Multi-source data fusion oriented comprehensive geological genesis analysis method and system

Abstract

The application discloses a comprehensive geological genesis analysis method and system for multi-source data fusion, and comprises the following steps: step one, generating semantic mapping data; step two, constructing a cause characteristic pedigree tree and outputting cause main chain data; step three, generating geological space-time interleaving data; step four, limiting a time window and a space window, and generating spectrum fusion modeling data through a spectrum guide fusion transformer; step five, constructing a three-dimensional voxel grid, and performing dynamic clustering to generate a three-dimensional structure model; step six, generating geological evolution field data; and step seven, inputting the geological evolution field data into an improved CrossViT model, setting an interlayer cross attention unit, reconstructing a Token interaction mode, introducing a cause field constraint attention modulation mechanism, and outputting a geological genesis analysis result. Through multi-source data semantic mapping, cause pedigree reasoning and the improved CrossViT model, high-precision analysis of complex geological genesis is realized.

Description

Technical Field

[0001] This invention relates to the field of geological information processing and intelligent analysis technology, and in particular to a comprehensive geological genetic analysis method and system for multi-source data fusion. Background Technology

[0002] With the increasing depth of mineral resource exploration and the growing demand for research on the genetic mechanisms under complex geological conditions, comprehensive analysis and genetic identification techniques for multi-source geological data have received widespread attention. Existing geological genetic analysis methods mainly rely on empirical interpretation using a single data source or a limited number of data types. However, in practical applications, due to the complexity of geological body formation processes and the multi-scale coupling characteristics, these methods generally suffer from the following problems:

[0003] Geological data from different sources vary significantly in spatial resolution, sampling methods, and physical significance. Remote sensing hyperspectral data, geochemical spectral data, mineral X-ray spectral data, and drilling record data are difficult to align effectively within a unified spatial framework. Existing multi-source data fusion methods often employ simple overlay or statistical normalization, which can easily lead to overlapping genetic information or distortion of local features, making it difficult to accurately characterize the true genetic structure within geological bodies. Traditional geological genetic analysis relies heavily on human experience or rule-based reasoning, lacking systematic modeling of the transmission relationships between various genetic elements such as tectonic activity, material sources, and spatial distribution. Existing machine learning-based analysis methods typically ignore hierarchical constraints and temporal characteristics in the genetic evolution process, resulting in unstable and uninterpretable genetic path inferences. Furthermore, existing geological modeling techniques are mostly focused on geometric structure reconstruction, lacking a unified expression of the spatiotemporal interrelationships of geological events and the synergistic effects of multi-source spectral features, making it difficult to support the refined requirements of genetic discrimination and evolutionary analysis in complex geological environments.

[0004] Therefore, how to provide a comprehensive geological genetic analysis method and system for multi-source data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a comprehensive geological genetic analysis method and system for multi-source data fusion. This invention constructs a genetic feature phylogenetic tree and performs genetic path reasoning by semantic mapping and unified spatial alignment of multi-source geological data. On this basis, an improved CrossViT model is introduced to perform genetic discrimination of geological evolution fields, realizing fine modeling and high-precision analysis of complex geological genetic processes. It has the advantages of strong result stability, high spatial consistency and outstanding genetic interpretation ability.

[0006] The comprehensive geological genetic analysis method for multi-source data fusion according to embodiments of the present invention includes the following steps:

[0007] Step 1: Perform semantic analysis on heterogeneous geological data to form semantic substructure data, and use a multi-semantic transformation network to perform spatial domain projection on the semantic substructure data to generate semantic mapping data under a unified geological spatial coordinate system;

[0008] Step 2: Based on semantic mapping data, construct a causal feature phylogenetic tree according to the causal pathology rules, and perform correlation analysis and screening on each causal path in the causal feature phylogenetic tree through the phylogenetic association propagation module to output the causal main chain data;

[0009] Step 3: Based on the causal main chain data, construct a three-axis solid cell of time-space-geological type, and establish an intersecting event record in the three-axis solid cell to generate geological spatiotemporal intersecting data;

[0010] Step 4: Based on the geological spatiotemporal interleaving data, the time and space windows are defined. The remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors included in the semantic mapping data are converted into corresponding two-dimensional spectral vector fields. The two-dimensional spectral vector fields are fused through a spectral guided fusion transformer to generate spectral fusion modeling data.

[0011] Step 5: Construct a 3D voxel mesh based on the spectral fusion modeling data, and use the dynamic voxel clustering module to perform dynamic clustering to generate a 3D construction model;

[0012] Step Six: Based on the three-dimensional structural model, construct the local environmental field, regional tectonic field, and genetic composite field to generate geological evolution field data;

[0013] Step 7: Input the geological evolution field data into the improved CrossViT model, set up interlayer cross attention units, reconstruct the token interaction method, introduce the genetic field-constrained attention modulation mechanism, and output the geological genetic analysis results.

[0014] Optionally, the heterogeneous geological data specifically includes remote sensing hyperspectral data, geochemical spectral data, mineral X-ray spectral data, and drilling record data.

[0015] Optionally, step one specifically includes:

[0016] Data source identification and structural analysis are performed on heterogeneous geological data. Remote sensing hyperspectral data is divided into grids according to the spatial location of pixels. The corresponding multi-band spectral response vector is extracted for each grid unit to generate remote sensing hyperspectral semantic substructure data indexed by spatial grids.

[0017] Geochemical spectrum data is mapped according to the spatial coordinates of sample collection points, and the elemental spectrum values ​​of each collection point are vectorized to generate geochemical spectrum semantic substructure data indexed by the coordinates of the sampling points.

[0018] The diffraction angle position and peak intensity information in the mineral X-ray spectrum data are numerically expanded and spatially labeled in combination with the sample source position to generate mineral X-ray spectrum semantic substructure data indexed by the sample spatial position.

[0019] The drilling record data is sequentially decomposed according to borehole number, depth range and corresponding lithology information to generate drilling semantic substructure data with depth coordinates;

[0020] Various semantic substructure data are input into a multi-semantic transformation network, which includes four independently configured transformation subnetworks. Each transformation subnetwork includes an input encoding layer, a feature mapping layer, and an output alignment layer.

[0021] In each transformation sub-network, the corresponding semantic substructure data is converted into a feature vector of a unified dimension, and spatial location encoding is introduced in the output alignment layer to map the feature vector to a unified geological spatial coordinate system.

[0022] Various features projected in the spatial domain are vectorized and aggregated to generate semantic mapping data, which includes remote sensing hyperspectral vectors, geochemical spectral vectors, mineral X-ray spectral vectors, and drilling record vectors.

[0023] Optionally, step two specifically includes:

[0024] The semantic mapping data is decomposed into causal elements to obtain structural causal elements, material causal elements and spatial causal elements. The structural causal elements, material causal elements and spatial causal elements are then vectorized to generate structural causal element vectors, material causal element vectors and spatial causal element vectors, respectively.

[0025] According to the rules of causal pathology, the causal element vectors are divided into levels: the structural causal element vectors are divided into the first level, the material causal element vectors are divided into the second level, and the spatial causal element vectors are divided into the third level.

[0026] Based on the hierarchical division results, the connection methods between causal element vectors are limited. Only directed connections are allowed between the first-level causal element vector and the second-level causal element vector, and only directed connections are allowed between the second-level causal element vector and the third-level causal element vector.

[0027] Using the causal element vector as the node set and the directed connection relationship as the edge set, a causal feature spectrum tree is constructed. The causal feature spectrum tree is a three-level directed acyclic structure, and the causal path is formed by the first-level nodes, the second-level nodes, and the third-level nodes in sequence.

[0028] The causal feature phylogenetic tree is input to the phylogenetic association propagation module. The phylogenetic association propagation module takes the causal element vectors in the node set as node inputs, performs layer-by-layer propagation calculations according to the hierarchical structure of the causal feature phylogenetic tree, and uses cosine similarity calculation to obtain correlation values ​​between causal element vectors at adjacent levels.

[0029] The correlation values ​​between nodes at each level in the same causal path are sequentially accumulated to obtain the path correlation score of the corresponding causal path, which is used as the output of the genealogical association propagation module.

[0030] Sort the path relevance scores of all causal paths, select the causal paths with scores higher than the preset score threshold as the causal main chain data and output them.

[0031] Optionally, step three specifically includes:

[0032] Each genetic path in the genetic main chain data is analyzed to extract the corresponding genetic time information, spatial location information and geological type information.

[0033] The formation time information is divided into time segments according to a preset time resolution, the spatial location information is divided into spatial units according to a unified geological spatial coordinate system, and the geological type information is divided into geological type units according to geological categories.

[0034] Using the time segment, spatial unit, and geological type unit as three orthogonal dimensions, a three-dimensional cell with time-space-geological type three-axis is constructed, where each three-dimensional cell corresponds to a unique combination of time, space, and geological type.

[0035] Each causal path in the causal main chain data is mapped to a corresponding three-dimensional cell according to its corresponding time segment, spatial unit, and geological type unit. When different causal paths correspond to different time segments within the same spatial unit and geological type unit, or correspond to different geological type units within the same time segment and spatial unit, an interleaved event record body is established in the corresponding three-dimensional cell. The interleaved event record body includes a causal path identifier, a time segment identifier, a spatial unit identifier, a geological type unit identifier, and an interleaved type identifier. When the interleaved type identifier is 1, it indicates time interleaving; when it is 2, it indicates geological type interleaving.

[0036] All intersecting event records are summarized to generate geological spatiotemporal intersecting data.

[0037] Optionally, step four specifically includes:

[0038] Based on geological spatiotemporal interleaving data, each interleaving event record is parsed to read the corresponding time segment identifier and spatial unit identifier. The time window for data selection is limited by the time segment identifier and the spatial window for data selection is limited by the spatial unit identifier. Remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors that fall within the time window and spatial window are filtered from semantic mapping data.

