaussian directed acyclic model is used to perform causal analysis on node features and generate a causal weight matrix.

[0006] S3. Optimize the graph structure constructed based on geochemical data based on the causal weight matrix, and make causal adjustments to the node features;

[0007] S4. Input the causal-adjusted graph structure into the graph attention network, and use the attention mechanism to weighted aggregate node information;

[0008] S5. Train the graph attention network model to complete the geochemical anomaly identification.

[0009] Furthermore, in S1, before constructing a graph structure with sampling points as nodes based on exploration geochemical data, the exploration geochemical data is standardized.

[0010] Furthermore, the adjacency matrix is ​​constructed using the Delaunay triangulation method.

[0011] Furthermore, the causal weight matrix is ​​generated by the formula A = Bχ + e, where χ represents the observed variable, B is a lower triangular matrix representing the causal weights between variables, and e is the non-Gaussian noise term.

[0012] Furthermore, a threshold method is used to adjust the causal relationship of node features, eliminating node features with causal weights lower than the set threshold, and assigning different weights to node features with causal weights higher than the set threshold based on the causal weight values.

[0013] Furthermore, the attention mechanism of graph attention networks calculates attention weights using the following formula:

[0014]

[0015] Where, α ij This represents the attention weights between the i-th node and the j-th node. Let represent a trainable parameter vector, and W represent a trainable weight matrix. This represents the feature vector of the i-th node. Let N represent the feature vector of the j-th node, ∥ denote the concatenation operation, LeakyReLU denotes the linear activation function, and N... i Let represent the set of neighboring nodes of the i-th node.

[0016] Furthermore, the training data for the graph attention network model consists of positive samples and negative samples. The positive samples are sampling points of known ore deposits, and the negative samples are random sampling points far away from the positive sample areas, which are about the same number as the positive samples.

[0017] This invention also proposes a geochemical anomaly identification system, comprising:

[0018] The graph structure building unit is used to construct a graph structure with sampling points as nodes based on exploration geochemical data. The node features are the elemental concentration values ​​of the sampling points, and edges are constructed based on geographical proximity.

[0019] The causal analysis unit is used to perform causal analysis on node features using a linear non-Gaussian directed acyclic model and generate a causal weight matrix.

[0020] The causal adjustment unit is used to optimize the graph structure constructed based on geochemical data based on the causal weight matrix and to perform causal adjustment on the node features;

[0021] The node information weighted aggregation unit is used to input the causally adjusted graph structure into the graph attention network and weighted aggregate the node information through the attention mechanism.

[0022] The geochemical anomaly identification unit is used to train the graph attention network model to complete the geochemical anomaly identification.

[0023] The beneficial effects of the technical solution provided by this invention are:

[0024] This invention first constructs a graph structure based on exploration geochemical data, then introduces a causal discovery algorithm to remove spurious correlation features from the geochemical data in the graph structure and strengthens causal correlation features. This effectively reduces the impact of noisy data on the identification results and improves the expressive power of the graph structure. Combined with a graph attention mechanism, node information is weighted and aggregated, improving the model's ability to identify geochemical anomaly patterns. This invention not only possesses good robustness and interpretability but also strong adaptability and generalizability, providing reliable technical support for geological prospecting and mineral resource evaluation. Attached Figure Description

[0025] Figure 1 This is a flowchart of a geochemical anomaly identification method based on causal discovery and deep learning, according to an embodiment of the present invention.

[0026] Figure 2 This is a causal relationship diagram of geochemical element data for a certain region constructed using a causal analysis method in an embodiment of the present invention;

[0027] Figure 3 This is a causal relationship diagram of geochemical element data (using lithium as an example) selected from a certain region in this embodiment of the invention, showing the absolute values ​​of causal weights between elements.

[0028] Figure 4 This is a comparison diagram between the method of this invention and a graph attention network that does not introduce causality. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0030] The flowchart of the geochemical anomaly identification method of this invention is as follows: Figure 1Specifically, it includes the following steps:

[0031] S1. Based on exploration geochemical data, construct a graph structure with sampling points as nodes, where the node characteristics are the element concentration values ​​of the sampling points, and construct edges based on geographical proximity.

[0032] Raw exploration geochemical data requires standardization. Data standardization is a process of converting data of different dimensions or magnitudes into a unified standard to facilitate comparison, analysis, and weighting. The choice of standardization method significantly impacts the analytical results. The following are some commonly used data standardization methods:

[0033] Conventional standardization methods:

[0034] (1) Range standardization: linearly transform the data to the [0,1] interval. This method is simple and easy to implement, but it is sensitive to outliers. If there are extremely large or small values ​​in the data, most of the data may be concentrated in a small interval, which will affect the subsequent analysis.

