Three-dimensional seismic fault identification method based on double-attention multi-scale fusion U-Net

Abstract

The invention provides a three-dimensional seismic fault identification method based on double-attention multi-scale fusion U-Net. The method comprises the following specific steps: constructing a U-shaped network architecture comprising an encoder, a decoder and jump connection; a constructed multi-scale feature fusion module is embedded in the first level of the encoder, and the extraction capability of fault features of different scales is enhanced through a multi-branch structure; introducing the constructed hole fusion modules into the second and third levels of the encoder, designing and expanding a receptive field by using multiple expansion rates, and capturing fault structures of different scales; a double-attention parallel mechanism is integrated in jump connection, and the sensitivity of channel attention and space attention to fault features is improved; constructing a combined loss function; and finally, performing three-dimensional seismic data training and reasoning based on the optimized model to realize high-precision fault identification. The method has high generalization and accuracy, and especially has good performance in the aspect of seismic image fault identification containing a large fault scale span.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary field of geophysical exploration and artificial intelligence, and specifically relates to a three-dimensional seismic fault identification method based on deep learning, in particular a U-Net model that combines a dual-attention mechanism with multi-scale feature fusion to improve the accuracy and efficiency of fault detection in seismic data. Background Art

[0002] Traditional methods mainly rely on manual interpretation and seismic attribute analysis. Coherence volume analysis identifies faults by calculating the local similarity of seismic data, but it is sensitive to noise and easily affected by stratum folds; curvature attribute analysis detects faults based on the geometric curvature of the seismic reflection surface, but has poor ability to identify tiny cracks; ant volume tracking technology simulates the ant colony algorithm to enhance fault continuity, but relies on manual parameter tuning and has low computational efficiency; using principal component analysis (PCA) after dimensionality reduction and combining it with the nearest neighbor algorithm for fault identification improves computational efficiency, but has limited feature expression capabilities and is difficult to adapt to complex geological structures.

[0003] In recent years, the introduction of deep learning has significantly improved the automation and accuracy of fault identification. CNN was introduced into fault identification, but it was limited by the local receptive field and had difficulty modeling long-distance fault structures. U-Net became a mainstream framework due to its encoder-decoder structure and skip connections. Attention mechanisms were introduced into this network to suppress noise interference. In order to alleviate gradient vanishing and improve the continuity of low-order fault identification, residual connections and multi-layer supervision were adopted. Densely connected networks enhance feature reuse through dense skip connections to improve the detection rate of small faults. However, these methods are still insufficient in multi-scale feature extraction. A pure Transformer architecture is used to capture global context through a self-attention mechanism, which can handle long-distance modeling well, but the computational cost is high. A multi-task learning framework is used in combination with seismic data reconstruction tasks to improve feature utilization, but the model complexity is high.

[0004] At present, most identification methods still face problems such as insufficient multi-scale feature coverage, inadequate long-distance continuity modeling, and poor robustness to noise interference, which restrict the effectiveness of fault detection under complex geological conditions; using complex network structures to identify faults requires too high computational costs and low computational efficiency; therefore, a method is needed to balance the accuracy and computational efficiency of fault identification. Summary of the Invention

[0005] 1. Purpose of the invention:

[0006] The purpose of the present invention is to propose a three-dimensional earthquake fault identification method based on dual-attention multi-scale fusion U-Net, which can better improve the accuracy and efficiency of fault detection in seismic data, while improving the computational efficiency of fault identification.

[0007] 2. Technical solution:

[0008] 1. A 3D seismic fault recognition method based on dual-attention multi-scale fusion U-Net, characterized by comprising the following steps:

[0009] Step 1: After reading the data volume and the corresponding fault labels, the data dimension is initialized to a single channel of 128×128×128. The seismic data is then normalized and the amplitude values ​​are mapped to the interval [0, 255]. The samples marked as faults are randomly rotated and flipped along the x / y axis to perform data augmentation to maintain the synchronization between the seismic data and the fault labels. A dynamic batching mechanism is then used to treat the generated sample set as a batch. The sample order is automatically shuffled after each round, and different domain labels are assigned to the data. Adversarial training is implemented through a gradient reversal layer. The read 3D data is divided into a training set and a validation set. The areas with the densest fault distribution in two real seismic data are cut out, and the 2D seismic data are then reconstructed into 3D seismic data blocks as the test set.

