Crack network generation method

Abstract

The invention discloses a fracture network generation method, and relates to the technical field of underground reservoir development. The method comprises the steps of determining center point position information, length information and angle information of cracks in an underground reservoir of a sample area by adopting a scale value point process method based on a statistical distribution rule of the cracks in the underground reservoir, and generating candidate crack network image samples according to the center point position information, the length information and the angle information. Adding known cracks determined based on prior information into the candidate crack network image sample, and removing cracks which are not communicated with other cracks to obtain a first crack network image sample; and training a generative adversarial network model according to the Gaussian noise sample of the target dimension and the first crack network image sample until a training stop condition is satisfied, and obtaining a crack network generation model for generating a crack network image according to the Gaussian noise of the target dimension. According to the embodiment of the invention, the authenticity of the generated crack network image can be improved.

Description

Technical Field

[0001] This application belongs to the technical field of underground reservoir development, and particularly relates to a method for generating a fracture network. Background Art

[0002] With the proposal of the dual-carbon strategy, the research and development of underground reservoirs have become an important topic in the current fields of geology and geophysics. The discrete fracture network (DFN) widely existing in underground reservoir rocks is the main seepage channel of fluids. Therefore, the characterization of the fracture network is crucial for exploring issues such as energy extraction efficiency and the safety of carbon dioxide sequestration. Since the fracture network in the underground reservoir cannot be directly observed, it can only be indirectly inversely characterized through geological and geophysical data. Due to the complexity of the fracture network, the existing traditional parameterization methods will result in a huge number of parameters, making this inverse characterization very difficult. Therefore, it is important to develop an effective low-dimensional parameterization method for the fracture network. One of the ideas for developing such a method is to generate fractures from a low-dimensional parameter space based on a deep learning model.

[0003] Existing methods use low-dimensional parameters to generate fracture network images, and it is difficult to obtain high-quality and high-fidelity fracture network images. In addition, the generated fracture network cannot reflect the statistical distribution characteristics of the naturally existing fracture network, and the authenticity of the generated fracture network image is not high. Summary of the Invention

[0004] The embodiments of this application provide a method for generating a fracture network, which can generate a fracture network based on low-dimensional parameter data through a trained fracture network generation model, and improve the authenticity of the fracture network generation.

[0005] In a first aspect, the embodiments of this application provide a method for generating a fracture network, and the method includes:

[0006] Based on the statistical distribution law of fractures in the underground reservoir, using the marked point process method, determine the position information, length information, and angle information of the center points of fractures in the underground reservoir of the sample area;

[0007] Generate a candidate fracture network image sample according to the position information, length information, and angle information of the center points;

[0008] Add a first fracture and remove a second fracture in the candidate fracture network image sample to obtain a first fracture network image sample; the first fracture represents a known fracture determined based on prior information, and the second fracture represents a fracture that is not connected to other fractures;

[0009] Train a preset generative adversarial network model based on Gaussian noise samples of the target dimension and the first fracture network image samples until the training stop condition is met, obtaining a fracture network generation model for generating fracture network images according to Gaussian noise of the target dimension; wherein, the target dimension is lower than a preset dimension threshold.

[0010] In some embodiments, the statistical distribution laws include fractal distribution, power-law distribution, and von Mises distribution;

[0011] Based on the statistical distribution law of fractures in the underground reservoir, use the marked point process method to determine the central point position information, length information, and angle information of fractures in the underground reservoir of the sample area, including:

[0012] Mark the positions of fractures based on fractal distribution to generate a two-dimensional point set that satisfies fractal distribution;

[0013] Randomly sample two-dimensional points that meet a preset quantity threshold in the two-dimensional point set;

[0014] Determine the central point position information of each fracture in the underground reservoir of the sample area according to the two-dimensional points;

[0015] Determine the length information of the fractures based on the power-law distribution according to the central point positions;

[0016] Determine the angle information of the fractures based on the von Mises distribution according to the central point positions.

