A Tomographic Identification Method and System Based on Spatial-Axial Dual Attention Mechanism

Abstract

This invention discloses a fault identification method and system based on a spatial-axial dual attention mechanism. The method includes: constructing an original training dataset using a high-precision forward convolution algorithm; designing a URNet basic network integrating UNet and ResNet architectures and performing preliminary training to obtain a pre-trained model; augmenting and expanding the original training dataset to form an enhanced training dataset; based on the pre-trained model, integrating the spatial-axial dual attention mechanism and introducing a multi-level residual connection structure to construct a URNet-TSA neural network containing a bidirectional enhanced decoder; performing secondary training on the URNet-TSA neural network using the enhanced training dataset; and using the trained URNet-TSA neural network for fault identification. This invention improves the ability to capture and identify complex geological fault features by constructing a novel URNet-TSA neural network architecture.

Description

Technical Field

[0001] This invention relates to the field of intelligent earthquake fault identification technology, and in particular to a fault identification method and system based on a spatial-axial dual attention mechanism. Background Technology

[0002] In recent years, deep learning technology has made significant progress in the field of earthquake data analysis. Through end-to-end model training, deep learning automatically extracts fault features from seismic reflection data, greatly improving detection efficiency and accuracy. Convolutional Neural Networks (CNNs), as a classic deep learning method, excels in capturing local features and extracting fault boundaries, and has been widely applied in fault detection tasks. For example, semantic segmentation networks such as FCN and U-Net achieve pixel-level classification of faults through an encoder-decoder structure, demonstrating particularly outstanding performance in detecting low-order faults and complex geological structures. Among them, the U-Net network, due to its encoder-decoder structure and skip connection mechanism, excels in image segmentation and is widely used in earthquake fault identification.

[0003] However, existing U-Net and its variants still face the following technical challenges when processing 3D seismic data: First, insufficient spatial correlation. Traditional U-Net networks primarily focus on planar feature extraction, failing to capture the spatial continuity of faults across different planes in 3D seismic data, resulting in fragmented identification results. Second, limited feature representation capabilities. Existing networks are not adaptable to handling fault features at different scales, especially exhibiting low accuracy in identifying small-scale faults and tiny cracks at their edges. Third, low training efficiency. The large volume and high dimensionality of 3D seismic data lead to high computational resource consumption and slow convergence speed in traditional network architectures during training, making it difficult to meet practical application requirements. Fourth, high risk of overfitting. The complexity and nonlinearity of seismic data make neural networks prone to overfitting training samples during training, resulting in insufficient generalization ability. Therefore, there is an urgent need to develop a novel neural network architecture that can effectively capture the spatial continuity features of faults in 3D seismic data, improve identification accuracy, and reduce computational complexity to meet the needs of practical geological exploration and analysis. Summary of the Invention

[0004] The purpose of this invention is to provide a fault identification method and system based on a spatial-axial dual attention mechanism. By innovatively constructing the URNet basic network structure and deeply optimizing the decoder structure, and integrating the unique spatial-axial dual attention mechanism (Spatial-Temporal Self-Attention, TSA), a novel neural network architecture, URNet-TSA, is obtained, which significantly improves the neural network model's ability to capture complex geological fault features and its identification accuracy.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A tomographic identification method based on a spatial-axial dual attention mechanism, comprising the following steps:

[0007] S1, Dataset Construction: A high-precision forward convolution algorithm is used to generate multiple pairs of synthetic seismic reflection data with real geological characteristics and their corresponding fault labels to construct the original training dataset;

[0008] S2, Pre-trained model construction: Design the URNet base network that integrates the UNet and ResNet architectures. This network contains multiple optimized convolutional layers. Initial training is performed using the original training dataset to obtain the pre-trained model.

