Mineralization potential level evaluation method based on deep learning

Abstract

The invention discloses a mineralization potential level evaluation method based on deep learning, and relates to the technical field of mineral resource exploration, and the method comprises the steps: obtaining and preprocessing multi-source geological data of a target exploration region, and generating a standardized data cube of the target region; performing cause feature extraction on the standardized data cube to generate a cause correlation distribution map representing mineralization control elements; based on a small number of known ore occurrences in the target area, extracting a target area metallogenic mode prototype representing a regional metallogenic law from the cause correlation distribution map, and constructing a structured metallogenic knowledge map based on domain knowledge; taking the target area metallogenic mode prototype as guidance, querying a metallogenic knowledge graph to perform knowledge reasoning, and generating a knowledge-guided distant view area focus graph; and performing information enhancement on the standardized data cube by using the knowledge-guided distant view area focus image to generate a knowledge enhanced feature cube. According to the invention, the problem of insufficient evaluation reliability of traditional deep learning under the condition of sample scarcity in the initial stage of exploration is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mineral resources exploration, in particular to a mineralization potential level evaluation method based on deep learning. BACKGROUND

[0002] The core task of mineral resources exploration is to accurately evaluate the mineralization potential of a specific area. With the development of geological big data, deep learning-based prediction methods have become one of the mainstream technologies. Existing representative technologies automatically extract features of multi-source geoscience data through convolutional neural networks and perform supervised training based on known mine point labels to realize end-to-end mineralization potential evaluation. This method reduces subjective dependence and performs well in data-rich areas.

[0003] When facing the small sample scenario commonly existing in the early exploration stage, the existing deep learning method has significant limitations. Since there are few known mine points, data-driven models are prone to overfitting and are difficult to extract regional representative mineralization rules from limited samples. The key problem is that existing methods lack a mechanism to effectively integrate structured prior knowledge into the model learning process, making it difficult to establish a stable evaluation criterion when samples are scarce. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a mineralization potential level evaluation method based on deep learning, which solves the problem of insufficient robustness of evaluation results caused by sample scarcity under the condition of a small number of known mine points.

[0006] To solve the above technical problems, the present application provides the following technical solutions: The present application provides a mineralization potential level evaluation method based on deep learning, which comprises: acquiring and preprocessing multi-source geological data of a target exploration area to generate a standardized data cube of the target area; Extracting genetic features from the standardized data cube to generate a genetic correlation degree distribution map representing ore-controlling elements; Based on a small number of known mine points in the target area, a target area mineralization mode prototype representing regional mineralization rules is extracted from the genetic correlation degree distribution map, and a structured mineralization knowledge graph is constructed based on domain knowledge; Using the target area mineralization mode prototype as a guide, querying the mineralization knowledge graph for knowledge reasoning to generate a knowledge-guided prospective area focus map; Using the knowledge-guided prospective area focus map to enhance the information of the standardized data cube to generate a knowledge-enhanced feature cube; Fusing the knowledge-enhanced feature cube with the known mine points, and establishing a mineralization potential level evaluation criterion for the target area through a fast learning mechanism; The knowledge-enhanced feature cube is evaluated by using the mineralization potential level evaluation criterion, and a mineralization potential level distribution map of the target area is output.

[0007] As a preferred scheme of the mineralization potential level evaluation method based on deep learning, the multi-source geological data of the target exploration area is acquired and preprocessed to generate a standardized data cube of the target area, including the following steps: The multi-source geological data of the target exploration area is collected, and the multi-source geological data includes geological map data, geophysical exploration data, geochemical sampling data, and remote sensing image data. The multi-source geological data is spatially registered and unified to the same coordinate system and grid resolution. The geophysical exploration data and the geochemical sampling data after spatial registration are standardized based on the geological background field to generate standardized geophysical exploration data and geochemical sampling data. The geological map data, the standardized geophysical exploration data, the standardized geochemical sampling data, and the remote sensing image data are fused according to the grid position to form a multi-dimensional data cube of the target area.

[0008] As a preferred scheme of the mineralization potential level evaluation method based on deep learning, the standardized data cube is subjected to genetic feature extraction to generate a genetic correlation degree distribution map representing the ore-controlling elements, including the following steps: High-level spatial features are extracted from the standardized multi-dimensional data cube using a convolutional neural network to generate a high-level spatial feature map. Based on a pre-defined ore-forming element combination, the geochemical channel in the high-level spatial feature map is information-enhanced to generate an enhanced geochemical feature, and the normalized distance from each grid cell in the standardized multi-dimensional data cube to the main ore-controlling structure is calculated to generate a structure proximity index. The enhanced geochemical feature and the structure proximity index are weighted and fused to generate a genetic correlation degree distribution map representing the ore-controlling elements.

[0009] As a preferred scheme of the mineralization potential level evaluation method based on deep learning, the target area metallogenic model prototype representing the regional metallogenic regularity is extracted from the genetic correlation degree distribution map based on a small number of known ore deposits in the target area, including the following steps: According to the spatial position information of the known ore deposits in the target exploration area, the grid cells corresponding to the spatial position information of the known ore deposits are located on the genetic correlation degree distribution map, and the feature values of the grid cells corresponding to the spatial position information of the known ore deposits in the genetic correlation degree distribution map are extracted to obtain a genetic correlation feature set of the known ore deposits. The feature space density peak clustering is performed on the genetic correlation characteristic set of the known ore spot, and a clustering center vector is determined as a prototype of the ore-forming pattern of the target area.

[0010] As a preferred scheme of the mineralization potential hierarchical evaluation method based on deep learning, wherein: a structured ore-forming knowledge graph is constructed based on domain knowledge, including the following steps: Collect and organize the structured information in the geological literature and deposit database to form the geological knowledge raw data; Identify and extract key geological entities from the geological knowledge raw data to form a set of geological entities; Identify and extract the relationship between the geological entities from the geological knowledge raw data to form a set of geological relationships; Construct a graph structure with the set of geological entities as nodes and the set of geological relationships as edges to form a pre-constructed ore-forming knowledge graph.

