Wavelet domain multi-scale seismic data missing region adaptive intelligent repairing method and device

Abstract

The embodiment of the application discloses a wavelet domain multi-scale seismic data missing region adaptive intelligent repairing method and device, wherein the method comprises the following steps: acquiring seismic data to be repaired; inputting the seismic data into a pre-trained first generation network to obtain repaired seismic data; the first generation network comprises an encoder, a residual module and a decoder; wherein the encoder and the decoder are provided with a wavelet convolution module. The generator network is designed as a network structure embedded with the wavelet convolution, which is helpful to enhance the smoothness and accuracy of the interpolation result, improve the reconstruction precision of low-frequency information and retain details.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of seismic data processing, in particular to the field of wavelet domain multi-scale seismic data repair, and more particularly to a wavelet domain multi-scale seismic data missing area adaptive intelligent repair method and device. BACKGROUND

[0002] In actual seismic data acquisition, the data obtained under ideal conditions should have the characteristics of high quality, regularity and density to support subsequent imaging, interpretation and modeling tasks. However, due to the influence of economic cost, environmental conditions and equipment limitations, the acquired seismic data is usually sparse, irregular or missing. Not only does it affect the usability of seismic data, but also it brings challenges to data processing and geological interpretation. SUMMARY

[0003] The present application provides a wavelet domain multi-scale seismic data missing area adaptive intelligent repair method and device, which can optimize the acquired original seismic data and repair the seismic data after optimization to obtain high-quality seismic data.

[0004] In one aspect, the present application provides a wavelet domain multi-scale seismic data missing area adaptive intelligent repair method, which is applied to a wavelet domain multi-scale seismic data missing area adaptive intelligent repair system, and the method comprises:

[0005] acquiring seismic data to be repaired;

[0006] inputting the seismic data into a pre-trained first generation network to obtain repaired seismic data;

[0007] The first generation network comprises an encoder, a residual module and a decoder.

[0008] Among them, wavelet convolution is used in the encoder and the decoder to restore and extract data of multi-level complex architecture.

[0009] The training method of the first generation network is:

[0010] acquiring training data, the training data comprising missing seismic data and corresponding complete seismic data;

[0011] inputting the missing seismic data in the training data into a generator to generate first generation data; calculating the loss of a discriminator using the first generation data and the complete seismic data and updating the parameters of the discriminator;

[0012] calculating the loss of the generator and updating the parameters of the generator until the model reaches a preset performance indicator;

[0013] using the trained generator as the first generation network.

[0014] Through wavelet transform, the signal can be analyzed locally at different scales, which is beneficial to extract seismic signals of different frequency levels; after introducing wavelet convolution in the model, the detailed features can be extracted at different scales, which helps to improve the fine-grained recovery and multi-level complex structure performance of seismic data interpolation. Therefore, for the seismic signal data with multi-scale complex patterns, the generator network is designed as a unet structure embedded with wavelet convolution, which helps to enhance the smoothness and accuracy of the interpolation results, improves the low-frequency information reconstruction accuracy while preserving the details.

[0015] In a possible implementation, the first generation network specifically includes four layers of down-sampling modules of an encoder and four layers of up-sampling modules of a decoder, uses a wavelet convolution layer, specifically adopts a db2 wavelet basis, and sets the decomposition layer number to 2.

[0016] In a possible implementation, the four layers of down-sampling modules of the encoder are implemented as follows: 2-level decomposition is performed on each layer of down-sampling by using a db2 wavelet basis to obtain low-frequency and high-frequency feature maps, and 1024-channel feature maps are obtained through layer-by-layer calculation; the residual module uses a 3x3 convolution; and the four layers of up-sampling modules of the decoder are implemented as follows: up-sampling is performed on each layer of up-sampling by using a db2 wavelet basis, and data images are restored to the same spatial resolution as the original image through layer-by-layer calculation.

[0017] The network of the present application adopts a db2 wavelet basis, and sets the decomposition layer number to 2. The db2 wavelet basis (Daubechies 2) can better extract the details and overall trend of the seismic signal through its balanced low-frequency and high-frequency characteristics. Compared with higher-order wavelet bases (such as db4 or db6), the db2 has lower computational complexity, is suitable for processing large-scale seismic data, the two-level decomposition helps to extract multi-scale features, enhances the noise suppression capability of the model, and provides a balance between computational efficiency and feature extraction capability, so that the generator can more effectively capture key features when processing complex seismic data.

