Elastic parameter high frequency component recovery method based on conditional generative adversarial network

Abstract

The application discloses an elastic parameter high-frequency component recovery method based on a conditional generative adversarial network, and comprises the following steps: acquiring large-scale elastic parameter information of a research area; inputting the large-scale elastic parameter information of the research area into a pre-constructed conditional generative adversarial network; the conditional generative adversarial network comprises a discriminator D and a generator G, and the generator G outputs corresponding elastic parameter information containing high-frequency components. After the conditional generative adversarial network is successfully trained, high-precision elastic parameter information containing large-scale elastic parameter information and small-scale elastic parameter information conforming to the geological law of the research area can be output by the generator only by providing the large-scale elastic parameter information of the research area, so that the recovery of the high-precision elastic parameter is realized. Higher-precision seismic data is provided, and seismic inversion is better carried out.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological exploration and artificial intelligence, and particularly relates to an elastic parameter high-frequency component recovery method based on a conditional generative adversarial network. BACKGROUND

[0002] The generative adversarial network (GAN) is a machine learning neural network architecture proposed by Ian Goodfellow of the University of Montreal in 2014, and is one of the most promising methods for machine learning on complex distribution in recent years. The generative adversarial network architecture includes a generation module and a discrimination module, and a good output is generated through mutual game learning of a generative model and a discriminative model.

[0003] The conditional generative adversarial network (Conditional GAN) adds a type label to the input of the generator and the discriminator on the basis of the original. The quality of the data generated by the conditional GAN is generally higher than that of the data generated by the same GAN without using any label information. By specifying the label information, the conditional generative adversarial network can generate data of a specified type.

[0004] Random modeling generally applies small-scale random perturbations on the basis of deterministic modeling, and has higher resolution than deterministic modeling. If the random perturbations satisfy certain prior conditions, the small-scale random perturbations can reflect certain geological distribution rules, and have certain reference value in guiding engineering practice and resource development of the distribution of underground geological bodies.

[0005] Elastic parameter high-frequency component modeling is an elastic parameter modeling technology with high resolution, and is a key research direction in the field of elastic modeling. Application of the modeling method can realize recovery of the high-frequency component of the elastic parameter, and has high application prospect and research value. SUMMARY

[0006] The purpose of the present application is to realize recovery of the high-frequency component of the elastic parameter by training the conditional generative adversarial network (Conditional GAN) to output high-precision elastic parameter information containing large-scale elastic parameter information and small-scale elastic parameter information conforming to the geological rules of the study area only by providing large-scale elastic parameter information of the study area. Meanwhile, the conditional generative adversarial network obtained by training has strong generalization performance, and can be applied to the entire study area for recovery of the high-frequency component of the elastic parameter only by providing local prior data of the study area as training data.

[0007] The application discloses an elastic parameter high-frequency component recovery method based on a conditional generative adversarial network, comprising the following steps:

[0008] Large-scale elastic parameter information of a research area is acquired, and the large-scale elastic parameter information of the research area is input into a pre-constructed conditional generative adversarial network, wherein the conditional generative adversarial network comprises a discriminator D and a generator G, and the generator G outputs corresponding elastic parameter information containing high-frequency components.

[0009] Further, the construction process of the conditional generative adversarial network comprises the following steps:

[0010] A local random elastic parameter model is established according to detailed geological data of a local research area, and elastic parameter data are selected from the local random elastic parameter model as training data; and large-scale elastic parameter information is extracted from the training data as a label to mark the training data, and the training data marked by the large-scale elastic parameter information are regarded as real data (A;m).

[0011] Further, the construction process of the conditional generative adversarial network further comprises the following steps:

[0012] Random large-scale elastic parameter information labels are used for marking to obtain random seeds (b;n).

[0013] Further, the training process of the conditional generative adversarial network comprises the following steps:

[0014] 1) The real data (A;m) are input into the discriminator D, and the discriminator D is trained so that the output of the discriminator D is 1;

[0015] 2) The random seeds (b;n) are input into the generator G, the generator G outputs generated data B marked by the same random large-scale elastic parameter information n, the generated data (B;n) are input into the discriminator D, and the discriminator D is trained so that the output of the discriminator D is 0;

[0016] 3) The random seeds (b;n) marked by the random large-scale elastic parameter information n are input into the generator G again, and the generator G is trained so that the generated data (B;n) generated by the generator G make the output of the discriminator D be 1 when the generated data (B;n) are input into the discriminator D;

[0017] 4) The processes 1) to 3) are repeated, and the generative adversarial network is continuously trained for multiple rounds.

[0018] The training is ended when the loss values of the generator G and the discriminator D are relatively stable.

