Generative adversarial network sedimentary facies modeling method and device based on single training image

Abstract

The present invention relates to a method and device for modeling sedimentary phases based on a generative adversarial network with a single training image. The method constructs a sedimentary phase simulation model with a multi-level generative adversarial network of a pyramid structure, obtains a single real sedimentary phase training image, and generates a multi-scale hierarchical training image based on the training image to train each level of the generative adversarial network. During the training process, the error between the output image of the generator in each layer of the network and the preset condition is judged, and the output image is corrected when the error threshold is exceeded. Finally, the sedimentary phase simulation model with complete training is used to simulate the sedimentary phase based on the given river channel training image and sedimentary phase generation conditions. The present invention reduces the acquisition cost of training data, has a fast modeling speed and high efficiency, and can perform random modeling of sedimentary phases and simulate the uncertainty of sedimentary phases on the premise of meeting given conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of geological modeling, and in particular to a method and device for generating adversarial network sedimentary phase modeling based on a single training image. Background Art

[0002] Sedimentary facies simulation refers to the use of well data measured at sparse physical locations to infer the sedimentary facies of the entire study area. Sedimentary facies modeling is an important discipline in geology and is of great significance for oil exploration, geological disaster assessment, geological evolution research, etc. Through sedimentary facies modeling, we can better understand the spatial distribution, reserve changes, lithological characteristics, etc. of sediments in the strata, providing a scientific basis for geological exploration, engineering planning and environmental assessment.

[0003] Early geological modeling mainly used two-point statistics to simulate geological patterns through variograms. Since only the correlation between two points in space is considered, it is difficult to reproduce complex geological structural characteristics. Based on this, the improved multi-point geological statistics obtains the spatial structural information of geological variables by scanning training images, and the simulation effect is better than the two-point geological statistics. However, since the expression of spatial structure (geological pattern) is limited to multi-point statistical information within the data template, when the reservoir in the study area is non-stationary and strongly heterogeneous, the effect of multi-point geological statistics is not ideal. At the same time, multi-point geological statistics has shortcomings in modeling efficiency and memory occupancy, and cannot meet actual use needs.

[0004] In recent years, deep learning methods have been gradually introduced into the field of reservoir modeling and have quickly become a research hotspot. Generative adversarial networks (GANs) can abstract and reproduce complex spatial pattern features, and have been successfully applied in many fields such as image generation. They have also been introduced into reservoir modeling in recent years. Traditional modeling methods based on generative adversarial networks are all based on multi-image training and are mainly used in sedimentary phase modeling of simple reservoirs, such as river channel simulation of two phase types. For sedimentary phase modeling with complex structures, it is necessary to develop a corresponding data set construction algorithm, which requires a large amount of geological information and training data. However, it is difficult and costly to obtain a large number of data sets that meet the characteristics of reservoir deposition. Traditional modeling methods based on generative adversarial networks are difficult to achieve large-scale application.

[0005] Therefore, it is necessary to propose a generative adversarial network sedimentary facies modeling method and device based on a single training image, which can reduce the cost and difficulty of acquiring model sample data, extract geological features of different scales using a single training image, generate high-quality and diverse data samples, and accurately simulate sedimentary facies with complex structures, providing better guidance for oilfield development. Summary of the invention

[0006] In view of this, the present invention provides a generative adversarial network sedimentary facies modeling method and device based on a single training image, so as to solve the technical problem that it is difficult and costly to obtain a large number of data sets that meet the reservoir sedimentary characteristics, resulting in the inability to give full play to the advantages of the generative adversarial network to model the sedimentary facies and provide theoretical guidance for oil field development.