[0018] In a possible implementation, the discriminator includes a convolution layer and seven spectral normalization convolution blocks connected in sequence; each convolution block includes a convolution layer, an SN layer and a Leaky ReLU activation function; the kernel size of each convolution is 3x3, the number of feature maps increases from 64 to 512, and a convolution with a step of 2 is used.

[0019] The seismic data sequence of the application passes through a convolution layer and seven spectral normalization convolution blocks (Conv SD-Block). Each convolution block includes a convolution layer, an SN layer and a LeakyReLU activation function. The kernel size of each convolution is 3*3, the number of feature maps increases from 64 to 512, and a convolution with a step of 2 is used to make the network pay attention to a larger range of data and extract more global features; it should be noted that the SN layer is used to replace the BN layer, which can make the discriminator more sensitive to data details, reduce overfitting and improve the learning effect of local features by limiting the spectral norm of each layer weight.

[0020] In a possible implementation, the loss function in the training manner of the first generation network comprises:

[0021] l G =l rec +αl per +βl adv +λ ltv

[0022] Wherein, l G is the loss function of the generator, and α, β and λ are weights corresponding to each loss term, used to balance the influence of the loss on training; l rec represents the reconstruction loss calculated using the mean square error (MSE), used to measure the pixel difference between the generated data and the target data, and the calculation is the MSE of the generated data and the corresponding complete data d R , which is expressed as:

[0023]

[0024] W and H represent the dimensions of the input missing seismic data d M , and t represents the scale of resolution improvement from d M to d R . is the corresponding complete data;

[0025] The perceptual loss l per is a loss function for measuring the perceptual quality difference between the generated data and the target data, and is expressed as:

[0026]

[0027] Wherein φ i，j represents the operation of extracting the feature map of the jth convolution before the ith max pooling layer, W i，j and H i，j represent the dimensions of each feature map in the VGG network;

[0028] l adv is the reconstruction loss; ltv is the total variation loss;

[0029] The reconstruction loss is represented as:

[0030]

[0031] where N represents the missing data d M the number of samples, represents the probability that the discriminator considers the reconstructed data as real complete seismic data; to obtain better gradient behavior, the minimization instead of

[0032] The total variation loss l tv is represented as:

[0033]

[0034] where M is the total number of pixels of the data image, used for normalization; X is the input data or feature map; (i,j) represents the pixel coordinates; the first term (X i+1，j -X i，j ) 2 calculates the amount of pixel change in the vertical direction; the second term (X i，j+1 -X i，j ) 2 calculates the amount of pixel change in the horizontal direction.

[0035] It should be noted that the loss function of the generator is crucial for the quality of the generated data. Using only pixel-level loss (such as L1 loss) in seismic data interpolation often produces blurred and over-smoothed results, which will lead to inaccuracy in subsequent imaging. Therefore, the VGG loss is integrated into the loss function of the generator as a perceptual loss. The loss is calculated using the feature maps extracted by a pre-trained 16-layer VGG network. It maps the seismic data to a higher-dimensional feature space and compares the distance between the generated data and the target data in this space. This process effectively utilizes the high-level features of the VGG network in the feature representation space to more accurately evaluate the structural similarity between the reconstructed data and the real data. Therefore, the VGG loss can encourage the model to generate reconstructed results with more rich details and higher perceptual quality during the optimization process. The total variation loss l tv is a loss function used to measure the smoothness of data or feature maps, mainly used to reduce noise and artifacts in data while maintaining the smoothness and structural continuity of data. When used in generative adversarial networks, the total variation loss can reduce the clutter and irregular textures of generated data, improving data quality.

[0036] In another aspect, the embodiment of the present application provides a wavelet domain multi-scale seismic data missing region adaptive intelligent repairing device,

[0037] The seismic data processing device executes the wavelet domain multi-scale seismic data missing region adaptive intelligent repairing method, and the steps include: obtaining seismic data to be repaired; inputting the seismic data into a pre-trained first generation network to obtain repaired seismic data; the first generation network includes an encoder, a residual module and a decoder; wherein the encoder and the decoder are provided with a wavelet convolution module; wherein the training method of the first generation network is: obtaining training data, the training data including missing seismic data and corresponding complete seismic data; inputting the missing seismic data in the training data into a generator to generate first generation data; calculating the loss of a discriminator using the first generation data and the complete seismic data and updating the parameters of the discriminator; calculating the loss of the generator through the discriminator and updating the parameters of the generator until the model reaches a preset performance index; and taking the trained generator as the first generation network.