[0039] Within the time and space windows, the remote sensing hyperspectral vectors are arranged according to the two-dimensional grid coordinates of the spatial units, so that each grid position corresponds to a remote sensing hyperspectral vector, generating a remote sensing hyperspectral two-dimensional spectral vector field; the geochemical spectral vectors are interpolated and mapped according to the two-dimensional projection positions of the sampling points within the spatial window, generating a geochemical two-dimensional spectral vector field; the mineral X-ray spectral vectors are arranged according to the correspondence between the spatial positions of the samples on the two-dimensional plane, generating a mineral X-ray two-dimensional spectral vector field.

[0040] Obtain the stratigraphic interface data corresponding to the spatial window, and convert the stratigraphic interface data into a two-dimensional stratigraphic interface guide map. Different stratigraphic units in the two-dimensional stratigraphic interface guide map are distinguished by boundary lines, and the stratigraphic unit identifier is marked for each two-dimensional spatial location.

[0041] The two-dimensional stratigraphic interface guidance map is introduced as a structural guidance factor into the spectral guidance fusion transformer. In the spectral guidance fusion transformer, the feature interaction range in the two-dimensional spectral vector field is limited according to the stratigraphic unit identifier, so that the spectral vectors can only interact between two-dimensional spatial positions within the same stratigraphic unit, and feature interaction across stratigraphic interfaces is blocked.

[0042] In the spectral-guided fusion converter, the spectral vectors of the remote sensing hyperspectral two-dimensional spectral vector field, the geochemical two-dimensional spectral vector field, and the mineral X-ray two-dimensional spectral vector field located in the same stratigraphic unit and at the same two-dimensional spatial location are spliced ​​together to obtain the spectral fusion feature vector.

[0043] The spectral fusion feature vectors are aggregated to generate spectral fusion modeling data.

[0044] Optionally, step five specifically includes:

[0045] Based on spectral fusion modeling data, on the basis of two-dimensional spatial coordinates determined by a unified geological spatial coordinate system, corresponding depth coordinates are introduced, and corresponding depth information is added to each spectral fusion feature vector to construct a three-dimensional spectral fusion data volume jointly characterized by two-dimensional spatial coordinates, depth coordinates and spectral fusion feature vectors.

[0046] The three-dimensional spectral fusion data volume is spatially discretized according to a preset voxel size, dividing the three-dimensional geological space into a regularly arranged three-dimensional voxel grid, and assigning a corresponding spectral fusion feature vector and spatial coordinate index to each voxel to form a three-dimensional voxel set.

[0047] The three-dimensional voxel set is input to the dynamic voxel clustering module, which uses voxels as the basic processing unit and uses the voxel's spectral fusion feature vector, spatial coordinates, and spatial adjacency relationship between voxels as joint inputs.

[0048] In the dynamic voxel clustering module, a voxel adjacency graph is constructed based on the spatial adjacency relationship of voxels in the three-dimensional voxel grid. Under the constraint of the voxel adjacency graph, the cosine similarity of the spectral fusion feature vectors between adjacent voxels is calculated. When the cosine similarity is greater than the preset clustering threshold, the corresponding voxels are assigned to the same candidate clustering unit.

[0049] A region-growing-based clustering algorithm is used to iteratively update the clustering labels of voxels, and to merge and adjust candidate clustering units so that the voxel clustering results simultaneously satisfy the spectral fusion feature similarity constraint and the spatial continuity constraint, until the voxel clustering results reach a stable state.

[0050] The stable voxel clustering results are used as the structural cluster partitioning results, and the voxel sets belonging to the same structural cluster are spatially combined to generate a three-dimensional construction model.

[0051] Optionally, step six specifically includes:

[0052] Read the voxel set, voxel three-dimensional spatial coordinates, and voxel spectrum fusion feature vector corresponding to each structural cluster in the three-dimensional construction model;

[0053] Taking the continuous voxel region corresponding to a single structure cluster as the computational object, the voxel spectrum fusion feature vector is rearranged according to the voxel space coordinates, and the spectral fusion feature value of each dimension is normalized by the min-max normalization method. The normalized voxel spectrum fusion feature value set is constructed into a three-dimensional scalar field, and the three-dimensional scalar field is defined as the local environment field.

[0054] Taking multiple structural clusters with spatial contact relationships in a three-dimensional construction model as the calculation object, the principal component analysis method is used to calculate the principal extension direction vector of the structural clusters based on the three-dimensional spatial coordinates of all voxels within the structural clusters.

[0055] The spectral fusion feature vectors of each voxel within the structure cluster combination are weighted statistically calculated. The weight of each voxel is determined based on the Euclidean distance to the geometric center of the structure cluster combination. The voxel weight is inversely proportional to the Euclidean distance. All voxel weights are normalized so that the sum of the voxel weights is 1.

[0056] The spectral fusion feature vectors of the corresponding voxels are weighted and summed using the normalized voxel weights to obtain the weighted spectral fusion feature vectors of the structural cluster combination. These feature vectors are then concatenated with the main extension direction vectors in three-dimensional space to construct a three-dimensional vector field, which is defined as a regional construction field.

[0057] Under a unified three-dimensional spatial coordinate system, the three-dimensional scalar field in the local environmental field and the three-dimensional vector field in the regional tectonic field are aligned on a voxel-by-voxel basis. At each voxel spatial location, the corresponding scalar value and vector value are concatenated and encoded to generate a set of joint feature vectors on a voxel basis. The set of joint feature vectors is then constructed into a three-dimensional joint field, which is defined as a genetic synthesis field.

[0058] The local environmental field, regional tectonic field, and genetic composite field are organized and aggregated according to a unified voxel spatial coordinate index to generate geological evolution field data.

[0059] Optionally, step seven specifically includes:

[0060] Geological evolution field data are encoded according to field type and spatial level to form corresponding input token sequences. The token sequences are then input into the improved CrossViT model. The token sequences corresponding to local environmental fields are used as first-level input tokens, the token sequences corresponding to regional tectonic fields are used as second-level input tokens, and the token sequences corresponding to genetic synthesis fields are used as third-level input tokens. Each token includes voxel spatial coordinate encoding and corresponding field feature vector.

[0061] The improved CrossViT model includes a first feature branch, a second feature branch, and a third feature branch. The first feature branch receives a first-level input token, the second feature branch receives a second-level input token, and the third feature branch receives a third-level input token. Each feature branch is composed of multiple layers of self-attention units connected sequentially.

[0062] Interlayer cross-attention units are set between the first feature branch and the second feature branch, and between the second feature branch and the third feature branch. The interlayer cross-attention units are connected sequentially in the order from the first feature branch to the second feature branch and from the second feature branch to the third feature branch, forming a progressive interlayer structure of cross-attention.

[0063] In the inter-layer cross-attention unit, the token from the lower-level feature branch serves as the key token and the value token, and the token from the higher-level feature branch serves as the query token. Cross-attention calculation is then performed based on the query token, the key token, and the value token.

[0064] In the process of inter-layer cross-attention calculation, a causal field-constrained attention modulation mechanism is introduced. The causal field-constrained attention modulation mechanism uses the causal comprehensive field Token in the third-level input Token as the constraint Token. Based on the feature similarity between the constraint Token and the query Token and key Token participating in the cross-attention calculation, the corresponding weight element in the cross-attention weight matrix is ​​subjected to numerical scaling operation.

[0065] At the output of the third feature branch, all tokens output by the third feature branch are aggregated and calculated to generate a cause discrimination token.

[0066] A classification calculation is performed on the gene formation discrimination token to generate a gene formation category identifier corresponding to the voxel spatial coordinates. The gene formation category identifier is bound and stored with the corresponding voxel spatial coordinates to generate a voxel-by-voxel gene formation discrimination result set. The gene formation category identifier set is output as the geological gene formation analysis result.

[0067] The comprehensive geological genetic analysis system for multi-source data fusion according to an embodiment of the present invention includes the following modules:

[0068] The semantic mapping module is used to perform semantic analysis on heterogeneous geological data, form semantic substructure data, and use a multi-semantic transformation network to perform spatial domain projection to generate semantic mapping data.

[0069] The causal feature spectrum construction module is used to construct a causal feature spectrum tree based on the semantic mapping data and in accordance with causal pathology rules.

[0070] The genealogical association propagation module is used to perform correlation analysis and screening on each causal path in the causal feature genealogical tree and output the causal main chain data.

[0071] The geological spatiotemporal intersection construction module is used to construct a three-dimensional cell based on the genetic main chain data, and to establish an intersection event record body to generate geological spatiotemporal intersection data;

[0072] The spectrum-guided fusion modeling module is used to convert the remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors included in the semantic mapping data into corresponding two-dimensional spectral vector fields within the time and space windows defined by the geological spatiotemporal interleaved data. The two-dimensional spectral vector fields are then fused through the spectrum-guided fusion transformer to generate spectrum fusion modeling data.

[0073] A voxel mesh construction module is used to construct a three-dimensional voxel mesh based on the spectral fusion modeling data;

[0074] The dynamic voxel clustering module is used to perform dynamic clustering processing on the voxels in the three-dimensional voxel mesh to generate a three-dimensional construction model;

[0075] The geological evolution field construction module is used to construct local environmental fields, regional tectonic fields, and genetic integrated fields based on the three-dimensional structural model, and generate geological evolution field data.

[0076] The causal discrimination module is used to input the geological evolution field data into the improved CrossViT model for causal discrimination calculation and output the geological causal analysis results corresponding to the voxel space coordinates.

[0077] The beneficial effects of this invention are:

[0078] This invention addresses the inconsistencies in spatial resolution, representation, and physical semantics among remote sensing hyperspectral data, geochemical spectral data, mineral X-ray spectral data, and drilling record data by performing semantic analysis on heterogeneous geological data and constructing semantic substructure data. It then combines this with a multi-semantic transformation network to complete spatial domain projection and generate semantic mapping data. This achieves semantic alignment and unified spatial representation of multi-source geological information. In the genetic analysis stage, based on the semantic mapping data, this invention constructs a genetic feature phylogenetic tree according to genetic pathology rules. A phylogenetic correlation propagation module is used to perform correlation analysis and screening of each genetic path, constraining complex multi-source genetic relationships into a genetic main chain data that conforms to the logic of geological genetic transmission. This avoids inference noise caused by invalid genetic combinations and improves the stability and interpretability of genetic path discrimination.

