A Mineral Resource Distribution Prediction Method Based on Big Data Analysis

By constructing an anchored representation system based on geological semantic units and a dynamic semantic alignment mechanism driven by graph neural networks, the problem of insufficient adaptability in the fusion of multi-source heterogeneous geological data is solved, achieving efficient and flexible mineral resource distribution prediction and improving prediction accuracy and interpretability.

CN123117428APending Publication Date: 2026-11-17GUANGDONG PROVINCIAL INSTITUTE OF MINERAL RESOURCES EXPLORATION (GUANGDONG PROVINCIAL INSTITUTE OF NUCLEAR GEOLOGICAL EXPLORATION)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611231121.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies have limited adaptability in the fusion of multi-source heterogeneous geological data, making it difficult to achieve deep semantic fusion and dynamic weight allocation driven by geological knowledge, resulting in insufficient prediction accuracy, especially in areas with complex geological structures.

Method used

By acquiring multi-source heterogeneous raw geological data, implicit geological semantic units are extracted and an initial geological semantic anchor point library is generated. A graph neural network is used to construct a reasonable anchor point relationship graph, calculate the dynamic confidence weight of the anchor points, generate the anchor point response spectrum, and generate a global semantic consistency vector through recursive propagation processing of stress transmission path, ultimately realizing the prediction of mineral resource distribution.

Benefits of technology

It significantly improves the relevance and physical consistency of cross-modal data fusion, enhances the extrapolation ability and robustness of prediction models under conditions of no mineral deposits or weak annotation, reduces the dependence on high-cost mineral deposit annotation data, and supports rapid deployment in small-sample or even zero-sample migration scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117428A_ABST
    Figure CN123117428A_ABST
Patent Text Reader

Abstract

This invention relates to a method for predicting mineral resource distribution based on big data analysis. It addresses the semantic extraction and uncertainty annotation of raw geological data, constructing a spatially and temporally consistent geological semantic anchor point library, and establishing a coordinate-independent, inferential anchor point relationship graph based on geological evolution logic using a graph neural network. Newly added local geological features can be efficiently matched with the anchor point library. Combined with geological evolution stages and stress transfer processing, dynamic weighting responses are applied to achieve feature-to-continuous spectral line mapping, further aggregating and generating a global semantic vector reflecting physical and topological consistency. Finally, a geological rule decoder is used to perform high-dimensional prediction of mineral resource distribution, outputting interpretive prediction results containing key anchor point combinations and evolutionary logic chains. This technology improves the accuracy of geological information expression, reasoning, and resource prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of mineral resource distribution prediction and geological big data fusion technology, and in particular to a method for predicting mineral resource distribution based on big data analysis. Background Technology

[0002] Current mineral resource distribution prediction primarily relies on the comprehensive analysis of multi-source geological data, achieving intelligent prediction through the fusion of various data sources such as remote sensing, geophysical exploration, geochemical analysis, and borehole core data. Existing technologies typically employ end-to-end deep learning models or manual feature engineering to project multi-source data onto the same raster or feature space for unified analysis. Some schemes utilize graph neural networks and convolutional neural networks for geological structure pattern recognition and resource target area inference. In recent years, data sources have become increasingly abundant and complementary. Mainstream solutions achieve data assimilation and integration through methods such as manually designed features, data-driven end-to-end models, raster registration, and feature stitching.

[0003] However, existing technologies have significant shortcomings in the fusion of multi-source heterogeneous geological data. Differences in spatial resolution, acquisition scale, and physical significance among different modalities make it difficult for traditional models to directly align and fuse them. Normalization and grid registration methods are prone to losing crucial geological genetic information. While deep learning end-to-end fusion models can automatically learn features, they lack interpretability and physical logic support, failing to demonstrate the key geological evidence and evolutionary chains upon which prediction results depend. Current technologies mostly employ static weight settings, making it difficult to dynamically adjust the fusion mechanism based on regional geological evolution and tectonic background, resulting in insufficient generalization ability and prediction reliability. Furthermore, existing methods generally rely on a large number of labeled mineral deposit samples, which has limited effectiveness in areas with complex geological structures.

[0004] Therefore, there is an urgent need for an innovative method that can overcome the limited adaptability of existing models to multi-source heterogeneous data and achieve geological knowledge-driven, semantically deep fusion, and dynamic weight allocation. Mapping geological information from different modalities to a unified and interpretable semantic space, and dynamically aligning and fusing it with geological evolution logic and topological relationships, has become a key technical requirement for improving prediction accuracy and interpretability. Summary of the Invention

[0005] This application provides a method for predicting the distribution of mineral resources based on big data analysis, which aims to solve one of the problems or issues of the existing technology mentioned in the background section.

[0006] This application provides a method for predicting the distribution of mineral resources based on big data analysis, specifically including:

[0007] S1: Obtain multi-source heterogeneous raw geological data, extract implicit geological semantic units for each type of raw geological data and label them with confidence labels to generate an initial geological semantic anchor point library;

[0008] S2: Based on the nodes in the initial geological semantic anchor point library and the known geological causal chain relationship, construct a reasonable anchor point relationship graph using a graph neural network;

[0009] S3: Match the newly accessed local geological features extracted based on the original geological data to the neighboring anchors in the initial geological semantic anchor point library, and activate the corresponding geological relationship sub-graph according to the current regional geological evolution stage to calculate the dynamic confidence weight of the anchor points.

[0010] S4: Use the dynamic confidence weight of the anchor points to distribute and map the response intensity of each data source on the inferable anchor point relationship graph to generate the anchor point response spectrum;

[0011] S5: Based on the inferable anchor point relationship graph, the response spectra of multiple anchor points are aggregated through recursive propagation of stress transmission paths to generate a global semantic consistency vector.

[0012] S6: Based on the global semantic consistency vector and its spatiotemporal derivative, perform mineral resource distribution deduction and output resource distribution prediction results.

[0013] This application provides a mineral resource distribution prediction method based on big data analysis, which has the following beneficial effects:

[0014] (1) This application effectively overcomes the problems of information alignment difficulties and semantic loss caused by inconsistent coordinate systems, large resolution differences, and strong modal heterogeneity in traditional multi-source geological data fusion by constructing an anchor-based representation system based on geological semantic units. Existing technologies usually rely on unified grid resampling or pixel-level registration for data preprocessing, which easily introduces spatial distortion and destroys the topological relationship of the original geological structure. However, this scheme abandons the forced alignment at the numerical level and instead extracts "semantic anchors" with clear geological significance from various types of data, such as alteration zone boundaries, magnetic gradient extreme lines, and hydrothermal activity intensity triplets, to form high-level abstract units with geological interpretability. These anchors not only carry location, time, and uncertainty information, but also contain key causal clues in the mineralization process, enabling data from different sources to achieve semantic-level matching and correlation reasoning under a unified geological logic framework, significantly improving the relevance and physical consistency of cross-modal data fusion, especially showing stronger adaptability in complex tectonic backgrounds or multi-stage superimposed mineralization areas.

[0015] (2) By introducing a dynamic semantic alignment mechanism driven by a graph neural network, this scheme achieves adaptive weight adjustment and relational reasoning that are sensitive to geological evolution background, breaking through the bottlenecks of poor generalization and lack of interpretability caused by static weighting, manually set rules, or end-to-end black box models in traditional fusion methods. Unlike fixed weight averaging or deep learning implicit feature splicing, this invention constructs a topological graph structure with semantic anchors as nodes and geological causal chains as edges, and dynamically activates the corresponding subgraph paths according to the current tectonic stage of the region (such as extension or compression environment), and then calculates the confidence weight of each anchor in this fusion—for example, automatically enhancing the influence of fracture density anchors in extensional environment and suppressing thrust fault related signals. On this basis, a recursive aggregation function based on stress transmission path is used to generate a global semantic consistency vector, avoiding information dilution caused by simple linear combination, so that the fusion result can truly reflect the spatial response pattern of regional geological dynamic processes, and greatly improve the extrapolation ability and robustness of the prediction model under conditions of no mining area or weak labeling.

[0016] (3) Since the input itself is a structured expression of geological semantic state, the model output can be directly traced back to the key anchor point combination and its evolutionary logic chain that supports its judgment, realizing full-link transparency from data to semantics to reasoning to decision-making. This mechanism not only reduces the dependence on high-cost mineral deposit annotation data and supports rapid deployment in small sample or even zero sample migration scenarios, but also provides exploration personnel with a clear path for analyzing mineralization control factors, greatly enhancing the trust foundation and practicality of human-machine collaborative decision-making. Overall, this scheme uses geological knowledge as the skeleton and data as the muscle to construct an efficient, flexible, and scalable multi-source heterogeneous geological information dynamic fusion architecture, providing solid technical support for deep mineral exploration prediction, tectonic-magmatic event reconstruction, and intelligent delineation of resource potential areas. Attached Figure Description

[0017] Figure 1 This is the main flowchart of a mineral resource distribution prediction method based on big data analysis;

[0018] Figure 2 This is a sub-flowchart of a mineral resource distribution prediction method based on big data analysis;

[0019] Figure 3 This is another sub-flowchart of a mineral resource distribution prediction method based on big data analysis. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0022] like Figure 1 As shown, this application provides a method for predicting the distribution of mineral resources based on big data analysis, specifically including:

[0023] S1: Obtain multi-source heterogeneous raw geological data, extract implicit geological semantic units for each type of raw geological data and label them with confidence labels to generate an initial geological semantic anchor point library;

[0024] S2: Based on the nodes in the initial geological semantic anchor point library and the known geological causal chain relationship, construct a reasonable anchor point relationship graph using a graph neural network;

[0025] S3: Match the newly accessed local geological features extracted based on the original geological data to the neighboring anchors in the initial geological semantic anchor point library, and activate the corresponding geological relationship sub-graph according to the current regional geological evolution stage to calculate the dynamic confidence weight of the anchor points.

[0026] S4: Use the dynamic confidence weight of the anchor points to distribute and map the response intensity of each data source on the inferable anchor point relationship graph to generate the anchor point response spectrum;

[0027] S5: Based on the inferable anchor point relationship graph, the response spectra of multiple anchor points are aggregated through recursive propagation of stress transmission paths to generate a global semantic consistency vector.