[0009] S3, Data Augmentation and Expansion: Introduce diverse geological deformations and noise interference into the original training dataset to augment and expand the data, forming an enhanced training dataset;

[0010] S4, Improved Bidirectional Enhanced Decoder: Based on the pre-trained model, a spatial-axial dual attention mechanism is integrated, and a multi-level residual connection structure is introduced to construct an URNet-TSA neural network containing a bidirectional enhanced decoder; the spatial-axial dual attention mechanism is used to capture the continuous correlation of fault features in different spatial dimensions;

[0011] S5, Secondary Training: The URNet-TSA neural network is trained a second time using the enhanced training dataset;

[0012] S6, Intelligent Fault Recognition: Fault recognition is performed using a trained URNet-TSA neural network.

[0013] Furthermore, in S1, the dataset construction specifically includes:

[0014] A high-precision forward convolution algorithm was used to generate 980 pairs of synthetic seismic reflection data with real geological characteristics and their corresponding fault labels.

[0015] The original training dataset is 128×128×128 in size.

[0016] Furthermore, in S2, the URNet base network includes 20 optimized convolutional layers, and the convolution formula of the URNet base network is as follows:

[0017]

[0018] Where l represents the current convolutional layer number, and k represents the convolutional kernel; represents the feature map bias of the seismic reflection data points output by the j-th convolutional kernel in the l-th layer; f(·) represents the ReLU activation function, used to introduce nonlinearity so that the network can learn complex patterns; This represents the feature map corresponding to the j-th convolutional kernel sequence in layer (l-1). M represents the weight of the i-th channel of the j-th convolutional kernel in layer l, used to process the i-th input feature map; j The value represents the total number of input channels that the j-th convolutional kernel needs to process; i represents the convolutional kernel channel index, used to traverse all input channels; j represents the convolutional kernel index in the current layer, which determines the number of output feature maps.

[0019] The residual connection formula for the URNet basic network is as follows:

[0020] N = (W - F + 2P) + E

[0021] Where N is the size of the output feature map, W is the size of the input feature map, F is the size of the convolution kernel, P is the padding value, S is the convolution stride, and E is the feature value of the previous layer.

[0022] Furthermore, in step S3, data augmentation and expansion expands the original training dataset from 980 pairs of data volumes to 1004 pairs, resulting in an enhanced training dataset.

[0023] Furthermore, in S4, the spatial-axial dual attention mechanism includes:

[0024] Spatial attention mechanism:

[0025] SA(X) = X + Dropout(Dense(Z))

[0026]

[0027] Where Q = W Q X flat K = W k X flat V = W V X flat X flat =Reshape(X)∈R B×(H×W×D)×C B: Batchsize, representing the number of samples processed at one time; H: Height, representing the height dimension of the 3D volumetric data; W: Width, representing the width dimension of the 3D volumetric data; D: Depth, representing the depth dimension of the 3D volumetric data; C: Channels, representing the number of feature channels, indicating the feature dimension at each spatial location; d krepresents the dimension of each head in the attention process; h represents the number of heads in the multi-head attention process; X represents the input tensor with shape [B,H,W,D,C], representing a batch of three-dimensional data; X flat This represents the flattened tensor of input X, with shape [B, (H×W×D), C]; W Q W represents the weight matrix used to generate the query; k W represents the weight matrix that generates the key; V The weight matrix represents the generated value; Softmax(·) transforms a set of real vectors into a probability distribution; Dropout(·) is a regularization technique used for a subset of neurons in a random neural network; Dense(Z) represents the linear transformation operation of a fully connected layer, where Z is the output of the attention mechanism.

[0028] Axial attention mechanism:

[0029] AA(X) = X + Dropout(Dense(o x +o y +o z )

[0030] In the formula, O x Indicates vertical directional attention:

[0031]

[0032] O y Indicates attention in the width direction:

[0033]

[0034] O z Indicates depth-direction attention:

[0035]

[0036] Where, d k Q represents the dimension of each head in the attention process; x ,K x V x It is the query, key, and value matrix reorganized along the height dimension H; Q y ,K y V y It is the query, key, and value matrix reorganized along the width dimension W; Q z ,K z V z It is a matrix of queries, keys, and values ​​reorganized along the depth dimension D.