[0035] (2) Z-score standardization: By subtracting the mean and dividing by the standard deviation, the data is converted into a standard normal distribution. This method can eliminate the differences in the dimensions and magnitudes of the data, making different indicators comparable, but requires the data to follow or approximately follow a normal distribution.

[0036] Geochemical regionalization standardization method:

[0037] Standardizing ore-forming and halo-forming elements according to geochemical zones can effectively suppress non-mineral anomalies in high background areas, enhance mineral-forming anomalies in low background areas, and highlight mineral exploration information. This method is particularly suitable for extracting mineral exploration information from large amounts of regional geochemical data.

[0038] ILR-transformed factor partitioning normalization method:

[0039] First, ILR data transformation is performed on the geochemical data to eliminate the data closure effect. Then, factor analysis is performed. Based on the factor partitioning results, the sample data is partitioned and standardized using EDA. This method can remove the influence of high background values ​​in geological bodies, so that the delineated anomalies correspond well with the spatial location of the deposits and are less affected by human factors.

[0040] The Delaunay triangulation method is used to construct the adjacency matrix. For example, if there are four points A, B, C, and D, two triangles, ABC and BCD, can be constructed with a suitable threshold distance. However, it is necessary to ensure that the circumcircle of any triangle does not contain other points. Adjacency relationships are established based on the edges of the triangles, such as AB connected, BC connected, CD connected, and AD not connected. In this case, 1 represents that the two points are connected and 0 represents that they are not adjacent, and the corresponding matrix is ​​constructed.

[0041] S2. A linear non-Gaussian directed acyclic model (LiNGAM) is used to perform causal analysis on node features and generate a causal weight matrix.

[0042] Specifically, the LiNGAM algorithm assumes that these elements all contain a set of non-Gaussian distributed noise, such as interference caused by errors in the sampling process or measurement. These are then substituted into the LiNGAM causal weight matrix generation formula, A = Bχ + e, where A represents the causal weight matrix, χ is a vector representing the values ​​of all observed variables, B is a lower triangular matrix representing the causal weights between variables, and e is the non-Gaussian noise term. We can then substitute each pair of elements into this formula. Whether a pair of elements can be represented by this formula determines whether an interpretable causal relationship exists between them, and determines the causal direction and the magnitude of the causal weights.

[0043] refer to Figure 2 and Figure 3 , Figure 2 This is a causal relationship diagram of geochemical element data for a certain region constructed using a causal analysis method in an embodiment of the present invention, wherein the arrows indicate the direction of the causal relationship between data variables. Figure 3 This is a causal relationship diagram of geochemical element data, selected from a certain region in this embodiment of the invention, displaying the absolute values ​​of causal weights between elements. The arrows also indicate the causal direction, and the line thickness represents the absolute values ​​of the causal weights between a pair of elements. From Figure 3 As can be seen, there is a strong causal relationship between K2O and Li, indicating that K2O has an important influence on the lithium mineralization process.

[0044] S3. Optimize the graph structure constructed based on geochemical data based on the causal weight matrix and make causal adjustments to the node features.

[0045] A threshold method is used to perform causal adjustment on node features. Node features with causal weights below a set threshold are removed. For node features with causal weights above the set threshold, variables with higher causal weights are assigned greater weight based on their causal weight values. The specific threshold can be set according to the distribution characteristics of the geochemical data. Causal adjustment weakens the influence of spurious correlation features and highlights key variables with strong causal relationships, thereby enhancing the expressive power of key variables in node features. The geochemical map after causal adjustment can more accurately reflect the intrinsic causal relationship between geochemical elements and mineral deposits, eliminate spurious correlation features in the data, and improve the expressive power of the geochemical map.

[0046] S4. Input the causal-adjusted graph structure into the graph attention network, and use the attention mechanism to weight and aggregate the node information to generate an optimized node feature representation.

[0047] The attention mechanism of graph attention networks calculates attention weights using the following formula:

[0048]

[0049] Where, α ij This represents the attention weights between the i-th node and the j-th node. Let represent a trainable parameter vector, and W represent a trainable weight matrix. This represents the feature vector of the i-th node. Let N represent the feature vector of the j-th node, ∥ denote the concatenation operation, LeakyReLU denotes the linear activation function, and N... i This represents the neighboring nodes of the i-th node.

[0050] By using an attention mechanism, the central node performs weighted fusion of information from its neighboring nodes, making key variables more representative in the aggregated feature representation.

[0051] S5. Train the graph attention network model to complete the geochemical anomaly identification.

[0052] The training data for the graph attention network model consists of positive and negative samples. Positive samples are sampling points from known mineral deposits, while negative samples are random sampling points located far from the positive sample areas, with a similar number of negative samples. By repeatedly training and iterating to optimize the graph attention network model, node feature information can be accurately aggregated, effectively improving the accuracy of geochemical anomaly identification and enhancing its resilience to disturbances.