[0017] In some embodiments, the generative adversarial network model includes a generator network; training a preset generative adversarial network model based on Gaussian noise samples of the target dimension and the first fracture network image samples until the training stop condition is met, obtaining a fracture network generation model, including:

[0018] Obtain a training sample set, where the training sample set includes multiple training samples, and each training sample includes a Gaussian noise sample of the target dimension and its corresponding first fracture network image sample;

[0019] Perform the following steps for each training sample respectively:

[0020] Input the Gaussian noise sample of the target dimension into the generative adversarial network model to obtain a second fracture network image corresponding to the Gaussian noise sample of the target dimension;

[0021] Determine the training loss of the generative adversarial network model according to the first fracture network image sample and the second fracture network image;

[0022] When the training loss does not meet the training stop condition, adjust the parameters of the generator network to obtain an updated generative adversarial network model, and return the second fracture network image corresponding to the Gaussian noise sample of the target dimension by inputting the Gaussian noise sample of the target dimension into the generative adversarial network model until the training stop condition is met, and obtain a fracture network generation model.

[0023] In some embodiments, the generative adversarial network model includes a discriminator network; determining the training loss of the generative adversarial network model according to the first fracture network image sample and the second fracture network image includes:

[0024] Input the first fracture network image sample and the second fracture network image into the discriminator network to obtain a discrimination result;

[0025] Calculate the Wasserstein distance and gradient penalty between the first fracture network image sample and the second fracture network image based on the discrimination result;

[0026] Obtain the training loss of the generative adversarial network model according to the sum of the Wasserstein distance and the gradient penalty.

[0027] In some embodiments, the discriminator network includes a convolutional layer, a downsampling residual network module, and a fully connected layer;

[0028] Among them, the downsampling residual network module is composed of a normalization layer, a ReLU activation function layer, a convolutional layer, and an average pooling layer.

[0029] In some embodiments, after training a preset generative adversarial network model according to the Gaussian noise sample of the target dimension and the first fracture network image sample until the training stop condition is met to obtain a fracture network generation model, the method further includes:

[0030] Randomly sample from the standard Gaussian distribution to obtain Gaussian noise of the target dimension, and the target dimension is lower than the preset dimension threshold;

[0031] Generate a fracture network image through the generator network of the fracture network generation model;

[0032] Increase the resolution of the fracture network image to the first preset resolution through an enhanced deep super-resolution network to obtain a target fracture network image;

[0033] Identify the target fracture network image through the probabilistic Hough transform method to obtain the fracture network information of the target area of the underground reservoir.

[0034] In some embodiments, the generator network includes a fully connected layer, an upsampling residual network module, a batch normalization layer, a ReLU activation function layer, a convolutional layer, and a Tanh activation function layer;

[0035] Among them, the upsampling residual network module is composed of a batch normalization layer, a ReLU activation function layer, a convolutional layer, and a depth-to-space layer.

[0036] The embodiment of the present application provides a method for generating a fracture network. The fracture network generation model provided by the embodiment of the present application is trained based on Gaussian noise samples with dimensions lower than a preset threshold and the first fracture network image samples. In the process of generating the first fracture network image samples, the corresponding prior information is added by using the marked point process method. For example, the center point position, direction, and length of the fractures all satisfy the statistical laws observed in the real underground reservoir, ensuring the existence of known fractures and the connectivity of the fracture network. Therefore, the fracture network image samples used in training the model of the present application can meet the prior information of pre-guidance. Therefore, the embodiment of the present application can improve the authenticity of the fracture network generated based on the fracture network generation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required to be used in the embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0038] Figure 1 is a schematic flowchart of a method for generating a fracture network provided by an embodiment of the present application;

[0039] Figure 2 is a schematic diagram of the principle of a method for training a fracture network generation model provided by an embodiment of the present application;

[0040] Figure 3 is a schematic structural diagram of a discriminator network provided by an embodiment of the present application;

[0041] Figure 4 is a schematic structural diagram of a downsampling residual network module provided by an embodiment of the present application;

[0042] Figure 5 is a schematic flowchart of a method for generating a fracture network provided by an embodiment of the present application;

[0043] Figure 6 is a schematic structural diagram of a generator network provided by an embodiment of the present application;

[0044] Figure 7 is a schematic structural diagram of an upsampling residual network module provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by illustrating the examples of the present application.

[0046] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the existence of other identical elements in the process, method, article or equipment including the elements.