[0007] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0008] In a first aspect, the present invention provides a sedimentary facies modeling method based on a generative adversarial network of a single training image, comprising:

[0009] Constructing a sedimentary facies simulation model with a pyramid structure and a multi-level generative adversarial network; each level of the generative adversarial network in the pyramid structure corresponds to a group of generators and discriminators;

[0010] Acquire a single real sedimentary phase training image, generate a multi-scale hierarchical training image according to the sedimentary phase training image, and use each of the hierarchical training images as a training image for a corresponding level of generative adversarial network;

[0011] The hierarchical training images are used to train the generative adversarial networks of each level of the sedimentary facies simulation model to obtain a fully trained sedimentary facies simulation model; wherein, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, and the corrected image is used as the input of the current network until the error between the output image of the current generator and the preset condition is less than the preset threshold;

[0012] A river channel training image and sedimentary phase generation conditions are obtained, and sedimentary phase modeling is performed according to the river channel training image and sedimentary phase generation conditions using the well-trained sedimentary phase simulation model.

[0013] Furthermore, the sedimentary phase simulation model includes N-level to 0-level generative adversarial networks connected in sequence; the n-th level generative adversarial network includes mutually coupled generators G n and the discriminator D n , n∈[0,1,…,N];

[0014] The hierarchical training images are used to train the generative adversarial networks of various levels of the sedimentary facies simulation model, including:

[0015] Using the classified training image x n Training the discriminator D n ,The training process is carried out level by level, and the parameters do not change after each level of training is completed;

[0016] The network level N passes a random vector Z N As input, after the generator GN After getting the initial sample Except for the Nth level, each level of generator upsamples the output image of the generator at the previous level and adds random noise as the input of the network at this level.

[0017] Furthermore, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, including:

[0018] When the error between the well point location data of the generator output image in any level of the generative adversarial network and the prior well point observation data exceeds a preset threshold, the well point location data in the output image is replaced.

[0019] Furthermore, the generator G in each level of the generative adversarial network n and the discriminator D n Both include a convolutional network, which includes several convolutional layers and batch normalization layers.

[0020] Furthermore, the convolutional layer uses LReLU as an activation function, and the number of filters in the convolutional layer increases with the increase of the level of the pyramid structure.

[0021] Furthermore, the loss function of the n-th level generative adversarial network is:

[0022]

[0023] Among them, L adv (G n ,D n ) is the conditional adversarial network loss. The training goal of the generator is to maximize the gap between the generated target and the real sample under the preset conditions, and the training goal of the discriminator is to minimize the distance between the real sample and the generated sample.

[0024] Furthermore, the output image of the generator at the previous level is upsampled, including:

[0025] Upsample the output image of the generator at the previous level according to the preset sampling coefficient.

[0026] In a second aspect, the present invention also provides a generative adversarial network sedimentary facies modeling device based on a single training image, comprising:

[0027] A model building module, used to build a sedimentary facies simulation model with a pyramid structure and a multi-level generative adversarial network; each level of the generative adversarial network in the pyramid structure corresponds to a group of generators and discriminators;

[0028] A data acquisition module is used to acquire a single real sedimentary phase training image, generate a multi-scale hierarchical training image according to the sedimentary phase training image, and use each of the hierarchical training images as a training image for a corresponding level of generative adversarial network;

[0029] A training module, used to train the generative adversarial networks of each level of the sedimentary facies simulation model using the hierarchical training images to obtain a fully trained sedimentary facies simulation model; wherein, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, and the corrected image is used as the input of the current network until the error between the output image of the current generator and the preset condition is less than the preset threshold;

[0030] The output module is used to obtain the river channel training image and the sedimentary phase generation conditions, and use the well-trained sedimentary phase simulation model to perform sedimentary phase modeling according to the river channel training image and the sedimentary phase generation conditions.

[0031] In a third aspect, the present invention also provides an electronic device comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the generative adversarial network sedimentary phase modeling method based on a single training image as described in any of the above technical solutions is implemented.

[0032] In a fourth aspect, the present invention also provides a computer-readable storage medium for storing computer-readable programs or instructions. When the program or instructions are executed by a processor, the steps of the generative adversarial network sedimentary phase modeling method based on a single training image in any of the above-mentioned implementation methods can be implemented.