[0028] S6: Based on the global semantic consistency vector and its spatiotemporal derivative, perform mineral resource distribution deduction and output resource distribution prediction results.

[0029] Step S1: Acquire multi-source heterogeneous raw geological data, extract implicit geological semantic units for each type of raw geological data and label them with confidence levels to generate an initial geological semantic anchor point library. This initial geological semantic anchor point library carries explicit geological definitions, spatial locations, and temporal attributes. Specifically, it includes:

[0030] S1.1: Modal identification and preprocessing of multi-source heterogeneous raw geological data to generate a standardized geological data stream; specifically, this includes acquiring remote sensing image data, geophysical profile data, geochemical spectral data, and borehole core description text, using multimodal data cleaning to remove noise interference and performing format unification processing, and outputting a standardized geological data stream with a unified timestamp and spatial reference frame as the execution object for subsequent semantic extraction.

[0031] Acquire multi-source heterogeneous raw geological data, including remote sensing images, geophysical profiles, geochemical spectral lines, and borehole core texts, as input objects for modality recognition and preprocessing.

[0032] Radiometric and atmospheric corrections are performed on remote sensing image data to eliminate sensor noise and environmental scattering interference, and a normalized surface reflectance matrix is ​​generated.

[0033] Wavelet thresholding is applied to geophysical profile data to separate high-frequency random noise from low-frequency tectonic signals and extract effective magnetic or gravity anomaly sequences.

[0034] Baseline drift correction and smoothing filtering were performed on the geochemical spectral data to eliminate instrument background fluctuations and generate standardized concentration vectors of elemental content.

[0035] Natural language preprocessing was performed on the borehole core description text, including word segmentation, stop word removal, and entity recognition, and lithology, tectonics, and alteration keywords were extracted in a structured manner.

[0036] Establish a unified spatial reference frame and use coordinate transformation to map the geographic coordinates of all data to the same projected coordinate system to ensure the consistency of spatial location.

[0037] A timestamp alignment mechanism is introduced to normalize the collection time of each data source to a standard geological time unit or UTC time format, and a spatiotemporal synchronization index is constructed.

[0038] Through the above-mentioned multimodal data cleaning and format unification processing, heterogeneous raw data are transformed into standardized geological data streams with a unified spatiotemporal benchmark, enabling the subsequent semantic extraction module to achieve efficient compatibility and accurate parsing of multi-source data.

[0039] S1.2: Based on the standardized geological data stream, perform implicit geological feature extraction to generate a candidate geological semantic unit set; use a deep learning feature encoder to perform deep feature mining on different modal data in the standardized geological data stream, identify alteration zone boundary features from remote sensing images, extract gradient extreme value line features from geophysical profiles, and structurally extract lithological combination features from borehole core descriptions, and encapsulate the above extraction results into a candidate geological semantic unit set carrying original attribute information.

[0040] The system receives standardized geological data streams preprocessed by S1.1, including multispectral remote sensing images, gravity, magnetic and electric geophysical profiles, and borehole core texts, as input for implicit feature extraction.

[0041] For remote sensing image modalities, a semantic segmentation network based on the U-Net architecture is deployed. The input is a normalized multi-band pixel matrix. The encoder extracts spatial texture and spectral absorption features, and the decoder outputs a binary mask of the alteration zone boundary and a confidence heatmap to identify the spatial distribution of mineral assemblages such as silicification and argillaceization.

[0042] For geophysical profile data, the horizontal and vertical derivatives of the potential field data are calculated using the sliding window gradient operator to construct a three-dimensional gradient tensor field. The gradient extremum lines are extracted using the Canny edge detection algorithm to locate the geometric traces of fracture structures and lithological interfaces, and to generate a linear feature set carrying intensity attributes.

[0043] For unstructured borehole core description text, a pre-trained BERT model in the geological domain is loaded to perform named entity recognition task, extract key entities such as lithology name, degree of fragmentation, and cementation type, construct the modification relationship between entities through dependency parsing, and restructure them into a "lithology-structure-genesis" triple sequence.

[0044] The aforementioned visual mask, linear trace, and text triple are uniformly encapsulated into candidate geological semantic units. Each unit is appended with the original data source ID, extraction time, and preliminary spatial coordinates, forming a multimodal heterogeneous set of candidate geological semantic units.

[0045] By using multimodal deep learning feature encoding, the standardized data stream from the previous step is transformed into a set of candidate geological semantic units carrying deep geological attribute information, thus achieving a leap from raw signals to semi-structured geological features and providing a high-dimensional feature base for subsequent semantic alignment.

[0046] S1.3: Perform geological definition mapping and uncertainty quantification on the candidate geological semantic unit set to generate geological semantic anchor instance with confidence label; based on the pre-set geological knowledge graph, perform semantic alignment on each unit in the candidate geological semantic unit set, give it a clear geological definition name, and use the Bayesian probability model to calculate the spatial location deviation probability or attribute confidence of each unit to generate geological semantic anchor instance containing geological definition name, spatial location coordinates, time attribute field and confidence label.

[0047] A set of candidate geological semantic units and a pre-built geological knowledge graph are obtained. Graph matching is used to calculate the similarity between the deep feature vector of each candidate unit and the standard entity nodes in the knowledge graph. Based on the cosine similarity threshold, the standard geological concepts with the closest topological structure are selected, and a mapping relationship between candidate units and standard geological terms is established, giving them clear geological definitions such as "silicification alteration zone" or "ore-controlling fault".

[0048] A Bayesian probabilistic model is constructed, and for the mapped and defined geological semantic anchor points, the spatial coordinate error distribution parameters and attribute recognition confidence scores of their original data sources are extracted. A prior probability distribution function is defined, and historical sensor accuracy data is used as the likelihood function. The posterior probability distribution is then updated through Bayesian inference.

[0049] The probability of spatial position deviation is calculated using the following formula: Where P(δ|D) is the posterior spatial bias probability, P(D|δ) is the likelihood function based on the observed data D, P(δ) is the prior distribution of spatial location, and P(D) is the evidence factor. This calculation quantifies the range of positioning uncertainty of the anchor point in three-dimensional geological space.

[0050] The system calculates attribute confidence indices and performs weighted fusion of the consistency between remote sensing spectral matching, geophysical inversion residuals, and borehole lithological descriptions. It uses Monte Carlo simulation to generate probability density functions for attribute values ​​and extracts the 95% confidence interval width as the core parameter for confidence labels. The system encapsulates the geological definition name, central spatial coordinates, formation era timestamp, and calculated confidence labels into a structured object.

[0051] By using geological definition mapping and Bayesian uncertainty quantification, the candidate features from the previous step are transformed into geological semantic anchor instances with clear semantic identities and credibility indicators. This achieves standardized alignment and explicit quality labeling of multi-source data at the semantic level, laying a data foundation for the subsequent construction of a highly reliable inference anchor relationship graph.

[0052] S1.4: Based on the geological semantic anchor point instances labeled with confidence level tags, perform spatiotemporal attribute verification and topological association initialization to generate geological semantic anchor points with spatiotemporal consistency to be added to the database; use spatiotemporal constraint verification to perform logical consistency verification on the spatial location coordinates and time attribute fields in the geological semantic anchor point instances, remove abnormal instances with spatiotemporal conflicts, and initialize adjacency pointers for the geological semantic anchor point instances that pass the verification to form geological semantic anchor points with spatiotemporal consistency to be added to the database.

[0053] Receive the set of geological semantic anchor point instances generated by S1.3, which includes geological definition name, spatial location coordinates, time attribute field and confidence label, as input objects for spatiotemporal verification and topology initialization.

[0054] Perform spatiotemporal constraint verification, traverse each geological semantic anchor point instance, extract its spatial location coordinate vector and geological time stamp, and construct a four-dimensional spatiotemporal index tree.

[0055] Based on a pre-set geological sequence law and tectonic evolution model, the spatiotemporal logical consistency score between adjacent anchor point instances is calculated, and anomalous instances with stratigraphic inversion or spatiotemporal paradoxes are identified and marked.

[0056] Remove anomalous instances marked as spatiotemporal conflicts, retain valid geological semantic anchors that pass logical consistency verification, and form a cleaned subset of candidate anchors.

[0057] For the retained valid geological semantic anchor points, potential spatial adjacent nodes are retrieved based on their spatial Euclidean distance threshold and geological genetic correlation rules.

[0058] For each effective qualitative anchor point, initialize an adjacency pointer, establish reference links to its spatially neighboring and geologically compatible nodes, and construct a local topological connection structure.

[0059] The initialized adjacency pointers are embedded into the data structure of the geological semantic anchor instance, giving each instance explicit topological context information.

[0060] Through the above spatiotemporal constraint verification and topological association initialization process, the discrete geological semantic anchor point instances generated in the previous step are transformed into geological semantic anchor points to be included in the database with spatiotemporal consistency and local topological connections. This realizes the transformation of data from isolated features to structured geological knowledge units, providing high-quality basic node data for the subsequent construction of a globally inferable anchor point relationship graph.

[0061] S1.5: Structure and index the geological semantic anchor points to be added to the database to generate an initial geological semantic anchor point library; use a graph database storage engine to classify and store the geological semantic anchor points to be added to the database according to geological definitions, and establish a spatial index based on spatial location coordinates and a quality index based on confidence labels. Finally, output the initial geological semantic anchor point library containing complete node attributes and index structure for subsequent steps to build a reasonable anchor point relationship graph.

[0062] Obtain a set of geological semantic anchor point instances with spatiotemporal consistency to be added to the database, which will serve as input objects for structured storage and index construction.

[0063] The graph database storage engine is invoked to initialize the node schema, defining the geological definition name, spatial coordinate triplet, time attribute field, and confidence label as mandatory attribute fields.

[0064] Iterate through the anchor point instances to be added to the database, map each instance to an independent node object in the graph database, and write the corresponding attribute key-value pairs to complete the basic data persistence.

