A 3D tomographic segmentation method based on local-global feature enhancement and geometric constraints

Abstract

This invention relates to a 3D fault segmentation method based on local-global feature enhancement and geometric constraints, belonging to the fields of geophysical exploration and artificial intelligence. The method includes: standardizing and preprocessing the original 3D seismic data volume and cropping it into local sub-blocks; constructing an encoder-decoder network with 3DU-Net as the backbone, where the encoder extracts deep semantic features through 3D convolutional blocks, and the terminal layer aggregates multi-scale geological context information via MC-Bridge; the decoder embeds LG-Module and ACO-Module at each level; training the model using a structured loss function including binary cross-entropy loss, similarity coefficient loss, edge smoothing loss, and geometric constraint loss; and generating a fault prediction probability map by convolution and Softmax activation of the decoded enhanced features. This invention balances local texture details and global long-range dependencies, explicitly constrains fault geometry, and improves segmentation accuracy and geological plausibility.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration and artificial intelligence technology, specifically relating to a three-dimensional fault segmentation method based on local-global feature enhancement and geometric constraints. Background Technology

[0002] Faults, as widely developed geological discontinuities in the Earth's crust, are key structural elements controlling the migration of underground fluids and the accumulation of oil and gas. In the field of oil and gas exploration and development, accurate identification of fault structures is crucial for reducing exploration risks. Traditional fault interpretation methods mainly rely on manual picking on seismic profiles. This method is not only time-consuming and labor-intensive, but also heavily influenced by the subjective experience of the interpreters, resulting in poor consistency of results. With technological advancements, semi-automatic identification methods based on seismic attributes such as coherence and curvature have been proposed. However, these methods often face multiple interpretation problems in areas with complex geological conditions or strong noise interference, and struggle to accurately characterize small faults and diffuse fracture zones.

[0003] In recent years, deep learning technologies, represented by convolutional neural networks (CNNs), have made significant progress in fault identification. However, due to the massive volume of 3D seismic data (typically reaching TB levels), existing computing resources are insufficient to support the input of the entire data volume, leading to the widespread adoption of a block-based parallel processing strategy. While this block-based processing paradigm solves the memory limitation problem, it disrupts the spatial context between data blocks, resulting in a lack of global consistency in the model's prediction of adjacent sub-block boundaries. This leads to noticeable stitching artifacts and fragment noise at the stitching points. Furthermore, existing 3D segmentation networks (such as 3D U-Net) typically rely solely on voxel-level loss functions for optimization, lacking explicit modeling of the inherent geometric smoothness and topological continuity of the geological surface of faults. This results in predictions often exhibiting fractures, voids, rough edges, and incomplete topological structures, failing to meet the demands of high-precision industrial interpretation. Summary of the Invention

[0004] The purpose of this invention is to address the problems in existing large-scale three-dimensional seismic fault intelligent identification, such as cross-block inconsistency and severe stitching artifacts caused by block-based parallel processing, as well as fault topological fractures, blurred edges, and poor continuity caused by the lack of geometric constraints. This invention provides a three-dimensional fault segmentation method based on local-global feature enhancement and geometric constraints, which can simultaneously take into account local texture details and global long-range dependencies, and can explicitly constrain fault geometry.

[0005] The method includes:

[0006] S1. Perform Z-score normalization preprocessing on the original 3D seismic data volume, and then cut the preprocessed data into local sub-blocks as network input; S2. Using local sub-blocks as network input, an encoder-decoder network with 3D U-Net as the backbone is constructed. The encoder extracts deep semantic features through multi-level 3D convolutional blocks, and a multi-scale context bridge MC-Bridge is implanted at the end of the encoder to aggregate multi-scale geological context information. The decoder embeds a local-global feature joint enhancement module LG-Module and an attention-guided continuity optimization module ACO-Module at each level to perform local detail and global semantic fusion, fault topological continuity repair, and output decoded enhanced features. S3. Construct a structured loss function to train the encoder-decoder network with 3D U-Net as the backbone. The structured loss function includes binary cross-entropy loss, similarity coefficient loss, edge smoothing loss and geometric constraint loss. S4. The decoded and enhanced features are then processed by a 1×1×1 convolution and Softmax activation in the output layer to generate the final tomographic prediction probability map.

[0007] Furthermore, the encoder described in step S2 extracts deep semantic features through multi-level three-dimensional convolutional blocks, and the specific steps for embedding a multi-scale context bridge (MC-Bridge) at the encoder end to aggregate multi-scale geological context information are as follows: S201. The encoder extracts deep semantic features from the preprocessed data through progressive downsampling using multi-level 3D convolutional blocks. The formula is as follows: ; in, For encoder number i The deep semantic feature map output by the hierarchy. i For hierarchical indexes; This is a 2×2×2 pooling operation with a step size of 2, used to downsample the feature map to expand the receptive field and reduce the computational cost; The basic convolutional block operation consists of two consecutively stacked units, each containing a 3×3×3 three-dimensional convolutional layer, batch normalization, and a ReLU activation function to extract nonlinear local spatial features. S202. A multi-scale context bridge (MC-Bridge) is implanted at the end of the encoder. Taking the deep semantic features at the encoder end as input, multi-scale geological context information is captured through parallel multi-dilation rate dilated convolutions to generate global deep features. The formula is: ; in, This is a global deep feature that aggregates multi-scale contextual information. This is a hybrid operation function that includes dilated convolutions with different dilation rates and global average pooling.

