Mine geological exploration data analysis method and system

Abstract

The invention discloses a mine geological exploration data analysis method and system, and the method comprises the steps: constructing a multimodal geological knowledge map based on a metallogenic law, fusing space coordinates and geological features, and achieving the dynamic increment updating; performing spatial gridding main reserve prediction by adopting a three-dimensional convolutional neural network, and quantifying abnormal features of a geological space in combination with an anomaly perception proxy model based on knowledge distillation and comparative learning; an abnormal region is automatically judged and positioned through multi-modal feature fusion and a dynamic threshold mechanism, then semantic alignment and parameter compensation based on a knowledge graph are executed, and local correction is carried out on a grade prediction result; according to the method, the accuracy of reserve prediction and the abnormal region detection capability are improved, and efficient collaboration and adaptive evolution of geological knowledge and a deep learning model are realized.

Description

Technical Field

[0001] This invention relates to the field of geological resource exploration and reserve prediction technology, and in particular to a method and system for analyzing geological exploration data in mines. Background Technology

[0002] In the current field of mine geological exploration data analysis and resource reserve estimation, mainstream technical solutions are mostly based on data-driven machine learning models, supplemented by multi-source geological data (such as geophysical, borehole core, and geochemical data) for spatial feature modeling to predict the spatial distribution of ore bodies and resource grades. In recent years, with the development of artificial intelligence and deep learning technologies, 3D convolutional neural networks (3D-CNN), U-Net variants, and graph convolutional networks have been gradually applied to reserve estimation tasks. Through end-to-end automatic feature extraction and spatial relationship modeling, they have significantly improved prediction efficiency and automation levels. At the same time, the industry has also begun to focus on the role of domain knowledge in the modeling process, attempting to incorporate expert experience such as geological mineralization patterns and prior models into machine learning models. Some studies have introduced geological knowledge graphs as auxiliary information sources to achieve feature enhancement or prior constraints. However, existing data-driven reserve estimation methods generally have significant limitations. First, in areas with complex geological structures or sparse effective exploration data, the main model relies heavily on existing observational data, making it difficult to accurately characterize high-order geological processes and spatial anomalies, leading to localized biases in reserve predictions. Second, although there have been preliminary attempts in the industry to encode knowledge graphs as model feature inputs to achieve "knowledge pre-injection," this static injection method has failed to dynamically link with the model inference process. Knowledge information only participates in feature enhancement in the early stages of the model; it cannot make targeted corrections based on actual local anomalies during the inference stage. This disconnect between knowledge and the data-driven model makes it difficult to leverage the timely guiding value of domain knowledge. Especially when facing geological anomalies or complex tectonic areas such as fault intersections, fold structures, and heterogeneous alteration, mainstream models often exhibit problems such as insufficient generalization ability, insensitivity to anomaly detection, and unreliable predictions. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method and system for analyzing geological exploration data in mines.

[0004] The technical solution of this invention is implemented as follows: a method for analyzing geological exploration data in mines, comprising: S1: Construct a domain knowledge graph based on geological mineralization laws. The knowledge graph uses a multimodal graph structure to store geological constraint rules including structural features, lithological combinations, mineralization and alteration, etc., and establishes a mapping relationship between spatial coordinates and geological features. S2: Based on a three-dimensional convolutional neural network, a main reserve estimation model is constructed. After dividing the exploration area into regular spatial grids, geophysical data, borehole core data and geochemical data are input, and the initial resource grade prediction value and confidence distribution of each grid cell are output. S3: Construct an anomaly perception proxy model based on knowledge distillation. The proxy model extracts high-order semantic rules from the geological knowledge graph through a contrastive learning strategy, generates an anomaly scoring function that includes fault intersection features, lithofacies combination patterns and alteration superposition effects, and establishes a quantitative mapping between geological anomaly degree and spatial coordinates. S4: Simultaneously run the main reserve estimation model and the anomaly perception proxy model to generate two sets of output data for the same spatial grid cell: the grade prediction tensor output by the main model and the geological anomaly tensor output by the proxy model, forming a decision space for multimodal feature fusion. S5: The geological anomaly tensor is binarized and segmented based on a dynamic threshold determination mechanism. When the anomaly of a local area exceeds the adaptive threshold, a knowledge intervention process is triggered. The adaptive threshold is dynamically adjusted according to the density of regional exploration data and the prediction confidence of the master model. S6: Perform geological semantic alignment operation, perform graph matching between the spatial feature vector of the triggered intervention area and similar structural scenes in the knowledge graph, and generate a semantic alignment parameter set including fault strike correction coefficient, lithological weight adjustment factor and alteration superposition compensation value; S7: Based on the semantic alignment parameter set, the prediction results of the main model are locally corrected. A weighted fusion strategy is used to nonlinearly combine the knowledge-guided parameters with the original prediction values ​​to generate a corrected resource grade distribution model. The correction weight coefficient is positively correlated with the geological anomaly. S8: Establish a closed-loop feedback mechanism, input the corrected prediction results verified by drilling and the corresponding geological feature labels into the knowledge graph update module, and optimize the geological constraint rules and anomaly scoring function parameters in the knowledge graph through incremental learning strategy.

[0005] The present invention also provides a mining geological exploration data analysis system, which uses the above-mentioned mining geological exploration data analysis method to perform mining geological exploration data analysis.

[0006] The present invention provides a method and system for analyzing geological exploration data in mines, which has the following beneficial effects: (1) This invention achieves knowledge-guided dynamic reasoning optimization by constructing a collaborative architecture between the main model and a lightweight anomaly perception proxy model. The dynamic knowledge distillation mechanism proposed in this invention enables the proxy model to extract high-order mineralization semantic rules from the geological knowledge map and transform them into a computable anomaly scoring function, thereby identifying spatial units that deviate from typical mineralization patterns in real time during the reasoning stage. When a high-confidence anomaly is detected, the system automatically triggers a knowledge intervention process, adjusting the prediction weight of the main model in that region based on semantic alignment and contrastive learning strategies, thereby achieving knowledge-driven correction of local prediction results. This design breaks through the rigid mode of knowledge utilization in traditional methods and significantly enhances the adaptability and estimation reliability of the model in areas with sparse data or structural anomalies. (2) This invention further introduces a closed-loop optimization mechanism of "data-knowledge-model", which significantly improves the system's self-evolution capability and long-term stability. Unlike existing technologies that only use knowledge graphs to assist modeling in a one-way manner, this invention, after completing on-site exploration and verification, will back-inject confirmed and valid amendment examples into the knowledge graph to achieve incremental updates and continuous enrichment of the knowledge base. This mechanism not only ensures that the knowledge system evolves with new discoveries, but also enables subsequent estimation tasks to make judgments based on a more comprehensive empirical foundation, forming a virtuous feedback loop. At the same time, since anomaly perception is performed by a dedicated lightweight proxy model, the main model maintains its original efficient structure without increasing significant computational burden, ensuring the real-time response capability of the overall system in large-scale three-dimensional space modeling. Attached Figure Description

[0007] Figure 1 This is a flowchart of a mining geological exploration data analysis method according to the present invention; Figure 2 This is a sub-flowchart of a mining geological exploration data analysis method according to the present invention; Figure 3 This is another sub-flowchart of a mining geological exploration data analysis method according to the present invention. Detailed Implementation

[0008] 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.

[0009] 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.

[0010] like Figure 1 As shown, this invention provides a method for analyzing geological exploration data in mines, specifically including: S1: Construct a domain knowledge graph based on geological mineralization laws. The knowledge graph uses a multimodal graph structure to store geological constraint rules including structural features, lithological combinations, mineralization and alteration, etc., and establishes a mapping relationship between spatial coordinates and geological features. S2: Based on a three-dimensional convolutional neural network, a main reserve estimation model is constructed. After dividing the exploration area into regular spatial grids, geophysical data, borehole core data and geochemical data are input, and the initial resource grade prediction value and confidence distribution of each grid cell are output. S3: Construct an anomaly perception proxy model based on knowledge distillation. The proxy model extracts high-order semantic rules from the geological knowledge graph through a contrastive learning strategy, generates an anomaly scoring function that includes fault intersection features, lithofacies combination patterns and alteration superposition effects, and establishes a quantitative mapping between geological anomaly degree and spatial coordinates. S4: Simultaneously run the main reserve estimation model and the anomaly perception proxy model to generate two sets of output data for the same spatial grid cell: the grade prediction tensor output by the main model and the geological anomaly tensor output by the proxy model, forming a decision space for multimodal feature fusion. S5: The geological anomaly tensor is binarized and segmented based on a dynamic threshold determination mechanism. When the anomaly of a local area exceeds the adaptive threshold, a knowledge intervention process is triggered. The adaptive threshold is dynamically adjusted according to the density of regional exploration data and the prediction confidence of the master model. S6: Perform geological semantic alignment operation, perform graph matching between the spatial feature vector of the triggered intervention area and similar structural scenes in the knowledge graph, and generate a semantic alignment parameter set including fault strike correction coefficient, lithological weight adjustment factor and alteration superposition compensation value; S7: Based on the semantic alignment parameter set, the prediction results of the main model are locally corrected. A weighted fusion strategy is used to nonlinearly combine the knowledge-guided parameters with the original prediction values ​​to generate a corrected resource grade distribution model. The correction weight coefficient is positively correlated with the geological anomaly. S8: Establish a closed-loop feedback mechanism, input the corrected prediction results verified by drilling and the corresponding geological feature labels into the knowledge graph update module, and optimize the geological constraint rules and anomaly scoring function parameters in the knowledge graph through incremental learning strategy.

