An intelligent identification method for detecting geological anomalies

Abstract

The present application relates to a kind of geological anomaly detection intelligent identification method, by accessing and standardizing seismic wave, remote sensing image and drilling core and other multi-source heterogeneous geological data, structured semantic unit is extracted based on field ontology library, and the computable geological semantic atlas is constructed.Further extraction and embedding local geometric invariant, fusion into multi-source semantic vector, realize space-time topology constraint coding and unified optimization.According to the geological perception encoder constructed, the new geological data collected are dynamically aligned, mapped to implicit feature manifold, drive geological anomaly intelligent identification, and output geological anomaly probability distribution and early warning signal.The scheme realizes the structured semantic expression of multi-source complex geological information, feature adaptive alignment and anomaly accurate detection, improves the intelligentization of geological analysis and space-time unified representation ability.

Description

Technical Field

[0001] This invention relates to the field of geological anomaly detection and intelligent identification technology, and in particular to an intelligent identification method for geological anomaly detection. Background Technology

[0002] Current geological anomaly detection and intelligent identification systems mainly rely on the spatiotemporal alignment and feature fusion of multi-source heterogeneous geological data. Mainstream technologies achieve data association through pre-setting geological structural blocks, unified coordinate system transformation, and timestamp-based resampling or weight allocation strategies, enabling the synchronous integration of seismic wave signals, remote sensing images, and drilling core data in spatial coordinate systems or time axes, providing standardized input for deep learning anomaly detection models.

[0003] However, existing technologies have significant limitations. Static rule-based alignment paradigms lack adaptability to the dynamic evolution of geological systems and cannot adjust alignment strategies in real time according to the causes of geological events and spatial topological changes, leading to systematic biases in feature extraction. Methods relying on explicit grid mapping and temporal resampling lose local geometric invariants, reducing the accuracy of data in depicting the true geological state. Traditional statistical models or fusion networks employ parameter weights or template matching, which easily leads to parameter redundancy, learns fixed alignment relationships, and has insufficient generalization ability for new scenarios and unseen geological structures. Summary of the Invention

[0004] This application provides an intelligent identification method for geological anomaly detection, which aims to solve one of the problems or issues of the existing technology mentioned in the background section.

[0005] This application provides an intelligent identification method for geological anomaly detection, specifically including:

[0006] S1: Acquire multi-source heterogeneous geological raw data, extract structured semantic units based on the geological ontology library, and construct a computable geological semantic map;

[0007] S2: Based on the multi-source heterogeneous geological raw data, retain the inherent spatiotemporal sampling characteristics of each data source, extract local geometric invariants, and map the local geometric invariants into low-dimensional semantic embedding vectors embedded in the geological semantic map to generate a multi-source semantic embedding vector set;

[0008] S3: Based on the multi-source semantic embedding vector set, extract the geological evolution time features based on the temporal attention branch in the pre-constructed spatiotemporal topology encoder, extract the spatial neighborhood strata continuity features based on the spatial topology constraint branch, and generate the geological topology constraint features through dual-branch coupling operation.

[0009] S4: Define a differentiable topological loss function based on the geological topological constraint features to generate an implicitly aligned geological feature manifold;

[0010] S5: Convert the newly acquired geological data to be detected into a new semantic embedding vector and input it into the spatiotemporal topology encoder. Calculate the position coordinates and time position of the new semantic embedding vector in the implicitly aligned geological feature manifold in real time, and output a dynamically aligned geological feature sequence.

[0011] S6: Perform geological anomaly detection based on the dynamically aligned geological feature sequence and generate a geological anomaly probability distribution map;

[0012] S7: Determine whether there are any abnormal areas exceeding a preset threshold in the geological anomaly probability distribution map. If so, trigger a geological early warning signal and record the anomaly spatiotemporal label. If not, update the evolution parameters of the geological semantic map.

[0013] The intelligent identification method for geological anomaly detection provided in this application has the following beneficial effects:

[0014] (1) Construct a computable geological semantic map, transforming qualitative knowledge such as tectonic types, lithological combination rules, and stress response modes into structured semantic units, and generating low-dimensional semantic embedding vectors using local geometric invariants (such as waveform phase singularities, edge curvature extrema, and bedding offset vectors). Design a spatiotemporal topology encoder to model the temporal position probability distribution and spatial topological compatibility constraints respectively, so as to achieve lossless preservation and cross-source alignment of the inherent spatiotemporal characteristics of multi-source data.

[0015] (2) A differentiable topological loss function is introduced to minimize the consistency entropy value of cross-source semantic embeddings on the evolutionary order distribution, and to simultaneously penalize feature alignment results that violate the topological constraints of stratum continuity, so that multimodal data can be automatically clustered in the latent space to the same geological event evolution trajectory. This mechanism does not require labeled sample supervised training and preset fusion rules, and still maintains strong robustness and high generalization ability in non-stationary geological environments such as faults and folds.

[0016] (3) New data only needs to have its local invariant features extracted and mapped to semantic embeddings to decode its position in the shared semantic coordinate system in real time, without the need for post-processing such as interpolation or filtering. The aligned unified representation is directly used as the input of the downstream model, eliminating the cross-modal learning burden, reducing parameter redundancy, improving classification accuracy and convergence speed, and realizing end-to-end adaptive alignment from multi-source observation to geological cognition. Attached Figure Description

[0017] Figure 1 This is the main flowchart of an intelligent identification method for geological anomaly detection;

[0018] Figure 2 This is a sub-flowchart of an intelligent identification method for geological anomaly detection;

[0019] Figure 3This is another sub-flowchart of a method for intelligent identification of geological anomalies. Detailed Implementation

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

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

[0022] like Figure 1 As shown, this application provides an intelligent identification method for geological anomaly detection, specifically including:

[0023] S1: Acquire multi-source heterogeneous geological raw data, extract structured semantic units based on the geological ontology library, and construct a computable geological semantic map;

[0024] S2: Based on the multi-source heterogeneous geological raw data, retain the inherent spatiotemporal sampling characteristics of each data source, extract local geometric invariants, and map the local geometric invariants into low-dimensional semantic embedding vectors embedded in the geological semantic map to generate a multi-source semantic embedding vector set;

[0025] S3: Based on the multi-source semantic embedding vector set, extract the geological evolution time features based on the temporal attention branch in the pre-constructed spatiotemporal topology encoder, extract the spatial neighborhood strata continuity features based on the spatial topology constraint branch, and generate the geological topology constraint features through dual-branch coupling operation.

[0026] S4: Define a differentiable topological loss function based on the geological topological constraint features to generate an implicitly aligned geological feature manifold;

[0027] S5: Convert the newly acquired geological data to be detected into a new semantic embedding vector and input it into the spatiotemporal topology encoder. Calculate the position coordinates and time position of the new semantic embedding vector in the implicitly aligned geological feature manifold in real time, and output a dynamically aligned geological feature sequence.

[0028] S6: Perform geological anomaly detection based on the dynamically aligned geological feature sequence and generate a geological anomaly probability distribution map;

[0029] S7: Determine whether there are any abnormal areas exceeding a preset threshold in the geological anomaly probability distribution map. If so, trigger a geological early warning signal and record the anomaly spatiotemporal label. If not, update the evolution parameters of the geological semantic map.

[0030] Step S1: Acquire multi-source heterogeneous geological raw data and extract structured semantic units based on a geological domain ontology library to construct a computable geological semantic map. The multi-source heterogeneous geological raw data includes seismic wave signals, remote sensing image data, and drilling core records; the structured semantic units include: structural types, lithological combination rules, and stress response patterns. The geological domain ontology library refers to a pre-constructed set of semantic rules encompassing structural geology, sedimentary petrology, and tectonic stress field theory. It is used to define geological entity types (such as normal faults, folds, interbedded sand and mudstone, etc.), entity attributes (such as dip angle, radius of curvature, sedimentary environment), and relationships between entities (such as spatial adjacency, genetic inheritance, stress coupling, and interbedded lithology) in a formal knowledge base. The geological domain ontology library is obtained by: structurally abstracting concepts, attributes, and relationships within the geological domain based on structural geology, sedimentary petrology, and tectonic stress field theory to form a set of semantic rules that can be parsed by a computer and storing it in a database.

[0031] Step S1 specifically includes: S1.1: Accessing and standardizing the format of multi-source heterogeneous geological raw data such as seismic wave signals, remote sensing image data and drilling core records to generate a standardized geological data stream with a unified interface call, which serves as the basic input object for subsequent semantic parsing.

[0032] This system integrates multi-source, heterogeneous geological raw data, including seismic wave signals, remote sensing image data, and drilling core records, identifying the physical attributes and storage format characteristics of each data source. For seismic wave signals, it performs sampling rate unification and detrending processing to eliminate baseline shifts caused by instrument drift, generating a temporally continuous seismic waveform sequence. For remote sensing image data, it performs radiometric correction and geometric fine correction to eliminate atmospheric scattering effects and sensor attitude errors, generating a raster image matrix with a unified geographic coordinate reference. For drilling core records, it performs text structured parsing and depth alignment, extracting lithological description fields and mapping them to a standard stratigraphic depth index, generating a discretized core attribute vector. A unified interface protocol for multimodal data is established, defining standardized metadata header information including timestamps, spatial coordinates, and data types. The processed seismic waveform sequence, raster image matrix, and core attribute vector are encapsulated into a standardized geological data stream object.

[0033] S1.2: Based on the standardized geological data stream, call the geological domain ontology library to perform entity mapping operations to extract structured semantic units containing structural types of normal faults and folds, lithological combination rules indicating sedimentary environments of interbedded sand and mud, and stress response patterns of high-density joints associated with compression zones, forming a set of geological prior knowledge to be aggregated.

[0034] Receive the standardized geological data stream generated by S1.1, and parse the seismic waveform phase spectrum, remote sensing image spectral reflectance, and drill core mineral composition table within it, using them as the original input objects for entity mapping.

[0035] The pre-built geological ontology library is invoked, and a semantic rule set containing structural geology, sedimentary petrology, and tectonic stress field theory is loaded to initialize the matching context of the entity recognition engine.

[0036] For seismic wave signals, spectral clustering is used to identify waveform discontinuities, which are then mapped to nodes in the ontology library that are "fault fracture zones" or "folded axial surfaces". The dip angle of normal faults and the radius of curvature of folds are extracted as structural type attribute values.

[0037] For remote sensing image data, a multi-scale edge detection operator is used to extract linear structural traces, and a vegetation cover anomaly area is calculated by combining spectral index. The stress response pattern rule of "compression zone with high-density joints" in the ontology is matched to generate a stress concentration index vector.

[0038] For drilling core records, natural language processing technology is used to analyze lithological description texts, identify sand-mud interbedded sequences, and deduce paleocurrent direction and sedimentary facies type based on the lithological combination rules of "sand-mud interbedded indices sedimentary environment" in the ontology database, thus forming lithological semantic tags.