[0065] Based on the geospatial coordinate attributes of nodes, an R-Tree spatial index structure is constructed, and a minimum bounding rectangle (MBR) splitting threshold is set to optimize spatial query efficiency.

[0066] By calculating the confidence variance in the node confidence labels, an inverted index based on quality grading is established, and high-confidence anchors and low-confidence anchors are stored hierarchically.

[0067] For the geological definition name field, a B+ tree classification index is constructed to enable fast retrieval and aggregation by semantic categories such as lithology and structural type.

[0068] The initial adjacency pointer is converted into a directed edge entity in the graph database, associated with the source node and the target node ID, and a topological connection strength weight is attached.

[0069] Verify the integrity of the index structure, ensuring that the spatial index, quality index, and classification index all point to valid node storage addresses, and generate an initial geological semantic anchor point library.

[0070] Through the above-mentioned structured storage and multi-dimensional index construction process, the discrete anchor point instances of the previous step are transformed into a graph database cluster with efficient retrieval capabilities, realizing the expected technical effects of millisecond-level random access and topological traversal of massive geological semantic data in subsequent steps.

[0071] Step S2: Based on the nodes in the initial geological semantic anchor point library and the known geological causal chain relationships, a reasonable anchor point relationship graph is constructed using a graph neural network. This reasonable anchor point relationship graph is independent of coordinate system alignment and uses geological evolution logic as edges. Specifically, it includes:

[0072] S2.1: Obtain the geological semantic unit nodes and predefined geological causal chain relationship data from the initial geological semantic anchor point library, perform structured parsing processing on the geological causal chain relationship data to extract source nodes, target nodes and relationship type triples, and generate a standardized geological causal association set.

[0073] The process begins by reading geological semantic unit node data and predefined geological causal chain relationship records from an initial geological semantic anchor point library as input for structured parsing. Syntax tree traversal and pattern matching are performed on the predefined geological causal chain relationship data to identify source node identifiers, target node identifiers, and relationship type description fields. Source node identifiers are extracted and mapped to their corresponding geological semantic unit node IDs, establishing the starting entity of the causal chain. Target node identifiers are extracted and mapped to their corresponding geological semantic unit node IDs, establishing the ending entity of the causal chain. The relationship type description fields are parsed and standardized into enumerated geological action mechanism labels, such as 'control', 'associated', or 'inhibition'. The extracted source node IDs, target node IDs, and standardized relationship type labels are combined to construct a triplet data structure. Logical consistency checks are performed on all generated triples, eliminating abnormal records where the source and target nodes are the same or the relationship type is undefined. Through this structured parsing process, unstructured geological causal chain relationship data is transformed into a standardized set of geological causal associations, realizing a digital explicit expression of geological evolution logic and providing a standardized edge data foundation for subsequent graph neural network construction.

[0074] S2.2: Based on the geological causal association set, graph embedding processing is used to perform low-dimensional vector mapping processing on each geological semantic unit node to capture the hidden geological attribute features of the node and generate a node feature vector matrix carrying semantic information.

[0075] Receive the standardized geological causal association set generated by S2.1, extract the source node ID, target node ID and relation type label contained therein, and construct the initial node-edge index table.

[0076] For each geological semantic unit node in the initial geological semantic anchor point library, its multidimensional attribute data is parsed, including lithology code, geophysical field strength value, geochemical element abundance and spatial topological coordinates.

[0077] The TransE graph embedding model is used to map discrete geological semantic nodes to a low-dimensional continuous vector space. The node feature vector matrix is ​​initialized and the embedding dimension is set to 128 dimensions to balance computational efficiency and information retention.

[0078] Define a loss function to optimize the node vector representation, ensuring that in the vector space, the sum of the head node vector and the relation vector with causal relationship is as close as possible to the tail node vector. The specific formula is as follows:

[0079]

[0080] Where h, r, and t represent the vector representations of the head node, relation, and tail node, respectively, S is the set of positive sample triples, and γ is the margin hyperparameter. , The error node vector generated by negative sampling. Indicates the distance between positive samples. Indicates the distance to negative samples. Let max(0,x) represent the hinge loss function, which guarantees that the loss is non-negative.

[0081] The node vector and relation vector are iteratively updated using stochastic gradient descent to minimize the loss function L, so that the geometric distance in the vector space reflects the logical tightness of the geological causal chain.

[0082] L2 regularization is introduced to constrain the vector norm, prevent model overfitting, and ensure that the generated node feature vectors have generalization ability and can capture the implicit geological attribute similarity.

[0083] After convergent training, the output is a node feature vector matrix carrying deep semantic information. Each row of the matrix corresponds to a low-dimensional dense vector representation of a geological semantic anchor point.

[0084] By using the TransE graph embedding method, the structured causal relationships of the previous step are transformed into a low-dimensional dense node feature vector matrix, realizing the quantitative mapping of geological semantics from discrete symbols to continuous vector space, and providing a semantically rich input basis for subsequent graph neural network message passing.

[0085] S2.3: Based on the node feature vector matrix and the geological causal chain relationship type, construct a message passing mechanism for a graph neural network, aggregate and update the feature information between adjacent nodes to simulate the stress transmission and attribute diffusion logic in the geological evolution process, and generate a node hidden state representation with context awareness.

[0086] The system receives the node feature vector matrix generated in S2.2 and the geological causal chain relationship type parsed in S2.1, and initializes the message passing mechanism of the graph neural network. An attention-based edge weight calculation module is constructed to calculate the matching score of geological attribute features for each pair of source and target nodes with causal connections. Cosine similarity is used to quantify the directional consistency of the source and target node feature vectors in the semantic space, generating an initial association strength value. A geological evolution time decay factor is introduced, and the initial association strength value is nonlinearly corrected based on the time attribute difference between the source and target nodes to simulate the weakening effect of geological processes over time. The corrected association strength value is normalized using a Softmax function to generate a dynamic message passing weight matrix reflecting the degree of geological causal dependence. Based on the dynamic message passing weight matrix, the feature information of adjacent source nodes is weighted and aggregated. A linear transformation layer is used to map the aggregated neighbor node features to the same latent dimension space as the current node, ensuring the uniformity of the feature space. The mapped neighbor aggregated features are concatenated and fused with the original feature vector of the current node to form a composite feature representation containing local topological context information. By applying nonlinear activation processing to the composite feature representation through a multilayer perceptron network, high-order geological evolution law features are extracted, and the node hidden state representation with context awareness is updated. Through the aforementioned message passing and feature update mechanism, static node features are transformed into dynamic hidden states containing the logic of geological stress transmission, realizing deep modeling and enhanced reasoning ability for complex geological causal relationships.

[0087] S2.4: Calculate the geological evolution logic similarity score between node pairs using the node hidden state representation, filter high-confidence connection edges according to a preset threshold to eliminate redundant or erroneous causal inference paths, and generate an optimized set of geological evolution logic edges.

[0088] Obtain the context-aware hidden state representation of the nodes generated by S2.3, and use it as the basic input data for calculating the similarity of geological evolution logic.

[0089] For any two geological semantic unit nodes, the initial semantic association strength is calculated using cosine similarity to quantify the directional consistency between nodes in the feature space.

[0090] By introducing node confidence labels as a penalty factor, a weighted similarity correction mechanism is constructed to reduce the contribution weight of highly uncertain nodes to the connection relationship and suppress noise interference.

[0091] The geological evolution logical similarity score between node i and node j is calculated using the following formula:

[0092]

[0093] Where S(i,j) is the similarity score between nodes i and j, h i and h j Let U(i,j) be the hidden state vectors of the two nodes, U(i,j) be the joint entropy value of the confidence labels of the two nodes, λ be the uncertainty decay coefficient, and ε be the smoothing term to prevent the denominator from being zero.

[0094] A dynamic threshold screening mechanism is set up to determine the confidence cutoff point based on the prior probability distribution of the geological causal chain, and to remove redundant or erroneous connection edges with similarity scores below the threshold.

[0095] The high-confidence connecting edges are subjected to topological connectivity verification to ensure that the generated edge set conforms to the non-cyclic and transitive constraints of geological evolution logic.

[0096] By using the weighted similarity calculation and dynamic threshold filtering methods described above, the hidden state of the nodes in the previous step is transformed into an optimized set of geological evolution logic edges, thereby achieving the expected technical effect of eliminating false causal inferences and improving the inference accuracy of the anchor point relationship graph.

[0097] S2.5: Perform topological assembly processing based on the geological semantic unit nodes and the geological evolution logical edge set to construct a graph structure model that does not depend on geographic coordinate system alignment, and generate an anchor point relationship graph with reasoning capabilities.

[0098] Obtain the optimized set of geological evolution logic edges and geological semantic unit nodes, and initialize the node container and edge connection table of the graph structure model.

[0099] The unique identifier of the geological semantic unit node is mapped to the vertex index in the graph topology, establishing a direct reference relationship between node attributes and graph locations.

[0100] Traverse the set of edges in the geological evolution logic, resolve the source node ID and target node ID of each edge, and register the directed connection relationship in the graph connection table.

[0101] Initial weight values ​​are assigned based on edge type, and a weighted adjacency matrix reflecting the intensity of geological causality is constructed to ensure that edge directions conform to the temporal constraints of geological evolution.

[0102] Perform graph connectivity detection processing to identify and remove isolated nodes or unreachable weakly connected components, ensuring the overall topological integrity of the anchor point relationship graph.

[0103] The graph structure is standardized to unify the dimension of node features and the weight of edge weights, thus eliminating numerical scale bias caused by differences in data sources.

[0104] Through the above topological assembly process, discrete nodes and logical edges are transformed into anchor point relationship graphs with rigorous causal structures, realizing the construction of geological semantic space independent of geographic coordinate systems, and providing a reasonable structural basis for subsequent dynamic subgraph activation.