[0047] With the introduction of the dual carbon strategy, the research and development of underground reservoirs has become an important topic in the current field of geology and geophysics, which is related to many engineering applications such as clean energy extraction and geological storage of carbon dioxide. Due to the high density of underground reservoir rocks, the widespread fracture network (Discrete Fracture Network, DFN) in the rock is the main seepage channel for fluids. The characterization of the fracture network is crucial to studying issues such as energy extraction efficiency and carbon dioxide storage safety. Therefore, the accurate characterization of the fracture network has become a key problem in the development of underground reservoirs. Since it cannot be directly observed, the fracture network in the underground reservoir can only be characterized indirectly through geological and geophysical data.

[0048] However, this characterization faces many challenges: First, it is very difficult to effectively parameterize the fracture network. Natural fractures in underground reservoirs are scattered and densely distributed, and the fractures intersect with each other, with very strong discontinuous characteristics. Directly using the coordinates of each fracture endpoint as its parameter will form a high-dimensional parameter space, and because the number of fractures in the underground reservoir is unknown, the dimension of the parameter space is also uncertain, which brings great difficulties to the characterization of the fracture network. Secondly, due to the high cost of drilling and geophysical testing, the observation data used for fracture network characterization is usually very sparse, and it is difficult to effectively constrain the high-dimensional parameter space. The characterization of complex fracture networks in underground reservoirs urgently needs to develop effective low-dimensional parameterization methods.

[0049] The existing low-dimensional parameterization methods for fracture networks can be roughly divided into two categories. One method is to make assumptions about the relevant characteristics of the fracture network through prior information such as outcrops, well logging, and cores, such as fixing the orientation or position of the fracture, thereby reducing the number of parameters required to describe the fracture network. The other method uses the powerful data analysis and complex structure generation capabilities of deep learning models to train the model using randomly generated fracture network images, thereby achieving the purpose of generating complex fracture network images from a low-dimensional latent variable space. The latent variable space can be used as the parameter space of the fracture network. The deep learning algorithms currently used include Deep Sparse Autoencoder (DSAE), Variational Autoencoder (VAE), Variational Autoencoder-Generative Adversarial Networks (VAE-GAN), etc.

[0050] However, the prior art has the following disadvantages:

[0051] In contrast to the dimensionality reduction method based on prior information constraints such as outcrops, well logging, and cores, this type of method imposes certain constraints on DFN through prior information, such as fixing the orientation of the fracture in two main directions and the growth nodes of the fracture, so that only a small number of parameters are needed to describe the number of fractures, the distance from the growth nodes, and other factors to determine the DFN; or discretizing the endpoint coordinates of the fracture to reduce the complexity of the parameter space to improve the computational efficiency of subsequent inversion characterization. Although the constraints of this type of method are based on certain prior information, the DFN generated by its parameterization is far from the real DFN characteristics. The real DFN in nature does not completely have fixed directions or growth nodes, so this type of parameterization method will distort the generated DFN, and the authenticity and reliability of the fracture network inversion characterization results based on this method will also be reduced.

[0052] For deep learning-based methods using a fracture parameter dataset as the training set, such methods can generate a DFN with high-dimensional parameters from a low-dimensional space. However, in the training dataset composed of fracture parameter data, the dimension of each data is determined. Therefore, such methods can only parameterize fracture networks with a fixed number of fractures, which is a constraint for actual inversion and characterization work and may lead to difficulty in making the characterization results approach the true results. For example, each two-dimensional fracture requires four parameters (the horizontal and vertical coordinates of two endpoints) to represent. If the dimension of the data in the training set is 40, then both the training set and the model after training can only represent a fracture network with 10 fractures. In actual application scenarios, the number of fractures actually existing in underground reservoirs is often an unknown quantity, and the parameterization method with a fixed number of fractures may reduce the reliability of the inversion and characterization results.

[0053] For deep learning-based methods using DFN images as the training set, such methods can generate high-dimensional DFN images from a low-dimensional space. However, the currently proposed methods based on the VAE model still lack the ability to generate high-quality fracture network images. There is still much room for improvement in the clarity, morphology, and continuity of the DFN images parameterized and generated by existing methods. In the VAE structure, the fracture network image is compressed into a low-dimensional space (latent variable space) by the encoder and then expanded from the low-dimensional space to the high-dimensional image space by the decoder. Although the KL constraint is added to this low-dimensional space during the training process to ensure its proximity to the standard normal distribution, there are still differences from the standard normal distribution space. Therefore, during the process of random sampling from the low-dimensional space and random generation through the decoder, due to the deviation between the randomly sampled data and the distribution of the latent variable space, it is possible to cause deviations between the generated images and the training set images, resulting in poor generated images.

