Geological reservoir fracture segmentation method and system fusing 3DU-Net graph attention mechanism and graph theory connectivity detection

Abstract

The invention provides a geological reservoir fracture segmentation method fusing a 3DU-Net graph attention mechanism and graph theory connectivity detection and a system thereof, and aims at solving the technical problems of multi-modal data registration, 3D U-Net graph attention feature extraction, graph theory connectivity detection and the like. The problem of insufficient geological reservoir fracture segmentation precision caused by data heterogeneity and connectivity abnormity is effectively solved, high-precision segmentation of a complex geological reservoir fracture network is realized through accurate alignment of multi-modal data and a graph attention feature extraction and graph theory connectivity detection technology based on 3D U-Net, and the method has the advantages of high precision and high efficiency. The problems of local misjudgment and abnormal connectivity caused by the fact that a traditional method only depends on amplitude or texture features are effectively solved, and therefore segmentation precision and robustness are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of geological exploration, and in particular to a geological reservoir fracture segmentation method and system that integrates a 3DU-Net graph attention mechanism with graph theory connectivity detection. Background Art

[0002] Accurate identification of fracture networks in geological reservoirs is a key step in oil and gas resource assessment and development, requiring the coordinated analysis of multimodal data such as seismic data, well logs, and core scans. However, multimodal data suffers from spatial registration errors due to differences in acquisition methods and formation stress and deformation, resulting in blurred fracture boundaries in the fused data. Traditional segmentation methods rely solely on amplitude or texture features, ignoring the topological relationship between the fracture network and the formation stress field and lithologic zoning. This can lead to misidentification of local noise or non-connected pores as valid fractures, or the omission of low-permeability microfractures. This can cause segmentation results to conflict with geomechanical laws, seriously impacting reservoir modeling accuracy and development planning.

[0003] Current technologies primarily employ two approaches: 1) deep learning-based 3D U-Net and its variants, which extract multi-scale features through an encoder-decoder architecture; 2) graph-theory-based segmentation methods, which utilize fractures to construct a graph structure for connectivity analysis. Some studies have attempted to incorporate graph attention mechanisms (GATs) to capture long-range spatial dependencies, but these approaches still suffer from significant drawbacks: multimodal registration relies on rigid transformations, which inadequately compensate for nonlinear formation deformation (such as fault movement), leading to cross-modal feature misalignment; 3) graph models fail to incorporate lithologic zoning rules and fracture curvature constraints, which can easily generate anomalous cross-layer connectivity; and 4) segmentation networks are decoupled from topological analysis, making it difficult to optimize feature learning directions based on geological constraints. Summary of the Invention

[0004] In response to the needs raised in the above background technology, an embodiment of the present invention provides a geological reservoir fracture segmentation method and system that integrates the 3DU-Net graph attention mechanism and graph connectivity detection. The method aims to effectively solve the problem of insufficient accuracy in geological reservoir fracture segmentation caused by data heterogeneity and connectivity anomalies by adopting technical means such as multimodal data alignment, 3D U-Net graph attention feature extraction and graph connectivity detection.

[0005] A geological reservoir fracture segmentation method that integrates the 3DU-Net graph attention mechanism and graph connectivity detection, specifically including the following steps:

[0006] Step 1: Multimodal data integration and alignment. This step uses data from different acquisition devices (such as seismic 3D volumes, well logging curves, and core CT images) to achieve spatial and depth correction between data through affine transformation and elastic registration methods (such as B-spline registration), thereby generating a three-dimensional fused voxel matrix containing multimodal numerical features, achieving unified data registration and feature alignment.

[0007] Step 2: Deep Feature Extraction and Graph Attention Enhancement. In this step, the 3D fused voxel matrix obtained in step 1 is fed into an improved 3D U-Net model. Multi-scale features are extracted through multiple layers of 3D convolutions, while a graph attention mechanism is embedded. This mechanism includes spatial attention (capturing edge and texture details within a local neighborhood) and channel attention (assigning dynamic weights to different modalities) to enhance the ability to distinguish cracks from background.