[0011] As a preferred scheme of the mineralization potential hierarchical evaluation method based on deep learning, wherein: the target area ore-forming pattern prototype is used as a guide to query the ore-forming knowledge graph for knowledge reasoning, and a knowledge-guided prospective area focus map is generated, including the following steps: The target area ore-forming pattern prototype is input into the ore-forming knowledge graph as a query vector; Perform graph attention network reasoning in the ore-forming knowledge graph to activate the most relevant ore-forming knowledge path to the query vector; Map the activated ore-forming knowledge path back to the spatial dimension of the target exploration area to generate a knowledge-guided prospective area focus map.

[0012] As a preferred scheme of the mineralization potential hierarchical evaluation method based on deep learning, wherein: the knowledge-guided prospective area focus map is used to enhance the information of the standardized data cube to generate a knowledge-enhanced feature cube, including the following steps: Each pixel value in the knowledge-guided prospective area focus map is used as a weight coefficient; Each feature channel in the corresponding spatial position of the standardized data cube is multiplied by the weight coefficient using an element-by-element multiplication operation; The standardized data cube processed by the element-by-element multiplication operation is defined as a knowledge-enhanced feature cube.

[0013] As a preferred scheme of the mineralization potential hierarchical evaluation method based on deep learning, wherein: the knowledge-enhanced feature cube is fused with the known ore spot, including the following steps: According to the spatial position information of the known ore spot, the grid unit corresponding to the known ore spot in the knowledge-enhanced feature cube is located; extracting all feature values of the grid cells corresponding to the known ore spots in the knowledge-enhanced feature cube to generate a knowledge-enhanced feature sample set of the known ore spots; extracting feature values of the non-ore spot grid cells in the knowledge-enhanced feature cube to generate a knowledge-enhanced feature sample set of the non-ore spots; combining the knowledge-enhanced feature sample set of the known ore spots and the knowledge-enhanced feature sample set of the non-ore spots to generate a training sample set for the fast learning mechanism.

[0014] As a preferred scheme of the mineralization potential level evaluation method based on deep learning, the mineralization potential level evaluation criterion of the target area is established by the fast learning mechanism, including the following steps: inputting the training sample set into the meta-learning framework pre-trained by the multi-mineralization module to perform gradient descent update; the internal parameters of the meta-learning framework after the gradient descent update are solidified as the mineralization potential level evaluation criterion.

[0015] As a preferred scheme of the mineralization potential level evaluation method based on deep learning, the mineralization potential level evaluation criterion is applied to evaluate the knowledge-enhanced feature cube, and a mineralization potential level distribution map of the target area is output, including the following steps: inputting the feature vector of each grid cell of the knowledge-enhanced feature cube into the mineralization potential level evaluation criterion to calculate a continuous potential value for each feature vector; mapping the continuous potential value to a discrete mineralization potential level, and combining the mineralization potential levels of all grid cells according to the spatial positions to generate a mineralization potential level distribution map The present application has the advantages that: the multi-source geological data is standardized to generate a data cube, and an ore-forming pattern prototype is extracted from a small number of known ore spots, a knowledge-guided prospective area focusing map is generated in combination with a pre-constructed ore-forming knowledge graph, the data cube is information-enhanced by using the focusing map to form a knowledge-enhanced feature cube, an evaluation criterion is established by a meta-learning mechanism, and a mineralization potential level distribution map is output, so that the problem of insufficient evaluation reliability of traditional deep learning under the condition of sample scarcity in the early exploration stage is solved through deep fusion of knowledge guidance and data driving. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0017] Fig. 1 A flowchart of a deep learning-based mineralization potential hierarchical assessment method.

[0018] Fig. 2 A flowchart for generating a genetic correlation degree distribution diagram.

[0019] Fig. 3 A flowchart for constructing a metallogenic knowledge graph.

[0020] Fig. 4 A flowchart for generating a training sample set. DETAILED DESCRIPTION

[0021] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0022] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0023] Secondly, the "one embodiment" or "embodiment" referred to herein means that a specific feature, structure or characteristic can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is separate or alternative to other embodiments.

[0024] REFERENCE Figs. 1-4 For one embodiment of the present application, the embodiment provides a deep learning-based mineralization potential hierarchical assessment method, comprising the following steps: S1, obtaining and preprocessing multi-source geological data of a target exploration area to generate a standardized data cube of the target area.

[0025] S1.1, collecting multi-source geological data of the target exploration area, the multi-source geological data including geological map data, geophysical exploration data, geochemical sampling data and remote sensing image data.

[0026] Further, the geological map data provides lithology distribution and tectonic framework information, the geophysical exploration data reflects the underground physical property difference, the geochemical sampling data indicates the element enrichment regularity, and the remote sensing image data captures the comprehensive information of the ground. The multi-source data collection method ensures the completeness of the information dimension, and through the integration of observation data with different physical meanings, a more three-dimensional description of the geological body is constructed, avoiding the multi-solution problem that may exist in a single data source. The multi-source geological data collection realizes the all-around characterization of the ore-forming environment.

[0027] S1.2, spatially register multi-source geological data, unify to the same coordinate system and grid resolution.

[0028] Further, the geological map data is rasterized through a vector-to-raster operation, the geophysical exploration data is regularized through an interpolation method, the geochemical sampling data is used to generate a continuous field through Kriging interpolation, the remote sensing image data is corrected to eliminate distortion, and spatial registration ensures that different source data strictly correspond in spatial position, solves the spatial mismatch problem caused by differences in collection scale and resolution of multi-source data, enables data with different physical meanings to be analyzed collaboratively under a unified grid framework, and establishes a spatial reference for multi-source data fusion.

[0029] S1.3, standardize the geophysical exploration data and the geochemical sampling data based on the geological background field after spatial registration, to generate standardized geophysical exploration data and geochemical sampling data.

[0030] Further, each grid position corresponds to a multi-dimensional feature vector, the feature vector includes lithology code, geophysical anomaly value, geochemical element content, and remote sensing spectral value, and the feature channels of different data sources are sequentially arranged according to physical meaning to form a multi-dimensional data structure with clear geological interpretation. This grid position-based fusion method maintains spatial correlation and preserves independent feature information of each data source, and the data cube fusion forms a target area multi-dimensional data cube containing multi-source information.