[0008] Furthermore, in step S2, the decoder sequentially embeds a local-global feature joint enhancement module (LG-Module) and an attention-guided continuity optimization module (ACO-Module) at each layer to perform local detail and global semantic fusion and fault topological continuity repair. The specific steps are as follows: S211. Enter the decoder path on the right, and restore the voxel space resolution step by step. At each decoding level, the deep semantic features are upsampled by trilinear interpolation and then concatenated with the shallow features of the encoder at the same layer along the channel dimension through skip connections to compensate for the high-frequency geometric loss caused by downsampling. The concatenation result is completed by 1×1×1 convolution to achieve channel dimensionality reduction and cross-semantic layer feature fusion. ; in, For the first i+ Deep features output by the Level 1 decoder For trilinear interpolation upsampling, For feature stitching along the channel dimension, For channel dimensionality reduction and feature fusion, a 1×1×1 convolution is used. For encoder number i Shallow features extracted hierarchically; S212. At each decoding level, a Local-Global Feature Joint Enhancement Module (LG-Module) is further constructed. This module captures local high-frequency textures and global channel responses in parallel through a dual-branch structure, as shown in the formula: ; in, As input fusion features, This is a composite mapping function that integrates local residual extraction and global attention calibration. For refined blended features of the output; S213. Construct an attention-guided continuity optimization module (ACO-Module) at each decoding level. Establish cross-voxel long-range dependencies using 3D window self-attention, and combine it with a feature continuity optimization unit (FCO-Unit). This is then fused with the fine features output by the LG-Module to repair the spatial geometry of the fault lines. The formula is as follows: ; in, Window attention operations for capturing long-range topological dependencies, This is a feature continuity optimization function used to nonlinearly fuse the local details provided by the LG-Module with the topology provided by the WA, outputting the final decoded enhanced features. .

[0009] Further, in step S211, the encoder's first...i The process of extracting shallow features in hierarchical extraction is represented as follows: ; Where X is the original input data volume. This is a max-pooling downsampling operation with a step size of 2. It is a dual convolution module that includes two consecutive convolutions, batch normalization, and ReLU activation.

[0010] Furthermore, the specific steps of step S212 are as follows: S2121. Construct a local residual convolution branch that directly acts on the input features. By mining geometric features within a local neighborhood through cascaded convolutional operations, a 3×3×3 convolutional layer is first used to capture fine edge structures in three-dimensional space, and then a 1×1×1 convolutional layer is used for cross-channel linear projection and feature integration.

[0011] in, and These are the weight parameters for 3×3×3 and 1×1×1 convolutional layers, respectively; S2122. Introduce a global branch that includes channel attention and spatial attention. In the channel dimension, global average pooling is used to compress spatial information, and the non-linear dependencies between channels are explicitly modeled through two cascaded 1×1×1 convolutional layers to generate channel weights. : ; in, and These are the ReLU and Sigmoid activation functions, respectively. S2123. In the spatial dimension, the same cascaded 1×1×1 convolutional structure is directly applied to the feature map, gradually mapping the high-dimensional features into pixel-level spatial weights. : ; S2124. The channel attention and spatial attention are added element-wise, and then the features are smoothed and reorganized through a residual block containing 3×3×3 convolutions and 1×1×1 convolutions. Finally, the refined hybrid features are fused with the output of the local residual branch. .

[0012] Furthermore, the specific steps of step S213 are as follows: S2131.ACO-Module first introduces a window-based self-attention mechanism (WA), which divides the 3D feature volume into non-overlapping local cube windows, and then flattens the voxels in each window and maps them into query, key, and value vectors. S2132. Within each independent window, a multi-head self-attention mechanism is used to calculate the semantic association matrix between voxels, inferring the potential extension trajectory of the fault within the local region. The calculation formula for the multi-head self-attention is as follows: ; in, These are the query, key, and value matrices generated by linear projection of the voxel features within the current window. The feature dimension of the key vector, and the denominator. This is a scaling factor used to adjust the magnitude of the dot product value; This is a matrix transpose operation used to calculate the similarity score between voxels; This represents the learnable relative position offset.

[0013] S2133. Input features Divide into N non-overlapping windows For each window Using linear projection matrix , , Generate the corresponding queries, keys, and values, and perform multi-head attention computation. Then, restore all enhanced windows to their original spatial positions through the inverse operation: ; in, This is the inverse operation to reshape the block features back into a 3D tensor. This is a splicing operation along a spatial dimension; Features that provide global information; S2134. The ACO-Module calls the Feature Continuity Optimization Unit (FCO-Unit) and fuses the components using an element-wise addition method. and The system performs semantic complementation between local texture and global topology. The fused features are then fed into the deep residual refinement module within the FCO-Unit. This module consists of two stacked 3×3×3 convolutional layers and ReLU activation. The FCO-Unit uses a sliding overlap mechanism to break through the WA window boundaries, performing continuous kernel weighting on the tomographic fragments in physical space to obtain the final decoded enhanced features. : ; in, For convolution weights, This is the ReLU activation function.

[0014] Furthermore, the structured loss function in step S3 Includes four loss items: ; in, and This constitutes the basic segmentation loss. For binary cross-entropy loss, For Dice similarity coefficient loss; and It is a structured regularization term that constrains the fault geometry and edge quality. For edge smoothing loss, For geometric constraint loss; and It is used for balance and Hyperparameters of loss contribution.

[0015] Furthermore, the structured loss function middle: ; in, For the i-th prediction probability map, This is the i-th real label; ; in, For smoothing terms; ; in, and The weighting coefficients and edge consistency are calculated using a custom 3D Sobel operator to determine the predicted probability map. and real labels 3D gradient magnitude and And by minimizing and The L1 distance between them forces the learning of accurate boundary positions, and local smoothing is achieved using the three Laplacian operators. Convolve the predicted probability map, calculate and penalize the mean square value of the second derivative of the predicted surface. For predicting probability graphs P exist v The second derivative at that point; ; in, It is a line template dynamically generated based on geometric properties such as aspect ratio and area.B For training batch size, For the first b A predicted probability plot for each sample.