[0105] like Figure 2 As shown, step S3 involves matching newly accessed local geological features extracted from the original geological data to neighboring anchor points in the initial geological semantic anchor point library, activating the corresponding geological relationship sub-graph based on the current regional geological evolution stage, and calculating the dynamic confidence weight of the anchor points. The dynamic confidence weight of the anchor points reflects the dynamic changes in the tectonic background. Specifically, this includes:

[0106] S3.1: Obtain the newly accessed local geological feature vector set, and perform proximity retrieval processing on each feature unit in the local geological feature vector set based on multi-dimensional spatial indexing to select a set of candidate geological semantic anchor points that meet preset conditions from the initial geological semantic anchor point library. The preset conditions include: spatial overlap and highest semantic similarity.

[0107] The local geological features refer to multimodal information units extracted from raw geological data that characterize geological attributes at a specific spatial location, including remote sensing spectral reflectance, gravity and magnetic anomaly gradient values, and geochemical element concentrations. The local geological feature vector set refers to a structured data set organized according to a unified geodetic coordinate system after spatial coordinate standardization of multiple local geological feature units. Its acquisition method is as follows: preprocessing is performed on the raw geological data to extract multimodal data such as remote sensing spectral reflectance, gravity and magnetic anomaly gradient values, and geochemical element concentrations at each spatial location; spatial coordinate standardization is then performed on each data unit, mapping it to a geodetic coordinate system consistent with the initial geological semantic anchor point library, thereby generating the local geological feature vector set.

[0108] A newly acquired local geological feature vector set is obtained, containing preprocessed multimodal data units such as remote sensing spectral reflectance, gravity and magnetic anomaly gradient values, and geochemical element concentrations. Spatial coordinate standardization is performed on each feature unit in the local geological feature vector set, mapping it to a geodetic coordinate system consistent with the initial geological semantic anchor point library, eliminating spatial reference bias caused by differences in data acquisition equipment. A multidimensional spatial index structure based on R-tree is constructed, hierarchically partitioning and storing anchor nodes in the initial geological semantic anchor point library according to their spatial location coordinates, establishing a spatial topology grid for fast retrieval. For each local geological feature unit to be matched, a radius neighborhood search is performed in the initial geological semantic anchor point library using multidimensional spatial indexing, filtering out all candidate anchor points with spatial distances less than a preset threshold, forming a preliminary subset of spatially overlapping anchor points. The semantic similarity between the local geological feature unit and each spatially overlapping candidate anchor point is calculated, and the cosine similarity metric is used to evaluate the directional consistency of the feature vectors in the implicit geological semantic space. A semantic similarity threshold is set to eliminate pseudo-matching anchor points that are spatially adjacent but have significantly different geological attributes, retaining several anchor points with overlapping spatial locations and the highest semantic similarity. The selected anchor points and their corresponding similarity scores are encapsulated into structured data objects, generating a candidate geological semantic anchor point set carrying spatial association attributes and semantic confidence. Through a multi-dimensional spatial index retrieval and semantic similarity dual verification mechanism, the newly accessed local features from the previous step are transformed into a candidate anchor point set with high-confidence spatial-semantic association, achieving accurate mapping and preliminary alignment of multi-source heterogeneous data to a unified geological semantic space. For example, for a newly accessed hyperspectral remote sensing data unit in a mining area, its center coordinates are latitude and longitude (105.23, 30.45), and its spectral feature vector is 32-dimensional. An R-tree index is constructed, with a node capacity of 10 and a minimum overlap area of ​​0.1. A neighborhood search with a radius of 500 meters is performed, retrieving 15 spatially adjacent anchor points in the initial anchor point database. The cosine similarity between the remote sensing unit and 15 candidate anchor points was calculated. Anchor point A (silicification alteration zone) had a similarity of 0.92, anchor point B (fracture zone) had a similarity of 0.88, and the similarity of the remaining anchor points was all below 0.75. A similarity threshold of 0.8 was set, and anchor points A and B were selected as the final set of candidate geological semantic anchor points. This process effectively eliminated interference terms that were spatially adjacent but had inconsistent lithology, significantly improving the accuracy of subsequent dynamic weight calculations.

[0109] S3.2: Obtain the candidate geological semantic anchor point set and the current regional geological evolution stage identifier parameters, perform temporal consistency verification on the node attributes in the candidate geological semantic anchor point set, and generate a list of active geological relationship subgraph nodes that conform to the current regional geological evolution stage identifier parameters.

[0110] The system receives a set of candidate geological semantic anchor points and the current regional geological evolution stage identifier parameters as input data for temporal consistency verification. It parses the ontology definition library in the geological period mapping rule engine, extracts the allowed geological time interval constraints for each semantic anchor point type, and constructs a time validity judgment matrix. It iterates through the candidate geological semantic anchor point set, reading the time attribute field carried in each anchor point instance, which contains the absolute age value of the strata or the relative geological time label. It logically compares the read time attribute with the allowed time interval of the corresponding anchor point type, calculating the time deviation metric. The time consistency score is calculated using the following formula:

[0111]

[0112] Where S is the time consistency score. For anchor point time attributes, This is the central time of the current evolutionary stage. This represents the characteristic duration span of the geological period. The calculated scores are filtered based on a preset temporal consistency threshold, eliminating spatiotemporal conflict anomalies where scores are below the threshold. The verified anchor point instances are retained, and their unique node identifiers are extracted to form a list of active geological relationship sub-graph nodes that conform to the identifier parameters of the current regional geological evolution stage. Through temporal consistency verification processing by the geological period mapping rule engine, the candidate set from the previous step is transformed into an active node list with temporal rationality, achieving the expected technical effect of feature filtering and precise activation under dynamic geological backgrounds.

[0113] S3.3: Obtain the list of nodes in the active geological relationship subgraph and the topological connection data of the inferable anchor point relationship graph. Based on the graph traversal algorithm, perform subgraph reconstruction processing on the nodes in the list of nodes in the active geological relationship subgraph and their edge relationships in the inferable anchor point relationship graph to generate a dynamic geological relationship subgraph structure limited to the geological evolution stage identifier parameters of the current region.

[0114] Obtain the list of active geological relationship subgraph nodes and the topological connection data of the inferable anchor point relationship graph, as the initial input object for subgraph reconstruction.

[0115] Based on breadth-first search, starting from each node in the list of active nodes, a hierarchical traversal operation is performed in the inferable anchor relationship graph to retrieve the set of first-order neighbor nodes directly connected to the starting node.

[0116] A second temporal consistency check is performed on the retrieved first-order neighbor nodes, and nodes from different periods that do not belong to the time window allowed by the current region's geological evolution stage identifier parameter are removed to ensure the temporal logic self-consistency within the subgraph.

[0117] Based on the preset maximum hop count threshold, the traversal continues to expand into the deeper network, capturing second- and third-order neighbor nodes, and constructing a candidate subgraph node pool containing multi-hop geological causal chains.

[0118] Extract the original edge weights and relationship type attributes of all node pairs in the candidate subgraph node pool in the inference anchor point relationship graph to form a local topological connection matrix.

[0119] A graph pruning strategy is adopted to remove weakly correlated edges whose connection strength is lower than the dynamic confidence threshold, thereby eliminating the impact of noise interference on the stability of the subgraph structure.

[0120] The retained strongly correlated nodes and edges are re-indexed and mapped to construct a compact adjacency list structure, generating a dynamic geological relationship subgraph structure limited to the identifier parameters of the current regional geological evolution stage.

[0121] By using the graph traversal and pruning methods described above, the discrete active nodes from the previous step are transformed into a dynamic geological relationship subgraph structure with complete causal chains and spatiotemporal consistency. This enables the accurate reconstruction of the correlation network of key ore-forming elements under complex geological backgrounds, providing a high signal-to-noise ratio topological foundation for subsequent stress transfer calculations.

[0122] S3.4: Obtain the local geological feature response intensity values ​​of the dynamic geological relationship subgraph structure and each data source. Perform influence diffusion calculation on the nodes in the dynamic geological relationship subgraph structure through stress transfer path recursive propagation processing to generate the original confidence score sequence characterizing the importance of each candidate geological semantic anchor point in a specific tectonic background.

[0123] The local geological feature response intensity value refers to the quantitative value extracted from the local geological features, characterizing the intensity of geological attributes of each data source at a specific spatial location. It is obtained by performing amplitude normalization processing on each feature unit in the local geological feature vector set, including remote sensing spectral reflectance, gravity and magnetic anomaly gradient values, and geochemical element concentrations, mapping the original measurement values ​​of different dimensions to a unified numerical range, thereby generating the local geological feature response intensity value.

[0124] The dynamic geological relationship subgraph structure and the local geological feature response intensity values ​​of each data source are obtained as the initial boundary conditions for influence diffusion calculation. The topological adjacency matrix of each semantic anchor node in the subgraph is extracted to identify direct causal connection paths and indirect transmission links between nodes. The original confidence score of each node is initialized as a zero vector, and the local geological feature response intensity values ​​are mapped to the corresponding source nodes, setting them as the initial energy injection points for the diffusion process. A tectonic stress transmission attenuation function is defined, which quantifies the energy loss characteristics of stress propagation in the rock medium based on the type weight of the geological causal chain between nodes and the spatial topological distance. Recursive iteration is used to execute multiple rounds of message passing. In each iteration, the weighted sum of responses received by the current node from all neighboring nodes is calculated. For the confidence update of any target node i in the t-th iteration, the following formula is used for calculation:

[0125]

[0126] in, This represents the cumulative confidence score of node i after the t-th iteration. Let represent the initial value of the local geological feature response intensity of node i, λ be the global damping factor used to suppress noise accumulation, and N(i) represent the set of direct neighbor nodes of node i. This represents the edge weight parameter from node j to node i, reflecting the efficiency of geological stress transfer. This represents the confidence score of neighbor node j in the previous iteration. A convergence threshold and a maximum number of iterations are set, and the Euclidean distance change of the confidence vector of all nodes in the whole map is monitored between adjacent iterations. The recursive propagation process terminates when the distance change is less than the preset threshold or the maximum number of iterations is reached. The confidence scores of each candidate geological semantic anchor point in the final converged state are extracted to form the original confidence score sequence. Through recursive propagation processing via stress transfer paths, the dynamic subgraph structure and local response values ​​generated in the previous step are transformed into the original confidence score sequence characterizing the importance of each anchor point in a specific tectonic context, thus realizing the dynamic constraint of geological evolution logic on data weights.

