Training method and device of network model for determining underground reflectivity

Abstract

The invention provides a training method and device for determining a network model of underground reflectivity. The method comprises the following steps: acquiring observation system information and seismic source wavelets of seismic imaging; generating an initial migration image according to the Sigbee speed model, the observed seismic data and the initial data; determining an approximate fuzzy operator according to the smooth velocity model, the scattering point model and the initial data; performing deblurring processing on the initial offset image through a convolutional neural network model to obtain approximate subsurface reflectivity; performing non-stationary two-dimensional convolution processing on the approximate subsurface reflectivity and the approximate fuzzy operator to obtain an approximate offset image; and iteratively updating the network parameters of the convolutional neural network model to minimize the error between the approximate offset image and the initial offset image to obtain a target convolutional neural network model. By applying the method, the imaging precision of the underground shallow layer and the model boundary is improved, and artifacts between bottom layers in migration imaging are inhibited.

Description

Technical Field

[0001] The present disclosure relates to the field of underground exploration, and in particular to a training method and device for a network model for determining underground reflectivity. Background Art

[0002] Seismic imaging is one of the important methods for detecting and evaluating oil and gas reservoirs. Its ultimate goal is to obtain the most realistic underground reflectivity. As oil and gas exploration develops towards complex geological environments such as deep-sea oil and gas and onshore mountains, traditional accompanying imaging methods are difficult to accurately generate amplitude-balanced, high-resolution, and high-fidelity seismic images. Therefore, the existing technology uses least squares migration to approximate the inverse operator of forward modeling using a linear method to improve imaging accuracy. Least squares migration includes data domain least squares migration and image domain least squares migration. If high-resolution imaging is obtained, the use of data domain least squares migration will lead to an increase in the number of iterations and increase the computational cost; while the use of image domain least squares migration has problems such as uneven imaging amplitude, insufficiently illuminated areas cannot be fully compensated, offset imaging has artifacts, and image resolution is not high. Summary of the invention

[0003] The present disclosure provides a training method, device, electronic device and storage medium for a network model for determining underground reflectivity, so as to at least solve the above technical problems existing in the prior art.

[0004] According to a first aspect of the present disclosure, a training method for a network model for determining underground reflectivity is provided, characterized in that the method comprises: acquiring initial data of seismic imaging, the initial data comprising observation system information and source wavelet; generating an initial offset image according to a Sigbee velocity model, observed seismic data and the initial data; determining an approximate fuzzy operator according to a smooth velocity model, a scattering point model and the initial data, the approximate fuzzy operator being a point spread function; inputting the initial offset image into a convolutional neural network model for defuzzification processing to obtain an approximate underground reflectivity; performing non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate fuzzy operator to obtain an approximate offset image; constructing an objective function, the objective function being the error between the approximate offset image and the initial offset image; iteratively updating the network parameters of the convolutional neural network model based on the objective function to minimize the objective function, thereby obtaining a target convolutional neural network model, the output of which is a target underground reflectivity.

[0005] In one embodiment, the convolutional neural network model includes an encoder part and a decoder part, the encoder part includes five first convolution blocks, each first convolution block includes two convolution layers, a discard layer is set after the second convolution layer of each first convolution block, a convolution attention module is set after each first convolution block, the convolution attention module includes a channel attention module and a spatial attention module, each layer of the encoder part downsamples the initial offset image through a preset convolution kernel, and obtains an enhanced feature map corresponding to the encoder part after layer-by-layer processing; the decoder part includes five second convolution blocks, each second convolution block includes two convolution layers, a discard layer is set after the second convolution layer of each second convolution block, each layer of the decoder part upsamples the enhanced feature map corresponding to the encoder part, and obtains an approximate underground reflectivity after layer-by-layer processing.

[0006] In one possible implementation manner, the approximate underground reflectivity is subjected to non-stationary two-dimensional convolution processing with the approximate blur operator to obtain an approximate offset image, including: interpolating a point spread function to obtain a point spread function of the entire space; and convolving the point spread function of the entire space with the approximate underground reflectivity to obtain an approximate offset image.

[0007] In one possible implementation, the interpolation processing of the point spread function obtains the point spread function of the entire space, including: determining the position of the filter and the corresponding point spread function value according to the sampling points of the filter in the first direction and the second direction; determining the offset of the current point in the first direction and the second direction relative to the position of the four nearest filters, the four nearest filters are respectively the filter in the upper left corner closest to the current point, the filter in the upper right corner closest to the current point, the filter in the lower left corner closest to the current point, and the filter in the lower right corner closest to the current point; determining four interpolation weights of the four nearest filters in the first direction and the second direction; determining the point spread function value of the current point based on the four interpolation weights, the offset of the current point in the first direction and the second direction relative to the position of the four nearest filters, and the point spread function values ​​of the four filters.

