Mineral deposit prediction method, device and storage medium based on deep learning

Abstract

The present application relates to the field of data processing technology and discloses a method, device, and storage medium for mineral prediction based on deep learning. The method comprises: obtaining a standardized data set by processing multi-source geological data through data correction and normalization, extracting and fusing features using a deep learning model, constructing a mineralization knowledge graph to obtain knowledge vectors, constructing a prediction model based on a Transformer network and a Bayesian framework to obtain prediction results, and after spatial optimization, using a deep generative model to construct a three-dimensional visualization model and perform explanatory analysis to obtain prediction verification results. The present application can effectively integrate multi-source geological data, simultaneously complete multiple prediction tasks, and ensure that the prediction results are both consistent with geological laws and have high accuracy.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a mineral deposit prediction method, device and storage medium based on deep learning. Background Art

[0002] Existing mineral deposit prediction methods primarily include traditional geostatistical methods and machine learning-based methods. Traditional geostatistical methods perform spatial analysis for mineralization prediction through methods such as kriging interpolation and multi-point geostatistics, but these methods overly rely on expert experience and prior knowledge. Machine learning methods, such as support vector machines and random forest algorithms, have been introduced into the field of mineral deposit prediction. These methods predict mineralization potential by mapping geological features to mineralization. These methods are capable of processing multi-source geological data and automatically learning feature representations.

[0003] However, existing technologies have the following problems: First, current methods often process different types of geological data (such as remote sensing images, geophysical and geochemical data) separately, and lack an effective multi-source data fusion mechanism; second, traditional machine learning methods have limited ability to extract data features, making it difficult to capture complex geological characteristics and mineralization laws; third, existing prediction models usually only focus on a single prediction task (such as mineralization type prediction) and cannot handle multiple related prediction tasks at the same time; fourth, the prediction results lack interpretability, making it difficult to provide geologists with an intuitive prediction basis; finally, the constraints of geological knowledge are not fully considered in the prediction process, resulting in the prediction results possibly violating basic geological laws. Summary of the Invention

[0004] The present application provides a mineral prediction method, device and storage medium based on deep learning, which are used to effectively integrate multi-source geological data, complete multiple prediction tasks simultaneously, and make the prediction results conform to geological laws and have high accuracy.

[0005] In the first aspect, the present application provides a mineral deposit prediction method based on deep learning, which includes: preprocessing remote sensing image data, geophysical exploration data and geochemical data through data correction and normalization algorithms to obtain a standardized multi-source data set; based on the standardized multi-source data set, constructing a feature extraction model through an improved ResNet structure, a three-dimensional convolutional neural network and a graph neural network, and using a variational autoencoder to perform feature fusion to obtain deep coding features; based on the deep coding features, constructing a mineralization knowledge graph through a multi-head self-attention mechanism and a message passing mechanism, and obtaining a knowledge vector using a graph embedding algorithm; based on the deep coding features and the knowledge vector, constructing a multi-task prediction model through a Transformer network, and performing uncertainty assessment through a Bayesian neural network framework to obtain a mineralization prediction result; based on the mineralization prediction result, iterative optimization is performed through spatial clustering and Bayesian optimization algorithms to obtain an optimized prediction result; based on the optimized prediction result, a three-dimensional visualization model is constructed using a deep generative model, and explanatory analysis is performed through the Grad-CAM algorithm to obtain a target prediction verification result.

[0006] In a second aspect, the present application provides a mineral deposit prediction device based on deep learning, the mineral deposit prediction device based on deep learning comprising:

[0007] The processing module is used to pre-process remote sensing image data, geophysical exploration data and geochemical data through data correction and normalization algorithms to obtain standardized multi-source data sets;

[0008] A fusion module is used to construct a feature extraction model based on the standardized multi-source dataset through an improved ResNet structure, a three-dimensional convolutional neural network, and a graph neural network, and to perform feature fusion using a variational autoencoder to obtain deep coding features;

[0009] A construction module is used to construct a metallogenic knowledge graph based on the deep encoding features through a multi-head self-attention mechanism and a message passing mechanism, and obtain a knowledge vector using a graph embedding algorithm;

[0010] An evaluation module is used to construct a multi-task prediction model through a Transformer network based on the deep coding features and the knowledge vector, and perform uncertainty evaluation through a Bayesian neural network framework to obtain a mineralization prediction result;

[0011] An optimization module is used to perform iterative optimization based on the mineralization prediction result through spatial clustering and Bayesian optimization algorithm to obtain an optimized prediction result;

[0012] The analysis module is used to construct a three-dimensional visualization model based on the optimized prediction results using a deep generative model, and perform explanatory analysis through the Grad-CAM algorithm to obtain target prediction verification results.

[0013] A third aspect of the present application provides a computer-readable storage medium, which stores instructions that, when executed on a computer, enable the computer to execute the above-mentioned deep learning-based mineral prediction method.

[0014] In the technical solution provided by this application, multi-source geological data are pre-processed through data correction and normalization algorithms to ensure the standardization and consistency of different types of data and improve the accuracy of subsequent feature extraction. The feature extraction model constructed by using an improved ResNet structure, a three-dimensional convolutional neural network and a graph neural network can perform deep feature extraction based on the characteristics of remote sensing images, geophysical exploration and geochemical data, respectively, while the feature fusion mechanism of the variational autoencoder realizes the effective integration of multi-source data. The mineralization knowledge graph is constructed using a multi-head self-attention mechanism and a message passing mechanism, which fully explores the correlation between geological elements. The knowledge vector obtained by the graph embedding algorithm effectively encodes complex geological relationships. The multi-task prediction model constructed based on the Transformer network realizes the joint prediction of mineralization type, scale and grade, and the introduction of the Bayesian neural network framework provides uncertainty assessment of the prediction results. The iterative optimization process of spatial clustering and Bayesian optimization algorithm improves the spatial continuity and geological rationality of the prediction results, making the prediction results more in line with actual geological laws. Finally, a three-dimensional visualization model was constructed through a deep generative model and the Grad-CAM algorithm was used for explanatory analysis, which not only provided an intuitive display of the prediction results, but also revealed the key basis in the prediction process, enhanced the interpretability and credibility of the prediction results, and improved the accuracy, reliability and practicality of mineral prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 This is a schematic diagram of an embodiment of a mineral deposit prediction method based on deep learning in an embodiment of the present application;

[0017] Figure 2 A schematic diagram of a node-edge relationship structure in an embodiment of the present application;

[0018] Figure 3Schematic diagram of the mineralization knowledge map in the embodiment of this application;

[0019] Figure 4 This is a schematic diagram of an embodiment of a mineral deposit prediction device based on deep learning in an embodiment of the present application. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a method, device and storage medium for mineral prediction based on deep learning. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or apparatus.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the mineral deposit prediction method based on deep learning includes:

[0022] Step S101: pre-processing remote sensing image data, geophysical exploration data, and geochemical data through data correction and normalization algorithms to obtain a standardized multi-source data set;

[0023] Step S102: Based on the standardized multi-source dataset, a feature extraction model is constructed by improving the ResNet structure, a three-dimensional convolutional neural network, and a graph neural network, and feature fusion is performed using a variational autoencoder to obtain deep coding features;

[0024] Step S103: Based on the deep coding features, a metallogenic knowledge graph is constructed through a multi-head self-attention mechanism and a message passing mechanism, and a knowledge vector is obtained using a graph embedding algorithm;

[0025] Step S104: Based on the deep coding features and knowledge vectors, a multi-task prediction model is constructed through the Transformer network, and uncertainty assessment is performed through the Bayesian neural network framework to obtain a mineralization prediction result;

[0026] Step S105: Based on the mineralization prediction result, iterative optimization is performed through spatial clustering and Bayesian optimization algorithm to obtain an optimized prediction result;

[0027] Step S106: Based on the optimized prediction results, a three-dimensional visualization model is constructed using a deep generative model, and an explanatory analysis is performed using the Grad-CAM algorithm to obtain a target prediction verification result.