[0039] Establish a cross-modal semantic association matrix, calculate the co-occurrence probability between structural type nodes, stress response mode nodes and lithological semantic labels, and screen semantic triples with confidence scores higher than a preset threshold.

[0040] The filtered semantic triples are aggregated into structured units, each containing an entity ID, semantic category, physical attribute parameters, and source data identifier, forming a set of geological prior knowledge to be aggregated.

[0041] S1.3: Utilize the aforementioned set of geological prior knowledge to perform node definition and relation edge construction processing, in order to establish a geological semantic graph skeleton that describes the evolutionary logic and spatial topological associations between geological entities, and output an unweighted geological semantic network with a preliminary topological structure.

[0042] Receive the set of geological prior knowledge generated by S1.2, which contains structured semantic units of tectonic types, lithological combination rules and stress response modes.

[0043] The structured semantic units are assigned entity unique identifiers, and a globally unique URI identifier is generated for each structural type node, lithological rule node, and stress mode node, thus establishing the atomic basis of the graph nodes.

[0044] Based on the principles of geological evolution dynamics, semantic relationship types between nodes are defined, including four basic topological edges: "spatial adjacency", "genetic inheritance", "stress coupling" and "lithological interlayering", and a relation ontology schema is constructed.

[0045] By traversing the set of prior geological knowledge, natural language processing technology is used to analyze the implicit logical relationships between semantic units, transforming qualitative descriptions into specific edge connection instructions, such as establishing a "stress coupling" edge connection between "normal fault" and "tension stress zone".

[0046] Based on the defined topological edge type, the adjacency matrix structure is initialized, and all identified node pairs are mapped to the corresponding positions in the matrix. The initial weights are set to unit values ​​to form an unweighted sparse adjacency representation.

[0047] A connectivity test is performed on the generated preliminary topology to remove isolated nodes and invalid branches that cannot form closed loops, ensuring the integrity and accessibility of the map skeleton in terms of geological logic.

[0048] S1.4: Perform differentiable encoding on the node attributes and edge relationships in the unweighted geological semantic network to transform qualitative geological evolution laws into quantitative topological constraint parameters that can be recognized by the deep learning framework, and generate a computable geological semantic map containing a numerical weight matrix.

[0049] Receive an unweighted geological semantic network, which includes structural type nodes, lithological combination rule nodes, and stress response mode nodes, as well as initial edge relationships describing the evolution logic between entities.

[0050] Vectorization encoding is performed on the node attributes in the network, and a pre-trained geological domain word embedding model is used to map qualitative text descriptions into high-dimensional dense vectors, giving each node initial semantic representation capabilities.

[0051] Based on the laws of geophysics, the edge weights are defined as differentiable calculation rules. For the topological association between adjacent nodes, a geological compatibility coefficient is introduced as the weight initialization benchmark to ensure that the edge weights reflect the true geological constraint strength.

[0052] A differentiable graph attention mechanism layer is constructed to dynamically update edge weights. The attention coefficient between node i and node j is calculated using the following formula:

[0053]

[0054] in, Let be the attention weight of node i to node j, LeakyReLU be the non-linear activation function, and W be the learnable weight matrix. and Let i and j be the feature vectors of nodes i and j, respectively. Let i be the set of neighbors of node i.

[0055] By using the calculated attention coefficients to weight and aggregate node features, an updated node representation containing neighborhood topological information is generated, thereby achieving information fusion of local geological context.

[0056] The updated node representation is input into the multilayer perceptron module, which outputs a numerical edge weight matrix. This matrix quantifies the strength of the evolutionary logic between geological entities, forming computable quantitative topological constraint parameters.

[0057] The generated weight matrix is ​​sparsified to remove weak connections below a preset threshold, retain strong geological correlation paths, and optimize the computational efficiency and structural clarity of the map.

[0058] By using a differentiable coding process, the unweighted geological semantic network from the previous step is transformed into a computable geological semantic map containing a numerical weight matrix. This enables the transformation of qualitative geological laws into quantitative parameters that can be identified by a deep learning framework, providing a physically consistent topological constraint basis for subsequent local geometric invariant mapping.

[0059] For example, an unweighted network containing 50 structural types such as normal faults and folds, and 30 lithological regularities such as interbedded sand and mud layers, is used to generate 128-dimensional node initial vectors using the Word2Vec model. The geological compatibility coefficient is initially set to 0.5, and a two-layer graph attention network is constructed. When calculating the attention coefficients, W is set as a 128x256-dimensional matrix, activated by LeakyReLU, and then normalized using Softmax to obtain weight values ​​between 0 and 1. A sparsity threshold of 0.1 is set on the weight matrix; after removing weak connections, the remaining strong connections account for approximately 40%. In the final generated computable geological semantic map, the weights between nodes accurately reflect the physical correlation strength between fault and joint development, significantly improving the robustness of subsequent feature alignment.

[0060] S1.5: Perform global consistency verification and index optimization based on the computable geological semantic map to eliminate semantic conflicts and ensure the logical self-consistency of the map in the spatiotemporal dimension, and output the final version of the computable geological semantic map to drive subsequent local geometric invariant mapping.

[0061] It receives the computable geological semantic map containing a numerical weight matrix generated by S1.4 as the input object for global consistency verification and index optimization.

[0062] The transitive closure is calculated for the semantic association strength of all node pairs in the graph. The adjacency matrix is ​​updated iteratively using matrix multiplication. Weakly connected edges with logical contradictions are detected and marked, and a set of potential conflict edges is generated.

[0063] For potential conflict edge sets, a causal constraint mechanism based on geological evolution time sequence is introduced. By comparing the timestamp attributes between nodes with the order of tectonic evolution, reverse connections that violate causality are eliminated, and a preliminarily purified acyclic semantic subgraph is output.

[0064] The optimized embedding vector of node i is calculated by Laplacian smoothing using the average aggregation method on the node embedding vectors in the acyclic semantic subgraph, according to the following formula:

[0065]

[0066] in, Let be the embedding vector of node i. Let N(i) be the degree of node i, and N(i) be the set of neighbors of node i. Let (i,j) be the weight of edge (i,j). Let be the embedding vector of neighbor node j.

[0067] Based on the optimized node embedding vectors, a multi-dimensional spatial index structure is constructed. The KD-Tree algorithm is used to partition the high-dimensional semantic space and establish a fast mapping index table from semantic features to graph nodes.

[0068] Redundancy pruning is performed on the index table, similar nodes with a spatial distance less than a preset threshold are merged, map fragmentation caused by data noise is eliminated, and a final version of the computationally comprehensible geological semantic map with a compact topology is generated.

[0069] Through the above-mentioned global consistency verification and index optimization, the preliminary map generated in the previous step is transformed into a final computable geological semantic map with logical self-consistency and efficient retrieval capabilities, achieving the expected technical effect of eliminating semantic conflicts and ensuring the logical consistency of the map in the spatiotemporal dimensions.

[0070] Step S2: Based on the multi-source heterogeneous geological raw data, retain the inherent spatiotemporal sampling characteristics of each data source to extract local geometric invariants, and map these local geometric invariants into low-dimensional semantic embedding vectors embedded in the geological semantic map, generating a multi-source semantic embedding vector set. Specifically, this includes:

[0071] S2.1: Perform modal-specific analysis on multi-source heterogeneous geological raw data such as seismic wave signals, remote sensing image data and drilling core records, construct a non-resampled data preservation strategy based on the inherent spatiotemporal sampling characteristics of each data source, and generate a set of modal independent data streams that retain the original spatiotemporal resolution.

[0072] The S2.1 sub-step aims to perform modality-specific analysis on multi-source heterogeneous geological raw data, construct a non-resampled data preservation strategy, and generate a set of modality-independent data streams that retain the original spatiotemporal resolution.

[0073] It receives seismic wave signals, remote sensing image data and drilling core records that have been normalized by S1, identifies the inherent sampling frequency and spatial grid dimension of each data source, and establishes a modal metadata index table.

[0074] For seismic wave signals, time series integrity checks are performed, missing sampling points are marked, and linear interpolation is used to fill in tiny gaps while maintaining the original millisecond-level timestamp accuracy, generating a continuous seismic waveform data stream.

[0075] For remote sensing image data, the geographic reference coordinate system and pixel resolution are analyzed, radiometric correction is performed to eliminate the effects of atmospheric scattering, the original spectral band information is preserved, and a high-fidelity remote sensing image data stream is generated.

[0076] For drilling core records, depth intervals and lithological description text are extracted, a depth-attribute mapping relationship is established, the original centimeter-level depth sampling interval is preserved, and a structured core attribute data stream is generated.

[0077] Define a non-resampled data preservation strategy to prohibit uniform gridding or timestamp alignment operations on each modality of data, ensuring that each data stream maintains its spatiotemporal independence at the time of physical acquisition during subsequent processing.

[0078] Each modal data stream is assigned a unique identifier and a spatiotemporal reference label, encapsulated as an independent data object, forming a set of modal-independent data streams, which serve as the basic input for extracting local geometric invariants. The non-resampling prohibits cross-modal uniform meshing or timestamp alignment operations on each modal data, but allows data integrity restoration within a modality (such as linear interpolation to fill small gaps).

[0079] S2.2: Based on the modal independent data stream set, perform local geometric feature extraction operation, perform phase singularity point detection on the processed seismic wave signal, obtain waveform phase singularity point sequence, obtain image edge curvature extreme point set based on the remote sensing image data, process the drilling core record using the bedding offset vector calculation method to obtain core bedding offset vector set, and generate an original feature element library containing waveform phase singularity point sequence, image edge curvature extreme point set and core bedding offset vector set.

[0080] The system receives a set of modal-independent data streams, performs a Hilbert transform on the seismic wave signals to construct an analytic signal, and extracts the instantaneous phase components. It then uses a phase gradient operator to calculate the spatial derivative of the phase field, identifies phase abrupt change points as waveform phase singularities, and generates a sequence of phase singularities.

[0081] Multi-scale edge detection is performed on remote sensing image data using the difference of Gaussian operator to extract pixel-level edge contours. The local curvature values ​​of the edge curves are calculated using the curvature formula in differential geometry, and extreme points with absolute curvature values ​​exceeding a threshold are selected to form a set of extreme curvature points for the image edge.

[0082] Automatic bedding plane identification is performed on drill core records, and bedding normal vectors are extracted. The angle and displacement deviation of normal vectors within adjacent depth intervals are calculated to construct a bedding offset vector set describing the characteristics of formation faulting.

[0083] The geometric invariants of the three modalities mentioned above are uniformly encapsulated into a raw feature library. Through local geometric invariant extraction processing, the modal-independent data streams from the previous step are transformed into structured feature data containing phase singularities, curvature extrema, and bedding offset vectors, realizing standardized representation of multi-source data at the geometric topology level and providing an invariant basis for subsequent semantic mapping.