[0079] In the spatiotemporal modeling stage, by constructing a three-axis 3D cell based on time, space, and geological type and introducing interleaved event records, the genetic main chain data is transformed into structured geological spatiotemporal interleaved data. This allows the genetic evolution process to be characterized within a unified spatiotemporal framework, effectively expressing the superposition and interleaving relationships of different genetic events in time and space. In the spectral feature fusion stage, this invention limits the time and space windows based on the geological spatiotemporal interleaved data and uses a spectral-guided fusion transformer to fuse the two-dimensional spectral vector field under structural constraints, avoiding the mixing of spectral features across stratigraphic interfaces and improving the genetic consistency of the spectral fusion modeling data. In the 3D modeling stage... The system dynamically clusters three-dimensional voxel meshes using a dynamic voxel clustering module to generate stable three-dimensional structural models. It further constructs local environmental fields, regional tectonic fields, and genetic comprehensive fields, transforming discrete structural results into continuous geological evolution field data. Finally, the geological evolution field data is input into the improved CrossViT model. Through a progressive interlayer cross-attention structure and a genetic field-constrained attention modulation mechanism, it achieves collaborative discrimination of multi-level genetic information and outputs geological genetic analysis results corresponding to voxel spatial coordinates. This significantly improves the accuracy, stability, and automation level of comprehensive geological genetic analysis in complex geological environments. Attached Figure Description

[0080] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0081] Figure 1 This is an overall flowchart of the comprehensive geological genetic analysis method for multi-source data fusion proposed in this invention;

[0082] Figure 2 This is a schematic diagram of the structure of the comprehensive geological genetic analysis system for multi-source data fusion proposed in this invention. Detailed Implementation

[0083] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0084] Example 1: Reference Figure 1 A comprehensive geological genetic analysis method for multi-source data fusion includes the following steps:

[0085] Step 1: Perform semantic analysis on heterogeneous geological data to form semantic substructure data, and use a multi-semantic transformation network to perform spatial domain projection on the semantic substructure data to generate semantic mapping data under a unified geological spatial coordinate system;

[0086] Step 2: Based on semantic mapping data, construct a causal feature phylogenetic tree according to the causal pathology rules, and perform correlation analysis and screening on each causal path in the causal feature phylogenetic tree through the phylogenetic association propagation module to output the causal main chain data;

[0087] Step 3: Based on the causal main chain data, construct a three-axis solid cell of time-space-geological type, and establish an intersecting event record in the three-axis solid cell to generate geological spatiotemporal intersecting data;

[0088] Step 4: Based on the geological spatiotemporal interleaving data, the time and space windows are defined. The remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors included in the semantic mapping data are converted into corresponding two-dimensional spectral vector fields. The two-dimensional spectral vector fields are fused through a spectral guided fusion transformer to generate spectral fusion modeling data.

[0089] Step 5: Construct a 3D voxel mesh based on the spectral fusion modeling data, and use the dynamic voxel clustering module to perform dynamic clustering to generate a 3D construction model;

[0090] Step Six: Based on the three-dimensional structural model, construct the local environmental field, regional tectonic field, and genetic composite field to generate geological evolution field data;

[0091] Step 7: Input the geological evolution field data into the improved CrossViT model, set up interlayer cross attention units, reconstruct the token interaction method, introduce the genetic field-constrained attention modulation mechanism, and output the geological genetic analysis results.

[0092] In this embodiment, the heterogeneous geological data specifically includes remote sensing hyperspectral data, geochemical spectrum data, mineral X-ray spectrum data, and drilling record data;

[0093] Remote sensing hyperspectral data is acquired through airborne or satellite remote sensing platforms equipped with hyperspectral imagers, collecting surface reflectance spectra. After atmospheric and radiometric correction, it is used to identify surface mineral distribution and alteration characteristics. Primarily, in step four, it is converted into a two-dimensional spectral vector field for spectral fusion modeling, revealing genetically related spectral anomaly regions. Geochemical spectral data is obtained through regional sample collection and chemical analysis, reflecting geochemical anomaly distributions. Also in step four, it is used to construct a two-dimensional spectral vector field and fuse with other spectra to identify the distribution characteristics of ore-controlling element combinations. Mineral X-ray spectral data is acquired through laboratory X-ray diffraction techniques, characterizing the crystal structure and composition of minerals in samples. In step four, it provides spectral features as an important information source for mineral phase identification, enhancing the phase interpretation capability of spectral fusion results. Drilling record data is obtained from actual drilling operations, recording stratigraphic depth, lithological changes, and structural information. It is used for spatial mapping in step one and supports structural modeling and depth positioning in step five. Through collaborative use across the entire process, a unified spatial spectral structure fusion system is generated.

[0094] In this embodiment, step one specifically includes:

[0095] Data source identification and structural analysis are performed on heterogeneous geological data. Remote sensing hyperspectral data is divided into grids according to the spatial location of pixels. The corresponding multi-band spectral response vector is extracted for each grid unit to generate remote sensing hyperspectral semantic substructure data indexed by spatial grids.

[0096] Geochemical spectrum data is mapped according to the spatial coordinates of sample collection points, and the elemental spectrum values ​​of each collection point are vectorized to generate geochemical spectrum semantic substructure data indexed by the coordinates of the sampling points.

[0097] The diffraction angle position and peak intensity information in the mineral X-ray spectrum data are numerically expanded and spatially labeled in combination with the sample source position to generate mineral X-ray spectrum semantic substructure data indexed by the sample spatial position.

[0098] The drilling record data is sequentially decomposed according to borehole number, depth range and corresponding lithology information to generate drilling semantic substructure data with depth coordinates;

[0099] Various semantic substructure data are input into a multi-semantic transformation network, which includes four independently configured transformation subnetworks. Each transformation subnetwork includes an input encoding layer, a feature mapping layer, and an output alignment layer.

[0100] In each transformation sub-network, the corresponding semantic substructure data is converted into a feature vector of a unified dimension, and spatial location encoding is introduced in the output alignment layer to map the feature vector to a unified geological spatial coordinate system.

[0101] The various features projected in the spatial domain are vectorized and aggregated to generate semantic mapping data, which includes remote sensing hyperspectral vectors, geochemical spectral vectors, mineral X-ray spectral vectors, and drilling record vectors.

[0102] In this invention, heterogeneous geological data are first identified and structurally analyzed. Different data sources are processed according to their spatial representation and data structure. Remote sensing hyperspectral data is divided into regular grids with a fixed spatial resolution of 30m × 30m. Each grid cell corresponds to a set of pixels. The reflectance values ​​of all bands in this set are arranged in band order to form a multi-band spectral response vector of length 128, serving as the semantic substructure data of the remote sensing hyperspectral data, with the spatial coordinates of the corresponding grid center point as the index. Geochemical spectral data is spatially mapped according to the latitude and longitude coordinates of the sample collection points. The elemental content data measured at each sampling point are arranged in a fixed elemental sequence to form an elemental spectral vector of length 32, serving as the semantic substructure data of the geochemical spectrum. Mineral X-ray spectral data has a diffraction angle range of 5 degrees to 70 degrees and a step size of 0.02 degrees. The peak intensities at corresponding angle positions are arranged sequentially to form... X-ray spectral vectors with a length of 3250 are bound to the spatial location of the sample source to form mineral X-ray spectral semantic substructure data. Drilling record data are indexed sequentially according to borehole number, and the lithological code and structural identifier of the corresponding depth segment are expanded into vectors of length 24 at 1-meter depth intervals to serve as drilling semantic substructure data. The above four types of semantic substructure data are respectively input into four independent transformation subnetworks in a multi-semantic transformation network. Each transformation subnetwork contains an input encoding layer, a feature mapping layer, and an output alignment layer. The input encoding layer maps the original vectors to a unified 64-dimensional feature space. The feature mapping layer performs nonlinear transformation on the features. The output alignment layer introduces spatial location encoding and completes spatial domain projection, ultimately forming remote sensing hyperspectral vectors, geochemical spectral vectors, mineral X-ray spectral vectors, and drilling record vectors. The four vectors are spatially aligned under a unified geological spatial coordinate system to form semantic mapping data.

[0103] In this embodiment, step two specifically includes:

[0104] The semantic mapping data is decomposed into causal elements to obtain structural causal elements, material causal elements and spatial causal elements. The structural causal elements, material causal elements and spatial causal elements are then vectorized to generate structural causal element vectors, material causal element vectors and spatial causal element vectors, respectively.

[0105] Specifically, the genetic element decomposition of semantic mapping data first involves regular analysis based on data source and physical meaning. Feature terms related to fault distribution, tectonic boundaries, and spatial continuity changes are extracted and grouped into tectonic genetic elements; feature terms related to elemental combination characteristics, mineral facies composition, and material source indications are extracted and grouped into material genetic elements; and feature terms related to the location of genetic processes, enrichment zones, and spatial distribution range are extracted and grouped into spatial genetic elements. During the decomposition process, each type of genetic element maintains its spatial index under a unified geological spatial coordinate system, thereby ensuring the consistency of subsequent genetic path construction. After the genetic elements are decomposed, each type of genetic element is vectorized. The structural genetic elements are formed by concatenating the corresponding feature terms in a fixed order and then normalizing the values. The material genetic elements are formed by standardizing and combining the element spectrum feature values ​​and mineral feature values. The spatial genetic elements are formed by encoding the spatial coordinate information and the segment identification information. The above vectorized expression results serve as the basic input data for the construction of the genetic feature phylogenetic tree and the calculation of the phylogenetic association propagation module.