[0033] The present invention provides a method and device for modeling sedimentary phases based on a generative adversarial network with a single training image. The method of the present invention constructs a sedimentary phase simulation model with a pyramid structure and a multi-level generative adversarial network. By acquiring a single real sedimentary phase training image and generating a multi-scale hierarchical training image according to the training image, each level of the generative adversarial network is trained. During the training process, the error between the output image of the generator in each layer of the network and the preset condition is judged, and the output image is corrected when the error threshold is exceeded. Finally, the sedimentary phase simulation model with complete training is used to simulate the sedimentary phase based on the given river channel training image and sedimentary phase generation conditions. The present invention uses a pyramid-shaped generative adversarial network structure, and the hierarchical training network learns geological features of different scales from a single training image; in the model generation stage, based on the idea of ​​image editing, the intermediate model is modified to achieve conditional phase simulation; the acquisition cost of training data is reduced, the modeling speed is fast, and the efficiency is high. The sedimentary phase can be randomly modeled on the premise of meeting the given conditions, and the uncertainty of the sedimentary phase can be simulated. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of the process of applying the GAN model provided by the present invention to sedimentary phase modeling;

[0035] Figure 2 A schematic diagram of the flow chart of the sedimentary facies modeling method based on a single training image generated by a generative adversarial network provided by the present invention;

[0036] Figure 3 A schematic structural diagram of an embodiment of a sedimentary phase simulation model provided by the present invention;

[0037] FIG4( a ) is a river channel phase training diagram used in the model training of the present invention;

[0038] FIG4( b ) is a non-stationary simulation training diagram used in training the model of the present invention;

[0039] FIG4( c ) is a porous medium training image used in training the model of the present invention;

[0040] FIG5( a) is a schematic diagram of the training results of the unconditional simulation provided by the present invention;

[0041] FIG5( b ) is a schematic diagram of training results of conditional simulation provided by the present invention;

[0042] Figure 6 A schematic diagram of the modeling effect of the sedimentary phase simulation model provided by the present invention;

[0043] Figure 7 A schematic structural diagram of an embodiment of a generative adversarial network sedimentary facies modeling device based on a single training image provided by the present invention;

[0044] Figure 8 A schematic structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0046] Before introducing the embodiments of the present invention, the relevant terms and inventive concepts of the present application are explained first.

[0047] GAN: The Generative Adversarial Networks (GAN) framework was first proposed by (Goodfellow et al., 2014). It is a framework consisting of deep learning models that aims to generate realistic data samples through two competing neural networks. The typical architecture of GAN consists of two neural networks: the discriminator and the and the generator G θ . and θ represent the trainable weight values ​​in the network. Figure 1 As shown, Figure 1 Shown is the typical workflow when GAN is applied to sedimentary facies modeling.

[0048] Generator G θ The discriminator transforms the input random noise into potential high-dimensional data (such as images, sedimentary facies models, etc.). The generated samples and real data are mapped into two scores. The goal of the discriminator is to distinguish whether the sample comes from real data or from the sample generated by the generator. The goal of the generator is to confuse the discriminator by gradually optimizing the results it generates, so that it cannot determine the true source of the sample. The generator and the discriminator are trained and optimized in sequence, and finally reach a dynamic equilibrium called Nash equilibrium. After training, the samples generated by the generator are realistic enough that the discriminator has a probability of 0.5 for all samples input (whether generated by the generator or real).

[0049] The training of the GAN network is achieved by optimizing the following loss function:

[0050]

[0051] Among them, x represents the real image, z represents the input G network noise, G θ (z) represents the image generated by the G network, Represents the probability of a real image, output 1 indicates a 100% real image, and output 0 indicates an image that is not real.