[0016] Furthermore, the formula for generating the final fault prediction probability map in step S4 is as follows: ; in, For output layer convolution, The Softmax activation function is used. This is the final fault prediction probability map.

[0017] As can be seen from the above technical solutions, the present invention has the following advantages: This invention improves the accuracy and completeness of fault segmentation by introducing a local-global feature joint enhancement module. While preserving high-frequency details at fault edges, it effectively suppresses background noise, making the model more robust in low signal-to-noise ratio data. It also solves the stitching artifact problem caused by block processing by establishing long-range dependencies across voxels through an attention-guided continuity optimization module. This explicitly stitches together block boundaries at the feature level, ensuring seamless continuity of the predicted fault surface at arbitrary sub-cube boundaries, eliminating the slice-like artifacts and fragment noise common in traditional methods. Furthermore, it enhances the geological rationality of fault interpretation by introducing a structured loss function, upgrading simple numerical fitting to constraints on geometric morphology and topology. This forces the model to output predictions with sharp edges, smooth surfaces, and topological closure, enabling accurate identification and tracking of micro-faults and hidden fractures, providing a high-fidelity three-dimensional fault skeleton for geological modeling. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints of the present invention.

[0020] Figure 2 This is a schematic diagram of the overall network architecture of the three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints of the present invention, wherein, Figure 2 (a) in the diagram represents the overall flow of GeoLG-3DFaultNet. Figure 2 (b) in the diagram represents the architecture of MC-Bridge. Figure 2(c) in the diagram represents the architecture of the LG-Module. Figure 2 The architecture of (d)ACO-Module in the text.

[0021] Figure 3 This is a schematic diagram comparing the convergence curves of the loss function Loss as a function of Epoch during the training process of the method of this invention and the comparison model.

[0022] Figure 4 This is a schematic diagram comparing the two-dimensional slice segmentation results of the method of this invention with those of existing mainstream models on a synthetic dataset. Figure 4 (a) in the image is a Seismic image. Figure 4 (b) in the text represents the ground truth. Figure 4 (c) in the text refers to a 3D U-Net. Figure 4 (d) in the text represents MA3DSeg. Figure 4 (e) in the text refers to nnFormer. Figure 4 (f) in the text refers to SwinUNETR. Figure 4 (g) in the text refers to TransUNet. Figure 4 (h) in the text stands for UNETR++. Figure 4 (i) in the text refers to the present invention.

[0023] Figure 5 This is a schematic diagram comparing the predicted fault boundary morphology of the method and the comparison model in three-dimensional space, wherein... Figure 5 (a) in the image is a Seismic image. Figure 5 (b) in the text represents the ground truth. Figure 5 (c) in the text refers to a 3D U-Net. Figure 5 (d) in the text represents MA3DSeg. Figure 5 (e) in the text refers to nnFormer. Figure 5 (f) in the text refers to SwinUNETR. Figure 5 (g) in the text refers to TransUNet. Figure 5 (h) in the text stands for UNETR++. Figure 5 (i) in the text refers to the present invention.

[0024] Figure 6 This is a schematic diagram showing the fault segmentation results of the method of the present invention on real field seismic data in the F3 work area of ​​the Netherlands, and its comparison with the comparative model. Figure 6 (a) in the image represents a 3D U-Net. Figure 6 (b) in the text is MA3DSeg. Figure 6 (c) in the text refers to SwinUNETR. Figure 6 (d) in the text refers to the present invention. Detailed Implementation

[0025] The various embodiments of the invention will be described more fully in the detailed steps of the three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints described below. The invention may have various embodiments, and adjustments and modifications may be made therein. However, it should be understood that there is no intention to limit the various embodiments of the invention to the specific embodiments disclosed herein, but rather the invention should be understood to cover all modifications, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of the invention.

[0026] It should be understood that, when used in this specification, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0027] The terms "one embodiment" or "some embodiments" used in this invention mean that one or more embodiments of the invention include the specific features, structures, or characteristics described in that embodiment. Therefore, the terms "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of the invention do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0028] To make the objectives, features, and advantages of this invention more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solutions protected by this invention. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0029] Please see Figure 1 The diagram shows a flowchart of a 3D tomographic segmentation method based on local-global feature enhancement and geometric constraints. The method includes the following steps: S1. Perform Z-score normalization preprocessing on the original 3D seismic data volume, and then cut the preprocessed data into local sub-blocks as network input; S2. Using local sub-blocks as network input, an encoder-decoder network with 3D U-Net as the backbone is constructed. The encoder extracts deep semantic features through multi-level 3D convolutional blocks, and a multi-scale context bridge MC-Bridge is implanted at the end of the encoder to aggregate multi-scale geological context information. The decoder embeds a local-global feature joint enhancement module LG-Module and an attention-guided continuity optimization module ACO-Module at each level to perform local detail and global semantic fusion, fault topological continuity repair, and output decoded enhanced features. S3. Construct a structured loss function to train the encoder-decoder network with 3D U-Net as the backbone. The structured loss function includes binary cross-entropy loss, similarity coefficient loss, edge smoothing loss and geometric constraint loss. S4. The decoded and enhanced features are then processed by a 1×1×1 convolution and Softmax activation in the output layer to generate the final tomographic prediction probability map.