[0008] In one possible implementation manner, the iteratively updating the network parameters of the convolutional neural network model based on the objective function includes: iteratively updating the network parameters of the convolutional neural network model according to the following formula: ;in, are learnable network parameters, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, represents the learning rate of the convolutional neural network model, represents the gradient of the objective function with respect to the network parameters, Indicates the number of iterations.

[0009] According to a second aspect of the present disclosure, a training device for a network model for determining underground reflectivity is provided, characterized in that the device comprises: an acquisition module for acquiring initial data of seismic imaging, wherein the initial data comprises observation system information and source wavelet; a generation module for generating an initial migration image according to a Sigbee velocity model, observed seismic data and the initial data; a determination module for determining an approximate fuzzy operator according to a smooth velocity model, a scattering point model and the initial data, wherein the approximate fuzzy operator is a point spread function; a first processing module for inputting the initial migration image into a convolutional neural network model for defuzzification processing to obtain an approximate underground reflectivity; a second processing module for performing a non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate fuzzy operator to obtain an approximate migration image; a construction module for constructing an objective function, wherein the objective function is an error between the approximate migration image and the initial migration image; and an update module for iteratively updating the network parameters of the convolutional neural network model based on the objective function to minimize the objective function and obtain a target convolutional neural network model, wherein the output of the target convolutional neural network model is a target underground reflectivity.

[0010] In one possible implementation, the second processing module includes: a first sub-processing module, used to interpolate the point spread function to obtain a point spread function of the entire space; and a second sub-processing module, used to convolve the point spread function of the entire space with the approximate underground reflectivity to obtain an approximate offset image.

[0011] In one possible implementation manner, the first sub-processing module is specifically used to determine the position of the filter and the corresponding point spread function value according to the sampling points of the filter in the first direction and the second direction; determine the offset of the current point in the first direction and the second direction relative to the positions of the four nearest filters, and the four nearest filters are respectively the filter in the upper left corner closest to the current point, the filter in the upper right corner closest to the current point, the filter in the lower left corner closest to the current point, and the filter in the lower right corner closest to the current point; determine four interpolation weights of the four nearest filters in the first direction and the second direction; determine the point spread function value of the current point based on the four interpolation weights, the offset of the current point in the first direction and the second direction relative to the positions of the four nearest filters, and the point spread function values ​​of the four filters.

[0012] In one possible implementation, the updating module is specifically used to iteratively update the network parameters of the convolutional neural network model according to the following formula: ;in, are learnable network parameters, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, represents the learning rate of the convolutional neural network model, represents the gradient of the objective function with respect to the network parameters, Indicates the number of iterations.

[0013] According to a third aspect of the present disclosure, there is provided an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in the present disclosure.

[0014] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable a computer to execute the method described in the present disclosure.

[0015] The present invention discloses a training method and device for a network model for determining underground reflectivity, which obtains initial data of seismic imaging, wherein the initial data includes observation system information and source wavelet; generates an initial offset image according to a Sigbee velocity model, observed seismic data and the initial data; determines an approximate fuzzy operator according to a smooth velocity model, a scattering point model and the initial data; inputs the initial offset image into a convolutional neural network model for defuzzification processing to obtain an approximate underground reflectivity; performs non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate fuzzy operator to obtain an approximate offset image; constructs an objective function, wherein the objective function is the error between the approximate offset image and the initial offset image; iteratively updates the network parameters of the convolutional neural network model based on the objective function to minimize the objective function, thereby obtaining a target convolutional neural network model, wherein the output of the target convolutional neural network model is the target underground reflectivity.

[0016] This method is applied to determine the underground reflectivity by training a convolutional neural network model. The complexity and irregularity of the underground medium are taken into account, and the underground imaging results can be accurately characterized. The problem of uneven illumination of deep structures is solved. The event axis fracture caused by inaccurate inversion is taken into account, which can make the event axis more continuous. The imaging accuracy of the shallow underground layer and the model boundary is improved, and the artifacts between strata in the offset imaging are suppressed.

[0017] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which: In the drawings, the same or corresponding reference numerals represent the same or corresponding parts.