[0106] According to the rules of causal pathology, the causal element vectors are divided into levels: the structural causal element vectors are divided into the first level, the material causal element vectors are divided into the second level, and the spatial causal element vectors are divided into the third level.

[0107] Based on the hierarchical division results, the connection methods between causal element vectors are limited. Only directed connections are allowed between the first-level causal element vector and the second-level causal element vector, and only directed connections are allowed between the second-level causal element vector and the third-level causal element vector.

[0108] Using the causal element vector as the node set and the directed connection relationship as the edge set, a causal feature spectrum tree is constructed. The causal feature spectrum tree is a three-level directed acyclic structure, and the causal path is formed by the first-level nodes, the second-level nodes, and the third-level nodes in sequence.

[0109] The causal feature phylogenetic tree is input to the phylogenetic association propagation module. The phylogenetic association propagation module takes the causal element vectors in the node set as node inputs, performs layer-by-layer propagation calculations according to the hierarchical structure of the causal feature phylogenetic tree, and uses cosine similarity calculation to obtain correlation values ​​between causal element vectors at adjacent levels.

[0110] The correlation values ​​between nodes at each level in the same causal path are sequentially accumulated to obtain the path correlation score of the corresponding causal path, which is used as the output of the genealogical association propagation module.

[0111] Sort the path relevance scores of all causal paths, select the causal paths with scores higher than the preset score threshold as the causal main chain data and output them;

[0112] In step two, the semantic mapping data is decomposed into causal elements, explicitly divided into tectonic, material, and spatial causal elements. This ensures that geological information from different sources and semantic levels is structured and standardized before entering the computational model, avoiding semantic confusion and correlation distortion caused by direct fusion of multi-source data. Each type of causal element is represented by a fixed 128-dimensional vector. This dimension ensures the integrity of the causal semantic expression while balancing computational efficiency and model stability. Following the objective geological laws of tectonic control-material response-spatial distribution in genetic pathology, the tectonic causal element vector is set as the first level, the material causal element vector as the second level, and the spatial causal element vector as the third level. By limiting unidirectional connections between the first and second levels and between the second and third levels, a directed acyclic structure conforming to the actual causal transmission logic is forced, eliminating unreasonable reverse or cross-level correlations at the structural level. Based on these constraints, a model is constructed... The three-layer causal feature phylogenetic tree effectively compresses the invalid combination space, transforming the causal path search from a fully connected problem into a controlled path reasoning problem. In the phylogenetic association propagation module, cosine similarity is used as the inter-layer correlation measure, with its value range limited to 0~1. This can stably reflect the consistency of causal element vectors in direction and reduce the interference caused by amplitude differences. By sequentially accumulating the correlation values ​​of the three-layer nodes in the same path, a path correlation score is formed, and a score threshold of 2.1 is set. Only causal paths with scores higher than the score threshold are retained, thereby realizing the automatic identification and output of causal main chain data and improving the reliability and interpretability of the comprehensive geological causal analysis results.

[0113] In this embodiment, step three specifically includes:

[0114] Each genetic path in the genetic main chain data is analyzed to extract the corresponding genetic time information, spatial location information and geological type information.

[0115] The formation time information is divided into time segments according to a preset time resolution, the spatial location information is divided into spatial units according to a unified geological spatial coordinate system, and the geological type information is divided into geological type units according to geological categories.

[0116] Using the time segment, spatial unit, and geological type unit as three orthogonal dimensions, a three-dimensional cell with time-space-geological type three-axis is constructed, where each three-dimensional cell corresponds to a unique combination of time, space, and geological type.

[0117] Each causal path in the causal main chain data is mapped to a corresponding three-dimensional cell according to its corresponding time segment, spatial unit, and geological type unit. When different causal paths correspond to different time segments within the same spatial unit and geological type unit, or correspond to different geological type units within the same time segment and spatial unit, an interleaved event record body is established in the corresponding three-dimensional cell. The interleaved event record body includes a causal path identifier, a time segment identifier, a spatial unit identifier, a geological type unit identifier, and an interleaved type identifier. When the interleaved type identifier is 1, it indicates time interleaving; when it is 2, it indicates geological type interleaving.

[0118] All interleaved event records are summarized to generate geological spatiotemporal interleaved data;

[0119] In step three, the construction of geological spatiotemporal interleaved data uses the genetic main chain data as direct input. The genetic main chain data is obtained by the phylogenetic association propagation module and contains multiple highly correlated genetic paths. Each genetic path has completed the hierarchical organization of genetic element vectors and path correlation discrimination in step two, thus possessing a clear genetic evolution sequence and spatial orientation. Based on this, each causal path is further analyzed to obtain its corresponding causal time information, spatial location information, and geological type information, transforming the causal path from a causal correlation structure into an evolutionary event that can fall within a spatiotemporal framework. By discretizing time information into time segments, mapping spatial location information into spatial units, and classifying geological type information into geological type units, a three-dimensional cell structure of time-space-geological type is constructed, thus providing a unified spatiotemporal carrying structure for multiple causal paths. When different causal paths present different time segments within the same spatial unit and geological type unit, or different geological types within the same time segment and spatial unit, the above-mentioned interlaced relationships are structurally recorded through interlaced event recorders. This allows the causal main chain relationship formed in step two to be further transformed into geological spatiotemporal interlaced data with clear spatiotemporal location and type identification in step three, providing basic data support for the construction of the geological evolution field.

[0120] In this embodiment, step four specifically includes:

[0121] Based on geological spatiotemporal interleaving data, each interleaving event record is parsed to read the corresponding time segment identifier and spatial unit identifier. The time window for data selection is limited by the time segment identifier and the spatial window for data selection is limited by the spatial unit identifier. Remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors that fall within the time window and spatial window are filtered from semantic mapping data.

[0122] Within the time and space windows, the remote sensing hyperspectral vectors are arranged according to the two-dimensional grid coordinates of the spatial units, so that each grid position corresponds to a remote sensing hyperspectral vector, generating a remote sensing hyperspectral two-dimensional spectral vector field; the geochemical spectral vectors are interpolated and mapped according to the two-dimensional projection positions of the sampling points within the spatial window, generating a geochemical two-dimensional spectral vector field; the mineral X-ray spectral vectors are arranged according to the correspondence between the spatial positions of the samples on the two-dimensional plane, generating a mineral X-ray two-dimensional spectral vector field.

[0123] The construction of the two-dimensional spectral vector field is based on a defined spatial window. Spectral vectors from different sources are processed using a consistent two-dimensional spatial alignment method. For remote sensing hyperspectral vectors, based on the two-dimensional grid coordinates of spatial units in a unified geological spatial coordinate system, the remote sensing hyperspectral vector corresponding to the center position of each spatial unit is directly filled into the corresponding two-dimensional grid position, thus forming a remote sensing hyperspectral two-dimensional spectral vector field with a regular grid structure. For geochemical spectral vectors, since the sampling points are discretely distributed within the spatial window, the spatial coordinates of each sampling point are first projected onto a two-dimensional plane coordinate system. Then, using the spatial unit grid as the target, geochemical spectral vectors falling within the same grid neighborhood are resampled. The spectral vectors of nearby sampling points are averaged by distance to calculate the spectral vector value at the corresponding grid position, thus forming a continuously distributed geochemical two-dimensional spectral vector field. For mineral X-ray spectral vectors, the two-dimensional projected coordinates of each sample are matched with the spatial grid according to the spatial source position of the mineral sample, and the corresponding mineral X-ray spectral vector is assigned to the matched grid position. If there are multiple samples in the same grid, their spectral vectors are sequentially merged and the average value is taken to represent the result. Finally, a mineral X-ray two-dimensional spectral vector field with the same spatial resolution as the remote sensing hyperspectral two-dimensional spectral vector field is obtained. The above processing ensures the correspondence between spectral vectors from different sources in two-dimensional space, providing a unified two-dimensional input structure for spectral-guided fusion transformation.

[0124] Obtain the stratigraphic interface data corresponding to the spatial window, and convert the stratigraphic interface data into a two-dimensional stratigraphic interface guide map. Different stratigraphic units in the two-dimensional stratigraphic interface guide map are distinguished by boundary lines, and the stratigraphic unit identifier is marked for each two-dimensional spatial location.

[0125] The two-dimensional stratigraphic interface guidance map is introduced as a structural guidance factor into the spectral guidance fusion transformer. In the spectral guidance fusion transformer, the feature interaction range in the two-dimensional spectral vector field is limited according to the stratigraphic unit identifier, so that the spectral vectors can only interact between two-dimensional spatial positions within the same stratigraphic unit, and feature interaction across stratigraphic interfaces is blocked.

[0126] In the spectral-guided fusion converter, the spectral vectors of the remote sensing hyperspectral two-dimensional spectral vector field, the geochemical two-dimensional spectral vector field, and the mineral X-ray two-dimensional spectral vector field located in the same stratigraphic unit and at the same two-dimensional spatial location are spliced ​​together to obtain the spectral fusion feature vector.

[0127] The spectral fusion feature vectors are aggregated to generate spectral fusion modeling data;

[0128] In the spectral-guided fusion transformer, the spectral vector fusion process uses remote sensing hyperspectral two-dimensional spectral vector fields, geochemical two-dimensional spectral vector fields, and mineral X-ray two-dimensional spectral vector fields as inputs, and uses the stratigraphic unit identifiers marked on the stratigraphic interface guidance map as structural constraints. For any two-dimensional spatial location, the fusion operation is only performed if the spectral vectors corresponding to the three types of two-dimensional spectral vector fields are located within the same stratigraphic unit, thereby avoiding the mixing of spectral features across stratigraphic interfaces. During the fusion process, the remote sensing hyperspectral vectors, geochemical spectral vectors, and mineral X-ray spectral vectors at the same two-dimensional spatial location are dimensionally aligned and spliced ​​in the feature dimension direction to generate spectral fusion feature vectors. Subsequently, the spectral fusion feature vectors from all two-dimensional spatial locations are uniformly collected to form spectral fusion modeling data. This spectral fusion modeling data maintains consistency with the original two-dimensional spectral vector fields in the spatial dimension and integrates multi-source spectral information in the feature dimension, providing highly consistent input data for subsequent three-dimensional structural modeling and geological evolution field construction.