[0052] SinGAN: Single Image Generative Adversarial Network (SinGAN) is a non-conditional generative model that can learn image distribution characteristics from a single natural image. Traditional single image generative models usually "overfit" deep models to a single training example. These methods are designed for specific tasks, such as super-resolution and texture extension rather than for drawing random samples. The SinGAN framework is purely generative, that is, mapping noise to image samples, which conforms to the idea of ​​random modeling and can be used to describe reservoir uncertainty. Based on a single training image, the training image construction cost is low, the efficiency is high, and there are more ways, such as dense well network dissection, mature oil field constructed model, outcrop 3D dissection, seismic information conversion, etc., which can greatly reduce the application cost of generative adversarial networks.

[0053] CGAN: Conditional Generative Adversarial Network (CGAN) follows the network architecture in Generative Adversarial Network, but adds conditional probabilities to allow conditional information as input, such as labels, text, and images.

[0054] Since it is difficult and costly to obtain a large number of data sets that meet the characteristics of reservoir deposition, existing model research is objectively limited, so it is necessary to develop intelligent modeling technology based on a single training image. Inspired by the conditional generative adversarial network, we applied the generative adversarial network trained with a single image to reservoir sedimentary phase modeling. Although single-image training has no statistical advantage over multi-image training, this method is a good alternative when data sets are scarce. At the same time, the goal of network learning is the distribution characteristics of small blocks inside the image, which is similar to the idea of ​​image splicing and style-based multi-point statistical modeling. Compared with the above two methods, the generative adversarial network method has more powerful pattern reproduction and non-stationarity processing capabilities.

[0055] Based on the SinGAN network and the CGAN network, this application proposes a three-dimensional conditional simulation method CSinGAN based on a single training image. By constructing a pyramid-shaped generative adversarial network structure, the hierarchical training network learns geological features of different scales from a single training image. In the model generation stage, based on the idea of ​​image editing, conditional phase simulation is achieved by modifying the intermediate model. After the network training is completed, the generative adversarial network method has a very fast modeling speed, and it usually takes only a few seconds to build a model. For the conditioning of well point data, we have explored a more convenient and transparent method that allows the generated image to be modified directly to achieve conditional simulation.

[0056] The present invention provides a generative adversarial network sedimentary phase modeling method, device, electronic device and computer-readable storage device based on a single training image, which are described below respectively.

[0057] Combination Figure 2 As shown, a specific embodiment of the present invention discloses a sedimentary facies modeling method based on a generative adversarial network of a single training image, comprising:

[0058] Step S101, constructing a sedimentary facies simulation model with a multi-level generative adversarial network in a pyramid structure; each level of the generative adversarial network in the pyramid structure corresponds to a group of generators and discriminators;

[0059] Step S102, obtaining a single real sedimentary phase training image, generating a multi-scale hierarchical training image according to the sedimentary phase training image, and using each of the hierarchical training images as a training image for a corresponding level of generative adversarial network;

[0060] Step S103, using the hierarchical training images to train the generative adversarial networks of each level of the sedimentary facies simulation model to obtain a fully trained sedimentary facies simulation model; wherein, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, and the corrected image is used as the input of the current network until the error between the output image of the current generator and the preset condition is less than the preset threshold;

[0061] Step S104, obtaining a river channel training image and sedimentary phase generation conditions, and using the well-trained sedimentary phase simulation model to perform sedimentary phase modeling according to the river channel training image and sedimentary phase generation conditions.

[0062] Compared with the prior art, the method of this embodiment constructs a pyramid-shaped sedimentary phase simulation model with a multi-level generative adversarial network. By acquiring a single real sedimentary phase training image and generating a multi-scale hierarchical training image based on the training image, each level of the generative adversarial network is trained. During the training process, the error between the output image of the generator in each layer of the network and the preset condition is judged. When the error threshold is exceeded, the output image is corrected. Finally, the sedimentary phase simulation model with complete training is used to simulate the sedimentary phase based on the given river channel training image and sedimentary phase generation conditions. The method of this embodiment uses a pyramid-shaped generative adversarial network structure, and the hierarchical training network learns geological features of different scales from a single training image; in the model generation stage, based on the idea of ​​image editing, the intermediate model is modified to achieve conditional phase simulation; the acquisition cost of training data is reduced, the modeling speed is fast, and the efficiency is high. It can randomly model the sedimentary phase under the premise of meeting the given conditions and simulate the uncertainty of the sedimentary phase.