[0054] In summary, the existing fracture network generation methods have the following problems: (1) The generated fracture network images are not clear and of low quality. (2) High-quality fracture network images cannot be generated with low-dimensional parameters, that is, the dimension reduction of parameters cannot be achieved, and thus the difficulty of the inversion work cannot be well reduced. (3) Due to the lack of consideration of the true statistical laws or prior information that the fracture network should have in the training set or the generation process, the authenticity of the generated fracture network is not high.

[0055] To solve the problems of the prior art, an embodiment of the present application provides a method for generating a fracture network. The fracture network generation model provided by the embodiment of the present application is trained based on Gaussian noise samples with dimensions lower than a preset threshold and first fracture network image samples. In the process of generating the first fracture network image samples, prior information is added by using the marked point process method. For example, the central point position, direction, and length of the fractures all satisfy the statistical laws observed in real underground reservoirs, ensuring the existence of known fractures and the connectivity of the fracture network. Therefore, the fracture network image samples used in training the model of the present application can satisfy the prior information of pre-guidance. Furthermore, the fracture network generation model trained therefrom can also satisfy the corresponding prior information. Therefore, the embodiment of the present application can improve the authenticity of the fracture network generated based on the fracture network generation model.

[0056] First, the fracture network generation method provided by the embodiment of the present application will be introduced below.

[0057] Figure 1 The flowchart of the fracture network generation method provided by an embodiment of the present application is shown. As Figure 1 shown, the method may include the following steps: S101 to S104.

[0058] S101: Based on the statistical distribution law of fractures in the underground reservoir, use the marked point process method to determine the central point position information, length information, and angle information of the fractures in the underground reservoir of the sample area.

[0059] A large number of field outcrop and core observation data show that the distribution of fractures in the underground reservoir conforms to a certain statistical law, that is, the distribution of underground fractures is not completely independent.

[0060] The statistical distribution law may include fractal distribution, power-law distribution, and von Mises distribution, etc. Among them, the fractal distribution is used to characterize the position characteristics of fractures, the power-law distribution is used to characterize the length characteristics of fractures, and the von Mises distribution is used to characterize the angle characteristics of fractures.

[0061] The marked point process method (MPP) can be used to determine the characteristic information of the fracture network.

[0062] S102: Generate candidate fracture network image samples according to the central point position information, length information, and angle information.

[0063] By considering the fracture statistical law of the fracture network, the generated fracture network image can satisfy the corresponding statistical law. In this way, the authenticity of the finally obtained fracture network image can be improved, and further, the reliability of the generated fracture network image can be improved.

[0064] S103: Add a first crack and remove a second crack from the candidate fracture network image sample to obtain a first fracture network image sample; the first crack represents a known crack determined based on prior information, and the second crack represents a crack that is not connected to other cracks.

[0065] In the problem of representing real underground fracture networks, prior knowledge about known fractures and fracture network connectivity can usually be obtained from geological / geophysical data and in-situ hydrogeological experiments (such as water injection / pumping and tracer tests). Although this prior knowledge is limited, it provides important hard constraints on the fracture network and is crucial for subsequent fracture network inversion work. Existing generation methods do not impose any hard constraints on known fractures or fracture network connectivity in the fracture network. Therefore, the generated DFNs cannot guarantee the existence of known fractures or the connectivity of the DFNs. In inversion based on observed flow rates and tracer data, etc., since the fluid flow and tracer transport in the DFN model require connectivity between injection points and monitoring points, the generated fracture network should also be connected. Therefore, directly using existing parameterization methods for fracture network inversion will inevitably result in the generation of a large number of "invalid" DFN models.

[0066] Therefore, in the process of S103, other prior information is added. One is to ensure the existence of known fractures by adding known existing fractures to each training sample. The other is to ensure the connectivity of the fracture network by judging the connectivity of the fracture network through a calculation program in the above-generated samples, and the connected fracture networks are retained. Finally, a fracture network dataset that can satisfy the statistical distribution law, the existence of known fractures, and ensure connectivity is formed.

[0067] The first fracture network image sample is a fracture network image for model training. Since the first fracture network image sample is generated based on the statistical law of fractures, using the first fracture network image sample for model training can ensure the effectiveness of model training.