[0127] S3.5: Obtain the original confidence score sequence and confidence labels, perform attenuation correction processing on each score value in the original confidence score sequence, and output the final anchor dynamic confidence weight vector that reflects the dynamic changes in the constructed background. In this embodiment, attenuation correction processing is performed on each score value in the original confidence score sequence based on a Bayesian probability model.

[0128] Obtain the original confidence score sequence output from step S3.4 and the confidence labels for each candidate geological semantic anchor point generated in step S1.3, and use them as input parameters for the Bayesian probability model. Construct a Bayesian correction framework based on a Gaussian prior distribution, treating the original confidence scores as observational likelihoods and transforming the confidence labels into prior variance parameters. Calculate the corrected dynamic confidence weights using the following formula:

[0129]

[0130] in, The corrected anchor point dynamic confidence weight. The original confidence score. To observe signal strength, Let μ be the prior variance. prior The prior mean, To measure the observation noise variance, the prior variance is determined by inverse mapping based on the attribute confidence values ​​in the confidence labels. Lower confidence values ​​result in larger variances, indicating more ambiguous knowledge of the initial state of the anchor point. Based on the statistical characteristics of the residual distribution during multi-source data fusion, the observation noise variance is estimated to quantify the level of random error in the current local geological feature matching process. Element-wise matrix operations are performed, substituting each scalar value from the original scoring sequence into the Bayesian update formula to calculate the attenuated posterior probability estimate. The calculated posterior probability estimate is then normalized to ensure that the sum of the dynamic confidence weights of all candidate anchor points is a unit of 1, forming a weight vector with a valid probability distribution. Through Bayesian probability correction, the original confidence scores from the previous step are transformed into a statistically robust dynamic confidence weight vector for anchor points, achieving automatic suppression of low-quality or highly uncertain geological semantic anchor points and significantly improving feature selection accuracy and model generalization ability during multi-source heterogeneous data fusion.

[0131] like Figure 3 As shown, step S4 involves using the dynamic confidence weights of the anchor points to distribute the response intensity of each data source on the inferable anchor point relationship graph, generating an anchor point response spectrum. The anchor point response spectrum characterizes the activation state of each data source for each anchor point in the semantic anchor point library. Specifically, it includes:

[0132] S4.1: Based on the node topology in the inferable anchor point relationship graph, perform neighbor anchor point matching processing on the local geological features extracted from various original geological data to generate an initial anchor point matching set carrying spatial correlation attributes, which serves as the basic input object for subsequent weight application.

[0133] The output of step S3 is obtained as follows: the dynamic confidence weight vector of anchor points reflecting the dynamic changes in the tectonic background, and the set of local geological feature vectors extracted from multi-source heterogeneous raw geological data. For each semantic anchor point node in the inferable anchor point relationship graph, its spatial coordinate attributes and topological adjacency matrix are extracted, and a multi-dimensional spatial index structure based on KD-Tree is constructed to support fast proximity retrieval in high-dimensional feature space.

[0134] For each type of raw geological data (including remote sensing images, geophysical profiles, geochemical spectral lines, and borehole core descriptions), a standardized preprocessing method is performed to eliminate dimensional differences and generate a feature query vector with a unified dimension. Using multidimensional spatial indexing, the Euclidean distance between the feature query vector and the spatial coordinates of each anchor point in the initial geological semantic anchor point database is calculated, and a set of candidate anchor points that overlap or are adjacent in spatial location is selected.

[0135] Based on the candidate anchor point set, the cosine similarity between the local geological feature vector and the historical feature prototype of each candidate anchor point is calculated to quantify the semantic matching degree. A similarity threshold is set to eliminate low-confidence matches and retain high-similarity anchor points as valid matching objects. The inverse of spatial distance and semantic similarity are weighted and fused to generate an initial matching score, establishing a preliminary mapping relationship between local geological features and semantic anchor points.

[0136] Each valid match is labeled with a spatial association attribute, recording its data source type, original spatial coordinates, and matching timestamp, forming an initial anchor point matching set carrying complete contextual information. This set serves as the base input for subsequent weighting, ensuring the accurate placement of multi-source data in the semantic space.

[0137] Through the aforementioned neighboring anchor point matching process, discrete multi-source local geological features are transformed into a structured initial anchor point matching set with spatial attributes, achieving preliminary alignment from the original data domain to the geological semantic domain, and providing an accurate data basis for subsequent dynamic weight modulation.

[0138] S4.2: Using the spatial correlation attributes in the initial anchor point matching set and the dynamic confidence weight of the anchor points calculated in the current regional geological evolution stage, a weighted modulation operation is performed on the initial anchor point matching set to generate a weighted anchor point response intensity matrix that reflects the dynamic changes in the tectonic background.

[0139] Obtain the spatial association attribute data from the initial anchor point matching set and the anchor point dynamic confidence weight vector output from step S3, and construct the data input environment for weighted modulation operation.

[0140] The spatial topological index of each match in the initial anchor matching set is parsed, the node identifier of the corresponding semantic anchor in the inferable anchor relationship graph is extracted, and a mapping index table between discrete feature points and semantic space is established.

[0141] Read the weight coefficient corresponding to the node identifier in the dynamic confidence weight vector of the anchor point. This coefficient represents the contribution of the semantic anchor point to the mineralization control at the current geological evolution stage.

[0142] The Hadamard product operation is used to multiply the original local geological feature response intensity values ​​element-wise with the corresponding dynamic confidence weights, thereby achieving dynamic modulation of data features by the tectonic background. The specific formula is as follows:

[0143]

[0144] in, Here, W represents the weighted anchor response strength matrix, and W is the dynamic confidence weight vector of the anchor points. Given the original response intensity vector in the initial anchor matching set, diag(W) represents converting the weight vector into a diagonal matrix to perform a linear transformation.

[0145] The weighted response intensity values ​​are subjected to non-negativity constraint verification to eliminate negative responses caused by noise interference, ensuring the rationality of the physical meaning.

[0146] The verified weighted response strength values ​​are matrix-reorganized according to the data source type and anchor point ID to generate a two-dimensional weighted anchor point response strength matrix with rows and columns corresponding to the data source and semantic anchor point, respectively.

[0147] By using a weighted modulation operation, the discrete matching result of the previous step is transformed into a weighted anchor point response intensity matrix that reflects the dynamic changes of the construction background, thereby realizing adaptive quantitative representation of multi-source data in a unified semantic space.

[0148] S4.3: Based on the weighted anchor response intensity matrix, the discrete response values ​​within the matrix are processed through continuous distribution mapping to generate a local anchor response density field characterizing the probability distribution characteristics of single-source data in the semantic space. In this embodiment, the Gaussian kernel function density estimation method is used to process the discrete response values ​​within the matrix through continuous distribution mapping.

[0149] Obtain the weighted anchor response intensity matrix generated in step S4.2. This matrix contains discretely distributed semantic anchors and their corresponding weighted response values, which serve as the input data source for density estimation.

[0150] A set of Gaussian kernel functions is constructed with the coordinates of semantic anchor points in a unified geological semantic space as independent variables, and the kernel bandwidth parameter is set to control the smoothness of local response and the balance between resolution.

[0151] For each discrete anchor point response value in the matrix, calculate its Euclidean distance to any sampling point in the semantic space, and substitute it into the Gaussian kernel function formula to perform probability density mapping:

[0152]

[0153] Where x is an arbitrary position vector in the semantic space, x i Let K(x, x) be the coordinate vector of the i-th discrete anchor point, σ be the kernel bandwidth parameter, d be the dimension of the semantic space, and K(x, x) be the coordinate vector of the i-th discrete anchor point. i ): Gaussian kernel function value (i.e. probability density value).

[0154] By traversing the semantic space grid, the kernel function contribution values ​​of all discrete anchor points are linearly superimposed to form a continuous probability density distribution surface.

[0155] An adaptive bandwidth adjustment mechanism is introduced to dynamically adjust the σ value of the local region based on the confidence label of the anchor point. In the low confidence region, the smoothing range is expanded to suppress noise, and in the high confidence region, the bandwidth is reduced to preserve detailed features.

[0156] The density field after superposition is truncated at the boundary to remove artifacts caused by the tailing effect of the kernel function at the edge of the semantic space, thus ensuring the physical authenticity of the density field.

[0157] By using the Gaussian kernel function density estimation method, the discrete weighted response matrix from the previous step is transformed into a continuous and smooth local anchor point response density field, realizing the mapping from discrete sampling points to a continuous probability distribution space and eliminating the quantization error caused by data sparsity.

[0158] S4.4: Perform a normalized integral transformation operation on the local anchor point response density field to eliminate the dimensional differences between different data sources, thereby generating a standardized anchor point activation probability distribution vector with cross-modal comparability.

[0159] The local anchor point response density field data generated by S4.3 is obtained. This data represents the discrete probability distribution characteristics of single-source geological data in semantic space and is used as the input object for normalized integral transform.

[0160] A unified dimensional mapping function for multi-source heterogeneous data is constructed. Based on the physical attribute definitions of each anchor point in the geological semantic anchor point library, the upper and lower boundary conditions of the integral transformation are determined to ensure the comparability of different modal data on the same semantic dimension.

[0161] The Riemann integral approximation is performed on the local anchor point response density field to transform the continuous density distribution curve into a cumulative probability value, thus eliminating the dimensional differences between remote sensing spectral intensity, geophysical field strength, and geochemical concentration.

[0162] The standardized probability distribution transformation is performed using the following formula:

[0163]

[0164] Among them, P i Let f(t) be the standardized activation probability of the i-th semantic anchor, f(t) be the local anchor response density function, and x be the normalized activation probability of the i-th semantic anchor. i is the response intensity value of the current anchor point, and M is the maximum theoretical threshold of the response intensity.

[0165] The calculated cumulative probability value is standardized using Z-Score, which subtracts the mean and divides by the standard deviation to further eliminate statistical bias caused by different sampling densities of the data source, generating an intermediate state vector with zero mean and unit variance.

