Conditional generative adversarial network-based missing seismic data reconstruction method

Abstract

The invention relates to the technical field of seismic data reconstruction, in particular to a missing seismic data reconstruction method based on a conditional generative adversarial network, which comprises the following steps: S1, obtaining corresponding missing seismic data from complete seismic data through a data missing method; s2, establishing a conditional generative adversarial network model, forming a pair of conditions by the missing seismic data and the complete seismic data, and inputting the conditions to the conditional generative adversarial network model for training and verification; and S3, inputting missing seismic data to be reconstructed into the trained conditional generative adversarial network model for data reconstruction. The method has higher reconstruction precision and has a certain denoising effect, so that the reconstruction result has good effectiveness.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data reconstruction, and more specifically, to a missing seismic data reconstruction method based on a conditional generative adversarial network. Background Art

[0002] Seismic exploration is a geophysical exploration method that measures the response of the earth to artificially excited seismic waves through detectors, obtains underground structure information, and then analyzes the underground medium layer to determine the location of the oil and gas circle. It is an important means of oil exploration. Due to the limitations of environmental and economic conditions in actual exploration, the collected seismic data will have problems such as missing seismic traces or insufficient spatial sampling. For example, when encountering complex geological conditions such as mountains and rivers that make it impossible to place detectors, or when the detectors fail, the collected data will be missing. The lack of data will have a serious impact on the subsequent processing links. For example, insufficient sampling rate will lead to the generation of false frequencies, which will eventually cause low imaging resolution, reduce the accuracy of seismic data processing, and people cannot make a correct interpretation of the underground structure. In order to be able to reconstruct the missing data, although there is a seismic data reconstruction method in the prior art, its reconstruction result is not ideal, and the effectiveness of the reconstruction result is not high enough. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art in that the seismic data reconstruction results are not ideal, and to provide a missing seismic data reconstruction method based on a conditional generative adversarial network, which can have higher reconstruction accuracy and a certain denoising effect, so that the reconstruction results have good effectiveness.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A missing seismic data reconstruction method based on a conditional generative adversarial network is provided, comprising the following steps:

[0006] S1. Obtain the corresponding missing seismic data from the complete seismic data through the data missing method;

[0007] S2. Establish a conditional generative adversarial network model, and then input a pair of conditions consisting of the missing earthquake data and the complete earthquake data into the conditional generative adversarial network model for training and verification;

[0008] S3. Input the missing seismic data to be reconstructed into the trained conditional generative adversarial network model for data reconstruction.

[0009] The present invention is a method for reconstructing missing seismic data based on a conditional generative adversarial network. A conditional mechanism is added to the conditional generative adversarial network model. The missing seismic data and the complete seismic data are combined into a pair of conditions and input into the conditional generative adversarial network model for training. The present invention can use the conditional mechanism to control the output of the conditional generative adversarial network model, and can have higher reconstruction accuracy and a certain denoising effect, which is beneficial to removing spatial aliasing.

[0010] Furthermore, in step S2, the conditional generative adversarial network model includes a generator and a discriminator. The generator introduces a HED module based on the U-Net network structure. The HED module adds a side output layer after the last convolution layer in each upsampling stage of the U-Net network structure, concatenates the side output layer of each level in the channel dimension, and then uses 1×1 convolution to fuse the level outputs.

[0011] Furthermore, the discriminator is a pyramid structured fully convolutional network, and the number of network input channels of the discriminator is 2; the number of network input channels of the generator is 1.

[0012] Furthermore, in step S2, the training of the conditional generative adversarial network model is optimized using the Adam optimizer, and the cross entropy loss function is selected for evaluation, wherein the loss function L is:

[0013] L=L adv +λL rec ;

[0014] Where, L adv represents the adversarial loss, L rec represents the reconstruction loss, λ represents the reconstruction loss L rec The weight of .

[0015] Furthermore, against the loss L adv for:

[0016] L adv =log(1-D(G(x));

[0017] Wherein, D(G(x)) represents the result of inputting the sample G(x) generated by the generator into the discriminator.

[0018] Furthermore, the adversarial loss L adv The optimization of is jointly controlled by the generator and the discriminator.

[0019] Further, in step S1, the missing seismic data is input as a training set, and the complete seismic data is used as a label of the training set;

[0020] In step S2, the reconstruction loss Lrec The ideal relationship between the gradient, label, and generator output is:

[0021]

[0022] Wherein, y represents the label, represents the output of the generator, and ε represents the reconstruction loss L rec gradient.