[0063] As a preferred embodiment, in step S101, the sedimentary facies simulation model includes N-level to 0-level generative adversarial networks connected in sequence; the n-th level generative adversarial network includes mutually coupled generators G n and the discriminator D n , n∈[0,1,…,N];

[0064] The hierarchical training images are used to train the generative adversarial networks of various levels of the sedimentary facies simulation model, including:

[0065] Using the classified training image x n Training the discriminator D n ,The training process is carried out level by level, and the parameters do not change after each level of training is completed;

[0066] The network level N passes a random vector Z N As input, after the generator G N After getting the initial sample Except for the Nth level, each level of generator upsamples the output image of the generator at the previous level and adds random noise as the input of the network at this level.

[0067] As a preferred embodiment, upsampling the output image of the generator at the previous level includes:

[0068] Upsample the output image of the generator at the previous level according to the preset sampling coefficient.

[0069] As a specific example, Figure 3 As shown, the sedimentary phase simulation model of this embodiment is designed as a pyramid structure, which consists of a generator pyramid {G0, ...G N} and the discriminator pyramid {D0,…D N The network pyramid is trained from level N to level 0 in sequence, and the training process is carried out level by level. After each level of training is completed, the parameters will not change. N} represents training images of different levels, which are used in the training of each level respectively, {Z0,…Z N} represents different levels of network input, Represents the generated samples at different levels.

[0070] The first layer of the network consists only of random vectors Z N As input, after the generator G N After getting the initial sample The generated image is then upsampled (↑ r , r is the sampling coefficient). Except for the first layer, the subsequent generators need to use the generation results of the previous layer as input and add random noise.

[0071] Single-image generative adversarial networks have good performance in image editing tasks, that is, inputting edited images (such as adding, deleting, and moving image blocks) into an intermediate-scale generator, and the output result can restore the realism of the image. Inspired by the image editing task, when there is a large error between the feature data of the output model and the observed actual situation, we can correct the feature data of the output model.

[0072] As a preferred embodiment, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, including:

[0073] When the error between the well point location data of the generator output image in any level of the generative adversarial network and the prior well point observation data exceeds a preset threshold, the well point location data in the output image is replaced.

[0074] Specifically, the modified image is used as the input of the current layer network again, and then the output of the generator is tested. If there is still a mismatch, the result is returned to the input of the generator and passes through the generator again until the network numerical error is less than a given threshold.

[0075] See also Figure 3 , Figure 3 The structure of one of the generative adversarial networks is partially displayed. In the figure, R represents the use of well point observation data to replace the data of the corresponding well point location in the model, where the observation data is sampled according to the grid scale. It is worth noting that the network only combines conditional data in the generation process, while the training process is carried out in an unconditional manner.

[0076] As a preferred embodiment, the generator G in each level of the generative adversarial network n and the discriminator D n Both include a convolutional network, which includes several convolutional layers and batch normalization layers.

[0077] As a preferred embodiment, the convolution layer adopts LReLU as an activation function, and the number of filters in the convolution layer increases with the increase of the level of the pyramid structure.

[0078] As a preferred embodiment, the loss function of the n-th level generative adversarial network is:

[0079]

[0080] Among them, L adv (G n ,D n ) is the conditional adversarial network loss. The training goal of the generator is to maximize the gap between the generated target and the real sample under the preset conditions, and the training goal of the discriminator is to minimize the distance between the real sample and the generated sample.

[0081] Different from the loss function in the original generative adversarial network, this application modifies the input image whose error is greater than the set threshold through preset conditions, re-enters the model to generate the image again, and transforms the ordinary probability into conditional probability.