[0084] S2.3: Based on the original feature element library and the computable geological semantic map constructed in the previous steps, perform semantic association mapping. Using the graph neural network embedding mechanism, the waveform phase singular point sequence, the image edge curvature extreme point set and the core bedding offset vector group are mapped to the structural type node, lithological combination rule node and stress response mode node in the geological semantic map, respectively, to generate an initial semantic association vector group with geological ontology constraints.

[0085] The system receives the original feature library generated in the previous steps, which includes waveform phase singular point sequences, image edge curvature extreme point sets, and core bedding offset vector groups, and loads the constructed computable geological semantic map as a mapping benchmark.

[0086] A semantic association mapping network based on graph attention mechanism is constructed. The multimodal geometric invariants in the original feature element library are defined as query vectors, and the structural type nodes, lithological combination rule nodes, and stress response mode nodes in the geological semantic map are defined as key vectors and value vectors.

[0087] The cosine similarity between the query vector and the key vector is calculated to generate an initial attention weight matrix, which quantifies the potential association strength between each local geometric feature and a specific geological semantic unit.

[0088] A geological ontology constraint mask is introduced to sparsify the initial attention weight matrix, eliminating non-physical connections that violate the logic of geological evolution, and ensuring that the mapping process conforms to prior knowledge such as fault-associated joints and fold-controlled sedimentation.

[0089] By employing a weighted summation aggregation strategy, the attention weights processed by masking are applied to the value vector to generate a preliminary semantic embedding representation for each data source feature, thus achieving a preliminary projection from the geometric space to the semantic space.

[0090] A semantic alignment loss function is defined. By minimizing the distribution distance of semantic embedding vectors from different data sources under the same geological event, the parameters of the mapping network are optimized, enhancing the consistency of cross-modal features in the semantic space. The specific formula is as follows:

[0091]

[0092] Where L is the semantic alignment loss value, and N and M are the number of features from different data sources, respectively. and Let represent the embedding vectors of the i-th and j-th features in the semantic space, respectively, and ||·|| denote the Euclidean norm.

[0093] The weight parameters of the graph attention network are updated using the backpropagation algorithm, and the optimization is iteratively performed until the loss function converges, generating an initial semantic association vector set with geological ontology constraints.

[0094] S2.4: Based on the initial semantic association vector set, perform dimensionality reduction and manifold projection processing. Using manifold learning, the high-dimensional initial semantic association vector set is compressed into a low-dimensional latent space while maintaining the local geometric invariant topological structure, generating a low-dimensional semantic embedding vector set that conforms to the geological evolution logic.

[0095] The system receives an initial semantic association vector set with geological ontology constraints from the output of S2.3. This vector set contains high-dimensional sparse features and exhibits dimensionality differences between modes. A manifold learning dimensionality reduction model based on Laplacian eigenmaps is constructed to compress data dimensionality while preserving local geometric topology. The similarity matrix of the initial semantic association vector set is calculated, and a Gaussian kernel function is used to measure the Euclidean distance between any two vectors, generating a weight matrix reflecting the geological semantic proximity. A graph Laplacian operator is constructed based on the weight matrix, and low-dimensional embedding coordinates are extracted by solving the generalized eigenvalue problem. The target dimensionality reduction dimension is set to the number of principal components of geological evolution, typically selecting the eigenvectors corresponding to the k smallest non-zero eigenvalues. The extracted eigenvectors are orthogonally normalized to eliminate the influence of dimensions and enhance numerical stability. A geological topology consistency regularization term is introduced to constrain the degree of manifold distortion during the dimensionality reduction process, ensuring that the relative positions of seismic wave phase singularities and remote sensing edge curvature remain unchanged in low-dimensional space. The low-dimensional embedding parameters are updated until convergence by iterative optimization to minimize the sum of the reconstruction error and the regularization term. The output is a low-dimensional semantic embedding vector set that conforms to the logic of geological evolution. Its dimension is significantly lower than that of the input vector and it retains key topological information.

[0096] S2.5: Perform multi-source data fusion encapsulation based on the low-dimensional semantic embedding vector set, and restructure the low-dimensional semantic embedding vectors according to the data source identifier and spatiotemporal label to generate a multi-source semantic embedding vector set for driving the geological sensing spatiotemporal topology encoder.

[0097] like Figure 2 As shown, step S3: Based on the multi-source semantic embedding vector set, the geological evolution time features are extracted based on the temporal attention branch in the pre-constructed spatiotemporal topology encoder, and the spatial neighborhood strata continuity features are extracted based on the spatial topology constraint branch. Geological topology constraint features are then generated through a dual-branch coupling operation. According to an embodiment of the present invention, the spatiotemporal topology encoder adopts a dual-branch parallel architecture, consisting of a temporal attention branch, a spatial topology constraint branch, and a dual-branch feature coupling module. During the training phase (step S3), the spatiotemporal topology encoder receives the multi-source semantic embedding vector set and outputs the geological topology constraint features; during the inference phase (step S5), it receives new semantic embedding vectors to be mapped and outputs temporal evolution state representations and spatial topological compatibility representations for subsequent manifold projection calculations.

[0098] Specifically, the temporal attention branch is a relative order probability distribution branch, composed of a multi-layer temporal attention network. This temporal attention branch takes the input semantic embedding vector set as a query term, calculates a similarity score with the geological historical periodization, normalizes it using Softmax, and smooths it with a Gaussian kernel to output a relative order probability distribution of geological evolution. In the training phase (S3.1), this relative order probability distribution of geological evolution is used to subsequently couple and generate geological topological constraint features; in the inference phase (S5.2), this relative order probability distribution of geological evolution is encoded as a temporal evolution state representation vector, quantifying the relative position and uncertainty of newly acquired data in the geological time dimension.

[0099] The spatial topological constraint branch is a stratum continuity constraint matrix generation module, which embeds a graph convolutional network and connects to a known regional tectonic ore-controlling model parameter library. This graph convolutional network consists of two consecutive hidden layers and one output layer. During the training phase (S3.2), this branch generates a spatial neighborhood stratum continuity constraint matrix based on the probability distribution output by the temporal branch and the tectonic ore-controlling model. During the inference phase (S5.3), this branch uses the temporal evolution state representation as semantic guidance input, retrieves stratigraphic deposition and tectonic deformation rules corresponding to the geological history stages, extracts the lithological similarity and occurrence consistency of adjacent grid cells using the graph convolutional network, constructs a sparse spatial adjacency matrix, and outputs a spatial topological compatibility representation after normalization.

[0100] The dual-branch feature coupling module consists of a tensor product operation unit and a normalized output layer, and is activated only during the training phase (S3.3). Its tensor product operation unit uses the Kronecker product operator to perform an outer product operation on the temporal dimension feature vector and the spatial dimension adjacent feature tensor to construct a spatiotemporal joint feature hypercube. The normalized output layer outputs standardized geological topological constraint features after projection dimensionality reduction, ReLU activation and normalization, which are used for subsequent construction of differentiable topological loss functions and implicit manifold optimization.

[0101] It should be noted that the spatiotemporal topology encoder uses the same network architecture in the training phase (step S3) and the inference phase (step S5) but executes different forward computation logic—the training phase outputs geological topological constraint features to construct the loss function, and the inference phase outputs temporal state representation and spatial topological representation for manifold projection.

[0102] S3 specifically includes: S3.1: Based on the multi-source semantic embedding vector set, the geological evolution time axis mapping process is performed through the temporal attention mechanism in the spatiotemporal topology encoder to calculate the relative position weight of each semantic embedding vector in the geological history period and generate the relative order probability distribution of geological evolution.

[0103] The system receives a multi-source semantic embedding vector set generated in the preceding steps. This vector set contains low-dimensional geological semantic features from seismic waves, remote sensing, and drilling data. A time-series mapping module based on a multi-head self-attention mechanism is constructed to divide the geological history into several discrete time slices, which serve as a reference system for relative order calculation.

[0104] A linear projection transformation is performed on the input multi-source semantic embedding vector to generate a query vector, a key vector, and a value vector, which are used to capture the dynamic correlation features of different data sources during geological evolution. The dot product similarity between the query vector and the key vector is calculated and divided by a scaling factor to stabilize the gradient distribution, resulting in the original attention score matrix.

[0105] A geological evolution prior mask is introduced to constrain the original attention score matrix, shielding non-causal temporal connections that violate stratigraphic sedimentary laws and ensuring the physical rationality of temporal logic. The masked attention scores are then normalized using the Softmax function to generate probability distribution weights for each semantic embedding vector across different geological time slices.

[0106] A weighted summation method is used to multiply the normalized probability weights by the encoding vectors of the corresponding time slices and accumulate them to calculate the continuous temporal position representation of each semantic embedding vector. A fully connected layer is used to perform a nonlinear mapping on the continuous temporal position representation to output the final geological evolution relative order probability distribution.

[0107] S3.2: Based on the known regional tectonic ore-controlling model and the relative order probability distribution of geological evolution, perform spatial neighborhood topology analysis, use graph convolutional networks to extract the stratum occurrence correlation features between adjacent grid cells, and generate a spatial neighborhood stratum continuity constraint matrix.

[0108] According to one embodiment of the present invention, the known regional tectonic ore-controlling model is a pre-constructed geological prior knowledge parameter library, consisting of a tectonic geometry parameter module, a stratigraphic sequence module, and a tectonic deformation rule module. This module provides a known geological tectonic framework and sedimentary evolution constraints for spatial neighborhood topological analysis of the study area. Specifically, the tectonic geometry parameter module stores the spatial distribution coordinates of known fault traces, fault attitude elements, and geometric parameters of fold axis traces within the study area, used to construct undirected graph adjacency relationships of geological grid units and define spatial connection weights. The stratigraphic sequence module stores the sedimentary thickness, lithological assemblage, sedimentary facies type, and stratigraphic contact relationships of strata at each geological historical stage, used to constrain the reasonable range of lithological similarity and attitude consistency between adjacent grid units. The tectonic deformation rule module stores the regional tectonic stress field evolution sequence and the tectonic patterns of deformation events at each stage, used to provide prior discrimination criteria for tectonic deformation types in the calculation of topological compatibility weights, ensuring that the construction of the spatial adjacency matrix conforms to the regional tectonic evolution law. The model is invoked by the spatial topology constraint branch during the training phase (S3.2) and the inference phase (S5.3), providing geophysical rule constraints and parameter initialization benchmarks for the spatial feature aggregation of graph convolutional networks.