[0023] Furthermore, the reconstruction loss L rec The gradient of is solved by the indefinite integral to obtain the reconstruction loss L of the sample G(x) at a certain point r for:

[0024]

[0025] The reconstruction loss function of the overall data is:

[0026]

[0027] Wherein, y represents the label, represents the output of the generator, N w 、N h 、N d Respectively represent the size of the data in the width, height, and depth directions, y i,j,k represents the value of point (i,j,k) in the label, Represents the value of the point (i, j, k) in the output of the generator.

[0028] Furthermore, the reconstruction loss L rec Controlled solely by the discriminator.

[0029] Furthermore, in step S1, the corresponding missing seismic data are obtained by any one of the data missing methods including random missing, continuous missing, and a combination of random missing and continuous missing.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. A conditional mechanism is added to the conditional generative adversarial network model, and a pair of conditions consisting of missing seismic data and complete seismic data is input into the conditional generative adversarial network model for training, so that the present invention can use the conditional mechanism to control the output of the conditional generative adversarial network model, which can have higher reconstruction accuracy and a certain denoising effect, and is conducive to removing spatial aliasing;

[0032] 2. The missing earthquake data and the complete earthquake data are combined into a pair of conditions and input into the conditional generative adversarial network model for training, so that the generator and the discriminator are constrained by the conditions at the same time during the training process, and can learn the target features more easily;

[0033] 3. The selected cross entropy loss function has a smooth gradient, which can avoid the huge gradient fluctuations caused by the L1 norm, and can avoid the gradient reversal caused by the L2 norm in some cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart of a missing seismic data reconstruction method based on a conditional generative adversarial network according to the present invention;

[0035] Figure 2 A structural diagram of a generator for a conditional generative adversarial network model of the present invention;

[0036] Figure 3 A structural diagram of a discriminator for a conditionally generated adversarial network model of the present invention;

[0037] Figure 4 The overall structure diagram of the conditional generation adversarial network model of the present invention;

[0038] Figure 5 is a loss function curve diagram of the present invention;

[0039] Figure 6 The present invention provides 70% random missing pre-stack raw seismic data, missing seismic data, generated seismic data, and residual graphs;

[0040] Figure 7 This is a single-channel fitting comparison diagram of 70% random missing pre-stack seismic data of the present invention;

[0041] Figure 8 The present invention includes 75 consecutive missing pre-stack original seismic data, missing seismic data, generated seismic data, and residual graphs;

[0042] Fig. 9 This is a single-channel fitting comparison diagram of 75 consecutive missing pre-stack seismic data of the present invention;

[0043] Fig.10 The present invention provides 70% random missing post-stack raw seismic data, missing seismic data, generated seismic data, and residual graphs;

[0044] Fig.11 This is a single-channel fitting comparison diagram of 70% random missing post-stack seismic data of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further described below in conjunction with specific implementation methods. In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0046] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or apparatuses.

[0048] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0049] Embodiment 1

[0050] like Figure 1 The first embodiment of a missing seismic data reconstruction method based on a conditional generative adversarial network of the present invention is shown, comprising the following steps:

[0051] S1. Obtain the corresponding missing seismic data from the complete seismic data through the data missing method;

[0052] S2. Establish a conditional generative adversarial network model, and then input a pair of conditions consisting of missing earthquake data and complete earthquake data into the conditional generative adversarial network model for training and verification;

[0053] S3. Input the missing seismic data to be reconstructed into the trained conditional generative adversarial network model for data reconstruction.

[0054] The present invention is a method for reconstructing missing seismic data based on a conditional generative adversarial network. A conditional mechanism is added to the conditional generative adversarial network model. The missing seismic data and the complete seismic data are combined into a pair of conditions and input into the conditional generative adversarial network model for training. The present invention can use the conditional mechanism to control the output of the conditional generative adversarial network model, and can have higher reconstruction accuracy and a certain denoising effect, which is beneficial to removing spatial aliasing.

[0055] Embodiment 2

[0056] This embodiment is a second embodiment of a method for reconstructing missing seismic data based on a conditional generative adversarial network. This embodiment is similar to the first embodiment, except that in step S1, the missing seismic data is used as a training set input, and the complete seismic data is used as a label of the training set.

[0057] In step S1, the corresponding missing seismic data are obtained by any one of the data missing methods including random missing, continuous missing, and a combination of random missing and continuous missing.