[0028] It is understandable that the execution subject of this application can be a mineral deposit prediction device based on deep learning, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0029] Specifically, it starts with multi-source data preprocessing. For remote sensing image data, atmospheric correction algorithms are used to remove the effects of atmospheric scattering and absorption. Specifically, the original remote sensing data is subjected to radiometric calibration and atmospheric optical thickness calculations. Surface reflectance data is then jointly inverted with atmospheric parameters to eliminate atmospheric interference with the spectral characteristics of ground objects. Geometric correction is then performed to establish a correspondence between ground control points and image points. Terrain correction is performed using a digital elevation model, and the image data is unified into a standard coordinate system. Geophysical exploration data processing includes noise removal for gravity and magnetic data. Median filtering is used to remove random noise and sudden changes, and the data is mapped to the [0,1] interval using minimum-maximum normalization. For geochemical data, outliers are identified using boxplots, the interquartile range is calculated, data points outside the normal range are marked as abnormal, and missing and abnormal data are filled using interpolation.

[0030] During the feature extraction and fusion phase, standardized remote sensing image data is fed into an improved ResNet architecture, where spatial features are extracted through convolutional layers and residual connections. Specifically, each convolutional layer applies a sliding window operation to the input data, extracting local feature maps. Residual connections preserve the original feature information. Geophysical exploration data is processed using a three-dimensional convolutional neural network, performing hierarchical feature extraction on the 3D data volume to capture geophysical anomalies at different depth levels. Geochemical data is constructed into a node-edge relationship structure, where information is aggregated using a graph neural network. Nodes represent elemental content, and edges represent correlations between elements. All extracted features are fused using a variational autoencoder. The encoder compresses multi-source features into a latent space, while the decoder reconstructs the features to ensure information preservation. During the knowledge graph construction phase, multi-head self-attention computation is performed on the deeply encoded features, dividing the feature space into multiple subspaces for parallel processing and calculating the association weights between different features. Feature information is transferred and updated between nodes through a message passing mechanism, constructing a mineralization knowledge graph, where nodes represent geological elements and edges represent relationships between elements. The graph embedding algorithm generates node sequences through random walk sampling and converts graph structure data into vector representation.

[0031] During the multi-task prediction model construction phase, deep encoding features and knowledge vectors are concatenated and fed into a Transformer network, capturing sequential features through positional encoding and self-attention mechanisms. The model comprises a shared encoding layer and dedicated prediction branches tailored to mineralization type, scale, and grade. A Bayesian neural network framework estimates the uncertainty of predictions through sampling, generating a distribution of predictions rather than a single value. During the iterative optimization phase, spatial clustering is used to identify high-probability prediction areas, and a spatial distance matrix is calculated between these areas to group them. A Bayesian optimization algorithm optimizes prediction parameters through sampling point selection and objective function evaluation, continuously updating and supplementing exploration locations.

[0032] During the visualization verification phase, the optimized prediction results are reconstructed and voxelized in 3D, and a 3D visualization representation is constructed using a deep generative model. The Grad-CAM algorithm performs interpretive analysis on the model, calculating gradients and activation maps at the feature layer to identify regions that are crucial to the prediction results.

[0033] For example, remote sensing imagery collected in a study area increased surface reflectivity by 15% after removing atmospheric influences, and the signal-to-noise ratio of gravity anomaly data increased from the original 3:1 to 8:1 after filtering. The fused depth features captured the spatial correlations between geological elements such as faults, rock masses, and alteration zones, resulting in a knowledge graph containing over 500 geological element nodes and over 1,000 relationship edges. The multi-task prediction model simultaneously outputs predictions for mineralization type, scale, and grade, with a prediction uncertainty interval estimated to be ±10% using a Bayesian framework. Three high-potential target areas were identified during the iterative optimization process, and after additional exploration verification, the predictions achieved an 85% agreement with the actual situation. A 3D visualization model clearly illustrates the spatial distribution of mineralization predictions, and Grad-CAM analysis indicates that fault intersection zones and alteration halos contribute most significantly to the predictions.

[0034] In the embodiment of the present application, multi-source geological data is pre-processed by data correction and normalization algorithms to ensure the standardization and consistency of different types of data and improve the accuracy of subsequent feature extraction. The feature extraction model constructed by using improved ResNet structure, three-dimensional convolutional neural network and graph neural network can perform deep feature extraction based on the characteristics of remote sensing images, geophysical exploration and geochemical data, respectively, while the feature fusion mechanism of variational autoencoder realizes the effective integration of multi-source data. The mineralization knowledge map is constructed using multi-head self-attention mechanism and message passing mechanism, which fully explores the correlation between geological elements. The knowledge vector obtained by the graph embedding algorithm effectively encodes complex geological relationships. The multi-task prediction model constructed based on Transformer network realizes the joint prediction of mineralization type, scale and grade, and the introduction of Bayesian neural network framework provides uncertainty assessment of prediction results. The iterative optimization process of spatial clustering and Bayesian optimization algorithm improves the spatial continuity and geological rationality of the prediction results, making the prediction results more in line with actual geological laws. Finally, a three-dimensional visualization model was constructed through a deep generative model and the Grad-CAM algorithm was used for explanatory analysis, which not only provided an intuitive display of the prediction results, but also revealed the key basis in the prediction process, enhanced the interpretability and credibility of the prediction results, and improved the accuracy, reliability and practicality of mineral prediction.

[0035] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0036] (1) Through atmospheric correction, atmospheric scattering removal and radiation correction are performed on remote sensing image data to obtain preliminary corrected remote sensing image data;

[0037] (2) Based on the preliminary corrected remote sensing image data, geometric distortion correction and geographic coordinate registration are performed through the digital elevation model to obtain the corrected remote sensing image data;

[0038] (3) The geophysical exploration data are divided into gravity data and magnetic data according to the data type. The gravity data and magnetic data are noise-removed by median filtering to obtain the de-noised geophysical data.

[0039] (4) Based on the de-noised geophysical data, perform numerical normalization using the minimum-maximum method to obtain normalized geophysical data;

[0040] (5) Based on the element content data in the geochemical data, the outliers are identified using the box plot method and the data are filled in using the interpolation method to obtain the processed geochemical data;

[0041] (6) The corrected remote sensing image data, normalized geophysical data and processed geochemical data are fused to obtain a standardized multi-source data set.