[0109] The system receives the relative order probability distribution of geological evolution and the known tectonic ore-controlling model of the region generated in the previous steps as the initial input conditions for spatial topology analysis. Based on the geometric parameters of fault traces and fold axes in the regional tectonic ore-controlling model, an undirected graph structure reflecting the adjacency relationships of geological grid units in the study area is constructed. Nodes are defined as discretized geological grid units, and edges represent the spatial connections between adjacent units. A graph convolutional network is used to aggregate spatial features of the multi-source semantic embedding vector set. A message passing mechanism is used to capture the stratigraphic attitude correlation features between adjacent grid units, extracting high-dimensional spatial topological feature vectors containing strike, dip, and curvature variations. The graph Laplacian operator is used to calculate the smoothness constraints of feature vectors between adjacent nodes, quantifying the continuity deviation of stratigraphic distribution in the spatial neighborhood, and generating a local constraint matrix characterizing the physical continuity of the stratigraphy. Combined with the relative order probability distribution of geological evolution, the local constraint matrix is ​​temporally weighted to eliminate discontinuous noise interference caused by sedimentary discontinuities or tectonic abrupt changes, and to strengthen the spatial topological connection weights that conform to the logic of geological evolution. The weighted constraint matrix is ​​subjected to symmetry and normalization operations to ensure the stability and physical interpretability of the matrix eigenvalues, and the final spatial neighborhood strata continuity constraint matrix is ​​generated.

[0110] By using message passing and graph Laplacian regularization in graph convolutional networks, multi-source semantic embedding vectors are transformed into stratum continuity constraint matrices with spatial topological consistency, thereby achieving a quantitative characterization of the spatial continuity of complex geological structures and providing a precise spatial constraint basis for subsequent bi-branch feature coupling.

[0111] For example, for a coal-bearing basin exploration area, an undirected graph containing 500 grid cells is constructed, with each node corresponding to a 100m × 100m geological unit. A three-layer graph convolutional network is used. The three-layer graph convolutional network includes two consecutive hidden layers (first convolutional layer and second convolutional layer) and one output layer (third convolutional layer). The first layer graph convolution uses the directly adjacent grid cells in the undirected graph as the message passing domain, and performs a weighted summation of the multi-source semantic embedding vector of the current node (grid cell) and the multi-source semantic embedding vectors of the adjacent nodes to obtain a first-order spatial aggregation feature; then, after linear mapping and ReLU nonlinear activation, the first-layer node embedding is output (dimension set to 128→64).

[0112] The second layer graph convolution takes the first layer node embedding output from the previous layer as input and is still computed on the same undirected graph structure. However, the receptive field is expanded through the second propagation of the adjacency matrix, so that the current node can simultaneously perceive the continuation trend of the more distant co-phase strata and the truncation effect brought by the fault. After linear mapping and ReLU, the second layer node embedding is output.

[0113] The third-layer graph convolution takes the second-layer node embeddings as input and introduces the graph Laplacian operator (L=DA, where D is the degree matrix and A is the adjacency matrix) as a regularization guide. A smoothing penalty term, "difference from the neighborhood mean," is superimposed on the message-weighted summation term, ensuring that the output third-layer node embeddings (64-dimensional) retain edges at structural abrupt changes while suppressing noise jitter within continuous sedimentary zones. The third-layer output is the 64-dimensional spatial topological feature vector corresponding to each grid cell. After obtaining the 64-dimensional spatial topological feature vectors of all nodes, the sum of squared differences between the 64-dimensional spatial topological feature vectors of adjacent nodes is minimized to generate a 500×500 spatial neighborhood stratum continuity constraint matrix.

[0114] S3.3: Based on the relative order probability distribution of geological evolution and the spatial neighborhood strata continuity constraint matrix, a two-branch feature coupling operation is performed, and the temporal order information and spatial topological constraints are fused using tensor product operation to generate geological topological constraint features.

[0115] The relative order probability distribution of geological evolution and the spatial neighborhood strata continuity constraint matrix are received as dual-branch input data.

[0116] Tensor reshaping is performed on the relative order probability distribution to convert it into a time-dimensional feature vector representation to meet the needs of subsequent high-dimensional fusion operations.

[0117] Sparse encoding is performed on the spatial neighborhood stratum continuity constraint matrix to extract the topological connection weights between adjacent grid cells and generate a spatial adjacency feature tensor.

[0118] We construct a spatiotemporal joint feature hypercube by performing an outer product operation on the temporal dimension feature vector and the spatial dimension adjacent feature tensor using the Kronecker product operator.

[0119] The spatiotemporal coupled feature tensor is calculated using the following formula:

[0120]

[0121] Where T represents the geological topological constraint feature. This represents the relative order probability distribution vector of geological evolution. This is the spatial neighborhood strata continuity constraint matrix. This represents the Kronecker product operator.

[0122] Multidimensional projection dimensionality reduction is performed on the generated spatiotemporal joint feature hypercube, and a fully connected layer is used to map the high-dimensional coupled features to the low-dimensional latent space.

[0123] Nonlinear activation transformation is performed in the latent space, and the ReLU activation function is introduced to filter negative topological conflict signals while retaining positive constraint features that conform to the logic of geological evolution.

[0124] Normalization is performed on the activated feature vectors to eliminate numerical biases caused by different dimensions, forming standardized geological topological constraint features.

[0125] like Figure 3 As shown, step S4: Define a differentiable topological loss function based on the geological topological constraint features to generate an implicitly aligned geological feature manifold. Specifically, this includes:

[0126] S4.1: The relative order probability distribution of geological evolution and the multi-source semantic embedding vector are jointly encoded to extract the distribution dispersion of each data source on the geological evolution timeline, generating a cross-source evolution order consistency entropy value as a basic quantitative indicator to measure the degree of temporal logical conflict of multi-source data. In this embodiment, the Shannon information entropy formula is used to calculate the distribution dispersion of each data source on the geological evolution timeline, and the distribution dispersion is mapped to the cross-source evolution order consistency entropy value. In other embodiments, statistical indicators such as variance or relative entropy (KL divergence) can also be used to quantify the degree of distribution dispersion of data sources on the timeline.

[0127] Specifically, it receives the relative order probability distribution of geological evolution and the multi-source semantic embedding vectors in the multi-source semantic embedding vector set as the basic input data for temporal consistency quantization.

[0128] The probability density reconstruction process is performed on the relative order probability distribution, and the discrete time axis segment weights are mapped to continuous probability density functions to ensure that the distributions of different data sources are comparable in the time dimension.

[0129] Based on the reconstructed probability density function, the Shannon information entropy metric is used to calculate the uncertainty measure of each data source on the geological evolution time axis, quantifying the degree of time distribution dispersion within a single data source.

[0130] A cross-source joint probability distribution matrix is ​​constructed, and the relative order probabilities of seismic wave, remote sensing and drilling data are fused by tensor outer product operation to characterize the joint occurrence probability of multi-source data on a unified time axis.

[0131] The cross-source mutual information entropy is calculated using the joint probability distribution matrix to measure the statistical dependence and logical conflict level between different geological data sources in the evolutionary time series.

[0132] The independent entropy values ​​of each data source are weighted and differentially calculated with the cross-source mutual information entropy to generate the cross-source evolutionary order consistency entropy value. This indicator directly reflects the logical deviation of multi-source data in the geological time series alignment process.

[0133] S4.2: Based on the spatial neighborhood strata continuity constraint matrix and the multi-source semantic embedding vector, perform topological compatibility discrimination operation, use the graph Laplacian regularization method to calculate the gradient difference of stratigraphic distribution characteristics between adjacent grid cells, and generate a spatial topological violation penalty term to characterize the feature alignment deviation that violates the continuity of geological structures.

[0134] By receiving the spatial neighborhood strata continuity constraint matrix and multi-source semantic embedding vectors, a geological topological compatibility discrimination model is constructed.

[0135] Based on the spatial adjacency relationship defined by the regional tectonic ore-controlling model, the multi-source semantic embedding vector is mapped to graph structure node features, where nodes correspond to geological grid units and edge weights are determined by the stratum continuity constraint matrix.

[0136] The smoothness of the stratigraphic distribution characteristics between adjacent grid cells is calculated using the graph Laplacian operator, and the spatial topology violation penalty term is calculated using the following formula:

[0137]

[0138] in, This is a penalty item for spatial topology violations. The connection weight between the i-th and j-th grid cells in the rock stratum continuity constraint matrix is... and These are the multi-source semantic embedding vectors of the i-th and j-th grid cells, respectively.

[0139] Gradient analysis is performed on the graph Laplacian regularization results to calculate the rate of change of local curvature of the embedding vector in the manifold space, and to identify anomalous alignment deviations that violate the continuity of geological structures.

[0140] Regions where the rate of change of local curvature exceeds a preset threshold are marked as topological conflict zones, and spatial topological violation penalty vectors are generated to quantify structural errors in the feature alignment process.

[0141] S4.3: The cross-source evolutionary order consistency entropy value and the spatial topology violation penalty term are weighted and fused, and an adaptive balance coefficient is introduced to adjust the contribution weights of temporal consistency and spatial continuity in order to construct a differentiable topology loss function and form a unified optimization objective that can simultaneously constrain temporal evolution logic and spatial topology.

[0142] The cross-source evolutionary order consistency entropy value and spatial topology violation penalty term generated in the preceding steps are received as the basic input data for constructing the composite loss function.

[0143] Dynamic gradient monitoring is performed on the cross-source evolutionary sequence consistency entropy value, and its rate of change in the current training batch is calculated to quantify the convergence difficulty and optimization urgency of the temporal alignment task.

[0144] The adaptive balance coefficients are initialized based on the rate of change. The variance of the loss components of the most recent N iteration steps is recorded using a sliding window statistical method to evaluate the difference in the numerical magnitude between the two constraints of temporal consistency and spatial continuity.

[0145] By using the Sigmoid activation function to map the normalized variance ratio to the interval between 0 and 1, a dynamic balance factor is generated to adjust the contribution weights of the two losses, ensuring that global temporal alignment is emphasized in the early stage of training and local spatial topological constraints are strengthened in the later stage.

[0146] The dynamic balance factor is multiplied by the cross-source evolutionary order consistency entropy and the spatial topology violation penalty term, respectively, to achieve a weighted fusion of the two heterogeneous loss indicators and eliminate the optimization dominance bias caused by different dimensions.

[0147] Construct the following formula for the differentiable topological loss function:

[0148]

[0149] Where L is the differentiable topological loss function, and α is the adaptive balance coefficient. The entropy value represents the consistency of cross-source evolutionary order. This is a penalty item for violations of spatial topology.

[0150] Automatic differentiation is performed on the composite loss function to calculate its partial derivatives with respect to the coordinates of each vector in the multi-source semantic embedding vector set, forming a gradient field pointing to the convergence direction of the geological event evolution trajectory.