[0082] In order to verify the actual modeling effect of this application, as shown in Figures 4(a), 4(b) and 4(c), we used three different types of training images to evaluate the performance of the sedimentary phase simulation model of this application (hereinafter referred to as CSinGAN model). Figure 4(a) is a river channel training image widely used in sedimentary phase modeling algorithm testing (Strebelle et al., 2000, 2002; Zhang et al., 2006), Figure 4(b) is used to test the performance of the algorithm in non-stationary simulation; Figure 4(c) is a porous media training image, which is often used in pore reconstruction algorithm testing. The grid size of the three training images is 250×250.

[0083] As shown in Table 1, Table 1 shows the two-dimensional network parameters of the sedimentary phase simulation model CSinGAN used in this implementation. Each layer of the network pyramid contains a set of generative adversarial networks, and both the generative network and the discriminative network contain five convolutional layers. The slope of the LReLU activation function is set to 0.2, the initial value of the number of filters is 32, and the number of filters doubles every four levels of the network pyramid. The output scale (N×N) of each layer of the network pyramid is 25×25, 33×33, 44×44, 60×60, 80×80, 108×108, 144×144, 200×200 respectively; there are nine layers in total.

[0084] Table 1 Two-dimensional architecture of the generator and discriminator

[0085]

[0086] In Table 1, 2D convolution 5x5 represents a 2D convolution layer with a 5×5 convolution kernel, and N×N represents the model size of the current layer in the network pyramid.

[0087] CSinGAN is based on the idea of ​​image editing to achieve conditionalization. In order to verify whether the practice of modifying the intermediate results during the generation process is reasonable and effective, we fixed the network input and compared the difference in simulation effects between modified and unmodified. Please refer to Figure 5(a) and Figure 5(b). Four representative layers of output were selected from the network pyramid for comparison. Figure 5(a) shows the result of unconditional simulation, and Figure 5(b) shows the result after conditional simulation.

[0088] Comparing Figure 5(a) and Figure 5(b), we can see that when we modify the phase type of the well point location, the model is not only correctly simulated at the well point location, but also the river channel structure around the well point is reasonably represented, indicating that the conditional strategy based on image editing is a simple and feasible approach.

[0089] In order to further verify the feasibility of the conditioning strategy, we still use Figure 4(a), Figure 4(b), and Figure 4(c) as different training images and increase the number of well points for testing. Figure 6 As shown, Figure 6 Rows 1 to 3 in the middle are the CSinGAN conditional simulation implementations of the three training images in Figure 4(a), Figure 4(b), and Figure 4(c), respectively.

[0090] The first column is the conditional data, and columns 2 to 4 are three sets of conditional realizations obtained using different initial inputs. As can be seen from the figure, CSinGAN can generate different random realizations under the premise of satisfying the conditional data, further demonstrating the feasibility of the conditionalization strategy.

[0091] In order to better implement the generative adversarial network sedimentary facies modeling method based on a single training image in the embodiment of the present invention, based on the generative adversarial network sedimentary facies modeling method based on a single training image, correspondingly, please refer to Figure 7 , Figure 7 A schematic diagram of the structure of an embodiment of a generative adversarial network sedimentary facies modeling device based on a single training image provided by the present invention. A generative adversarial network sedimentary facies modeling device based on a single training image 700 provided by an embodiment of the present invention comprises:

[0092] A model building module 701 is used to build a sedimentary facies simulation model with a pyramid structure and a multi-level generative adversarial network; each level of the generative adversarial network in the pyramid structure corresponds to a group of generators and discriminators;

[0093] The data acquisition module 702 is used to acquire a single real sedimentary phase training image, generate a multi-scale hierarchical training image according to the sedimentary phase training image, and use each of the hierarchical training images as a training image for a corresponding level of generative adversarial network;

[0094] The training module 703 is used to train the generative adversarial networks of each level of the sedimentary facies simulation model using the hierarchical training images to obtain a fully trained sedimentary facies simulation model; wherein, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, and the corrected image is used as the input of the current network until the error between the output image of the current generator and the preset condition is less than the preset threshold;

[0095] The output module 704 is used to obtain the river channel training image and the sedimentary phase generation conditions, and use the well-trained sedimentary phase simulation model to perform sedimentary phase modeling according to the river channel training image and the sedimentary phase generation conditions.