[0037] Furthermore, in S4, the multi-level residual connection structure is represented as follows:

[0038] Y = ReLU(Conv) 1×1×1 (R)+BatchNorm(Conv 3×3×3 (ReLU(BatchNorm(Conv 3×3×3 (R))))))

[0039] Where R is the residual input feature; Conv 1×1×1 It's a 1×1×1 convolution, used as a shortcut path; Conv 3×3×3 It is a 3×3×3 convolution used for feature extraction; BatchNorm is a batch normalization operation; ReLU is the activation function; Y is the output of the residual block.

[0040] This invention also provides a tomography system based on a spatial-axial dual attention mechanism, applied to the above-described tomography method based on a spatial-axial dual attention mechanism, comprising:

[0041] The dataset construction module is used to generate multiple pairs of synthetic seismic reflection data with real geological features and their corresponding fault labels using a high-precision forward convolution algorithm, thereby constructing the original training dataset.

[0042] The pre-trained model building module is used to design the URNet base network that integrates the UNet and ResNet architectures. This network contains multiple optimized convolutional layers and is initially trained using the original training dataset to obtain the pre-trained model.

[0043] The data augmentation and expansion module introduces diverse geological deformations and noise interference into the original training dataset to augment and expand the data, forming an enhanced training dataset.

[0044] The improved bidirectional enhanced decoder module, based on a pre-trained model, integrates a spatial-axial dual attention mechanism and introduces a multi-level residual connection structure to construct an URNet-TSA neural network containing a bidirectional enhanced decoder; the spatial-axial dual attention mechanism is used to capture the continuity correlation of fault features in different spatial dimensions.

[0045] The secondary training module uses the enhanced training dataset to perform secondary training on the URNet-TSA neural network;

[0046] The intelligent fault identification module uses a trained URNet-TSA neural network for fault identification.

[0047] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The fault identification method based on a spatial-axial dual attention mechanism provided by the present invention: First, it uses a high-precision forward convolution algorithm to generate multiple pairs of synthetic seismic reflection data with real geological features and corresponding fault labels to construct a high-quality original training dataset; Second, it innovatively integrates the UNet and ResNet architectures to design a URNet basic network structure containing multiple optimized convolutional layers, and performs preliminary training using a high-quality original training dataset to construct a pre-trained model with high generalization ability; Third, it designs and implements data augmentation and expansion strategies, and expands the sample set by introducing diverse geological deformations and noise interference, significantly enhancing the model's anti-interference ability and adaptability; Fourth, it proposes a bidirectional enhanced decoder structure improvement scheme: on the one hand, it significantly improves training efficiency through pre-trained model loading technology; on the other hand, it innovatively integrates a spatial-axial dual attention mechanism to achieve adaptive association of cross-scale features, and introduces multi-level residual connections to enhance feature flow and gradient propagation efficiency, fundamentally solving the problem of information loss in traditional networks when processing complex fault data.

[0048] The URNet-TSA neural network constructed in this invention significantly outperforms existing technologies in fault identification, especially in achieving a qualitative leap in the spatial continuity between fault sections. For large-scale complex faults, the identification accuracy is improved, and the detection precision of tiny cracks at the edges is enhanced. Compared to traditional UNet and its variants, the URNet-TSA architecture of this invention achieves significant improvements in identification accuracy, computational efficiency, and generalization ability, providing a novel technical solution for the field of intelligent geological fault identification. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0050] Figure 1 This is a flowchart of the tomography identification method based on the spatial-axial dual attention mechanism of the present invention;

[0051] Figure 2 This is a schematic diagram of the URNet basic network structure according to an embodiment of the present invention, wherein (a) is a schematic diagram of the URNet network structure and (b) is a schematic diagram of the residual connection.