[0019] Further, the construction process of the conditional generative adversarial network comprises the following steps: constructing an elastic parameter high-frequency component model.

[0020] Furthermore, the construction process of the high-frequency component model of the elastic parameters includes: inputting the large-scale elastic parameter information at different locations in the study area into the trained generator, outputting the high-frequency component information of the elastic parameters at different locations in the study area, and synthesizing the high-frequency component information of the elastic parameters at different locations in the study area to obtain the high-frequency component model of the elastic parameters.

[0021] The beneficial effects of this invention are as follows:

[0022] Once successfully trained, the conditional generative adversarial network (GAN) only needs to provide large-scale elastic parameter information for the study area. The generator can then output high-precision elastic parameter information containing both large-scale and small-scale elastic parameter information consistent with the geological characteristics of the study area, thus achieving high-precision elastic parameter recovery. This provides more accurate seismic data and facilitates better seismic inversion. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the process of the present invention;

[0024] Figure 2 This is a schematic diagram of the training process of the conditional generative adversarial network of the present invention.

[0025] Figure 3 This is a schematic diagram of the training data and its large- and small-scale elastic parameters of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0027] like Figure 1 As shown, the method for recovering high-frequency components of elastic parameters based on conditional generative adversarial networks of the present invention includes: inputting large-scale elastic parameter information of the study area into a pre-constructed conditional generative adversarial network, wherein the conditional generative adversarial network includes a discriminator D and a generator G, the generator G outputs corresponding elastic parameter information containing high-frequency components, and the discriminator D outputs elastic parameter information containing high-frequency components after discrimination.

[0028] The construction process of conditional generative adversarial networks includes the following parts:

[0029] (1) Establish a high-frequency component model of local elastic parameters as the basis for prior data:

[0030] According to the local detailed geological data of the research area, a local elastic parameter high-frequency component model is established. Due to the need of method research, it is assumed that the small-scale random disturbance existing in the random elastic parameter model established by the fast Fourier transform moving average method (FFT-MA) used in the patent specification is consistent with the geological distribution law of the research area, so that a local elastic parameter high-frequency component model is established. In actual application, a local elastic parameter high-frequency component model containing reliable large and small scale elastic parameter information is needed as a prerequisite for the method. The local elastic parameter high-frequency component model established here will serve as the prior data basis for the subsequent technical solutions.

[0031] The fast Fourier transform moving average method (FFT-MA) is a frequency domain implementation of the moving average model (MA) and can realize random elastic parameter modeling. The autocorrelation function can represent the small-scale non-uniformity of the earth medium, and some commonly used analytical expressions are Gaussian type, exponential type, etc. Taking the two-dimensional case as an example, the random medium model can be represented by two autocorrelation lengths (a, b), autocorrelation angle (θ), and roughness factor r, and its general expression is:

[0032]

[0033] Where Δz, Δx are the spatial offsets in the horizontal and vertical directions, r=0 represents the Gaussian type, r=1 represents the exponential type, and the model will become rougher as r changes from 0 to 1. By setting different autocorrelation coefficients and model parameters, non-uniform geological random modeling can be realized, and thus a local elastic parameter high-frequency component model is established as a prior data basis.

[0034] (2) Selecting training data, extracting labels and marking data from the local elastic parameter high-frequency component model:

[0035] According to the local random elastic parameter model established by the FFT-MA method, elastic parameter data can be selected as training data. Usually, elastic parameters A(x, y) in the horizontal direction along the depth direction are selected as training data. The large-scale elastic parameter information of the training data is extracted as a label to mark the training data. The training data marked by the large-scale elastic parameter information will be used as real data to participate in the training of the conditional generative adversarial network.

[0036] The training of the generative adversarial network also needs a random seed marked randomly. A random large-scale geological information label is used to mark the random seed, and the random seed marked randomly is input into the generator of the conditional generative adversarial network. The generator will output generated data with the same random large-scale geological information label as the random seed.

[0037] The conditional generative adversarial network can be trained by using real data marked by large-scale elastic parameter information and random seeds marked by random large-scale elastic parameter information.

[0038] (3) Training the conditional generative adversarial network:

[0039] The conditional generative adversarial network is composed of a generator G and a discriminator D, and the training process of the conditional generative adversarial network is actually the training process of the generator G and the discriminator D. Compared with other neural network architectures, the generator G and the discriminator D in the training process are mutually influenced, and the performance of the two neural networks is improved through mutual confrontation and game.