[0030] As a refinement and extension of the specific implementation methods described above, and to fully illustrate the specific implementation process in this embodiment, this method addresses the common challenges of cross-block inconsistency, blurred fault boundaries, and missing topological structures that are prevalent in block-based inference methods for 3D fault segmentation. It constructs a three-dimensional fault segmentation network (GeoLG-3DFaultNet) based on local-global feature enhancement and geometric constraints. By establishing cross-block consistency regularization and embedding long-range geological constraints into feature learning, the model maintains the continuity of fault surface space and topology at block boundaries. This includes three strategies: a local-global feature joint enhancement module for achieving detailed refinement and global consistency modeling; a continuity optimization module for suppressing fragment noise and artifacts; and geometric constraints and edge smoothing losses. GeoLG-3DFaultNet uses the classic 3D U-Net architecture as its backbone, structurally reshaping faults in three-dimensional space to address geological structural characteristics such as sparsity and long range, achieving geological voxel-level topological integrity and cross-block consistency. The overall network architecture is as follows: Figure 2 As shown, the given network input is a preprocessed 3D seismic data volume. The encoder consists of a series of 3D convolutional blocks, which are responsible for extracting deep semantic features from the raw volume data through stepwise downsampling.

[0031] The method includes the following steps: S1. Perform Z-score normalization preprocessing on the original 3D seismic data volume, and then cut the preprocessed data into local sub-blocks as network input; S2. Using local sub-blocks as network input, an encoder-decoder network with 3D U-Net as the backbone is constructed. The encoder extracts deep semantic features through multi-level 3D convolutional blocks, and a multi-scale context bridge MC-Bridge is implanted at the end of the encoder to aggregate multi-scale geological context information. The decoder embeds a local-global feature joint enhancement module LG-Module and an attention-guided continuity optimization module ACO-Module at each level to perform local detail and global semantic fusion, fault topological continuity repair, and output decoded enhanced features. The encoder described in step S2 extracts deep semantic features through multi-level three-dimensional convolutional blocks, and the specific steps for embedding a multi-scale context bridge (MC-Bridge) at the encoder end to aggregate multi-scale geological context information are as follows: S201. The encoder extracts deep semantic features from the preprocessed data through progressive downsampling using multi-level 3D convolutional blocks. The formula is as follows: ; in, For encoder number i The deep semantic feature map output by the hierarchy. i For hierarchical indexes; This is a 2×2×2 pooling operation with a step size of 2, used to downsample the feature map to expand the receptive field and reduce the computational cost; The basic convolutional block operation consists of two consecutively stacked units, each containing a 3×3×3 three-dimensional convolutional layer, batch normalization, and a ReLU activation function to extract nonlinear local spatial features. S202. A multi-scale context bridge (MC-Bridge) is implanted at the end of the encoder. Taking the deep semantic features at the encoder end as input, multi-scale geological context information is captured through parallel multi-dilation rate dilated convolutions to generate global deep features. The formula is: ; in, This is a global deep feature that aggregates multi-scale contextual information. This is a hybrid operation function that includes dilated convolutions with different dilation rates and global average pooling.

[0032] In step S2, the decoder sequentially embeds a Local-Global Feature Joint Enhancement Module (LG-Module) and an Attention-Guided Continuity Optimization Module (ACO-Module) at each layer to perform local detail and global semantic fusion and fault topological continuity repair. The specific steps are as follows: S211. Enter the decoder path on the right, and restore the voxel space resolution step by step. At each decoding level, the deep semantic features are upsampled by trilinear interpolation and then concatenated with the shallow features of the encoder at the same layer along the channel dimension through skip connections to compensate for the high-frequency geometric loss caused by downsampling. The concatenation result is completed by 1×1×1 convolution to achieve channel dimensionality reduction and cross-semantic layer feature fusion. ; in, For the first i+ Deep features output by the Level 1 decoder For trilinear interpolation upsampling, For feature stitching along the channel dimension, For channel dimensionality reduction and feature fusion, a 1×1×1 convolution is used. For encoder number i Shallow features extracted hierarchically; In step S211, the encoder's first... i The process of extracting shallow features in hierarchical extraction is represented as follows: ; Where X is the original input data volume. This is a max-pooling downsampling operation with a step size of 2. It is a dual convolution module that includes two consecutive convolutions, batch normalization, and ReLU activation.

[0033] S212. At each decoding level, a Local-Global Feature Joint Enhancement Module (LG-Module) is further constructed. This module captures local high-frequency textures and global channel responses in parallel through a dual-branch structure, as shown in the formula: ; in, As input fusion features, This is a composite mapping function that integrates local residual extraction and global attention calibration. For refined blended features of the output; The specific steps of step S212 are as follows: S2121. Construct a local residual convolution branch that directly acts on the input features. By mining geometric features within a local neighborhood through cascaded convolutional operations, a 3×3×3 convolutional layer is first used to capture fine edge structures in three-dimensional space, and then a 1×1×1 convolutional layer is used for cross-channel linear projection and feature integration.

[0034] in, and These are the weight parameters for 3×3×3 and 1×1×1 convolutional layers, respectively; S2122. Introduce a global branch that includes channel attention and spatial attention. In the channel dimension, global average pooling is used to compress spatial information, and the non-linear dependencies between channels are explicitly modeled through two cascaded 1×1×1 convolutional layers to generate channel weights. : ; in, and These are the ReLU and Sigmoid activation functions, respectively. S2123. In the spatial dimension, the same cascaded 1×1×1 convolutional structure is directly applied to the feature map, gradually mapping the high-dimensional features into pixel-level spatial weights. : ; S2124. The channel attention and spatial attention are added element-wise, and then the features are smoothed and reorganized through a residual block containing 3×3×3 convolutions and 1×1×1 convolutions. Finally, the refined hybrid features are fused with the output of the local residual branch. .