[0042] Specifically, atmospheric scattering primarily includes Rayleigh scattering and Mie scattering. Rayleigh scattering is caused by atmospheric molecules, and its scattering intensity is inversely proportional to the fourth power of the wavelength, primarily affecting shortwave bands. Mie scattering is caused by aerosols, and its scattering intensity is approximately inversely proportional to the wavelength. Atmospheric correction involves establishing a radiative transfer equation to calculate the relationship between the radiance at the top of the atmosphere and the reflectance of the surface. Radiometric correction converts image grayscale values into physical radiometric quantities and involves two steps: sensor calibration and solar irradiance correction. Sensor calibration converts digital numbers (DN values) into radiance values. Solar irradiance correction considers the effects of the solar zenith angle and the distance between the sun and the earth. Digital elevation models (DEMs) are used for geometric distortion correction, primarily to address image distortion caused by terrain. During processing, a network of ground control points is established, each containing both image and field coordinates. The DEM data is used to calculate the terrain elevation for each pixel, and combined with the sensor imaging geometry model, a correspondence between pixels in image space and geographic space is established. The corrected image is reprojected into the standard geographic coordinate system through the resampling method, and the grayscale value of the resampled pixel is determined by bilinear interpolation or cubic convolution.

[0043] Geophysical exploration data is first processed by classifying it into gravity and magnetic data. Gravity data reflects the distribution of underground density, while magnetic data reflects the distribution of magnetic minerals. Both types of data are then subjected to median filtering. Median filtering is a nonlinear filtering method that takes the median value of all points in its neighborhood, sorted by size, for each data point. This method effectively removes impulse noise and effectively preserves the edge characteristics of the data. For one-dimensional time series data, the sliding window size is set to an odd number to ensure the uniqueness of the median. Data normalization uses the minimum-maximum method to map data of different dimensions to a unified interval. For each data point x, its normalized value y is calculated: y = (x-min) / (max-min), where min and max are the minimum and maximum values of the data series, respectively. This processing method preserves the relative distribution characteristics of the data and facilitates comprehensive analysis of data from different physical quantities.

[0044] Outliers in geochemical data are identified using a boxplot method. The quartiles Q1 (25% and 75% of the data) and Q3 (interquartile range IQR) are calculated, with the interquartile range (IQR) being Q3-Q1. Data points outside the interval [Q1-1.5IQR, Q3+1.5IQR] are marked as outliers. These outliers and missing values are imputed using distance-weighted inverse distance interpolation. This interpolation considers valid data points that are spatially close, with weights inversely proportional to distance.

[0045] The data fusion phase integrates the three processed data types, unifying their spatial resolution and coordinate system. A regular grid is established, and all data types are resampled onto the same grid. A weighted average method is used for data synthesis, with weight coefficients determined based on the reliability of each data type and its contribution to mineralization prediction.

[0046] For example, during atmospheric correction of remote sensing data for a specific region, atmospheric optical depth (AOD) is calculated using the radiation transfer equation. This DN value is then converted to surface reflectance using sensor calibration parameters. During topographic correction, a control point network is established using 20-meter resolution DEM data for geometric correction. Gravity data exhibits random noise, but a five-point median filter significantly suppresses this noise, making geological anomalies more distinct. Magnetic data is normalized to a range of [0,1]. Geochemical data with copper content exceeding the normal range is spatially interpolated to obtain reasonable estimates. The resulting standardized multi-source dataset exhibits good spatial consistency, fully integrating various geological information.

[0047] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0048] (1) The remote sensing image data in the standardized multi-source dataset is input into the improved ResNet structure, the spatial features are extracted through the convolution layer, and the gradient is transferred through the residual connection to obtain the surface feature data;

[0049] (2) Reconstruct the geophysical exploration data in the standardized multi-source data set into three-dimensional data, and extract hierarchical features through a three-dimensional convolutional neural network to obtain geophysical feature data;

[0050] (3) The geochemical data in the standardized multi-source dataset is constructed into a node-edge relationship structure, and the information of the nodes is aggregated through the graph neural network to obtain chemical characteristic data;

[0051] (4) Construct a joint feature matrix based on surface feature data, geophysical feature data, and chemical feature data, and use attention weights to fuse different features to obtain fused feature data;

[0052] (5) Input the fused feature data into the encoder part of the variational autoencoder, perform dimensionality reduction compression through the multi-layer perceptron, and obtain the potential feature vector;

[0053] (6) The reconstruction error and distribution constraints of the latent feature vector are calculated, and the features are re-encoded through nonlinear mapping to obtain deep coding features.

[0054] Specifically, the improved ResNet architecture extracts surface features through repeated convolution operations and residual connections. Standardized remote sensing image data is input to the first convolutional layer of the improved ResNet architecture. The convolution kernel size is set to 3×3 with a stride of 1. A sliding window operation is performed on the input data to extract local features. The output of the convolutional layer is processed by batch normalization and ReLU activation. Residual connections pass the input data directly to the next layer, where it is element-wise added to the convolutional layer output to avoid the vanishing gradient problem as the network deepens. By stacking multiple convolutional blocks, surface features from low-level textures to high-level semantics are extracted layer by layer, resulting in surface feature data containing information such as geological structure and lithologic distribution. In the three-dimensional reconstruction of geophysical exploration data, gravity and magnetic data are reorganized into a three-dimensional data volume based on spatial location information, with x and y representing plane coordinates and z representing depth information. The three-dimensional convolutional neural network uses 3D convolution kernels to simultaneously extract features in three dimensions. The network architecture comprises multiple 3D convolutional layers, each with a 3×3×3 kernel size. Downsampling occurs layer by layer in the depth direction, extracting geophysical anomaly features at different depths. The pooling layer employs a max pooling operation to reduce the spatial dimensionality of the feature map while retaining significant features. This combination of multiple layers of 3D convolution and pooling yields geophysical feature data that characterizes the three-dimensional distribution of subsurface geological bodies.

[0055] The graph structure of geochemical data is constructed based on the spatial distribution of element content and the correlation between elements. Each sampling point is a node in the graph, and the node attributes contain the element content information of the point. The edges between nodes are determined based on the spatial distance and the correlation of element content. Connections are established between points with a distance less than a threshold and high element correlation. Figure 2 , which is a schematic diagram of the node-edge relationship structure in an embodiment of the present application, Figure 2The circular nodes in the middle represent sampling points, and the copper (Cu) and gold (Au) content values for each sampling point are displayed within the nodes. The lines connecting the nodes represent correlation connections, where w represents the connection weight, which ranges from 0 to 1, with larger values indicating stronger correlation. This figure illustrates how discrete geochemical sampling data is converted into a spatially correlated graph structure, providing the input data format for subsequent feature extraction using a graph neural network. The graph neural network aggregates node features through a message passing mechanism, with each node collecting information from its neighbors and updating its own state. After multiple rounds of message passing, each node's state encompasses the geochemical characteristics of the local area, resulting in aggregated chemical signature data. The construction of a joint feature matrix combines surface, geophysical, and chemical signature data along the feature dimension. The attention mechanism calculates the weighted relationships between different features, mapping them into the query, key, and value spaces. The attention score is calculated by taking the dot product of the query and key. The weight coefficient is normalized using the softmax function and then weighted summed with the value vector to produce the attention output. This mechanism dynamically adjusts the importance of different feature types, enabling adaptive feature fusion.