[0011] Step S1: Construct a domain knowledge graph based on geological mineralization laws. This knowledge graph uses a multimodal graph structure to store geological constraint rules including dimensions such as tectonic features, lithological assemblage, and mineralization alteration, and establishes a mapping relationship between spatial coordinates and geological features. Specifically, this includes: S1.1: Acquire multi-source heterogeneous geological data, including regional tectonic maps, borehole lithology logs, geochemical anomaly maps, orebody boundary data, and alteration zone distribution information, to construct the original knowledge base of the geological knowledge graph and provide basic semantic entities and relationships for subsequent graph structure modeling; The input objects include regional tectonic maps, borehole lithology logs, geochemical anomaly maps, orebody boundary data, and alteration zone distribution information. The original knowledge base of the geological knowledge graph needs to be constructed from multi-source heterogeneous datasets. A multi-source data interface integration method (parameters: structural map vector file, lithological log structured table, anomaly map raster data, orebody boundary Shapefile, alteration zone GeoTIFF) is adopted to achieve unified reading and formatting processing of different data sources, and to establish a unified coordinate reference system to eliminate spatial reference differences. Furthermore, by using a spatial resampling algorithm (parameters: target grid resolution = 50m, interpolation strategy = bilinear interpolation), the pixel size of the raster data is aligned with the spatial grid cell, and a geochemical anomaly grid and alteration zone distribution grid with consistent spatial resolution are obtained. Furthermore, through a structured parsing algorithm (parameters: lithology coding dictionary, structural element regularity matching rules), the text hierarchical parsing of the borehole lithology log is realized, transforming lithology, structural features and depth information into standardized field patterns, and generating a borehole lithology table containing coordinates, depth and lithology type; Furthermore, through a spatial vector fusion algorithm (parameters: topology tolerance = 0.001, coordinate system EPSG: 4490), the spatial overlay of ore body boundary vector data and structural map vector data is achieved, the intersection information of the boundary and structural line is extracted, and the corresponding spatial relationship table is generated; Furthermore, attribute standardization processing (parameters: normalization range [0,1] for numerical data, One-Hot encoding for categorical data) is adopted to unify the numerical scale of the original data and provide an input feature matrix for subsequent entity recognition and semantic annotation; Through the above algorithm processing method, the original spatial heterogeneous data is transformed into an original knowledge base with unified spatial reference, unified format structure and unified attribute scale, so that multi-source geological data can be directly called in the knowledge graph modeling stage; For example, the exploration area is 120 The regional tectonic map, spanning km², was imported in DXF format, containing 234 main fault lines and 512 secondary faults. The borehole lithology log is a CSV file, recording 184 boreholes at an average depth of 600m. Lithology fields are mapped to 12 categories based on a lithology coding dictionary, with a depth resolution of 1m. The geochemical anomaly map has a resolution of 25m×25m, with elemental concentrations in ppm, and was resampled to 50m×50m using bilinear interpolation. The orebody boundary data is in Shapefile format, containing closed polygons of 24 main orebodies. Alteration zone distribution information is in GeoTIFF format, with spectral features normalized and mapped to [0,1]. The coordinate system is EPSG:4490. A spatial vector fusion algorithm yielded 398 intersection points between tectonic lines and orebody boundaries. After attribute standardization, the original knowledge base size is approximately 28,000 spatial nodes and 38-dimensional attribute fields, which can be directly used for subsequent graph structure entity recognition and relationship extraction. S1.2: Based on the geological mineralization regularity, entity recognition and semantic annotation are performed on the original geological data to extract core geological elements, including fault strike, fold axis, lithological unit, alteration type, and mineralization intensity, as node representations in the knowledge graph, so as to realize the semantic structuring of geological features; Using the multi-source heterogeneous geological data obtained in step S1.1 as input, an automatic geological entity recognition algorithm oriented towards the domain (parameters: regional tectonic graphic element matrix, lithological log text sequence, geochemical anomaly map thermal value, ore body boundary vector set, alteration zone distribution raster set) is adopted to realize the automatic detection and classification of potential geological elements in the original data. Furthermore, by using a hybrid method of named entity recognition and rule matching (rule parameters: fault strike pattern template set, fold axis geometric template library, lithological combination feature code table, alteration type coding table, mineralization intensity grading standard), geological semantic annotation of detected entities is achieved, and a structured annotation metadata dataset is obtained. Furthermore, the fault strike and fold axis are parametrically represented by a spatial geometric feature extraction algorithm (parameters: three-dimensional coordinate system reference, strike calculation formula, axis fitting function), and the output includes quantitative indicators such as strike angle, included angle, and curvature. Furthermore, the mineralization intensity is numerically calculated using a geochemical anomaly feature interpretation algorithm (parameters: anomaly concentration threshold, enrichment coefficient calculation formula) to form a mineralization intensity quantification parameter vector; By using geological feature encoding processing, the above identification and annotation results are transformed into node representations of a knowledge graph, forming a set of geological entity nodes with multi-dimensional attributes, thereby realizing the semantic structuring of geological features. For example, in data processing for a mining area, the input regional structural map has a resolution of 5 m / pixel, the lithological log contains a sequence of 500 core records, and the gold concentration in the geochemical anomaly map is recorded in ppm. An automatic geological entity recognition algorithm is used to perform fault line detection on the structural map. The parameters are set with a minimum strike detection length of 50 m and a curvature tolerance of 5°. The mean strike angle of the extracted faults is... Fault elements of degree were identified. The lithological log was annotated using a rule-matching method, identifying three lithological units: sandstone, shale, and diorite, and denoted as nodes with codes 00A, 00B, and 00C. The fold axes of the tectonic map were fitted using a spatial geometric feature extraction algorithm, and the axis curvature was calculated. / m, mean angle Degree. Mineralization intensity is given by the formula.