[0151] The gradient field is transmitted to the spatiotemporal topology encoder through the backpropagation mechanism, and the latent space coordinate parameters of the embedded vector are updated, so that the multi-source data spontaneously satisfy the dual constraints of temporal evolution logic and spatial topology in the feature space.

[0152] By using a weighted fusion processing method, the discrete loss index of the previous step is transformed into a unified and optimized differentiable topological loss function, thereby achieving the expected technical effect of high-precision spatiotemporal alignment of multi-source heterogeneous geological data on implicit manifolds.

[0153] S4.4: The backpropagation optimization iteration is performed on the latent space coordinates of the multi-source semantic embedding vector set using the differentiable topological loss function. The loss function value is minimized and the embedding vector parameters are updated by the gradient descent algorithm, so as to drive the data from different sources to spontaneously cluster in the feature space to the same geological event evolution trajectory, and generate a preliminary convergent implicitly aligned geological feature manifold.

[0154] The latent space coordinates refer to the positional parameters of each embedding vector in the multi-source semantic embedding vector set within the low-dimensional latent space. These coordinates characterize the distribution of different geological data sources (seismic signals, remote sensing image data, and drilling core records) within a unified semantic space. They are obtained as follows: First, the high-dimensional initial semantic association vector set is compressed into a low-dimensional latent space through manifold learning dimensionality reduction in S2.4, generating a low-dimensional semantic embedding vector set. Then, multi-source data fusion and encapsulation in S2.5 forms a multi-source semantic embedding vector set. Finally, in S4.4, a differentiable topological loss function is used to perform backpropagation optimization iterations on the latent space coordinates of the multi-source semantic embedding vector set. The coordinate parameters are updated using a gradient descent algorithm, causing data from different sources to spontaneously cluster in the feature space towards the same geological event evolution trajectory, thereby generating the latent space coordinates.

[0155] Receive the differentiable topological loss function generated by S4.3 and the multi-source semantic embedding vector set output by S2, and initialize the latent space coordinate parameter matrix.

[0156] We construct a computation graph based on automatic differentiation and incorporate multi-source semantic embedding vectors as trainable variables into the backpropagation link.

[0157] The Adam optimizer is used to set the initial learning rate and momentum parameters, and the gradient descent iterative process is initiated to minimize the value of the composite loss function.

[0158] The partial derivatives of the composite loss function with respect to the latent space coordinate parameters are calculated using the chain rule to obtain the gradient direction vector. The specific formula is as follows:

[0159]

[0160] Where θ is the latent space coordinate parameter, η is the adaptive learning rate, L is the differentiable topological loss function, and t is the subscript representing the current iteration step (time step).

[0161] Updating the position coordinates of the embedding vector in the hidden manifold based on the gradient direction encourages cross-source data to cluster towards low-entropy regions.

[0162] Gradient correction of the spatial neighborhood topological constraint penalty term is applied simultaneously to suppress anomalous displacements that violate the rules of rock stratum continuity.

[0163] Monitor the rate of change of the loss function of adjacent iteration steps, and stop the optimization iteration when the rate of change is lower than the preset convergence threshold.

[0164] The output is a set of multi-source semantic embedding vectors after parameter updates, forming a preliminary convergent implicitly aligned geological feature manifold.

[0165] S4.5: Based on the initially converged implicitly aligned geological feature manifold, perform manifold smoothness verification and topological stability evaluation, apply manifold learning dimensionality reduction technology to remove local distortions caused by noise interference, and output the finally determined implicitly aligned geological feature manifold to ensure that it has continuity and differentiability in accordance with geological and physical laws, directly supporting the inference operation of the subsequent intelligent recognition model.

[0166] Receive the preliminary convergent implicitly aligned geological feature manifold generated in step S4.4. This manifold contains the coordinate distribution of multi-source data in the implicit space and the corresponding topological constraint residuals.

[0167] Perform local curvature analysis on the preliminary convergent manifold, calculate the geodesic distance deviation in the neighborhood of each embedding vector, and identify non-physical geometric distortion points caused by noise interference.

[0168] Based on the identified distortion points, a Laplacian Beltrami operator is constructed. By solving the eigenvalue problem, the low-frequency intrinsic modes of the manifold are extracted, and the high-frequency noise components are separated from the low-frequency geological structure signals.

[0169] We employ the thermonuclear diffusion method from spectral theory, define a diffusion time parameter to control the smoothing intensity, and perform anisotropic diffusion processing on high-dimensional embedded coordinates to suppress local mutations while preserving the global topological structure.

[0170] We reconstruct the low-dimensional embedding space using Isomap manifold learning, maintaining the invariance of geodesic distances between neighboring points and eliminating linear projection errors introduced by static rule alignment.

[0171] Perform topological persistence homology analysis on the reconstructed manifold, calculate the Betti numbers of the one-dimensional and two-dimensional homology groups, and verify whether there are topological defects such as pores or fractures that violate geological continuity in the manifold.

[0172] If a topological defect is detected, a Tikhonov regularization term is introduced to fine-tune the embedding vector, minimizing the manifold curvature energy function, until the Betti number matches the prior topological expectation of the regional geological structure.

[0173] The output is a final implicitly aligned geological feature manifold that has been smoothed and topologically stable, ensuring its smoothness and physical consistency in the sense of differential geometry.

[0174] Step S5: Convert the newly acquired geological data to be detected into a new semantic embedding vector and input it into the spatiotemporal topology encoder. Calculate the position coordinates and temporal position of the new semantic embedding vector in the implicitly aligned geological feature manifold in real time, and output a dynamically aligned geological feature sequence. Specifically, this includes:

[0175] S5.1: Local geometric invariant extraction processing is performed on the newly acquired geological raw data to be detected. Based on the phase singularity points of the seismic wave signal, the extreme values ​​of the edge curvature of the remote sensing image, and the bedding offset vector of the drilling core, a new semantic embedding vector to be mapped is generated as the input object of the subsequent encoder.

[0176] The system receives newly acquired geological raw data to be tested, including seismic wave signals, remote sensing image data, and drilling core records, maintaining the inherent spatiotemporal sampling characteristics of each data source without resampling. Hilbert-Huang transform is performed on the seismic wave signals to extract instantaneous phase sequences and locate phase abrupt change points, generating a set of waveform phase singularities to characterize the discontinuous characteristics of the subsurface medium's wave impedance interface. Gaussian curvature operators are applied to the remote sensing image data for multi-scale edge detection, calculating the principal curvature extrema within the pixel neighborhood and selecting a set of image edge curvature extrema points reflecting surface structural deformation characteristics. Automatic bedding plane identification is performed on the drilling core records, calculating the angle and displacement deviation between the normal vectors of adjacent bedding planes to construct a core bedding offset vector set reflecting changes in stratigraphic attitude. The waveform phase singularity set, the image edge curvature extrema set, and the core bedding offset vector set are uniformly mapped to the computable geological semantic map potential space constructed in the previous steps. Using a graph attention network mechanism, the semantic association weights between each local geometric invariant and the structural type nodes and lithological combination rule nodes in the geological semantic map are calculated. Based on these semantic association weights, the local geometric invariants are weighted and aggregated to generate a high-dimensional initial semantic association vector with geological ontology constraints. t-SNE manifold learning is used to perform nonlinear dimensionality reduction on the high-dimensional initial semantic association vector, projecting it to a low-dimensional latent space while preserving the local topological structure. L2 norm normalization is then applied to the dimensionality-reduced vector to eliminate numerical bias caused by differences in the dimensions of data from different modalities, generating a new semantic embedding vector to be mapped.

[0177] S5.2: Perform geological perception coding operation on the new semantic embedding vector to be mapped, and use the relative order probability distribution branch in the spatiotemporal topology encoder to calculate the relative order probability distribution of the new data on the geological evolution time axis, and output the temporal evolution state representation.

[0178] Receive the new semantic embedding vector to be mapped generated by S5.1, which contains low-dimensional feature representations of seismic wave phase singularities, remote sensing image edge curvature extrema, and drilling core bedding offset vectors.

[0179] The relative order probability distribution branch in the geological sensing spatiotemporal topology encoder is invoked. This branch consists of a multi-layer temporal attention network and is used to capture long-range dependencies in the geological evolution process.

[0180] The new semantic embedding vector to be mapped is input into the query layer of the temporal attention mechanism, and its similarity score with key nodes of each historical period on the geological evolution timeline is calculated to generate the original attention weight matrix.

[0181] The original attention weight matrix is ​​normalized using the Softmax function to eliminate dimensional differences and establish probability distribution characteristics, thereby obtaining the relative position confidence level under each historical period.

[0182] The relative order probability distribution of the new data on the geological evolution timeline is calculated using the following formula:

[0183]

[0184] in, Let be the relative ordinal probability of the new data in the t-th geological period. Let t be the attention similarity score for the t-th period, and N be the total number of geological history periods.

[0185] The calculated relative order probability distribution is smoothed, and the Gaussian kernel function is used to suppress probability abrupt changes caused by local noise, so as to ensure the continuity of the temporal evolution state.

[0186] The smoothed relative order probability distribution is encoded as a temporal evolution state characterization vector, which quantifies the relative position and uncertainty of newly acquired data in the geological time dimension.

[0187] S5.3: Based on the temporal evolution state representation, perform spatial topological constraint matching, call the rock stratum continuity constraint matrix generation module in the spatiotemporal topology encoder, combine the known regional tectonic ore-controlling model to calculate the spatial neighborhood rock stratum continuity constraint matrix, and output the spatial topological compatibility representation.

[0188] The temporal evolution state representation output by S5.2 is received and used as the semantic guidance input for spatial topological constraint matching.

[0189] The rock strata continuity constraint matrix generation module in the geological sensing spatiotemporal topology encoder is invoked to load the parameters of the known tectonic ore-controlling model for the region.

[0190] Based on the geological history stages determined by the temporal evolution state characterization, the stratigraphic sedimentary sequences and tectonic deformation rules of the corresponding periods in the tectonic ore-controlling model are retrieved.

[0191] Graph convolutional networks are used to extract grid cell features within the spatial neighborhood of newly acquired data, and lithological similarity and attitude consistency between adjacent cells are calculated.

[0192] The topological compatibility weights between any two grid cells i and j within a spatial neighborhood are calculated using the following formula:

[0193]

[0194] in, and σ represents the geological semantic embedding vectors of grid cells i and j, respectively, and σ is the bandwidth parameter used to control the decay rate of neighborhood influence.

[0195] Based on the calculated topological compatibility weights, a sparse spatial adjacency matrix is ​​constructed to characterize the spatial continuity constraints of local geological structures.

[0196] Tensor product operation is performed on the sparse spatial adjacency matrix and the temporal evolution state representation to integrate the evolution probability in the time dimension and the topological constraints in the spatial dimension.