[0031] S1.4, fuse the geological map data, the standardized geophysical exploration data, the standardized geochemical sampling data, and the remote sensing image data according to the grid position to form a target area multi-dimensional data cube.

[0032] Further, each grid position corresponds to a multi-dimensional feature vector, the feature vector includes lithology code, geophysical anomaly value, geochemical element content, and remote sensing spectral value, and the feature channels of different data sources are sequentially arranged according to physical meaning to form a multi-dimensional data structure with clear geological interpretation. This grid position-based fusion method maintains spatial correlation and preserves independent feature information of each data source, and the data cube fusion forms a target area multi-dimensional data cube containing multi-source information.

[0033] S2, extract genetic characteristics from the standardized data cube to generate a genetic correlation degree distribution map representing ore-controlling elements.

[0034] S2.1, use a convolutional neural network to extract high-level spatial features from the standardized multi-dimensional data cube to generate a high-level spatial feature map.

[0035] Further, the convolutional neural network performs layer-by-layer abstract processing on the standardized multi-dimensional data cube through the alternating stacking structure of multiple convolutional layers and pooling layers, identifies local basic patterns such as linear structure and lithological boundary in the shallow convolutional layer, and identifies more complex spatial combination relationships such as the intersection position of specific rock mass and fracture in the deep convolutional layer through the expansion of the receptive field. This hierarchical processing mechanism can automatically learn the complex spatial covariant relationships between multi-source geosciences data and encode them into highly condensed feature representations. Finally, the feature map output by the last layer of the convolutional neural network is a high-level spatial feature map, which carries the semantic information related to the spatial pattern of mineralization after deep refinement.

[0036] S2.2, information enhancement is performed on the geochemical channel in the high-level spatial feature map based on a predefined ore-forming element combination, and an enhanced geochemical feature is generated. A normalized distance from each grid cell in the standardized multi-dimensional data cube to the main ore-controlling structure is calculated, and a structure proximity index is generated.

[0037] Further, the predefined ore-forming element combination is a group of key elements determined in advance according to the element paragenetic rule of a specific deposit model. For example, for a porphyry copper system, copper, molybdenum, gold and silver are determined as the core element combination. In operation, the geochemical feature channels corresponding to these predefined elements are identified and separated from the high-level spatial feature map, and the feature responses of these selected element channels are superimposed and enhanced by using linear weighting or nonlinear fusion. The weighting coefficients can be allocated according to the importance of each element in the ore-forming process. This targeted enhancement based on prior knowledge can effectively amplify the abnormal signal of the element combination directly related to mineralization, and relatively suppress the background response of irrelevant or interfering elements, thereby generating an enhanced geochemical feature that highlights the ore-induced anomaly.

[0038] Specifically, the ore-controlling structure is a geological structure that controls mineralization, such as regional faults and structural intersections, which is interpreted from geological map data. The shortest distance from the center point of each grid cell to all main ore-controlling structures is calculated as The minimum value and the maximum value of all grid cells in the entire study area are calculated. The original distance of each grid cell is mapped to a structure proximity index value between zero and one through the formula, and the normalization processing eliminates the influence of the absolute distance dimension and scale, so that the structure ore-controlling effect of different scale exploration areas is comparable. The closer the value is to one, the closer the grid cell is to the ore-controlling structure in space, i.e. the more favorable the ore-forming geological conditions.

[0039] The expression of the structure proximity index is: ; wherein, the distance from the grid cell to the nearest major ore-controlling structure, a structure proximity index, the distance from the grid cell to the nearest major ore-controlling structure, the minimum value of the distance from all grid cells in the study area to the nearest major ore-controlling structure, the maximum value of the distance from all grid cells in the study area to the nearest major ore-controlling structure, the maximum value of the distance from all grid cells in the study area to the nearest major ore-controlling structure, the grid cell.

[0040] S2.3, the enhanced geochemical feature and the structure proximity index are weighted and fused to generate a genetic correlation degree distribution map representing the ore-controlling elements.

[0041] Further, the enhanced geochemical feature reflects material composition anomalies, and the structure proximity index reflects the favorable relationship between spatial position and ore-controlling structure, both of which depict ore-forming potential from two independent and complementary dimensions of material basis and spatial conditions; the weighted fusion process uses a linear weighted sum model, assigns a weight coefficient to the enhanced geochemical feature, and assigns another weight coefficient to the structure proximity index, the sum of the two weight coefficients is one, and the specific value of the weight coefficient is determined according to the dominant ore-forming factor of the target deposit type.

[0042] Specifically, for example, for a deposit type dominated by structural ore control, the weight of the structure proximity index can be appropriately increased, and for a deposit type dominated by lithology and lithofacies ore control, the weight of the enhanced geochemical feature can be correspondingly increased; the enhanced geochemical feature value and the structure proximity index value of each grid cell are multiplied by the corresponding weight and then added to obtain the genetic correlation degree value of the grid cell; the genetic correlation degree values of all grid cells constitute a spatial grid map, i.e., a genetic correlation degree distribution map, which organically combines material anomalies and spatial relationships to generate a comprehensive prospecting indicator.

[0043] S3, based on a small number of known ore spots in the target area, a target area ore-forming mode prototype representing the regional ore-forming regularity is extracted from the genetic correlation degree distribution map.

[0044] S3.1, according to the spatial position information of the known ore spots in the target exploration area, the grid cells corresponding to the spatial position information of the known ore spots are located on the genetic correlation degree distribution map, the feature values of the grid cells corresponding to the spatial position information of the known ore spots in the genetic correlation degree distribution map are extracted, and the genetic correlation feature set of the known ore spots is obtained.

[0045] Further, the known ore spot spatial position information is provided in the form of geographic coordinates, the coordinates of each known ore spot are matched with the grid coordinate system of the genetic correlation degree distribution map, the specific grid cell position where each ore spot coordinate is located is determined, the values stored in the genetic correlation degree distribution map of the occupied grid cells are read, which represent the comprehensive genetic correlation degree scores obtained in the previous assessment, and the genetic correlation degree values of the corresponding grid cells of all known ore spots are collected to form a value set, and the value set is the genetic correlation feature set of the known ore spots.