[0010] Step 2: Construct a U-shaped network architecture including an encoder, a decoder, and a skip connection. The improved U-shaped network includes the following structure:

[0011] Step 2.1: Construct a multi-scale feature fusion module to effectively extract multi-scale information of input features through a parallel convolution branch structure. The multi-scale feature fusion module includes the following structure:

[0012] First, an automatic filter allocation mechanism is adopted in this module. This allocation mechanism can automatically configure the parameters of each branch by only specifying the number of basic filters. A total of four branches are designed. The first branch is a 1×1×1 convolution layer, which can realize linear transformation and fast channel compression of channel information. The second and fourth branches both adopt the structure of 1×1×1 convolution + 3×3×3 convolution, where the 1×1×1 convolution is used to reduce the dimension and reduce the amount of calculation, and the 3×3×3 convolution further extracts local spatial features. The third branch has the same structure as the first branch, which is a 1×1×1 convolution layer to further enrich the shallow information representation. Then, all branch outputs are spliced ​​according to the channel dimension to form a multi-scale fusion feature map, and then feature reorganization and channel compression are performed through 3×3×3 convolution and ReLU activation, and finally a feature representation of unified dimension is output.

[0013] Among them, for the input feature map X∈R D×H×W×C , D, H, W, C represent the length, height, width and number of channels of the input feature map respectively, and the feature change process is: Y = σ(W6*Concat(σ(W1X),σ(W3σ(W2X)),σ(W4X),σ(W6σ(W5X)),W iRepresents the weight matrix of each convolutional layer, i∈(1,2,3...6), * represents the 3D convolution operation, σ· is the ReLU activation function, Concat represents the channel dimension splicing, and Y is the output feature after the input feature is extracted by the multi-scale fusion module;

[0014] Step 2.2: Construct a hole fusion module to capture both local details and global context information in 3D data, solving the problem of limited receptive field of traditional convolution. The hole fusion module includes the following structure:

[0015] This module combines multi-scale dilated convolution and residual fusion mechanism, and has two parallel dilated convolution branches: in branch one, a 3×3×3 convolution with a dilation rate of 2 is used to expand the receptive field to a 7×7×7 area, and in branch two, a 3×3×3 convolution with a dilation rate of 4 is used to expand the receptive field to a 15×15×15 area. The application of dual-branch dilated convolution can capture the spatial dependencies of medium range and long distance respectively without increasing the number of parameters. After the outputs of the two dilated convolution branches are channel-grouped and normalized and activated by the Relu function, the features of the two branches are concatenated in the channel dimension. After concatenation, 1×1×1 convolution and Relu activation function are used for feature fusion and channel expansion. Then, a 3×3×3 intermediate convolution and Relu function are used to enhance the nonlinear expression ability. In order to avoid the gradient disappearance problem, a residual connection is designed. The input features are adjusted by 1×1×1 convolution and then added to the features fused by the dual-branch dilated convolution. Finally, the Relu activation function is used to ensure nonlinearity.

[0016] Step 2.3: Construct a dual attention parallel module to dynamically enhance important channels and spatial regions in the 3D feature map. The dual attention parallel module includes the following structure:

[0017] First, in the channel attention branch, global average pooling and global maximum pooling are first performed to comprehensively summarize the global information of the channel. Then, 1×1×1 convolution and Relu activation function are used to reduce the dimension, reduce the number of channels to form bottleneck information, and force the network to learn key features. Then, a 1×1×1 convolution and Sigmoid function are used to increase the dimension, restore the original number of channels, generate weight values ​​of 0 to 1, and learn the importance relationship between channels through the combination of two 1×1×1 convolutions. Finally, the original input features are multiplied by the generated channel weights channel by channel. The entire process of dynamically adjusting the channel weights is: Att channnel =σ(MLP(GAP(X))+MLP(GMP(X)), where σ(·) is the Sigmoid activation function, MLP is a multi-layer perceptron implemented by two 1×1×1 convolutions with ReLU and Sigmoid, GAP is global average pooling, GMP is global maximum pooling, X is the input feature, Attchannnel is the output feature after channel attention;