[0056] The encoder portion of the variational autoencoder consists of a multi-layer perceptron that performs dimensionality reduction and compression on the fused feature data. The encoder maps the input features to a mean vector and a variance vector through multiple fully connected layers, and samples the resulting latent feature vector. During training, a KL divergence loss term is introduced to constrain the distribution of the latent space to a standard normal distribution while minimizing reconstruction error. During the feature recoding phase, the latent feature vector is reconstructed, transforming the features into a new representation space through a nonlinear mapping. The reconstruction process uses inverse transposed convolutional layers to gradually restore the spatial structure of the features. The error between the reconstructed result and the original features is calculated and, combined with the KL divergence constraint, guides the model to learn more effective feature representations.

[0057] Taking data processing from a mining area as an example, multispectral remote sensing imagery covering a 20-square-kilometer area was fed into a modified ResNet architecture. Surface features were extracted using five residual blocks, each consisting of two convolutional layers and a short-circuit connection. Gravity and magnetic data, covering a depth of 500 meters, were extracted using a four-layer 3D convolutional network, reducing the feature map size by half layer by layer. Geochemical data, containing 10 elemental content indicators, was constructed using a graph structure based on the spatial distribution of sampling points. Edge weights between adjacent nodes were determined based on the correlation coefficient of elemental content. When fusion of the three feature types was performed, an attention mechanism was used to calculate weights of 0.4, 0.35, and 0.25 for surface, geophysical, and chemical features, respectively. A variational autoencoder compressed the 2048-dimensional fused features into a 128-dimensional latent space. The reconstruction error converged to a stable value on the validation set, demonstrating that the encoded features retained key information from the original data.

[0058] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0059] (1) The depth coding features are divided into fault features, rock mass features, and alteration zone features according to the geological element type, and the initial geological element data are obtained through feature matrix reconstruction;

[0060] (2) Perform multi-head mapping decomposition on the initial geological element data, calculate the attention weights between elements through the multi-head self-attention mechanism, and obtain correlation feature data;

[0061] (3) Establish a set of geological element nodes based on the associated feature data, transfer and update features between nodes through the message passing mechanism, and obtain node status data;

[0062] (4) Convert node status data into an entity-relationship-attribute structure, perform knowledge representation through semantic coding, and obtain a mineralization knowledge graph;

[0063] (5) Perform random walk sampling on the nodes in the metallogenic knowledge graph, generate node sequences through graph embedding algorithm, and obtain graph sequence data;

[0064] (6) The graph sequence data is vectorized and dimensionally unified, and the knowledge vector is obtained through nonlinear transformation.

[0065] Specifically, the classification processing of deep-coded features is based on the characteristic attributes of geological elements, separating feature vectors according to the characteristic expressions of different geological element types. Fault features primarily contain linear structural information, manifested as linear textures and geophysical anomalies in remote sensing imagery; rock mass features include lithologic combinations and spatial distribution; and alteration zone features reflect mineral alteration information. Feature submatrices are constructed for each of these three feature types, and the initial geological element data is reconstructed through matrix concatenation. During the multi-head mapping decomposition process, the initial geological element data is mapped into multiple subspaces. Specifically, multiple feature mapping heads are set up, each of which independently performs a linear transformation on the input features to generate a query vector, a key vector, and a value vector. A multi-head self-attention mechanism calculates attention weights in each subspace. Each geological element calculates a similarity score with other elements, and the attention coefficient is obtained after softmax normalization. These coefficients represent the strength of the association between different geological elements, thereby generating correlation feature data.

[0066] The establishment of node sets is based on associative feature data. Each geological element is regarded as a node, and the node attributes contain the feature information of the element. The message passing mechanism transmits information between nodes. In each round of transmission, the node collects information from its neighbors and updates its own status. The association weights between nodes are considered in the process of information transmission. The neighbor nodes with greater weights contribute more information. After multiple rounds of transmission, node status data containing local structural information is obtained. The construction of the entity-relationship-attribute structure converts the node status data into a knowledge graph representation. Entities correspond to geological element nodes, relationships represent the interactions between entities, and attributes describe the characteristics of entities. Semantic coding maps entities, relationships, and attributes to a unified semantic space to construct a mineralization knowledge graph. For example Figure 3 As shown in the figure, it is a schematic diagram of the mineralization knowledge map in the embodiment of the present application. The circular nodes in the figure represent different types of geological element entities, including fault characteristics (main faults, secondary faults), rock mass characteristics (intrusive rock mass) and alteration zone characteristics (alteration center, alteration halo); the key attribute information of the entity is displayed in the node; the lines between the nodes represent the relationship between geological elements, such as "derivation", "cutting", "control of mineralization", "contact zone alteration" and "peripheral development". The figure shows the complex associations between geological elements and constitutes a complete mineralization knowledge representation system. This structured representation captures the complex relationship network between geological elements.

[0067] Random walk sampling is performed on the mineralization knowledge graph. Starting from each node, random walks are performed according to the weighted probabilities of the edges, generating node visit sequences. These sequences preserve the neighborhood structure of the nodes in the graph. The graph embedding algorithm processes these sequences, mapping each node into a fixed-dimensional vector representation. Similar nodes are placed closer together in the embedding space. Vectorization transforms the graph sequence data into numeric vectors. The vector dimensions are adjusted through nonlinear transformations to ensure that all knowledge vectors have a uniform dimensional representation. This process preserves the structural information and semantic relationships in the knowledge graph.

[0068] For example, by classifying depth-encoded features, we extract fault characteristics (including regional main faults and secondary faults), rock mass characteristics (including intrusive rocks and surrounding rocks), and alteration zone characteristics (including alteration halos and alteration centers). Multi-head self-attention calculations reveal a strong correlation between faults and alteration zones, as reflected by their high attention weights. During node information transmission, fault nodes convey structural information to their surroundings, alteration zone nodes convey mineral assemblage information, and rock mass nodes convey lithologic information. This information forms directed relationships in the knowledge graph, such as "faults control alteration" and "alteration occurs in the rock mass contact zone." During random walk sampling, starting from the main fault node and following high-weight edges, we frequently reach the alteration center node, indicating a close correlation between the two. The resulting knowledge vector encodes these complex geological relationships.

[0069] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0070] (1) Concatenate the deep coding features and knowledge vectors, perform sequence conversion through position encoding, and obtain input sequence data;

[0071] (2) Decompose the input sequence data into tasks according to mineralization type, ore body size and grade prediction, and obtain multi-task input data through task identifier encoding;

[0072] (3) Construct a multi-task prediction model based on multi-task input data, where the multi-task prediction model includes a shared encoding layer, a task-specific decoding layer, and a prediction head layer. Feature transformation is performed through the multi-layer self-attention calculation of the Transformer network to obtain task feature data;

[0073] (4) Input the task characteristic data into the mineralization type prediction branch, ore body size prediction branch and grade prediction branch respectively, and perform decoding processing through the task specific decoding layer to obtain branch prediction data;

[0074] (5) Construct probability distribution parameters based on branch prediction data, perform sampling and uncertainty evaluation through the Bayesian neural network framework, and obtain prediction distribution data;

[0075] (6) The predicted distribution data is subjected to multi-task weighted fusion, and the mineralization prediction results are obtained through confidence threshold screening and normalization processing.