[0096] It should be noted here that the corresponding device 700 provided in the above embodiment can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can refer to the corresponding contents in the above method embodiments, which will not be repeated here.

[0097] like Figure 8 As shown, the above-mentioned method for modeling sedimentary phases based on a generative adversarial network of a single training image, the present invention also provides an electronic device 800 accordingly, which can be a computing device such as a mobile terminal, a desktop computer, a notebook, a PDA, and a server. The electronic device includes a processor 801, a memory 802, and a display 803.

[0098] In some embodiments, the memory 802 may be an internal storage unit of a computer device, such as a hard disk or memory of a computer device. In other embodiments, the memory 802 may also be an external storage device of a computer device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the computer device. Further, the memory 802 may also include both an internal storage unit of a computer device and an external storage device. The memory 802 is used to store application software and various types of data installed on the computer device, such as program codes for installing the computer device. The memory 802 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a program 804 of a sedimentary phase modeling method based on a single training image for generating adversarial networks is stored on the memory 802, and the program 804 of a sedimentary phase modeling method based on a single training image for generating adversarial networks can be executed by the processor 801, thereby realizing a sedimentary phase modeling method based on a single training image for generating adversarial networks in various embodiments of the present invention.

[0099] In some embodiments, the processor 801 can be a central processing unit (CPU), a microprocessor or other data processing chip, which is used to run the program code or process data stored in the memory 802, such as executing a generative adversarial network sedimentary phase modeling method program based on a single training image.

[0100] In some embodiments, the display 803 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 803 is used to display information on the computer device and to display a visual user interface. The components 801-803 of the computer device communicate with each other through a system bus.

[0101] This embodiment also provides a computer-readable storage medium, on which a generative adversarial network sedimentary phase modeling program based on a single training image is stored. When the generative adversarial network sedimentary phase modeling program based on a single training image is executed by a processor, the steps in the above embodiment can be implemented.

[0102] The present invention provides a method and device for modeling sedimentary phases based on a generative adversarial network of a single training image. The method of the present application constructs a sedimentary phase simulation model with a multi-level generative adversarial network of a pyramid structure, obtains a single real sedimentary phase training image, and generates a multi-scale hierarchical training image according to the training image to train each level of the generative adversarial network. During the training process, the error between the output image of the generator in each layer of the network and the preset condition is judged, and the output image is corrected when the error threshold is exceeded. Finally, the sedimentary phase simulation model with complete training is used to simulate the sedimentary phase based on the given river channel training image and sedimentary phase generation conditions.

[0103] The present invention uses a pyramid-shaped generative adversarial network structure to hierarchically train the network to learn geological features of different scales from a single training image; in the model generation stage, based on the idea of ​​image editing, the intermediate model is modified to achieve conditional phase simulation; the cost of obtaining training data is reduced, the modeling speed is fast and efficient, and the sedimentary phase can be randomly modeled on the premise of meeting given conditions, and the uncertainty of the sedimentary phase can be simulated.