[0197] By performing matrix normalization, the dimensional differences in spatial constraint strength between data sources of different scales are eliminated, and a standardized spatial topological compatibility representation is generated.

[0198] S5.4: Perform implicit manifold projection calculation on the temporal evolution state representation and the spatial topological compatibility representation. By minimizing the consistency entropy value of cross-source semantic embedding and applying the spatial neighborhood topological constraint penalty term, map the new semantic embedding vector to the implicitly aligned geological feature manifold and output the manifold projection coordinate parameters.

[0199] The temporal evolution state representation output from S5.2 and the spatial topological compatibility representation output from S5.3 are received as input data for implicit manifold projection calculation.

[0200] We construct a projection optimization objective function based on geological semantic constraints, which aims to map new semantic embedding vectors onto implicitly aligned geological feature manifolds generated during the training phase.

[0201] Define a consistency entropy loss term for cross-source semantic embedding to measure the difference between the dispersion of new data distribution on the geological evolution timeline and the prior distribution.

[0202] The time-series consistency cost is calculated using the Shannon information entropy formula, which is expressed as:

[0203]

[0204] in Let be the relative ordinal probability of the new data in the i-th geological evolution stage, and N be the total number of evolution stages. This comes at the cost of time-series consistency.

[0205] Define a spatial topology violation penalty term to constrain the continuity of rock strata within the spatial neighborhood of new data to conform to geological structural laws.

[0206] The spatial topological cost is calculated using the graph Laplacian regularization method, expressed by the following formula:

[0207]

[0208] in For spatial neighborhood rock strata continuity constraint matrix elements, and represents the implicit space coordinate vector of adjacent grid cells. M is the total number of grid cell nodes in the spatial topology graph.

[0209] Construct a differentiable topological loss function, and here we will use the differentiable topological loss function defined in S4.3.

[0210] Initialize the initial projection coordinates of the new semantic embedding vector in the implicitly aligned geological feature manifold, usually set to the coordinates of the manifold center or the nearest neighbor sample.

[0211] The gradient of the composite loss function with respect to the projected coordinates is calculated using automatic differentiation techniques to obtain the parameter update direction.

[0212] The Adam optimizer is used to perform gradient descent iterations, gradually updating the projected coordinates to minimize the value of the composite loss function.

[0213] The convergence of the loss function is monitored in each iteration, and optimization is stopped when the gradient norm falls below a preset threshold or the maximum number of iterations is reached.

[0214] Output the final converged manifold projection coordinate parameters, including position coordinate components and time position components.

[0215] S5.5: Based on the manifold projection coordinate parameters, reconstruct the geological feature sequence, extract the position coordinate components and time position components in the manifold projection coordinate parameters, and integrate them to generate a dynamically aligned geological feature sequence with a unified geological semantic coordinate system.

[0216] Receive the manifold projection coordinate parameters output by S5.4, which contain the multidimensional positional components and temporal positional probability distribution of the new semantic embedding vector in the implicitly aligned geological feature manifold.

[0217] The manifold projection coordinate parameters are subjected to dimensional decoupling processing. Tensor slicing operations are used to separate the spatial location coordinate components and the geological evolution time location components, and independent spatial topological indexes and time sequence labels are constructed.

[0218] Based on the spatial location coordinate components, the regional geological grid mapping table is invoked to discretize the implicit spatial continuous coordinates into standard geological exploration grid unit indices, ensuring that the spatial positioning is consistent with the physical coordinate system of the existing geological database.

[0219] Based on the time location component and combined with the geochronological scale, the specific interval to which the newly collected data belongs in the geological history period is calculated, and a timestamp label with a clear geological age attribute is generated.

[0220] The grid cell index, timestamp label, and original multi-source semantic embedding vector are structurally recombined and the data is encapsulated in a "space-time-semantic" triple format to eliminate the spatiotemporal heterogeneity between different data sources.

[0221] Step S6: Perform geological anomaly detection based on the dynamically aligned geological feature sequence to generate a geological anomaly probability distribution map. Specifically, this includes:

[0222] S6.1: Obtain the dynamically aligned geological feature sequence as the initial input data. Based on the position coordinates and time position output by the spatiotemporal topology encoder, perform multi-scale convolutional feature extraction processing on the dynamically aligned geological feature sequence to capture the subtle texture changes and spatial neighborhood associations of local geological structures, and generate multi-scale geological texture feature maps.

[0223] This system receives dynamically aligned geological feature sequences, analyzes the manifold projection coordinate components and temporal location components contained within them, and constructs a multi-channel input tensor with a unified geological semantic coordinate system. Based on the multi-resolution characteristics of geological structures, a parallel multi-scale convolutional neural network branch is designed, with convolutional kernels configured with different receptive fields to capture spatial correlation features ranging from microscopic lithological textures to macroscopic fault distribution. Shallow convolution operations are performed on the input tensor, using small 3x3 convolutional kernels to extract high-frequency local details, generating high-resolution feature maps reflecting subtle changes in core bedding. Simultaneously, deep convolution operations are performed on the input tensor, using large 7x7 convolutional kernels combined with dilated convolution techniques to expand the receptive field, extract low-frequency global structure, and generate low-resolution semantic feature maps reflecting the distribution of seismic wave phase singularities. A spatial attention mechanism is introduced to weightedly fuse the multi-scale feature maps, calculating the feature importance weights for each spatial location, suppressing background noise interference, and enhancing the response intensity of anomalous structural regions. The weighted multi-scale feature maps are concatenated along the channel dimension, and a 1x1 convolutional layer is used to achieve cross-scale feature interaction and information integration, eliminating the semantic gap caused by scale differences. Batch normalization is applied to the integrated feature maps to accelerate model convergence and improve the stability of feature representation, outputting multi-scale geological texture feature maps containing rich spatiotemporal topological information. Through multi-scale convolutional feature extraction, the dynamically aligned geological feature sequence from the previous step is transformed into multi-scale geological texture feature map data, achieving efficient capture of subtle texture changes and spatial neighborhood associations in complex geological structures, significantly improving the feature representation capability of subsequent anomaly detection models for multi-source heterogeneous data.

[0224] For example, a multi-scale convolutional network with three parallel branches is constructed for dynamically aligned geological feature sequences. The first branch uses 64 3x3 convolutional kernels with a stride of 1 and padding of 1 to extract high-resolution local texture features; the second branch uses 128 5x5 convolutional kernels with a stride of 1 and padding of 2 to extract mesoscale structural features; and the third branch uses 256 7x7 convolutional kernels with a dilation rate of 2 to extract large-scale semantic features. The ReLU activation function is applied to the output of each branch, and then a weight map is calculated using a spatial attention module, as shown in the formula: Where A is the attention weight, σ is the sigmoid function, and F is the feature map. The weighted feature maps are concatenated and then reduced to 256 channels via 1x1 convolution. After processing with BatchNorm, a multi-scale geological texture feature map is output. This processing significantly improves the recognition rate of minor faults and greatly optimizes edge localization accuracy.

[0225] S6.2: Receive multi-scale geological texture feature maps as intermediate processing objects, and use a pre-trained geological semantic attention mechanism to perform cross-modal feature weighted fusion processing on the multi-scale geological texture feature maps to suppress irrelevant background noise and enhance the semantic response intensity of abnormal areas, thereby generating a highly robust geological semantic fusion feature vector.

[0226] It receives multi-scale geological texture feature maps as input data, which contain local structural texture information of seismic waves, remote sensing and drilling data in a unified geological semantic coordinate system.

[0227] A geological semantic attention query matrix is ​​constructed, which maps the node embedding vectors in the computable geological semantic map to the query vectors of the attention mechanism to represent the semantic priors of different geological structure types.

[0228] A linear projection transformation is performed on the multi-scale geological texture feature map to generate a Key vector and a Value vector, which respectively represent the spatial location information and the texture intensity information of the channel dimension of the feature map.

[0229] The dot product similarity between the query vector and the key vector is calculated, and the original attention weight matrix is ​​generated by normalizing it through a scaling factor, thereby quantifying the degree of association between each texture region and specific geological semantics.

[0230] A spatial topological mask matrix is ​​introduced, which is transformed from the rock layer continuity constraint matrix generated in the previous step, and is used to shield invalid attention connections that violate geological topological continuity.

[0231] The spatial topological mask is applied to the original attention weight matrix, and negative infinity is assigned to illegal connections. After processing by the Softmax function, an effective attention distribution that conforms to the geophysical laws is generated.

[0232] By utilizing an effective attention distribution to perform weighted summation on the Value vector, cross-modal texture information is fused to highlight anomalous response features that are highly correlated with the current geological semantics and suppress background noise interference.

[0233] Residual connectivity and layer normalization are performed on the weighted fused feature vectors to maintain the stability of gradient flow and enhance the model's sensitivity to minor geological anomalies.

[0234] The above calculation process is repeated by using a multi-head parallel attention mechanism to capture the geological semantic-texture coupling relationship in different subspaces, splice the multi-head outputs and project them through a linear layer.

[0235] By using a geological semantic attention-weighted fusion processing method, the multi-scale geological texture feature map of the previous step is transformed into a highly robust geological semantic fusion feature vector, thereby achieving the expected technical effects of cross-modal feature alignment enhancement and noise suppression.

[0236] For example, we set the number of attention heads to 8, with each head having a dimension of 64. The query vector has a dimension of 512, and both the key and value vectors have a dimension of 512. In the spatial topological mask, adjacent grid cells are marked as 0 if the difference in the dip angle of the rock strata exceeds 15 degrees, and 1 otherwise. After calculating the dot product similarity, we multiply it by a scaling factor: After Softmax processing, the attention weight of the anomalous region increased from 0.1 to 0.85, while the background noise weight decreased to below 0.05. The signal-to-noise ratio of the final output fused feature vector was significantly improved, effectively enhancing the characterization ability of weak anomalous signals such as fault fracture zones.

[0237] S6.3: Input the highly robust geological semantic fusion feature vector into the fully connected classification network layer. Based on the constraint of the shared geological semantic coordinate system, perform nonlinear mapping and dimensionality reduction transformation on the highly robust geological semantic fusion feature vector to project the high-dimensional feature space to the low-dimensional anomaly discrimination subspace and generate geological anomaly discrimination latent variables.

[0238] It receives a highly robust geological semantic fusion feature vector, which integrates spatiotemporal topological constraint information from multi-source heterogeneous data, and serves as the input benchmark for nonlinear mapping.

[0239] A fully connected classification network layer with a multilayer perceptron structure is constructed, and the number of hidden layer nodes and the type of activation function are set to establish a mapping channel from the high-dimensional semantic space to the low-dimensional discriminative subspace.

[0240] A linear weighted transformation is performed on the input high-dimensional feature vector, and the original features are projected onto the intermediate hidden layer using the weight matrix to achieve preliminary compression of feature dimensions and information recombination.

