A diffraction wave separation method based on a multi-feature reuse hollow convolutional neural network

Abstract

The present application relates to a kind of diffracted wave separation methods based on multi-feature reuse hollow convolutional neural network.Based on hollow convolution and multi-feature reuse module, neural network is built, training data set is constructed by synthesizing data and actual data, neural network is trained, L1-multiple scale structural similarity index hybrid function is used to measure the distance between network output and label in training to obtain optimal network model;Seismic common offset data or poststack data are input into optimal network model, and separated diffracted wave data are obtained;According to seismic migration velocity file, the diffracted wave data separated by multi-feature reuse hollow convolutional neural network are migrated and imaged.This method can effectively remove steep dip reflection while preserving diffracted wave energy, and can solve the problems of diffracted energy loss and reflection residue in the process of diffracted wave separation.For the case where diffracted wave and reflection event cross, the separated diffracted wave is more complete.

Description

Technical Field

[0001] This invention relates to the field of seismic reflection wave exploration technology, specifically a diffraction wave separation method based on a multi-feature reuse holed convolutional neural network. Background Technology

[0002] Accurate characterization of small-scale discontinuous geological bodies such as collapse columns, pinch-outs, karst caves, and minor faults is crucial for safe coal production and oil and gas reservoir exploration. Traditional seismic reflection wave exploration is limited by the Fresnel zone assumption and cannot accurately detect these small-scale subsurface discontinuous geological bodies. While conventional processing techniques such as coherence analysis can highlight fractures and fissures on imaging profiles, they are prone to producing structural artifacts. Diffraction waves carry information about small-scale subsurface discontinuous geological bodies, and diffraction wave imaging is of great significance for improving the resolution of seismic exploration.

[0003] Diffracted and reflected waves coexist in seismic records. Diffracted waves have significantly less energy than reflected waves, making the extraction of diffracted waves crucial for high-resolution diffracted wave imaging. Traditional diffracted wave separation methods include using plane wave prediction filtering, median filtering in dip convergence to separate reflected and diffracted waves, and singular value decomposition to extract diffracted waves. However, these traditional methods are ineffective at removing reflections at steep dip angles, and the removal of reflected waves often results in the loss of diffracted wave energy, hindering high-resolution diffracted wave imaging. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention discloses a diffraction wave separation method based on a multi-feature reuse dilated convolutional neural network, which removes reflected waves while protecting the amplitude of diffraction waves, and solves the problems of difficult removal of reflected waves at steep angles and severe energy loss of diffraction waves separated when the phase axes of reflected waves and diffraction waves intersect.

[0005] To achieve the above objectives, the process of diffraction wave separation in this invention includes the following steps:

[0006] Step 1: Obtain actual co-offset data and post-stack data;

[0007] Step 2: Generate simulated common offset data using the forward modeling algorithm for explosion reflector surfaces based on Fast Fourier Transform;

[0008] Step 3: Perform data augmentation on the actual co-offset data, post-stack data, and simulated co-offset data. Divide the augmented data into training dataset, validation dataset, and test dataset, and construct the corresponding label dataset using the plane wave destruction method.

[0009] Step 4: Construct a multi-feature reuse dilated convolutional neural network framework based on dilated convolution and multi-feature reuse modules;

[0010] Step 5: Train a multi-feature reuse dilated convolutional neural network using the training dataset, and use the L1-multi-scale structural similarity index mixture function as the loss function to evaluate the training results. The training network's loss function value must meet the preset requirements.

[0011] Step 6: Use the multi-feature reuse dilated convolutional neural network with the loss function value meeting the preset requirements as the diffraction wave separation network model, and perform diffraction wave separation on the common offset data of the diffraction wave to be separated to obtain the diffraction wave common offset data.

[0012] Step 7: Perform offset imaging on the separated diffracted waves.