[0040] As shown in Figure 2 , the training process of the conditional generative adversarial network is as follows:

[0041] 1) input the real data (A; m) marked by the large-scale elastic parameter information m into the discriminator D, and train the discriminator D so that its output is 1 (TRUE);

[0042] 2) input the random seed (b; n) marked by the random large-scale elastic parameter information n into the generator G, and the generator G will output a generated data B marked by the same random large-scale elastic parameter information n. Input the generated data (B; n) into the discriminator D, and train the discriminator D so that its output is 0 (FALSE);

[0043] 3) input the random seed (b; n) marked by the random large-scale elastic parameter information n into the generator G again, and train the generator G so that the generated data (B; n) generated by the generator G makes the output of the discriminator D be 1 when input into the discriminator D;

[0044] 4) repeat processes (1) to (3) to continuously train the generative adversarial network for multiple rounds;

[0045] 5) observe the loss value change curve of the generator G and the discriminator D until the loss values of the two are relatively stable, and end the training.

[0046] (4) Elastic parameter high-frequency component recovery

[0047] By inputting the large-scale elastic parameter information of different positions in the study area into the generator G, high-precision elastic parameter information containing large-scale elastic parameter information and small-scale elastic parameter information (high-frequency component) conforming to the geological rules of the study area can be obtained. By synthesizing the high-precision elastic parameter information of different positions in the modeling area, elastic parameter high-frequency component modeling can be realized, thereby realizing elastic parameter high-frequency component recovery.

[0048] Meanwhile, the method can be used to restore the high-frequency component of the elastic parameter, and the area is not limited to the modeling area of the prior high-frequency component model of the elastic parameter, and the method can be used to restore the high-frequency component of the elastic parameter in the whole study area, and the modeling effect depends on the complexity of the geological conditions of the study area, the training effect of the conditional generative adversarial network, and the like.

[0049] The process of applying the method to restore the high-frequency component of the elastic parameter is as follows: inputting large-scale elastic parameter information of the study area into the trained generator, outputting elastic parameter high-frequency component information, and synthesizing elastic parameter high-frequency component information at different positions to obtain an elastic parameter high-frequency component model, thereby realizing the restoration of the high-frequency component of the elastic parameter.

[0050] The above only describes the preferred embodiments of the present application, and it should be noted that those skilled in the art can make several improvements and modifications without departing from the technical principles of the present application, and these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1.A method for recovering high-frequency components of elastic parameters based on a conditional generative adversarial network, characterized in that, Comprise: Expressed by two autocorrelation lengths (a, b), autocorrelation angle (θ), and roughness factor r: Where △z, △x are the spatial offsets in horizontal and vertical directions, r=0 represents Gaussian type, r=1 represents exponential type, and the model will become rougher as r changes from 0 to 1; non-uniform geological random modeling is realized by setting different autocorrelation coefficients and model parameters, thereby establishing a local elastic parameter high-frequency component model as a prior data basis; Obtaining large-scale elastic parameter information of the research area, inputting the large-scale elastic parameter information of the research area into a conditional generative adversarial network constructed in advance, the conditional generative adversarial network comprising a discriminator D and a generator G, and the generator G outputting corresponding elastic parameter information containing high-frequency components The construction process of the conditional generative adversarial network comprises: According to the local detailed geological data of the research area, a local random elastic parameter model is established, and elastic parameter data is selected from the local random elastic parameter model as training data, and large-scale elastic parameter information is extracted from the training data as a label to mark the training data, and the training data marked by the large-scale elastic parameter information is taken as real data (A; m); The construction process of the conditional generative adversarial network further comprises: Using random large-scale elastic parameter information labels to mark, obtaining random seeds (b; n); The training process of the conditional generative adversarial network comprises: 1) inputting the real data (A; m) into the discriminator D, and training the discriminator D so that its output is 1; 2) inputting the random seed (b; n) into the generator G, the generator G will output a generated data B marked by the same random large-scale elastic parameter information n, inputting the generated data (B; n) into the discriminator D, and training the discriminator D so that its output is 0; 3) inputting the random seed (b; n) marked by the random large-scale elastic parameter information n into the generator G again, and training the generator G so that the generated data (B; n) generated by the generator G makes the output of the discriminator D as 1 when input into the discriminator D; 4) repeating processes 1) to 3), and continuously training the generative adversarial network for multiple rounds; When the loss values of the generator G and the discriminator D are relatively stable, the training is ended; The construction process of the conditional generative adversarial network comprises: constructing an elastic parameter high-frequency component model; The construction process of the elastic parameter high-frequency component model comprises: inputting large-scale elastic parameter information of different positions in the research area into the trained generator respectively, outputting elastic parameter high-frequency component information of different positions in the research area, and synthesizing the elastic parameter high-frequency component information of different positions in the research area to obtain an elastic parameter high-frequency component model.

Images