[0035] It should be noted that in the decoding and reconstruction stage of 3D seismic data, although skip connections introduce shallow features to supplement spatial details, simple channel stitching is insufficient to effectively bridge the feature gap between deep semantics and shallow details. Fault targets often possess the dual characteristics of weak local texture and large global span. Conventional convolutional operations are limited by fixed local receptive fields, making it difficult to accurately locate minute edges amidst complex background noise, and also unable to adaptively establish long-range semantic dependencies between channels. This results in significant blurring and artifacts remaining after feature fusion. Therefore, a local-global feature joint enhancement module, LG-Module, was constructed. This module operates on the same voxel tensor through a dual-branch structure: the local branch uses cascaded convolution to model fine-grained geometric gradients, while the global branch models long-range geological context through channel-space self-attention. The outputs of the two branches are fused using learnable weighted fusion, simultaneously suppressing background noise and enhancing the saliency of fault normal discontinuities within the feature domain, achieving isomorphic mapping between deep and shallow semantics.

[0036] S213. Construct an attention-guided continuity optimization module (ACO-Module) at each decoding level. Establish cross-voxel long-range dependencies using 3D window self-attention, and combine it with a feature continuity optimization unit (FCO-Unit). This is then fused with the fine features output by the LG-Module to repair the spatial geometry of the fault lines. The formula is as follows: ; in, Window attention operations for capturing long-range topological dependencies, This is a feature continuity optimization function used to nonlinearly fuse the local details provided by the LG-Module with the topology provided by the WA, outputting the final decoded enhanced features. .

[0037] The specific steps of step S213 are as follows: S2131.ACO-Module first introduces a window-based self-attention mechanism (WA), which divides the 3D feature volume into non-overlapping local cube windows, and then flattens the voxels in each window and maps them into query, key, and value vectors. S2132. Within each independent window, a multi-head self-attention mechanism is used to calculate the semantic association matrix between voxels, inferring the potential extension trajectory of the fault within the local region. The calculation formula for the multi-head self-attention is as follows: ; in, These are the query, key, and value matrices generated by linear projection of the voxel features within the current window. The feature dimension of the key vector, and the denominator. This is a scaling factor used to adjust the magnitude of the dot product value; This is a matrix transpose operation used to calculate the similarity score between voxels; This represents the learnable relative position offset.

[0038] It should be noted that, since self-attention operations themselves possess permutation invariance (i.e., shuffling the input order does not affect the output), this bias term is introduced to allow the network to perceive the relative distance and orientation of faults in three-dimensional space. B This means that the attention score depends not only on the similarity of the content, but also on the spatial geometric relationship between voxels.

[0039] S2133. Input features Divide into N non-overlapping windows For each window Using linear projection matrix , , Generate the corresponding queries, keys, and values, and perform multi-head attention computation. Then, restore all enhanced windows to their original spatial positions through the inverse operation: ; in, This is the inverse operation to reshape the block features back into a 3D tensor. This is a splicing operation along a spatial dimension; Features that provide global information; It should be noted that the resulting It not only preserves local details, but also establishes long-range semantic relationships between voxels within the window.

[0040] S2134. To address the potential boundary truncation effect introduced by non-overlapping window partitioning, the ACO-Module invokes the Feature Continuity Optimization Unit (FCO-Unit) and fuses elements by addition. and The system performs semantic complementation between local texture and global topology. The fused features are then fed into the deep residual refinement module within the FCO-Unit. This module consists of two stacked 3×3×3 convolutional layers and ReLU activation. The FCO-Unit uses a sliding overlap mechanism to break through the WA window boundaries, performing continuous kernel weighting on the tomographic fragments in physical space to obtain the final decoded enhanced features. : ; in, For convolution weights, This is the ReLU activation function.

[0041] It should be noted that, considering the unique elongated extension and geometric continuity of 3D seismic faults, this embodiment further constructs an attention-guided continuity optimization module, ACO-Module, after LG-Module. This module adopts a long-range capture-local stitching collaborative strategy, and is coupled with a window-based multi-head attention (WA) branch and a feature continuity optimization unit (FCO-Unit). It establishes global context associations across windows and forcibly repairs topological fractures that may be caused by block processing.

[0042] By deeply integrating attention guidance and convolutional stitching through FCO-Unit, ACO-Module effectively overcomes the spatial fracturing problem in complex geological backgrounds, ensuring that the output features maintain sharp edges while achieving high topological continuity. ACO-Module achieves seamless coupling between global context and local texture across the entire image, ensuring that the network can output topologically complete and sharp-edged fault prediction results when faced with strong background noise and large-scale structural spans.

[0043] S3. Construct a structured loss function to train the encoder-decoder network with 3D U-Net as the backbone. The structured loss function includes binary cross-entropy loss, similarity coefficient loss, edge smoothing loss and geometric constraint loss. It should be noted that the loss function, as a supervisory signal, not only constrains the pixel-level classification accuracy, but also forces the network to approximate the real fault geological structure in terms of macroscopic morphology and microscopic boundaries, thereby achieving end-to-end morphological perception optimization.

[0044] The structured loss function in step S3 Includes four loss items: ; in, and This constitutes the basic segmentation loss. For binary cross-entropy loss, For Dice similarity coefficient loss; and It is a structured regularization term that constrains the fault geometry and edge quality. For edge smoothing loss, For geometric constraint loss; and It is used for balance and Hyperparameters of loss contribution.