[0018] Then, in the spatial attention branch, first, through maximum pooling and average pooling, maximum pooling retains the most significant features of each local area, average pooling smoothes regional features to avoid noise interference, and uses these two pooling methods to extract spatial information from different perspectives; then the pooled features are subjected to a 3x3x3 convolution and Sigmoid activation function operation to learn the correlation of spatial regions, and then upsampling is used to restore the pooled low-resolution feature map to the input size and align it with the original features; finally, the original input features are multiplied by the generated spatial weights position by position to highlight the key areas; the process of combining maximum pooling, average pooling and 3D convolution to generate a spatial weight map is: Att spatial =Upsample(σ(Conv3D(Pool max (X)))+σ(Conv3D(Pool avg (X)))), where σ(·) is the Sigmoid activation function, Upsample is the upsampling operation, and Pool max For maximum pooling, Pool avg is average pooling, Conv3D is 3D convolution, X is input feature, Att spatial is the output feature after spatial attention;

[0019] Finally, the outputs of the channel attention branch and the spatial attention branch are directly added together to retain the complementary information of the two attention mechanisms; a lightweight convolution is performed on the fused features through a 1×1×1 convolution and a Sigmoid activation function to further adjust the feature distribution and output a result of the same size as the input; after the weighted fusion of the dual attention outputs, the integration process through a 1×1×1 convolution is: Y = σ(Conv3D 1x1x1 (X⊙Att channel +X⊙Att spatial )), where ⊙ represents element-by-element multiplication, σ(·) is the Sigmoid function, and Y is the output feature after parallel dual attention. Conv3D 1x1x1 is a 1×1×1 convolution, X is the input feature, Att channel is the output feature of channel attention, Att spatial is the output feature of spatial attention;

[0020] Step 2.4: Construct a dual-attention multi-scale fusion U-Net model. The dual-attention multi-scale fusion U-Net model includes the following structure:

[0021] While retaining the symmetric structure of the classic U-Net encoder and decoder, the modules constructed in steps 2.1, 2.2, and 2.3 are introduced. The network adopts a 4-layer deep design, and the input feature map size is 128×128×128×1. After downsampling and upsampling, the final output is a fault prediction map, and the output feature map size is the same as the input feature map size.

[0022] In the encoder path, a multi-scale feature fusion module is first used at the first level to extract initial features from the input. The four-branch parallel structure can effectively capture fault features of different scales. Then, a dilated convolution fusion module is introduced at the second and third levels. This module can expand the receptive field by using parallel convolutions with different dilation rates. After each encoding level, a 2×2×2 maximum pooling is also connected for downsampling. After reaching the bottleneck layer, two stacked 3×3×3 convolutions and ReLU activation functions are used to maintain the feature dimension. The number of channels in the entire encoder path is 1, 16, 32, and 64, respectively.

[0023] The decoder path adopts a symmetrical upsampling structure. First, the bottleneck layer output is upsampled by 2×2×2, and then the upsampled feature map is spliced ​​with the third-layer encoding feature map processed by the dual attention module, and then two stacked 3×3×3 convolutions and Relu activation functions are used for feature fusion; then the feature map output by the third layer of the decoder is used as the feature map of the second layer, and then spliced ​​with the second-layer encoder feature map after dual attention processing, and similarly, two stacked 3×3×3 convolutions and Relu activation functions are used for feature fusion; then the feature map output by the second layer of the decoder is used as the feature map of the first layer, and then spliced ​​with the first-layer encoder feature map after dual attention processing, and then two stacked 3×3×3 convolutions and Relu activation functions are used for feature fusion; finally, the feature map output by the first layer of the decoder is output as a fault probability map through 1×1×1 convolution and Sigmoid activation. The number of channels in this path is 128, 96, 48, 16 and 1 respectively;

[0024] Step 3: Parameter optimization, the process is as follows:

[0025] Step 3.1: Construct a hybrid loss function, the calculation formula is as follows:

[0026] The Dice Loss component is used to display the spatial continuity of the optimized tomographic voxels. The Dice Loss function formula is: Where N is the number of pixels in the 3D seismic image, y i represents the label of pixel i, p i is the prediction of pixel i, ∈=10 -6 Prevent division by zero, L Diceis the Dice Loss function; FocalLoss focuses on difficult samples through dynamic weights. The FocalLoss function formula is:

[0027] Where N is the number of pixels in the 3D seismic image, y i represents the label of pixel i, p i It represents the prediction of pixel i. α and γ are high parameters and the final values ​​are obtained through multiple experiments. Focal (α,γ) is the Focal Loss function; the combination of FocalLoss and DiceLoss is used as the loss function: L Focal_Dice =L Focal (α,γ)+L Dice , where L Dice is the Dice Loss function, L Focal (α,γ) is the FocalLoss function, L Focal_Dice It is the combined loss function of FocalLoss and Dice Loss. α and γ are both high parameters and the optimal values ​​are obtained through multiple experiments.

[0028] Step 3.2: Introduce the dual-attention multi-scale fusion U-Net model constructed in step 2.4 into the loss function of step 3.1 and design and combine it into a complete training model;

[0029] Step 3.3: Set the learning rate lr, batch size batch_size, decay rate decay, training period epoch, and optimizer adam. All parameters are optimized after multiple experiments.

[0030] Step 3.4: Use the training data from step 1 to train the model from step 3.2, evaluate the model performance using the validation data, and save the model after a complete training cycle.

[0031] Step 4: Use the two sets of test data in step 1 for the model prediction saved in step 3.4, and output a 3D fault prediction probability map;

[0032] 3. Innovation:

[0033] This paper proposes a method that significantly improves the accuracy and efficiency of fault identification based on the traditional U-Net architecture through the innovative design of a multi-scale feature fusion module, a void fusion module, and a dual-attention parallel mechanism. Its core innovations are as follows:

[0034] (1) To address the problem of large fault spans, a multi-scale feature fusion module with a four-branch parallel structure is designed. This module extracts features in parallel through convolution kernels of different sizes, and combines channel splicing and reshaping operations to achieve joint detection of small cracks and large faults.

[0035] (2) To address the problem of limited receptive field of traditional convolution and difficulty in modeling long-distance fault continuity, a fusion module based on multi-dilation rate dilated convolution is designed. By using convolutions with different dilation rates in parallel, the module expands the receptive field area without increasing the number of parameters, captures fault structures of different scales in seismic data, and avoids local overfitting.

[0036] (3) To address the boundary blurring caused by noise interference and stratum folds, a parallel dual attention module is introduced into the skip connection. Channel attention dynamically weights important feature channels through global pooling and multi-layer perceptron, while spatial attention focuses on key areas by combining maximum pooling and average pooling. The two work together to improve the model's sensitivity to fault boundaries in low signal-to-noise ratio environments.

[0037] (4) A combined loss function of FocalLoss and Dice Loss is used to focus on difficult samples through dynamic weights, while optimizing the spatial continuity of tomographic voxels to solve the problem of fragmented prediction caused by extreme sample imbalance; 4. Beneficial effects:

[0038] The present invention discloses a three-dimensional seismic fault recognition method based on deep learning, and provides a three-dimensional seismic fault recognition method based on dual-attention multi-scale fusion U-Net. First, through a multi-scale feature fusion module, global and local features in seismic data are captured simultaneously, and the model's recognition ability of fault features of different scales is enhanced; then, by using a void fusion module, the model expands the receptive field without reducing the resolution, can capture fault structures of different scales in seismic data, and maintain the continuity of long-distance faults in seismic data; finally, through parallel dual-attention modules, the expression of important feature channels is enhanced, unimportant channels are suppressed, and the focus is on the spatial area where the fault exists, thereby improving the model's sensitivity to fault features; this method ensures that the model effectively recognizes faults while also improving the model's recognition accuracy and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flowchart for identifying faults in the dual-attention multi-scale fusion U-Net model;

[0040] Figure 2Figure 2 shows the structure of the multi-scale feature fusion module. This module extracts multi-scale information of input features through a parallel convolutional branch structure. An automated filter allocation mechanism is used in this module, which automatically configures the parameters of each branch by simply specifying the number of basic filters. Two 1x1x1 convolution branches directly extract features, while two 1x1x1+3x3x3 convolution branches perform dimensionality reduction before extracting spatial features. The outputs of all branches are concatenated in the channel dimension and finally reshaped through a 3x3x3 convolution.