[0038] In another aspect, the present application provides an electronic device, comprising: a processor adapted to implement one or more instructions; and a computer storage medium, the computer storage medium storing one or more instructions, the one or more instructions being adapted to be loaded and executed by the processor to implement the above-mentioned wavelet domain multi-scale seismic data missing region adaptive intelligent repairing method.

[0039] In another aspect, the present application provides a computer storage medium, characterized in that the computer storage medium stores one or more instructions, the one or more instructions being adapted to be loaded and executed by a processor to implement the above-mentioned wavelet domain multi-scale seismic data missing region adaptive intelligent repairing method.

[0040] Therefore, the present application extracts detailed features on different scales by using wavelet convolution in the network, which helps the seismic data interpolation to recover in fine granularity and complex multi-level structure. Specifically, for the characteristics of seismic data, the db2 wavelet basis is used, and the decomposition layer is set to 2, which can better extract the details and overall trend of the seismic signal. Compared with higher-order wavelet bases (such as db4 or db6), the calculation complexity of db2 is lower, which is suitable for processing large-scale seismic data, and two-level decomposition helps to extract multi-scale features, enhances the noise suppression ability of the model, and provides a balance between calculation efficiency and feature extraction capability, so that the generator can more effectively capture key features when processing complex seismic data.

[0041] Furthermore, the spectral normalization convolution block is used in the present application, which can make the discriminator more sensitive to data details by limiting the spectral norm of each layer of weights, reduce overfitting, and improve the learning effect of local features.

[0042] In addition, in the setting of the loss function, the problem that only pixel-level loss (such as L1 loss) is used in the interpolation of seismic data tends to produce blurred and over-smoothed results. The VGG loss is integrated into the loss function of the generator as a perception loss, which effectively utilizes the high-level features of the VGG network in the feature representation space to more accurately evaluate the structural similarity between the reconstructed data and the real data. When the total variation loss is used in the generative adversarial network, the clutter and irregular texture of the generated data can be reduced, and the data quality can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0044] Figure 1 is a schematic diagram of the generator structure of the WTCGAN network designed by the present application provided by the embodiments of the present application;

[0045] Figure 2 is a flowchart of a wavelet domain multi-scale seismic data missing area adaptive intelligent repair method provided by the embodiments of the present application;

[0046] Figure 3 is a schematic diagram of a discriminator network structure provided by the embodiments of the present application;

[0047] Figure 4 is a schematic diagram of seismic data provided by the embodiments of the present application; Figure 4 (a) is ideal seismic data; Figure 4 (b) is seismic data with 60% random missing; Figure 4 (c) is seismic data with 70% random missing;

[0048] Figure 5 is a schematic diagram of seismic data F-K spectrum provided by the embodiments of the present application; Figure 5 (a)-(c) are F-K spectra in (a)-(c), respectively; Figure 4 (a)-(c) are F-K spectra in (a)-(c), respectively;

[0049] Figure 6 is a reconstruction result of 60% random missing provided by the embodiments of the present application; Figure 6 (a) is a result of traditional U-Net reconstruction, SNR=16.34dB; Figure 6 (b) is a result of U-NetGAN reconstruction, SNR=23.07dB; Figure 6(c) is the result of reconstruction by WTCGAN, SNR = 29.11dB;

[0050] Figure 7 is the reconstruction result of 70% random missing provided by the embodiment of the application; Figure 7 (a) is the result of traditional U-Net reconstruction, SNR = 12.60dB; Figure 7 (b) is the result of reconstruction by U-NetGAN, SNR = 21.04dB; Figure 7 (c) is the result of reconstruction by WTCGAN, SNR = 28.18dB.

[0051] Figure 8 is a structural schematic diagram of an electronic device provided by an embodiment of the application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0053] It should be noted that the terms "first", "second", and the like in the specification and claims of the embodiments of the application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order.