[0045] It should be noted that the basic segmentation loss ( and The goal is to achieve pixel-level classification accuracy. Binary cross-entropy loss ( The Dice similarity coefficient loss measures the difference in information entropy between the predicted probability and the true label. It encourages the model to push the predictions towards the extremes of 0 or 1, thereby accelerating network convergence. This was used to address the class imbalance problem commonly found in fault identification tasks, since fault voxels constitute a very small percentage of the entire seismic body. By directly optimizing the overlap between the predicted results and the true labels (F1 score), the model focuses more on the accurate identification of the minority class (tomographic breaks). Edge smoothing loss ( This aims to suppress local high-frequency noise and jagged spikes at fault boundaries, improving the accuracy of fault interpretation. Geometric constraint loss ( This loss function aims to constrain the macroscopic geometry and extensibility of fault surfaces in 3D space, thereby enhancing the spatial continuity of faults. This loss function introduces dynamic geometric feature templates. Integrate prior knowledge of macroscopic morphology into the training process. The core mechanism is to penalize the extent to which the predicted result deviates from the ideal value of 1 in areas identified as conforming to fault geometry.

[0046] Structured loss function middle: ; in, For the i-th prediction probability map, This is the i-th real label; ; in, For smoothing terms; ; in, and The weighting coefficients and edge consistency are calculated using a custom 3D Sobel operator to determine the predicted probability map. and real labels 3D gradient magnitude and And by minimizing and The L1 distance between them forces the learning of accurate boundary positions, and local smoothing is achieved using the three Laplacian operators. Convolve the predicted probability map, calculate and penalize the mean square value of the second derivative of the predicted surface. For predicting probability graphs P exist v The second derivative at that point; ; in, It is a line template dynamically generated based on geometric properties such as aspect ratio and area. B For training batch size, For the first b A predicted probability plot for each sample.

[0047] S4. The decoded and enhanced features are then processed by a 1×1×1 convolution and Softmax activation in the output layer to generate the final tomographic prediction probability map.

[0048] The formula for generating the final fault prediction probability map in step S4 is: ; in, For output layer convolution, The Softmax activation function is used. This is the final fault prediction probability map.

[0049] It's important to note that the structured loss function upgrades the optimization objective of fault identification from a single voxel-level probability fitting to a multi-scale joint constraint encompassing geometric continuity, edge fidelity, and pixel accuracy. This hybrid supervision strategy breaks the statistical assumption of traditional loss functions that treat voxels as independent events. By introducing explicit morphological priors, it forces the network to follow the physical manifold distribution of the faults during gradient descent: the basic segmentation loss establishes the class discrimination benchmark under sparse samples, the edge smoothing loss corrects the jagged edges at the microscopic level, and the geometric constraint loss safeguards the topological connectivity at the macroscopic level. Thus, this function constructs a rigorous gradient-guided mechanism throughout the entire training cycle, ensuring that the GeoLG-3DFaultNet prediction results not only numerically approximate the true labels but also structurally strictly conform to the topological completeness and natural continuity of geological structures.

[0050] In some embodiments, to comprehensively evaluate the overall advantages of the present invention in 3D seismic fault segmentation tasks, this embodiment selects a series of mainstream 3D segmentation models for systematic comparison, including 3D U-Net, MA3DSeg, nnFormer, SwinUNETR, TransUNet, and UNETR++. To ensure the fairness of the comparison, all comparison models are implemented based on their official open-source code and strictly follow the hyperparameter configuration and training strategies recommended in the original literature to ensure that each model is in its optimal performance state. All models are trained using the same 200 pairs of 3D seismic data volumes and evaluated on 20 pairs of independent test data volumes. The prediction results are quantitatively analyzed along three orthogonal directions: Inline, Crossline, and Timeline, to ensure the statistical reliability of the evaluation results.

[0051] like Figure 3The figure shows the curves of the loss function of different models changing with Epoch during training. Experimental results show that the method of this invention (GeoLG-3DFaultNet, red curve in the figure) exhibits significantly better convergence characteristics and stability than the comparison models. Specifically, the comparison models generally have higher loss values ​​in the initial stage, while the method of this invention is at a significantly lower loss level from the first iteration and maintains the global minimum throughout the entire 300 training rounds, demonstrating a better convergence starting point and overall stability. This convergence advantage of low starting point and high stability is due to the unique feature enhancement architecture and geometric constraint strategy of this invention: LG-Module extracts local edge details and global semantic information simultaneously in the shallow network through a dual-stream parallel mechanism, providing a highly discriminative initial gradient for the optimization process; while edge smoothing and geometric topology constraints explicitly guide parameter updates throughout the training process, enabling the model to quickly enter the low-loss region and avoid large oscillations and long search times. Therefore, this invention can efficiently and stably fit the complex nonlinear mapping between seismic data and fault structures. In contrast, other comparative models converged more slowly overall, and their final loss values ​​were still significantly higher than those of this invention, reflecting their shortcomings in fault detail modeling and structural consistency learning.

[0052] Table 1. Comparison of different models on synthetic seismic datasets

[0053] This embodiment uses Dice coefficient, Intersection over Union (IoU), precision, recall, and F1 score as key evaluation metrics to quantitatively evaluate the segmentation performance of the proposed method (GeoLG-3DFaultNet) and comparative models. Table 1 shows the comparison results of different models on a synthetic seismic dataset. The proposed method achieved optimal values ​​across all evaluation metrics: Dice coefficient of 88.65%, IoU of 79.61%, precision of 88.05%, recall of 89.26%, and F1 score of 88.65%. Compared with traditional 3D U-Net and models such as SwinUNETR and TransUNet, the proposed method exhibits significant performance advantages. Specifically, although SwinUNETR achieves near-optimal performance in recall and Dice metrics, its IoU metric, which is sensitive to edge accuracy, is only 69.41%, approximately 10 percentage points lower than the proposed method. This indicates that existing global attention or dense convolution techniques are insufficient for effectively supervising fault features with few voxels and weak geometric topology. This invention achieves the highest Recall and Precision simultaneously, demonstrating that the model, under extremely unbalanced sample distribution, ensures zero missed detection of minute faults while successfully suppressing background noise, proving its robustness in mining fine-grained fault features and restoring topological structures.