[0129] In this embodiment, step five specifically includes:

[0130] Based on spectral fusion modeling data, on the basis of two-dimensional spatial coordinates determined by a unified geological spatial coordinate system, corresponding depth coordinates are introduced, and corresponding depth information is added to each spectral fusion feature vector to construct a three-dimensional spectral fusion data volume jointly characterized by two-dimensional spatial coordinates, depth coordinates and spectral fusion feature vectors.

[0131] The three-dimensional spectral fusion data volume is spatially discretized according to a preset voxel size, dividing the three-dimensional geological space into a regularly arranged three-dimensional voxel grid, and assigning a corresponding spectral fusion feature vector and spatial coordinate index to each voxel to form a three-dimensional voxel set.

[0132] The three-dimensional voxel set is input to the dynamic voxel clustering module, which uses voxels as the basic processing unit and uses the voxel's spectral fusion feature vector, spatial coordinates, and spatial adjacency relationship between voxels as joint inputs.

[0133] In the dynamic voxel clustering module, a voxel adjacency graph is constructed based on the spatial adjacency relationship of voxels in the three-dimensional voxel grid. Under the constraint of the voxel adjacency graph, the cosine similarity of the spectral fusion feature vectors between adjacent voxels is calculated. When the cosine similarity is greater than the preset clustering threshold, the corresponding voxels are assigned to the same candidate clustering unit.

[0134] A region-growing-based clustering algorithm is used to iteratively update the clustering labels of voxels, and to merge and adjust candidate clustering units so that the voxel clustering results simultaneously satisfy the spectral fusion feature similarity constraint and the spatial continuity constraint, until the voxel clustering results reach a stable state.

[0135] The region-growing-based clustering algorithm introduced in the dynamic voxel clustering module uses a 3D voxel mesh as the basic data structure, treating each voxel as a potential growth unit. The region growth process begins with an unlabeled voxel, which is initialized as a new growth seed, and its spectral fusion feature vector is used as the initial feature description of the current growth region. During iteration, the algorithm visits neighboring voxels at the boundary of the current growth region one by one based on the voxel adjacency graph. The similarity between the spectral fusion feature vector of the neighboring voxel and the feature statistics of the current growth region is calculated. When the similarity is greater than the preset clustering threshold of 0.85, the neighboring voxel is merged into the current growth region, and the feature statistics of the region are updated synchronously. The above growth operation is continuously performed under spatial adjacency constraints, allowing the region to expand along the direction of continuous voxels in 3D space, thereby ensuring the spatial connectivity of the clustering results.

[0136] When a growth region no longer has any voxels satisfying the similarity condition within its neighboring voxels, the growth process of that region ends, forming a stable candidate clustering unit. Subsequently, the algorithm moves on to other unlabeled voxels and repeats the above process until all voxels are assigned to their corresponding clustering units. Based on this, the boundary voxels between adjacent clustering units are checked. If the similarity between the spectral fusion feature statistics of adjacent clustering units satisfies the merging condition, the corresponding clustering unit is merged and adjusted. The stable state refers to the state where, during two consecutive iterations, the clustering labels of voxels no longer change, and there are no clustering units that satisfy the merging or splitting conditions. At this point, the voxel clustering results have reached a convergent state under the constraints of feature consistency and spatial continuity.

[0137] The stabilized voxel clustering results are used as the structural cluster partitioning results, and the voxel sets belonging to the same structural cluster are spatially combined to generate a three-dimensional construction model.

[0138] In this invention, steps one through four do not directly employ three-dimensional geological spatial representation. Instead, they are processed based on two-dimensional space within a unified geological spatial coordinate system, primarily due to considerations of data source characteristics, processing objectives, and computational consistency. Firstly, regarding data acquisition, the remote sensing hyperspectral data, geochemical spectrum data, and mineral X-ray spectrum data involved in steps one and four are all collected in planar distribution forms on the Earth's surface or samples. Their original spatial representation naturally corresponds to two-dimensional spatial coordinates. Directly extending them to a three-dimensional form would introduce unrealistic spatial assumptions and reduce the physical consistency of the data itself. Secondly, the core objective of steps two and three is to construct the correlation structure and spatiotemporal interrelationships between genetic elements. This stage focuses on the logical connections, temporal order, and spatial relative positions of genetic paths, rather than precise three-dimensional geometric shapes. Using two-dimensional spatial coordinates can effectively express the spatial orientation relationships between genetic elements, avoiding... This avoids introducing deep uncertainty before structural identification is completed. Secondly, in step four, when constructing the spectral vector field and performing spectral guidance fusion, the two-dimensional spatial expression is beneficial for aligning, fusing, and structurally constraining multi-source spectral features at a unified planar scale, enabling the stratigraphic interface guidance factor to play a constraining role in the form of a clear two-dimensional boundary. Finally, introducing the three-dimensional expression into step five helps to uniformly introduce the depth dimension for voxel modeling on the basis of the genetic correlation screening, spatiotemporal cross-limitation, and spectral feature fusion already completed in the previous steps. This ensures that the three-dimensional structural model is built on stable and convergent two-dimensional genetic and spectral information, improving the reliability and feasibility of the overall technical process.

[0139] In this embodiment, step six specifically includes:

[0140] Read the voxel set, voxel three-dimensional spatial coordinates, and voxel spectrum fusion feature vector corresponding to each structural cluster in the three-dimensional construction model;

[0141] Taking the continuous voxel region corresponding to a single structure cluster as the computational object, the voxel spectrum fusion feature vector is rearranged according to the voxel space coordinates, and the spectral fusion feature value of each dimension is normalized by the min-max normalization method. The normalized voxel spectrum fusion feature value set is constructed into a three-dimensional scalar field, and the three-dimensional scalar field is defined as the local environment field.

[0142] Taking multiple structural clusters with spatial contact relationships in a three-dimensional construction model as the calculation object, the principal component analysis method is used to calculate the principal extension direction vector of the structural clusters based on the three-dimensional spatial coordinates of all voxels within the structural clusters.

[0143] The spectral fusion feature vectors of each voxel within the structure cluster combination are weighted statistically calculated. The weight of each voxel is determined based on the Euclidean distance to the geometric center of the structure cluster combination. The voxel weight is inversely proportional to the Euclidean distance. All voxel weights are normalized so that the sum of the voxel weights is 1.

[0144] The spectral fusion feature vectors of the corresponding voxels are weighted and summed using the normalized voxel weights to obtain the weighted spectral fusion feature vectors of the structural cluster combination. These feature vectors are then concatenated with the main extension direction vectors in three-dimensional space to construct a three-dimensional vector field, which is defined as a regional construction field.

[0145] Under a unified three-dimensional spatial coordinate system, the three-dimensional scalar field in the local environmental field and the three-dimensional vector field in the regional tectonic field are aligned on a voxel-by-voxel basis. At each voxel spatial location, the corresponding scalar value and vector value are concatenated and encoded to generate a set of joint feature vectors on a voxel basis. The set of joint feature vectors is then constructed into a three-dimensional joint field, which is defined as a genetic synthesis field.

[0146] The local environmental field, regional tectonic field, and genetic composite field are organized and aggregated according to a unified voxel spatial coordinate index to generate geological evolution field data;

[0147] In this invention, the geological evolution field is constructed using structural clusters that have been stably formed in a three-dimensional structural model as the basic data units. Through further calculation and organization of voxel-level information, the discrete structural cluster results are transformed into continuous three-dimensional field data. When constructing the local environmental field, the continuous voxel region corresponding to a single structural cluster is taken as the processing object. The voxel spectral fusion feature vectors are rearranged in an orderly manner according to the coordinate position of the voxels in three-dimensional space, so that the spectral fusion features and spatial positions form a one-to-one correspondence. The dimensional differences between different spectral feature dimensions are eliminated by min-max normalization, thereby forming a continuously changing three-dimensional scalar field within the structural cluster. This local environmental field describes the spatial distribution of spectral feature intensity within the same structural cluster in a voxel-by-voxel manner, providing a basis for characterizing the fine-scale changes of the local geological environment.

[0148] In the process of constructing the regional tectonic field, multiple structural clusters with contact relationships in three-dimensional space are selected, and the principal component analysis method is used to statistically analyze the voxel spatial coordinates to obtain the main extension direction of the structural cluster combination, which is used to characterize the overall trend of the regional structure. At the same time, a weighted statistical method based on the distance from the voxel to the geometric center of the structural cluster combination is introduced, so that voxels closer to the core area of ​​the structure occupy a higher weight in the spectral feature statistics, thereby forming a three-dimensional vector field reflecting the intensity and directional characteristics of the regional structure. Finally, in the stage of constructing the genetic comprehensive field, the scalar information of the local environmental field and the vector information of the regional tectonic field are aligned and spliced ​​at the voxel level to achieve a unified expression of the local spectral feature distribution and the regional structural morphology. By organizing the above three types of fields according to a unified voxel index, a geological evolution field data with a clear structure that can be directly input into the analysis model is formed, providing a complete three-dimensional evolution information foundation for geological genetic identification.

[0149] In this embodiment, step seven specifically includes:

[0150] Geological evolution field data are encoded according to field type and spatial level to form corresponding input token sequences. The token sequences are then input into the improved CrossViT model. The token sequences corresponding to local environmental fields are used as first-level input tokens, the token sequences corresponding to regional tectonic fields are used as second-level input tokens, and the token sequences corresponding to genetic synthesis fields are used as third-level input tokens. Each token includes voxel spatial coordinate encoding and corresponding field feature vector.