[0008] Step 3: Graph-theoretic connectivity topology modeling. This step, based on the multi-level feature graph output from Step 2, extracts the fracture centerlines through threshold segmentation and Gaussian smoothing. These centerline voxels are then abstracted into graph nodes. Lie algebraic path integrals and curvature constraints (e.g., using the Nijenhuis operator) are used to calculate distances between nodes and the rationality of connections. This constructs a connectivity topology that reflects the true structure of the fracture network. Topological constraints are also used to exclude anomalous connections.

[0009] Step 4: Initial Segmentation and Connectivity Constraint Optimization. This step uses the decoder to perform deconvolution on the multi-scale feature map, restoring the original resolution and generating an initial fracture network probability map. A connectivity constraint loss function (ensuring consistent probabilities of connected nodes in the graph) and a joint Dice-Focal loss (accounting for class imbalance and edge sensitivity) are introduced to optimize the initial segmentation results, ensuring they conform to geological seepage patterns and fracture network connectivity requirements.

[0010] Step 5: Non-backtracking graph walk post-processing. This step performs post-processing on the initial segmentation results. By constructing a hypergraph based on spatial proximity and feature similarity and adopting a non-backtracking random walk strategy, isolated noise areas are removed and only voxels connected to the main crack network are retained, thereby further improving the segmentation accuracy.

[0011] Step 6: Dynamic training strategy and regularization control. In this step, an exponentially decaying learning rate strategy and L2 regularization are used during the model training phase to ensure rapid convergence in the early stages and fine-tuning of parameters in the later stages. At the same time, weight decay is applied to the attention module to prevent the model from over-relying on local or single modalities and improve the overall generalization ability.

[0012] Step 7: Multimodal cross-validation and result fusion. This step evaluates the model performance through four-fold cross-validation, integrates the prediction results under different modes, and finally calculates and optimizes the average Dice coefficient and sensitivity index to verify and ensure the robustness and high-precision performance of the model in the segmentation of complex geological reservoir fracture networks.

[0013] Further: A geological reservoir fracture segmentation system that integrates the 3DU-Net graph attention mechanism and graph connectivity detection consists of several modules, including:

[0014] Multimodal data acquisition and preprocessing module: This module is responsible for collecting and preprocessing seismic 3D volume data, well logging curves, and core CT scan images. It also uses affine transformation to adjust the vertical depth and spatial position between different data to align the modal data. It also uses B-spline elastic registration to compensate for the nonlinear deformation caused by stratum fault movement and outputs a 3D fused voxel matrix containing multimodal numerical features.

[0015] The feature extraction and graph attention module inputs the preprocessed 3D fused voxel matrix into the improved 3D U-Net and extracts multi-scale features through multi-layer 3D convolution. It also uses the graph attention mechanism to enhance the boundary response between cracks and background, introduces a spatial attention mechanism to capture edge and texture features within the local neighborhood, and introduces a channel attention mechanism to dynamically weight the contributions of different modalities and suppress noise.

[0016] The graph-theoretic connectivity modeling module uses threshold segmentation to extract crack centerlines based on the multi-scale feature map output from the encoder, and constructs a crack connectivity topology map using centerline voxels as graph nodes. It uses Lie algebraic path integrals and curvature constraints to calculate reasonable connection distances between nodes, and applies topological constraints (such as the Nijenhuis operator) to block abnormal connections, ensuring that the constructed graph structure truly reflects the crack network.

[0017] The fracture network segmentation and connectivity optimization module uses a decoder to restore the multi-layer feature map to its original resolution, generates an initial fracture network probability map, and optimizes the segmentation results by introducing connectivity constraints and a Dice-Focal joint loss function. The connectivity constraint loss function ensures that the predicted probabilities of connected nodes in the graph are consistent and conform to geological seepage laws. The Dice-Focal joint loss is introduced to balance the class imbalance problem and enhance edge sensitivity.