[0053] In one exemplary embodiment, a geochemical anomaly identification system is included, the system comprising:

[0054] The graph structure building unit is used to construct a graph structure with sampling points as nodes based on exploration geochemical data. The node features are the elemental concentration values ​​of the sampling points, and edges are constructed based on geographical proximity.

[0055] The causal analysis unit is used to perform causal analysis on node features using a linear non-Gaussian directed acyclic model and generate a causal weight matrix.

[0056] The causal adjustment unit is used to optimize the graph structure constructed based on geochemical data based on the causal weight matrix and to perform causal adjustment on the node features;

[0057] The node information weighted aggregation unit is used to input the causally adjusted graph structure into the graph attention network and weighted aggregate the node information through the attention mechanism.

[0058] The geochemical anomaly identification unit is used to train the graph attention network model to complete the geochemical anomaly identification.

[0059] To verify the effectiveness of the method of the present invention, in this embodiment, the method of the present invention is compared with a graph attention network that does not introduce causality, and the comparison results are as follows. Figure 4 As shown. From Figure 4 As can be seen from the above, the method of the present invention has strong anti-disturbance ability and high accuracy in identifying geochemical anomalies.

[0060] Through the above steps, this invention achieves the measurement of the causal relationship between geochemical data and mineral deposits, as well as the effective identification and extraction of geochemical spatial patterns. The model not only effectively eliminates spurious correlation features and improves the data's resistance to interference, but also further enhances the identification effectiveness of geochemical anomalies through an optimized graph attention network, providing scientific support for geological prospecting and resource evaluation.

[0061] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying geochemical anomalies based on causal discovery and deep learning, characterized in that, Includes the following steps: S1. Based on exploration geochemical data, construct a graph structure with sampling points as nodes, where the node characteristics are the element concentration values ​​of the sampling points, and construct edges based on geographical proximity. S2. A linear non-Gaussian directed acyclic model is used to perform causal analysis on node features and generate a causal weight matrix. S3. Optimize the graph structure based on the causal weight matrix and adjust the node features causally. S4. Input the causal-adjusted graph structure into the graph attention network, and use the attention mechanism to weighted aggregate node information; S5. Train the graph attention network model to complete the geochemical anomaly identification.

2. The geochemical anomaly identification method based on causal discovery and deep learning according to claim 1, characterized in that, In S1, before constructing a graph structure with sampling points as nodes based on exploration geochemical data, the exploration geochemical data is standardized.

3. The geochemical anomaly identification method based on causal discovery and deep learning according to claim 1, characterized in that, The adjacency matrix is ​​constructed using the Delaunay triangulation method.

4. The geochemical anomaly identification method based on causal discovery and deep learning according to claim 1, characterized in that, The causal weight matrix is ​​generated by the formula A = Bχ + e, where A represents the causal weight matrix, χ represents the observed variable, B is a lower triangular matrix representing the causal weights between variables, and e is the non-Gaussian noise term.

5. The geochemical anomaly identification method based on causal discovery and deep learning according to claim 1, characterized in that, A threshold method is used to adjust the causal relationship of node features, eliminating node features with causal weights lower than the set threshold, and assigning different weights to node features with causal weights higher than the set threshold based on the causal weight values.

6. The geochemical anomaly identification method based on causal discovery and deep learning according to claim 1, characterized in that, The attention mechanism of graph attention networks calculates attention weights using the following formula: Where, α ij This represents the attention weights between the i-th node and the j-th node. Let represent a trainable parameter vector, and W represent a trainable weight matrix. This represents the feature vector of the i-th node. Let N represent the feature vector of the j-th node, ∥ denote the concatenation operation, LeakyReLU denotes the linear activation function, and N... i This represents the neighboring nodes of the i-th node.

7. The geochemical anomaly identification method based on causal discovery and deep learning according to claim 1, characterized in that, The training data for the graph attention network model consists of positive and negative samples. Positive samples are sampling points from known mineral deposits, while negative samples are random sampling points from non-positive sample regions with a number of positive samples that are comparable to the number of positive samples.

8. A geochemical anomaly identification system based on causal discovery and deep learning, characterized in that, include: The graph structure building unit is used to construct a graph structure with sampling points as nodes based on exploration geochemical data. The node features are the elemental concentration values ​​of the sampling points, and edges are constructed based on geographical proximity. The causal analysis unit is used to perform causal analysis on node features using a linear non-Gaussian directed acyclic model and generate a causal weight matrix. The causal adjustment unit is used to optimize the graph structure based on the causal weight matrix and perform causal adjustment on the node features. The node information weighted aggregation unit is used to input the causally adjusted graph structure into the graph attention network and weighted aggregate the node information through the attention mechanism; The geochemical anomaly identification unit is used to train the graph attention network model to complete the geochemical anomaly identification.

Images