[0041] Figure 3 This is the structural diagram of the dilated fusion module. This module is a feature extraction module that combines multi-scale dilated convolution and residual fusion mechanism. It has two parallel 3x3x3 convolutions with expansion rates of 2 and 4 respectively. The number of output channels of each branch is half of the input. Each branch is followed by group normalization and ReLU activation. The dual-branch features are concatenated in the channel dimension. The number of channels is expanded by 1x1x1 convolution. A 3x3x3 intermediate convolution is added to enhance nonlinearity. The input is dimensionalized by 1x1x1 convolution and ReLU activation is performed after adding the fused features.

[0042] Figure 4 This is the structure diagram of the dual-attention parallel module. The channel attention branch of this module obtains channel statistical information through global average pooling GAP and global maximum pooling GMP, uses a bottleneck structure composed of two 1x1x1 convolutions to first reduce the dimension and then restore the generated channel weights, multiplying the weights with the original features to highlight important channels; spatial attention branch obtains spatial information through maximum pooling and average pooling downsampling, uses 3x3x3 convolution to process the downsampled features, and generates a spatial weight map after upsampling to restore the original size. Multiply the weights with the original features to highlight important spatial areas; finally, the outputs of the two attention branches are directly added, and the final output is adjusted through 1x1x1 convolution;

[0043] Figure 5 This is the structure of the dual-attention multi-scale fusion U-Net model. While retaining the symmetrical structure of the classic U-Net encoder and decoder, it introduces two feature extraction modules and an attention module to enhance the ability to extract fault features. The network uses a 4-layer deep design, with an input feature map size of 128×128×128×1. After downsampling and upsampling, the final output is a fault prediction map, and the output feature map size is the same as the input feature map size.

[0044] Figure 6 This is a diagram showing the effect of the dual-attention multi-scale fusion U-Net model on fault recognition on the 3D synthetic seismic data validation set. The six sub-images are respectively the fault recognition effects on the six 3D synthetic seismic data validation sets;

[0045] Figure 7The first figure shows the predicted effect on the F3 block seismic data offshore the Netherlands, and the second figure shows the predicted effect on the Kerry-3D data. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0047] A three-dimensional seismic fault recognition method based on dual-attention multi-scale fusion U-Net is characterized by comprising the following steps:

[0048] Step 1: After reading the 3D seismic data volume and the corresponding fault labels in the synthetic seismic dataset provided by Wu, the data dimension is initialized to 128×128×128 and single channel. Then, the training mode, test mode or domain adaptation mode is entered. The seismic data is normalized and the amplitude value is mapped to the interval [0, 255]. The samples marked as faults are randomly rotated and randomly flipped along the x / y axis to perform data enhancement to maintain the synchronization between the seismic data and the fault labels. A dynamic batching mechanism is then adopted to treat the generated sample set as a batch. The sample order is automatically shuffled after each round. Different domain labels (1 / 0) are assigned to the synthetic data and the real data. Adversarial training is implemented through the gradient reversal layer. The read 3D data is divided into a training set and a validation set. The area with the densest fault distribution in the two real seismic data is cut out, and the 2D seismic data is reconstructed into a 3D seismic data block as the test set.

[0049] Among them, the synthetic seismic dataset provided by Wu comes from [1]: WU X, LIANG L, SHI Y et al. FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation [J]. Geophysics 2019, 84 (3): IM35–IM45. Two sets of real test data: the first set of test datasets is the F3 block seismic data of the Netherlands offshore with a volume size of 512×384×128, which is provided by NLOG and dGB Earth Sciences and has a low signal-to-noise ratio; the second set of data is the Kerry3D data provided by New Zealand Petroleum and Minerals with a volume size of 287×735×1252. The sub-area with the densest faults is extracted and a sub-volume of 128×512×256 is cut out as the effective test area;

[0050] Step 2: Construct a U-shaped network architecture including an encoder, a decoder, and a skip connection. The improved U-shaped network includes the following structure:

[0051] Step 2.1: Construct a multi-scale feature fusion module to effectively extract the multi-scale information of the input features through a parallel convolution branch structure. Figure 2 As shown;

[0052] Step 2.2: Construct a hole fusion module to capture both local details and global context information in 3D data. Figure 3 As shown;

[0053] Step 2.3: Construct a dual attention parallel module to dynamically enhance the important channels and spatial regions in the 3D feature map. Figure 4 As shown;

[0054] Step 2.4: Build a dual-attention multi-scale fusion U-Net model that introduces the modules of steps 3.1, 3.2 and 3.3. Figure 5 As shown;

[0055] Step 3: Parameter optimization, the process is as follows:

[0056] Step 3.1: Construct a hybrid loss function with hyperparameters α = 0.9 and γ = 2.0. All parameters are optimized after multiple experiments.