[0018] The non-backtracking graph walk post-processing module performs post-processing on the initial probability map by constructing a hypergraph and performing non-backtracking random walks to remove isolated noise areas and retain voxels connected to the main crack network.

[0019] Dynamic training and multimodal validation module, which is used to manage the model training process, including dynamic learning rate adjustment and weight decay, to ensure rapid convergence of the model in the early stage and fine-tuning of parameters in the later stage; at the same time, cross-validation is used to fuse multimodal prediction results to evaluate model performance.

[0020] Use exponential decay learning rate strategy and L2 regularization: used to prevent overfitting and weight bias.

[0021] Four-fold cross validation was used to improve model robustness and aggregate prediction results.

[0022] Further: the terminal device may include: a processor, a storage medium and a bus, the storage medium stores machine-readable instructions executable by the processor, when the terminal device is running, the processor and the storage medium communicate through the bus, and the processor executes the machine-readable instructions to execute the steps of the deep learning model training method described in the above embodiments.

[0023] Further: A storage medium storing a computer program, wherein the computer program executes the steps of the above method when executed by a processor.

[0024] Further: A computer program product, comprising a computer program, wherein the computer program executes the method described above when executed by a processor.

[0025] Beneficial effects of the present invention: Compared with the existing technology, the advancement of this solution lies in that, through the precise alignment of multimodal data, graph attention feature extraction based on 3D U-Net and graph connectivity detection technology, high-precision segmentation of complex geological reservoir fracture networks is achieved, which effectively solves the local misjudgment and connectivity anomaly problems caused by traditional methods that rely only on amplitude or texture features, thereby significantly improving the segmentation accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 A flow chart of the method of the present invention is shown. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be understood that the drawings in the present invention only serve the purpose of illustration and description and are not used to limit the scope of protection of the present invention. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the present invention illustrate the operations implemented according to some embodiments of the present invention. It should be understood that the operations of the flowchart can be implemented out of sequence, and steps that have no logical context relationship can be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of the present invention, can add one or more other operations to the flowchart, or remove one or more operations from the flowchart.

[0029] In addition, the embodiments described in the present invention are only some of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0030] It should be noted that the term "including" will be used in the embodiments of the present invention to indicate the presence of the features declared thereafter, but does not exclude the addition of other features. It should also be noted that similar numbers and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of the present invention, it should also be noted that the terms "first", "second", "third", etc. are only used to distinguish the description and are not to be understood as indicating or implying relative importance.

[0031] The following is a detailed description of this case with reference to the relevant drawings in the specification.

[0032] See Figure 1 , shows a geological reservoir fracture segmentation method that integrates the 3DU-Net graph attention mechanism and graph connectivity detection. The specific steps include:

[0033] Step 1: Multimodal geological data registration and feature alignment

[0034] 1.1. Due to differences in data collected by different data acquisition devices, directly fusing multimodal data will lead to errors in the positioning of the fracture network area. This step collects seismic 3D volume data, well logging curves, and core CT scan images, and uses affine transformation to adjust the vertical depth calibration of the well logging data to align it with the spatial reference of the seismic data volume; B-spline elastic registration is used to compensate for nonlinear deformation caused by formation fault dislocation, and the final output is a 3D fused voxel matrix containing seismic wave impedance, well logging porosity, and core permeability;

[0035] Assume that the earthquake data volume is (Dimensions are H×W×D, representing height, width, and depth), well logging curve image (dimensions are H′×W′×D′), the core CT scan image is (dimensions are H″×W″×D″), through the affine transformation matrix Adjustment I Core The spatial orientation of (Formula 1):

[0036] I′ Core =T affine I Core (1)

[0037] Where R is the rotation matrix; t is the translation vector, which ensures that the core sample is aligned with the underground structure;