[0241] A nonlinear activation function is introduced to perform element-wise mapping on the linear transformation result. Negative responses are suppressed and the sparse representation capability of key features is enhanced by using ReLU or Sigmoid functions to generate nonlinear hidden layer output.

[0242] Based on the constraint of a shared geological semantic coordinate system, a manifold regularization term is embedded in the loss function to force adjacent geological units to maintain topological proximity in low-dimensional space, thereby preventing feature distortion.

[0243] Dimensionality reduction projection operation is performed, and the high-dimensional nonlinear features are compressed into an anomaly discrimination subspace of a preset dimension through the terminal mapping matrix of the fully connected layer, so as to extract key latent variables that can distinguish between normal and abnormal geological states.

[0244] Calculate the latent variable for geological anomaly discrimination. This variable represents the degree to which the current geological unit deviates from the normal geological evolution trajectory and serves as the direct basis for subsequent probability estimation.

[0245] By using nonlinear mapping and dimensionality reduction transformation of the fully connected classification network layer, the high-robust geological semantic fusion feature vector from the previous step is transformed into a low-dimensional geological anomaly discrimination latent variable, realizing the accurate projection of high-dimensional complex features onto discriminable indicators, and significantly improving the sensitivity and computational efficiency of anomaly detection.

[0246] For example, for seismic and remote sensing fusion data of a mining area, a robust geological semantic fusion feature vector with a 512-dimensional input is used. A fully connected classification network layer with two hidden layers is constructed, with 256 nodes in the first hidden layer and 64 nodes in the second hidden layer. The activation function is Leaky ReLU, and the alpha parameter is set to 0.01. The weight matrix is ​​initialized using the He initialization method. After linear transformation and nonlinear activation, the features are mapped to a 64-dimensional intermediate space. Subsequently, the output layer maps them to an 8-dimensional anomaly discrimination subspace. During training, a manifold regularization coefficient lambda of 0.05 is introduced to ensure that feature points of the same lithological unit are clustered in the low-dimensional space. The final output 8-dimensional geological anomaly discrimination latent variable, with its Euclidean distance, can significantly distinguish fault fracture zones from intact rock masses. Compared with the traditional PCA dimensionality reduction method, the inter-class separation is significantly improved, effectively supporting subsequent probability distribution estimation.

[0247] S6.4: Use the latent variables for geological anomaly discrimination as the basis for probability calculation, and use the soft maximum normalization function to perform probability density distribution estimation on the latent variables for geological anomaly discrimination, so as to quantify the confidence level of each spatial grid cell belonging to various types of geological anomalies and generate the original geological anomaly probability tensor.

[0248] The system receives latent variables for geological anomaly discrimination as input data. These latent variables represent the projected coordinates of high-dimensional features in a low-dimensional anomaly discrimination subspace. A probability normalization calculation module based on the Softmax function is constructed to map discrete, non-standardized logits values ​​to a continuous probability distribution interval. The total number of categories, K, is defined to correspond to a predefined set of geological anomaly types, including typical anomaly patterns such as fault fracture zones, lithological abrupt change zones, and stress concentration zones. An exponential operation is performed on the latent variable vector corresponding to each spatial grid cell to eliminate the influence of negative values ​​and amplify significant feature differences. The original probability value of the i-th grid cell belonging to the k-th geological anomaly type is calculated using the following formula:

[0249]

[0250] in, This represents the confidence level of the probability that the i-th grid cell belongs to the k-th type of anomaly. This represents the component value corresponding to the k-th class in the latent variables for geological anomaly discrimination, where K is the total number of geological anomaly types. The probability density estimation process described above is performed in parallel across all spatial grid cells, ensuring that the sum of the probability values ​​of each cell strictly converges to 1. The calculated probability values ​​are then reorganized according to the spatial topology of the original geological data to construct a three-dimensional tensor data structure. The dimensions of this tensor correspond to the spatial row index, spatial column index, and anomaly category channel, forming the original geological anomaly probability tensor. Through Softmax normalization, the abstract latent variables generated in the previous step are transformed into probability confidence indices with clear physical meaning, thus establishing the foundation for the quantitative characterization and spatial distribution visualization of geological anomaly types.

[0251] S6.5: Read the original geological anomaly probability tensor and combine it with prior knowledge of regional geological structure. Perform spatial smoothing filtering and threshold truncation on the original geological anomaly probability tensor to eliminate isolated noise and optimize the continuity of anomaly boundaries. Finally, output the geological anomaly probability distribution map.

[0252] Read the original geological anomaly probability tensor generated by S6.4. This tensor contains the initial confidence values ​​of each spatial grid cell belonging to different anomaly types.

[0253] By calling upon the prior knowledge base of regional geological structures, the continuity parameters of strata attitude and the topological constraints of fault zone distribution in the current exploration area are extracted, and a spatially smooth weight matrix is ​​constructed.

[0254] Based on the spatial smoothing weight matrix, anisotropic diffusion filtering is performed on the original geological anomaly probability tensor. The smoothing intensity is dynamically adjusted by utilizing the geological semantic similarity between neighboring pixels to suppress isolated noise points of non-geological origin.

[0255] The smoothed probability value is calculated using the following formula:

[0256]

[0257] in, Let (i,j) be the anomaly probability after smoothing the grid. Here, N(i,j) is the normalization coefficient, and N(i,j) is the set of geological neighborhoods of grid (i,j). For adaptive weights calculated based on geological semantic similarity, is the original probability value of the neighborhood grid k.

[0258] Threshold truncation is performed on the filtered probability tensor. A lower confidence threshold is set to remove low-confidence background noise, and an upper confidence threshold is set to retain high-confidence core anomaly regions.

[0259] Morphological closing operations are applied to optimize the boundary continuity of anomalous regions, fill in internal voids caused by threshold truncation, and ensure the integrity of the anomalous body in spatial topology.

[0260] Generate the final binary or hierarchical geological anomaly probability distribution map, clearly identifying the anomaly target areas with geological and physical significance.

[0261] Step S7: Determine whether there are any abnormal regions exceeding a preset threshold in the geological anomaly probability distribution map. If so, trigger a geological early warning signal and record the anomaly spatiotemporal label. If not, update the evolution parameters of the geological semantic map to optimize subsequent inference accuracy. Specifically, this includes:

[0262] S7.1: Perform threshold comparison processing on the probability values ​​of each grid cell in the geological anomaly probability distribution map to filter out a set of candidate anomaly areas that exceed the preset confidence threshold, thereby generating a list of anomaly areas to be confirmed as input objects for subsequent decision execution.

[0263] Receive the geological anomaly probability distribution map generated in the previous steps. This distribution map is a three-dimensional tensor structure that includes spatial grid coordinate indices and corresponding anomaly confidence probability values.

[0264] Traverse each spatial grid cell in the geological anomaly probability distribution map and read the anomaly confidence probability value of the current grid cell as the indicator to be judged.

[0265] The preset dynamic confidence threshold parameter is invoked. The parameter is adaptively set based on the statistical distribution characteristics of historical geological exploration data and prior knowledge of the current regional tectonic complexity.

[0266] The anomaly confidence probability value of the current grid cell is compared with the dynamic confidence threshold, and it is determined whether the probability value is strictly greater than the dynamic confidence threshold.

[0267] If the comparison result is true, the current grid cell is marked as a candidate anomaly point, and its spatial coordinate index and corresponding probability value are extracted and stored in a temporary cache queue.

[0268] If the comparison result is false, the current grid cell is ignored, and the process continues to traverse the next grid cell until the full image scan is completed.

[0269] Connectivity analysis is performed on candidate outliers in the temporary cache queue, and the eight-neighbor algorithm is used to identify the set of spatially adjacent candidate outliers.

[0270] Connected candidate outliers are aggregated into an independent candidate outlier region, and the geometric center coordinates, coverage area, and average confidence level of the region are calculated.

[0271] Construct a list of anomaly regions to be confirmed, and encapsulate the identifier, geometric center coordinates, coverage area, and average confidence score of each independent candidate anomaly region into a structured data object.

[0272] By using the threshold comparison and connected component aggregation methods described above, discrete grid-level probability values ​​are transformed into a list of anomaly regions to be confirmed with clear spatial boundaries. This achieves a data granularity transformation from pixel-level detection to region-level decision input, significantly reducing computational redundancy in subsequent early warning processing and improving the spatial continuity of anomaly localization.

[0273] S7.2: Perform spatiotemporal label association operation based on the spatial coordinate index of the list of anomaly areas to be confirmed, so as to map the high probability areas in the geological anomaly probability distribution map into structured data units with clear timestamps and geographical location information, thereby generating a geological early warning signal containing complete spatiotemporal dimensions.

[0274] Obtain a list of anomaly regions to be confirmed, and extract the set of two-dimensional grid coordinate indices of each candidate anomaly region in the geological anomaly probability distribution map, which will serve as the spatial reference data for spatiotemporal label association.

[0275] Based on the spatial reference data, the implicitly aligned geological feature manifolds output by the spatiotemporal topology encoder in the preceding steps are queried, and the manifold projection temporal location components that match the grid coordinate index are retrieved to construct a spatial-temporal mapping pair.

[0276] The system calls upon the regional geological exploration metadata database, inverts the corresponding geographic information system latitude and longitude coordinates and elevation data based on the spatial coordinate index, and converts the abstract grid index into absolute geographical location parameters with physical meaning.

[0277] Linear interpolation is used to make the discrete time position components continuous. Combined with the system clock reference at the time of data acquisition, the precise absolute timestamp of the abnormal event is calculated to eliminate the temporal ambiguity caused by the sampling interval.

[0278] Define the data structure for geological early warning signals, which includes four core fields: anomaly type identifier, confidence probability value, absolute geographical location parameter, and precise absolute timestamp, and perform structured data encapsulation operations.

[0279] Integrity checks are performed on the encapsulated structured data units to ensure the non-emptiness and logical consistency of spatiotemporal dimension information, and to generate geological early warning signals containing complete spatiotemporal dimensions.

[0280] S7.3: Perform residual feature extraction processing on the probability distribution map of geological anomalies that are determined to not exceed the preset threshold, so as to obtain the deviation vector between the current inference result and the prior knowledge of the geological semantic map, thereby generating a model error feedback sequence for parameter optimization.

[0281] The system receives a probability distribution map of geological anomalies that is determined to not exist in any abnormal areas exceeding a preset threshold, and uses it as the initial input for residual feature extraction.

[0282] The geological anomaly probability distribution map is subjected to background field reconstruction processing. The theoretical probability distribution model under ideal conditions is constructed using prior knowledge in the geological semantic map. The pixel-by-pixel deviation between the actual output probability and the theoretical probability is calculated.