[0013] As a preferred embodiment of the present invention, the plane wave destruction method in step three is specifically as follows:

[0014] o=U(σ)q

[0015] Among them, o=[o1,o2,Ko N ] T The prediction error is the predicted diffracted wave. T is the matrix transpose operation, q is the matrix representation of the seismic co-offset data or post-stack data, and U is the plane wave failure operator, in the following form:

[0016]

[0017] Where I is the identity matrix, and C is the low-order plane wave destruction operator, whose scalar form is as follows:

[0018]

[0019] Among them, Z t Z x These are the Z-transforms of the variables in the time direction t and the spatial direction x, respectively; p i This represents the slope of the earthquake dip angle.

[0020] As a preferred embodiment of the present invention, the specific process of constructing the multi-feature reuse dilated convolutional neural network framework in step four is as follows:

[0021]

[0022] Where, a i Let represent the output feature map of the i-th layer of the multi-feature reuse dilated convolutional neural network, y be the input seismic data, θ represent the learnable parameters of the multi-feature reuse dilated convolutional neural network, ReLU(·) represent the corrected linear unit, Dconv(·,v[i]) represent the two-dimensional dilated convolution, and v[i] represent the dilation rate of the i-th layer of the multi-feature reuse dilated convolutional neural network. The dilation rate of each layer is set as follows:

[0023] v=(1,1,2,5,9,1,2,5,9,1,2,5,9,1,2,5,9,1,1,1,1)

[0024] MFM(·) is a multi-feature reuse module, consisting of modified linear units, batch normalization, two-dimensional dilated convolution, and concatenated feature concatenation, as shown below:

[0025] MFM(a1,a2,K,a i-1 ,d[i])=ReLU(BN(Dconv(Cat(a1,a2,K,a i-1 ),v[i])))

[0026] Where BN(·) represents batch normalization, and Cat(·) represents splicing along the channel direction of the feature map.

[0027] As a preferred embodiment of the present invention, the specific process of training the multi-feature reuse dilated convolutional neural network in step five is as follows:

[0028] Using seismic data y containing reflected and diffracted waves as input, and diffracted wave data d as the desired output, the relationship between y and d is established using the following formula:

[0029] d = Net(y; θ),

[0030] Where θ = {W, b}, W and b are network parameters, where W represents the weight matrix and b represents the bias matrix, and Net(·) represents the network processing procedure; the seismic data y contains reflected wave data r and diffracted wave data d, y = r + d;

[0031] During network training, the error between the actual network output Net(y; θ) and the expected output d continuously decreases. The L1-multiscale structural similarity index mixture function is used to measure the error between the actual network output and d:

[0032]

[0033] Where L(θ) represents the mixture loss function, and G is a Gaussian filter. Let L be the mean absolute error function (MAE), with α set to 0.89. MS-SSIM The multi-scale structural similarity index takes the following form:

[0034]

[0035] Where M = 5 is the total number of scales, μ n ,μ d and σ n ,σ dThe mean and standard deviation of the network outputs Net(y; θ) and d are respectively, σ nd Let C1 = 0.01 be the covariance of these variables. 2 C2 = 0.03 2 ω[m] and κ[m] define the relative importance of the two components at each scale. ω[m] and κ[m] are set as follows for the five scales:

[0036] ω=κ=(0.0448,0.2856,0.3001,0.2363,0.1333)

[0037] The process of network training is a continuous process of updating the network parameters θ. The following optimization formula is used to update the parameters:

[0038]

[0039] Where, θ j Let A represent the neural network parameters at the j-th iteration. j Let B be the first moment of the gradient of L(θ) at the j-th iteration. j Let be the second moment of the gradient of L(θ) at the j-th iteration, where ε is a very small constant to prevent the denominator from being 0, and β j Let be the learning rate at the j-th iteration, which has the following form:

[0040]

[0041] in, Let be the minimum learning rate during the s-th cycle. T represents the maximum learning rate during the s-th cycle. s For the s-th restart cycle, T c This represents the number of rounds elapsed since the (s-1)th restart cycle ended.