[0019] Figure 1 The following is a schematic diagram showing the implementation process of a training method for a network model for determining underground reflectivity according to an embodiment of the present disclosure. Figure 1 ; Figure 2 A schematic diagram of the network structure of a convolutional neural network model according to an embodiment of the present disclosure is shown; Figure 3 The following is a schematic diagram showing the implementation process of a training method for a network model for determining underground reflectivity according to an embodiment of the present disclosure. Figure 2 ; Figure 4 A schematic diagram of a module of a training device for a network model for determining underground reflectivity according to an embodiment of the present disclosure is shown; Figure 5 A schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0020] In order to make the purpose, features, and advantages of the present disclosure more obvious and easy to understand, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0021] Figure 1 The following is a schematic diagram showing the implementation process of a training method for a network model for determining underground reflectivity according to an embodiment of the present disclosure. Figure 1 ,include: Step 101, obtaining initial data of seismic imaging, the initial data including observation system information and source wavelet.

[0022] Seismic imaging is a geophysical imaging method that analyzes the kinematic and dynamic characteristics of various phases of observed seismic waves and uses inversion methods to solve information such as the velocity structure and physical parameters of underground media. This application first obtains the initial data of seismic imaging, which includes observation system information and source wavelets, which are source signals that generate seismic waves. The observation system information includes: source location, detection point location, frequency of seismic waves in the forward transmission process, time of seismic waves in the transmission process, and other information.

[0023] Step 102: Generate an initial migration image based on the Sigbee velocity model, the observed seismic data and the initial data.

[0024] Using the Kirchhoff operator to simulate the reflection coefficient model, a seismic record containing the wave propagation information between the source and the receiver can be generated. The Kirchhoff forward modeling observation data can be expressed as: (1); in, represents the detection point position, represents the earthquake source location, Represents the frequency of the seismic wave in the forward transmission process, is the Sigbee speed model, represents the coordinates of underground space, For Stimulate The observation data received at Represents the coordinates in underground space Located at The Green's function generated by a point source is In the uniform background velocity field, The point source is at the detection point The Green's function generated at Source signal The frequency domain representation of is a frequency-dependent filter, Represents the subsurface reflectivity.

[0025] Then, the observed data is migrated by the conjugate transpose of the Kirchhoff operator to reconstruct the position and shape of the underground reflection interface and obtain the initial migration image. According to the dot product test, the Kirchhoff migration can be expressed as: (2); in, represents the initial offset image, represents the complex conjugate, Represents observed earthquake data.

[0026] Step 103: determining an approximate fuzzy operator according to the smooth velocity model, the scattering point model and the initial data, wherein the approximate fuzzy operator is a point spread function.

[0027] According to formula (1), the observed earthquake data can be approximated as follows: (3); in is the true reflectivity of the underground, Represents underground space coordinates .

[0028] Combining formula (2) and (3), we can get: (4); In the above formula Represents the Hessian matrix corresponding to the adjoint operator. The Hessian matrix is ​​converted into the Green function as follows: (5); in and represents the coordinates of underground space, Represents the modulus of a complex signal.

[0029] Furthermore, the initial offset image and Hessian matrix can be converted from the form of Green's function to the form of amplitude and travel time according to the asymptotic expression of Green's function. The following is the asymptotic expression of Green's function: ; ; ; ; and They represent amplitude and travel time respectively. In wave problems, amplitude usually refers to the maximum displacement of the wave, which describes the intensity of the wave. In Green's function theory, amplitude is closely related to the asymptotic behavior of Green's function; travel time usually refers to the time required for the wave to propagate from the source point to the observation point. In Green's function theory, travel time is related to the phase information of Green's function, which describes the speed and direction of wave propagation. Substituting the asymptotic expression of Green's function into formula (2) yields: (6); Performing inverse Fourier transform on formula (6) yields the following formula: (7) Similarly, substituting the asymptotic expression of the Green function into formula (5) yields: (8); Defined as Source signal The autocorrelation function after filtering: (9); The filtered version of the frequency-dependent source signal is defined as , its inverse Fourier transform is , we can get the following derivation: (10).

[0030] It is understandable that calculating the adjoint matrix of the Hessian matrix will bring an extremely high computational burden. Therefore, the point spread function can be used to replace the Hessian matrix as an approximate blur operator. The point spread function (PSF) is the response of an imaging system to a point light source. In image processing, the PSF is used to describe the blur degree of the image, so it can be used for image restoration tasks. The Green's function contained in formula (5) represents the wave field from the source to the Spread to Then by Propagate to the detection point The process of wave field from the detection point Spread to Then by Propagation to the source These two sets of physical processes describe the process of wave field propagation in underground space for a single scattering point. This is highly consistent with the meaning of PSF, so each position in the PSF corresponds to a column of the Hessian matrix, which is the response of the underground imaging point.