[0058] Embodiment 3

[0059] This embodiment is the third embodiment of a missing seismic data reconstruction method based on a conditional generative adversarial network. This embodiment is similar to the first or second embodiment, except that in step S2, the conditional generative adversarial network model includes a generator and a discriminator. The generator introduces a HED module based on the U-Net network structure. The HED module adds a side output layer after the last convolution layer of each upsampling stage of the U-Net network structure, splices the side output layer of each level in the channel dimension, and then uses 1×1 convolution to fuse the level output, and replaces the original output of the U-Net network structure with the fused output. The HED module can learn images at multiple scales and directly operate on the entire image, so it can effectively perceive multi-scale features and capture the intrinsic features of the input data with different step sizes at different stages.

[0060] like Figures 2 to 4As shown in the figure, the number of network input channels of the generator is 1. The generator receives the complete seismic data as a condition while receiving the missing seismic data, and then generates a sample G(x) related to the condition; the discriminator receives the sample G(x), the complete seismic data, and the missing seismic data. The discriminator judges the complete seismic data while judging the sample G(x). The discriminator will use the missing seismic data as a condition during the learning process, so that the discriminator can judge whether the sample G(x) and the complete seismic data are the original data corresponding to the missing seismic data, thereby improving the classification accuracy of the discriminator. The complete seismic data is introduced into the conditional generative adversarial network model as a condition, so that the generator and the discriminator are constrained by the condition at the same time during the training process. First, through two convolution operations, the data size remains unchanged and the number of channels becomes 32; then four pooling-convolution-convolution operations are performed, each time the input size is reduced by half and the number of channels is doubled. After four times, the number of channels is 512; then four upsampling-linking-convolution-convolution operations are performed, each time the input size is doubled and the number of channels is halved. In this embodiment, the discriminator is a pyramid structure full convolution network. The number of network input channels of the discriminator is 2, and after four convolutions, the number of channels is 512; the last convolution outputs the discrimination result.

[0061] Specifically, in step S2, the training of the conditional generative adversarial network model is optimized using the Adam optimizer, and the cross entropy loss function is selected for evaluation.

[0062] Among them, the loss function L is:

[0063] L=L adv +λL rec ;

[0064] Where, L adv represents the adversarial loss, L rec represents the reconstruction loss; λ represents the reconstruction loss L rec The weight of , weight λ, is used to adjust it to the same order of magnitude as the adversarial loss.

[0065] Among them, the adversarial loss L adv for:

[0066] L adv =log(1-D(G(x));

[0067] In the formula, D(G(x)) represents the result of the sample G(x) generated by the generator and input into the discriminator.

[0068] Specifically, in step S2, regarding the reconstruction loss L rec The ideal relationship between the gradient, label, and generator output is:

[0069]

[0070] In the formula, y represents the label, that is, the complete earthquake data; represents the output of the generator, i.e., predicted earthquake data; ε represents the reconstruction loss L rec gradient.

[0071] Then, the reconstruction loss L rec The gradient of is solved by the indefinite integral to obtain the reconstruction loss L of the sample G(x) at a certain point r for:

[0072]

[0073] The reconstruction loss function of the overall data is:

[0074]

[0075] In the formula, y represents the label, that is, the complete earthquake data; represents the output of the generator, i.e., predicted earthquake data; N w 、N h 、N d Respectively represent the size of the data in the width, height, and depth directions; y i,j,k represents the value of point (i,j,k) in the complete seismic data, Indicates the value of the predicted seismic data point (i, j, k). In this embodiment, the loss function L has a smooth gradient, which can avoid the huge gradient oscillation caused by the L1 norm and the gradient reversal caused by the L2 norm in some cases.

[0076] During training, the adversarial loss L adv The optimization is controlled by the generator and the discriminator. The purpose of the generator is to make the adversarial loss L adv Minimum, and the purpose of the discriminator is to make the adversarial loss L adv Maximum, the generator and the discriminator compete with each other to make the adversarial loss L adv Convergence is reached. The reconstruction loss L rec Controlled by the discriminator alone, when the reconstruction loss L rec When the minimum value is reached, it means that the discriminant ability of the discriminator has reached the maximum. During the training process, the Adam optimizer is used for optimization, the learning rate is selected as 0.0005, the training batch size is 128, and after 25 training iterations, the loss function reaches the extreme value convergence, as shown in Figure 5 shown.