[0052] Figure 3 This is a schematic diagram of the URNet-TSA neural network structure according to an embodiment of the present invention;

[0053] Figure 4 The figures show a comparison of tomography results in an embodiment of the present invention, where (a) is the tomography result predicted by the URNet-TSA neural network, and (b) is the tomography result predicted by the URNet basic network. Detailed Implementation

[0054] The embodiments of the present invention are described in detail below. These embodiments are intended to explain the present invention and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they are performed according to the techniques or conditions described in the literature in the art or according to the product instructions. Materials or instruments whose manufacturers are not specified are all conventional products that can be obtained commercially.

[0055] This invention proposes a fault identification method based on a spatial-axial dual attention mechanism. By constructing an URNet-TSA neural network, intelligent fault identification is achieved, fundamentally solving the spatial correlation problem in fault identification of 3D seismic data and improving the accuracy and continuity of fault identification.

[0056] Example 1

[0057] like Figures 1-4 As shown, the tomography identification method based on a spatial-axial dual attention mechanism provided in this embodiment of the invention includes the following steps:

[0058] S1, Dataset Construction: A high-precision forward convolution algorithm was used to generate 980 pairs of synthetic seismic reflection data with real geological characteristics and their corresponding fault labels to construct a high-quality original training dataset with a size of 128×128×128.

[0059] S2, Pre-trained model construction: Designing the URNet base network, which integrates the UNet and ResNet architectures, such as Figure 2 As shown, the network contains 20 optimized convolutional layers and is initially trained using the original training dataset to obtain a pre-trained model. This step innovatively integrates the UNet and ResNet architectures, resulting in a pre-trained model with high generalization ability.

[0060] Specifically, the convolution formula for the URNet base network is as follows:

[0061]

[0062] Where l represents the current convolutional layer number, and k represents the convolutional kernel; represents the feature map bias of the seismic reflection data points output by the j-th convolutional kernel in the l-th layer; f(·) represents the ReLU activation function, used to introduce nonlinearity so that the network can learn complex patterns; This represents the feature map corresponding to the j-th convolutional kernel sequence in layer (l-1). M represents the weight of the i-th channel of the j-th convolutional kernel in layer l, used to process the i-th input feature map; j The value represents the total number of input channels that the j-th convolutional kernel needs to process; i represents the convolutional kernel channel index, used to traverse all input channels; j represents the convolutional kernel index in the current layer, which determines the number of output feature maps.

[0063] The residual connection formula for the URNet basic network is as follows:

[0064] N = (W - F + 2P) + E

[0065] Where N is the size of the output feature map, W is the size of the input feature map, F is the size of the convolution kernel, P is the padding value, S is the convolution stride, and E is the feature value of the previous layer.

[0066] S3, Data Augmentation and Expansion: Diverse geological deformations and noise interference are introduced into the original training dataset to augment and expand the data, forming an enhanced training dataset.

[0067] This step designs and implements data augmentation and expansion strategies. By introducing diverse geological deformations and noise interference, the data pairs are expanded to 1004 pairs, which helps to enhance the model's anti-interference ability and adaptability.

[0068] S4, Improved Bidirectional Enhanced Decoder: Based on a pre-trained model, a spatial-axial dual attention mechanism is integrated, and a multi-level residual connection structure is introduced to construct an URNet-TSA neural network containing a bidirectional enhanced decoder, such as... Figure 3 As shown, the spatial-axial dual attention mechanism is used to capture the continuity and correlation of fault features in different spatial dimensions.

[0069] This step proposes an improved bidirectional enhanced decoder structure: on the one hand, it significantly improves training efficiency through pre-trained model loading technology; on the other hand, it innovatively integrates a spatial-axial dual attention mechanism to achieve adaptive association of cross-scale features, while introducing multi-level residual connections to enhance feature flow and gradient propagation efficiency, fundamentally solving the problem of information loss in traditional networks when processing complex tomographic data.

[0070] Specifically, the spatial-axial dual attention mechanism includes a spatial attention mechanism and an axial attention mechanism.