[0068] S104: Train a preset generative adversarial network model according to the Gaussian noise sample of the target dimension and the first fracture network image sample until the training stop condition is satisfied, to obtain a fracture network generation model for generating a fracture network image according to the Gaussian noise of the target dimension; wherein, the target dimension is lower than a preset dimension threshold.

[0069] In the process of generating fracture networks in the prior art, the dimension of the low-dimensional parameter data used is still relatively high. Compared with the traditional method of using high-dimensional parameter data to generate fracture networks, its dimensionality reduction efficiency is still low. For this reason, in the embodiments of the present application, Gaussian noise with a dimension lower than the preset dimension threshold, that is, a low-dimensional Gaussian noise sample with a lower amount of parameter data, is used for model training to achieve a better dimensionality reduction effect.

[0070] The generative adversarial network model can be used for crack network generation training to obtain a crack network generation model. As Figure 2 shown, the generative adversarial network model consists of a generator network 201 and a discriminator network 202. Gaussian noise can be input into the generative adversarial network model, and the generator network 201 of the generative adversarial network model generates a crack network image. Then, the generated crack network image and the real crack network image sample are simultaneously input into the discriminator network 202 of the generative adversarial network model, and the model is trained through the discriminator network 202 to obtain a model capable of generating real crack network images.

[0071] The crack network generation model is used to generate crack network images, obtaining crack network images with high authenticity.

[0072] The crack network generation model provided by the embodiments of this application is trained based on Gaussian noise samples with dimensions lower than a preset threshold and the first crack network image samples. And corresponding prior information is added to the first crack network image samples by using the marked point process method during the generation process, for example, making the center point position, direction, and length of the cracks all satisfy the statistical laws observed in the real underground reservoir, ensuring the existence of known cracks, and ensuring the connectivity of the crack network. Therefore, the crack network image samples used in the training model of this application can meet the prior information of the pre-guidance. Furthermore, the crack network generation model trained from it can also meet the corresponding prior information. Therefore, the embodiments of this application can improve the authenticity of the crack network images generated based on the crack network generation model.

[0073] In some embodiments, S101 may include:

[0074] Mark the positions of the cracks based on the fractal distribution to generate a two-dimensional point set that satisfies the fractal distribution;

[0075] Randomly sample two-dimensional points that meet a preset quantity threshold in the two-dimensional point set;

[0076] Determine the center point position information of each crack in the underground reservoir of the sample area according to the two-dimensional points;

[0077] In some embodiments, S103 may include:

[0078] Determine the length information of the cracks based on the power-law distribution according to the center point position;

[0079] Determine the angle information of the cracks based on the von Mises distribution according to the center point position.

[0080] The power-law distribution can be expressed by the following formula:

[0081] n(l) = βl -a [1]

[0082] Where l is the crack length, n(l) is the number of cracks with lengths between l and l+dl, β is the proportionality coefficient, and a is the power law exponent.

[0083] The von Mises distribution can be expressed as follows:

[0084]

[0085] Where θ is the angle of the crack, f(θ) is the probability density of the crack angle distribution, μ is the average angle, κ is the concentration, and I 0 is the normalization constant.

[0086] By considering the crack statistical laws of the crack network, the parameterized generated crack network image can meet the corresponding statistical laws. In this way, the authenticity of the crack network image finally obtained can be improved, and then the reliability of the generated crack network can be improved.

[0087] In some embodiments, the center point location information of each fracture in the target area of ​​the underground reservoir is determined based on the statistical distribution law of the fractures in the underground reservoir. Alternatively, the center point location information of the fractures may be determined by using Poisson distribution.

[0088] In some embodiments, for each crack, the length information and angle information of the crack are determined based on the statistical distribution law according to the center point position information of the crack. The length information of the crack can also be determined by normal distribution or lognormal distribution, and the angle information of the crack can be determined by Fisher distribution, Fisher-von Mises distribution, or normal distribution.

[0089] In some embodiments, S104 may include:

[0090] Acquire a training sample set, the training sample set includes a plurality of training samples, each training sample includes a Gaussian noise sample of a target dimension and a corresponding first crack network image sample;

[0091] For each training sample, perform the following steps:

[0092] Inputting a Gaussian noise sample of the target dimension into a generative adversarial network model to obtain a second crack network image corresponding to the Gaussian noise sample of the target dimension;

[0093] Determining a training loss of a generative adversarial network model based on the first crack network image sample and the second crack network image;

[0094] When the training loss does not meet the training stop condition, the parameters of the generator network are adjusted to obtain an updated generative adversarial network model, and the Gaussian noise sample of the target dimension is returned to the generative adversarial network model to obtain a second crack network image corresponding to the Gaussian noise sample of the target dimension until the training stop condition is met to obtain a crack network generation model.