[0012] Calculation, where The average concentration in the test area. Given the background concentration, the enrichment coefficient is obtained as follows: The above parameter set is encoded as node attributes and input into the knowledge graph construction module to form standardized geological nodes containing multiple attributes such as fault strike angle, lithology type, fold curvature, and mineralization enrichment coefficient. This realizes the semantic structuring of geological features, which can be used for relation extraction and knowledge embedding processing to support the generation of high-order semantic rules in the dynamic knowledge distillation mechanism. S1.3: Using graph neural networks and knowledge extraction technology, relationships are extracted from labeled geological entities to identify the spatial association between faults and ore bodies, the causal relationship between lithological combinations and mineralization, and the synergistic effect of alteration superposition and mineralization enrichment, so as to construct edge relationships in a multimodal graph structure and realize the topological expression of geological knowledge. The input condition is the semantically labeled geological entity set output by step S1.2. This entity set contains core elements such as fault strike, fold axis, lithological unit, alteration type and mineralization intensity. A graph neural network relationship inference method based on named entity pair extraction (parameters: node feature dimension 128, number of layers 3, number of attention heads 8) is adopted to model the potential spatial relationships between labeled geological entities; Furthermore, by using a spatial proximity calculation algorithm (parameter: threshold distance set to 50m), the spatial correlation between fault entities and ore bodies is evaluated, and a preliminary set of fault-ore body relationship edges is obtained. Furthermore, a causal relationship inference algorithm based on the metallogenic evolution law (parameter: time series window length of 5 geological periods) is adopted to identify the causal relationship between lithological assemblage and metallogenesis, and to generate a causal parameter matrix of lithological assemblage-metallogenesis. Furthermore, using a multimodal synergistic effect analysis method (parameters: geochemical anomaly intensity weight 0.6, alteration type weight 0.4), the synergistic effect between alteration superposition mode and mineralization enrichment degree is quantified, and an alteration superposition-mineralization enrichment relationship weight vector is generated. By using graph structure construction algorithms, the spatial correlation edges of faults and ore bodies, the causal edges of lithological combinations and mineralization, and the edges of alteration superposition and mineralization enrichment synergistic effects are integrated into a multimodal graph structure to realize the topological expression of geological knowledge. For example, for a certain exploration area, the input geological entity set includes 15 faults, 28 ore bodies, 12 lithological units, and 7 alteration types. A relation inference is performed using a node feature dimension of 128, a 3-layer graph neural network, and an 8-attention head configuration. The Euclidean distance between faults and ore bodies is calculated. When the distance is less than... Spatial correlation edges were established at the meter level, resulting in 42 fault-orebody relationship edges. Based on time series windows of 5 geological periods, 18 edges showing a significant causal relationship between lithological assemblages and mineralization were inferred, forming a lithological assemblages-mineralization causal parameter matrix. Using weighted proportions... The synergistic effect between alteration type and geochemical anomaly intensity was quantified, resulting in 9 alteration superposition-mineralization enrichment relationship edges. The final multimodal graph structure contains 62 nodes and 69 edges, covering three types of relationships: spatial correlation, causal relationship, and synergistic effect. This achieves a complete topological expression of geological knowledge in the graph structure and significantly improves the computability and correlation of geological anomaly features in subsequent knowledge distillation processes. S1.4: Based on the graph embedding algorithm, the multimodal graph structure is represented by node vectorization, and the geological entities and their relationships are mapped to a low-dimensional continuous vector space to generate an embedding model of the geological knowledge graph, providing a computable semantic rule expression for subsequent knowledge distillation; The input condition is the multimodal geological knowledge graph constructed in step S1.3. Its nodes include entities such as fault strike, fold axis, lithological unit, alteration type and mineralization intensity. The edge relationships describe spatial association, causal relationship and synergistic effect. Both nodes and edges have high-dimensional feature attribute vectors, which need to be vectorized for subsequent knowledge distillation. The DeepWalk algorithm (parameters: random walk length of 64, number of walks of 10, window size of 5) is used to sample the local structural feature sequences of multimodal graph nodes and generate the corresponding set of node context sequences. Furthermore, by using the Skip-gram word vector training method (parameters: embedding dimension set to 128, negative sampling number of 5), the vector space mapping training of the sampled node context sequence is realized, and the basic node embedding matrix is ​​obtained. Furthermore, the GraphSAGE algorithm (parameters: number of sampled neighbors is [10,10], aggregation function type is mean aggregator) is used to realize graph convolutional neighborhood aggregation that combines the original attribute vector of the node with the structural context, and to generate an enhanced embedding representation of the node in the global topology. Furthermore, by using an edge type weighting mechanism (parameters: the weight of structural feature edges is set to 1.2, the weight of lithological combination edges is set to 1.0, and the weight of alteration edges is set to 1.3), the influence of different types of edges on graph embedding training is adjusted differentially, and a weighted node embedding matrix is ​​generated. The t-SNE dimensionality reduction algorithm (parameters: target dimension set to 64, perplexity set to 30) is used to map the above high-dimensional embedding matrix to a low-dimensional continuous vector space to improve the computational efficiency and visual interpretability of the embedding representation in the subsequent knowledge distillation process. By embedding matrix normalization (method: L2 norm normalization), the result of the previous step is transformed into a normalized low-dimensional node vector set, realizing the computable semantic rule expression of geological entities and their relationships; For example, a knowledge graph for a certain mining area contains 2000 nodes, with edge types including three categories: fault-orebody, lithology-mineralization, and alteration-enrichment. When executing the DeepWalk algorithm, the random walk length is configured to 64 steps, the number of walks per node is 10, the window size is 5, and a total of 128,000 node context sequences are generated. When training using the Skip-gram model, the embedding dimension is set to 128, and the negative sampling number is 5, resulting in an initial node vector matrix with a shape of (2000, 128). When running GraphSAGE aggregation, each node samples two layers of neighbors: 10 neighbors in the first layer and 10 neighbors in the second layer, with the aggregation function being the mean. Edge weights are configured according to structural features 1.2, lithological combinations 1.0, and alteration 1.3, used for weighted aggregation results. t-SNE was used to reduce the dimensionality to 64, with a perplexity of 30. The resulting embedding matrix had a shape of (2000, 64) and was normalized using the L2 norm to ensure that the magnitude of each row vector was 1. The final output embedding model was then used by the proxy model in subsequent knowledge distillation, achieving efficient semantic rule retrieval and matching based on node vectors in complex construction scenarios. This significantly improved the accuracy and stability of anomaly detection and bias correction. S1.5: Establish a mapping function between spatial coordinates and geological features, and associate the three-dimensional exploration grid with nodes in the knowledge graph through a spatial indexing mechanism, so that each spatial unit can retrieve its corresponding geological constraint rules to support the spatial positioning and semantic matching of the dynamic knowledge intervention mechanism. S1.6: An incremental update mechanism is executed on the knowledge graph. Based on the geological feature labels in the field verification data and the corrected reserve prediction results, a graph attention network is used to optimize the node representation and edge weights in the knowledge graph online to improve the adaptability and generalization ability of the knowledge graph.

[0013] Step S2: Constructing a main reserve estimation model based on a three-dimensional convolutional neural network, dividing the exploration area into regular spatial grids, inputting geophysical data, borehole core data, and geochemical data, and outputting the initial resource grade prediction value and confidence distribution of each grid cell. Specifically, this includes: S2.1: Spatial discretization processing is performed on the mining exploration area. Based on the geological coordinate system, the three-dimensional geological space is divided into equally spaced cubic grid units to establish a unified spatial reference frame for subsequent fusion input of multi-source heterogeneous data. For the three-dimensional spatial data of the mining exploration area, a regular discretization method based on the geological coordinate system is adopted (parameters: spatial resolution Δx, Δy, Δz correspond to the geological structure survey accuracy respectively) to achieve the standard grid division of geological space and the unification of spatial reference system; Furthermore, by using a spatial coordinate normalization algorithm (parameters: the origin of the coordinates is set to the southwest corner of the exploration area, and the normalization range is [0,1]), the original measuring point coordinates are mapped to the standardized coordinate system, and a normalized coordinate matrix is ​​obtained, ensuring that the spatial location encoding of different data sources is consistent; Furthermore, a three-dimensional subdivision algorithm is adopted (parameters: the number of subdivisions in the x, y, and z directions is automatically matched according to the sampling density of the exploration data) to discretize the regional space into equally spaced cubic units and generate a three-dimensional mesh index table for subsequent data cube construction; Furthermore, through spatial voxelization (parameter: matching the cell side length with the minimum feature size of geological survey), the voxel numbering and adjacency matrix construction of the discretized cells are realized, and a three-dimensional data skeleton that can support convolution operations is generated. Furthermore, a mapping function between spatial coordinates and voxel indexes is adopted to establish a geological feature retrieval index, ensuring that any coordinate point can be quickly located to the corresponding grid cell, thus realizing a spatial fusion benchmark when inputting multi-source heterogeneous geological data; By using meshing and voxelization, the original geological spatial coordinates from the previous step are transformed into cubic mesh index data with a uniform resolution, achieving the expected technical effect of spatial alignment and fusion input of multi-source data. For example, for an exploration area in a mining region with dimensions of 3000 meters (length), 2500 meters (width), and 1000 meters (height), and measurement accuracies of 10 meters, 10 meters, and 5 meters respectively, a regular discretization method is used to set Δx = 10 meters, Δy = 10 meters, and Δz = 5 meters, forming 300 × 250 × 200 cubic grid cells. The origin of the spatial coordinates is set as the southwest corner of the area, and the measured borehole coordinates (1500, 1250, 250) are normalized to (0.5, 0.5, 0.25), with the normalization matrix used as the spatial positioning input. A 3D partitioning algorithm is applied to divide the area into 300 segments in the x-direction, 250 segments in the y-direction, and 200 segments in the z-direction, generating a total of 150 million cubic indices. Spatial voxelization is used to assign a number to each cell, and a 6-adjacency relation matrix is ​​generated for each cell, supporting subsequent 3D convolution operations. Using the established coordinate and index mapping function, the geochemical data point (2000, 1500, 500) is quickly located to the index (200, 150, 100), realizing the normalization and fusion of heterogeneous data in this spatial unit. Then, in step S2.2, a data cube containing the physical, core and geochemical characteristics of this unit is generated for high-precision grade prediction of the main reserve estimation model. S2.2: Acquire geophysical data, borehole core data and geochemical data, and perform standardized preprocessing on the data, including missing value imputation, outlier removal and feature normalization, to generate a structured data cube that can be used as input to a three-dimensional convolutional neural network; S2.3: Construct a main reserve estimation model architecture based on a three-dimensional convolutional neural network, adopting a U-Net variant structure, including an encoder-decoder framework and a skip connection mechanism, to achieve multi-scale extraction and reconstruction of three-dimensional geological spatial features; For the standardized preprocessed 3D geological data cube, a 3D convolutional neural network architecture design method (parameters: kernel size 3×3×3, stride 1, fill mode same) is adopted to realize the continuous extraction of local spatial geological features and layer-by-layer feature mapping. Furthermore, the encoder part is constructed through a variant U-Net structure (parameter: four-layer convolutional downsampling module, each layer containing convolution, batch normalization and ReLU activation), to achieve layer-by-layer spatial resolution compression and high-dimensional semantic feature abstraction of the 3D input data, and to obtain feature tensors at different levels; Furthermore, by designing the decoder part symmetrically (parameters: four-layer deconvolution upsampling module, kernel size of 2×2×2, stride of 2), spatial resolution reconstruction of the high-dimensional features output by the encoder is achieved. In addition, a skip connection mechanism corresponding to the encoder layer is introduced in each decoder layer to fuse low-level local detail features with high-level semantic features, generating multi-scale feature reconstruction results. Furthermore, a feature fusion convolutional layer (parameters: kernel size 1×1×1, number of output channels is 1) is used to achieve channel compression and output mapping of the fused features at the end of the decoder, and to generate an initial resource quality prediction tensor, which retains the correspondence between spatial coordinates and predicted quality. Furthermore, a parallel confidence estimation branch is introduced (parameters: sharing decoder feature input with grade prediction output, convolution kernel size 1×1×1, output channel number 1, activation function is Sigmoid) to realize the quantitative output of prediction reliability for each spatial grid cell and form a confidence distribution with the same scale as the predicted grade tensor; By using the above-mentioned U-Net variant architecture and skip connection fusion mechanism, the data cube constructed in the previous step is transformed into a dual-channel output containing the initial resource grade prediction value and confidence distribution, thereby achieving the expected technical effect of multi-scale feature extraction and fine spatial feature reconstruction. For example, within the exploration area of ​​a certain mining area, the three-dimensional coordinate range is 0~100m in the X direction, 0~100m in the Y direction, and 0~50m in the Z direction. The spatial discretization resolution is 5m×5m×5m. Each grid cell collects 3 channels of input data (geophysical parameters, core chemical composition, and geochemical indicators), and the input data cube size is 20×20×10×3. The encoder part uses a 3×3×3 convolution kernel, a stride of 1, and 32 channels in the first layer. After batch normalization and ReLU activation, the output size is 20×20×10×32, which is then downsampled to 10×10×5×32 by max pooling. The second layer increases the number of channels to 64 and downsamples to 5×5×2×64. The third layer has 128 channels and a size of 3×3×1×128. The fourth layer is the bottleneck layer with 256 channels and a size of 3×3×1×256. The decoder layer uses a 2×2×2 deconvolution kernel with a stride of 2 for upsampling at each layer, and fuses with the corresponding encoder layer through skip connections. For example, when decoding the third layer, the output size is 5×5×2×128 after fusing the 128-channel features of the third layer encoder. The grade prediction branch ends with a 1×1×1 convolution kernel to compress the fused features into a single-channel prediction output, resulting in a grade prediction tensor of size 20×20×10. The confidence branch uses a 1×1×1 convolution with the same structure and is activated by a sigmoid function, outputting a confidence tensor ranging from 0 to 1. During the testing phase, the model generated predicted grade values ​​and corresponding confidence scores for each grid cell in the above-mentioned mining area, realizing joint modeling of multi-scale geological features and identification of high-confidence prediction areas, significantly improving the stability of reserve estimation in complex geological structures. S2.4: Supervised training of the main reserve estimation model is performed based on the labeled training sample set, wherein the training sample includes the structured data cube and the corresponding true grade value label, and the model parameters are optimized using the mean square error loss function to improve the grade prediction accuracy. After completing the construction of the main reserve estimation model architecture, the input conditions are the data cubes that have undergone spatial discretization and standardization preprocessing, and the corresponding real grade value label data; Supervised training method is used (parameters: training sample input data cube) Label vector Batch size Initial learning rate This enables the iterative optimization of model parameters to establish a nonlinear mapping relationship between input features and predicted grades. Furthermore, the mean squared error loss between the predicted output and the true label is calculated using the backpropagation algorithm (parameters: activation function ReLU, optimizer Adam), and the loss gradient is passed to the weights of each layer of the network to update the parameters. Furthermore, the mean squared error loss function is adopted:

[0014] in, For the first The predicted value for each sample, To correspond to the true taste label, The total number of samples; Furthermore, a dynamic learning rate adjustment strategy is employed (parameter: threshold for the rate of change of validation set loss). Adjustment factor This allows for adaptive reduction of the learning rate during training, preventing it from getting stuck in local optima and improving convergence stability. Furthermore, an early stopping mechanism is utilized (parameter: number of rounds in which validation set performance is not improved). This achieves overfitting suppression. When the loss on the validation set fails to improve after several consecutive iterations, training is stopped, and the optimal model weights are retained. The supervised training algorithm described above transforms the data from the previous step into a main reserve estimation model with converged parameters, enabling high-precision spatial grade prediction and outputting the weight file and validation set performance metrics of the trained model. For example, supervised training is performed on the input structured geophysical, borehole core, and geochemical data cubes, with the batch size configured as follows: The initial learning rate is set to The Adam optimizer is used in conjunction with the weight decay factor. To improve generalization performance, the mean squared error loss function is applied at the end of each iteration, and the calculation formula is as follows:

[0015] Among them, the total number of samples for The ratio of training set to validation set is During training to the [number]th During each round, when the validation set loss decreases to the convergence threshold, the early stopping mechanism is triggered, and the optimal weight file is output. After implementing this configuration, the mean absolute error of the model's validation set is significantly reduced compared to the initial state, and the stability and accuracy of spatial grade prediction are greatly improved. S2.5: During the inference phase, the structured data cube is input into the trained main reserve estimation model, and resource grade prediction is performed cell by cell to generate a three-dimensional grade distribution tensor. The prediction confidence index of each cell is calculated as the basis for subsequent anomaly detection and knowledge intervention.

[0016] like Figure 2 As shown, step S3 involves constructing an anomaly perception proxy model based on knowledge distillation. This proxy model extracts high-order semantic rules from the geological knowledge graph using a contrastive learning strategy, generating an anomaly scoring function that includes fault intersection features, lithofacies assemblage patterns, and alteration superposition effects. It also establishes a quantitative mapping between geological anomaly degree and spatial coordinates. Specifically, this includes: S3.1: Based on the multimodal geological structural rules stored in the geological knowledge graph, extract high-order semantic representations of fault intersection features, lithological combination patterns, and alteration superposition effects to construct a knowledge-guided template for geological anomaly features; The input condition is the geological knowledge map constructed by step S1. This map stores rules related to mineralization laws, including fault structure parameters, lithological combination characteristics, alteration type and intensity, in a multimodal graph structure, and establishes a mapping with the three-dimensional exploration grid through a spatial indexing mechanism. A multimodal feature parsing algorithm (parameters: construct feature node set, lithology node set, alteration node set) is used to realize the semantic parsing of three core geological structures of fault, lithology and alteration in the knowledge graph, extract the geometric attributes, associated edge weights and mineralization mode labels of the nodes, and form the original high-order semantic candidate set. Furthermore, by constructing a pattern recognition algorithm (parameters: fault strike vector, dip angle parameter, intersection node density), the characteristics of fault intersection are identified and extracted, and node attribute data of regional fault intersection radius, intersection angle and mineral enrichment probability are obtained. Furthermore, the lithological combination pattern matching algorithm (parameters: lithological sequence vector, combination relationship matrix) is used to extract lithological combination features and generate lithological combination pattern codes, including lithological types, spatial continuity index and matching degree of typical ore-forming rocks; Furthermore, by using an alteration superposition effect analysis algorithm (parameters: alteration type label, superposition number, spatial coverage), the alteration type and superposition mode are identified, and an alteration superposition intensity coefficient and a mineralization enhancement index are generated to quantify the promoting effect of alteration on mineralization. By using a feature-based multi-layer fusion processing method, fault intersection features, lithological combination patterns, and alteration superposition effects are integrated into a unified high-order semantic representation template, enabling the output of knowledge-guided templates for geological anomaly features. For example, in a geological knowledge map of a certain mining area, fault structure nodes include strike. = degrees, inclination angle = Degree, intersection node density Units / km²; the lithological assemblage consists of a continuous distribution of granodiorite and quartz diorite, with a spatial continuity index of [missing information]. Ore-forming rocks matching degree The alteration types are sericitization and chloritization, with a superposition number of times... coverage When executing the fault intersection identification algorithm, it is based on the intersection radius. m and angle Nodes in high enrichment probability regions are extracted, and those with a matching degree higher than [a certain value] are selected in lithological combination model matching. Combination mode encoding is used to calculate the superposition intensity coefficient in alteration superposition analysis. The final knowledge-guided template generated by the fusion includes fault intersection parameters, lithological pattern coding, and alteration enhancement coefficients. It has been verified that the identification results in anomaly detection in this mining area are significantly improved, and the training effect of the anomaly scoring function of the subsequent proxy model is significantly enhanced. S3.2: Perform graph embedding processing on the knowledge-guided template of the geological anomaly features, and use a graph neural network (GNN) to map it into a geological knowledge embedding representation in a low-dimensional semantic vector space to form a knowledge representation vector set that can be used for model training; The input is the geological anomaly feature knowledge guidance template constructed in step S3.1. This template contains structured attribute information such as the geometric positional relationship parameters of fault intersection features, the mineralization indicator coefficient of lithological combination mode, and the multi-stage mineralization intensity vector of alteration superposition effect. A graph embedding method (algorithm: graph neural network GNN, parameters: node feature dimension f, adjacency matrix and weight matrix W) is used to map the nodes and their relationships in the geological anomaly knowledge guidance template into a continuous vector representation in the embedding space. Furthermore, a weighted aggregation of neighborhood node features and current node features is achieved through a GNN layer based on a message passing mechanism (parameters: propagation step number k, nonlinear activation function ReLU), and node embedding vectors under topological constraints are obtained. Furthermore, by introducing edge weight normalization processing during the embedding computation (method: normalize the weight of each edge by degree), the scale consistency of different types of geological relationships in the embedding space is achieved, and a node embedding dataset that simultaneously reflects structural correlation and semantic rules is generated. Furthermore, graph-level feature aggregation is performed on the node embedding vectors through global average pooling to generate a knowledge representation vector set for model training, ensuring that each knowledge guidance template has a unique, low-dimensional semantic encoding in the embedding space; Through the above-mentioned embedding algorithm, the geological anomaly feature knowledge template from the previous step is transformed into a computable low-dimensional knowledge representation vector set, thereby realizing a unified quantitative expression of semantic features in machine learning model training. For example, in a mining exploration scenario, the input knowledge guidance template consists of 10 types of geological anomaly feature nodes, including fault intersection nodes, 5 different lithological combination nodes, and 4 alteration superposition type nodes. The node feature dimension f is set to 64 dimensions. The adjacency matrix is ​​weighted according to spatial distance and mineralization correlation. The edge weight between fault intersection nodes and lithological nodes within 50 meters is set to 0.8, and the rest are set to 0.4 based on the correlation strength. A 2-layer GNN structure is adopted, with a propagation step k of 2 in each layer. The node feature aggregation method is weighted summation followed by ReLU activation. The message passing formula is as follows:

[0017] in, This is the updated low-dimensional semantic embedding vector for node v. Let u be the feature vector of the l-th layer neighbor node. This is the weight matrix. It is the ReLU activation function. Let v be the set of neighbors of node v. Normalized edge weights. The calculation formula is as follows:

[0018] in, Let v be the degree matrix of node v. After graph embedding, a 64-dimensional embedding vector for each knowledge template is obtained, which is then compressed to 32 dimensions by global average pooling to form a knowledge representation vector set. This vector set is input into the subsequent contrastive learning framework, which significantly improves the surrogate model's ability to distinguish geological anomalies on training samples simulating structural anomalies and normal structural patterns, and enables reliable correction of resource reserve estimation results under complex structural conditions. S3.3: Construct a contrastive learning training framework based on the knowledge representation vector set, and use a Siamese network structure to compare the features of normal mineralization pattern samples and anomalous tectonic samples to optimize the surrogate model's ability to discriminate geological anomaly patterns; S3.4: Deploy the trained surrogate model as a lightweight neural network structure, input the spatial coordinate feature vector and local geological attribute data of the current exploration area, and output the geological anomaly score of the corresponding spatial unit; S3.5: Construct a geological anomaly tensor based on the geological anomaly score to realize the quantitative mapping between geological anomaly degree and three-dimensional spatial coordinates, providing a basis for spatial positioning and anomaly intensity assessment for subsequent dynamic intervention mechanisms.

[0019] like Figure 3 As shown, step S4 involves simultaneously running the main reserve estimation model and the anomaly perception proxy model, generating two sets of output data for the same spatial grid cell: the grade prediction tensor output by the main model and the geological anomaly degree tensor output by the proxy model, forming a multimodal feature fusion decision space. Specifically, this includes: S4.1: Based on the main reserve estimation model, perform three-dimensional convolution operation on the divided regular spatial grid, input geophysical data, borehole core data and geochemical data, and output the initial resource grade prediction value and prediction confidence tensor of each grid cell to form the main model prediction output; The main reserve estimation model based on regular spatial grids receives multi-source fusion inputs of geophysical data, borehole core data and geochemical data, and establishes a spatial data cube compatible with three-dimensional convolution operations. A three-dimensional convolutional neural network convolutional layer (parameters: kernel size 3×3×3, stride 1, padding mode is same) is used to extract local three-dimensional features within a spatial data cube and retain its spatial correlation information with the geological structure direction. Furthermore, by utilizing the encoding channels in the encoder-decoder structure and employing multi-layer convolution and max pooling operations (parameters: pooling kernel size 2×2×2, stride 2), multi-scale compressed representation of the feature map is achieved, and noise interference from non-critical features on the grade prediction output is reduced. Furthermore, by employing transposed convolution operations (parameters: kernel size 3×3×3, stride 2) in the decoding channel, the spatial resolution of the feature map is restored. The high-resolution features from the encoding stage are then fused with the restored features from the decoding stage through skip connections, forming a multi-scale feature map that is sensitive to the details of geological structures. A fully connected layer and a Softmax normalization function are used to classify and regress the spatial features of the decoded output to obtain the initial resource quality prediction value corresponding to each grid cell, and a prediction confidence tensor is constructed based on the probability output. Through the mean square error loss function The formula for optimizing network parameters is as follows:

[0020] in, This represents the number of mesh cells in the X direction. This represents the number of mesh cells in the Y direction. This represents the true quality value. To predict grade values; Through the above-mentioned processing method of three-dimensional convolution—multi-scale encoding and decoding—jump connection—regression output, multi-source geological data is transformed into a two-dimensional tensor set containing the initial resource grade prediction value and prediction confidence, thereby realizing the construction of the main model prediction output; For example, in a spatial area of ​​1km × 1km × 500m within a mining exploration zone, the spatial discretization cube unit size is set to 10m × 10m × 10m, resulting in a grid of 100 × 100 × 50. Geophysical data includes magnetic anomaly intensity and gravity anomaly values; borehole core data includes the percentage of core chemical composition; and geochemical data includes elemental grade values. The convolution kernel size is selected as 3 × 3 × 3 with a stride of 1, and the pooling kernel size is 2 × 2 × 2 with a stride of 2. After encoder compression, the feature map size is reduced from 100 × 100 × 50 to 25 × 25 × 12. The decoder restores the original grid size through transposed convolution and uses skip connections to fuse features from the encoding stage. In the output layer, Softmax regression is used to obtain the grade probability distribution for each grid unit. The grade value corresponding to the highest probability is used as the prediction result, and the confidence index of this result is calculated. The output results in this area show that the predicted grades of the five grid cells near the fault intersection are significantly higher than those of the surrounding areas, and the confidence values ​​are greatly improved, which verifies the accurate identification ability of this step in grade prediction in complex tectonic areas. S4.2: Based on the anomaly perception proxy model, graph matching and semantic rule reasoning operations are performed on the same spatial grid. The high-order metallogenic mode rules extracted from the geological knowledge graph and the current spatial feature vector are input, and the geological anomaly score tensor is output to quantify the degree of deviation of the local area from the typical metallogenic features. S4.3: Perform spatiotemporal alignment processing on the grade prediction tensor output by the main model and the geological anomaly tensor output by the surrogate model. Align the prediction results and anomaly scores of each grid cell based on the spatial coordinate mapping relationship to generate synchronously aligned multimodal feature tensors. For the grade prediction tensor output by the main reserve estimation model and the geological anomaly degree tensor output by the anomaly perception proxy model, a spatial coordinate index matching algorithm (parameters: three-dimensional grid coordinate system, node unique identifier) ​​is used to calculate the positional correspondence between the two types of tensors in a unified three-dimensional reference system. Furthermore, by using the bilinear interpolation method (parameters: grid cell size, node neighborhood weight), the continuous mapping and completion of the predicted values ​​and anomalies in spatial location are achieved, and a preliminary alignment matrix with complete spatial grid coverage is obtained. Furthermore, by utilizing a timestamp synchronization mechanism (parameters: collection batch ID, model running time series), the serialization alignment of output tensors at different time points in the time dimension is achieved, and a four-dimensional aligned data structure containing time indexes is generated. Furthermore, a multidimensional index hash mapping algorithm (parameters: coordinate hash function, collision resolution strategy) is adopted to achieve fast lookup and access of the alignment matrix in terms of coordinate uniqueness and storage index, ensuring that the grade prediction value and anomaly score of each grid cell correspond one-to-one during retrieval; By performing spatiotemporal weight superposition operations (parameters: spatial similarity weight, temporal adjacency weight), a joint weight matrix is ​​calculated to optimize the alignment results in terms of spatial consistency and temporal smoothness, and a synchronously aligned multimodal feature tensor is generated to achieve high-precision input conditions for subsequent fusion processing. For example, in a mining exploration scenario, the 3D spatial grid cell size is set to 10m × 10m × 10m, and the grade prediction tensor size is 100 × 100 × 50, corresponding to spatial coordinate ranges of 0~1000m on the X-axis, 0~1000m on the Y-axis, and 0~500m on the Z-axis. The anomaly tensor has the same size, but the acquisition times differ by 2 hours, requiring time series synchronization. A spatial coordinate index matching algorithm is used to construct a grid coordinate hash table, and the hash function is set to... ,in and The grid numbers are denoted by X and Y, respectively. Bilinear interpolation is used to perform numerical interpolation when the coordinate difference does not exceed half a grid size, with interpolation weights determined by spatial relative distance. Temporal synchronization uses linear interpolation to align the time series of the anomaly tensor to the timestamp of the grade prediction tensor. During the weight optimization stage, spatial similarity is set to 0.7, and temporal adjacency is set to 0.3; the generated joint weight matrix is ​​used to adjust the alignment accuracy. The final output synchronized and aligned multimodal feature tensor shows a significant reduction in alignment deviation compared to the original data, ensuring the stability and accuracy of subsequent feature fusion and threshold determination. S4.4: The aligned grade prediction tensor and the geological anomaly tensor are spliced ​​and fused to generate a joint feature space containing predicted grade, prediction confidence and anomaly score, so as to construct a decision space tensor for multimodal feature fusion. The input is the multimodal feature tensor after the spatiotemporal alignment processing in step S4.3, which includes the grade prediction tensor and prediction confidence generated by the main reserve estimation model, and the geological anomaly tensor generated by the anomaly perception proxy model. A feature splicing and fusion algorithm (parameters: 3D spatial coordinate index, feature arrangement priority) is adopted to realize the cascade combination of the aligned grade prediction tensor and the geological anomaly tensor in the feature dimension, ensuring that features from different sources are arranged in a unified spatial grid order. Furthermore, through a multimodal weighted fusion method (parameters: prediction confidence weight, abnormal score weight), the predicted grade value, prediction confidence and abnormal score are vectorized and synthesized to obtain a joint vector containing three types of features, providing a complete input for subsequent threshold determination; Furthermore, the joint eigenvectors of each grid cell are organized into a three-dimensional tensor structure using the matrix constructor (parameters: spatial grid number, eigenvector length), achieving a strict binding between the spatial dimension and the feature dimension, and obtaining a preliminary decision space tensor. Furthermore, a data consistency verification algorithm (parameters: coordinate mapping table, feature value range setting) is adopted to automatically detect and fill out outliers and missing values ​​in the fused tensor, ensuring the numerical integrity and physical rationality of the decision space tensor; By using a multimodal feature fusion processing method, the aligned grade prediction tensor and geological anomaly tensor are transformed into a joint feature space containing predicted grade, prediction confidence and anomaly score, so as to realize the collaborative expression of multi-source information under a unified spatial reference framework. For example, in a 100×100×50 three-dimensional spatial grid scenario in a mining area, the grade prediction tensor output by the main model has a size of 100×100×50×1, the prediction confidence tensor has a size of 100×100×50×1, and the anomaly tensor output by the surrogate model has a size of 100×100×50×1. During feature splicing and fusion, the feature arrangement order is set to [predicted grade value, prediction confidence, anomaly score]. After fusion, each grid cell yields a feature vector of length 3. A weighted fusion method is used, with the prediction confidence weight set to... The abnormal rating weight is set to Generate the joint vector for each unit. ,in This is the predicted grade value. For anomaly scoring, during the matrix construction phase, all unit joint vectors are assembled into a decision space tensor with a size of 100×100×50×3 according to spatial indices. Feature components with values ​​exceeding the [0,1] range are removed through consistency checks, and missing terms are filled. In this scenario, the fused decision space tensor accurately reflects the mineralization prediction, confidence level, and anomaly degree of each grid unit, providing a stable input for S4.5's standardization processing and significantly improving the positioning accuracy of anomaly detection and knowledge intervention. S4.5: Based on the decision space tensor, perform data format standardization and normalization processing to unify the grade prediction value and anomaly score to the same numerical scale, so as to improve the stability and interpretability of subsequent threshold determination and semantic alignment operations.