[0151] When encoding the geological evolution field data into the model, the data is grouped according to field type. The local environmental field, regional tectonic field, and genetic composite field generated in step six are processed as three independent input sources. For each type of geological evolution field, a unified three-dimensional voxel spatial coordinate system is used as an index to traverse and read the voxel units within the field, obtaining the spatial coordinate information corresponding to each voxel and the field feature vector corresponding to that field type. The voxel spatial coordinates are encoded using a positional encoding method, mapping the three-dimensional coordinate values ​​to a fixed-dimensional coordinate encoding vector to represent the relative positional relationship of voxels in three-dimensional geological space.

[0152] After completing the spatial coordinate encoding, the coordinate encoding vector and the field feature vector corresponding to the voxel are concatenated along the feature dimension to form the basic token representation corresponding to a single voxel. All basic tokens within the same geological evolution field are sequentially arranged to construct the corresponding input token sequence. Among them, the token sequence generated by the local environmental field is defined as the first-level input token, which is used to characterize the fine-grained variation characteristics of local geological attributes; the token sequence generated by the regional tectonic field is defined as the second-level input token, which is used to characterize the tectonic orientation and combination characteristics at the structural cluster scale; and the token sequence generated by the genetic comprehensive field is defined as the third-level input token, which is used to characterize the comprehensive expression of multi-source genetic information in three-dimensional space.

[0153] After completing the above encoding and serialization processes, the first, second, and third level input token sequences are respectively input into the corresponding feature branches of the improved CrossViT model as the basic input data for hierarchical cross attention calculation and causal association discrimination.

[0154] The improved CrossViT model includes a first feature branch, a second feature branch, and a third feature branch. The first feature branch receives a first-level input token, the second feature branch receives a second-level input token, and the third feature branch receives a third-level input token. Each feature branch is composed of multiple layers of self-attention units connected sequentially.

[0155] Interlayer cross-attention units are set between the first feature branch and the second feature branch, and between the second feature branch and the third feature branch. The interlayer cross-attention units are connected sequentially in the order from the first feature branch to the second feature branch and from the second feature branch to the third feature branch, forming a progressive interlayer structure of cross-attention.

[0156] In the inter-layer cross-attention unit, the token from the lower-level feature branch is used as the key token and the value token, and the token from the higher-level feature branch is used as the query token. Cross-attention calculation is performed based on the query token, the key token, and the value token. Specifically, in the inter-layer cross-attention unit between the first feature branch and the second feature branch, the token of the first feature branch is used as the key token and the value token, and the token of the second feature branch is used as the query token.

[0157] In the interlayer cross-attention unit between the second feature branch and the third feature branch, the token of the second feature branch is used as the key token and the value token, and the token of the third feature branch is used as the query token. Cross-attention calculation is then performed based on the query token, the key token, and the value token.

[0158] In the process of inter-layer cross-attention calculation, a causal field-constrained attention modulation mechanism is introduced. The causal field-constrained attention modulation mechanism uses the causal comprehensive field Token in the third-level input Token as the constraint Token. Based on the feature similarity between the constraint Token and the query Token and key Token participating in the cross-attention calculation, the corresponding weight element in the cross-attention weight matrix is ​​subjected to numerical scaling operation.

[0159] When performing numerical scaling operations on the cross-attention weight matrix, the feature vector of the constraint token is used as a reference. Feature similarity is calculated between the constraint token and the feature vectors corresponding to the query token and the key token participating in the cross-attention calculation. The feature similarity is calculated using the vector cosine similarity method, that is, by summing the products of the two feature vectors dimension by dimension and dividing by the product of the magnitudes of the two feature vectors, a similarity value between 0 and 1 is obtained. Subsequently, the similarity values ​​between the constraint token and the query token and the constraint token and the key token are weighted and summed to obtain the constraint correlation coefficient corresponding to the query-key pair. The weight coefficients of the two similarity values ​​are preset and the sum of the weights is 1, preferably 0.5 and 0.5 respectively.

[0160] After obtaining the constraint correlation coefficient, a numerical scaling operation is performed on the original attention weights between the corresponding query token and key token in the cross-attention weight matrix. Specifically, the original attention weights are multiplied by the constraint correlation coefficient to obtain the constraint-modulated attention weights. When the constraint correlation coefficient is close to 1, the original attention weights are maintained or slightly enhanced; when the constraint correlation coefficient is close to 0, the original attention weights are significantly reduced, thereby reducing the information interaction intensity between tokens that do not conform to the causal constraint relationship. After completing the above element-wise multiplication scaling, the scaled attention weights are normalized so that the sum of all attention weights corresponding to the same query token is restored to 1, so as to ensure the numerical stability and comparability of the attention distribution.

[0161] At the output of the third feature branch, all tokens output by the third feature branch are aggregated and calculated to generate a cause discrimination token.

[0162] Perform classification calculations on the gene formation discrimination token to generate gene formation category identifiers corresponding to voxel spatial coordinates. Bind and store the gene formation category identifiers with the corresponding voxel spatial coordinates to generate a set of gene formation discrimination results for each voxel. Output the set of gene formation category identifiers as the geological gene formation analysis results.

[0163] When performing classification calculations on the causal discrimination tokens, the causal discrimination tokens output by the improved CrossViT model are used as classification input features. Each causal discrimination token corresponds to a unique voxel space coordinate and contains fused multi-level geological evolution field feature information. For each causal discrimination token, its feature vector is input to a fully connected classification layer. The output dimension of the fully connected classification layer is consistent with the number of predefined causal categories, used to map the continuous feature space to the discrete causal category space.

[0164] In the fully connected classification layer, matrix multiplication is performed on the feature vector of the cause discrimination token and the classification weight matrix, and a bias term is added to obtain the original classification score for each cause category. Then, the original classification score is normalized and the softmax calculation method is used to convert the score of each cause category into a probability distribution so that the sum of the probability values ​​of each cause discrimination token in all cause categories is 1. The category number corresponding to the cause category with the highest probability value is determined as the final cause category identifier of the cause discrimination token.

[0165] The improved CrossViT model used in this invention is derived from the existing CrossViT model by specifically improving it to take into account the hierarchical and spatial continuity characteristics of geological evolution field data. The existing CrossViT model mainly achieves feature interaction through a bidirectional cross-attention mechanism between tokens of different scales. Its design focuses on the collaborative modeling of local and global visual features in natural images, but it does not distinguish and model the inherent genetic hierarchical relationships and spatial constraints in multi-source geological data, which easily introduces invalid interference information across levels during feature interaction.

[0166] To address the aforementioned issues, this invention progressively reorganizes the cross-attention structure of the original CrossViT model, transforming the original parallel and symmetrical multi-scale cross-attention relationships into a hierarchical interaction structure consistent with the geological evolution field. This allows tokens at different levels to be modeled layer by layer according to the order of local environmental field, regional tectonic field, and genetic comprehensive field. Simultaneously, a genetic field-constrained attention modulation mechanism is introduced during the cross-attention calculation process. By numerically modulating the attention weights based on the consistency of the genetic field, the interaction intensity between tokens with inconsistent genetic attributes is weakened, thereby enhancing the stability of feature information transmission under the same genetic background.

[0167] Through the above structural improvements, the CrossViT model can more accurately express the hierarchical dependencies and spatial continuity characteristics in the geological genesis process, and significantly improve the stability and reliability of the genetic discrimination results in complex geological environments.

[0168] refer to Figure 2 A comprehensive geological genetic analysis system for multi-source data fusion includes the following modules:

[0169] The semantic mapping module is used to perform semantic analysis on heterogeneous geological data, form semantic substructure data, and use a multi-semantic transformation network to perform spatial domain projection to generate semantic mapping data.

[0170] The causal feature spectrum construction module is used to construct a causal feature spectrum tree based on the semantic mapping data and in accordance with causal pathology rules.

[0171] The genealogical association propagation module is used to perform correlation analysis and screening on each causal path in the causal feature genealogical tree and output the causal main chain data.

[0172] The geological spatiotemporal intersection construction module is used to construct a three-dimensional cell based on the genetic main chain data, and to establish an intersection event record body to generate geological spatiotemporal intersection data;

[0173] The spectrum-guided fusion modeling module is used to convert the remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors included in the semantic mapping data into corresponding two-dimensional spectral vector fields within the time and space windows defined by the geological spatiotemporal interleaved data. The two-dimensional spectral vector fields are then fused through the spectrum-guided fusion transformer to generate spectrum fusion modeling data.

[0174] A voxel mesh construction module is used to construct a three-dimensional voxel mesh based on the spectral fusion modeling data;

[0175] The dynamic voxel clustering module is used to perform dynamic clustering processing on the voxels in the three-dimensional voxel mesh to generate a three-dimensional construction model;

[0176] The geological evolution field construction module is used to construct local environmental fields, regional tectonic fields, and genetic integrated fields based on the three-dimensional structural model, and generate geological evolution field data.

[0177] The causal discrimination module is used to input the geological evolution field data into the improved CrossViT model for causal discrimination calculation and output the geological causal analysis results corresponding to the voxel space coordinates.

[0178] Example 2: To verify the feasibility of the present invention in practice, the present invention was applied to a multi-source geological data fusion analysis scenario, and a comprehensive geological genesis analysis experiment was carried out. The results were compared with existing methods to verify its accuracy, robustness and practical application capability.

[0179] In the experimental scenario, the study area is a typical tectonically complex region containing abundant geological information sources, such as remote sensing hyperspectral data, geochemical data, mineral X-ray diffraction data, and historical drilling records. It also has a complex geological background with multiple superimposed tectonic phases and multiple mineralization processes. Traditional methods often face problems such as scattered geological information dimensions, ambiguous determination of causal paths, and insufficient accuracy in tectonic identification during the analysis of this region, making it difficult to achieve highly reliable inference of causal mechanisms.