[0095] Among them, the training loss can be used to characterize the accuracy of the generative adversarial network model in generating crack network images. When the training loss meets the training stop condition, it can be considered that the accuracy of the current updated generative adversarial network model is the highest, and the current generative adversarial network model is used as the final required crack network generation model.

[0096] By continuously iteratively training the generative adversarial network model and continuously adjusting the parameters of the generator network of the generative adversarial network model, the crack network images generated by the generator network can be made more realistic, and the accuracy and authenticity of the crack network images generated by the crack network generation model can be improved.

[0097] As an example, a preset generative adversarial network model is trained according to the Gaussian noise sample of the target dimension and the first crack network image sample until the training stop condition is met to obtain a crack network generation model, which can be trained using an Adam optimizer. The initial learning rate of the Adam optimizer can be set to 0.0001, the batch size to 64, and the training iteration to 100,000 epochs, where epoch represents the number of updates in which all training sample data have been used once. The specific training process includes the following steps:

[0098] a. Randomly sample from the standard Gaussian distribution of the underground reservoir in the sample area to obtain a noise of a target dimension, input this noise into the generator network G, and output the generated DFN image.

[0099] b. Input the generated DFN image and the real image of the training sample set into the discriminator network D, and calculate the training loss using formula (1).

[0100] c. Feed the obtained training loss back to the discriminator network D, and update the weight of the discriminator network D. At the same time, set the initial value of the number of iterations i to 1 and the value of the iteration threshold n to 5.

[0101] d. Determine whether i≤n. If so, set i=i+1 and return to step c; if not, return to step d.

[0102] e. Feed the obtained training loss back to the generator network G and update the weights of the generator network G.

[0103] f. Repeat steps af until the training loss drops to a preset threshold and remains relatively stable.

[0104] In some embodiments, determining the training loss of the generative adversarial network model according to the first crack network image sample and the second crack network image includes:

[0105] Inputting the first crack network image sample and the second crack network image into the discriminator network to obtain a discrimination result;

[0106] Calculate the Wasserstein distance and gradient penalty between the first crack network image sample and the second crack network image based on the discrimination result;

[0107] According to the sum of Wasserstein distance and gradient penalty, the training loss of the generative adversarial network model is obtained.

[0108] The training loss represents the Wasserstein distance between the discriminator for the real image, i.e., the first crack network image sample, and the generated image of the generator network, i.e., the second crack network image.

[0109] By adding a gradient penalty to the training loss during the training process, a more stable training effect can be obtained, thereby improving the robustness of the trained model.

[0110] As an example, the training loss of a generative adversarial network model can be expressed as follows:

[0111]

[0112] Among them, D represents the discriminator network, x is the real sample, is the sample generated by the generator. In the gradient penalty part, ε is a parameter randomly sampled uniformly from the range [0,1], x r is data randomly sampled from the real sample, x g is data randomly sampled from the generated sample.

[0113] In some embodiments, the discriminator network may include a convolutional layer and a down-sampling residual network module; the first crack network image sample and the second crack network image are input into the discriminator network to obtain a discrimination result, including:

[0114] The features of the first crack network image sample and the second crack network image are respectively extracted through a convolution layer to obtain a first feature and a second feature;

[0115] Down-sample the first feature and the second feature respectively through a down-sampling residual network module to obtain a third feature and a fourth feature;

[0116] Based on the third feature and the fourth feature, the first crack network image sample and the second crack network image are discriminated to obtain a discrimination result.

[0117] The discriminator network is used to discriminate the authenticity of the second crack network image based on the first crack network image sample. Figure 3 As shown, the discriminator network may include a convolutional layer 301, a plurality of down-sampling residual network modules 302 and a fully connected layer 303, wherein the convolutional layer 301 is used to extract features of the first crack network image sample and the second crack network image, and the down-sampling residual network module 302 is used to down-sample the features of the first crack network image sample and the second crack network image. Figure 4 As shown, the down-sampling residual network module can be composed of several layers of normalization layer 401, ReLU activation function layer 402, convolution layer 403 and average pooling layer 404.