[0054] Based on the above description, the embodiment of the application provides a wavelet domain multi-scale seismic data missing area adaptive intelligent repair method, which is applied to a wavelet domain multi-scale seismic data missing area adaptive intelligent repair system,

[0055] The method comprises the following steps.

[0056] S201, acquiring seismic data to be repaired;

[0057] S202, inputting the seismic data into a pre-trained first generation network to obtain repaired seismic data;

[0058] The first generation network comprises an encoder, a residual module, and a decoder.

[0059] Among the encoder and the decoder, wavelet convolution is used to restore and extract data of a multi-level complex architecture; a first generation network schematic diagram is as shown in Figure 1 .

[0060] The training mode of the first generation network is as follows.

[0061] Obtaining training data, the training data comprising missing seismic data and corresponding complete seismic data;

[0062] inputting the missing seismic data in the training data into a generator to generate first generated data; calculating a loss of a discriminator using the first generated data and the complete seismic data and updating parameters of the discriminator;

[0063] calculating a loss of the generator and updating parameters of the generator until the model reaches a preset performance index;

[0064] using the trained generator as a first generation network.

[0065] Through wavelet transform, the signal can be analyzed locally at different scales, and it is particularly suitable for seismic signals containing different frequency levels; after introducing wavelet convolution in the model, the detailed features can be extracted at different scales, which is helpful for the performance of seismic data interpolation which needs fine-grained recovery and multi-level complex structure. Therefore, for seismic signals which have multi-scale complex patterns, the generator network is designed as a unet structure embedded with wavelet convolution, which helps to enhance the smoothness and accuracy of the interpolation results, to improve the reconstruction accuracy of low-frequency information while preserving details. The introduction of the residual module further enhances the non-linear modeling capability, making the interpolation results more consistent in high-frequency and low-frequency features, avoiding false oscillation or waveform distortion.

[0066] In a possible implementation, the first generation network specifically includes four layers of down-sampling modules of an encoder and four layers of up-sampling modules of a decoder, uses a wavelet convolution layer, specifically adopts a db2 wavelet basis, and sets the decomposition layer number to 2.

[0067] In a possible implementation, the four layers of down-sampling modules of the encoder are implemented as follows: at each layer of down-sampling, 2-level decomposition is performed through a db2 wavelet basis to obtain low-frequency and high-frequency feature maps, and a 1024-channel feature map is calculated layer by layer; the residual module uses a 3x3 convolution; and the four layers of up-sampling modules of the decoder are implemented as follows: at each layer of up-sampling, up-sampling is performed through a db2 wavelet basis, and the data image is restored to the same spatial resolution as the original image layer by layer.

[0068] The network of the application adopts a db2 wavelet basis, sets the decomposition layer number to 2, and through the balanced low-frequency and high-frequency characteristics of the db2 wavelet basis (Daubechies 2), the details and overall trend of the seismic signal can be better extracted. Compared with higher-order wavelet bases (such as db4 or db6), the calculation complexity of db2 is lower, which is suitable for processing large-scale seismic data, two-level decomposition helps to extract multi-scale features and enhance the noise suppression capability of the model, and a balance is provided between the calculation efficiency and the feature extraction capability, so that the generator can more effectively capture key features when processing complex seismic data.

[0069] In a possible implementation, the discriminator includes one convolutional layer and seven spectral normalization convolutional blocks connected in sequence; each convolutional block includes a convolutional layer, an SN layer and a Leaky ReLU activation function; the kernel size of each convolution is 3*3, the number of feature maps increases from 64 to 512, and a convolution with a step of 2 is used to make the network pay attention to a larger range of data.

[0070] The seismic data sequence of the application passes through a convolutional layer and seven spectral normalization convolutional blocks (Conv SD-Block). Each convolutional block includes a convolutional layer, an SN layer and a Leaky ReLU activation function. The kernel size of each convolution is 3*3, the number of feature maps increases from 64 to 512, and a convolution with a step of 2 is used to make the network pay attention to a larger range of data and extract more global features. It should be noted that the SN layer is used to replace the BN layer, which can make the discriminator more sensitive to data details, reduce overfitting and improve the learning effect of local features by limiting the spectral norm of each layer weight. The network structure of the discriminator is as shown in Figure 3 .