[0031] Since each PSF needs to be obtained through one migration, for large models, the construction of the full-space PSF will require the execution of migration operations of the same order of magnitude as the model size, resulting in a waste of computing resources. In order to effectively reduce the computing cost and storage requirements, the Hessian matrix is ​​approximately constructed by setting scattering points to solve the corresponding underground imaging response, and the process of blurring the full-space reflection model by the Hessian operator is converted into a series of PSF blurring processes in the local space. This method not only greatly reduces the computational complexity, but also greatly reduces the storage requirements.

[0032] In the actual calculation process, the point spread function is obtained using the smoothed velocity model and the scattering point model, wherein the smoothed velocity model is the velocity model of the Sigbee velocity model after the dual-axis moving average smoothing process using the filtfilt function.

[0033] Step 104: Input the initial offset image into the convolutional neural network model for deblurring to obtain an approximate underground reflectivity.

[0034] The initial offset image is input into the convolutional neural network model for deblurring. The convolutional neural network model used in this application is a variant of the U-Net network, the U-Net-CBAM network. The U-Net-CBAM network is a deep learning architecture commonly used for image segmentation. Its symmetrical encoder-decoder structure and jump connection help capture multi-scale features and detail information in the image, and can be used for deblurring. When used for deblurring, the U-Net-CBAM network can learn the mapping relationship from blurred images to clear images. Through multiple training and learning, the U-Net-CBAM network can grasp the feature change rules in the blurred image and generate high-quality clear images, that is, obtain approximate underground reflectivity.

[0035] Step 105, performing non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate blur operator to obtain an approximate migration image.

[0036] After obtaining the approximate underground reflectivity, the approximate underground reflectivity and the approximate blur operator are subjected to non-stationary two-dimensional convolution processing to blur the approximate underground reflectivity to obtain an approximate offset image, wherein the non-stationary two-dimensional convolution means that when processing an image, the parameters or characteristics of the convolution kernel will change with different positions or times of the image, and this change can reflect local features or dynamic changes in the image.

[0037] Step 106: construct an objective function, where the objective function is the error between the approximate offset image and the initial offset image.

[0038] The mean square error is used as the loss function to measure the difference between the approximate offset image and the corresponding initial offset image to determine the objective function. Specifically, the square of the bi-norm of the approximate offset image and the initial offset image can be calculated as the difference between the approximate offset image and the initial offset image, that is, the objective function ,in represents the approximate offset image, represents the initial offset image, It is the approximate underground reflectivity output by the convolutional neural network model. For ease of processing, this application uses the dot product of the Hessian matrix and the approximate underground reflectivity to represent the approximate offset image.

[0039] Step 107, iteratively update the network parameters of the convolutional neural network model based on the objective function to minimize the objective function and obtain a target convolutional neural network model, the output of which is the target underground reflectivity.

[0040] In order to make the approximate underground reflectivity output by the convolutional neural network model more accurate and closer to the real underground reflectivity, the inverse problem is solved, the model weights are adjusted using the loss value to minimize the prediction error, the gradient is calculated by back propagation, and the network parameters of the convolutional neural network model are iteratively updated using the Adam optimizer. The smaller the difference between the approximate offset image and the initial offset image, the closer the approximate offset image is to the initial offset image, that is, the more accurate the underground reflectivity output by the convolutional neural network model. Therefore, the network parameters of the convolutional neural network model are iteratively updated based on the objective function. After the network parameters are updated, the initial offset image is processed by the updated convolutional neural network model to obtain the approximate offset image until the difference between the obtained approximate offset image and the initial offset image is minimized, that is, the objective function is minimized. The convolutional neural network model corresponding to the minimum objective function is determined as the target convolutional neural network model, and the output of the target convolutional neural network model is the target underground reflectivity, which is closest to the real underground reflectivity.

[0041] The present application obtains observation system information and source wavelets of seismic imaging; generates an initial offset image according to the Sigbee velocity model, observed seismic data and initial data; determines an approximate fuzzy operator according to a smooth velocity model, a scattering point model and initial data; deblurs the initial offset image through a convolutional neural network model to obtain an approximate underground reflectivity; performs non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate fuzzy operator to obtain an approximate offset image; iteratively updates the network parameters of the convolutional neural network model to minimize the error between the approximate offset image and the initial offset image, and obtains a target convolutional neural network model. This method is applied to determine the underground reflectivity by training a convolutional neural network model, thereby improving the imaging accuracy of the shallow underground layer and the model boundary, suppressing artifacts between the bottom layers in the offset imaging, and improving the illumination of the deep underground structure in the complex structure.