[0046] Specifically, the discrete geological entity position is converted into a continuous feature value set, the genetic correlation degree value of the known ore spot position is directly extracted, the comprehensive ore-prospecting information generated in the previous processing stage is effectively utilized, and it is ensured that the subsequent analysis is directly established on the features closely related to mineralization.

[0047] S3.2, the genetic correlation feature set of the known ore spots is subjected to feature space density peak clustering, and the cluster center vector is determined as the target area metallogenic pattern prototype.

[0048] Further, the local density of each data point in the genetic correlation feature set of the known ore spots and the minimum distance to a higher density point are determined, the points with high local density represent the sample aggregation area in the feature space, and the points with large distance to the higher density points are likely to be the center of a cluster; by finding the data points with high local density and large distance as candidate cluster centers, the natural clusters existing in the data set can be automatically discovered without pre-specifying the number of clusters; for example, different feature patterns corresponding to the main mineralization period and the secondary mineralization period; the feature vector corresponding to the cluster center with the highest local density is selected as the target area metallogenic pattern prototype, and the prototype vector integrates the most representative mineralization feature information.

[0049] Specifically, the problem that the simple arithmetic mean can be excessively affected by the abnormally high or low ore spot features is overcome, and the most typical metallogenic feature pattern in the region can be stably extracted, the density peak clustering algorithm automatically identifies the natural distribution center in the feature space in a data-driven manner, ensures the representativeness and reliability of the target area metallogenic pattern prototype, and determines the cluster center vector as the target area metallogenic pattern prototype, thereby completing the extraction process from discrete samples to representative patterns.

[0050] S4, a structured metallogenic knowledge graph is constructed based on domain knowledge.

[0051] S4.1, structured information in geological literature and ore deposit databases is collected and sorted to form geological knowledge raw data.

[0052] Further, the geological literature includes regional geology, deposit exploration reports and academic papers, and the deposit database contains global or regional standardized deposit information library; key texts and data tables describing deposit genetic types, wall rock characteristics, ore-controlling structures and mineral assemblages are extracted from the sources; the extracted information is standardized, such as unifying the naming standards of rock names, standardizing the classification scheme of deposit types, and standardizing the representation units of geochemical data; the information from different sources is associated and merged according to the deposit instances to ensure that each deposit instance has complete and consistent attribute descriptions; a structured data table with standardized attributes is formed, which is the geological knowledge raw data.

[0053] S4.2, identify and extract key geological entities from the geological knowledge raw data to form a geological entity set.

[0054] Further, the geological knowledge raw data contains structured text descriptions and data tables, and the identification process is based on a pre-defined geological entity type dictionary, including rock types, mineral types, structural elements, deposit types, and geochemical elements; the text content in the geological knowledge raw data is scanned by string matching and natural language processing technology to identify and label professional term instances belonging to these pre-defined categories; for specific fields in the data table, the values are taken as instances of corresponding types of geological entities; each identified geological entity instance is assigned a unique identifier and records the name, type and source information; all identified geological entity instances are collected to form a geological entity set.

[0055] Specifically, unstructured geological text information is converted into discrete geological concept units with clear semantics, laying the foundation for the topology of the knowledge graph; through the extraction process guided by the pre-defined type dictionary, the accuracy and consistency of geological entity identification are ensured, and the formation of the geological entity set realizes the preliminary conversion from raw data to structured knowledge.

[0056] S4.3, identify and extract relationships between geological entities from the geological knowledge raw data to form a geological relationship set.

[0057] Further, the relationship extraction between geological entities is based on predefined relationship type patterns, which define typical geological semantic relationships such as ore-controlling, paragenesis, occurrence, and genetic correlation. Through analyzing the sentence structure and context of the original geological knowledge data describing the co-occurrence of geological entities, the entity pairs with the above predefined relationships are identified. For structured data tables, the relationship between entities is inferred by analyzing the relevance between fields, for example, when the deposit record and the wall rock record belong to the same exploration area, the occurrence relationship is established. Each identified relationship instance is recorded in the form of a triple, including the subject geological entity, the relationship type, and the object geological entity. All extracted relationship instances are collected to form a geological relationship set.

[0058] S4.4, construct a graph structure with the geological entity set as nodes and the geological relationship set as edges, and form a pre-constructed ore-forming knowledge graph.

[0059] Further, each geological entity instance in the geological entity set is mapped to an independent node in the graph structure, and each node is labeled with the corresponding geological entity type and attribute information. Each relationship instance in the geological relationship set is mapped to a directed edge connecting two corresponding nodes, and each edge is labeled with the specific relationship type. Through this mapping relationship, the complex semantic association between geological entities is transformed into the topological connection in the graph structure, forming a network knowledge structure composed of nodes and edges. The graph structure can intuitively express the multi-relationship between geological concepts, for example, a deposit type node can be connected to different wall rock nodes, structure nodes, and mineral nodes through multiple edges, fully describing the ore-forming geological background. The pre-constructed ore-forming knowledge graph formed is a machine-readable and computable knowledge base, which supports complex semantic queries and reasoning based on graph traversal and graph algorithms.

[0060] S5, guided by the target area ore-forming mode prototype, query the ore-forming knowledge graph for knowledge reasoning, and generate a knowledge-guided prospective area focus map.

[0061] S5.1, input the target area ore-forming mode prototype as a query vector into the ore-forming knowledge graph.

[0062] Further, the target area metallogenic model prototype is a feature vector extracted from a small number of known ore spots in the target exploration area, and carries the specific metallogenic feature mode information of the target area. The feature vector of the target area metallogenic model prototype is semantically aligned with the feature representation of the nodes in the metallogenic knowledge graph, ensuring that they are in the same feature space and are comparable. The semantically aligned target area metallogenic model prototype vector is used as a query input, and the semantic similarity between the query vector and the feature representation of all nodes in the metallogenic knowledge graph is calculated to start the retrieval and matching process of the entire knowledge graph. The regional pattern derived from data is used as the active query condition, which changes the passive retrieval mode of the traditional knowledge base and realizes the active semantic mapping from specific data to abstract knowledge.

[0063] Specifically, through semantic alignment and similarity calculation, the specific geological characteristics of the target area are converted into query semantics understandable by the knowledge graph, thereby activating relevant knowledge nodes and paths in the graph, ensuring the pertinence and effectiveness of knowledge guidance.