[0071] In a possible implementation, the loss function in the training mode of the first generation network includes: lG=lrec+αlper+βladv+λltv

[0072] l G =l rec +αl per +βl adv +λl tv

[0073] Wherein, l G is the loss function of the generator, and α, β and λ are weights corresponding to each loss term, used to balance the influence of the loss on training; l rec represents the reconstruction loss calculated using the mean square error (MSE), used to measure the pixel difference between the generated data and the target data, and the MSE of the generated data and the corresponding complete data d R is calculated as follows:

[0074]

[0075] W and H represent the dimensions of the input missing seismic data d M , and t represents the scale of resolution improvement from d M to d R . is the corresponding complete data;

[0076] The perceptual loss l per is a loss function for measuring the perceptual quality difference between the generated data and the target data, and is represented as:

[0077]

[0078] where φ i,j denotes the operation of extracting the feature map of the jth convolution before the ith max-pooling layer, W i,j and H i，j denote the dimensions of each feature map within the VGG network;

[0079] l adv is the reconstruction loss; l tv is the total variation loss;

[0080] The reconstruction loss is expressed as:

[0081]

[0082] where N denotes the number of missing data d M samples, denotes the probability that the discriminator considers the reconstructed data as real complete seismic data; in order to obtain better gradient behavior, the minimization is not

[0083] The total variation loss l tv is expressed as:

[0084]

[0085] where M is the total number of pixels of the data image, used for normalization; X is the input data or feature map; (i,j) denotes the pixel coordinates; the first term (X i+1，j -X i，j ) 2 calculates the amount of change in the vertical direction of the pixel; the second term (X i，j+1 -X i，j ) 2 calculates the amount of change in the horizontal direction of the pixel.

[0086] It is noted that the loss function of the generator is crucial to the quality of the generated data. Using only pixel-level loss (such as L1 loss) in seismic data interpolation often produces blurred and over-smoothed results, which will lead to inaccuracy in subsequent imaging. Because, the VGG loss is integrated into the loss function of the generator as a perceptual loss. We use the feature maps extracted by the pre-trained 16-layer VGG network to calculate the loss. It maps the seismic data to a higher-dimensional feature space and compares the distance between the generated data and the target data in this space. This process effectively utilizes the high-level features of the VGG network in the feature representation space to more accurately evaluate the structural similarity between the reconstructed data and the real data. Therefore, the VGG loss can encourage the model to generate reconstructed results with more detailed and higher perceptual quality during the optimization process.

[0087] l adv is defined based on the probability of identifying the reconstructed data as real data, in order to achieve the purpose of deceiving the discriminator, the generator learns as much as possible the distribution of real complete and intact seismic data.

[0088] Finally, the total variation loss l tv is a loss function used to measure the smoothness of data or feature maps, mainly used to reduce noise and artifacts in data while maintaining the smoothness and structural continuity of data. When used in generative adversarial networks, the total variation loss can reduce the clutter and irregular texture of generated data, improving data quality.

[0089] In one possible implementation, the signal-to-noise ratio (SNR) and structural similarity (SSIM) indicators are used to evaluate the quality of the interpolated seismic data. The formulas for calculating SNR and SSIM are as follows:

[0090] SNR (dB) = 10 log y 2 / (y-G(x)) 2

[0091]

[0092] where y and G(x) represent ideal and recovered seismic data, respectively; μ y and μ G(x) represent the average of the data image pixels of y and G(x), respectively; σ y and σ G(x) represent the variance of the data image pixels of y and G(x), respectively; σ yG(x) represents the covariance of the data image pixels of y and G(x); and c1 and c2 represent the stability constants for calculation.