[0042] As a preferred embodiment of the present invention, the diffraction wave common offset data obtained in step six is ​​as follows:

[0043] d′=Net′(y′)

[0044] Where y′ represents the number of common offsets of the diffracted waves to be separated, d′ represents the separated diffracted wave data, and Net′(·) represents the multi-feature reuse dilated convolutional neural network when the loss function value meets the preset requirements.

[0045] The beneficial effects of this invention are as follows: This invention achieves adaptive diffraction wave separation of co-offset and post-stack data through a multi-feature reuse diffracted convolutional neural network. Diffraction wave data can be obtained by inputting co-offset or post-stack seismic data. The multi-feature reuse diffracted convolutional neural network can remove steeply dipping angle reflected waves while protecting the amplitude of diffraction waves, thus better protecting the energy of diffraction waves. In cases where the phase axes of reflected waves and diffracted waves intersect, the separated diffraction waves are more complete, improving the quality of diffraction wave separation and achieving high-resolution diffraction wave imaging. This enables accurate detection of small-scale discontinuous geological bodies such as coal seam collapse columns and fracture-cavities related to oil and gas reservoirs, and can be used for geological disaster prevention and control and oil and gas reservoir detection. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the workflow of the present invention.

[0047] Figure 2 This refers to the common offset data of the diffracted waves to be separated in this invention;

[0048] Figure 3 This is the separated diffraction wave data of the present invention. Detailed Implementation

[0049] Example 1

[0050] like Figure 1 As shown, a diffraction wave separation method based on a multi-feature reuse dilated convolutional neural network includes the following steps:

[0051] Step 1: Obtain actual co-offset data and post-stack data;

[0052] Step 2: Generate simulated common offset data using the forward modeling algorithm for explosion reflector surfaces based on Fast Fourier Transform;

[0053] Step 3: Perform data augmentation on the actual co-offset data, post-stack data, and simulated co-offset data. Divide the augmented data into training datasets, validation datasets, and test datasets, and construct corresponding label datasets using the plane wave destruction method. The plane wave destruction method specifically includes:

[0054] o=U(σ)q

[0055] Among them, o=[o1,o2,Ko N ] T The prediction error is the predicted diffracted wave. T is the matrix transpose operation, q is the matrix representation of the seismic co-offset data or post-stack data, and U is the plane wave failure operator, in the following form:

[0056]

[0057] Where I is the identity matrix, and C is the low-order plane wave destruction operator, whose scalar form is as follows:

[0058]

[0059] Among them, Z t Z x These are the Z-transforms of the variables in the time direction t and the spatial direction x, respectively; p i Indicates the slope of the earthquake dip angle;

[0060] Step 4: Construct a multi-feature reuse dilated convolutional neural network framework based on dilated convolution and multi-feature reuse modules, specifically as follows:

[0061]

[0062] Where, a i Let represent the output feature map of the i-th layer of the multi-feature reuse dilated convolutional neural network, y be the input seismic data, θ represent the learnable parameters of the multi-feature reuse dilated convolutional neural network, ReLU(·) represent the corrected linear unit, Dconv(·,v[i]) represent the two-dimensional dilated convolution, and v[i] represent the dilation rate of the i-th layer of the multi-feature reuse dilated convolutional neural network. The dilation rate of each layer is set as follows:

[0063] v=(1,1,2,5,9,1,2,5,9,1,2,5,9,1,2,5,9,1,1,1,1)

[0064] MFM(·) is a multi-feature reuse module, consisting of modified linear units, batch normalization, two-dimensional dilated convolution, and concatenated feature concatenation, as shown below:

[0065] MFM(a1,a2,K,a i-1 ,d[i])=ReLU(BN(Dconv(Cat(a1,a2,K,a i-1 ),v[i])))

[0066] Where BN(·) represents batch normalization, and Cat(·) represents splicing along the channel direction of the feature map;

[0067] Step 5: Train a multi-feature reuse dilated convolutional neural network using the training dataset. Use the L1-multi-scale structural similarity index mixture function as the loss function to evaluate the training results. The training network's direct loss function value must meet preset requirements. Specifically, seismic data containing reflected and diffracted waves can be represented as:

[0068] y = r + d

[0069] Where y represents data containing reflected and diffracted waves, d represents diffracted wave data, and r represents reflected wave data;

[0070] The multi-feature reuse dilated convolutional neural network takes seismic data y containing reflected and diffracted waves as input and diffracted waves d as the desired output. That is, it establishes the relationship between y and d using the following formula:

[0071] d = Net(y; θ)

[0072] Where θ = {W, b}, W and b are network parameters, where W represents the weight matrix, b represents the bias matrix, and Net(·) represents the network processing procedure;

[0073] During network training, the error between the actual network output Net(y; θ) and the expected output d continuously decreases. The L1-multiscale structural similarity index mixture function is used to measure the error between the actual network output and d:

[0074]

[0075] Where L(θ) represents the mixture loss function, and G is a Gaussian filter. Let L be the mean absolute error function (MAE), with α set to 0.89. MS-SSIM The multi-scale structural similarity index takes the following form:

[0076]

[0077] Where M = 5 is the total number of scales, μ n ,μ d and σ n ,σ d The mean and standard deviation of the network outputs Net(y; θ) and d are respectively, σ nd Let C1 = 0.01 be the covariance of these variables. 2 C2 = 0.03 2 ω[m] and κ[m] define the relative importance of the two components at each scale. ω[m] and κ[m] are set as follows for the five scales:

[0078] ω=κ=(0.0448,0.2856,0.3001,0.2363,0.1333)

[0079] The process of network training is essentially a process of continuously updating the network parameters θ, using the following optimization formula to update the parameters:

[0080]

[0081] Where, θ j Let A represent the neural network parameters at the j-th iteration. j Let B be the first moment of the gradient of L(θ) at the j-th iteration. jLet be the second moment of the gradient of L(θ) at the j-th iteration, where ε is a very small constant to prevent the denominator from being 0, and β j Let be the learning rate at the j-th iteration, which has the following form:

[0082]

[0083] in, Let be the minimum learning rate during the s-th cycle. T represents the maximum learning rate during the s-th cycle. s For the s-th restart cycle, T c This represents the number of rounds elapsed since the end of the (s-1)th restart cycle;

[0084] Step Six: Using the multi-feature reuse dilated convolutional neural network whose loss function value meets preset requirements as the diffraction wave separation network model, perform diffraction wave separation on the common offset data of the diffraction waves to be separated, obtaining diffraction wave common offset data; specifically, the diffraction wave separation is performed on the common offset data of the diffraction waves to be separated.

[0085] d′=Net′(y′)

[0086] Where y′ represents the number of common offsets of the diffracted waves to be separated, d′ represents the separated diffracted wave data, and Net′(·) represents the multi-feature reuse dilated convolutional neural network when the loss function value meets the preset requirements; the common offset data of the diffracted waves to be separated are as follows: Figure 2 As shown, the obtained diffracted wave co-offset data are as follows: Figure 3 As shown;

[0087] Step 7: Perform offset imaging on the separated diffracted waves.

[0088] The parts not described in detail in this article are existing technologies.

[0089] While the specific embodiments of the present invention have been described in detail above, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and modifications or variations without creative effort are still within the protection scope of the present invention.

Claims

1. A diffraction wave separation method based on a multi-feature reuse dilated convolutional neural network, characterized in that, The steps include the following: Step 1: Obtain actual co-offset data and post-stack data; Step 2: Generate simulated common offset data using the forward modeling algorithm for explosion reflector surfaces based on Fast Fourier Transform; Step 3: Perform data augmentation on the actual co-offset data, post-stack data, and simulated co-offset data. Divide the augmented data into training dataset, validation dataset, and test dataset, and construct the corresponding label dataset using the plane wave destruction method. Step 4: Construct a multi-feature reuse dilated convolutional neural network framework based on dilated convolution and multi-feature reuse modules; Step 5: Train a multi-feature reuse dilated convolutional neural network using the training dataset, and use the L1-multi-scale structural similarity index mixture function as the loss function to evaluate the training results. The training network's loss function value must meet the preset requirements. Step 6: Use the multi-feature reuse dilated convolutional neural network with the loss function value meeting the preset requirements as the diffraction wave separation network model, and perform diffraction wave separation on the common offset data of the diffraction wave to be separated to obtain the diffraction wave common offset data. Step 7: Perform offset imaging on the separated diffracted waves.