[0038] 1.2. B-spline elastic registration generates a non-rigid displacement field through Equation (2), optimizes the control grid parameters to minimize the local mutual information difference (Equation 3), and compensates for the change of formation dip and fault movement:

[0039]

[0040] In formula (2), c ijk represents the control grid, which determines the shape of the displacement field; θ represents the spacing between control points, which affects the local deformation granularity of the displacement field;

[0041] In formula (3), c controls the grid parameters, which are optimized to minimize the mutual information difference; It represents the image after the displacement field acts;

[0042] The final output is the 3D fused voxel matrix In the generated three-dimensional fusion voxel matrix V(x), each voxel (three-dimensional pixel) simultaneously contains the numerical features of three modalities, such as seismic amplitude, well logging gamma value and core CT grayscale.

[0043] Step 2: 3D U-Net graph attention feature extraction

[0044] 2.1. Input the registered 3D fusion voxel matrix V(x) into the improved 3DU-Net encoder, extract multi-scale features through 3D convolution, and set the feature map of the first layer to be (height, width, depth, number of channels), used to capture the macroscopic structure and microscopic fracture texture of the reservoir;

[0045] 2.2. Then integrate the spatial attention mechanism A s (Enhanced Crack-Matrix Boundary Response) and Channel Attention Mechanism A c (dynamically weighted seismic / well log / core modal contributions);

[0046] Spatial AttentionA s It is used to analyze the feature differences within the 3×3×3 neighborhood around each voxel. The calculation process is shown in formula (4):

[0047]

[0048] In formula (4), K independent 3D convolution kernels (parameters ) Extract multi-scale local features. For example, when K = 3, it is possible to capture edges, textures, and smooth areas respectively, and perform weighted summation of K convolution results, with weight ω k Through back propagation learning, normalization (such as Softmax) ensures ∑ω k =1; Generate spatial attention map A through Sigmoid function σ s (F l )∈[0,1] H×W×D , high-value areas (such as fracture network boundaries) are enhanced.

[0049] Channel Attention A c Used to evaluate the importance of different modes to the current area, suppress the influence of noise modes, and channel attention through global average pooling GAP3D and fully connected layers Generate modal weights (Equation 5):

[0050]

[0051] in Map the C-dimensional features to 3 modal weights. If a modal weight is lower than the threshold, its gradient is attenuated during back propagation.

[0052] Implement gradient attenuation for low-contribution modes (such as logging sections contaminated by drilling disturbances) to enhance the characteristic expression of fracture-sensitive modes;

[0053] Through the synergistic effect of these two attention mechanisms, the model is able to adaptively focus on key voxels related to the crack network and reduce background interference.

[0054] 2.3. Finally, output a multi-level feature map with functional sensitivity

[0055] Multi-level feature maps

[0056] Step 3: Reservoir fracture graph connectivity modeling

[0057] Multi-level feature maps based on encoder output (Contains abstract representation of crack branches), extracts the crack centerline through threshold segmentation and Gaussian smoothing, and defines each centerline voxel as a graph node to form a node set The curvature-constrained distance between nodes is calculated by combining Lie algebraic path integral (the Nijenhuis operator quantifies the rationality of fracture branching), and a fracture connectivity topology map is constructed (path constraint). Abnormal connections across faults or lithologic interfaces are blocked (topology constraint). The resulting fracture connectivity map reflects the true structure of the fracture system and is used to constrain the rationality of fracture network segmentation.