[0064] S5.2, execute graph attention network reasoning in the metallogenic knowledge graph to activate the most relevant metallogenic knowledge path to the query vector.

[0065] Further, the graph attention network reasoning first injects the target area metallogenic model prototype query vector as initial information into the corresponding nodes of the metallogenic knowledge graph, propagates and aggregates information among the graph nodes through a multi-round message passing mechanism. In each round of message passing, each node dynamically obtains attention weights based on the features and connection relationships of neighboring nodes. The size of the attention weight reflects the closeness of the semantic association between adjacent nodes, and nodes and edges with high semantic relevance to the query vector will obtain higher attention weights. After multiple iterations, the graph attention network will form one or more activated paths in the metallogenic knowledge graph, starting from the initial query node and composed of nodes and edges with high attention weights. These activated paths represent the most relevant metallogenic knowledge subgraph to the target area metallogenic model prototype.

[0066] Specifically, the reasoning mechanism based on dynamic attention weights can adaptively focus on the most relevant knowledge subnetwork to the current query, realizing intelligent filtering and directed activation of massive geological knowledge. The graph attention network reasoning overcomes the rigid restrictions of traditional rule matching, realizes flexible knowledge retrieval based on semantic similarity, and ensures that the activated metallogenic knowledge path not only conforms to geological laws but also is close to the specific characteristics of the target area. The activated metallogenic knowledge path provides precise knowledge guidance for subsequent spatial mapping.

[0067] S5.3, map the activated metallogenic knowledge path back to the spatial dimension of the target exploration area to generate a knowledge-guided prospective area focus map.

[0068] Further, the activated mineralization knowledge path contains a series of nodes and edges with high attention weight, and the nodes represent specific geological entity types such as granite body, fault structure or alteration type; the actual spatial distribution of geological bodies matching the activated node types in the target exploration area is identified, for example, all granite body distribution areas are extracted through a geological map, and linear structures are interpreted through geophysical data; then the spatial superposition rule is determined according to the semantic relationship strength between the nodes in the activated path, for example, when the knowledge path shows that the intersection of the granite body and the fault structure has a high ore-controlling weight, the rock body distribution area and the fault zone buffer area are superimposed and analyzed in space; the spatial attention degrees corresponding to different activated paths are weighted and fused to generate a continuous spatial distribution map, and the pixel value of each position represents the matching degree of the geological conditions at the position and the activated mineralization knowledge path, that is, the knowledge-guided prospective area focus map.

[0069] Specifically, the mapping method converts abstract semantic associations into specific spatial attention weights, so that prior mineralization knowledge can directly guide the selection of field exploration areas. Through spatial superposition and weighted fusion, the knowledge-guided prospective area focus map effectively integrates the spatial correlation of multi-source geological information and highlights the most favorable mineralization section. The generation of the knowledge-guided prospective area focus map provides an accurate spatial guidance signal for subsequent data enhancement.

[0070] S6, information enhancement is performed on the standardized data cube by using the knowledge-guided prospective area focus map to generate a knowledge-enhanced feature cube.

[0071] S6.1, each pixel value in the knowledge-guided prospective area focus map is used as a weight coefficient.

[0072] Further, the knowledge-guided prospective area focus map is a spatial raster map, and the value stored in each pixel position in the map represents the correlation degree score of the spatial position and the mineralization potential based on the reasoning of the mineralization knowledge graph; these correlation degree scores are continuous values, reflecting the judgment strength of prior knowledge on the mineralization prospect of the spatial position; in subsequent processing, each pixel value of the knowledge-guided prospective area focus map is directly extracted and used as a weight coefficient for the corresponding spatial position, and a high correlation degree score corresponds to a high weight coefficient, and a low correlation degree score corresponds to a low weight coefficient; the value range of the weight coefficient is consistent with the pixel value range of the knowledge-guided prospective area focus map, ensuring the fidelity of information transmission. This direct mapping method converts the semantic output of knowledge reasoning into a calculable mathematical parameter, providing an accurate quantitative control signal for data modulation.

[0073] Specifically, by directly using spatialized knowledge correlation as weight coefficients, a seamless transition from knowledge judgment to data manipulation is achieved, enabling prior geological knowledge to be applied to multi-source data in a quantitative manner. The establishment of weight coefficients provides a knowledge-based modulation benchmark for subsequent feature enhancement.

[0074] S6.2. Perform element-wise multiplication on each feature channel corresponding to the spatial location in the standardized data cube using weighting coefficients.

[0075] Furthermore, for each spatial location in the standardized data cube, the weight coefficient value of the knowledge-guided prospective area focus map at the same spatial location is extracted; the weight coefficient value is then multiplied by the values ​​of all feature channels of the standardized data cube at that spatial location; this operation is mathematically represented as a feature channel expression, where the value of each feature channel is modulated by the weight coefficient of the same spatial location; this element-wise multiplication operation ensures that the modulation effect of knowledge guidance is uniformly applied to all feature dimensions, whether it is geophysical field features, geochemical element content features, or remote sensing spectral features, all are enhanced or suppressed to the same degree based on mineralization knowledge.

[0076] Specifically, the global feature modulation method effectively injects knowledge signals into multi-source data. Through spatially aligned channel-wise weighting, the knowledge-enhanced feature cube retains all information dimensions of the original data while improving the relative intensity of signals related to mineralization. This overcomes the information loss problem that may be caused by traditional feature selection methods and achieves full-dimensional data enhancement guided by knowledge.

[0077] The feature channel expression is: ; in, To enhance knowledge, feature cubes in spatial location No. The result values ​​on each feature channel To standardize the spatial location in the data cube No. The original values ​​on each feature channel For the feature channel index in the standardized data cube, This refers to the row index of the spatial location coordinates within the data cube. The column index for the spatial location coordinates in the data cube. A knowledge-guided vision area focusing map in spatial location The corresponding weighting coefficients.

[0078] S6.3. The standardized data cube that has undergone element-wise multiplication is defined as the knowledge-enhanced feature cube.