[0042] In one embodiment, the convolutional neural network model includes an encoder part and a decoder part. The encoder part includes five first convolution blocks, each of which includes two convolution layers. A discard layer is set after the second convolution layer of each first convolution block. A convolution attention module is set after each first convolution block. The convolution attention module includes a channel attention module and a spatial attention module. Each layer of the encoder part downsamples the initial offset image through a preset convolution kernel, and an enhanced feature map corresponding to the encoder part is obtained after layer-by-layer processing; the decoder part includes five second convolution blocks, each of which includes two convolution layers. A discard layer is set after the second convolution layer of each second convolution block, and each layer of the decoder part upsamples the enhanced feature map corresponding to the encoder part, and the approximate underground reflectivity is obtained after layer-by-layer processing.

[0043] Through multiple training and learning, the U-Net-CBAM network can grasp the feature change rules in the blurred image and generate high-quality clear images to obtain the underground reflectivity.

[0044] Figure 2 A schematic diagram of the network structure of the convolutional neural network model of an embodiment of the present disclosure is shown.

[0045] like Figure 2 As shown, the convolutional neural network model used in this application is a variant of the U-Net network, the U-Net-CBAM network, and the convolutional neural network model includes an encoder part and a decoder part. The encoder part contains five first convolution blocks, each of which has a structure of two 3×3 convolution layers, and batch normalization and activation functions are connected after the convolution layer. The activation function in this application is the Leaky ReLU activation function; and a drop layer is set after the second convolution layer of each first convolution block. Setting the drop layer can be used to prevent overfitting in the image processing process. In order to achieve better image processing effects, the drop rate of the drop layer is set to 0.005.

[0046] A convolutional attention module is set after each first convolutional block. The convolutional attention module includes a channel attention module and a spatial attention module. The output of each first convolutional block can be enhanced by the channel attention module and the spatial attention module. Specifically, the channel attention module extracts information through adaptive average pooling and adaptive maximum pooling, and generates channel weights through a shared multilayer perceptron (MLP). The spatial attention module generates two single-channel feature maps by calculating the average and maximum values ​​on each channel and splices them, and then generates spatial weights through a 7×7 convolution layer. Finally, the channel weights and spatial weights are multiplied to obtain the output image corresponding to each layer of the encoder part. The encoder part downsamples the output image corresponding to each layer through a preset convolution kernel to reduce the size of the image. This application uses a maximum pooling layer with a preset convolution kernel of 2×2 and a step size of 2 for downsampling, so that the output image corresponding to each layer of the encoder part is halved in size relative to the output image of the previous layer. Each layer of the encoder part is processed through the above process, and the enhanced feature map corresponding to the encoder part is finally obtained after five first convolutional blocks are processed layer by layer. When the size of the initial offset image input to the convolutional neural network model is 1×256×192, the image obtained after processing by the encoder part is reduced to 1024×24×16. In the encoder part, the convolutional attention module is introduced to enhance the ability of feature extraction and information fusion.

[0047] The decoder part also includes five second convolution blocks. Before processing the image through the second convolution block, the decoder part will first upsample the image obtained by the previous convolution block to expand the size of the image, such as first upsampling the enhanced feature map corresponding to the encoder part, and then input the upsampled image into the second convolution block of the decoder part; the upsampling method adopted in this application is the nearest neighbor interpolation upsampling, and the size of the feature map after upsampling is twice the size of the feature map before upsampling, which is used to restore the spatial resolution. The upsampled feature map is spliced ​​with the feature map of the corresponding layer of the encoder part in the channel dimension, and then processed by the second convolution block of the next layer. Each second convolution block also includes two convolution layers, and a discard layer is also set after the second convolution layer of each second convolution block to prevent overfitting during image processing. The operation of each upsampling layer is similar, the number of channels of the output feature map gradually decreases, and the image size of the feature map doubles. The second convolution block of the last layer is used to generate the final prediction result, which is the approximate underground reflectivity.

[0048] The U-Net network is a deep learning architecture commonly used for image segmentation. Its symmetrical encoder-decoder structure and jump connection help capture multi-scale features and detail information in the image, and can be used for deblurring. When used for deblurring, the U-Net network can learn the mapping relationship from blurry images to clear images. The convolutional neural network model of the present application, namely the U-Net-CBAM network, integrates the attention module CBAM into the U-Net network, that is, adding a CBAM module after each layer of the U-Net network, which can better focus on the key areas and features in the image, improve the accuracy of image processing, and generate high-quality clear images to obtain underground reflectivity.

[0049] In one possible implementation, the approximate underground reflectivity and the approximate blur operator are subjected to non-stationary two-dimensional convolution processing to obtain an approximate migration image, including: Interpolating the point spread function to obtain the point spread function of the entire space; The point spread function of the whole space is convolved with the approximate underground reflectivity to obtain the approximate migration image.