[0057] Step 3.2: Introduce the dual-attention multi-scale fusion U-Net model constructed in step 2.4 into the loss function of step 3.1 and design and combine it into a complete training model;

[0058] Step 3.3: Set the learning rate lr = 0.0001, batch size batch_size = 1, decay rate decay = 0.0002, training period epoch = 50, and optimizer adam. All parameters are optimized after multiple experiments.

[0059] Step 3.4: Use the training data from step 1 to train the model from step 3.2, evaluate the model performance using the validation data, and save the model after a complete training cycle.

[0060] Step 4: Randomly synthesize six data points of the 3D seismic data validation set, use the model of the complete training cycle saved in step 3.4 to predict these six data points, and visualize and generate a 3D fault prediction probability map. Figure 6 As shown; use the two sets of test data in step 1 for the model prediction saved in step 3.4, and output a three-dimensional fault prediction probability map. Figure 7 As shown;

[0061] The following table compares the fault identification method between this method and different methods. The evaluation index is accuracy, and the calculation formula is: Where TP represents the number of correctly identified fault bodies, TN represents the number of correctly identified non-fault bodies, FP represents the number of non-fault bodies mistakenly predicted as fault bodies, FN represents the number of fault bodies mistakenly predicted as non-fault bodies, and Accuracy represents the accuracy rate. The accuracy rate calculation formula is: TP represents the number of correctly identified fault bodies, FP represents the number of non-fault bodies mistakenly predicted as fault bodies, and Precision represents the accuracy rate.

[0062] Table 1 Comparison of evaluation indicators and training time of the method of the present invention and different models Table 1.ComparisonTable of Evaluation Metrics and Training TimeBetween the Proposed Method and Different Models

[0063] Accuracy precision time UNet 0.9486 0.6216 4h10min MARU-Net 0.9664 0.7944 5h21min AHRFaultSegNet 0.9489 0.7512 4h36min ESU-Net 0.9570 0.7269 5h30min Algorithm of the present invention 0.9753 0.8146 4h48min