[0058] Specifically, based on the characteristics of reservoir fractures, the following rules are used to construct edge connections:

[0059] Path constraint: If there is a continuous crack path between two nodes (the characteristic gradient on the path changes smoothly), then a connection is established and the node v is defined i With v j The rational distance between curve (v i ,v j ), d curve (v i ,v j ) The calculation process is shown in formula (6):

[0060]

[0061] Where γ(t) is the crack path; λ curve is the weight coefficient, which is used to balance the contribution of the curvature integral and the Nijenhuis operator;

[0062] Assume that the crack path γ(t) is a parameterized curve with a curvature and the torsion τ(t) are calculated from the Frenet frame:

[0063]

[0064] Topological constraints: block abnormal connections across faults or lithologic interfaces (calculate path curvature using the Nijenhuis operator, and determine that the curvature is too large to be disconnected). The Nijenhuis operator quantifies the abnormal value of branch curvature. If d curve (v i ,vj )>ψ, it is determined as an abnormal connection across a fault or lithologic interface and is removed from the topological map, where ψ represents the curvature distance threshold used to determine abnormal connections.

[0065] Step 4: Fracture network segmentation and connectivity constraint optimization

[0066] Decoder deconvolution is used to restore the multi-level feature map resolution and generate the initial segmentation result (initial crack network probability map P∈[0,1] H×W×D ), two types of errors may occur at this time:

[0067] Low-permeability barrier layers (such as calcareous cemented zones next to fractures) are misjudged as effective fracture systems.

[0068] The segmented areas are not connected to the dominant seepage channel network (violating the fluid migration pattern dominated by Darcy's law).

[0069] To this end, the connectivity constraint loss function (Equation 7) and the Dice-Focal joint loss function (Equation 8) are introduced to optimize the segmentation accuracy of the model for the edges of the fracture network and the seepage channel network.

[0070]

[0071] In formula (7), P(v i ) represents the crack node v i The crack network probability; N(v i ) represents node v i The neighbor set of (according to the crack connectivity topology);

[0072] The introduction of a connectivity constraint loss function requires that the fracture network probabilities corresponding to connected fracture nodes be consistent (for example, if nodes A and B are connected, the fracture network probability at A must be close to the value at B). This constraint ensures that the segmentation results conform to the principle that "fracture networks are attached to seepage channel networks" and reduces isolated error areas.

[0073]

[0074] In formula (8), P represents the crack network probability map predicted by the model; Y represents the true label (gold standard); α represents the FocalLoss weight, which balances the difficult and easy samples; π represents the modulation factor, which reduces the loss contribution of easy-to-classify samples;

[0075] The Dice-Focal joint loss function is used to balance category imbalance and edge sensitivity;

[0076] The total loss Represents the connectivity constraint loss function l connThe weight coefficient is used to adjust the contribution ratio of the crack topology constraint to the total loss.

[0077] Step 5: Post-processing of non-backtracking graph walks

[0078] A non-backtracking hypergraph walk (defining a hyperedge as a set of spatially adjacent voxels with similar characteristics) is performed on the initial fracture network probability map. Isolated noise (such as false fractures caused by rock cuttings interference) is eliminated through a probabilistic transfer model, and voxels connected to the main fracture network area are retained.

[0079] After threshold segmentation of the initial crack network probability map, isolated small areas may remain (such as noise or crack truncation artifacts). To this end, a hypergraph model is constructed:

[0080] Define hyperedge e k ={x i |||x i -x j ||2<δ,||F(x i )-F(x j )||2<ε}, where δ represents the spatial proximity threshold, which represents the maximum spatial distance allowed to form a hyperedge, ||x i -x j ||2<δ is a mathematical constraint to ensure that the voxels within the hyperedge are adjacent in three-dimensional space; ε represents the feature similarity threshold, which represents the maximum percentage of feature differences allowed, ||F(x i )-F(x j )||2<ε is a mathematical constraint; spatially adjacent (26-neighborhood) voxels with similar features (such as ADC value difference <5%) are grouped as the same hyperedge, and then a hypergraph is constructed, and the probability is updated through non-backtracking random walk.

[0081] Non-backtracking walks simulate the process of random walkers spreading outward from the core of the crack network, prohibiting repeated visits to traversed hyperedges. If a region cannot be connected to the main crack network area through a hyperedge, it is determined to be isolated noise and removed. This step can eliminate <5mm 3 Isolated false-positive regions.