[0118] By inputting the first crack network image sample and the second crack network image into a discriminator network including a convolutional layer and a down-sampling residual network module to discriminate the authenticity of the second crack network image, the accuracy of authenticity discrimination of the second crack network image can be improved.

[0119] In some embodiments, the down-sampling residual network module in the discriminator network can be replaced by a U-Net network module, a RU-Net network module, an MLP network module, a transformer network module, etc.

[0120] In some embodiments, after the crack network generation model is trained, the trained crack network generation model can be used to generate a crack network image. Figure 5 The crack information is extracted by the method to complete further analysis, verification and specific application. Figure 5 . Figure 5 Schematic diagram of the process of fracture network generation method.

[0121] like Figure 5 As shown, after S104, the following steps may also be included: S501 to S504.

[0122] S501: Randomly sample from a standard Gaussian distribution to obtain Gaussian noise of a target dimension, where the target dimension is lower than a preset dimension threshold.

[0123] S502: Generate a crack network image through a trained generator network based on a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) model.

[0124] Specifically, by inputting Gaussian noise of the target dimension into the WGAN-GP model, the crack network image can be obtained through the generator network of the WGAN-GP model.

[0125] S503: Improving the resolution of the crack network image to a first preset resolution through an enhanced deep super-resolution network to obtain a target crack network image.

[0126] Specifically, by inputting the crack network image into an enhanced deep super-resolution network, the details of the crack network image are enriched by the enhanced deep super-resolution network, and the resolution of the crack network image is improved to a first preset resolution.

[0127] S504: Identify the target fracture network image by using the probabilistic Hough transform method to obtain fracture network information of the underground reservoir.

[0128] Specifically, the probabilistic Hough transform method is used to identify and extract crack information from the target crack network image, and the center point position information, length information and angle information of the cracks in the crack image information are identified to obtain the crack information in the crack network.

[0129] The enhanced deep super-resolution network (EDSR) is used to improve the resolution of a single crack network while retaining the necessary features of the image.

[0130] The first preset resolution may be manually set, for example, the first preset resolution may be 300×300 pixels.

[0131] The Probabilistic Hough Transform (PHT) method is used to perform feature recognition and extraction on a target fracture network image with a first preset resolution to obtain fracture network information of an underground reservoir in a target area, where the fracture network information may include the position, length and angle of the fractures.

[0132] By inputting the low-dimensional Gaussian noise of the underground reservoir obtained from the target area into the fracture network generation model trained by the above embodiment, a fracture network image with high authenticity can be generated through the generator network of the fracture network generation model, and then the resolution of the fracture network image is improved by the enhanced deep super-resolution network, and the probabilistic Hough transform method is used to perform feature recognition and extraction on the fracture network image with high resolution, so that the fracture network information of the underground reservoir with high authenticity in the target area can be obtained, thereby achieving accurate generation of the fracture network of the underground reservoir in the target area under low-dimensional parameters.

[0133] In some embodiments, the generator network may include an upsampling residual network module; S502 may include:

[0134] The Gaussian noise of the target dimension is upsampled through the upsampling residual network module to obtain the fifth feature;

[0135] A fracture network image satisfying a second preset resolution condition is generated based on the fifth feature.

[0136] The generator network is used to generate crack network images based on Gaussian noise of the target dimension. Figure 6 As shown, the generator network may include a fully connected layer 601, several upsampling residual network modules 602, a batch normalization layer 603, a ReLU activation function layer 604, a convolution layer 605 and a Tanh activation function layer 606, wherein the upsampling residual network module 602 is used to upsample the Gaussian noise of the target dimension. Figure 7 As shown, the upsampling residual network module 602 can be composed of several batch normalization layers 701, ReLU activation function layers 702, convolutional layers 703 and depth-to-space layers 704.

[0137] The second preset resolution may be manually set, for example, the second preset resolution may be 64×64 pixels.

[0138] By inputting Gaussian noise of the target dimension into a generator network including an upsampling residual network module to generate realistic crack network images, the accuracy of generating crack network images can be improved.

[0139] In some embodiments, the upsampling residual network module in the generator network can be replaced by a U-Net network module, a RU-Net network module, an MLP network module, a transformer network module, etc.

[0140] The above is only a specific implementation of the present application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the protection scope of the present application is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the protection scope of this application.