Claims

1. A 3D earthquake fault recognition method based on dual-attention multi-scale fusion U-Net, characterized by The following steps are involved: Step 1: After reading the data volume and the corresponding fault labels, the data dimension is initialized to a single channel of 128×128×128. The seismic data is then normalized and the amplitude values ​​are mapped to the interval [0, 255]. The samples marked as faults are randomly rotated and flipped along the x / y axis to perform data augmentation to maintain the synchronization between the seismic data and the fault labels. A dynamic batching mechanism is then used to treat the generated sample set as a batch. The sample order is automatically shuffled after each round, and different domain labels are assigned to the data. Adversarial training is implemented through a gradient reversal layer. The read 3D data is divided into a training set and a validation set. The areas with the densest fault distribution in two real seismic data are cut out, and the 2D seismic data are then reconstructed into 3D seismic data blocks as the test set. Step 2: Construct a U-shaped network architecture including an encoder, a decoder, and a skip connection. The improved U-shaped network includes the following structure: Step 2.1: Construct a multi-scale feature fusion module to effectively extract multi-scale information of input features through a parallel convolution branch structure. The multi-scale feature fusion module includes the following structure: First, an automatic filter allocation mechanism is adopted in this module. This allocation mechanism can automatically configure the parameters of each branch by only specifying the number of basic filters. A total of four branches are designed. The first branch is a 1×1×1 convolution layer, which can realize linear transformation and fast channel compression of channel information. The second and fourth branches both adopt the structure of 1×1×1 convolution + 3×3×3 convolution, where the 1×1×1 convolution is used to reduce the dimension and reduce the amount of calculation, and the 3×3×3 convolution further extracts local spatial features. The third branch has the same structure as the first branch, which is a 1×1×1 convolution layer to further enrich the shallow information representation. Then, all branch outputs are spliced ​​according to the channel dimension to form a multi-scale fusion feature map, and then feature reorganization and channel compression are performed through 3×3×3 convolution and ReLU activation, and finally a feature representation of unified dimension is output. Among them, for the input feature map X∈R D×H×W×C , D, H, W, C represent the length, height, width and number of channels of the input feature map respectively, and the feature change process is: Y = σ(W6*Concat(σ(W1X),σ(W3σ(W2X)),σ(W4X),σ(W6σ(W5X)),W i Represents the weight matrix of each convolutional layer, i∈(1,2,3...6), * represents the 3D convolution operation, σ· is the ReLU activation function, Concat represents the channel dimension splicing, and Y is the output feature after the input feature is extracted by the multi-scale fusion module; Step 2.2: Construct a hole fusion module to capture both local details and global context information in 3D data, solving the problem of limited receptive field of traditional convolution. The hole fusion module includes the following structure: This module combines multi-scale dilated convolution and residual fusion mechanism, and has two parallel dilated convolution branches: in branch one, a 3×3×3 convolution with a dilation rate of 2 is used to expand the receptive field to a 7×7×7 area, and in branch two, a 3×3×3 convolution with a dilation rate of 4 is used to expand the receptive field to a 15×15×15 area. The application of dual-branch dilated convolution can capture the spatial dependencies of medium range and long distance respectively without increasing the number of parameters. After the outputs of the two dilated convolution branches are channel-grouped and normalized and activated by the Relu function, the features of the two branches are concatenated in the channel dimension. After concatenation, 1×1×1 convolution and Relu activation function are used for feature fusion and channel expansion. Then, a 3×3×3 intermediate convolution and Relu function are used to enhance the nonlinear expression ability. In order to avoid the gradient disappearance problem, a residual connection is designed. The input features are adjusted by 1×1×1 convolution and then added to the features fused by the dual-branch dilated convolution. Finally, the Relu activation function is used to ensure nonlinearity. Step 2.3: Construct a dual attention parallel module to dynamically enhance important channels and spatial regions in the 3D feature map. The dual attention parallel module includes the following structure: First, in the channel attention branch, global average pooling and global maximum pooling are first performed to comprehensively summarize the global information of the channel. Then, 1×1×1 convolution and Relu activation function are used to reduce the dimension, reduce the number of channels to form bottleneck information, and force the network to learn key features. Then, a 1×1×1 convolution and Sigmoid function are used to increase the dimension, restore the original number of channels, generate weight values ​​of 0 to 1, and learn the importance relationship between channels through the combination of two 1×1×1 convolutions. Finally, the original input features are multiplied by the generated channel weights channel by channel. The entire process of dynamically adjusting the channel weights is: Att channnel =σ(MLP(GAP(X))+MLP(GMP(X)), where σ(·) is the Sigmoid activation function, MLP is a multi-layer perceptron implemented by two 1×1×1 convolutions with ReLU and Sigmoid, GAP is global average pooling, GMP is global maximum pooling, X is the input feature, Att channnel is the output feature after channel attention; Then, in the spatial attention branch, first, through maximum pooling and average pooling, maximum pooling retains the most significant features of each local area, average pooling smoothes regional features