[0093] The experiment adopts a relevant data set, selects 80 shots of seismic data, divides the data set into a training set of 64 shots, a validation set of 8 shots, and a test set of 8 shots according to the proportions of 80%, 10%, and 10%, and fixes the random seed to perform 60% random loss on the data to test the interpolation ability of the model. In order to test the interpolation ability and robustness of the model under more stringent and more serious data loss conditions, the data is also randomly lost by 70%, and the training set and the validation set are divided into seismic data of 256x256 size. The actual data uses a marine acquisition data set, and 120 shots of data are randomly selected from the data set. Similarly, the data set is divided into a training set of 96 shots, a validation set of 12 shots, and a test set of 12 shots according to the proportions of 80%, 10%, and 10%, and the random seed is fixed to perform 60% and 70% loss on the data, and then the data set is divided into a data block of 512x120. The batch size of the experiment is set to 8, the training round is set to 100, and the learning rate of the generator and the discriminator is set to 0.0004 and 0.0008 respectively to ensure that the generator can be gradually optimized during the training process while reducing the risk of training instability and pattern collapse. In the loss function, the weight coefficient of the generator's adversarial loss is 0.05, the weight coefficient of the perception loss is 0.04, and the weight coefficient of the total variation loss is 0.001. In addition, the Adam optimizer is used for parameter update during the training process. The network input and output size is 256x256 slice data, the experiment is implemented based on the PyTorch deep learning framework, and the RTX 4060Ti GPU is used for training. In order to evaluate the performance of the proposed method, we designed two groups of experiments for 60% and 70% random loss of seismic data. On this basis, we use UNET and UNETGAN networks to perform the same interpolation task for comparative analysis.

[0094] The present application compares the UNET-based method, UNETGAN, and WTCGAN of the present application on complex synthetic seismic records. In order to simulate the loss of data, Figure 4 60% of the traces in (a) are randomly deleted, as shown in Figure 4 (b), and the more stringent 70% random loss case is shown in Figure 4 (c). Further analysis of the effect of random loss on frequency domain features, Figure 5 shows the F-K spectrum corresponding to Figure 4 (a), Figure 4 (b), and Figure 4 (c). Figure 5 (a) is the F-K spectrum of the complete seismic record, reflecting the frequency characteristics and energy distribution of the original data. Figure 5 (b) is the F-K spectrum of the 60% randomly lost data, and it can be observed that there is significant loss of frequency information and destruction of the spectral structure due to the loss of traces. Figure 5(c) is the F-K spectrum of 70% random missing data, the information loss in frequency domain is more serious compared with 60% random missing, especially in high frequency part, showing more obvious energy attenuation and aliasing phenomenon. Figure 6 The reconstruction results of 60% random missing are shown: Figure 6 (a) is the result of traditional U-Net reconstruction, SNR = 16.34 dB; Figure 6 (b) is the result of reconstruction by U-NetGAN, SNR = 23.07 dB; Figure 6 (c) is the result of reconstruction by the scheme of the application, SNR = 29.11 dB. Figure 7 The reconstruction results of 70% random missing are shown: Figure 7 (a) is the result of traditional U-Net reconstruction, SNR = 12.60 dB; Figure 7 (b) is the result of reconstruction by U-NetGAN, SNR = 21.04 dB; Figure 7 (c) is the result of reconstruction by WTCGAN, SNR = 28.18 dB. It can be seen that the method of the application exhibits superior reconstruction characteristics.

[0095] In another aspect, the embodiments of the application provide a wavelet domain multi-scale seismic data missing area adaptive intelligent repairing device,

[0096] The wavelet domain multi-scale seismic data missing area adaptive intelligent repairing system executes a wavelet domain multi-scale seismic data missing area adaptive intelligent repairing method, and the steps include:

[0097] Obtaining seismic data to be repaired;

[0098] Inputting the seismic data into a pre-trained first generation network to obtain repaired seismic data;

[0099] The first generation network includes an encoder, a residual module and a decoder:

[0100] Among them, the wavelet convolution module is arranged in the encoder and the decoder;

[0101] The training method of the first generation network is:

[0102] Obtaining training data, the training data including missing seismic data and corresponding complete seismic data;

[0103] Inputting the missing seismic data in the training data into the generator to generate first generation data; calculating the loss of the discriminator using the first generation data and the complete seismic data and updating the parameters of the discriminator;

[0104] The loss of the generator is calculated by the discriminator, and the parameters of the generator are updated until the model reaches a preset performance index.

[0105] The trained generator is used as the first generation network.

[0106] The embodiment of the present application further provides a computer storage medium (Memory), which is a memory device in an electronic device and is used for storing programs and data. It can be understood that the computer storage medium herein can include a built-in storage medium in the electronic device, and of course can also include an expansion storage medium supported by the electronic device. The computer storage medium provides a storage space, and the storage space stores an operating system of the electronic device. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory; and optionally, at least one computer storage medium located away from the processor.

