Seismic data reconstruction method based on mask-guided attention convolutional neural network

Abstract

The application discloses a seismic data reconstruction method based on a mask-guided attention convolutional neural network. The mask updating mechanism and attention are combined, the context feature relationship is established through the mask-guided attention mechanism, the relationship between local features is established by calculating the cross correlation between different local features, the local features can obtain global information, and the learning ability of the local features is strengthened by introducing more context information. Therefore, the mask-guided attention can establish long-distance feature relationship without significantly increasing the network depth, and effectively fuse local features and global features. The iterative updating operation of the mask guides the reconstruction process of the network, so that each local feature obtains more global information as input, and the problem that the global features are not fully utilized in the current deep learning interpolation method is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of earthquake data reconstruction, specifically a method for earthquake data reconstruction based on a mask-guided attention convolutional neural network. Background Technology

[0002] Seismic data plays a crucial role in oil and gas exploration, geological exploration, and other fields. However, economic costs and geographical barriers prevent the regular arrangement of seismic detectors in the field, often resulting in missing seismic traces in the acquired data, which complicates subsequent processing steps. To improve the completeness and quality of seismic data, it is necessary to reconstruct the missing seismic data.

[0003] Currently, commonly used methods for seismic data reconstruction can be categorized into three types: traditional methods, theory-driven methods, and data-driven methods. Traditional methods rely on certain physical assumptions. For example, Spitz, under the linear assumption of seismic events, proposed using filters estimated from low-frequency components in the frequency space (fx) domain to recover aliased high-frequency data. However, traditional methods require prior knowledge and incur high computational costs, limiting their practical application. Theory-driven methods are data-independent, utilizing mathematical or physical properties such as sparse transforms and low-rank assumptions. Methods based on sparse transforms, such as those proposed by Trickett et al., primarily utilize sparsity constraints, which assume that seismic traces in the transform domain are sparse. By eliminating small-amplitude coefficients, missing seismic traces are gradually recovered.

[0004] In recent years, with the enormous potential of deep learning in image processing and natural language processing, data-driven approaches represented by deep learning algorithms have been extensively studied. For example, Wang B et al. used residual networks, Fang et al. used U-Net networks, and Oliveira et al. used generative adversarial networks to implement seismic data interpolation. However, these methods are mainly for tasks with random missing traces, while the situation of continuously missing seismic traces often occurs, which is a challenge for most interpolation methods because when seismic data has large gaps between consecutive missing traces, the network can only capture local relationships. To solve this problem, Pan S et al. used partial convolution to extract features from the network to recover the signal, and He T et al. designed a multi-stage training process by analyzing the interpolation results of the U-Net training process, training multiple models and using multiple U-Nets to continuously reconstruct missing seismic data of different amplitude ranges at different stages. However, when the trajectory is continuously missing within large gaps, the lack of surrounding information leads to a decrease in interpolation performance and the inability to obtain reasonable results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a seismic data reconstruction method based on a mask-guided attention convolutional neural network.

[0006] The technical solution of this invention to solve the aforementioned technical problem is to provide a seismic data reconstruction method based on a mask-guided attention convolutional neural network, characterized in that the method includes the following steps:

[0007] Step 1: Process the original earthquake dataset to generate earthquake data images of uniform size and perform visualization processing; divide the earthquake data images into training set, validation set and test set;

[0008] The seismic data images include complete seismic images and their corresponding missing seismic images; the missing seismic images and their corresponding binary masks m are used as masks to guide the input of the attention convolutional neural network, and their corresponding complete seismic images are used as labels when calculating the loss function;

[0009] Step 2: Downsample the missing track seismic image and the corresponding binary mask m to generate a low-resolution seismic image LR and the corresponding downsampled binary mask M;

[0010] Step 3: Input the low-resolution seismic image LR and its corresponding downsampled binary mask M into a mask-guided attention convolutional neural network for feature extraction to obtain the high-resolution reconstructed seismic image F for this iteration. out ;

[0011] A mask-guided attention convolutional neural network consists of several downsampled convolutional layers, several residual blocks, a mask-guided attention module, and several upsampled convolutional layers connected in sequence.

[0012] Step 4: Obtain the high-resolution reconstructed seismic image F for this iteration. out Then, the mask-guided attention convolutional neural network for this iteration is optimized using the loss function;

[0013] Step 5: Repeat steps 2 to 4, continuously optimizing the mask-guided attention convolutional neural network through the loss function until the set number of iterations is reached, and obtain the trained mask-guided attention convolutional neural network; then input the missing seismic images in the test set and the corresponding binary masks into the trained mask-guided attention convolutional neural network to complete the seismic data reconstruction and obtain the reconstructed seismic data.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] (1) This invention combines a mask update mechanism with attention. By establishing contextual feature relationships through a mask-guided attention mechanism, and by calculating the cross-correlation between different local features, relationships between local features are established, enabling local features to acquire global information. By introducing more contextual information, the learning ability of local features is enhanced. Therefore, mask-guided attention can establish long-distance feature relationships without significantly increasing network depth, effectively integrating local and global features.