[0104] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A generative adversarial network sedimentary facies modeling method based on a single training image, characterized in that: include: Construct a pyramid-structured sedimentary facies simulation model with a multi-level generative adversarial network; Each level of the generative adversarial network in the pyramid structure corresponds to a group of generators and discriminators; Acquire a single real sedimentary phase training image, generate a multi-scale hierarchical training image according to the sedimentary phase training image, and use each of the hierarchical training images as a training image for a corresponding level of generative adversarial network; The generative adversarial networks of each level of the sedimentary facies simulation model are trained using the hierarchical training images to obtain a fully trained sedimentary facies simulation model; wherein, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, and the corrected image is used as the input of the current network until the error between the output image of the current generator and the preset condition is less than the preset threshold; including: when the error between the well point location data of the output image of the generator in any level of the generative adversarial network and the prior well point observation data exceeds a preset threshold, the well point location data in the output image is replaced; A river channel training image and sedimentary phase generation conditions are obtained, and sedimentary phase modeling is performed according to the river channel training image and sedimentary phase generation conditions using the well-trained sedimentary phase simulation model.

2. The method for sedimentary facies modeling based on a generative adversarial network using a single training image according to claim 1, characterized in that: The sedimentary facies simulation model comprises N-level to 0-level generative adversarial networks connected in sequence; the n-th level generative adversarial network comprises mutually coupled generators G n and the discriminator D n , n∈[0,1,…,N]; The hierarchical training images are used to train the generative adversarial networks of various levels of the sedimentary facies simulation model, including: Using the classified training image x n Training the discriminator D n ,The training process is carried out level by level, and the parameters do not change after each level of training is completed; The network level N passes a random vector Z N As input, after the generator G N After getting the initial sample Except for the Nth level, each level of generator upsamples the output image of the generator at the previous level and adds random noise as the input of the network at this level.

3. The method for sedimentary facies modeling based on a generative adversarial network using a single training image according to claim 2, characterized in that: In each level of the generative adversarial network, the generator G n and the discriminator D n Both include a convolutional network, which includes several convolutional layers and batch normalization layers.

4. The method for sedimentary facies modeling based on a generative adversarial network using a single training image according to claim 3, characterized in that: The convolution layer uses LReLU as an activation function, and the number of filters in the convolution layer increases with the increase of the level of the pyramid structure.

5. The method for sedimentary facies modeling based on a generative adversarial network using a single training image according to claim 2, characterized in that: The loss function of the n-th level generative adversarial network is: Among them, L adv (G n ,D n ) is the conditional adversarial network loss. The training goal of the generator is to maximize the gap between the generated target and the real sample under the preset conditions, and the training goal of the discriminator is to minimize the distance between the real sample and the generated sample.

6. The method for sedimentary facies modeling based on a generative adversarial network using a single training image according to claim 2, characterized in that: Upsample the output image of the previous generator, including: Upsample the output image of the generator at the previous level according to the preset sampling coefficient.

7. A generative adversarial network sedimentary facies modeling device based on a single training image, characterized in that: include: A model building module, which is used to construct a pyramid-structured sedimentary facies simulation model with a multi-level generative adversarial network; Each level of the generative adversarial network in the pyramid structure corresponds to a group of generators and discriminators; A data acquisition module is used to acquire a single real sedimentary phase training image, generate a multi-scale hierarchical training image according to the sedimentary phase training image, and use each of the hierarchical training images as a training image for a corresponding level of generative adversarial network; A training module is used to train the generative adversarial networks of each level of the sedimentary facies simulation model using the hierarchical training images to obtain a fully trained sedimentary facies simulation model; wherein, when the error between the output image of the generator in any level of the generative adversarial network and the preset condition exceeds a preset threshold, the output image is corrected, and the corrected image is used as the input of the current network until the error between the output image of the current generator and the preset condition is less than the preset threshold; including: when the error between the well point location data of the output image of the generator in any level of the generative adversarial network and the prior well point observation data exceeds a preset threshold, the well point location data in the output image is replaced; The output module is used to obtain the river channel training image and the sedimentary phase generation conditions, and use the well-trained sedimentary phase simulation model to perform sedimentary phase modeling according to the river channel training image and the sedimentary phase generation conditions.

8. An electronic device, characterized in that: It includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the sedimentary phase modeling method based on a generative adversarial network of a single training image as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the generative adversarial network sedimentary phase modeling method based on a single training image as described in any one of claims 1 to 6.

Images