[0166] The standardized vector is truncated to map outliers that exceed the preset confidence interval [0,1] to boundary values, ensuring that the output vector conforms to the physical constraints of the probability distribution and preventing noise from interfering with subsequent aggregation calculations.

[0167] By using a normalized integral transform method, the local anchor point response density field of the previous step is transformed into a standardized anchor point activation probability distribution vector with cross-modal comparability, thereby achieving dimensionless alignment of multi-source heterogeneous data in a unified semantic space.

[0168] S4.5: Integrate the standardized anchor activation probability distribution vectors generated by all data sources, encapsulate them in a structured manner according to the data source type identifier, and finally generate the anchor response spectrum representing the activation state of each anchor in the semantic anchor library for each data source, thus completing the morphological transformation from discrete features to continuous spectrum lines.

[0169] Step S5: Based on the inferable anchor point relationship graph, the response spectra of multiple anchor points are aggregated through recursive propagation of stress transfer paths to generate a global semantic consistency vector. This global semantic consistency vector possesses both physical interpretability and topological consistency. Specifically, it includes:

[0170] S5.1: Obtain the node connection matrix and the anchor response spectrum in the inferable anchor relationship graph, extract the set of direct neighbor nodes and the corresponding edge weight parameters of each target semantic anchor based on the graph theory adjacency traversal mechanism, and generate an anchor neighborhood structure descriptor containing local topological connection relationships.

[0171] The node connection matrix refers to a two-dimensional matrix structure generated after mathematically abstracting the connection relationships between semantic anchor nodes in the inferable anchor relationship graph. Its row and column indices correspond to different semantic anchor nodes in the graph, and the element values ​​in the matrix represent the edge weight parameters between the corresponding row and column nodes. These edge weight parameters characterize the strength or physical connectivity of the geological causal chain. It is obtained by traversing all node pairs in the inferable anchor relationship graph. If a direct connection edge exists between two nodes, the corresponding edge weight value is filled into the corresponding position in the matrix; if no direct connection edge exists, the corresponding position is set to zero, thus generating the node connection matrix.

[0172] Obtain the node connection matrix of the inferable anchor point relationship graph and the anchor point response spectrum generated by S4, and use them as the initial input data for topology aggregation. Parse the row and column indices of the node connection matrix to establish a mapping index table between semantic anchor points and their direct neighbor nodes, clarifying the local neighborhood range of each node in the graph structure. Traverse the mapping index table, and for each target semantic anchor point, extract the set of IDs of all its directly connected neighbor nodes to construct a local neighborhood node list. Read the edge weight parameters between the target node and each neighbor node from the node connection matrix; these values ​​represent the strength of the geological causal chain or physical connectivity. Structure the extracted neighbor node ID set with the corresponding edge weight parameters to form key-value pairs of local topological data units. Integrate the local topological data units of all target semantic anchor points to generate an anchor point neighborhood structure descriptor containing complete adjacency relationships and weight parameters.

[0173] S5.2: Using the edge weight parameters in the anchor point neighborhood structure descriptor, combined with the preset structural stress attenuation function, the local geological feature response intensity values ​​in the direct neighbor node set are subjected to distance-weighted correction processing to simulate the non-uniform diffusion effect of geological stress in the rock medium, and a weighted neighborhood response vector reflecting the local stress field distribution characteristics is generated.

[0174] Obtain the set of direct neighbor nodes and their corresponding edge weight parameters from the anchor point neighborhood structure descriptor, and use them as the topological basis input for stress diffusion calculation.

[0175] The geological causal edge types connecting the target anchor point and neighboring nodes in the inferable anchor point relationship graph are analyzed to identify them as physical media properties such as tensional fractures, compressive folds, or hydrothermal permeation channels.

[0176] Based on the identified medium properties, the corresponding tectonic stress attenuation coefficient is retrieved from the pre-set geophysical parameter library to establish a mapping relationship between edge weights and stress transmission efficiency.

[0177] The Euclidean distance between the target anchor point and each of its direct neighbor nodes in the three-dimensional geological space is calculated, and the equivalent geological transmission distance is obtained by combining the rock porosity parameter.

[0178] An exponential decay function is used to construct a structural stress transfer model. The equivalent geological transmission distance is substituted into the decay formula to calculate the stress retention rate of a single path. The specific formula is as follows:

[0179]

[0180] Where W is the corrected stress transfer weight, w edge λ represents the initial edge weights based on geological causality, d represents the equivalent geological transmission distance, and λ represents the characteristic attenuation length of the medium.

[0181] Multiply the calculated stress transfer weight by the local geological feature response intensity value of the neighbor node in the anchor point response spectrum to obtain the neighbor contribution value modulated by both distance and medium.

[0182] By traversing all direct neighbor nodes and summing the modulated contribution values ​​of each node, a weighted neighborhood response vector reflecting the non-uniform distribution characteristics of the local stress field is formed.

[0183] By constructing a stress attenuation function and performing distance-weighted correction on the responses of neighboring nodes, the topological description from the previous step is transformed into a physically meaningful weighted neighborhood response vector, thereby achieving a quantitative characterization of the geological stress diffusion effect.

[0184] S5.3: Based on the weighted neighborhood response vector and the self-response value of the current target semantic anchor point, perform multi-round recursive iterative propagation calculations to generate a convergent and stable multi-level stress transfer accumulation. The multi-round recursive iterative propagation calculation includes: using the intermediate aggregation state generated in the previous iteration as the input boundary condition for the next iteration; accumulating the contribution values ​​of neighboring nodes and introducing a damping factor to suppress noise accumulation to generate the convergent and stable multi-level stress transfer accumulation. The self-response value refers to the local geological feature response intensity value of the target semantic anchor point itself.

[0185] Obtain the weighted neighborhood response vector generated by S5.2 and the self-response value of the current target semantic anchor point as the initial boundary conditions for recursive iterative propagation calculation.

[0186] Initialize the multi-level stress transfer cumulative matrix, assign the self-response value of the target anchor point to the zeroth-order iterative state, and establish the energy benchmark of local geological features.

[0187] Construct a recursive propagation operator based on graph topology, and define the contribution coefficient of neighboring nodes to the central node based on the edge weight parameters in the inferable anchor node relationship graph.

[0188] The first round of iterative calculation is performed by traversing the direct neighbor nodes of the target anchor point, extracting their weighted neighborhood response intensity, and multiplying it by the corresponding structural stress attenuation factor.

[0189] The intermediate aggregation state of the k-th iteration is calculated using the following formula:

[0190]

[0191] in, Let be the cumulative amount in the k-th iteration. Let be the edge weight between nodes i and j, N(i) be the set of neighbors of node i, and α be the damping factor. This represents the cumulative stress transfer amount of neighbor node j at the (k-1)th iteration. Let be the initial self-response value of the target node i.

[0192] A damping factor is introduced to suppress the accumulation of long-range noise. The ratio of local self-response preservation to neighborhood information fusion is balanced by adjusting the α value to prevent over-smoothing.

[0193] The intermediate aggregate state generated in the previous round is used as the input boundary condition for the next round of iteration to update the response contribution value of neighboring nodes, thereby realizing the deep diffusion of the geological stress field.

[0194] The Euclidean distance between two adjacent iterations is monitored, and when the difference is less than the preset convergence threshold, the multi-stage stress transfer process is determined to have reached a stable state.

[0195] The output is the cumulative amount of multi-level stress transfer after convergence. This data integrates the geological semantic information of multi-hop neighbor nodes and tectonic background constraints.

[0196] By using recursive iterative propagation and damping suppression, the local weighted response of the previous step is transformed into globally convergent stress transfer accumulation data, achieving the expected technical effects of topological consistency aggregation of geological features in semantic space and enhanced noise robustness.

[0197] S5.4: Apply geological rule constraint mask to the accumulated amount of multi-level stress transmission for logical verification processing, remove abnormal transmission path components that violate the direction of known geological causal chain, and perform normalization mapping operation on the remaining effective components to generate a corrected global semantic fusion vector that conforms to the geological evolution logic and has topological consistency.

[0198] The geological rule constraint mask is a pre-constructed filtering structure based on known geological causal chain directions and topological reachability constraints. It is implemented in matrix or vector form, and its elements are used to identify the validity of stress / information transmission paths. The geological rule constraint mask is constructed as follows: based on the geological causal edge types and directional relationships recorded on each edge of the inferable anchor point relationship graph—that is, the stress / information transmission direction and medium connectivity attributes between anchor points—and combined with the time validity constraints of causal types on the current regional geological evolution stage, a set of causal transmission directions conforming to geological evolution laws is extracted offline and mapped to a mask representation of the same dimension as the multi-level stress transmission accumulation. Positions with a mask value of 1 correspond to valid transmission components that should be retained, while positions with a mask value of 0 correspond to invalid transmission components that should be masked.

[0199] Obtain the multi-level stress transfer cumulative vector and the preset geological rule constraint mask matrix generated by S5.3. The mask matrix is ​​based on the geological causal chain directionality definition and identifies the information flow between allowed and prohibited semantic nodes.

[0200] Element-wise multiplication is performed on the multi-level stress transfer cumulative vector. Abnormal transfer path components that violate the known geological causal chain direction are set to zero, while valid components that conform to the geological evolution logic are retained, forming a semantic feature vector after preliminary filtering.

[0201] Sparseness check is used to detect isolated high-value nodes in the vector after initial filtering. If the response intensity of a node exceeds a preset threshold but lacks upstream causal node support, it is judged as noise interference and is removed to ensure the continuity of the topology.

[0202] A normalized mapping function is constructed, and the L2 norm of the remaining effective components is calculated as a scaling factor to scale the values ​​of each dimension of the vector proportionally, eliminating the dimensional drift caused by the difference in the number of recursive iterations, and generating a corrected global semantic fusion vector with topological consistency.

[0203] By using logical verification and normalization mapping of geological rule constraint masks, the multi-level stress transfer accumulation in the previous step is transformed into a corrected global semantic fusion vector that conforms to geological evolution logic and has topological consistency, thereby significantly improving the physical interpretability and data purity of the prediction results.