[0016] (2) This invention introduces a masking mechanism to determine the region that needs to be reconstructed, allowing the network to focus only on the region in the mask, which effectively solves the problem of insufficient feature extraction in previous interpolation methods. In addition, the iterative update operation of the mask guides the reconstruction process of the network, so that each local feature can obtain more global information as input, which effectively solves the problem of insufficient utilization of global features in current deep learning interpolation methods.

[0017] (3) Compared with traditional reconstruction methods, the present invention can automatically update parameters without the need for other prior information, and can learn complex seismic data features, effectively avoiding spatial aliasing problems, and can quickly reconstruct clear high-resolution seismic data in a GPU environment.

[0018] (4) The present invention realizes an end-to-end image reconstruction method, avoiding excessive hyperparameters and manual operations.

[0019] (5) The present invention introduces a mean squared error loss function in the loss function part for optimization. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the mask-guided attention convolutional neural network structure of the present invention;

[0021] Figure 2 This is a schematic diagram of the mask-guided attention module of the present invention;

[0022] Figure 3 This is a seismic image with missing traces from Embodiment 1 of the present invention;

[0023] Figure 4 This is a complete seismic image of Embodiment 1 of the present invention;

[0024] Figure 5 The high-resolution reconstructed seismic image F obtained in Embodiment 1 of the present invention out ;

[0025] Figure 6 The image shows the complete seismic image and the residual image of the reconstructed seismic image in Embodiment 1 of the present invention.

[0026] Figure 7This is a graph showing the relationship between the number of iterations (Epoch) and the signal-to-noise ratio (SNR) obtained using the validation set during the training process of the network in Embodiment 1 of the present invention. Detailed Implementation

[0027] Specific embodiments of the present invention are given below. These specific embodiments are only used to further illustrate the present invention and do not limit the scope of protection of the claims of this application.

[0028] This invention provides a seismic data reconstruction method based on a mask-guided attention convolutional neural network (hereinafter referred to as the method), characterized by the following steps:

[0029] Step 1: Process the original earthquake dataset to generate earthquake data images of uniform size and perform visualization processing; divide the earthquake data images into training set, validation set and test set;

[0030] Seismic data images include clear, complete seismic images (referred to as complete seismic images) and corresponding consecutive missing seismic images (referred to as missing seismic images); the missing seismic images and their corresponding binary masks m are used as inputs to the attention convolutional neural network, and their corresponding complete seismic images are used as labels; when calculating the loss function, the smaller the difference between the reconstructed seismic image and the label, the smaller the loss function value.

[0031] Preferably, in step 1, during the process of generating seismic data images, image enhancement is performed by randomly flipping and rotating the seismic data in the original seismic dataset by 90°, 180°, and 270°, thereby increasing the amount of training data and improving the robustness of the model.

[0032] Step 2: In order to facilitate the training of the mask-guided attention convolutional neural network, the missing-channel seismic image and the corresponding binary mask m are downsampled to generate a low-resolution seismic image LR and the corresponding downsampled binary mask M.

[0033] Step 3: Input the low-resolution seismic image LR and its corresponding downsampled binary mask M into a mask-guided attention convolutional neural network for feature extraction to obtain the high-resolution reconstructed seismic image F for this iteration. out ;

[0034] Mask-guided attention convolutional neural networks (e.g.) Figure 1 (As shown) It consists of several downsampling convolutional layers, several residual blocks, a mask-guided attention module, and several upsampling convolutional layers connected in sequence;

[0035] Preferably, in step 3, the number of downsampling convolutional layers and upsampling convolutional layers are determined according to the image size; in this embodiment, there are two.

[0036] Preferably, in step 3, the number of residual blocks is determined based on the balance between the expressive power and computational efficiency of the implementation model, depending on the task, data, and performance requirements. In this embodiment, it is six.

[0037] Preferably, step 3 specifically involves the following steps: The low-resolution seismic image LR and its corresponding downsampled binary mask M are first downsampled several times (twice in this embodiment, the same as the number of downsampled convolutional layers) to compress the spatial size and extract low-level information; then, skip connections are introduced through several consecutive residual blocks (six in this embodiment, the same as the number of residual blocks) to more effectively propagate gradients, thereby refining the image and enabling deeper feature learning. This allows for more effective learning of more complex features and patterns, forming high-level information that aids in training deep networks; then, the mask-guided attention module aggregates global information based on correlation and outputs global features; finally, the global features are upsampled several times (twice in this embodiment, the same as the number of upsampled convolutional layers) to obtain the high-resolution reconstructed seismic image F for this iteration. out .