[0082] Step 6: Dynamic Learning Rate and Weight Decay

[0083] An exponentially decaying learning rate strategy (fast convergence in the early stage and fine-tuning parameters in the later stage) and L2 regularization are used to constrain the attention weight to prevent the model from over-relying on a single modality or local features and improve generalization ability.

[0084] A higher learning rate (1e-3) is used for rapid convergence in the early stages of training, and then gradually reduced to 1e-1 in the later stages to fine-tune the parameters. The weight parameters for the attention module are:

[0085] Weight filtering: L2 regularization is applied to the neighborhood weights in spatial attention to prevent certain areas from having too high weights (such as excessive attention to a single modality). The process is as follows:

[0086]

[0087] In formula (9), λ represents the regularization coefficient; Represents the neighborhood weight of the lth layer in spatial attention;

[0088] Channel suppression: If the channel attention weight of a modality is continuously lower than 0.1, its contribution is judged to be low, and the gradient update is reduced during back propagation. The process is shown in formula (10):

[0089] η t =η init ·exp(-κt)(10)

[0090] In formula (10), κ controls the speed of learning rate decrease, balancing the rapid convergence in the early stage and the fine tuning of parameters in the later stage; η init represents the initial learning rate; t represents the training round;

[0091] This strategy balances the importance of different modalities and regions and improves the generalization ability of the model.

[0092] Step 7: Multimodal cross-validation and result fusion

[0093] The model was evaluated through four-fold cross-validation, and multimodal prediction results were integrated (occlusion experiment quantified the contribution of CT / MRI). The average Dice coefficient (0.89±0.03) and sensitivity (92.5%) were calculated to verify the robustness of the algorithm for small fracture networks, ultimately achieving accurate segmentation of the reservoir fracture network.

[0094] Principle of the method of the present invention: The method of the present invention realizes the accurate segmentation of reservoir fracture network by integrating multimodal geological data preprocessing, 3DU-Net graph attention feature extraction and graph theory connectivity detection. First, the seismic data, well logging curves and core CT images are spatially aligned through affine transformation and B-spline elastic registration to generate a three-dimensional fused voxel matrix containing multimodal features; then, an improved 3D U-Net is used to extract multi-scale features in combination with spatial and channel attention mechanisms, and the fracture centerlines are extracted using threshold segmentation and Gaussian smoothing. They are abstracted into graph nodes to construct a topological graph, and the reasonable connection between nodes is determined by path constraints and topological constraints; then, the decoder is used to restore the resolution to generate the initial fracture network probability map, and the segmentation results are optimized by connectivity constraint loss and Dice-Focal joint loss; finally, the results are post-processed with the help of non-backtracking graph walks to remove noise areas. At the same time, the dynamic learning rate and weight decay strategy are combined to ensure model convergence and generalization, thereby fully realizing high-precision segmentation of complex geological reservoir fracture networks.

[0095] Combining the method proposed in this case, a geological reservoir fracture segmentation system is proposed that integrates the 3DU-Net graph attention mechanism and graph connectivity detection. It consists of several modules, including:

[0096] Multimodal data acquisition and preprocessing module: This module is responsible for collecting and preprocessing seismic 3D volume data, well logging curves, and core CT scan images. It also uses affine transformation to adjust the vertical depth and spatial position between different data to align the modal data. It also uses B-spline elastic registration to compensate for the nonlinear deformation caused by stratum fault movement and outputs a 3D fused voxel matrix containing multimodal numerical features.

[0097] The feature extraction and graph attention module inputs the preprocessed 3D fused voxel matrix into the improved 3D U-Net and extracts multi-scale features through multi-layer 3D convolution. It also uses the graph attention mechanism to enhance the boundary response between cracks and background, introduces a spatial attention mechanism to capture edge and texture features within the local neighborhood, and introduces a channel attention mechanism to dynamically weight the contributions of different modalities and suppress noise.