[0204] S5.5: The modified global semantic fusion vector is encapsulated by adding spatiotemporal derivative labels. The values ​​of each dimension in the modified global semantic fusion vector are bound and recombined with their corresponding geological semantic anchor point definitions and confidence labels. Finally, a global semantic consistency vector containing complete physical meaning and dynamic evolution information is output.

[0205] The spatiotemporal derivative label refers to a set of auxiliary annotation information added to each component in the corrected global semantic fusion vector corresponding to the geological semantic anchor point dimension, in order to characterize the rate of change of the anchor point response or confidence weight in the time dimension and the gradient direction / magnitude of change along the legitimate geological causal chain.

[0206] The method of obtaining it is as follows: the response value / weight value of the same geological semantic anchor point at the current time and the previous time or the previous sliding time window has been cached, and the difference is divided by the time interval between the two times to obtain the time derivative label; at the same time, within the inferable anchor point relationship graph, and only in the causal transmission direction allowed by the geological rule constraint mask, the spatial gradient direction label is generated by the difference in response value between the center anchor point and the upstream neighbor anchor point. Then, the above time derivative and spatial gradient direction label are used as the spatiotemporal derivative label.

[0207] Step S6: Based on the global semantic consistency vector and its spatiotemporal derivative, perform mineral resource distribution deduction and output resource distribution prediction results. The resource distribution prediction results include key anchor point combinations and evolutionary logic chains. Specifically, they include:

[0208] S6.1: Obtain the global semantic consistency vector and its spatiotemporal derivative, and use multidimensional feature projection to perform high-dimensional spatial mapping processing on the global semantic consistency vector to generate an initial mineralization potential feature tensor carrying dynamic changes in geological processes, providing a standardized data base for subsequent rule decoding.

[0209] The system receives the global semantic consistency vector and its spatiotemporal derivative from the output of step S5. This vector contains the comprehensive activation state and geological evolution rate information of each semantic anchor. A high-dimensional nonlinear projection space is constructed, using a multilayer perceptron architecture as the mapping core. The number of input layer nodes is kept consistent with the dimension of the semantic anchor library to ensure that each anchor feature enters the network independently. A linear transformation is performed on the input vector, mapping the low-dimensional semantic features to the hidden layer through a weight matrix. The weight parameters are initialized according to the Xavier distribution to maintain gradient stability. A nonlinear activation function is introduced to process the hidden layer output, using the modified linear unit ReLU to eliminate negative responses and preserve the sparsity of the geologically positive correlation features. A second nonlinear mapping is performed to project the hidden layer features to a higher-dimensional mineralization potential feature space. The dimension is set to three times the original number of anchors to capture higher-order interaction relationships. Batch normalization is performed on the high-dimensional features, calculating the mean and variance of the current batch of data to standardize the feature distribution, accelerating convergence and suppressing internal covariate shifts. By fusing spatiotemporal derivative information, the temporal rate of change and spatial gradient of the global semantic consistency vector are used as auxiliary channels and injected into specific slices of the high-dimensional feature tensor through a concatenation operation. An attention mechanism is applied to weight the spatiotemporal components, calculating the correlation coefficient between derivative information and static semantic features, and dynamically adjusting the contribution ratio of spatiotemporal information in the final tensor. An initial mineralization potential feature tensor is generated, which structurally stores geological feature data that has undergone high-dimensional mapping and spatiotemporal enhancement. Through multi-dimensional feature projection processing, the global semantic vector from the previous step is transformed into an initial mineralization potential feature tensor carrying dynamic information about geological processes, realizing the transformation from static semantic representation to a dynamic mineralization potential base, providing a standardized data foundation rich in evolutionary logic for subsequent rule decoding.

[0210] For example, the global semantic consistency vector has a dimension of 128, corresponding to 128 geological semantic anchor points. A three-layer fully connected neural network is constructed, with the hidden layer dimension set to 384. The input vector undergoes a first-layer linear transformation, with a weight matrix size of 128x384 and the bias term initialized to zero. After ReLU activation, a second-layer linear transformation maps the dimension to a high-dimensional space of 768. In batch normalization, the momentum parameter is set to 0.9, and the epsilon value is 1e-5. The spatiotemporal derivative vector is also 128-dimensional, and after concatenation with the static features, it forms an 896-dimensional intermediate vector. The attention mechanism calculates the spatiotemporal weight coefficient to be 0.35 and the static weight to be 0.65. The final generated initial mineralization potential feature tensor size is [batch size

[768] , where each element represents the activation intensity of a specific geological assemblage in high-dimensional space. This tensor effectively integrates the dynamic coupling information of the Yanshanian tectonic stress field and hydrothermal activity intensity, significantly improving the characterization ability of deep concealed ore bodies and greatly enhancing the matching accuracy of subsequent geological rule decoders.

[0211] S6.2: Based on the initial mineralization potential feature tensor, the tectonic stress transfer logic library in the pre-built geological rule decoder is invoked to perform causal chain matching processing on the abnormal response regions in the feature tensor, generating a set of candidate mineralization models that conform to the laws of geological evolution, ensuring that the prediction process is controlled by a clear geophysical mechanism.

[0212] The system receives the initial mineralization potential feature tensor generated by S6.1. This tensor contains the distribution of geological attributes and dynamic evolution gradient information in a high-dimensional semantic space. A pre-built geological rule decoder is invoked, and a tectonic stress transfer logic library is loaded. This library stores physical constraint rules and causal topological templates for geological processes such as fault activity, magma intrusion, and hydrothermal migration. Local extremum detection is performed on the feature tensor to identify abnormal response regions with response intensities exceeding a preset threshold, marking them as candidate nodes for potential mineralization areas. For each candidate node, the activation state sequence of its corresponding semantic anchor point in the global semantic consistency vector is extracted as the query key for causal chain matching. The geological evolution pattern template with the highest semantic similarity to the query key is retrieved from the tectonic stress transfer logic library, and the source-sink relationships, stress transmission paths, and temporal constraints defined in the template are obtained. The spatiotemporal derivative information of the candidate node is compared with the temporal constraints in the template, and a temporal consistency score is calculated, eliminating matches with temporal logical conflicts. Subgraph matching is performed between the local semantic topology around the candidate node and the causal chain structure in the template, and a structural similarity matrix is ​​calculated. Based on the structural similarity matrix and temporal consistency score, a multidimensional matching scoring function is constructed to weight and rank candidate mineralization models. Model combinations with scores above a confidence threshold are selected to form a preliminary set of candidate mineralization models. Physical mechanism verification is performed on each model in the set to verify whether it conforms to the principles of mass conservation, energy conservation, and geomechanical equilibrium, eliminating spurious correlation models that violate physical laws. Through this processing, the feature tensor from the previous step is transformed into a set of candidate mineralization models that conform to the laws of geological evolution, achieving the expected technical effect of controlling the prediction process with a clear geophysical mechanism.

[0213] S6.3: Based on the set of candidate mineralization modes, an attention weight allocation mechanism is used to reorder the importance of the activation states of each semantic anchor point in the global semantic consistency vector, so as to select the key anchor point combination that plays a decisive role in the current regional resource distribution, thereby reducing the model complexity and improving feature focus.

[0214] The system receives an initial mineralization potential feature tensor and a set of candidate mineralization patterns, and constructs an attention weight allocation network based on geological semantic association. The feature tensor is mapped to a query vector, and key geological elements in the candidate mineralization patterns are encoded as key vectors and value vectors.

[0215] The dot product similarity between the query vector and the key vector is calculated, and a scaling factor is introduced to suppress gradient vanishing, generating the original attention score matrix. The original score matrix is ​​then normalized using the Softmax function to obtain the probability distribution weights of each semantic anchor point under a specific mineralization pattern.

[0216] We use probability distribution weights to perform a weighted summation of the value vectors, fusing multi-source geological information to generate a context-aware feature representation for the current candidate pattern. We introduce a sparse regularization term to constrain the attention weight distribution, forcing the model to focus on a few high-contribution key semantic anchors and suppressing background noise interference.

[0217] A threshold truncation mechanism is used to filter semantic anchors with weights higher than a preset threshold, forming a key anchor combination that plays a decisive role in resource distribution. This combination retains the most geologically interpretable feature subset, significantly reducing the computational dimensionality in subsequent extrapolation stages.

[0218] By assigning attention weights and using sparse filtering, the high-dimensional feature tensor from the previous step is transformed into low-dimensional key anchor point combination data, achieving the expected technical effect of reducing model complexity and improving feature focus.

[0219] S6.4: Combining the key anchor point combination with the candidate mineralization mode set, the runtime space-time sequence inference engine iteratively calculates the mineralization probability field of the study area to generate a mineral resource distribution prediction map with spatial continuity and temporal evolution attributes, realizing quantitative inference from discrete feature points to continuous resource surfaces.

[0220] The system receives key anchor point combinations and a set of candidate mineralization modes, and constructs a spatiotemporal evolution state matrix. The spatial coordinates of the key anchor point combinations are mapped to discrete grid cells of the study area, and the prior mineralization probability of each grid cell is initialized to zero.

[0221] Based on the geological causal logic in the candidate metallogenic model set, a spatiotemporal transfer operator is defined. This operator includes a spatial diffusion coefficient and a temporal evolution weight, used to describe the migration patterns of ore-forming elements under specific geological backgrounds.

[0222] The runtime spacetime sequence inference engine performs iterative calculations. In the t-th iteration, it reads the mineralization probability field distribution data from the (t-1)-th iteration, applies the spacetime transition operator to perform convolution operations, and updates the probability distribution at the current moment.

[0223] The mineralization probability of grid cell i at time t is calculated using the following formula:

[0224]

[0225] in, Let be the mineralization probability of grid cell i at time t, α be the spatial diffusion attenuation factor, and N(i) be the neighborhood set of grid cell i. The spatiotemporal transfer weights from neighbor j to center i, Let be the probability value of neighbor j at the previous time step, and β be the contribution coefficient of the source term. The strength of direct evidence for the combination of key anchor points at the current location.