[0079] Further, after the element-by-element multiplication operation processing, each feature channel value at each spatial position in the standardized data cube has completed the weight modulation of the knowledge-guided prospective area focusing map; at this time, the data structure and dimension of the data cube remain unchanged, but the numerical distribution has undergone a substantial change, that is, the feature values of the high potential area determined by knowledge are retained or enhanced, and the feature values of the low potential area are inhibited; this data cube after knowledge weight modulation is given a new definition, that is, a knowledge-enhanced feature cube, to distinguish from the original standardized data cube; the knowledge-enhanced feature cube not only retains the multi-source feature dimension of the original data, but also integrates the spatial prior judgment of the ore-forming knowledge, forming a new feature representation of deep fusion of data and knowledge.

[0080] S7, fuse the knowledge-enhanced feature cube with the known ore spots.

[0081] S7.1, according to the spatial position information of the known ore spots, locate the grid cells corresponding to the known ore spots in the knowledge-enhanced feature cube.

[0082] Further, the spatial position information of the known ore spots exists in the form of geographic coordinates, and the knowledge-enhanced feature cube has a clear geographic reference and grid division; through coordinate conversion, the geographic coordinates of the known ore spots are mapped to the grid coordinate system of the knowledge-enhanced feature cube, to determine the specific grid cell row and column number where each known ore spot coordinate is located, to ensure that each known ore spot has only one corresponding grid cell in the knowledge-enhanced feature cube, and to establish an accurate correspondence between geological entities and data grids.

[0083] S7.2, extract all feature values of the grid cell corresponding to the known ore spots in the knowledge-enhanced feature cube to generate a knowledge-enhanced feature sample set of the known ore spots.

[0084] Further, in the knowledge-enhanced feature cube, each grid cell contains numerical values of multiple feature channels, which cover multidimensional information such as geophysical field features, geochemical element content features, and remote sensing spectral features; for each located grid cell corresponding to a known ore spot, the numerical values of the cell in all feature channels are systematically read to form a complete multidimensional feature vector; all multidimensional feature vectors corresponding to the known ore spots are collected in order to construct a feature matrix, the rows of which correspond to different known ore spot samples, and the columns correspond to different feature channels; this feature matrix is the knowledge-enhanced feature sample set of the known ore spots.

[0085] Specifically, the full feature extraction method ensures complete information representation of each known ore spot, avoids information loss caused by feature selection, and uses the knowledge-enhanced feature space. The knowledge-enhanced feature sample set of the known ore spot can more clearly reflect the ore-forming related signals, and the generation of the knowledge-enhanced feature sample set of the known ore spot completes the conversion process from spatial positioning to feature extraction.

[0086] S7.3, randomly select non-ore spot grid cells in the knowledge-enhanced feature cube, extract feature values of the non-ore spot grid cells to generate a knowledge-enhanced feature sample set of the non-ore spot.

[0087] Further, a set of all grid cells not occupied by the known ore spot in the knowledge-enhanced feature cube is determined as a candidate pool, and a random number generation algorithm is used to draw a specified number of grid cells from the candidate pool without replacement as negative sample points, ensuring that the probability of each grid cell being selected is equal and the spatial distribution is representative. For each selected non-ore spot grid cell, extract the values in all feature channels of the knowledge-enhanced feature cube to form a multi-dimensional feature vector; and all feature vectors of the non-ore spot grid cells are collected to form a negative sample feature matrix, i.e. the knowledge-enhanced feature sample set of the non-ore spot.

[0088] Specifically, the random sampling strategy ensures that the negative samples can comprehensively cover various geological background conditions in the study area, avoiding human selection bias. By using the same knowledge-enhanced feature space as the positive samples, the knowledge-enhanced feature sample set of the non-ore spot can accurately reflect the regional background feature pattern, providing high-quality negative samples for the classifier to learn the discrimination boundary between mineralization and non-mineralization. The establishment of the knowledge-enhanced feature sample set of the non-ore spot lays a foundation for building a balanced training data set.

[0089] S7.4, combine the knowledge-enhanced feature sample set of the known ore spot and the knowledge-enhanced feature sample set of the non-ore spot to generate a training sample set for the fast learning mechanism.

[0090] Further, all feature vectors in the knowledge-enhanced feature sample set of the known ore spot are taken as positive samples, and each positive sample is given a positive class label; all feature vectors in the knowledge-enhanced feature sample set of the non-ore spot are taken as negative samples, and each negative sample is given a negative class label; the feature vectors of the positive and negative samples are merged according to a unified data structure to form a complete feature matrix; the order of the merged samples is randomly rearranged to eliminate the training bias that may be caused by the sample arrangement order; and finally a complete training sample set containing feature vectors and corresponding class labels is generated. This merging operation ensures the balance of positive and negative samples in the training data.

[0091] Specifically, by merging the positive and negative samples enhanced by knowledge, the training sample set can contain both the ore-forming feature patterns and the background feature patterns, so that the fast learning mechanism can learn the effective classification boundary. This data preparation method overcomes the learning bias problem caused by the imbalance between the number of positive and negative samples under the small sample condition, and provides a high-quality training data basis for the fast learning mechanism. The generation of the training sample set for the fast learning mechanism makes good data preparation for the subsequent model adaptation.

[0092] S8. Establishing the mineralization potential level evaluation criteria of the target area through the fast learning mechanism.

[0093] S8.1, inputting the training sample set into the meta-learning framework pre-trained by the multi-ore-forming module, and performing gradient descent update.

[0094] Further, the meta-learning framework pre-trained by the multi-ore-forming module has the general ability to extract ore-forming related features from multi-source geoscience data; the training sample set contains the feature vectors of the positive and negative samples enhanced by knowledge and the corresponding mineralization potential labels; the meta-learning framework takes the feature vectors of the training sample set as input, calculates the prediction results through forward propagation, and compares the prediction results with the true labels to obtain the loss value; based on the loss value, the gradient of the internal parameters of the meta-learning framework with respect to the loss function is calculated using the back propagation algorithm; the internal parameters of the meta-learning framework are adjusted slightly in the opposite direction of the gradient using the gradient descent optimization algorithm, and the loss function value is gradually reduced.