[0107] Figure 8 is a structural schematic diagram of an electronic device provided by the embodiment of the present application. The electronic device comprises at least one processor 301, for example, a central processing unit (CPU), at least one memory 302, and at least one bus 303.

[0108] The memory 302 can store program instructions, and the processor 301 can be used to call the program instructions to execute a wavelet domain multi-scale seismic data missing region adaptive intelligent repair method.

[0109] Those skilled in the art can understand that all or part of the steps of various methods of the above embodiments can be completed by instructing the relevant hardware by a program, and the program can be stored in a computer readable storage medium, and the storage medium includes a read only memory (ROM), a random access memory (RAM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), a one-time programmable read only memory (OTPROM), an electrically-erasable programmable read only memory (EEPROM), a compact disc read-only memory (CD-ROM), a solid state disk (SSD) or other optical disk storage, a magnetic disk storage, a magnetic tape storage, or any other medium capable of carrying or storing data which can be read by a computer.

[0110] It should be noted that all the steps of the embodiments of the present application are executed under legal and compliant conditions, i.e., all the steps of the embodiments of the present application are executed under authorization.

[0111] The above only shows the preferred embodiments of the present application, and of course cannot limit the scope of the right of the present application, so the equivalent changes made according to the claims of the present application still fall within the scope of the present application.

Claims

1. A wavelet domain multi-scale seismic data missing region adaptive intelligent repair method, characterized in that, The method is applied to a wavelet domain multi-scale seismic data missing area adaptive intelligent repair system, and the method comprises: acquiring seismic data to be repaired; inputting the seismic data into a pre-trained first generation network to obtain repaired seismic data; the first generation network comprises an encoder, a residual module and a decoder: wherein a wavelet convolution module is arranged in the encoder and the decoder; wherein the training method of the first generation network is: obtaining training data, wherein the training data comprises missing seismic data and corresponding complete seismic data; inputting the missing seismic data in the training data into a generator to generate first generation data; calculating the loss of a discriminator using the first generation data and the complete seismic data and updating the parameters of the discriminator; updating the parameters of the generator through the discriminator until the model reaches a preset performance index; the trained generator is used as the first generation network; the first generation network specifically comprises four down-sampling modules of the encoder and four up-sampling modules of the decoder, uses a wavelet convolution layer, specifically adopts a db2 wavelet basis, and sets the decomposition layer number to 2; the implementation of the four down-sampling modules of the encoder is that 2-level decomposition is performed on each layer of down-sampling through the db2 wavelet basis to obtain low-frequency and high-frequency feature maps, and a 1024-channel feature map is calculated layer by layer; the residual module uses a 3x3 convolution; the implementation of the four up-sampling modules of the decoder is that up-sampling is performed on each layer of up-sampling through the db2 wavelet basis, and the data image is restored to the same spatial resolution as the original image through layer-by-layer calculation; the loss function in the training method of the first generation network comprises: l G = l rec + αl per + βl adv + λl tv wherein, l G is the loss function of the generator, and a, b, l are weights corresponding to each loss term, used to balance the influence of the loss on training; l rec represents the reconstruction loss calculated using the mean square error (MSE), used to measure the pixel difference between the generated data and the target data, and the MSE of the generated data and its corresponding complete data d R is represented as: W and H represent the input missing seismic data d M of dimension t, t represents the scale of resolution improvement from d M to d R ; is the corresponding complete data; perceptual loss l per is a loss function that measures the difference in perceptual quality between the generated data and the target data, represented as: where φ i,j represents an operation of extracting a feature map of the jth convolution before the ith max-pooling layer, W i，j and H ij represent dimensions of respective feature maps within the VGG network; x, y represent counting values; l adv is a reconstruction loss;l tv is a total variation loss.

2. The method of claim 1, wherein the method further comprises: the discriminator comprises a convolution layer and seven spectral normalization convolution blocks connected in sequence: each convolution block comprises a convolution layer, an SN layer and a Leaky ReLU activation function; the kernel size of each convolution is 3x3, the number of feature maps increases from 64 to 512, and a convolution with a step of 2 is used.