[0021] Step S5: Binarize the geological anomaly tensor based on a dynamic threshold determination mechanism. When the anomaly degree of a local area exceeds an adaptive threshold, a knowledge intervention process is triggered. The adaptive threshold is dynamically adjusted based on the regional exploration data density and the prediction confidence of the master model. Specifically, this includes: S5.1: Normalize the geological anomaly tensor output by the anomaly perception proxy model. Based on the min-max normalization algorithm, map the anomaly value of each spatial grid cell to the [0,1] interval to eliminate scale differences under different geological structure scenarios and obtain the normalized geological anomaly tensor as the input basis for subsequent threshold determination. S5.2: Based on the regional exploration data density and the master model prediction confidence, a dynamic threshold function is constructed. A linear weighted fusion algorithm is used to fuse the borehole data density and the master model prediction confidence within a unit spatial grid to generate an adaptive threshold associated with the local geological data quality, so as to reflect the differences in the sensitivity of different regions to anomaly responses. Based on the normalized geological anomaly tensor, the prediction confidence tensor from the main reserve estimation model, and the borehole data density per unit space grid, a linear weighted fusion algorithm (with weight parameters based on regional data quality assessment results) is adopted to construct a dynamic threshold function, which is then adaptively adjusted according to local geological data quality characteristics. Furthermore, using statistical analysis methods (parameters: spatial grid size, borehole data location coordinate set), the borehole data density index for each spatial unit is calculated, resulting in a borehole data density tensor. This density index serves as input to the fusion algorithm to characterize the completeness of observation data in different spatial units. Furthermore, through the machine learning confidence calculation module (parameter: the predicted confidence distribution output by the main model), the predicted confidence value of the corresponding spatial unit is extracted to generate a predicted confidence tensor, which is used to reflect the reliability of the main model's results in each region; Furthermore, the adaptive threshold function value is calculated using the following linear weighted fusion formula. :

[0022] in, For adaptive threshold function, For borehole data density, To predict confidence levels, and For the weighting coefficients, satisfying + = Constraints; Furthermore, an adaptive weighting algorithm based on the regional geological data quality level (parameters: mean regional data density, mean prediction confidence) is used to adjust the data accordingly. and Spatial differentiation optimization is performed to increase the confidence weight ratio in sparse data regions and the density weight ratio in dense data regions. By using the above algorithm, borehole data density and prediction confidence are fused into an adaptive threshold tensor that is associated with the quality of local geological data, thereby achieving a quantitative expression of the sensitivity to anomalies in different regions. For example, in a mining exploration area, the spatial grid size is set to 50m × 50m × 50m, with a total of 200 boreholes. The average distribution density is 8 boreholes per cubic kilometer, with some high-density areas reaching 15 boreholes per cubic kilometer and sparse areas having less than 3 boreholes per cubic kilometer. The main reserve estimation model has a mean prediction confidence level of 0.85 in high-density areas and 0.65 in sparse areas. The initial weight values ​​are set as follows: = , = After calculation using the adaptive weight adjustment algorithm, the weight of the high-density area is adjusted to... = , = The sparse region was adjusted to = , = A linear weighted fusion formula is then used to obtain adaptive thresholds for high-density areas and sparse areas, achieving the effect of dynamic threshold changes with spatial data quality, ultimately significantly improving the reliability and effectiveness of anomaly triggering area determination; S5.3: The normalized geological anomaly tensor is compared element by element with the dynamic threshold. Based on the binarization segmentation algorithm, a judgment operation is performed on each spatial grid unit. If its anomaly value is greater than the adaptive threshold, it is marked as an anomaly triggering area; otherwise, it is marked as a non-triggering area. A binarized geological anomaly triggering mask map is generated to guide the spatial positioning of subsequent knowledge intervention processes. The input data consists of the geological anomaly tensor after S5.1 minimum-maximum normalization and the adaptive threshold matrix generated by the S5.2 linear weighted fusion algorithm. The two maintain a one-to-one correspondence in the spatial coordinate system. An element-wise comparison algorithm (parameters: normalized anomaly tensor, adaptive threshold matrix) is used to realize the numerical determination function for each spatial grid cell. Furthermore, by constructing a binarization segmentation operator (parameters: threshold comparison result, output label set), grid cells with anomalies greater than the threshold are marked as anomaly triggering regions, and other cells are marked as non-triggering regions, thus obtaining a preliminary binarization label matrix; Furthermore, through a binary matrix mapping mechanism (parameters: label matrix, spatial coordinate index), the index transformation from label value to spatial position is realized, and a mask data layout conforming to the three-dimensional mesh structure is generated; Furthermore, by using a mask coding algorithm (parameters: anomaly flag index, non-anomaly flag index), the abnormal region is set to a logical value. Non-abnormal regions are set as logical values. And generate a complete geological anomaly trigger mask map; By using the above-mentioned binarization segmentation and mask coding processing methods, the anomaly degree and threshold comparison results of the previous step are transformed into spatially explicit binary mask data, thereby achieving accurate spatial positioning for subsequent knowledge intervention processes. For example, in a 3D exploration grid of a certain mining area, the normalized geological anomaly tensor has a value range of [0,1], and the adaptive threshold matrix in high data density regions is... The data density region is Low data density areas are The element-by-element comparison algorithm, in high-data-density regions, uses the following conditional formula: ,in The normalized outlier is used; in low data density regions, the judgment condition formula is as follows: Assign values ​​to the mesh cells that meet the conditions. Assign values ​​to the remaining units A three-dimensional mask array with a size of 200×200×50 was generated. After spatial coordinate indexing and mapping, the mask successfully indicated multiple high-anomaly blocks in the fault intersection zone and lithofacies transition zone, providing accurate and coherent spatial input for downstream S5.4 connectivity region detection, and significantly improving the accuracy and effectiveness of knowledge intervention positioning; S5.4: Perform connected region detection on the generated geological anomaly trigger mask map, identify continuous anomaly trigger regions based on morphological dilation and connected component labeling algorithms, extract their spatial boundary information, and generate anomaly region boundary tensors as input conditions for semantic alignment operations in subsequent knowledge intervention processes; S5.5: Based on the comparison between the boundary tensor of the abnormal region and the preset minimum intervention region threshold, a filtering operation is performed on isolated abnormal regions with an area smaller than the set threshold. The opening operation is used to remove small abnormal regions that are mistakenly triggered, and an optimized abnormal trigger mask map is generated to avoid invalid knowledge intervention caused by local noise or minor geological disturbances. S5.6: Spatially align the optimized anomaly trigger mask with the grade prediction tensor output by the main model, extract the predicted grade data and geological feature vectors corresponding to each anomaly trigger area, and generate an anomaly area prediction feature set, which serves as the key input basis for subsequent geological semantic alignment and prediction result correction.