[0180] Applying the method of this invention, firstly, a unified spatial domain projection is performed on all heterogeneous geological data through a semantic mapping module to construct remote sensing hyperspectral vectors, geochemical spectral vectors, mineral X-ray spectral vectors, and drilling record vectors under a unified geological spatial coordinate system, generating semantic mapping data; subsequently, a genetic feature phylogenetic tree is constructed according to the genetic pathology rules, and highly correlated genetic paths are selected through a phylogenetic association propagation module to extract genetic main chain data. Based on the main chain information, a three-axis three-dimensional cell of time-space-geological type is further constructed to generate geological spatiotemporal interleaved data.

[0181] In the spectral-guided fusion modeling module, the interaction range of spectral vectors is limited according to the stratigraphic interface guidance map to effectively avoid cross-strata interference, construct spectral fusion modeling data and use it to construct a three-dimensional voxel mesh; through the dynamic voxel clustering module, regional growth clustering is performed based on the spectral fusion feature vectors and the adjacency relationship of voxel space to construct a three-dimensional structural model, further constructing the local environmental field, regional structural field and genetic comprehensive field, and inputting them into the improved CrossViT model for genetic discrimination.

[0182] Furthermore, the existing CrossViT model and the improved CrossViT model in this invention are compared in a geological genetic analysis scenario, and the results are shown in Table 1 below.

[0183] Table 1. Performance Comparison of Existing CrossViT Model and Improved CrossViT Model

[0184] Project Indicators Existing CrossViT model Improved CrossViT model Average causal category discrimination accuracy 82.3% 91.7% Mean confidence level for causal category discrimination 0.74 0.89 Error in fuzzy region of cause boundary discrimination 12.6% 5.1%

[0185] As can be seen from the data in Table 1 above, the improved CrossViT model has achieved significant performance improvement in comprehensive geological genetic analysis. In terms of average genetic classification accuracy, the improved CrossViT model achieved 91.7%, an improvement of 9.4 percentage points compared to the existing CrossViT model. This indicates that it has stronger overall recognition ability in multi-category genetic identification and can more accurately correspond to the genetic types of different geological voxels. The mean confidence score of genetic classification increased from 0.74 to 0.89, indicating that the improved CrossViT model exhibits higher stability and decision confidence in classification output, avoiding fuzzy and uncertain predictions. In addition, the error in the fuzzy zone of genetic boundaries is only 5.1% for the improved CrossViT model, significantly lower than the 12.6% of the existing CrossViT model. This indicates a significant improvement in the discrimination accuracy in genetic transition regions or tectonic boundaries, enabling clearer characterization of the boundaries between different geological units. Overall, the improved CrossViT model not only improves overall accuracy but also demonstrates stronger advantages in model stability and boundary characterization ability, making it more suitable for geological genetic analysis tasks in complex tectonic settings.

[0186] This embodiment fully demonstrates the technical advantages of this invention in multi-source data fusion and comprehensive causal analysis under complex geological backgrounds. Through a semantic mapping module, heterogeneous geological data undergoes semantic parsing and spatial domain projection to establish semantically mapped data under a unified geological spatial coordinate system, resolving the problems of inconsistent expression and misalignment of multiple data sources. Through a causal feature spectrum construction module and a spectrum association propagation module, a causal feature spectrum tree is constructed according to causal pathology rules, and correlation screening is performed on each causal path to accurately extract the causal main chain data, enhancing the scientific rigor and structure of causal reasoning. In the geological spatiotemporal intersection construction module… In this study, a three-dimensional cell structure with time, space, and geological type is constructed to effectively map the intersecting events in the genetic process, providing clear temporal and spatial constraints for analysis. Through the spectral-guided fusion modeling module and the voxel grid construction module, the high-dimensional spectral features are fused and voxelized in three-dimensional geological space. The dynamic voxel clustering module generates a three-dimensional structural model, improving the accuracy of structure identification. Finally, the geological evolution field construction module, in conjunction with the improved CrossViT model, achieves high-precision identification of voxel-by-voxel genetic categories in the genetic discrimination module, ensuring the integrity and effectiveness of the output from multi-source information fusion to specific genetic classification.

[0187] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A comprehensive geological genetic analysis method for multi-source data fusion, characterized in that, Includes the following steps: Step 1: Perform semantic analysis on heterogeneous geological data to form semantic substructure data, and use a multi-semantic transformation network to perform spatial domain projection on the semantic substructure data to generate semantic mapping data under a unified geological spatial coordinate system; Step 2: Based on semantic mapping data, construct a causal feature phylogenetic tree according to the causal pathology rules, and perform correlation analysis and screening on each causal path in the causal feature phylogenetic tree through the phylogenetic association propagation module to output the causal main chain data; Step 3: Based on the causal main chain data, construct a three-axis solid cell of time-space-geological type, and establish an intersecting event record in the three-axis solid cell to generate geological spatiotemporal intersecting data; Step 4: Based on the geological spatiotemporal interleaving data, the time and space windows are defined. The remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors included in the semantic mapping data are converted into corresponding two-dimensional spectral vector fields. The two-dimensional spectral vector fields are fused through a spectral guided fusion transformer to generate spectral fusion modeling data. Step 5: Construct a 3D voxel mesh based on the spectral fusion modeling data, and use the dynamic voxel clustering module to perform dynamic clustering to generate a 3D construction model; Step Six: Based on the three-dimensional structural model, construct the local environmental field, regional tectonic field, and genetic composite field to generate geological evolution field data; Step 7: Input the geological evolution field data into the improved CrossViT model, set up interlayer cross attention units, reconstruct the token interaction method, introduce the genetic field-constrained attention modulation mechanism, and output the geological genetic analysis results; Step four specifically involves: Based on geological spatiotemporal interleaving data, each interleaving event record is parsed to read the corresponding time segment identifier and spatial unit identifier. The time window for data selection is limited by the time segment identifier and the spatial window for data selection is limited by the spatial unit identifier. Remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors that fall within the time window and spatial window are filtered from semantic mapping data. Within the time and space windows, the remote sensing hyperspectral vectors are arranged according to the two-dimensional grid coordinates of the spatial units, so that each grid position corresponds to a remote sensing hyperspectral vector, thereby generating a remote sensing hyperspectral two-dimensional spectral vector field. The geochemical spectral vector is interpolated and mapped according to the two-dimensional projection position of the sampling point within the spatial window to generate a two-dimensional geochemical spectral vector field. The mineral X-ray spectral vectors are arranged according to the correspondence between the spatial positions of the samples on a two-dimensional plane to generate a two-dimensional mineral X-ray spectral vector field. Obtain the stratigraphic interface data corresponding to the spatial window, and convert the stratigraphic interface data into a two-dimensional stratigraphic interface guide map. Different stratigraphic units in the two-dimensional stratigraphic interface guide map are distinguished by boundary lines, and the stratigraphic unit identifier is marked for each two-dimensional spatial location. The two-dimensional stratigraphic interface guidance map is introduced as a structural guidance factor into the spectral guidance fusion transformer. In the spectral guidance fusion transformer, the feature interaction range in the two-dimensional spectral vector field is limited according to the stratigraphic unit identifier, so that the spectral vectors can only interact between two-dimensional spatial positions within the same stratigraphic unit, and feature interaction across stratigraphic interfaces is blocked. In the spectral-guided fusion converter, the spectral vectors of the remote sensing hyperspectral two-dimensional spectral vector field, the geochemical two-dimensional spectral vector field, and the mineral X-ray two-dimensional spectral vector field located in the same stratigraphic unit and at the same two-dimensional spatial location are spliced ​​together to obtain the spectral fusion feature vector. The spectral fusion feature vectors are aggregated to generate spectral fusion modeling data.

2. The comprehensive geological genetic analysis method for multi-source data fusion according to claim 1, characterized in that, The heterogeneous geological data specifically includes remote sensing hyperspectral data, geochemical spectral data, mineral X-ray spectral data, and drilling record data.

3. The comprehensive geological genetic analysis method for multi-source data fusion according to claim 1, characterized in that, Step one specifically involves: Data source identification and structural analysis are performed on heterogeneous geological data. Remote sensing hyperspectral data is divided into grids according to the spatial location of pixels. The corresponding multi-band spectral response vector is extracted for each grid unit to generate remote sensing hyperspectral semantic substructure data indexed by spatial grids. Geochemical spectrum data is mapped according to the spatial coordinates of sample collection points, and the elemental spectrum values ​​of each collection point are vectorized to generate geochemical spectrum semantic substructure data indexed by the coordinates of the sampling points. The diffraction angle position and peak intensity information in the mineral X-ray spectrum data are numerically expanded and spatially labeled in combination with the sample source position to generate mineral X-ray spectrum semantic substructure data indexed by the sample spatial position. The drilling record data is sequentially decomposed according to borehole number, depth range and corresponding lithology information to generate drilling semantic substructure data with depth coordinates; Various semantic substructure data are input into a multi-semantic transformation network, which includes four independently configured transformation subnetworks. Each transformation subnetwork includes an input encoding layer, a feature mapping layer, and an output alignment layer. In each transformation sub-network, the corresponding semantic substructure data is converted into a feature vector of a unified dimension, and spatial location encoding is introduced in the output alignment layer to map the feature vector to a unified geological spatial coordinate system. Various features projected in the spatial domain are vectorized and aggregated to generate semantic mapping data, which includes remote sensing hyperspectral vectors, geochemical spectral vectors, mineral X-ray spectral vectors, and drilling record vectors.