[0098] The graph-theoretic connectivity modeling module uses threshold segmentation to extract crack centerlines based on the multi-scale feature map output from the encoder, and constructs a crack connectivity topology map using centerline voxels as graph nodes. It uses Lie algebraic path integrals and curvature constraints to calculate reasonable connection distances between nodes, and applies topological constraints (such as the Nijenhuis operator) to block abnormal connections, ensuring that the constructed graph structure truly reflects the crack network.

[0099] The fracture network segmentation and connectivity optimization module uses a decoder to restore the multi-layer feature map to its original resolution, generates an initial fracture network probability map, and optimizes the segmentation results by introducing connectivity constraints and a Dice-Focal joint loss function. The connectivity constraint loss function ensures that the predicted probabilities of connected nodes in the graph are consistent and conform to geological seepage laws. The Dice-Focal joint loss is introduced to balance the class imbalance problem and enhance edge sensitivity.

[0100] The non-backtracking graph walk post-processing module performs post-processing on the initial probability map by constructing a hypergraph and performing non-backtracking random walks to remove isolated noise areas and retain voxels connected to the main crack network.

[0101] Dynamic training and multimodal validation module, which is used to manage the model training process, including dynamic learning rate adjustment and weight decay, to ensure rapid convergence of the model in the early stage and fine-tuning of parameters in the later stage; at the same time, cross-validation is used to fuse multimodal prediction results to evaluate model performance.

[0102] Use exponential decay learning rate strategy and L2 regularization: used to prevent overfitting and weight bias.

[0103] Four-fold cross validation was used to improve model robustness and aggregate prediction results.

[0104] Based on the method of the present invention, a terminal device is proposed. The terminal device may include: a processor, a storage medium and a bus. The storage medium stores machine-readable instructions executable by the processor. When the terminal device is running, the processor and the storage medium communicate through the bus, and the processor executes the machine-readable instructions to perform the steps of the deep learning model training method described in the aforementioned embodiment.

[0105] Based on the method of the present invention, a storage medium is proposed, which stores a computer program. When the computer program is executed by a processor, the steps of the above method are executed.

[0106] Based on the method of the present invention, a computer program product is proposed, including a computer program, and the computer program executes the method described above when a processor is used.

[0107] The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A geological reservoir fracture segmentation method that integrates 3DU-Net graph attention mechanism and graph connectivity detection, characterized by: The specific steps include: Step 1: Multimodal data integration and alignment. This step uses data from different acquisition devices to achieve spatial and depth correction between the data through affine transformation and elastic registration methods, respectively. This generates a three-dimensional fused voxel matrix containing multimodal numerical features, achieving unified data registration and feature alignment. Step 2: Deep feature extraction and graph attention enhancement. In this step, the 3D fused voxel matrix obtained in step 1 is input into the improved 3D U-Net model. Multi-scale features are extracted through multi-layer 3D convolution and a graph attention mechanism is embedded. Step 3: Graph-theoretic connectivity topology modeling. This step is based on the multi-level feature graph output from step 2. Crack centerlines are extracted through threshold segmentation and Gaussian smoothing. Centerline voxels are abstracted as graph nodes. Lie algebraic path integrals and curvature constraints are used to calculate the distance between nodes and the rationality of connections. This constructs a connectivity topology graph that reflects the true structure of the crack network. Step 4: Initial segmentation and connectivity constraint optimization. In this step, the decoder performs deconvolution on the multi-scale feature map to restore the original resolution and generate an initial fracture network probability map. The connectivity constraint loss function and the Dice-Focal joint loss are introduced to optimize the initial segmentation result to ensure that the segmentation result conforms to the geological seepage law and fracture network connectivity requirements. Step 5: Non-backtracking graph walk post-processing. This step performs post-processing on the initial segmentation results. By constructing a hypergraph based on spatial proximity and feature similarity and using a non-backtracking random walk strategy, isolated noise regions are removed and only voxels connected to the main crack network are retained, thereby further improving the segmentation accuracy. Step 6: Dynamic training strategy and regularization control. In this step, an exponentially decaying learning rate strategy and L2 regularization are used during the model training phase to ensure rapid convergence in the early stages and fine-tuning of parameters in the later stages. At the same time, weight decay is applied to the attention module to prevent the model from over-relying on local or single modalities and improve overall generalization capabilities. Step 7: Multimodal cross-validation and result fusion. This step evaluates the model performance through four-fold cross-validation, integrates the prediction results under different modes, and finally calculates and optimizes the average Dice coefficient and sensitivity index to verify and ensure the robustness and high-precision performance of the model in the segmentation of complex geological reservoir fracture networks.