[0095] Specifically, the gradient descent update process fine-tunes the parameters of the meta-learning framework from general ore-forming feature representation to target area specific ore-forming patterns. Through supervised learning on the small sample training set, the meta-learning framework can quickly adapt to the geological feature distribution of the target area, realize the transfer from general knowledge to specific application, and the gradient descent update process completes the parameter optimization of the meta-learning framework for the target area.

[0096] S8.2, the internal parameters of the meta-learning framework after gradient descent update are solidified as the mineralization potential level evaluation criteria.

[0097] Further, after completing the gradient descent update, all parameters inside the meta-learning framework, including the weight matrix of the convolutional layer, the weight vector of the fully connected layer, and the bias term, etc., reach the optimal state for the target area data; these optimized parameters are extracted from the meta-learning framework and reorganized into an independent mathematical function structure, which maintains the same computational graph topology as the meta-learning framework; this independent mathematical function structure is defined as the mineralization potential level evaluation criterion, which can receive the knowledge-enhanced feature vector of any grid cell as input, perform a series of linear transformations and nonlinear activation function calculations, and output the continuous mineralization potential value of the location. Through the parameter solidification operation, the dynamic learning process is converted into a static evaluation function, enabling the mineralization potential level evaluation criterion to run independently.

[0098] Specifically, the solidification process allows the results of knowledge transfer to be saved and reused, avoiding the computational overhead of retraining for each evaluation. The establishment of the mineralization potential level evaluation criterion marks the transition from the learning process to the application tool, providing an efficient computational basis for subsequent batch prediction. The formation of the mineralization potential level evaluation criterion provides the core discriminant for the entire evaluation process.

[0099] S9. Applying the mineralization potential level evaluation criterion to the knowledge-enhanced feature cube to output the mineralization potential level distribution map of the target area.

[0100] S9.1, input the feature vector of each grid cell in the knowledge-enhanced feature cube into the mineralization potential level evaluation criterion, and calculate the continuous potential value for each feature vector.

[0101] Further, each grid cell in the knowledge-enhanced feature cube corresponds to a multi-dimensional feature vector that integrates multi-source information such as geophysics, geochemistry, and remote sensing after knowledge enhancement; the mineralization potential level evaluation criterion is a trained mathematical function that contains a computational graph composed of multiple layers of linear transformations and nonlinear activation functions; input the feature vector of each grid cell as input, and sequentially pass through the linear transformation, nonlinear activation, and final output layer calculation of the mineralization potential level evaluation criterion to obtain a continuous scalar value; this continuous scalar value is the continuous potential value of the grid cell, reflecting the similarity of the spatial location to the known ore-forming pattern.

[0102] Specifically, the mapping from the multi-source feature space to the one-dimensional potential score is achieved, converting complex geological feature combinations into intuitive potential indicators. Through point-by-point calculation of the mineralization potential level evaluation criterion, all spatial locations in the knowledge-enhanced feature cube obtain corresponding continuous potential values, and the calculation of continuous potential values completes the key conversion from feature analysis to potential quantification.

[0103] The continuous potential value expression is: ; wherein, is a continuous potential value of a grid cell, is a grid cell, is a distance metric function, is a scaling parameter, is a corresponding feature vector of the grid cell, is a target area ore-forming pattern prototype vector. S9.2, map the continuous potential value to a discrete mineralization potential level, and combine the mineralization potential levels of all grid cells according to the spatial positions to generate a mineralization potential level distribution map. Further, according to the actual geological background and known mineralization characteristics of the target exploration area, a set of threshold intervals for dividing the mineralization potential level are set, each threshold interval corresponds to a discrete mineralization potential level, such as low potential, medium potential, and high potential levels; the continuous potential value of each grid cell in the knowledge-enhanced feature cube is classified according to the preset threshold interval to determine the discrete mineralization potential level to which each grid cell belongs; the discrete mineralization potential level values of all grid cells are arranged according to their original spatial coordinates to form a spatial grid map, which is a mineralization potential level distribution map.

[0104] Specifically, the continuous numerical prediction result is converted into a classification level with clear geological significance, which facilitates intuitive interpretation and exploration decision-making by geologists. By setting reasonable threshold intervals, the mineralization potential level distribution map can clearly reflect the spatial distribution pattern of different mineralization potential areas. The generation of the mineralization potential level distribution map marks the completion of the entire mineralization potential level evaluation process, and realizes the visualization expression from multi-source data to final results.

[0105] The embodiment also provides a computer device suitable for the mineralization potential level evaluation method based on deep learning, which comprises a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the mineralization potential level evaluation method based on deep learning proposed in the above embodiment.

[0106] The embodiment also provides a computer device suitable for the mineralization potential level evaluation method based on deep learning, which comprises a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the mineralization potential level evaluation method based on deep learning proposed in the above embodiment.

[0107] ​The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0108] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for evaluating mineralization potential levels based on deep learning proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk or an optical disk.

[0109] In summary, the present application generates a data cube by standardizing a plurality of geological data, extracts a cause correlation distribution map, extracts a prototypical ore-forming pattern from a small number of known ore spots, generates a knowledge-guided prospective area focusing map in combination with a pre-constructed ore-forming knowledge graph, enhances information of the data cube by using the focusing map, forms a knowledge-enhanced feature cube, establishes an evaluation criterion by using a meta-learning mechanism, and outputs a mineralization potential level distribution map. Through deep fusion of knowledge guidance and data driving, the present application solves the problem of insufficient reliability of traditional deep learning in the condition of a small number of samples in the early exploration stage.

[0110] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based method for evaluating mineralization potential hierarchically, characterized in that: This includes acquiring and preprocessing multi-source geological data of the target exploration area to generate a standardized data cube of the target area; Genetic features are extracted from standardized data cubes to generate a genetic correlation distribution map representing the mineralization control elements; Based on a small number of known mineral occurrences within the target area, a prototype of the target area metallogenic model representing the regional metallogenic regularity is extracted from the genetic correlation distribution map, and a structured metallogenic knowledge graph is constructed based on domain knowledge. Guided by the metallogenic model prototype of the target area, the metallogenic knowledge graph is queried to perform knowledge reasoning and generate a knowledge-guided vision area focus map. Information enhancement is performed on standardized data cubes using knowledge-guided far-field focus maps to generate knowledge-enhanced feature cubes; By integrating knowledge-enhanced feature cubes with known mineral deposits, a hierarchical assessment criterion for the mineralization potential of a target area is established through a rapid learning mechanism. The knowledge-enhanced feature cube is evaluated using the mineralization potential hierarchy assessment criteria, and a mineralization potential hierarchy distribution map of the target area is output.