[0050] When the approximate fuzzy operator adopted in this application is a point spread function, in order to reduce the amount of calculation, the point spread function obtained by the Hessian matrix is ​​only the point spread function of the local space, which is not accurate or complete enough. At this time, a more accurate and complete point spread function of the whole space is obtained by interpolating the local point spread function. The point spread function of the whole space obtained by interpolation is locally convolved with the corresponding position of the approximate underground reflectivity output by the convolutional neural network model. The specific process of the local convolution is to use the point spread function at a certain position in the point spread function of the whole space as a filter and multiply it with the data of the corresponding area in the approximate underground reflectivity and sum it to obtain the result of the local convolution, and finally accumulate all the results of the local convolution to form an approximate offset image.

[0051] In one possible implementation, Figure 3 As shown, the point spread function is interpolated to obtain the point spread function of the entire space, including: Step 301, determining the position of the filter and the corresponding point spread function value according to the sampling points of the filter in the first direction and the second direction; Step 302, determining the offset of the current point in the first direction and the second direction relative to the positions of the four nearest filters, wherein the four nearest filters are the filter at the upper left corner closest to the current point, the filter at the upper right corner closest to the current point, the filter at the lower left corner closest to the current point, and the filter at the lower right corner closest to the current point; Step 303, determining four interpolation weights of the four nearest filters in the first direction and the second direction; Step 304 : determining a point spread function value of the current point based on four interpolation weights, the offset of the current point in the first direction and the second direction relative to the positions of the four nearest filters, and the point spread function values ​​of the four filters.

[0052] According to the sampling points of the filter in the first direction and the second direction, the position of the filter in the model and the point spread function corresponding to the filter are determined, wherein the first direction and the second direction are the x-axis direction and the z-axis direction. For each current point, the four filters closest to the current point in the x-axis direction and the z-axis direction are calculated. The four filters are the upper left corner filter closest to the current point, the upper right corner filter closest to the current point, the lower left corner filter closest to the current point, and the lower right corner filter closest to the current point. The offset of the current point relative to the position of the four nearest filters is determined according to the position of the four nearest filters. If the calculated filter index exceeds the valid range of the filter, it is limited to the valid range and the interpolation weight is adjusted. According to the calculated offset, the point spread function value of the current point is interpolated from the four nearest filters using the bilinear interpolation method. The interpolation formula is: ; in is the point spread function value of the filter nearest the upper left corner, is the point spread function value of the filter nearest the lower left corner, is the point spread function value of the filter nearest the upper right corner, is the point spread function value of the filter nearest the lower right corner.

[0053] is the x-axis interpolation weight of the left direction filter, and the calculation formula is: ; is the interpolation weight of the right direction filter in the x-axis direction, and the calculation formula is: ; is the interpolation weight of the upper filter in the z-axis direction, and the calculation formula is: ; is the interpolation weight of the lower filter in the z-axis direction, and the calculation formula is: ;in , Respectively represent the index of the model in the x-axis direction and the z-axis direction, , is the position of the first filter in the x-axis and z-axis directions, and are the filter indices closest to the current point (ix, iz) in the x-axis and z-axis directions, respectively. , is the spacing of the filters in the x-axis and z-axis directions.

[0054] In one possible implementation, the network parameters of the convolutional neural network model are iteratively updated based on the objective function, including: The network parameters of the convolutional neural network model are iteratively updated according to the following formula: ; in, are learnable network parameters, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, represents the learning rate of the convolutional neural network model, represents the gradient of the objective function with respect to the network parameters, Indicates the number of iterations.

[0055] The initial offset image is deblurred by the convolutional neural network model to obtain the approximate underground reflectivity, and the obtained approximate underground reflectivity is subjected to non-stationary two-dimensional convolution with the approximate blur operator to obtain the approximate offset image. The mean square error is used as the loss function to measure the difference between the approximate offset image and the corresponding initial offset image to determine the objective function. In order to make the approximate underground reflectivity output by the convolutional neural network model more accurate, the inverse problem is solved. Therefore, the model weights are adjusted using the loss value to minimize the prediction error, and the gradient is calculated by back propagation. The network parameters of the convolutional neural network model are iteratively updated using the Adam optimizer. According to the learning parameters of the convolutional neural network model corresponding to the current iteration, the learning rate of the convolutional neural network model corresponding to the current iteration, and the gradient of the objective function corresponding to the current iteration with respect to the network parameters, the network parameters of the convolutional neural network model for the next processing of the initial offset image are determined. By iteratively updating the learning parameters of the convolutional neural network model, the approximate underground reflectivity output by the convolutional neural network model can be made more accurate.