[0023] Step S6: Perform a geological semantic alignment operation, matching the spatial feature vector of the triggered intervention area with similar structural scenes in the knowledge graph to generate a semantic alignment parameter set containing fault strike correction coefficients, lithological weight adjustment factors, and alteration superposition compensation values. Specifically, this includes: S6.1: Based on the predefined metallogenic tectonic semantic tags in the geological knowledge graph, the three-dimensional spatial feature vector of the triggering intervention area is graph structure encoded. The three-dimensional spatial feature vector includes fault strike, lithological combination and alteration type. The topological features are extracted using a graph neural network to generate standardized geological semantic feature embedding vectors. S6.2: Based on the geological semantic feature embedding vector, execute the graph similarity matching algorithm to calculate the semantic distance between it and the typical metallogenic structural scenes stored in the knowledge graph. The typical metallogenic structural scenes include fault intersection areas, fold axis zones and lithofacies transition zones. Use cosine similarity as the matching metric to identify the most similar geological structural prototype. S6.3: Based on the semantic distance matching results, extract the geological constraint rule parameter set corresponding to the structural scene from the knowledge graph. The parameter set includes fault strike correction angle, lithological weight correction coefficient and alteration superposition compensation factor. Based on the graph structure semantic mapping mechanism, convert it into an executable knowledge guidance parameter template. S6.4: Perform local adaptive adjustments on the knowledge-guided parameter template. Based on the actual exploration data density and the prediction confidence of the master model in the triggering intervention area, use a weighted interpolation algorithm to fuse local data features and knowledge graph recommended parameters to generate a geological semantic alignment parameter set with spatial local consistency. S6.5: Output the geological semantic alignment parameter set to the main model correction module. The parameter set includes fault strike correction coefficient, lithology weight adjustment factor and alteration superposition compensation value, which are used to perform nonlinear weighted fusion of the main model prediction results in subsequent steps to achieve dynamic prediction correction based on geological knowledge. Based on the geological semantic alignment parameter set input generated by S6.4, the parameter encapsulation and interface mapping method (parameters: fault strike correction coefficient, lithology weight adjustment factor, alteration superposition compensation value) is adopted to realize the standardized encapsulation of the parameter set in the main model correction module. Furthermore, through parameter verification algorithms (parameters: tolerance threshold, data type constraints), the legality of each parameter in the geological semantic alignment parameter set is checked, and parameter set data that passes the verification is obtained; Furthermore, an asynchronous data channel transmission mechanism (parameters: buffer queue length, number of concurrent transmission threads) is adopted to achieve efficient transmission of parameter sets from the knowledge intervention module to the main reserve estimation and correction module in the intelligent analysis module, and to generate real-time code references that can be used for nonlinear weighted fusion. Furthermore, by utilizing the structure mapping algorithm (parameters: spatial coordinate index, grid cell ID), the semantic alignment parameter set is mapped to the spatial cell level of the prediction tensor inside the main model, and parameter metadata bound to each spatial cell is generated; By calling the method through the model interface (parameters: API function set, call priority), the parameter metadata from the previous step is input into the main model correction module, realizing the dynamic prediction and correction preparation stage based on geological knowledge; For example, in the northern exploration area of ​​a certain mine, the fault strike correction coefficient in the input geological semantic alignment parameter set is: Degree, lithology weight adjustment factor is Alteration superposition compensation value During the parameter encapsulation stage, the above values ​​are formatted as JSON and parameter tags are added to form standardized encapsulated data; during the verification stage, the tolerance threshold for the fault strike correction coefficient is set to [value missing]. The tolerance of the lithology weight adjustment factor is set to [value]. The tolerance of the alteration superposition compensation value is set to All verifications passed; during the transmission phase, the buffer queue length was set to [value missing]. 1 record, number of concurrent threads is To realize the parameter set in The parameter set is transmitted to the correction module within seconds; during the structure mapping stage, the parameter set is bound to the northern exploration area based on the spatial coordinate index. Each grid cell enables cell-level binding of parameters to the prediction tensor; ultimately, during the interface call phase, priority is used to determine the order of calls. The parameters were loaded, providing direct input for the subsequent nonlinear weighted fusion of S7, thus significantly improving the predicted grade in this region under complex tectonic conditions.

[0024] Step S7: Based on the semantic alignment parameter set, the prediction results of the main model are locally corrected. A weighted fusion strategy is used to nonlinearly combine the knowledge-guided parameters with the original predicted values ​​to generate a corrected resource grade distribution model, wherein the correction weight coefficient is positively correlated with the geological anomaly degree. Specifically, this includes: S7.1: Based on the fault strike correction coefficient in the geological knowledge graph, coordinate alignment correction is performed on the structural features of the corresponding spatial region in the grade prediction tensor output by the main model to eliminate the grade prediction offset caused by mismatch of structural strike. S7.2: The lithological weight adjustment factor is used to weight the lithological-related features in the corrected grade prediction tensor to enhance the contribution of specific metallogenic lithological combinations to the resource grade prediction results. S7.3: Nonlinear compensation is performed on the alteration-related parameters in the grade prediction tensor based on the alteration superposition compensation value to reflect the enhancing effect of multi-stage alteration superposition on mineralization enrichment. S7.4: Based on the anomaly score of each spatial unit in the geological anomaly tensor, a dynamic correction weight coefficient matrix that is positively correlated with it is constructed, and then the matrix is ​​fused with the predicted tensor after compensation for structure, lithology and alteration on an element-by-element weighted basis. S7.5: A nonlinear activation function is used to normalize the weighted and fused prediction tensor to generate a corrected resource grade distribution model with geophysical significance, and output the final grade prediction value and corrected confidence level of each grid cell. The input is the predicted grade distribution tensor obtained by element-wise weighted fusion with the dynamic correction weight coefficient matrix after fault strike correction, lithology weight adjustment and alteration superposition compensation. A nonlinear activation function normalization method (parameters: ReLU activation, standardized interval [0,1]) is used to perform threshold truncation and numerical compression on the grade values ​​in the fused prediction tensor, so as to eliminate the influence of local abnormal high values ​​or negative values ​​on the overall prediction distribution. Furthermore, by using a smoothing normalization algorithm (parameters: sliding window size w=5, mean smoothing weight α=0.7), the continuity constraint of the predicted value in the spatial neighborhood is realized, and the normalized grade tensor after neighborhood smoothing is obtained. Furthermore, a second correction is performed on the smoothed normalized tensor using a geophysical constraint formula, as follows:

[0025] in, To smooth and normalize the predicted grade value, To correct the output grade value, The geological feature influence coefficient, The coupling value for the construction of a spatial unit; Furthermore, by using the confidence-weighted normalization method (parameters: confidence tensor Q, weight β=0.5), the final confidence of the predicted output is adjusted to obtain a dual-channel output tensor containing the grade value and the corresponding confidence. By using nonlinear normalization and geological constraint smoothing, the results of the previous step are transformed into a modified resource grade distribution model that conforms to geophysical significance and has spatial continuity, thereby achieving the expected technical effect of improving the robustness and prediction reliability of the model in complex tectonic areas. For example, in a late-stage exploration area of ​​an open-pit mine, the original value range of the input predicted grade tensor is -3.5 to 8.2 g / t. After processing with the ReLU activation function, a non-negative value range is obtained, which is then normalized to the [0,1] interval. The sliding window size is set to 5 grid cells, and the mean smoothing weight α = 0.7. After smoothing, the fluctuation range of the grade value is significantly reduced. Based on the structural coupling value C = 0.15 of the fault intersection in this area, and setting the influence coefficient λ = 0.2, the formula is used... Calculations show that the grade value in some high-mineralization potential areas increases by approximately 0.03 g / t. Weighted normalization is performed using the confidence tensor Q (range 0.6 to 0.95) and weight β=0.5, resulting in enhanced output values ​​in high-confidence areas and moderate suppression in low-confidence areas. The final grade prediction values ​​exhibit a stable spatial distribution in complex tectonic zones, and the confidence distribution is highly consistent with the actual verification results, significantly improving the reliability of resource reserve estimation.