[0071] The spatial attention mechanism is represented as follows:

[0072] SA(X) = X + Dropout(Dense(Z))

[0073]

[0074] Where Q = W Q X flat K = W k X flat V = W V X flat X flat =Reshape(X)∈R B×(H×W×D)×C B: Batchsize, representing the number of samples processed at one time; H: Height, representing the height dimension of the 3D volumetric data; W: Width, representing the width dimension of the 3D volumetric data; D: Depth, representing the depth dimension of the 3D volumetric data; C: Channels, representing the number of feature channels, indicating the feature dimension at each spatial location; d k represents the dimension of each head in the attention process; h represents the number of heads in the multi-head attention process; X represents the input tensor with shape [B,H,W,D,C], representing a batch of three-dimensional data; X flat This represents the flattened tensor of input X, with shape [B, (H×W×D), C]; W Q W represents the weight matrix used to generate the query; k W represents the weight matrix that generates the key; V The weight matrix represents the generated value; Softmax(·) transforms a set of real vectors into a probability distribution; Dropout(·) is a regularization technique used for a subset of neurons in a random neural network; Dense(Z) represents the linear transformation operation of a fully connected layer, where Z is the output of the attention mechanism.

[0075] Axial attention mechanism:

[0076] AA(X) = X + Dropout(Dense(o x +o y +o z )

[0077] In the formula, O x Indicates vertical directional attention:

[0078]

[0079] O y Indicates attention in the width direction:

[0080]

[0081] O z Indicates depth-direction attention:

[0082]

[0083] Where, d k Q represents the dimension of each head in the attention process; x ,K x V x It is the query, key, and value matrix reorganized along the height dimension H; Q y ,K y V y It is the query, key, and value matrix reorganized along the width dimension W; Q z ,K z V z It is a matrix of queries, keys, and values ​​reorganized along the depth dimension D.

[0084] In S4, the multi-level residual connection structure is represented as follows:

[0085] Y = ReLU(Conv) 1×1×1 (R)+BatchNorm(Conv 3×3×3 (ReLU(BatchNorm(Conv 3×3×3 (R))))))

[0086] Where R is the residual input feature; Conv 1×1×1 It's a 1×1×1 convolution, used as a shortcut path; Conv 3×3×3 It is a 3×3×3 convolution used for feature extraction; BatchNorm is a batch normalization operation; ReLU is the activation function; Y is the output of the residual block.

[0087] S5, Secondary Training: The URNet-TSA neural network is trained a second time using the enhanced training dataset;

[0088] S6, Intelligent Fault Recognition: Fault recognition is performed using a trained URNet-TSA neural network.

[0089] Analysis of the effectiveness of intelligent tomography identification:

[0090] Validation results based on real-world data demonstrate that the fault intelligent identification model, trained a second time using the URNet-TSA architecture, significantly enhances the accuracy of identifying fractures between faults. This model not only optimizes the representation of the connectivity relationships within the fault system in the spatial dimension but also improves the ability to describe the spatial continuity of large faults. Real-world data confirms that the optimized model exhibits higher accuracy and spatial integrity in identifying geological structures in complex tectonic environments.

[0091] like Figure 4As shown in (a), the URNet model significantly improves the continuity of faults, optimizes the identification of fault intersections (elliptical region indicated by label 3), and greatly enhances the ability to characterize internal minute and complex structures (boxed region indicated by label 1). However, it still needs improvement in noise suppression at boundary fault points (boxed region indicated by label 2). This invention focuses on the identification capability of the URNet-TSA model at different fault scales, including the boundary accuracy of small faults and the overall morphological consistency of large faults. The results are as follows: Figure 4 As shown in (b): (1) The ability to maintain the spatial continuity of faults is significantly enhanced, effectively reducing the problem of missing fault lines (the area indicated by the box in label 5); (2) The ability to distinguish complex structural regions is significantly improved, enabling accurate identification of complex fault systems with intersecting internal structures (the area indicated by the box in label 4); (3) The accuracy of edge fracture detection is significantly improved, especially in boundary fault regions where traditional methods are prone to missed detection (the elliptical area indicated by label 6). These improvements make this model of significant application value in the field of automatic fault identification.