[0056] Figure 4 A module schematic diagram of a training device for a network model for determining underground reflectivity according to an embodiment of the present disclosure is shown.

[0057] See also Figure 4 According to a second aspect of an embodiment of the present disclosure, a training device for a network model for determining underground reflectivity is provided, the device comprising: an acquisition module 401, used to acquire initial data of seismic imaging, the initial data comprising observation system information and source wavelet; a generation module 402, used to generate an initial offset image according to a Sigbee velocity model, observed seismic data and initial data; a determination module 403, used to determine an approximate fuzzy operator according to a smooth velocity model, a scattering point model and initial data, the approximate fuzzy operator being a point spread function; a first processing module 404, used to input the initial offset image into a convolutional neural network model for defuzzification processing to obtain an approximate underground reflectivity; a second processing module 405, used to perform non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate fuzzy operator to obtain an approximate offset image; a construction module 406, used to construct an objective function, the objective function being the error between the approximate offset image and the initial offset image; an update module 407, used to iteratively update the network parameters of the convolutional neural network model based on the objective function to minimize the objective function and obtain a target convolutional neural network model, the output of the target convolutional neural network model being the target underground reflectivity.

[0058] In one embodiment, the second processing module 405 includes: a first sub-processing module 4051, which is used to interpolate the point spread function to obtain the point spread function of the entire space; and a second sub-processing module 4052, which is used to convolve the point spread function of the entire space with the approximate underground reflectivity to obtain an approximate offset image.

[0059] In one embodiment, the first sub-processing module 4051 is specifically used to determine the position of the filter and the corresponding point spread function value according to the sampling points of the filter in the first direction and the second direction; determine the offset of the current point in the first direction and the second direction relative to the position of the four nearest filters, the four nearest filters are the filter in the upper left corner closest to the current point, the filter in the upper right corner closest to the current point, the filter in the lower left corner closest to the current point, and the filter in the lower right corner closest to the current point; determine four interpolation weights of the four nearest filters in the first direction and the second direction; determine the point spread function value of the current point based on the four interpolation weights, the offset of the current point in the first direction and the second direction relative to the position of the four nearest filters, and the point spread function values ​​of the four filters.

[0060] In one embodiment, the updating module 407 is specifically configured to iteratively update the network parameters of the convolutional neural network model according to the following formula: ;in, are learnable network parameters, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, represents the learning rate of the convolutional neural network model, represents the gradient of the objective function with respect to the network parameters, Indicates the number of iterations.

[0061] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device and a readable storage medium.

[0062] Figure 5 A schematic block diagram of an example electronic device 500 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0063] like Figure 5As shown, the device 500 includes a computing unit 501, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 502 or a computer program loaded from a storage unit 508 into a random access memory (RAM) 503. In the RAM 503, various programs and data required for the operation of the device 500 can also be stored. The computing unit 501, the ROM 502, and the RAM 503 are connected to each other via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0064] A number of components in the device 500 are connected to the I / O interface 505, including: an input unit 506, such as a keyboard, a mouse, etc.; an output unit 507, such as various types of displays, speakers, etc.; a storage unit 508, such as a disk, an optical disk, etc.; and a communication unit 509, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 509 allows the device 500 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0065] The computing unit 501 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 501 performs the various methods and processes described above, such as a training method for a network model for determining underground reflectivity. For example, in some embodiments, a training method for a network model for determining underground reflectivity may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 500 via the ROM 502 and / or the communication unit 509. When the computer program is loaded into the RAM 503 and executed by the computing unit 501, one or more steps of the training method for a network model for determining underground reflectivity described above may be performed. Alternatively, in other embodiments, the computing unit 501 may be configured in any other appropriate manner (eg, by means of firmware) to execute a training method for a network model for determining underground reflectivity.

[0066] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0067] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0068] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0069] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0070] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0071] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0072] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0073] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present disclosure, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0074] The above is only a specific embodiment of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present disclosure, which should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A method for training a network model for determining underground reflectivity, characterized in that: The method comprises: Acquiring initial data of seismic imaging, wherein the initial data includes observation system information and source wavelet; generating an initial migration image based on a Sigbee velocity model, observed seismic data and the initial data; Determining an approximate fuzzy operator according to a smooth velocity model, a scattering point model and the initial data, wherein the approximate fuzzy operator is a point spread function; Inputting the initial offset image into a convolutional neural network model for deblurring to obtain an approximate underground reflectivity; Performing non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate fuzzy operator to obtain an approximate migration image; Constructing an objective function, wherein the objective function is an error between the approximate offset image and the initial offset image; Based on the objective function, the network parameters of the convolutional neural network model are iteratively updated to minimize the objective function, thereby obtaining a target convolutional neural network model, wherein the output of the target convolutional neural network model is the target underground reflectivity.