2. The deep learning-based mineralization potential hierarchical assessment method as described in claim 1, characterized in that: Acquire and preprocess multi-source geological data of the target exploration area to generate a standardized data cube of the target area, including the following steps: Collect multi-source geological data of the target exploration area, including geological map data, geophysical exploration data, geochemical sampling data, and remote sensing image data; Spatial registration of multi-source geological data is performed to unify them to the same coordinate system and grid resolution; The spatially registered geophysical exploration data and geochemical sampling data are standardized based on the geological background field to generate standardized geophysical exploration data and geochemical sampling data. Geological map data, standardized geophysical exploration data, standardized geochemical sampling data, and remote sensing image data are fused according to grid locations to form a multidimensional data cube of the target area.

3. The deep learning-based mineralization potential hierarchy assessment method as described in claim 2, characterized in that: Genetic feature extraction is performed on the standardized data cube to generate a genetic correlation distribution map representing the mineralization control elements, including the following steps: High-level spatial features are extracted from a standardized multidimensional data cube using a convolutional neural network to generate a high-level spatial feature map. Information enhancement of geochemical channels in high-level spatial feature maps is performed based on predefined ore-forming element combinations to generate enhanced geochemical features. Normalized distances from each grid cell in the standardized multidimensional data cube to the main ore-controlling structures are calculated to generate structural proximity indices. By weighted fusion of enhanced geochemical characteristics and tectonic proximity indicators, a genetic correlation distribution map characterizing the mineralization control elements is generated.

4. The deep learning-based mineralization potential hierarchy assessment method as described in claim 3, characterized in that: Based on a small number of known mineral occurrences within the target area, a prototype of the target area metallogenic model, representing the regional metallogenic regularity, is extracted from the genetic correlation distribution map, including the following steps: Based on the spatial location information of known mineral deposits in the target exploration area, the grid cells corresponding to the spatial location information of the known mineral deposits are located on the genetic correlation degree distribution map. The feature values ​​of the grid cells corresponding to the spatial location information of the known mineral deposits are extracted in the genetic correlation degree distribution map to obtain the genetic correlation feature set of the known mineral deposits. The feature space density peak clustering is performed on the genetic correlation feature set of known mineral deposits, and the cluster center vector is determined as the prototype of the mineralization mode of the target area.

5. The deep learning-based mineralization potential hierarchy assessment method as described in claim 4, characterized in that: Constructing a structured mineralization knowledge graph based on domain knowledge includes the following steps: Collect and organize structured information from geological literature and mineral deposit databases to form original data on geological knowledge; Key geological entities are identified and extracted from raw geological data to form a set of geological entities; Identify and extract the relationships between geological entities from raw geological data to form a set of geological relationships; Using sets of geological entities as nodes and sets of geological relationships as edges, a graph structure is constructed to form a pre-constructed metallogenic knowledge graph.

6. The deep learning-based mineralization potential hierarchy assessment method as described in claim 5, characterized in that: Guided by the metallogenic model prototype of the target area, the metallogenic knowledge graph is queried for knowledge reasoning to generate a knowledge-guided prospective area focus map, including the following steps: The metallogenic model prototype of the target area is used as a query vector to input the metallogenic knowledge graph; Perform graph attention network reasoning in the metallogenic knowledge graph to activate the metallogenic knowledge path most relevant to the query vector; The activated mineralization knowledge path is mapped back to the spatial dimension of the target exploration area, generating a knowledge-guided prospective area focus map.

7. The deep learning-based mineralization potential hierarchy assessment method as described in claim 6, characterized in that: The standardized data cube is augmented using a knowledge-guided prospective region focus map to generate a knowledge-enhanced feature cube, including the following steps: Each pixel value in the knowledge-guided vision area focus map is used as a weighting coefficient. Element-wise multiplication is performed on each feature channel at the corresponding spatial location in the standardized data cube using weighting coefficients. The standardized data cube that has undergone element-wise multiplication is defined as the knowledge-enhanced feature cube.

8. The deep learning-based mineralization potential hierarchy assessment method as described in claim 7, characterized in that: The process of integrating knowledge-enhanced feature cubes with known mineral deposits includes the following steps: Based on the spatial location information of the known mineral deposits, the corresponding grid cells of the known mineral deposits are located in the knowledge-enhanced feature cube; Extract all feature values ​​of the grid cells corresponding to known mining sites in the knowledge-enhanced feature cube to generate a knowledge-enhanced feature sample set of known mining sites; Randomly select non-mineral point grid cells in the knowledge-enhanced feature cube, and extract the feature values ​​of the non-mineral point grid cells to generate a knowledge-enhanced feature sample set for non-mineral points; The knowledge-enhanced feature sample set of known mining sites is merged with the knowledge-enhanced feature sample set of non-mining sites to generate a training sample set for the fast learning mechanism.

9. The deep learning-based mineralization potential hierarchical assessment method as described in claim 8, characterized in that: A rapid learning mechanism is used to establish a tiered assessment criterion for the mineralization potential of a target area, including the following steps: The training sample set is input into the meta-learning framework pre-trained by the multi-mineralization module, and gradient descent is performed to update it. The internal parameters of the meta-learning framework updated by gradient descent are solidified into the mineralization potential hierarchy evaluation criteria.

10. The deep learning-based mineralization potential hierarchy assessment method as described in claim 9, characterized in that: The knowledge-enhanced feature cube is evaluated using the mineralization potential hierarchy assessment criteria, and a mineralization potential hierarchy distribution map of the target area is output, including the following steps: The feature vector of each grid cell of the knowledge-enhanced feature cube is input into the mineralization potential hierarchy evaluation criterion, and a continuous potential value is calculated for each feature vector; By mapping continuous potential values ​​to discrete mineralization potential levels, and combining the mineralization potential levels of all grid cells according to their spatial location, a mineralization potential level distribution map is generated.

Images