[0038] Preferably, in step 3, the calculation formula for the mask-guided attention module is:

[0039]

[0040] In equation (1), h′ i v represents the i-th reconstructed patch (or simply the i-th reconstructed patch) in the missing trace portion H of the low-resolution seismic image LR; i D represents the i-th known patch (or simply the i-th known patch) within the complete portion V of the low-resolution seismic image LR; αi Initialize patch h for the i-th patch i Compared with the (i-1)th reconstructed patch h′ (i-1) Cross-correlation between them D βi Initialize patch h for the i-th patch i The i-th known patch v that is closest to its context region i Cross-correlation,

[0041] Preferably, in step 3, the specific process of masking the attention module is as follows (e.g.) Figure 2 As shown):

[0042] S3.1. The first initial patch h1 of the missing trace portion H in the low-resolution seismic image LR is used as the initial patch for the reconstruction process; since there are no reconstructed patches before the first initial patch h1, the cross-correlation degree D is... α1=0, therefore h1 is replaced with the first known patch v1 of the complete part V in the low-resolution seismic image LR, resulting in the first reconstructed patch h′1 (i.e. h′1 = v1 × 1);

[0043] S3.2. Perform a mask update operation on the first reconstructed patch h′1, removing the mask at position h′1 to obtain the second initialized patch h2; then, based on the second initialized patch h2, use h′1 as a convolution filter to obtain the cross-correlation D between h2 and h′1. α2 ;

[0044] Preferably, in step S3.2, the mask update operation specifically includes:

[0045] Where W is the weight of the convolutional filter. It is the i-th reconstructed plaque h′ i The corresponding eigenvalues, M is the downsampled binary mask of the low-resolution seismic image LR, and b is the corresponding bias, expressed as a scaling factor. To adjust the change in input to control the (i+1)th initial patch h i+1 The size of each patch h′ after reconstruction. i This will yield the characteristic values ​​of the reconstructed patch. When eigenvalues If it exists, perform a mask update operation and remove h′. i The location mask is used to obtain the initial patch h. i+1 .

[0046] S3.3. Based on the second initialized patch h2 obtained in step S3.2 and the second known patch v2 that is closest to h2 in its context region, obtain the cross-correlation degree D between h2 and v2. β2 ;

[0047] Preferably, in step S3.3, with h i The closest known patch v in the context region i The following method is used to determine the convolution filter: First, select all known patches v in the complete portion V of the low-resolution seismic image LR as the convolution filter, and then apply the filter to h in the missing portion H of the low-resolution seismic image LR. i By applying this method, we can obtain the values ​​of all known patches v and h in V. i All cross-correlation D between β ; Take all D β The maximum value of v in h is used as the value of h. i The closest known patch v in the context region i .

[0048] S3.4, According to the formula The second reconstructed patch h′2 is calculated, and the reconstruction of h′2 is completed;

[0049] S3.5 Repeat steps S3.2 to S3.4 until the i-th reconstructed patch h′ of the missing trace portion H in the low-resolution seismic image LR is obtained. i And remove h′ i After removing the mask for the location, the removal of all masks and the complete reconstruction operation are now finished.

[0050] Preferably, in step 3, each reconstructed patch h′ i The size of each is 1×1.

[0051] In summary, because the reconstruction operation is an iterative process, h′ i And all previous reconstructed patches (h′1 to h′) i-1 ) and v i Both are related; each reconstructed plaque h′ i Both can obtain more contextual information.

[0052] Preferably, in step 3, the kernel size of the downsampling convolutional layer is a 4×4 dilated convolutional layer, which reduces the spatial size by half while keeping the number of channels unchanged. The dilation expands the receptive field of the convolutional kernel, thereby obtaining more feature information.

[0053] Step 4: Obtain the high-resolution reconstructed seismic image F for this iteration. out Then, the mask-guided attention convolutional neural network for this iteration is optimized using the loss function;

[0054] Step 5: Repeat steps 2 to 4, continuously optimizing the mask-guided attention convolutional neural network through the loss function until the set number of iterations is reached, and obtain the trained mask-guided attention convolutional neural network; then input the missing seismic images in the test set and the corresponding binary masks into the trained mask-guided attention convolutional neural network to complete the seismic data reconstruction and obtain the reconstructed seismic data.