2. The method according to claim 1, characterized in that The convolutional neural network model includes an encoder part and a decoder part, the encoder part includes five first convolution blocks, each first convolution block includes two convolution layers, a discard layer is set after the second convolution layer of each first convolution block, and a convolution attention module is set after each first convolution block, and the convolution attention module includes a channel attention module and a spatial attention module. Each layer of the encoder part downsamples the initial offset image through a preset convolution kernel, and an enhanced feature map corresponding to the encoder part is obtained after layer-by-layer processing; The decoder part includes five second convolution blocks, each of which includes two convolution layers. A discard layer is set after the second convolution layer of each second convolution block. Each layer of the decoder part upsamples the enhanced feature map corresponding to the encoder part, and obtains the approximate underground reflectivity after layer-by-layer processing.

3. The method according to claim 1, characterized in that The step of performing non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate blur operator to obtain an approximate migration image includes: Interpolating the point spread function to obtain the point spread function of the entire space; The point spread function of the entire space is convolved with the approximate underground reflectivity to obtain an approximate offset image.

4. The method according to claim 3, characterized in that: The interpolation process of the point spread function to obtain the point spread function of the entire space includes: Determine the position of the filter and the corresponding point spread function value according to the sampling points of the filter in the first direction and the second direction; Determine an offset of the current point in the first direction and the second direction relative to the positions of the four nearest filters, wherein the four nearest filters are respectively a filter at the upper left corner closest to the current point, a filter at the upper right corner closest to the current point, a filter at the lower left corner closest to the current point, and a filter at the lower right corner closest to the current point; Determine four interpolation weights of the four nearest filters in the first direction and the second direction; The point spread function value of the current point is determined based on four interpolation weights, offsets of the current point in the first direction and the second direction relative to the positions of the four nearest filters, and the point spread function values ​​of the four filters.

5. The method according to claim 1, characterized in that: The iterative updating of the network parameters of the convolutional neural network model based on the objective function includes: The network parameters of the convolutional neural network model are iteratively updated according to the following formula: ； in, are learnable network parameters, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, Indicates The network parameters of the convolutional neural network model corresponding to the iteration, represents the learning rate of the convolutional neural network model, represents the gradient of the objective function with respect to the network parameters, Indicates the number of iterations.

6. A training device for a network model for determining underground reflectivity, characterized in that: The device comprises: An acquisition module, used to acquire initial data of seismic imaging, wherein the initial data includes observation system information and source wavelet; A generating module, used for generating an initial migration image according to a Sigbee velocity model, observed seismic data and the initial data; A determination module, used for determining an approximate fuzzy operator according to a smooth velocity model, a scattering point model and the initial data, wherein the approximate fuzzy operator is a point spread function; A first processing module is used to input the initial offset image into a convolutional neural network model for deblurring to obtain an approximate underground reflectivity; A second processing module is used to perform non-stationary two-dimensional convolution processing on the approximate underground reflectivity and the approximate fuzzy operator to obtain an approximate migration image; A construction module, used for constructing an objective function, wherein the objective function is an error between the approximate offset image and the initial offset image; An updating module is used to iteratively update the network parameters of the convolutional neural network model based on the objective function to minimize the objective function and obtain a target convolutional neural network model, wherein the output of the target convolutional neural network model is a target underground reflectivity.

7. The device according to claim 6, characterized in that The second processing module comprises: The first sub-processing module is used to perform interpolation processing on the point spread function to obtain the point spread function of the entire space; The second sub-processing module is used to perform convolution processing on the point spread function of the entire space and the approximate underground reflectivity to obtain an approximate offset image.

8. The device according to claim 7, characterized in that The first sub-processing module is specifically used to determine the position of the filter and the corresponding point spread function value according to the sampling points of the filter in the first direction and the second direction; Determine an offset of the current point in the first direction and the second direction relative to the positions of the four nearest filters, wherein the four nearest filters are respectively a filter at the upper left corner closest to the current point, a filter at the upper right corner closest to the current point, a filter at the lower left corner closest to the current point, and a filter at the lower right corner closest to the current point; Determine four interpolation weights of the four nearest filters in the first direction and the second direction; The point spread function value of the current point is determined based on four interpolation weights, offsets of the current point in the first direction and the second direction relative to the positions of the four nearest filters, and the point spread function values ​​of the four filters.

9. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 5.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to make a computer execute the method according to any one of claims 1-5.

Images