[0226] Boundary conditions are imposed on the probability field during the iteration process to ensure that the probability values ​​converge to the [0,1] interval. A damping mechanism is introduced to suppress high-frequency noise oscillations and ensure the physical stability of the evolution process.

[0227] The iteration loop terminates when the difference between the probability fields of two adjacent iterations is less than a preset threshold. The final converged mineralization probability field data is output, generating a spatially continuous mineral resource distribution prediction map.

[0228] Through iterative computation and processing by the spatiotemporal sequence inference engine, discrete key anchor point features are transformed into continuous resource distribution probability surfaces, realizing quantitative inference from point evidence to area prediction, and significantly improving the spatial resolution and geological rationality of the prediction results.

[0229] S6.5: For the mineral resource distribution prediction map, perform reverse tracing analysis to extract the key anchor point combination supporting each predicted high value area and its corresponding geological evolution logic chain, and finally output the resource distribution prediction results including specific spatial location, resource quantity estimation and causal explanation, thus completing the interpretable prediction closed loop.

[0230] Obtain the mineral resource distribution prediction map generated by S6.4, locate the local high-value areas in the map where the mineralization probability value exceeds the preset threshold, and extract the spatial coordinate index and probability intensity matrix corresponding to each high-value area.

[0231] Based on spatial coordinate indexing, the original geological semantic anchor point set relied upon in the process of generating the global semantic consistency vector is retrieved in reverse, and a reverse mapping path from high value area to initial anchor point is constructed.

[0232] The graph traversal algorithm is used to backtrack along the edge direction of the inferable anchor point relationship graph, and record all node sequences and edge weight parameters passed from the output layer to the input layer to form a candidate geological evolution logical chain topology.

[0233] Perform causal consistency checks on each path in the candidate logical chain, eliminate redundant branches that violate geological time series or physical stress transmission direction, and retain valid logical subgraphs that conform to geological mineralization laws.

[0234] The SHAP value calculation method is used to quantify the contribution of each key anchor point to the final prediction result. The importance score of anchor point i is calculated using the following formula:

[0235]

[0236] in, Let be the SHAP value of anchor point i, N be the total number of anchor points, S be the feature subset excluding anchor point i, M be the set of all features, and v(S) be the model's prediction on subset S. Denotes the set after removing feature i from the total feature set M. The cardinality of the feature subset.

[0237] Based on the calculated importance score, the anchors in the effective logic subgraph are sorted, and the Top-K key anchor combinations with a cumulative contribution reaching a preset ratio are selected.

[0238] By associating key anchor points with their corresponding geological definitions, spatial coordinates, and confidence labels, a structured data object containing causal explanations is generated.

[0239] By combining the probability density distribution in the prediction map, the geological volume estimation model is used to calculate the resource quantity integral in the high-value area and output the specific resource quantity estimation value.

[0240] By using a reverse tracing and analysis process, the prediction results of the black box model are transformed into an interpretable report that includes key anchor point combinations, evolutionary logic chains, and resource quantity estimates, thereby enabling the credibility verification of the prediction results and a closed-loop interpretation of their geological origins.

[0241] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0242] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0243] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the distribution of mineral resources based on big data analysis, characterized in that, Specifically, it includes: S1: Obtain multi-source heterogeneous raw geological data, extract implicit geological semantic units for each type of raw geological data and label them with confidence labels to generate an initial geological semantic anchor point library; S2: Based on the nodes in the initial geological semantic anchor point library and the known geological causal chain relationship, construct a reasonable anchor point relationship graph using a graph neural network; S3: Match the newly accessed local geological features extracted based on the original geological data to the neighboring anchors in the initial geological semantic anchor point library, and activate the corresponding geological relationship sub-graph according to the current regional geological evolution stage to calculate the dynamic confidence weight of the anchor points. S4: Use the dynamic confidence weight of the anchor points to distribute and map the response intensity of each data source on the inferable anchor point relationship graph to generate the anchor point response spectrum; S5: Based on the inferable anchor point relationship graph, the response spectra of multiple anchor points are aggregated through stress transmission path recursive propagation processing to generate a global semantic consistency vector. S6: Based on the global semantic consistency vector and its spatiotemporal derivative, perform mineral resource distribution deduction and output resource distribution prediction results.

2. The mineral resource distribution prediction method based on big data analysis according to claim 1, characterized in that, Step S3 specifically includes: The newly accessed local geological feature vector set is obtained, and the proximity retrieval process is performed on each feature unit in the local geological feature vector set based on multi-dimensional spatial indexing. A set of candidate geological semantic anchor points that meet the preset conditions is selected from the initial geological semantic anchor point library. Obtain the candidate geological semantic anchor point set and the current regional geological evolution stage identifier parameters, perform temporal consistency verification on the node attributes in the candidate geological semantic anchor point set, and generate an active geological relationship subgraph node list that conforms to the current regional geological evolution stage identifier parameters; Obtain the list of nodes in the active geological relationship subgraph and the topological connection data of the inferable anchor point relationship graph. Based on the graph traversal algorithm, perform subgraph reconstruction processing on the nodes in the list of nodes in the active geological relationship subgraph and their edge relationships in the inferable anchor point relationship graph to generate a dynamic geological relationship subgraph structure limited to the geological evolution stage identifier parameters of the current region. The dynamic geological relationship subgraph structure and the local geological feature response intensity values ​​of each data source are obtained. The influence diffusion calculation is performed on the nodes in the dynamic geological relationship subgraph structure through stress transfer path recursive propagation processing to generate the original confidence score sequence. Obtain the original confidence score sequence and confidence labels, perform attenuation correction processing on each score value in the original confidence score sequence, and output the anchor point dynamic confidence weight vector.

3. The mineral resource distribution prediction method based on big data analysis according to claim 2, characterized in that, The preset conditions include: overlapping spatial locations and the highest semantic similarity.

4. The mineral resource distribution prediction method based on big data analysis according to claim 2, characterized in that, The decay correction process for each score value in the original confidence score sequence is performed based on a Bayesian probability model.

5. The method for predicting mineral resource distribution based on big data analysis according to claim 1, characterized in that, Step S4 specifically includes: Based on the node topology in the inferable anchor point relationship graph, the local geological features extracted from various types of original geological data are subjected to neighbor anchor point matching processing to generate an initial anchor point matching set. Using the spatial correlation attributes in the initial anchor point matching set, and combining the dynamic confidence weight of the anchor points calculated at the current regional geological evolution stage, a weighted modulation operation is performed on the initial anchor point matching set to generate a weighted anchor point response intensity matrix. Based on the weighted anchor point response intensity matrix, the discrete response values ​​within the matrix are subjected to continuous distribution mapping to generate a local anchor point response density field. For the local anchor point response density field, a normalized integral transformation operation is performed to generate a standardized anchor point activation probability distribution vector; The standardized anchor point activation probability distribution vector is integrated and structured according to the data source type identifier to generate the anchor point response spectrum.

6. The method for predicting mineral resource distribution based on big data analysis according to claim 5, characterized in that, The continuous distribution mapping of discrete response values ​​within the matrix is ​​based on the Gaussian kernel function density estimation method.

7. The method for predicting mineral resource distribution based on big data analysis according to claim 1, characterized in that, Step S5 specifically includes: Obtain the node connection matrix and the anchor response spectrum in the inferable anchor relationship graph, extract the set of direct neighbor nodes and the corresponding edge weight parameters of each target semantic anchor based on the graph theory adjacency traversal mechanism, and generate an anchor neighborhood structure descriptor. Using the edge weight parameters in the anchor point neighborhood structure descriptor, combined with the preset structural stress attenuation function, the local geological feature response intensity values ​​within the set of direct neighbor nodes are subjected to distance-weighted correction processing to generate a weighted neighborhood response vector. Based on the weighted neighborhood response vector and the self-response value of the current target semantic anchor point, perform multiple rounds of recursive iterative propagation calculations to generate multi-level stress transfer accumulation; wherein, the self-response value refers to the local geological feature response intensity value of the target semantic anchor point itself; For the accumulated stress transmission at multiple levels, a geological rule constraint mask is applied for logical verification to remove abnormal transmission path components that violate the known geological causal chain direction. Then, a normalization mapping operation is performed on the remaining effective components to generate a corrected global semantic fusion vector. The modified global semantic fusion vector is encapsulated by adding a spatiotemporal derivative label. The value of each dimension in the modified global semantic fusion vector is bound and recombined with its corresponding geological semantic anchor point definition and the confidence label to output the global semantic consistency vector.

8. The method for predicting mineral resource distribution based on big data analysis according to claim 7, characterized in that, The execution of multi-round recursive iterative propagation calculation includes: using the intermediate aggregation state generated in the previous round of iteration as the input boundary condition for the next round of iteration, and generating the convergent and stable multi-level stress transfer accumulation by accumulating the contribution values ​​of neighboring nodes and introducing a damping factor to suppress noise accumulation.

9. The method for predicting mineral resource distribution based on big data analysis according to claim 7, characterized in that, Obtain the node connection matrix and the anchor response spectrum in the inferable anchor relationship graph. Based on the graph theory adjacency traversal mechanism, extract the set of direct neighbor nodes and corresponding edge weight parameters for each target semantic anchor. Generate an anchor neighborhood structure descriptor, including: Parse the row and column indices of the node connection matrix and establish a mapping index table between semantic anchors and their direct neighbor nodes; Traverse the mapping index table, and for each target semantic anchor point, extract the set of IDs of all its directly connected neighbor nodes to construct a local neighborhood node list; Read the edge weight parameters between the target node and each of its neighboring nodes from the node connection matrix; The extracted set of neighbor node IDs is structured and bound to the corresponding edge weight parameters to form a local topological data unit. Integrate the local topological data units of all target semantic anchor points to generate the anchor point neighborhood structure descriptor.

10. The method for predicting mineral resource distribution based on big data analysis according to claim 1, characterized in that, The resource distribution prediction results include key anchor point combinations and evolutionary logic chains.