to avoid noise interference, and uses these two pooling methods to extract spatial information from different perspectives; then the pooled features are subjected to a 3x3x3 convolution and Sigmoid activation function operation to learn the correlation of spatial regions, and then upsampling is used to restore the pooled low-resolution feature map to the input size and align it with the original features; finally, the original input features are multiplied by the generated spatial weights position by position to highlight the key areas; the process of combining maximum pooling, average pooling and 3D convolution to generate a spatial weight map is: Att spatial =Upsample(σ(Conv3D(Pool max (X)))+σ(Conv3D(Pool avg (X)))), where σ(·) is the Sigmoid activation function, Upsample is the upsampling operation, and Pool max For maximum pooling, Pool avg is average pooling, Conv3D is 3D convolution, X is input feature, Att spatial is the output feature after spatial attention; Finally, the outputs of the channel attention branch and the spatial attention branch are directly added together to retain the complementary information of the two attention mechanisms; a lightweight convolution is performed on the fused features through a 1×1×1 convolution and a Sigmoid activation function to further adjust the feature distribution and output a result of the same size as the input; after the weighted fusion of the dual attention outputs, the integration process through a 1×1×1 convolution is: Y = σ(Conv3D 1x1x1 (X⊙Att channel +X⊙Att spatial )), where ⊙ represents element-by-element multiplication, σ(·) is the Sigmoid function, and Y is the output feature after parallel dual attention. Conv3D 1x1x1 is a 1×1×1 convolution, X is the input feature, Att channel is the output feature of channel attention, Att spatial is the output feature of spatial attention; Step 2.4: Construct a dual-attention multi-scale fusion U-Net model. The dual-attention multi-scale fusion U-Net model includes the following structure: While retaining the symmetric structure of the classic U-Net encoder and decoder, the modules constructed in steps 2.1, 2.2, and 2.3 are introduced. The network adopts a 4-layer deep design, and the input feature map size is 128×128×128×1. After downsampling and upsampling, the final output is a fault prediction map, and the output feature map size is the same as the input feature map size. In the encoder path, a multi-scale feature fusion module is first used at the first level to extract initial features from the input. The four-branch parallel structure can effectively capture fault features of different scales. Then, a dilated convolution fusion module is introduced at the second and third levels. This module can expand the receptive field by using parallel convolutions with different dilation rates. After each encoding level, a 2×2×2 maximum pooling is also connected for downsampling. After reaching the bottleneck layer, two stacked 3×3×3 convolutions and ReLU activation functions are used to maintain the feature dimension. The number of channels in the entire encoder path is 1, 16, 32, and 64, respectively. The decoder path adopts a symmetrical upsampling structure. First, the bottleneck layer output is upsampled by 2×2×2, and then the upsampled feature map is spliced ​​with the third-layer encoding feature map processed by the dual attention module, and then two stacked 3×3×3 convolutions and Relu activation functions are used for feature fusion; then the feature map output by the third layer of the decoder is used as the feature map of the second layer, and then spliced ​​with the second-layer encoder feature map after dual attention processing, and similarly, two stacked 3×3×3 convolutions and Relu activation functions are used for feature fusion; then the feature map output by the second layer of the decoder is used as the feature map of the first layer, and then spliced ​​with the first-layer encoder feature map after dual attention processing, and then two stacked 3×3×3 convolutions and Relu activation functions are used for feature fusion; finally, the feature map output by the first layer of the decoder is output as a fault probability map through 1×1×1 convolution and Sigmoid activation. The number of channels in this path is 128, 96, 48, 16 and 1 respectively; Step 3: Parameter optimization, the process is as follows: Step 3.1: Construct a hybrid loss function, the calculation formula is as follows: The Dice Loss component is used to display the spatial continuity of the optimized tomographic voxels. The Dice Loss function formula is: Where N is the number of pixels in the 3D seismic image, y i represents the label of pixel i, p i is the prediction of pixel i, ∈=10 -6 Prevent division by zero, L Dice is the Dice Loss function; Focal Loss focuses on difficult samples through dynamic weights. The Focal Loss function formula is: Where N is the number of pixels in the 3D seismic image, y i represents the label of pixel i, p i It represents the prediction of pixel i. α and γ are high parameters and the final values ​​are obtained through multiple experiments. Focal (α,γ) is the Focal Loss function; the combination of FocalLoss and DiceLoss is used as the loss function: L Focal_Dice =L Focal (α,γ)+L Dice , where L Dice is the Dice Loss function, L Focal (α,γ) is the Focal Loss function, L Focal_Dice It is the combined loss function of FocalLoss and Dice Loss. α and γ are both high parameters and the optimal values ​​are obtained through multiple experiments. Step 3.2: Introduce the dual-attention multi-scale fusion U-Net model constructed in step 2.4 into the loss function of step 3.1 and design and combine it into a complete training model; Step 3.3: Set the learning rate lr, batch size batch_size, decay rate decay, training period epoch, and optimizer adam. All parameters are optimized after multiple experiments. Step 3.4: Use the training data from step 1 to train the model from step 3.2, evaluate the model performance using the validation data, and save the model after a complete training cycle. Step 4: Use the two sets of test data in step 1 for the model prediction saved in step 3.4, and output a 3D fault prediction probability map.

Images