Claims

1. A method for generating a fracture network, characterized in that: include: Based on the statistical distribution law of cracks in underground reservoirs, the standard value point process method is used to determine the center point position information, length information and angle information of the cracks in the underground reservoirs in the sample area; Generate a candidate crack network image sample according to the center point position information, the length information and the angle information; Adding a first crack to the candidate crack network image sample and removing a second crack to obtain a first crack network image sample; the first crack represents a known crack determined based on prior information, and the second crack represents a crack that is not connected to other cracks; A preset generative adversarial network model is trained according to the Gaussian noise samples of the target dimension and the first crack network image samples until the training stop condition is met, so as to obtain a crack network generation model for generating a crack network image according to the Gaussian noise of the target dimension; wherein the target dimension is lower than a preset dimension threshold.

2. The method according to claim 1, characterized in that The statistical distribution laws include fractal distribution, power law distribution and von-Mises distribution; The method of determining the center point position information, length information and angle information of the cracks in the underground reservoir in the sample area by using the standard value point process method based on the statistical distribution law of the cracks in the underground reservoir includes: Marking the position of the crack based on the fractal distribution, and generating a two-dimensional point set satisfying the fractal distribution; Randomly sampling two-dimensional points that meet a preset number threshold in the two-dimensional point set; Determine the center point position information of each fracture in the underground reservoir of the sample area according to the two-dimensional points; According to the position of the center point, determining the length information of the crack based on the power law distribution; According to the position of the center point, angle information of the crack is determined based on the von Mises distribution.

3. The method according to claim 1, characterized in that The generative adversarial network model includes a generator network; the preset generative adversarial network model is trained according to the Gaussian noise sample of the target dimension and the first crack network image sample until the training stop condition is met to obtain the crack network generation model, including: Acquire a training sample set, wherein the training sample set includes a plurality of training samples, each of the training samples includes a Gaussian noise sample of the target dimension and a corresponding first crack network image sample; The following steps are performed for each training sample: Inputting the Gaussian noise sample of the target dimension into the generative adversarial network model to obtain a second crack network image corresponding to the Gaussian noise sample of the target dimension; Determining a training loss of the generative adversarial network model according to the first crack network image sample and the second crack network image; When the training loss does not meet the training stop condition, the parameters of the generator network are adjusted to obtain an updated generative adversarial network model, and the Gaussian noise sample of the target dimension is returned to the generative adversarial network model to obtain a second crack network image corresponding to the Gaussian noise sample of the target dimension, until the training stop condition is met to obtain the crack network generation model.

4. The method according to claim 3, characterized in that The generative adversarial network model includes a discriminator network; and determining the training loss of the generative adversarial network model according to the first crack network image sample and the second crack network image includes: Inputting the first crack network image sample and the second crack network image into the discriminator network to obtain a discrimination result; Calculating the Wasserstein distance and gradient penalty between the first crack network image sample and the second crack network image based on the discrimination result; The training loss of the generative adversarial network model is obtained according to the sum of the Wasserstein distance and the gradient penalty.

5. The method according to claim 4, characterized in that The discriminator network includes a convolutional layer, a down-sampling residual network module and a fully connected layer; Among them, the downsampling residual network module consists of a normalization layer, a ReLU activation function layer, a convolution layer, and an average pooling layer.

6. The method according to any one of claims 1 to 5, characterized in that: After training a preset generative adversarial network model according to the Gaussian noise sample of the target dimension and the first crack network image sample until a training stop condition is met and a crack network generation model is obtained, the method further includes: Randomly sampling from a standard Gaussian distribution to obtain Gaussian noise of a target dimension, wherein the target dimension is lower than a preset dimension threshold; Generate a crack network image through a generator network of the crack network generation model; Improving the resolution of the crack network image to a first preset resolution through an enhanced deep super-resolution network to obtain a target crack network image; The target fracture network image is identified by a probabilistic Hough transform method to obtain fracture network information of the target area of ​​the underground reservoir.

7. The method according to claim 6, characterized in that The generator network includes a fully connected layer, an upsampling residual network module, a batch normalization layer, a ReLU activation function layer, a convolution layer and a Tanh activation function layer; Among them, the upsampling residual network module consists of a batch normalization layer, a ReLU activation function layer, a convolutional layer, and a depth-to-space layer.

Images