[0092] Example 2

[0093] This invention also provides a tomography system based on a spatial-axial dual attention mechanism, applied to the above-described tomography method based on a spatial-axial dual attention mechanism, comprising:

[0094] The dataset construction module is used to generate multiple pairs of synthetic seismic reflection data with real geological features and their corresponding fault labels using a high-precision forward convolution algorithm, thereby constructing the original training dataset.

[0095] The pre-trained model building module is used to design the URNet base network that integrates the UNet and ResNet architectures. This network contains multiple optimized convolutional layers and is initially trained using the original training dataset to obtain the pre-trained model.

[0096] The data augmentation and expansion module introduces diverse geological deformations and noise interference into the original training dataset to augment and expand the data, forming an enhanced training dataset.

[0097] The improved bidirectional enhanced decoder module, based on a pre-trained model, integrates a spatial-axial dual attention mechanism and introduces a multi-level residual connection structure to construct an URNet-TSA neural network containing a bidirectional enhanced decoder; the spatial-axial dual attention mechanism is used to capture the continuity correlation of fault features in different spatial dimensions.

[0098] The secondary training module uses the enhanced training dataset to perform secondary training on the URNet-TSA neural network;

[0099] The intelligent fault identification module uses a trained URNet-TSA neural network for fault identification.

[0100] Example 3

[0101] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the tomographic identification method based on the spatial-axial dual attention mechanism described above.

[0102] Matters not covered in this invention are common knowledge.

[0103] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0104] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A tomographic identification method based on a spatial-axial dual attention mechanism, characterized in that, Includes the following steps: S1, Dataset Construction: A high-precision forward convolution algorithm is used to generate multiple pairs of synthetic seismic reflection data with real geological characteristics and their corresponding fault labels to construct the original training dataset; S2, Pre-trained model construction: Design the URNet base network that integrates the UNet and ResNet architectures. This network contains multiple optimized convolutional layers. Initial training is performed using the original training dataset to obtain the pre-trained model. S3, Data Augmentation and Expansion: Introduce diverse geological deformations and noise interference into the original training dataset to augment and expand the data, forming an enhanced training dataset; S4, Improved Bidirectional Enhanced Decoder: Based on the pre-trained model, a spatial-axial dual attention mechanism is integrated, and a multi-level residual connection structure is introduced to construct an URNet-TSA neural network containing a bidirectional enhanced decoder; The spatial-axial dual attention mechanism is used to capture the continuous correlation of fault features in different spatial dimensions; S5, Secondary Training: The URNet-TSA neural network is trained a second time using the enhanced training dataset; S6, Intelligent Fault Recognition: Fault recognition is performed using a trained URNet-TSA neural network.

2. The tomographic identification method based on a spatial-axial dual attention mechanism according to claim 1, characterized in that, In S1, the dataset construction specifically includes: A high-precision forward convolution algorithm was used to generate 980 pairs of synthetic seismic reflection data with real geological characteristics and their corresponding fault labels. The original training dataset is 128×128×128 in size.

3. The tomographic identification method based on a spatial-axial dual attention mechanism according to claim 1, characterized in that, In S2, the URNet base network includes 20 optimized convolutional layers. The convolution formula of the URNet base network is as follows: ； in, Indicates the current convolutional layer number; Indicates the earthquake reflection data points after the first Layer j The feature map bias of the output of each convolutional kernel; f (·) denotes the ReLU activation function; It indicates the first -1st floor j Feature maps corresponding to each convolutional kernel sequence; Indicates the first The first in the layer j The first convolutional kernel i The weight of the i-th channel is used to process the i-th channel. i One input feature map; Indicates the first j The total number of input channels that each convolutional kernel needs to process; i This represents the kernel channel index, used to iterate through all input channels; j This indicates the kernel index in the current layer, which determines the number of output feature maps; The formula for calculating the output feature map size of the URNet base network is as follows: ； in, N To determine the size of the output feature map, W The size of the input feature map. F The kernel size is... P To fill in the value size, E These are the eigenvalues ​​of the previous layer.