[0076] Specifically, feature concatenation is performed on the deep encoding features and knowledge vectors. The concatenation operation is performed on the feature dimension. The deep encoding features contain geological feature information, and the knowledge vector contains information about the relationship between geological elements. The two are connected on the same dimension to form a joint feature vector. Position encoding processes the positional dependencies in the sequence by adding position information to the feature vector. Sine and cosine functions are used to generate position encoding vectors of different frequencies, which are added to the feature vector to obtain the input sequence data. The multi-task decomposition process targets three prediction objectives: mineralization type (including different genetic types such as porphyry and hydrothermal), ore body size (grading from small to very large), and grade prediction (numerical prediction of ore grade). Each task is assigned a unique task identifier, which is converted into vector form through one-hot encoding. The task identifier is combined with the input sequence data to construct multi-task input data containing task information.

[0077] The core architecture of the multi-task prediction model consists of three layers: a shared encoding layer that utilizes the Transformer structure to process common features across all tasks, a task-specific decoding layer that decodes features specific to each task, and a prediction head layer that generates predictions. The self-attention mechanism in the Transformer network captures long-range dependencies by calculating the correlation between each position in the sequence and all other positions. Stacking these layers of self-attention allows the model to learn feature representations at different levels, resulting in task-specific data.

[0078] The processing of the three prediction branches can be expressed as the following mathematical formula:

[0079] 1. Decoding calculation of mineralization type prediction branch:

[0080] ,

[0081] in: predicting probability distributions for mineralization types; Predict weight matrix for mineralization type; It is a hidden state of mineralization characteristics; is the bias term; is the softmax activation function.

[0082] 2. Decoding calculation of ore body size prediction branch:

[0083] ,

[0084] in: is the predicted value of ore body size; is the scale prediction transformation matrix; It is the scale feature representation; bias for scale prediction; is the ReLU activation function.

[0085] 3. Decoding calculation of the grade prediction branch:

[0086] ,

[0087] in: For grade prediction results; is the grade prediction coefficient matrix; is the grade feature tensor; is the bias term for grade prediction; is a linear activation function.

[0088] The Bayesian neural network framework characterizes prediction uncertainty through probability distributions. Each prediction branch outputs a probability distribution rather than a single value. Monte Carlo sampling is used to extract multiple samples from the prediction distribution, calculating the mean as the predicted value and the variance as the uncertainty measure. This approach accounts for the uncertainty of both model parameters and prediction results to generate prediction distribution data. Multi-task weighted fusion comprehensively considers the results of the three prediction branches, assigning weights to each task based on its importance and prediction reliability. High-confidence predictions are screened by setting confidence thresholds, and the predicted values are normalized to produce the mineralization prediction results.

[0089] For example, when processing input data from a study area, the deep coding features include surface alteration, structural linearity, and lithologic distribution characteristics, and the knowledge vector encodes mineralization patterns such as fault-controlled ore formation and rock contact zone-controlled ore formation. Through multi-task prediction model analysis, the mineralization type prediction branch identifies porphyry copper mineralization characteristics, which are consistent with the known regional geological background. The ore body size prediction branch estimates the ore body size based on the size of the ore-bearing rock body and the intensity of alteration, predicting it to be medium-sized. The grade prediction branch predicts the grade distribution based on the intensity of alteration and the intensity of element anomalies. The Bayesian framework evaluation shows that the prediction at the fault intersection has a high confidence level, while the prediction in the marginal area has a large uncertainty. The three prediction results are integrated to identify the most favorable mineralization target area.

[0090] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0091] (1) Sort the mineralization prediction results according to the prediction probability value, and obtain the high probability prediction area data through probability threshold segmentation;

[0092] (2) Based on the high-probability prediction area data, a spatial distance matrix is constructed, and the areas are grouped by spatial clustering to obtain potential target area data;

[0093] (3) Boundary extraction and spatial overlay analysis of potential target area data are performed, and optimization is performed through geological parameter constraints to obtain preliminary optimized data;

[0094] (4) Input the preliminary optimization data into the Bayesian optimization algorithm, optimize the parameters through sampling point selection and objective function evaluation, and obtain supplementary survey location data;

[0095] (5) Update geological parameters based on supplementary exploration location data, optimize data through iterative calculation and parameter adjustment, and obtain iteratively optimized data;

[0096] (6) The iterative optimization data is comprehensively evaluated and weighted, and the optimized prediction results are obtained through multiple rounds of iterative calculations.

[0097] Specifically, during the optimization process of mineralization prediction results, the predicted probability values are sorted from largest to smallest, and a probability threshold (typically determined based on the statistical distribution of known mineral deposits) is selected for data segmentation. The sorted data is reorganized according to spatial location information, and prediction areas above the threshold are extracted to form high-probability prediction area data. This data contains the spatial coordinates and predicted probability values of each prediction point. A spatial distance matrix is constructed based on the spatial coordinate information in the high-probability prediction area data, calculating the Euclidean distance between prediction points to form a distance matrix. Spatial clustering uses the DBSCAN (density-based spatial clustering algorithm) method. By setting two parameters, the neighborhood radius and the minimum number of points, spatially similar high-probability prediction points are grouped into potential target areas. This clustering method can identify target areas of any shape and automatically filter out noise points.

[0098] The boundary extraction process uses a convex hull algorithm to calculate the outer boundary of each target area, while also accounting for the complexity of the internal structure. Spatial overlay analysis overlays the target area boundary with geological parameters (including lithologic boundaries, fault distribution, and alteration zone extent), adjusting the target area boundary based on the geological parameters. This process considers the integrity of the geological units and the continuity of the geological boundaries.

[0099] The Bayesian optimization algorithm further optimizes the preliminary optimization data. Its core formula is:

[0100] ,

[0101] in: It is the optimal supplementary survey location; is the expected improvement function; is the probability improvement function; is the confidence upper bound function; is the weight coefficient of different acquisition functions; X is the feasible sampling space.

[0102] During the geological parameter update process, supplementary survey location data is integrated with the original geological parameters, and the regional geological parameter distribution is updated using spatial interpolation methods such as kriging interpolation. During the iterative calculation process, the prediction results are reevaluated after each update, and the model parameters are adjusted until the parameter change is less than the preset threshold. This process continuously optimizes the spatial distribution characteristics of the geological parameters. The comprehensive evaluation phase considers multiple evaluation indicators, including target area prediction probability, geological plausibility, and economic feasibility. By setting weight coefficients for different indicators, these multiple evaluation indicators are combined, and optimized prediction results are obtained through multiple rounds of iterative calculations.