[0026] Step S8: Establish a closed-loop feedback mechanism by inputting the corrected prediction results verified by drilling and the corresponding geological feature labels into the knowledge graph update module. This involves optimizing the geological constraint rules and anomaly scoring function parameters in the knowledge graph through an incremental learning strategy. Specifically, this includes: S8.1: Based on drilling verification data, obtain the verified accurate spatial prediction units and their corresponding geological feature labels in the corrected resource grade distribution model. The geological feature labels include fault strike, lithological combination, alteration type and spatial coordinate information, which serve as incremental input data for knowledge graph updates. S8.2: Perform graph structure mapping processing on the incremental input data, and use graph neural network to encode geological feature labels into multimodal graph nodes. The graph nodes contain spatial coordinates, structural features, lithological combinations and alteration information, and generate incremental graph units compatible with the knowledge graph topology. S8.3: Based on the graph matching algorithm, the incremental graph unit is semantically aligned with similar structural scenes in the existing geological knowledge graph, and graph substructures with topological similarity higher than a set threshold are identified to determine the geological constraint rules and anomaly scoring function associated nodes that need to be updated; S8.4: Perform parameter incremental learning processing on the geological constraint rules that need to be updated, update the embedding vector of the corresponding node in the graph neural network based on the contrastive learning strategy, and optimize the weight parameters related to fault intersection features, lithofacies combination patterns and alteration superposition effects in the geological anomaly scoring function. S8.5: Based on the updated geological knowledge graph, an incremental knowledge version containing the latest geological feature mapping relationships and anomaly scoring rules is generated, and it is deployed to the knowledge distillation module through a version control mechanism to support online parameter synchronization updates and inference performance optimization of the anomaly perception agent model.

[0027] The present invention also provides a mining geological exploration data analysis system, which uses the above-mentioned mining geological exploration data analysis method to perform mining geological exploration data analysis.

[0028] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0029] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for analyzing geological exploration data in mines, characterized in that, Includes the following steps: S1: Construct a geological knowledge graph based on geological mineralization laws and establish a quantitative mapping relationship between spatial coordinates and geological features; S2: Construct a main reserve estimation model. After dividing the exploration area into regular spatial grids, input geophysical data, borehole core data and geochemical data, and output the initial resource grade prediction value and confidence distribution of each grid unit. S3: Construct an anomaly perception proxy model, which extracts high-order semantic rules from the geological knowledge graph, generates anomaly scoring functions, and establishes a quantitative mapping between geological anomaly degree and spatial coordinates; S4: Simultaneously run the main reserve estimation model and the anomaly perception proxy model to generate two sets of data for the same spatial grid cell: the grade prediction tensor output by the main model and the geological anomaly degree tensor output by the proxy model. S5: The geological anomaly tensor is binarized and segmented based on a dynamic threshold determination mechanism. When the anomaly of a local area exceeds the adaptive threshold, the knowledge intervention process is triggered. S6: Perform geological semantic alignment operation, perform graph matching between the spatial feature vector of the triggered intervention area and similar structural scenes in the knowledge graph, and generate a semantic alignment parameter set; S7: Based on the semantic alignment parameter set, perform local correction on the prediction results of the main model, and use a weighted fusion strategy to nonlinearly combine the knowledge guidance parameters with the original prediction values ​​to generate a corrected resource grade distribution model.

2. The method for analyzing geological exploration data in a mine according to claim 1, characterized in that, Following step S7, the following is also included: S8: Establish a closed-loop feedback mechanism, input the corrected prediction results verified by drilling and the corresponding geological feature labels into the knowledge graph update module, and optimize the geological constraint rules and anomaly scoring function parameters in the knowledge graph through incremental learning strategy.

3. The method for analyzing geological exploration data in a mine according to claim 1, characterized in that, Step S1 specifically includes: Acquire multi-source heterogeneous geological data to construct the original knowledge base of the geological knowledge graph; Based on geological mineralization patterns, entity recognition and semantic annotation are performed on the multi-source heterogeneous geological data to extract core geological elements as node representations in the knowledge graph. Using graph neural networks and knowledge extraction techniques, we extract relationships from labeled geological entities, identify the spatial association between faults and ore bodies, the causal relationship between lithological assemblage and mineralization, and the synergistic effect of alteration superposition and mineralization enrichment, and construct edge relationships in a multimodal graph structure. The multimodal graph structure is represented by node vectorization based on graph embedding algorithm, and geological entities and their relationships are mapped to a low-dimensional continuous vector space to generate an embedding model of geological knowledge graph. Establish a mapping function between spatial coordinates and geological features, and associate the three-dimensional exploration grid with nodes in the knowledge graph through a spatial indexing mechanism; An incremental update mechanism is implemented on the geological knowledge graph. Based on the geological feature labels in the field verification data and the corrected reserve prediction results, a graph attention network is used to optimize the node representation and edge weights in the knowledge graph online.

4. The method for analyzing geological exploration data in a mine according to claim 3, characterized in that, The multi-source heterogeneous geological data includes regional tectonic maps, borehole lithology logs, geochemical anomaly maps, orebody boundary data, and alteration zone distribution information.

5. The method for analyzing geological exploration data in a mine according to claim 1, characterized in that, Step S2 specifically includes: Spatial discretization is performed on the mining exploration area, and the three-dimensional geological space is divided into equally spaced cubic grid units based on the geological coordinate system to establish a unified spatial reference frame; Geophysical data, borehole core data, and geochemical data are acquired, and the geophysical data, borehole core data, and geochemical data are standardized and preprocessed to generate structured data cubes. A main reserve estimation model is constructed based on a three-dimensional convolutional neural network. The main reserve estimation model is trained in a supervised manner based on a labeled training sample set, wherein the training sample includes the structured data cube and the corresponding true grade value label, and the model parameters are optimized using the mean square error loss function. The structured data cube is input into the trained main reserve estimation model, and resource grade prediction is performed cell by cell to generate a three-dimensional grade distribution tensor and calculate the prediction confidence index for each cell.

6. The method for analyzing geological exploration data in a mine according to claim 5, characterized in that, The standardized preprocessing includes missing value imputation, outlier removal, and feature normalization.

7. The method for analyzing geological exploration data in a mine according to claim 1, characterized in that, Step S3 specifically includes: Based on the multimodal geological structural rules stored in the geological knowledge graph, high-order semantic representations of fault intersection features, lithological combination patterns, and alteration superposition effects are extracted to construct knowledge-guided templates for geological anomaly features. The knowledge-guided template for the geological anomaly features is subjected to graph embedding processing, and a graph neural network is used to map it into a geological knowledge embedding representation in a low-dimensional semantic vector space to form a knowledge representation vector set. Based on the knowledge representation vector set, a contrastive learning training framework is constructed, and a Siamese network structure is used to compare the features of normal mineralization pattern samples and abnormal construction samples. The trained proxy model is deployed as a lightweight neural network structure, which takes into account the spatial coordinate feature vector and local geological attribute data of the current exploration area, and outputs the geological anomaly score of the corresponding spatial unit. Based on the geological anomaly score, a geological anomaly degree tensor is constructed.

8. The method for analyzing geological exploration data in a mine according to claim 1, characterized in that, Step S4 specifically includes: Based on the main reserve estimation model, a three-dimensional convolution operation is performed on the divided regular spatial grid. Geophysical data, borehole core data and geochemical data are input, and the initial resource grade prediction value and prediction confidence tensor of each grid cell are output to form the grade prediction tensor. Based on the anomaly perception agent model, graph matching and semantic rule reasoning operations are performed on the same spatial grid. The input is the high-order metallogenic mode rules extracted from the geological knowledge graph and the current spatial feature vector. The output is the geological anomaly score tensor. Spatiotemporal alignment processing is performed on the grade prediction tensor and the geological anomaly score tensor. Based on the spatial coordinate mapping relationship, the prediction results and anomaly scores of each grid cell are aligned to generate synchronously aligned multimodal feature tensors. The aligned grade prediction tensor and the geological anomaly tensor are spliced ​​and fused to generate a joint feature space, and a decision space tensor for multimodal feature fusion is constructed. Based on the decision space tensor, data format standardization and normalization are performed to unify the grade prediction value and the anomaly score to the same numerical scale.

9. The method for analyzing geological exploration data in a mine according to claim 8, characterized in that, The joint feature space includes predicted grade, prediction confidence, and anomaly score.

10. A mining geological exploration data analysis system, characterized in that: The geological exploration data analysis method described in any one of claims 1-9 is used to perform the geological exploration data analysis.

Images