4. The tomographic identification method based on a spatial-axial dual attention mechanism according to claim 2, characterized in that, In step S3, data augmentation and expansion expands the original training dataset from 980 pairs of data volumes to 1004 pairs, resulting in an augmented training dataset.

5. The tomographic identification method based on a spatial-axial dual attention mechanism according to claim 1, characterized in that, In S4, the spatial-axial dual attention mechanism includes: Spatial attention mechanism: ； ； in, , , , B represents the number of samples processed at one time; H represents the height dimension of the 3D volume data; W represents the width dimension of the 3D volume data; D represents the depth dimension of the 3D volume data; C is the number of feature channels, representing the feature dimension of each spatial location. X represents the dimension of each head in the attention process; X represents the input tensor with shape [B, H, W, D, C], representing a batch of three-dimensional data; X flat This represents the flattened tensor of input X, with shape [B, (H×W×D), C]; W Q W represents the weight matrix used to generate the query. k W represents the weight matrix for generating the bonds. V The weight matrix represents the generated values; Softmax(·) transforms a set of real vectors into a probability distribution; Dropout(·) is a regularization technique used for a subset of neurons in a random neural network; Dense(Z) represents the linear transformation operation of a fully connected layer, where Z is the output of the attention mechanism. Axial attention mechanism: ； In the formula, O x Indicates vertical directional attention: ； O y Indicates attention in the width direction: ； O z Indicates depth-direction attention: ； in, Q represents the dimension of each head in the attention process; x , K x V x It is the query, key, and value matrix reorganized along the height dimension H; Q y , K y V y It is the query, key, and value matrix reorganized along the width dimension W; Q z , K z V z It is a matrix of queries, keys, and values ​​reorganized along the depth dimension D.

6. The tomographic identification method based on a spatial-axial dual attention mechanism according to claim 1, characterized in that, In S4, the multi-level residual connection structure is represented as follows: Y=ReLU(Conv 1×1×1 (R)+BatchNorm(Conv 3×3×3 (ReLU(BatchNorm(Conv 3×3×3 (R)))))) Where R is the residual input feature; Conv 1×1×1 It's a 1×1×1 convolution, used as a shortcut path; Conv 3×3×3 It is a 3×3×3 convolution used for feature extraction; BatchNorm is a batch normalization operation; ReLU is the activation function; Y is the output of the residual block.

7. A tomographic identification system based on a spatial-axial dual attention mechanism, applied to the tomographic identification method based on a spatial-axial dual attention mechanism as described in any one of claims 1-6, characterized in that, include: The dataset construction module is used to generate multiple pairs of synthetic seismic reflection data with real geological features and their corresponding fault labels using a high-precision forward convolution algorithm, thereby constructing the original training dataset. The pre-trained model building module is used to design the URNet base network that integrates the UNet and ResNet architectures. This network contains multiple optimized convolutional layers and is initially trained using the original training dataset to obtain the pre-trained model. The data augmentation and expansion module introduces diverse geological deformations and noise interference into the original training dataset to augment and expand the data, forming an enhanced training dataset. The improved bidirectional enhanced decoder module, based on a pre-trained model, integrates a spatial-axial dual attention mechanism and introduces a multi-level residual connection structure to construct an URNet-TSA neural network containing a bidirectional enhanced decoder; The spatial-axial dual attention mechanism is used to capture the continuous correlation of fault features in different spatial dimensions; The secondary training module uses the enhanced training dataset to perform secondary training on the URNet-TSA neural network; The intelligent fault identification module uses a trained URNet-TSA neural network for fault identification.

Images