[0103] For example: sort the mineralization prediction results by prediction probability, set 0.75 as the probability threshold, and extract high-probability prediction areas. Through spatial cluster analysis, three potential target areas were identified at the intersection of faults, and each target area contained multiple high-probability prediction points. After boundary extraction, it was found that target area one and the known copper mineralization belt showed a zonal distribution feature in space, target area two was located in the core of the anticline structure, and target area three extended along the fault. The target area boundaries were optimized based on fault, lithology and alteration information to determine the best supplementary exploration position. Two drill holes were laid in target area one, and the geological parameters were updated after obtaining new core data. It was found that the mineralization intensity of target area one showed regular changes with increasing depth. After multiple rounds of iterative optimization, the priority exploration target area with the greatest exploration potential was identified.

[0104] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0105] (1) Reconstruct the optimized prediction results according to the spatial coordinates and obtain voxelized prediction data through three-dimensional interpolation processing;

[0106] (2) Input the voxelized prediction data into the deep generative model, and obtain the three-dimensional feature representation data through multi-layer encoding and decoding processing;

[0107] (3) Construct geometric topological relationships based on 3D feature representation data, optimize boundaries through spatial constraints, and obtain a 3D visualization model;

[0108] (4) Decompose the three-dimensional visualization model hierarchically, calculate the feature activation map using the Grad-CAM algorithm, and obtain prediction data;

[0109] (5) Compare and analyze the prediction basis data with geological prior knowledge, conduct rationality evaluation through rule verification, and obtain verification evaluation data;

[0110] (6) Perform multi-dimensional comprehensive analysis and weight calculation on the verification evaluation data, and obtain the target prediction verification results through confidence evaluation.

[0111] Specifically, optimizing the three-dimensional visualization of prediction results begins with spatial data reconstruction, converting the prediction results on a two-dimensional plane into a three-dimensional data structure based on spatial coordinate information. This process involves establishing a three-dimensional grid system, with the grid size determined based on the required prediction accuracy. Kriging interpolation is then used to spatially interpolate discrete prediction points to generate a continuous three-dimensional prediction field. Voxelization discretizes the continuous field into regular voxel units, each containing both the predicted value and spatial location information.

[0112] The deep generative model employs a 3D convolutional autoencoder architecture. The encoder compresses voxelized prediction data into a low-dimensional feature space through a multi-layer 3D convolutional network. Batch normalization and ReLU activation functions are applied after each convolution layer. The decoder gradually restores the spatial structure through transposed convolutional layers, recovering the 3D feature representation. During the encoding and decoding process, skip connections preserve feature information at different scales, resulting in a 3D feature representation containing multi-scale features. Geometric topological relationships are constructed based on the 3D feature representation data. Isosurfaces are extracted to construct the initial geometric boundary, and then connectivity relationships between voxels are established. Spatial constraints include geological constraints such as stratum occurrence and fault strike. The boundary is optimized and adjusted using the minimum surface energy method to ensure compliance with geological laws. The 3D visualization model constructs a geologically meaningful three-dimensional structure based on the topological relationships and spatial constraints.

[0113] The Grad-CAM algorithm slices the 3D visualization model layer by layer, generating 2D feature maps at different depths. For each feature map layer, a class activation map is calculated. The specific steps include calculating the gradient of the predicted class with respect to the feature map, performing global average pooling on the gradient to obtain a weight for each channel, adding the weights to the feature map, and performing ReLU activation to generate a heat map. This process reveals the key areas of focus for the model and generates prediction data. Comparative analysis of the prediction data with prior geological knowledge includes multiple aspects: spatial distribution rationality (whether it conforms to mineralization patterns), structural control relationships (whether it is consistent with major ore-controlling structures), and lithologic relevance (whether it corresponds to favorable mineralization lithologies). By establishing an evaluation rule library, each rule is quantified into a scoring metric, and comprehensive calculations are performed to generate validation evaluation data.

[0114] This multi-dimensional comprehensive analysis assesses prediction accuracy, geological plausibility, and economic feasibility. Accuracy is evaluated through cross-validation to assess the model's predictive capabilities, geological plausibility is determined based on geological patterns, and economic feasibility considers resource endowment and mining conditions. Indicators from these different dimensions are combined using a weighted approach, with weighting coefficients determined based on expert experience, to produce the target prediction verification results.

[0115] For example, the original prediction results included predicted probability values for planar distribution. A voxel model with a voxel size of 10 m × 10 m × 10 m was established within a depth range of 500 meters through 3D data reconstruction. The 3D feature representation obtained after processing the depth-generated model clearly depicts the spatial distribution characteristics of the ore body and shows the regularity of extension along the fault strike. Geometric topological relationship analysis determined the contact relationship between the ore body and the surrounding rock, as well as the spatial connectivity between different mineralized zones. Grad-CAM analysis showed that the model focused heavily on fault intersection zones and rock contact zones, which is consistent with regional mineralization patterns. Comparing these features with those of known ore deposits verified the rationality of the prediction results. Through multi-dimensional comprehensive analysis, the target areas with the greatest exploration value were identified.

[0116] In a specific embodiment, the process of performing the step of hierarchical decomposition of the three-dimensional visualization model may specifically include the following steps:

[0117] (1) Slice and decompose the 3D visualization model and obtain layered data through gridding;

[0118] (2) Sort and group the hierarchical data according to the depth information, and obtain multi-level structured data through hierarchical index encoding;

[0119] (3) Extract features from each level of the multi-level structure data and obtain feature weight data through target category gradient calculation;

[0120] (4) The feature weight data is weighted and superimposed with the feature map of the corresponding level, and the heat map data is obtained through the Grad-CAM algorithm;

[0121] (5) Perform threshold segmentation and regional connectivity analysis based on heat map data, and obtain key area data through spatial correlation calculation;

[0122] (6) The key regional data are sorted by feature importance and regionally integrated, and the prediction basis data is obtained through spatial combination relationship analysis.

[0123] Specifically, the analysis and processing of the three-dimensional visualization model is sliced and decomposed, and slices are taken along the three main planes of XY, YZ, and XZ to form a series of two-dimensional sections. Each section is gridded according to a fixed grid spacing, and the attribute values and spatial position information of the grid nodes are recorded. The gridding process takes into account the continuity of the geological body and adopts an adaptive grid division method. The grid is encrypted in areas where the geological characteristics change drastically to obtain hierarchical data reflecting the spatial distribution characteristics. The sorting and grouping of the hierarchical data are organized based on the depth information, and the slice data of different depths are arranged in order from shallow to deep. The hierarchical index encoding adopts a tree structure, where the parent node represents the larger-scale feature and the child node represents the local detail feature. Each level is given a unique coding identifier, and an inheritance relationship is established between the levels to form a multi-level structure data containing multiple hierarchical features.

[0124] During the feature extraction process, convolution operations are performed on the data at each level to extract feature maps of different scales. The target category gradient calculation reflects the sensitivity of the prediction result to the feature map, and the contribution of each feature map to the target prediction is calculated through backpropagation. The global average of the gradient is calculated for each feature channel to obtain feature weight data that represents the importance of the feature. The weighted superposition of feature weights and feature maps uses a pixel-by-pixel multiplication operation, multiplying the weight coefficient by the corresponding feature map. The Grad-CAM algorithm processes the weighted results using the ReLU activation function, highlighting areas of positive contribution and generating heat map data that reflects the basis for the prediction. High-value areas on the heat map indicate spatial locations that contribute significantly to the prediction result.