[0054] To visually evaluate the tomographic recognition performance, test samples were selected and visualized in two dimensions along the Inline, Crossline, and Timeline directions. For example... Figure 4 As shown, the comparative model exhibits significant topological breaks or spurious adhesions in fault intersection zones and weak signal regions (marked by red boxes in the figure). This is due to a mismatch between the receptive field and semantic granularity, failing to maintain edge sharpness under low signal-to-noise ratio conditions. In contrast, the segmentation results of the method in this invention demonstrate significant topological continuity and geometric smoothness in three-dimensional space. The predicted fault lines are sharp and clear, with complete closure at fault intersections, highly consistent with the true labels. Experimental results show that the model in this invention achieves an optimal balance in terms of edge sharpness preservation, long-range structural coherence, and background noise suppression.

[0055] A comparative analysis was conducted on the prediction results of three-dimensional fault planes in local areas. For example... Figure 5As shown in the green circle (marked area), existing models exhibit significant topological breaks and surface jaggedness in 3D space, resulting in discontinuous gaps and blurred edge artifacts on the fault planes. In contrast, the 3D fault planes segmented by this invention exhibit spatial smoothness and morphological continuity, effectively eliminating splicing marks and anisotropic artifacts caused by traditional block-based reasoning methods. Furthermore, it maintains natural transitions and structural consistency in complex geometric regions such as fault intersections and transitions. This demonstrates that this invention can collaboratively model multi-scale spatial dependencies, significantly improving the topological consistency of faults in 3D space.

[0056] To verify the model's generalization ability on complex field seismic data, the model of this invention, which was trained only on a synthetic dataset, was directly applied to the data of the F3 work area and tested with a comparison model.

[0057] This implementation applies the model to data from the Dutch F3 work area, which has a relatively complex geological structure. This dataset exhibits typical salt dome structures accompanied by a radial fault system, placing higher demands on the model's noise resistance and feature resolution capabilities. Figure 6 As shown, the comparative model is prone to producing scattered and patchy artifacts at the salt dome boundary and flanks, mistakenly identifying high-speed gradient changes as fractures. In contrast, the model of this invention segments a clear three-dimensional fault structure with a clean background, effectively suppressing structural artifacts around the salt dome and fully depicting the spatial distribution of radial faults.

[0058] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0059] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints, characterized in that, The method includes the following steps: S1. Perform Z-score normalization preprocessing on the original 3D seismic data volume, and then cut the preprocessed data into local sub-blocks as network input; S2. Using local sub-blocks as network input, construct an encoder-decoder network with 3D U-Net as the backbone. The encoder extracts deep semantic features through multi-level 3D convolutional blocks, and a multi-scale context bridge MC-Bridge is implanted at the end of the encoder to aggregate multi-scale geological context information. The decoder embeds a local-global feature joint enhancement module LG-Module and an attention-guided continuity optimization module ACO-Module at each level to perform local detail and global semantic fusion, fault topological continuity repair, and output decoded enhanced features. S3. Construct a structured loss function to train the encoder-decoder network with 3D U-Net as the backbone. The structured loss function includes binary cross-entropy loss, similarity coefficient loss, edge smoothing loss and geometric constraint loss. S4. The decoded and enhanced features are then processed by a 1×1×1 convolution and Softmax activation in the output layer to generate the final tomographic prediction probability map.

2. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 1, characterized in that, The encoder described in step S2 extracts deep semantic features through multi-level three-dimensional convolutional blocks, and the specific steps for embedding a multi-scale context bridge (MC-Bridge) at the encoder end to aggregate multi-scale geological context information are as follows: S201. The encoder extracts deep semantic features from the preprocessed data through progressive downsampling using multi-level 3D convolutional blocks. The formula is as follows: ； in, For encoder number i The deep semantic feature map output by the hierarchy. i For hierarchical indexes; This is a 2×2×2 pooling operation with a step size of 2, used to downsample the feature map to expand the receptive field and reduce the computational cost; The basic convolutional block operation consists of two consecutively stacked units, each containing a 3×3×3 three-dimensional convolutional layer, batch normalization, and a ReLU activation function to extract nonlinear local spatial features. S202. A multi-scale context bridge (MC-Bridge) is implanted at the end of the encoder. Taking the deep semantic features at the encoder end as input, multi-scale geological context information is captured through parallel multi-dilation rate dilated convolutions to generate global deep features. The formula is: ； in, This is a global deep feature that aggregates multi-scale contextual information. This is a hybrid operation function that includes dilated convolutions with different dilation rates and global average pooling.

3. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 1, characterized in that, In step S2, the decoder sequentially embeds a Local-Global Feature Joint Enhancement Module (LG-Module) and an Attention-Guided Continuity Optimization Module (ACO-Module) at each layer to perform local detail and global semantic fusion and fault topological continuity repair. The specific steps are as follows: S211. Enter the decoder path on the right, and restore the voxel space resolution step by step. At each decoding level, the deep semantic features are upsampled by trilinear interpolation and then concatenated with the shallow features of the encoder at the same layer along the channel dimension through skip connections to compensate for the high-frequency geometric loss caused by downsampling. The concatenation result is completed by 1×1×1 convolution to achieve channel dimensionality reduction and cross-semantic layer feature fusion. ； in, For the first i+ Deep features output by the Level 1 decoder For trilinear interpolation upsampling, For feature stitching along the channel dimension, For channel dimensionality reduction and feature fusion, a 1×1×1 convolution is used. For encoder number i Shallow features extracted hierarchically; S212. At each decoding level, a Local-Global Feature Joint Enhancement Module (LG-Module) is further constructed. This module captures local high-frequency textures and global channel responses in parallel through a dual-branch structure, as shown in the formula: ； in, As input fusion features, This is a composite mapping function that integrates local residual extraction and global attention calibration. For refined blended features of the output; S213. Construct an attention-guided continuity optimization module (ACO-Module) at each decoding level. Establish cross-voxel long-range dependencies using 3D window self-attention, and combine it with a feature continuity optimization unit (FCO-Unit). This is then fused with the fine features output by the LG-Module to repair the spatial geometry of the fault lines. The formula is as follows: ； in, Window attention operations for capturing long-range topological dependencies, This is a feature continuity optimization function used to nonlinearly fuse the local details provided by the LG-Module with the topology provided by the WA, outputting the final decoded enhanced features. .

4. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 3, characterized in that, In step S211, the encoder's first... i The process of extracting shallow features in hierarchical extraction is represented as follows: ； Where X is the original input data volume. This is a max-pooling downsampling operation with a step size of 2. It is a dual convolution module that includes two consecutive convolutions, batch normalization, and ReLU activation.

5. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 3, characterized in that, The specific steps of step S212 are as follows: S2121. Construct a local residual convolution branch that directly acts on the input features. By mining geometric features within a local neighborhood through cascaded convolutional operations, a 3×3×3 convolutional layer is first used to capture fine edge structures in three-dimensional space, and then a 1×1×1 convolutional layer is used for cross-channel linear projection and feature integration. in, and These are the weight parameters for 3×3×3 and 1×1×1 convolutional layers, respectively; S2122. Introduce a global branch that includes channel attention and spatial attention. In the channel dimension, global average pooling is used to compress spatial information, and the non-linear dependencies between channels are explicitly modeled through two cascaded 1×1×1 convolutional layers to generate channel weights. : ； in, and These are the ReLU and Sigmoid activation functions, respectively. S2123. In the spatial dimension, the same cascaded 1×1×1 convolutional structure is directly applied to the feature map to gradually map the high-dimensional features into pixel-level spatial weights. : ； S2124. The channel attention and spatial attention are added element-wise, and then the features are smoothed and reorganized through a residual block containing 3×3×3 convolutions and 1×1×1 convolutions. Finally, the refined hybrid features are fused with the output of the local residual branch. 。 6. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 3, characterized in that, The specific steps of step S213 are as follows: S2131.ACO-Module first introduces a window-based self-attention mechanism (WA), which divides the 3D feature volume into non-overlapping local cube windows, and then flattens the voxels in each window and maps them into query, key, and value vectors. S2132. Within each independent window, a multi-head self-attention mechanism is used to calculate the semantic association matrix between voxels, inferring the potential extension trajectory of the fault within the local region. The calculation formula for the multi-head self-attention is as follows: ； in, These are the query, key, and value matrices generated by linear projection of the voxel features within the current window. The feature dimension of the key vector, and the denominator. This is a scaling factor used to adjust the magnitude of the dot product value; This is a matrix transpose operation used to calculate the similarity score between voxels; Indicates the learnable relative position offset; S2133. Input features Divide into N non-overlapping windows For each window Using linear projection matrix , , Generate the corresponding queries, keys, and values, and perform multi-head attention computation. Then, restore all enhanced windows to their original spatial positions through the inverse operation: ； in, This is the inverse operation to reshape the block features back into a 3D tensor. This is a splicing operation along a spatial dimension; Features that provide global information; S2134. The ACO-Module calls the Feature Continuity Optimization Unit (FCO-Unit) and fuses the components using an element-wise addition method. and The system performs semantic complementation between local texture and global topology. The fused features are then fed into the deep residual refinement module within the FCO-Unit. This module consists of two stacked 3×3×3 convolutional layers and ReLU activation. The FCO-Unit uses a sliding overlap mechanism to break through the WA window boundaries, performing continuous kernel weighting on the tomographic fragments in physical space to obtain the final decoded enhanced features. : ； in, For convolution weights, This is the ReLU activation function.

7. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 1, characterized in that, The structured loss function in step S3 Includes four loss items: ； in, and This constitutes the basic segmentation loss. For binary cross-entropy loss, For Dice similarity coefficient loss; and It is a structured regularization term that constrains the fault geometry and edge quality. For edge smoothing loss, For geometric constraint loss; and It is used for balance and Hyperparameters of loss contribution.

8. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 7, characterized in that, Structured loss function middle: ； in, For the i-th prediction probability map, This is the i-th real label; ； in, For smoothing terms; ； in, and The weighting coefficients and edge consistency are calculated using a custom 3D Sobel operator to determine the predicted probability map. and real labels 3D gradient magnitude and And by minimizing and The L1 distance between them forces the learning of accurate boundary positions, and local smoothing is achieved using the three Laplacian operators. Convolve the predicted probability map, calculate and penalize the mean square value of the second derivative of the predicted surface. For predicting probability graphs P exist v The second derivative at point; ； in, A line template generated based on geometric properties such as aspect ratio and area. B For training batch size, For the first b A predicted probability plot for each sample.

9. The three-dimensional tomographic segmentation method based on local-global feature enhancement and geometric constraints according to claim 7, characterized in that, The formula for generating the final fault prediction probability map in step S4 is: ； in, For output layer convolution, The Softmax activation function is used. This is the final fault prediction probability map.

Images