3. The method of claim 2, wherein the method further comprises: the reconstruction loss is represented as: where N denotes the missing data d M the number of samples, denotes the probability that the discriminator considers the reconstructed data as real complete seismic data; to obtain better gradient behavior, the instead of The total variation loss l tv is represented as: where M is the total number of pixels of the data image for normalization; X is the input data or feature map; (i,j) represents the pixel coordinate; the first term (X i+1，j X i，j ) 2 calculates the pixel change amount in the vertical direction; the second term (X i,j+1 X i，j ) 2 calculates the pixel change amount in the horizontal direction.

4. The method of claim 1, wherein the method further comprises: the signal-to-noise ratio (SNR) and the structural similarity (SSIM) are used to evaluate the quality of the interpolated seismic data; the calculation method of the signal-to-noise ratio and the structural similarity is: SNR (dB) = 10 log y 2 (y - G(x)) 2 where y and G(x) represent the ideal and recovered seismic data, respectively; μ y and μ G(x) represent the mean of the data image pixels of y and G(x), respectively; σ y and σ G(x) represent the variance of the data image pixels of y and G(x), respectively; σ yG(x) represents the covariance of the data image pixels of y and G(x); and c1 and c2 represent stability constants for the calculation.

5. A device for adaptive intelligent repair of missing regions of multi-scale seismic data in wavelet domain, characterized in that, the wavelet domain multi-scale seismic data missing area adaptive intelligent repair device executes the wavelet domain multi-scale seismic data missing area adaptive intelligent repair method, and the steps comprise: acquiring seismic data to be repaired; inputting the seismic data into a pre-trained first generation network to obtain repaired seismic data; the first generation network comprises an encoder, a residual module and a decoder: wherein a wavelet convolution module is arranged in the encoder and the decoder; wherein the training method of the first generation network is: obtaining training data, wherein the training data comprises missing seismic data and corresponding complete seismic data; inputting the missing seismic data in the training data into a generator to generate first generation data; calculating the loss of a discriminator using the first generation data and the complete seismic data and updating the parameters of the discriminator; updating the parameters of the generator through the discriminator until the model reaches a preset performance index; The loss of the generator is calculated by the discriminator and the parameters of the generator are updated until the model reaches a preset performance index; The trained generator is used as a first generation network; The first generation network specifically includes four down-sampling modules of an encoder and four up-sampling modules of a decoder, uses a wavelet convolution layer, specifically adopts a db2 wavelet base, and sets the decomposition layer number to 2; The four down-sampling modules of the encoder are implemented in the following manner: at each layer of down-sampling, 2-level decomposition is performed by using the db2 wavelet base to obtain low-frequency and high-frequency feature maps, and a 1024-channel feature map is obtained by layer-by-layer calculation; the residual module uses a 3x3 convolution; the four up-sampling modules of the decoder are implemented in the following manner: at each layer of up-sampling, up-sampling is performed by using the db2 wavelet base, and the data image is restored to the same spatial resolution as the original image by layer-by-layer calculation; The loss function in the training mode of the first generation network includes: l G = l rec + αl per + βl adv + λl tv wherein, l G is the loss function of the generator, and a, b, l are weights corresponding to each loss term, used to balance the influence of the loss on training; l rec represents the reconstruction loss calculated using the mean square error (MSE), used to measure the pixel difference between the generated data and the target data, and the calculation is the MSE of the generated data and its corresponding complete data d R , denoted as: W and H represent the input missing seismic data d M of dimension t, t represents the scale of resolution improvement from d M to d R ; is the corresponding complete data; perceptual loss l per is a loss function that measures the difference in perceptual quality between the generated data and the target data, represented as: where φ i，j represents an operation of extracting a feature map of the jth convolution before the ith max-pooling layer, W i,j and H i,j represent dimensions of respective feature maps within the VGG network; x, y represent counting values; l adv is a reconstruction loss;l tv is a total variation loss.

6. An electronic device, comprising: including: a processor adapted to implement one or more instructions; and A computer storage medium stores one or more instructions, and the one or more instructions are adapted to be loaded and executed by the processor to implement the wavelet domain multi-scale seismic data missing region adaptive intelligent repair method according to any one of claims 1-4.

7. A computer storage medium, characterized in that The computer storage medium stores one or more instructions, and the one or more instructions are adapted to be loaded and executed by the processor to implement the wavelet domain multi-scale seismic data missing region adaptive intelligent repair method according to any one of claims 1-4.

Images