[0125] The threshold segmentation process uses the OTSU algorithm to adaptively determine the threshold and divide the thermal map data into high-contribution areas and low-contribution areas. The regional connectivity analysis uses the connected component labeling algorithm to identify spatially continuous high-contribution areas. By calculating the spatial distance and feature similarity between regions, the degree of association between regions is evaluated, and key area data with spatial correlation is obtained. Feature importance ranking is based on the average thermal value and area size of the region, and the importance of key areas is ranked. In the regional integration process, spatial proximity and geological similarity are considered, and spatially close areas with similar features are merged. By analyzing the spatial combination relationship of the merged areas, including spatial position relationship, geometric morphological characteristics and geological attribute correlation, the prediction basis data is obtained.

[0126] For example, a 3D visualization model covering a depth of 500 meters was sliced, with a horizontal section taken every 10 meters in the vertical direction, and vertical sections taken along the strike and dip directions. The layered data obtained after gridding clearly shows the spatial distribution characteristics of the mineralized zone. Hierarchical coding organizes the data into four levels, corresponding to geological features of different scales. Feature extraction found that deep fault zones and rock contact zones have higher feature weights. Thermal map analysis shows that favorable areas for mineralization are mainly concentrated at fault intersections and lithologic transition zones. Regional connectivity analysis identified three major key areas, corresponding to different mineralization environments. Through spatial combination relationship analysis, it was found that these key areas are spatially distributed in a banded manner, and are consistent with the direction of the regional tectonic lines, providing a reliable spatial basis for mineralization prediction.

[0127] The above describes the mineral deposit prediction method based on deep learning in the embodiment of the present application. The following describes the mineral deposit prediction device based on deep learning in the embodiment of the present application. Figure 4 In the embodiments of the present application, an embodiment of a mineral deposit prediction device based on deep learning includes:

[0128] The processing module 201 is used to pre-process the remote sensing image data, geophysical exploration data and geochemical data through data correction and normalization algorithms to obtain a standardized multi-source data set;

[0129] A fusion module 202 is configured to construct a feature extraction model based on the standardized multi-source dataset using an improved ResNet structure, a three-dimensional convolutional neural network, and a graph neural network, and perform feature fusion using a variational autoencoder to obtain deep coding features;

[0130] A construction module 203 is used to construct a metallogenic knowledge graph based on the deep coding features through a multi-head self-attention mechanism and a message passing mechanism, and obtain a knowledge vector using a graph embedding algorithm;

[0131] An evaluation module 204 is configured to construct a multi-task prediction model using a Transformer network based on the deep coding features and the knowledge vector, and perform uncertainty evaluation using a Bayesian neural network framework to obtain a mineralization prediction result;

[0132] An optimization module 205 is configured to perform iterative optimization based on the mineralization prediction result by using spatial clustering and Bayesian optimization algorithms to obtain an optimized prediction result;

[0133] The analysis module 206 is used to construct a three-dimensional visualization model based on the optimized prediction results using a deep generative model, and perform explanatory analysis using a Grad-CAM algorithm to obtain a target prediction verification result.

[0134] Through the collaborative efforts of these components, multi-source geological data are preprocessed using data correction and normalization algorithms, ensuring standardization and consistency across different data types and improving the accuracy of subsequent feature extraction. A feature extraction model constructed using an improved ResNet architecture, a three-dimensional convolutional neural network, and a graph neural network enables in-depth feature extraction tailored to the characteristics of remote sensing imagery, geophysical exploration, and geochemical data, respectively. The feature fusion mechanism of a variational autoencoder effectively integrates multi-source data. A mineralization knowledge graph is constructed using a multi-head self-attention mechanism and message passing, fully exploring the correlations between geological elements. The knowledge vectors derived through a graph embedding algorithm effectively encode complex geological relationships. A multi-task prediction model based on a Transformer network enables joint prediction of mineralization type, scale, and grade. The introduction of a Bayesian neural network framework provides uncertainty assessment for the prediction results. The iterative optimization process of spatial clustering and Bayesian optimization improves the spatial continuity and geological rationality of the prediction results, making them more consistent with actual geological laws. Finally, a three-dimensional visualization model was constructed through a deep generative model and the Grad-CAM algorithm was used for explanatory analysis, which not only provided an intuitive display of the prediction results, but also revealed the key basis in the prediction process, enhanced the interpretability and credibility of the prediction results, and improved the accuracy, reliability and practicality of mineral prediction.

[0135] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the deep learning-based mineral prediction method.

[0136] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0137] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0138] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above 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 application.

Claims

1. A mineral deposit prediction method based on deep learning, characterized in that: The deep learning-based mineral deposit prediction method includes: Preprocess remote sensing image data, geophysical exploration data, and geochemical data through data correction and normalization algorithms to obtain standardized multi-source data sets; According to the standardized multi-source data set, a feature extraction model is constructed through an improved ResNet structure, a three-dimensional convolutional neural network and a graph neural network, and feature fusion is performed using a variational autoencoder to obtain deep coding features, including: inputting the remote sensing image data in the standardized multi-source data set into the improved ResNet structure, extracting spatial features through the convolution layer, and performing gradient transfer through the residual connection to obtain surface feature data; performing three-dimensional data reconstruction on the geophysical exploration data in the standardized multi-source data set, performing hierarchical feature extraction through the three-dimensional convolutional neural network to obtain geophysical feature data; constructing the geochemical data in the standardized multi-source data set into a node-edge relationship structure, aggregating information on the nodes through the graph neural network to obtain chemical feature data; constructing a joint feature matrix based on the surface feature data, the geophysical feature data and the chemical feature data, and fusing different features using attention weights to obtain fused feature data; inputting the fused feature data into the encoder part of the variational autoencoder, performing dimensionality reduction compression through a multi-layer perceptron to obtain a potential feature vector; performing reconstruction error calculation and distribution constraint on the potential feature vector, and performing feature recoding through nonlinear mapping to obtain the deep coding features; Based on the deep coding features, a mineralization knowledge graph is constructed through a multi-head self-attention mechanism and a message passing mechanism, and a knowledge vector is obtained using a graph embedding algorithm; According to the deep coding features and the knowledge vectors, a multi-task prediction model is constructed through a Transformer network, and uncertainty assessment is performed through a Bayesian neural network framework to obtain a mineralization prediction result, including: feature splicing of the deep coding features and the knowledge vectors, and sequence conversion through position encoding to obtain input sequence data; task decomposition of the input sequence data according to mineralization type, ore body scale and grade prediction, and multi-task input data is obtained through task identifier encoding; based on the multi-task input data, the multi-task prediction model is constructed, wherein the multi-task prediction model includes a shared coding layer, a task-specific decoding layer and a prediction head layer, and feature transformation is performed through the multi-layer self-attention calculation of the Transformer network to obtain task feature data; the task feature data is respectively input into the mineralization type prediction branch, the ore body scale prediction branch and the grade prediction branch, and decoding processing is performed through the task-specific decoding layer to obtain branch prediction data; probability distribution parameters are constructed based on the branch prediction data, and sampling and uncertainty assessment are performed through the Bayesian neural network framework to obtain prediction distribution data; the prediction distribution data is subjected to multi-task weighted fusion, and the confidence threshold screening and normalization processing are performed to obtain the mineralization prediction result; Based on the mineralization prediction results, iterative optimization is performed through spatial clustering and Bayesian optimization algorithm to obtain optimized prediction results; Based on the optimized prediction results, a three-dimensional visualization model is constructed using a deep generative model, and an explanatory analysis is performed using the Grad-CAM algorithm to obtain target prediction verification results.