[0055] Preferably, in steps 4 and 5, the loss function is the mean squared error loss function:

[0056]

[0057] In equation (3), M represents the sample size, y m This represents the label of the m-th sample. This represents the m-th reconstructed data.

[0058] Preferably, during the training process in step 5, the SNR (signal-to-noise ratio) is calculated using the seismic data images in the validation set via equation (4), and the number of iterations is determined by observing the SNR value (the number of iterations is obtained when the optimal and stable SNR value is displayed) to avoid overfitting;

[0059]

[0060] In equation (4), I n I and I represent reconstructed data and original data, respectively.

[0061] Preferably, in step 5, SNR is used as an evaluation index to evaluate the reconstruction effect of the reconstructed seismic image. The smaller the difference between the reconstructed seismic image and the complete seismic data, the larger the SNR value, indicating a better reconstruction effect.

[0062] Example 1

[0063] Step 1: In this embodiment, the size is 1×128×128; the seismic data is divided into training set, validation set, and test set in a 7:2:1 ratio. The missing-channel seismic image is a single-channel image with 15% of the channels continuously missing (e.g., ...). Figure 3 As shown), its corresponding complete seismic image (e.g.) Figure 4 (As shown).

[0064] Step 2: Downsampling is performed at 2x to generate a low-resolution seismic image (LR) with dimensions of 256×64×64; the low-resolution seismic image (LR) has 256 channels.

[0065] Step 3: Input the results obtained in Step 2 into a masked guided attention convolutional neural network for feature extraction to obtain the high-resolution reconstructed seismic image F. out (like Figure 5 (As shown). By Figure 5 and Figure 3 The comparison shows that the mask-guided attention convolutional neural network reconstructed the missing parts of the image.

[0066] Six residual blocks serve to refine the information and form higher-level information. Each residual block consists of two convolutional layers and a residual connection. The convolutional kernel size of the convolutional layer is 3×3, the stride is 1, and the padding is 1. Each convolutional layer has 128 convolutional kernels.

[0067] Step 4: Obtain the high-resolution reconstructed seismic image F for this iteration. out Then, the mask-guided attention convolutional neural network for this iteration is optimized using the loss function;

[0068] Step 6: Continuously optimize the mask-guided attention convolutional neural network through the loss function. In order to speed up the convergence speed, the initial learning rate is set to 0.01, and a learning rate adjustment strategy is used. After every 30 iterations of training, the learning rate is changed to 0.1 times the original value. The optimizer is the ADAM optimizer, and the trained mask-guided attention convolutional neural network is obtained.

[0069] Figure 6 This shows the difference between the reconstructed seismic image and the complete seismic image. The closer the color is to white, the smaller the difference; the closer it is to black, the larger the difference. Figure 6 It can be seen that mask-guided attention convolutional neural networks can achieve good reconstruction results.

[0070] Depend on Figure 7 It can be seen that as the number of iterations increases, the SNR of the mask-guided attention convolutional neural network improves and can reach a stable value relatively quickly, with a small fluctuation range after reaching stability. This indicates that the method of using mask-guided attention convolutional neural networks for seismic data reconstruction can achieve better reconstruction results. While extracting features more effectively, it utilizes global image features, accelerates the convergence speed of the model, and thus improves the stability of model training.

[0071] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A seismic data reconstruction method based on a mask-guided attention convolutional neural network, characterized in that, The method includes the following steps: Step 1: Process the original earthquake dataset to generate earthquake data images of uniform size and perform visualization processing; divide the earthquake data images into training set, validation set and test set; The seismic data images include complete seismic images and their corresponding missing seismic images; the missing seismic images and their corresponding binary masks m are used as masks to guide the input of the attention convolutional neural network, and their corresponding complete seismic images are used as labels when calculating the loss function; Step 2: Downsample the missing track seismic image and the corresponding binary mask m to generate a low-resolution seismic image LR and the corresponding downsampled binary mask M; Step 3: Input the low-resolution seismic image LR and its corresponding downsampled binary mask M into a mask-guided attention convolutional neural network for feature extraction to obtain the high-resolution reconstructed seismic image for this iteration. ; A mask-guided attention convolutional neural network consists of several downsampled convolutional layers, several residual blocks, a mask-guided attention module, and several upsampled convolutional layers connected in sequence. The formula for calculating the mask-guided attention module is: （1） In equation (1), The i-th reconstructed patch in the missing trace portion H of the low-resolution seismic image LR; The i-th known patch in the complete portion V of the low-resolution seismic image LR; Initialize the i-th patch Compared with the (i-1)th reconstructed plaque Cross-correlation between them ; Initialize the i-th patch The i-th known patch closest to its context region Cross-correlation, ; The specific process of masking the attention module is as follows: S3.1 Initialize the first patch of the missing channel portion H in the low-resolution seismic image LR. As the initial patch in the reconstruction process; due to the first initial patch Previously, no plaques were reconstructed, therefore the cross-correlation was not considered. =0, therefore the first known patch of the complete portion V in the low-resolution seismic image LR is used. replace The first reconstructed plaque was obtained. ; S3.2, For the first reconstructed patch Perform a mask update operation to remove The location mask is used to obtain the second initialized patch. Then, based on the second initialized patch... ,by As a convolutional filter, obtain and Cross-correlation between ; S3.3, The second initialized patch obtained from step S3.2 and with The second known patch closest to the context region ,get and Cross-correlation between ; S3.4, According to the formula The second reconstructed patch was calculated. ,Finish The reconstruction; S3.