2. The diffraction wave separation method based on a multi-feature reuse dilated convolutional neural network according to claim 1, characterized in that: The plane wave destruction method described in step three is as follows: ， in, The prediction error is the predicted diffracted wave. This is the transpose operation of a matrix. This is a matrix representation of seismic co-offset data or post-stack data. The plane wave destruction operator has the following form: ， in, identity matrix The low-order plane wave destruction operator has the following scalar form: ， in, The time direction of the variables are respectively and spatial direction of Transformation; This represents the slope of the earthquake dip angle.

3. The diffraction wave separation method based on a multi-feature reuse dilated convolutional neural network according to claim 1, characterized in that: The specific process of constructing the multi-feature reuse dilated convolutional neural network framework described in step four is as follows: ， in, This represents a dilated convolutional neural network that reuses multiple features. The output feature mapping of the layer, To input earthquake data, The learnable parameters of a dilated convolutional neural network that allows for multi-feature reuse are represented. Indicates a corrected linear unit. Represents two-dimensional dilated convolution. This represents the dilation rate of the i-th layer in a multi-feature reuse dilated convolutional neural network. The dilation rate of each layer is set as follows: ， This multi-feature reuse module consists of modified linear units, batch normalization, two-dimensional dilated convolutions, and concatenated feature concatenation, and its form is as follows: ， in, For batch normalization, This indicates splicing along the channel direction of the feature map.

4. The diffraction wave separation method based on a multi-feature reuse dilated convolutional neural network according to claim 1, characterized in that: The specific process of training the multi-feature reuse dilated convolutional neural network in step five is as follows: using seismic data containing reflected waves and diffracted waves... As input, diffraction wave data To achieve the desired output, we can establish the following formula: and Relationship: ， in, , , These are all network parameters, among which Represents the weight matrix. Represents the bias matrix. This represents the network processing procedure; earthquake data. Contains reflected wave data and data of diffracted waves , ; During the training process of the network, the actual output of the network... With expected output The error between them is continuously reduced, and the L1-multiscale structural similarity index mixture function is used to measure the difference between the actual output of the network and the output of the network. Error between: ， in, Represents the mixed loss function. It is a Gaussian filter. The mean absolute error function, Set to 0.89, The multi-scale structural similarity index takes the following form: ， Where M=5 is the total number of scales. , and , Network outputs and The mean and standard deviation, Set their covariance , , and The relative importance of the two components at each scale is defined, and the importance of the components at all five scales is also defined. and Set to: ， The process of network training is to train network parameters. During the continuous updating process, the parameters are updated using the following optimization formula: ， in, This represents the neural network parameters at the j-th iteration. When it is the j-th iteration The first moment of the gradient, When it is the j-th iteration The second moment of the gradient, It is a very small constant to prevent the denominator from being 0. Let be the learning rate at the j-th iteration, which has the following form: ， in, Let be the minimum learning rate during the s-th cycle. This represents the maximum learning rate during the s-th cycle. For the s-th restart cycle, This represents the number of rounds elapsed since the (s-1)th restart cycle ended.

5. The diffraction wave separation method based on a multi-feature reuse dilated convolutional neural network according to claim 1, characterized in that: The diffracted wave common offset data obtained in step six are as follows: ， in, This represents the number of common offsets of the diffracted waves to be separated. This represents the separated diffraction wave data. This refers to a multi-feature reuse dilated convolutional neural network where the loss function value meets preset requirements.

Images