2. The mineral deposit prediction method based on deep learning according to claim 1, characterized in that: The remote sensing image data, geophysical exploration data and geochemical data are preprocessed by data correction and normalization algorithms to obtain a standardized multi-source data set, including: Performing atmospheric scattering removal and radiation correction on the remote sensing image data through atmospheric correction to obtain preliminary corrected remote sensing image data; According to the preliminary corrected remote sensing image data, geometric distortion correction and geographic coordinate registration are performed through a digital elevation model to obtain corrected remote sensing image data; The geophysical exploration data is divided into gravity data and magnetic data according to data type, and noise is eliminated on the gravity data and the magnetic data respectively by median filtering to obtain de-noised geophysical exploration data; performing numerical normalization on the denoised geophysical data using a minimum-maximum method to obtain normalized geophysical data; Based on the element content data in the geochemical data, outliers are identified using a box plot method and data are filled using an interpolation method to obtain processed geochemical data; The corrected remote sensing image data, the normalized geophysical data and the processed geochemical data are fused to obtain the standardized multi-source data set.

3. The mineral deposit prediction method based on deep learning according to claim 1, characterized in that: Based on the deep coding features, the metallogenic knowledge graph is constructed through a multi-head self-attention mechanism and a message passing mechanism, and a knowledge vector is obtained using a graph embedding algorithm, including: Dividing the depth coding features into fault features, rock mass features and alteration zone features according to geological element types, and obtaining initial geological element data through feature matrix reconstruction; Performing multi-head mapping decomposition on the initial geological element data, calculating the attention weights between elements through the multi-head self-attention mechanism, and obtaining correlation feature data; Establishing a geological element node set based on the correlation feature data, transmitting and updating features between nodes through the message transmission mechanism, and obtaining node status data; Converting the node state data into an entity-relationship-attribute structure, performing knowledge representation through semantic coding, and obtaining the metallogenic knowledge graph; Performing random walk sampling on the nodes in the metallogenic knowledge graph, generating a node sequence through the graph embedding algorithm, and obtaining graph sequence data; The graph sequence data is vectorized and dimensionally unified, and the knowledge vector is obtained through nonlinear transformation.

4. The mineral deposit prediction method based on deep learning according to claim 1, characterized in that: According to the mineralization prediction results, iterative optimization is performed through spatial clustering and Bayesian optimization algorithm to obtain optimized prediction results, including: Sorting the mineralization prediction results according to the prediction probability values, and obtaining high probability prediction area data by probability threshold segmentation; Constructing a spatial distance matrix based on the high-probability prediction area data, grouping the areas through spatial clustering, and obtaining potential target area data; Perform boundary extraction and spatial overlay analysis on the potential target area data, optimize through geological parameter constraints, and obtain preliminary optimized data; Inputting the preliminary optimization data into the Bayesian optimization algorithm, performing parameter optimization through sampling point selection and objective function evaluation, and obtaining supplementary survey position data; updating geological parameters based on the supplementary exploration location data, performing data optimization processing through iterative calculation and parameter adjustment to obtain iteratively optimized data; The iterative optimization data is comprehensively evaluated and weighted, and the optimization prediction result is obtained through multiple rounds of iterative calculations.

5. The mineral deposit prediction method based on deep learning according to claim 1, characterized in that: Based on the optimized prediction results, a three-dimensional visualization model is constructed using a deep generative model, and an explanatory analysis is performed using the Grad-CAM algorithm to obtain target prediction verification results, including: Reconstructing the optimized prediction results according to spatial coordinates and obtaining voxelized prediction data through three-dimensional interpolation processing; Inputting the voxelized prediction data into the deep generative model, and obtaining three-dimensional feature representation data through multi-layer encoding and decoding processing; Constructing a geometric topological relationship based on the three-dimensional feature representation data, and optimizing the boundary through spatial constraints to obtain the three-dimensional visualization model; Decomposing the three-dimensional visualization model hierarchically, calculating a feature activation map using the Grad-CAM algorithm, and obtaining prediction basis data; Comparing and analyzing the prediction basis data with prior geological knowledge, conducting rationality evaluation through rule verification, and obtaining verification evaluation data; The verification evaluation data is subjected to a multi-dimensional comprehensive analysis and weight calculation, and the target prediction verification result is obtained through confidence evaluation.

6. The mineral deposit prediction method based on deep learning according to claim 5, characterized in that: The step of performing hierarchical decomposition on the three-dimensional visualization model and calculating a feature activation map using the Grad-CAM algorithm to obtain prediction basis data includes: Slicing and decomposing the three-dimensional visualization model to obtain layered data through gridding; Sorting and grouping the hierarchical data according to depth information, and obtaining multi-level structure data through hierarchical index coding; Extracting features from each level of the multi-level structure data and obtaining feature weight data by target category gradient calculation; The feature weight data is weighted and superimposed with the feature map of the corresponding level, and the heat map data is obtained by the Grad-CAM algorithm; Performing threshold segmentation and regional connectivity analysis based on the heat map data, and obtaining key area data through spatial correlation calculation; The key area data are sorted by feature importance and regionally integrated, and the prediction basis data are obtained through spatial combination relationship analysis.

7. A mineral deposit prediction device based on deep learning, used to implement the mineral deposit prediction method based on deep learning as described in any one of claims 1 to 6, characterized in that: The deep learning-based mineral deposit prediction device includes: The processing module is used to pre-process remote sensing image data, geophysical exploration data and geochemical data through data correction and normalization algorithms to obtain standardized multi-source data sets; A fusion module is used to construct a feature extraction model based on the standardized multi-source dataset through an improved ResNet structure, a three-dimensional convolutional neural network, and a graph neural network, and to perform feature fusion using a variational autoencoder to obtain deep coding features; A construction module is used to construct a metallogenic knowledge graph based on the deep encoding features through a multi-head self-attention mechanism and a message passing mechanism, and obtain a knowledge vector using a graph embedding algorithm; An evaluation module is used to construct a multi-task prediction model through a Transformer network based on the deep coding features and the knowledge vector, and perform uncertainty evaluation through a Bayesian neural network framework to obtain a mineralization prediction result; An optimization module is used to perform iterative optimization based on the mineralization prediction result through spatial clustering and Bayesian optimization algorithm to obtain an optimized prediction result; The analysis module is used to construct a three-dimensional visualization model based on the optimized prediction results using a deep generative model, and perform explanatory analysis through the Grad-CAM algorithm to obtain target prediction verification results.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the mineral deposit prediction method based on deep learning as described in any one of claims 1 to 6 is implemented.

Images