4. The comprehensive geological genetic analysis method for multi-source data fusion according to claim 1, characterized in that, Step two specifically involves: The semantic mapping data is decomposed into causal elements to obtain structural causal elements, material causal elements and spatial causal elements. The structural causal elements, material causal elements and spatial causal elements are then vectorized to generate structural causal element vectors, material causal element vectors and spatial causal element vectors, respectively. According to the rules of causal pathology, the causal element vectors are divided into levels: the structural causal element vectors are divided into the first level, the material causal element vectors are divided into the second level, and the spatial causal element vectors are divided into the third level. Based on the hierarchical division results, the connection methods between causal element vectors are limited. Only directed connections are allowed between the first-level causal element vector and the second-level causal element vector, and only directed connections are allowed between the second-level causal element vector and the third-level causal element vector. Using causal element vectors as the node set and directed connection relationships as the edge set, a causal feature spectrum tree is constructed. The causal feature spectrum tree is a three-level directed acyclic structure, and the causal path is formed by the first-level nodes, the second-level nodes, and the third-level nodes in sequence. The causal feature phylogenetic tree is input to the phylogenetic association propagation module. The phylogenetic association propagation module takes the causal element vectors in the node set as node inputs, performs layer-by-layer propagation calculations according to the hierarchical structure of the causal feature phylogenetic tree, and uses cosine similarity calculation to obtain correlation values ​​between causal element vectors at adjacent levels. The correlation values ​​between nodes at each level in the same causal path are sequentially accumulated to obtain the path correlation score of the corresponding causal path, which is used as the output of the genealogical association propagation module. Sort the path relevance scores of all causal paths, select the causal paths with scores higher than the preset score threshold as the causal main chain data and output them.

5. The comprehensive geological genetic analysis method for multi-source data fusion according to claim 1, characterized in that, Step three specifically involves: Each genetic path in the genetic main chain data is analyzed to extract the corresponding genetic time information, spatial location information and geological type information. The formation time information is divided into time segments according to a preset time resolution, the spatial location information is divided into spatial units according to a unified geological spatial coordinate system, and the geological type information is divided into geological type units according to geological categories. Using the time segment, spatial unit, and geological type unit as three orthogonal dimensions, a three-dimensional cell with time-space-geological type three-axis is constructed, where each three-dimensional cell corresponds to a unique combination of time, space, and geological type. Each causal path in the causal main chain data is mapped to a corresponding three-dimensional cell according to its corresponding time segment, spatial unit, and geological type unit. When different causal paths correspond to different time segments within the same spatial unit and geological type unit, or correspond to different geological type units within the same time segment and spatial unit, an interleaved event record body is established in the corresponding three-dimensional cell. The interleaved event record body includes a causal path identifier, a time segment identifier, a spatial unit identifier, a geological type unit identifier, and an interleaved type identifier. When the interleaved type identifier is 1, it indicates time interleaving; when it is 2, it indicates geological type interleaving. All intersecting event records are summarized to generate geological spatiotemporal intersecting data.

6. The comprehensive geological genetic analysis method for multi-source data fusion according to claim 1, characterized in that, Step five specifically involves: Based on spectral fusion modeling data, on the basis of two-dimensional spatial coordinates determined by a unified geological spatial coordinate system, corresponding depth coordinates are introduced, and corresponding depth information is added to each spectral fusion feature vector to construct a three-dimensional spectral fusion data volume jointly characterized by two-dimensional spatial coordinates, depth coordinates and spectral fusion feature vectors. The three-dimensional spectral fusion data volume is spatially discretized according to a preset voxel size, dividing the three-dimensional geological space into a regularly arranged three-dimensional voxel grid, and assigning a corresponding spectral fusion feature vector and spatial coordinate index to each voxel to form a three-dimensional voxel set. The three-dimensional voxel set is input to the dynamic voxel clustering module, which uses voxels as the basic processing unit and uses the voxel's spectral fusion feature vector, spatial coordinates, and spatial adjacency relationship between voxels as joint inputs. In the dynamic voxel clustering module, a voxel adjacency graph is constructed based on the spatial adjacency relationship of voxels in the three-dimensional voxel grid. Under the constraint of the voxel adjacency graph, the cosine similarity of the spectral fusion feature vectors between adjacent voxels is calculated. When the cosine similarity is greater than the preset clustering threshold, the corresponding voxels are assigned to the same candidate clustering unit. A region-growing-based clustering algorithm is used to iteratively update the clustering labels of voxels, and to merge and adjust candidate clustering units so that the voxel clustering results simultaneously satisfy the spectral fusion feature similarity constraint and the spatial continuity constraint, until the voxel clustering results reach a stable state. The stable voxel clustering results are used as the structural cluster partitioning results, and the voxel sets belonging to the same structural cluster are spatially combined to generate a three-dimensional construction model.

7. The comprehensive geological genetic analysis method for multi-source data fusion according to claim 1, characterized in that, Step six specifically involves: Read the voxel set, voxel three-dimensional spatial coordinates, and voxel spectrum fusion feature vector corresponding to each structural cluster in the three-dimensional construction model; Taking the continuous voxel region corresponding to a single structure cluster as the computational object, the voxel spectrum fusion feature vector is rearranged according to the voxel space coordinates, and the spectral fusion feature value of each dimension is normalized by the min-max normalization method. The normalized voxel spectrum fusion feature value set is constructed into a three-dimensional scalar field, and the three-dimensional scalar field is defined as the local environment field. Taking multiple structural clusters with spatial contact relationships in a three-dimensional construction model as the calculation object, the principal component analysis method is used to calculate the principal extension direction vector of the structural clusters based on the three-dimensional spatial coordinates of all voxels within the structural clusters. The spectral fusion feature vectors of each voxel within the structure cluster combination are weighted statistically calculated. The weight of each voxel is determined based on the Euclidean distance to the geometric center of the structure cluster combination. The voxel weight is inversely proportional to the Euclidean distance. All voxel weights are normalized so that the sum of the voxel weights is 1. The spectral fusion feature vectors of the corresponding voxels are weighted and summed using the normalized voxel weights to obtain the weighted spectral fusion feature vectors of the structural cluster combination. These feature vectors are then concatenated with the main extension direction vectors in three-dimensional space to construct a three-dimensional vector field, which is defined as a regional construction field. Under a unified three-dimensional spatial coordinate system, the three-dimensional scalar field in the local environmental field and the three-dimensional vector field in the regional tectonic field are aligned on a voxel-by-voxel basis. At each voxel spatial location, the corresponding scalar value and vector value are concatenated and encoded to generate a set of joint feature vectors on a voxel basis. The set of joint feature vectors is then constructed into a three-dimensional joint field, which is defined as a genetic synthesis field. The local environmental field, regional tectonic field, and genetic composite field are organized and aggregated according to a unified voxel spatial coordinate index to generate geological evolution field data.

8. The comprehensive geological genetic analysis method for multi-source data fusion according to claim 1, characterized in that, Step seven specifically involves: Geological evolution field data are encoded according to field type and spatial level to form corresponding input token sequences. The token sequences are then input into the improved CrossViT model. The token sequences corresponding to local environmental fields are used as first-level input tokens, the token sequences corresponding to regional tectonic fields are used as second-level input tokens, and the token sequences corresponding to genetic synthesis fields are used as third-level input tokens. Each token includes voxel spatial coordinate encoding and corresponding field feature vector. The improved CrossViT model includes a first feature branch, a second feature branch, and a third feature branch. The first feature branch receives a first-level input token, the second feature branch receives a second-level input token, and the third feature branch receives a third-level input token. Each feature branch is composed of multiple layers of self-attention units connected sequentially. Interlayer cross-attention units are set between the first feature branch and the second feature branch, and between the second feature branch and the third feature branch. The interlayer cross-attention units are connected sequentially in the order from the first feature branch to the second feature branch and from the second feature branch to the third feature branch, forming a progressive interlayer structure of cross-attention. In the inter-layer cross-attention unit, the token from the lower-level feature branch serves as the key token and the value token, and the token from the higher-level feature branch serves as the query token. Cross-attention calculation is then performed based on the query token, the key token, and the value token. In the process of inter-layer cross-attention calculation, a causal field-constrained attention modulation mechanism is introduced. The causal field-constrained attention modulation mechanism uses the causal comprehensive field Token in the third-level input Token as the constraint Token. Based on the feature similarity between the constraint Token and the query Token and key Token participating in the cross-attention calculation, the corresponding weight element in the cross-attention weight matrix is ​​subjected to numerical scaling operation. At the output of the third feature branch, all tokens output by the third feature branch are aggregated and calculated to generate a cause discrimination token. A classification calculation is performed on the gene formation discrimination token to generate a gene formation category identifier corresponding to the voxel spatial coordinates. The gene formation category identifier is bound and stored with the corresponding voxel spatial coordinates to generate a voxel-by-voxel gene formation discrimination result set. The gene formation category identifier set is output as the geological gene formation analysis result.

9. A comprehensive geological genetic analysis system for multi-source data fusion, comprising the comprehensive geological genetic analysis method for multi-source data fusion as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The semantic mapping module is used to perform semantic analysis on heterogeneous geological data, form semantic substructure data, and use a multi-semantic transformation network to perform spatial domain projection to generate semantic mapping data. The causal feature spectrum construction module is used to construct a causal feature spectrum tree based on the semantic mapping data and in accordance with causal pathology rules. The genealogical association propagation module is used to perform correlation analysis and screening on each causal path in the causal feature genealogical tree and output the causal main chain data. The geological spatiotemporal intersection construction module is used to construct a three-dimensional cell based on the genetic main chain data, and to establish an intersection event record body to generate geological spatiotemporal intersection data; The spectrum-guided fusion modeling module is used to convert the remote sensing hyperspectral vectors, geochemical spectral vectors and mineral X-ray spectral vectors included in the semantic mapping data into corresponding two-dimensional spectral vector fields within the time and space windows defined by the geological spatiotemporal interleaved data. The two-dimensional spectral vector fields are then fused through the spectrum-guided fusion transformer to generate spectrum fusion modeling data. A voxel mesh construction module is used to construct a three-dimensional voxel mesh based on the spectral fusion modeling data; The dynamic voxel clustering module is used to perform dynamic clustering processing on the voxels in the three-dimensional voxel mesh to generate a three-dimensional construction model; The geological evolution field construction module is used to construct local environmental fields, regional tectonic fields, and genetic integrated fields based on the three-dimensional structural model, and generate geological evolution field data. The causal discrimination module is used to input the geological evolution field data into the improved CrossViT model for causal discrimination calculation and output the geological causal analysis results corresponding to the voxel space coordinates.

Images