[0077] like Figure 6 and Figure 7 The reconstruction results for 70% randomly missing prestack seismic data are as follows: Figure 8 and Fig. 9The reconstruction results of 75 consecutive missing pre-stack seismic data are shown in Figure 2. Fig.10 and Fig.11 This is the reconstruction result of 70% randomly missing post-stack seismic data, which can verify the effectiveness of the missing seismic data reconstruction method based on conditional generative adversarial network of the present invention.

[0078] The present invention first uses complete seismic data as the label of the training set, and uses missing seismic data as the input of the training set, and then establishes a conditional generation adversarial network model. The conditional generation adversarial network model has a conditional mechanism, and can use the conditional control network's feature learning and output. The conditional generation adversarial network uses the missing seismic data as a condition and complete seismic data to form a pair of inputs to the discriminator for discrimination, which can allow the discriminator to discriminate the true or false of the reconstructed seismic data in a targeted manner, improve the accuracy of discrimination, and also help to improve the generation ability of the generator; after the training of the conditional generation adversarial network model is completed, the trained conditional generation adversarial network model is used to reconstruct the missing seismic data, and the reconstruction result of the missing seismic data can be obtained. It plays a very important role in a series of seismic data processing processes such as improving resolution, offset, suppressing multiple waves, etc., and verifies the effectiveness of the present invention through 70% random missing prestack, 75 continuous missing prestack, and 70% random missing poststack.

[0079] In the specific contents of the above-mentioned specific implementation methods, the various technical features can be combined in any non-contradictory manner. In order to make the description concise, not all possible combinations of the above-mentioned technical features are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A missing seismic data reconstruction method based on conditional generative adversarial network, characterized in that: The steps include: S1. Obtain the corresponding missing seismic data from the complete seismic data through the data missing method; S2. Establish a conditional generative adversarial network model, and then input a pair of conditions consisting of the missing earthquake data and the complete earthquake data into the conditional generative adversarial network model for training and verification; S3. Input the missing seismic data to be reconstructed into the trained conditional generative adversarial network model for data reconstruction.

2. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 1 is characterized in that: In step S2, the conditional generative adversarial network model includes a generator and a discriminator. The generator introduces a HED module based on the U-Net network structure. The HED module adds a side output layer after the last convolution layer in each upsampling stage of the U-Net network structure, concatenates the side output layer of each level in the channel dimension, and then uses 1×1 convolution to fuse the level outputs.

3. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 2 is characterized in that: The discriminator is a pyramid structured fully convolutional network, and the number of network input channels of the discriminator is 2; the number of network input channels of the generator is 1.

4. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 1 is characterized in that: In step S2, the training of the conditional generative adversarial network model is optimized using the Adam optimizer, and the cross entropy loss function is selected for evaluation, where the loss function L is: L=L adv +λL rec ; Where, L adv represents the adversarial loss, L rec represents the reconstruction loss, λ represents the reconstruction loss L rec The weight of .

5. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 4 is characterized in that: Adversarial loss L adv for: L adv =log(1-D(G(x)); Wherein, D(G(x)) represents the result of inputting the sample G(x) generated by the generator into the discriminator.

6. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 5, characterized in that: The adversarial loss L adv The optimization of is jointly controlled by the generator and the discriminator.

7. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 5, characterized in that: In step S1, the missing seismic data is input as a training set, and the complete seismic data is used as a label of the training set; In step S2, the reconstruction loss L rec The ideal relationship between the gradient, label, and generator output is: Wherein, y represents the label, represents the output of the generator, and ε represents the reconstruction loss L rec gradient.

8. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 7, characterized in that: The reconstruction loss L rec The gradient of is solved by the indefinite integral to obtain the reconstruction loss L of the sample G(x) at a certain point r for: The reconstruction loss function of the overall data is: Wherein, y represents the label, represents the output of the generator, N w 、N h 、N d Respectively represent the size of the data in the width, height, and depth directions, y i,j,k represents the value of point (i,j,k) in the label, Represents the value of the point (i, j, k) in the output of the generator.

9. The missing seismic data reconstruction method based on conditional generative adversarial network according to claim 8, characterized in that: The reconstruction loss L rec Controlled solely by the discriminator.

10. The missing seismic data reconstruction method based on conditional generative adversarial network according to any one of claims 1 to 9, characterized in that: In step S1, the corresponding missing seismic data are obtained by any one of the data missing methods including random missing, continuous missing, and a combination of random missing and continuous missing.

Images