[0283] A sliding window mechanism is used to traverse the deviation matrix and extract statistical moment features in the local spatial neighborhood, including mean offset, variance volatility and skewness coefficient, to quantify the degree of deviation between local geological structure and global semantic constraints.

[0284] Based on the extracted statistical moment features, residual feature vectors are constructed. Mahalanobis distance is used to measure the degree of dispersion of the residual vector of each grid cell from the center of the standard normal distribution, and potential error hotspots with significant structural biases are identified.

[0285] Principal component analysis is performed on the residual characteristics of potential error hotspots to extract the top K principal component directions with the highest contribution rates, forming a low-dimensional residual subspace basis that characterizes the systematic bias of the model.

[0286] The low-dimensional residual subspace basis is mapped back to the geological semantic map node space, the projection weight of each semantic node on the residual subspace is calculated, and a deviation vector reflecting the intensity of the conflict between the current reasoning result and prior knowledge is generated.

[0287] S7.4: Based on the model error feedback sequence, gradient update calculations are performed on the weights of structural type nodes and the connection strength of lithological combination rules in the geological semantic map to adjust the evolution parameters of structured semantic units in the geological ontology library, thereby generating an updated geological semantic map to support the improvement of the next round of inference accuracy.

[0288] The system receives the model error feedback sequence generated by S7.3, analyzes the geological semantic deviation vector and gradient direction information contained therein, and uses them as the driving signal for updating the map parameters.

[0289] The deviation components in the error feedback sequence are decoupled to separate the classification confidence residual for structural type nodes and the topological consistency residual for lithological combination rule connections.

[0290] Based on the classification confidence residual, a node weight update objective function is constructed to calculate the sensitivity coefficient of each type of node in the current inference cycle.

[0291] An adaptive learning rate is used to dynamically adjust the step size of the weights of the constructed type nodes based on the sensitivity coefficient, and gradient descent updates are performed on the node attribute values. The specific formula is as follows:

[0292]

[0293] Among them, W i Let be the weight of the i-th construction type node, η be the adaptive learning rate, and L be the loss function. This represents the updated node weights.

[0294] Based on topological consistency residuals, the correction amount of the connection strength of lithological combination rules is calculated to assess the change in the tightness of semantic association between adjacent nodes.

[0295] A regularization constraint is introduced to prevent overfitting or numerical divergence during the connection strength update process, ensuring the stability of the spectral structure. The specific formula is as follows:

[0296]

[0297] in, Let λ be the connection strength between nodes i and j, λ be the regularization coefficient, and β be the decay factor. This represents the connection strength between nodes i and j after the update.

[0298] The updated node weights and connection strengths are re-injected into the geological semantic map, and global normalization is performed to eliminate dimensional differences.

[0299] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0300] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0301] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent identification of geological anomalies, characterized in that, include: S1: Acquire multi-source heterogeneous geological raw data, extract structured semantic units based on the geological ontology library, and construct a computable geological semantic map; S2: Based on the multi-source heterogeneous geological raw data, retain the inherent spatiotemporal sampling characteristics of each data source, extract local geometric invariants, and map the local geometric invariants into low-dimensional semantic embedding vectors embedded in the geological semantic map to generate a multi-source semantic embedding vector set; S3: Based on the multi-source semantic embedding vector set, extract the geological evolution time features based on the temporal attention branch in the pre-constructed spatiotemporal topology encoder, extract the spatial neighborhood strata continuity features based on the spatial topology constraint branch, and generate the geological topology constraint features through dual-branch coupling operation. S4: Define a differentiable topological loss function based on the geological topological constraint features to generate an implicitly aligned geological feature manifold; S5: Convert the newly acquired geological data to be detected into a new semantic embedding vector and input it into the spatiotemporal topology encoder. Calculate the position coordinates and time position of the new semantic embedding vector in the implicitly aligned geological feature manifold in real time, and output a dynamically aligned geological feature sequence. S6: Perform geological anomaly detection based on the dynamically aligned geological feature sequence and generate a geological anomaly probability distribution map; S7: Determine whether there are any abnormal areas exceeding a preset threshold in the geological anomaly probability distribution map. If so, trigger a geological early warning signal and record the anomaly spatiotemporal label. If not, update the evolution parameters of the geological semantic map.

2. The intelligent identification method for geological anomaly detection according to claim 1, characterized in that, The multi-source heterogeneous geological raw data includes seismic wave signals, remote sensing image data, and drilling core records.

3. The intelligent identification method for geological anomaly detection according to claim 2, characterized in that, Step S2 specifically includes: Modal specificity analysis was performed on the seismic wave signals, remote sensing image data and drilling core records. Based on the inherent spatiotemporal sampling characteristics of each data source, a non-resampled data preservation strategy was constructed to generate a set of modally independent data streams. Based on the modal independent data stream set, local geometric feature extraction operation is performed to generate an original feature element library containing waveform phase singular point sequence, image edge curvature extreme point set and core bedding offset vector group; Based on the original feature library and the geological semantic map, semantic association mapping is performed. Using the graph neural network embedding mechanism, the waveform phase singular point sequence, the image edge curvature extreme point set and the core bedding offset vector group are respectively mapped to the structural type node, lithological combination rule node and stress response mode node in the geological semantic map to generate the initial semantic association vector group. Based on the initial semantic association vector set, dimensionality reduction and manifold projection are performed to compress the high-dimensional initial semantic association vector set into a low-dimensional latent space while maintaining the local geometric invariant topological structure, thereby generating a low-dimensional semantic embedding vector set. Based on the low-dimensional semantic embedding vector set, multi-source data fusion encapsulation is performed, and the low-dimensional semantic embedding vectors are structurally reorganized according to the data source identifier and spatiotemporal label to generate the multi-source semantic embedding vector set.

4. The intelligent identification method for geological anomaly detection according to claim 3, characterized in that, Based on the modal independent data stream set, local geometric feature extraction is performed to generate an original feature element library containing waveform phase singular point sequences, image edge curvature extreme point sets, and core bedding offset vector sets, including: Phase singularity point detection is performed on the processed seismic wave signal to obtain the waveform phase singularity point sequence. The extreme value point set of the image edge curvature is obtained based on the remote sensing image data. The core record is processed using the bedding offset vector calculation method to obtain the core bedding offset vector set.

5. The intelligent identification method for geological anomaly detection according to claim 3, characterized in that, Step S3 specifically includes: Based on the multi-source semantic embedding vector set, the geological evolution time axis mapping process is performed through the temporal attention mechanism in the spatiotemporal topology encoder to calculate the relative position weight of each semantic embedding vector in the geological history period and generate the relative order probability distribution of geological evolution. Based on the preset regional known tectonic ore-controlling model and the relative order probability distribution of geological evolution, spatial neighborhood topology analysis is performed. The rock strata occurrence correlation features between adjacent grid cells are extracted using graph convolutional networks to generate a spatial neighborhood rock strata continuity constraint matrix. Based on the relative order probability distribution of geological evolution and the spatial neighborhood strata continuity constraint matrix, a two-branch feature coupling operation is performed, and the temporal order information and spatial topological constraints are fused using tensor product operation to generate the geological topological constraint features.

6. The intelligent identification method for geological anomaly detection according to claim 5, characterized in that, The graph convolutional network consists of two consecutive hidden layers and one output layer. The first and second graph convolutional layers are hidden layers, used to perform layer-by-layer aggregation and nonlinear transformation on the input multi-source semantic embedding vector. The third graph convolutional layer is the output layer, used to introduce the graph Laplacian operator for smoothing constraints and output the spatial topological feature vector.

7. The intelligent identification method for geological anomaly detection according to claim 5, characterized in that, Step S4 specifically includes: The relative order probability distribution of geological evolution and the multi-source semantic embedding vector are jointly encoded to extract the distribution dispersion of each data source on the geological evolution time axis and generate cross-source evolution order consistency entropy value. Based on the spatial neighborhood strata continuity constraint matrix and the multi-source semantic embedding vector, a topological compatibility discrimination operation is performed. The gradient difference of stratigraphic distribution characteristics between adjacent grid cells is calculated using the graph Laplacian regularization method to generate a spatial topological violation penalty term. The cross-source evolution sequence consistency entropy value and the spatial topology violation penalty term are weighted and fused, and an adaptive balance coefficient is introduced to adjust the contribution weights of temporal consistency and spatial continuity, thereby constructing a differentiable topology loss function. The backpropagation optimization iteration is performed on the latent space coordinates of the multi-source semantic embedding vector set using the differentiable topological loss function. The loss function value is minimized and the embedding vector parameters are updated by the gradient descent algorithm, generating a preliminary converged implicitly aligned geological feature manifold. Based on the initially converged implicitly aligned geological feature manifold, manifold smoothness verification and topological stability evaluation are performed. Manifold learning dimensionality reduction techniques are applied to remove local distortions caused by noise interference, and the implicitly aligned geological feature manifold is output.

8. The intelligent identification method for geological anomaly detection according to claim 7, characterized in that, Extracting the distribution dispersion of each data source on the geological evolution time axis and generating cross-source evolutionary order consistency entropy values ​​includes: calculating the distribution dispersion of each data source on the geological evolution time axis using the Shannon information entropy formula, and mapping the distribution dispersion to the cross-source evolutionary order consistency entropy values.

9. The intelligent identification method for geological anomaly detection according to claim 5, characterized in that, Step S5 specifically includes: Local geometric invariant extraction processing is performed on the newly acquired geological raw data to be detected. Based on the phase singularity points of seismic wave signals, the extreme values ​​of edge curvature of remote sensing images, and the bedding offset vector of drilling cores, a new semantic embedding vector to be mapped is generated. Geological sensing encoding operation is performed on the new semantic embedding vector to be mapped. The relative order probability distribution branch in the spatiotemporal topology encoder is used to calculate the relative order probability distribution of the new data on the geological evolution time axis and output the temporal evolution state representation. Based on the temporal evolution state representation, spatial topological constraint matching is performed. The rock strata continuity constraint matrix generation module in the spatiotemporal topology encoder is called. Combined with the known tectonic ore-controlling model of the region, the spatial neighborhood rock strata continuity constraint matrix is ​​calculated, and the spatial topological compatibility representation is output. Implicit manifold projection calculations are performed on the temporal evolution state representation and the spatial topological compatibility representation. By minimizing the consistency entropy value of cross-source semantic embedding and applying the spatial neighborhood topological constraint penalty term, the new semantic embedding vector is mapped to the implicitly aligned geological feature manifold, and the manifold projection coordinate parameters are output. Based on the manifold projection coordinate parameters, the geological feature sequence is reconstructed, and the position coordinate components and time position components in the manifold projection coordinate parameters are extracted and integrated to generate the dynamically aligned geological feature sequence.

10. The intelligent identification method for geological anomaly detection according to claim 1, characterized in that, The structured semantic units include: tectonic type, lithological combination rules, and stress response mode.

Images