5. Repeat steps S3.2 to S3.4 until the i-th reconstructed patch of the missing trace portion H in the low-resolution seismic image LR is obtained. and remove After removing the mask for the location, the removal of all masks and the complete reconstruction operation are now finished. Step 4: Obtain the high-resolution reconstructed seismic image for this iteration. Then, the mask-guided attention convolutional neural network for this iteration is optimized using the loss function; Step 5: Repeat steps 2 to 4, continuously optimizing the mask-guided attention convolutional neural network through the loss function until the set number of iterations is reached, and obtain the trained mask-guided attention convolutional neural network; then input the missing seismic images in the test set and the corresponding binary masks into the trained mask-guided attention convolutional neural network to complete the seismic data reconstruction and obtain the reconstructed seismic data.

2. The seismic data reconstruction method based on a mask-guided attention convolutional neural network according to claim 1, characterized in that, In step 1, during the process of generating seismic data images, image enhancement is performed by randomly flipping and rotating the seismic data in the original seismic dataset by 90°, 180°, and 270°.

3. The seismic data reconstruction method based on a mask-guided attention convolutional neural network according to claim 1, characterized in that, Step 3 involves the following steps: The low-resolution seismic image LR and its corresponding downsampled binary mask M are first downsampled several times; then, skip connections are introduced through several consecutive residual blocks; next, the mask-guided attention module aggregates global information based on correlations and outputs global features; finally, the global features are upsampled several times to obtain the high-resolution reconstructed seismic image for this iteration. .

4. The seismic data reconstruction method based on a mask-guided attention convolutional neural network according to claim 1, characterized in that, In step S3.2, the mask update operation is specifically as follows: ,in These are the convolutional filter weights. It is the i-th reconstructed plaque The corresponding eigenvalues, It is the downsampled binary mask of the low-resolution seismic image corresponding to LR. This is the corresponding deviation, expressed through a scaling factor. To adjust the change in input to control the (i+1)th initial patch. Size; Each patch is reconstructed This will yield the characteristic values ​​of the reconstructed patch. When the eigenvalue If it exists, perform a mask update operation to remove it. The location mask is used to obtain the initial patch. .

5. The seismic data reconstruction method based on a mask-guided attention convolutional neural network according to claim 1, characterized in that, In step S3.3, with The closest known patch in the context region The following method is used to determine this: First, select all known patches in the complete portion V of the low-resolution seismic image LR. As a convolutional filter, it is applied to the missing trace portion H in the low-resolution seismic image LR. By applying this method, all known patches in V can be obtained. and All cross-correlation between Take all The maximum value in As with The closest known patch in the context region .

6. The seismic data reconstruction method based on a mask-guided attention convolutional neural network according to claim 1 or 3, characterized in that, In step 3, the kernel size of the downsampling convolutional layer is a 4×4 dilated convolutional layer, which reduces the spatial size by half while keeping the number of channels unchanged.

7. The seismic data reconstruction method based on a mask-guided attention convolutional neural network according to claim 1, characterized in that, In steps 4 and 5, the mean squared error loss function is used for both steps: （3） In equation (3), M represents the sample size. This represents the label of the m-th sample. This represents the m-th reconstructed data.

8. The seismic data reconstruction method based on a mask-guided attention convolutional neural network according to claim 1, characterized in that, During the training process in step 5, the SNR is calculated using the seismic data images in the validation set via equation (4), and the number of iterations is determined by observing the value of the SNR. （4） In equation (4), I n I and I represent reconstructed data and original data, respectively; SNR is used as an evaluation index to evaluate the reconstruction effect of reconstructed seismic images. The smaller the difference between the reconstructed seismic image and the complete seismic data, the larger the SNR value, indicating a better reconstruction effect.

Images