2. The method according to claim 1, characterized in that The graph attention mechanism in step 2 includes spatial attention and channel attention. Spatial attention is used to capture edge and texture details in the local neighborhood, and channel attention is used to assign dynamic weights to different modalities.

3. The method according to claim 1, characterized in that Step three involves eliminating abnormal connections using topological constraints.

4. The method according to claim 1, wherein The connectivity constraint loss function is used to ensure that the probabilities of connected nodes in the graph are consistent.

5. A geological reservoir fracture segmentation system that integrates the 3DU-Net graph attention mechanism and graph connectivity detection, characterized by: Specifically include: Multimodal data acquisition and preprocessing module, which is responsible for collecting and preprocessing seismic 3D volume data, well logging curves, and core CT scan images. It also uses affine transformation to adjust the vertical depth and spatial position between different data to align the modal data. It also uses B-spline elastic registration to compensate for nonlinear deformation caused by formation fault movement and outputs a 3D fused voxel matrix containing multimodal numerical features. The feature extraction and graph attention module inputs the preprocessed 3D fused voxel matrix into the improved 3DU-Net and extracts multi-scale features through multi-layer 3D convolution. It also uses a graph attention mechanism to enhance the boundary response between cracks and background, introduces a spatial attention mechanism to capture edge and texture features within a local neighborhood, and introduces a channel attention mechanism to dynamically weight the contributions of different modalities and suppress noise. A graph-theoretic connectivity modeling module, based on the multi-scale feature map output from the encoder, uses threshold segmentation to extract crack centerlines and constructs a crack connectivity topology graph using centerline voxels as graph nodes. Using Lie algebraic path integrals and curvature constraints, it calculates reasonable connection distances between nodes and applies topological constraints to block abnormal connections, ensuring that the constructed graph structure truly reflects the crack network. The fracture network segmentation and connectivity optimization module uses a decoder to restore the multi-layer feature map to its original resolution, generating an initial fracture network probability map. It then optimizes the segmentation results by introducing connectivity constraints and a joint Dice-Focal loss function. The connectivity constraint loss function ensures that the predicted probabilities of connected nodes in the map are consistent and conform to geological seepage patterns. The Dice-Focal joint loss is introduced to balance class imbalance and enhance edge sensitivity. The non-backtracking graph walk post-processing module performs post-processing on the initial probability map by constructing a hypergraph and performing non-backtracking random walks to remove isolated noise regions and retain voxels connected to the main fracture network. Dynamic training and multimodal validation module, which is used to manage the model training process, including dynamic learning rate adjustment and weight decay, to ensure rapid convergence of the model in the early stage and fine-tuning of parameters in the later stage; at the same time, cross-validation is used to fuse multimodal prediction results to evaluate model performance.

6. A terminal device, characterized in that: include: A processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the terminal device is running, the processor and the storage medium communicate via the bus, and the processor executes the machine-readable instructions to perform the steps of the method according to any one of claims 1 to 4.

7. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, executes the steps of the method according to any one of claims 1 to 4.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the computer program implements the method according to any one of claims 1 to 4.

Images