Random medium reconstruction method based on generative adversarial network (ELGAN)

Abstract

The invention discloses a random medium reconstruction method based on a generative adversarial network (ELGAN). The random medium reconstruction method comprises the following working steps: S1, acquiring three-dimensional random medium image data; s2, carrying out binarization or segmentation preprocessing on the image in the step 1; s3, constructing a deep learning network model which is constructed based on a generative adversarial network architecture and comprises a generator, a discriminator and a neural network pruning module; the generator is constructed based on InfoGAN, the discriminator is constructed based on a deep convolutional neural network, and the neural network pruning module is constructed based on an LRP technology. According to the method, the controllability of a reconstruction result can be effectively improved, pore distribution is constrained by condition data, and the generated result can be directionally adjusted to meet specific physical attribute requirements; the calculation and parameter storage cost in the training process can be effectively reduced through the neural network pruning technology.

Description

Technical Field

[0001] The present invention relates to the field of random medium reconstruction technology, and in particular to a random medium reconstruction method based on a generative adversarial network (ELGAN). Background Art

[0002] The microstructure of random media (such as porous rocks and composite materials) directly influences their macroscopic physical properties (such as permeability and thermal conductivity). Traditional reconstruction methods rely on numerical simulations (such as multi-point geostatistics and simulated annealing), which suffer from the following drawbacks: 1) High computational complexity: Traditional methods require iterative optimization of statistical features, resulting in low reconstruction efficiency; 2) Limited feature capture: Accurately reproducing complex pore topologies and multi-scale correlations is difficult; 3) Inadequate control over randomness: The generated results are prone to deviating from the physical properties of the real data (such as porosity distribution).

[0003] Although the generative adversarial network (GAN) based on deep learning has performed well in the field of image generation, it still faces challenges: 1) the training generation results are highly random and cannot effectively restore the physical properties of the input data; 2) the model has a large number of parameters, and the calculation and parameter storage costs are high.

[0004] As mentioned above, we designed a random medium reconstruction method based on generative adversarial network (ELGAN) to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the prior art and propose a random medium reconstruction method based on a generative adversarial network (ELGAN).

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A random medium reconstruction method based on generative adversarial network (ELGAN) includes the following steps:

[0008] Step S1, obtaining three-dimensional random medium image data;

[0009] Step S2, performing binarization or segmentation preprocessing on the image in step 1;

[0010] Step S3: Construct a deep learning network model based on a generative adversarial network architecture, including a generator, a discriminator, and a neural network pruning module; the generator is constructed based on InfoGAN, the discriminator is constructed based on a deep convolutional neural network, and the neural network pruning module is constructed based on LRP technology;

[0011] Step S4: training a generative adversarial network model to obtain a random medium reconstruction network model; wherein the training set of the generator is the input random medium image, and the training set of the discriminator is the real image and the generated image;

[0012] Step S5: take the random medium image as input, reconstruct the generative adversarial network model of random medium obtained in step 4, and output the reconstructed image.

[0013] Preferably, the data preprocessing includes:

[0014] The dataset was trimmed to 64*64*64 pixels and 64 images were used as input data. The porosity distribution of the 64 images was extracted.

[0015] Preferably, the neural network pruning technology adopts a layer-relevance propagation pruning framework (LRP), the specific principle of which is:

[0016] The optimal subset selection of the parameter space is completed by the following formula, removing redundant network units whose contribution to the target task is lower than the threshold ∈:

[0017]

[0018] θ masked =θ⊙m,m∈{0,1} d

[0019] Where m is a binary mask vector and ∈ is the preset accuracy tolerance threshold.

[0020] Preferably, the specific operation steps of the neural network pruning module include:

[0021] Step s1, forward propagation and activation recording: perform standard forward propagation and record the activation values of each layer and weight w ij ;

[0022] Step s2, back-propagation relevance score: starting from the output layer, assuming the predicted category is c, initialize the output layer relevance Other output units are 0; back propagation layer by layer: according to the LRP-α1β0 rule, the output layer correlation is distributed to each unit of the previous layer. For the convolution layer, the sum of all channel correlations of the same filter is used as the total importance of the filter; aggregate importance: the correlation scores calculated for each reference sample are averaged to obtain the global importance score;

[0023] Step s3, pruning and connection cleaning: (1) Sorting and pruning: Arrange all units in ascending order of importance scores and remove the units with the lowest ranking of r%; (2) Clean invalid connections: Delete the input / output connections connected to the pruned units to ensure the integrity of the network structure;

[0024] Step s4, fine-tuning: Use the pruned model to generate data.

[0025] Preferably, the generative adversarial network architecture uses InfoGAN’s prior information to guide data training:

[0026] A Q classifier is added to the discriminator to extract conditional information to ensure that the conditional information can constrain the generator. Similarly, when reconstructing 3D porous media images, to ensure that the reconstructed model has a similar pore structure to the input image, the porosity distribution of the 3D digital rock in the Z direction is extracted as conditional information to control the porosity of the synthetic model. The Wasserstein-1 distance based on InfoGAN is defined as the following problem with mutual information variational regularization and hyperparameters:

[0027]

[0028] Where c is the conditional parameter.

[0029] Preferably, the generative adversarial network architecture uses the gradient penalty term of WGAN-GP to stabilize training and optimizes the naturalness of the pore structure by combining multi-scale information; the specific loss function is:

[0030]

[0031] Among them, P data (x) is the probability distribution of the real sample, P z c represents the probability distribution corresponding to the samples generated by the generator G; Represents the mathematical expectation value of the discriminator D on the generated samples; Represents the mathematical expectation value of the discriminator D on the real sample; Indicates that in P data (x) and P z Random interpolation sampling is performed between c;

[0032] The loss function of the conditional information constraint part is:

[0033]

[0034] The Q classifier constrains the porosity distribution f(G(z,c)) of the generated samples to be consistent with the input condition information c through the loss of the above formula.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention proposes a random medium reconstruction method based on a generative adversarial network (ELGAN), which can effectively improve the controllability of the reconstruction results. Conditional data constrains the pore distribution, and the generated results can be directionally adjusted to meet specific physical property requirements.

[0037] 2. The present invention proposes a random medium reconstruction method based on generative adversarial network (ELGAN), which can effectively compress the computational and parameter storage costs during the training process through neural network pruning technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the pruning process of a random medium reconstruction method based on generative adversarial network (ELGAN) proposed in the present invention;

[0039] Figure 2 This is a schematic diagram of the network structure of a random medium reconstruction method based on generative adversarial network (ELGAN) proposed in the present invention;

[0040] Figure 3 This paper compares the reconstruction results of the Shale shale dataset using a random medium reconstruction method based on generative adversarial network (ELGAN) proposed in this invention. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0042] Reference Figure 1-Figure 3 , a random medium reconstruction method based on generative adversarial network (ELGAN), comprising the following steps:

[0043] Step S1, obtaining three-dimensional random medium image data;

[0044] Step S2, performing binarization or segmentation preprocessing on the image in step 1;

[0045] Step S3: construct a deep learning network model based on a generative adversarial network architecture, including a generator, a discriminator, and an LRP neural network pruning module; the generator is constructed based on InfoGAN; the discriminator is constructed based on a deep convolutional neural network; and the neural network pruning module is constructed based on LRP technology;

[0046] The structure and parameter settings of the generator are shown in the following table:

[0047]

[0048]

[0049] Continued:

[0050] stage Function parameter b ApplyNoise AdaIN in_channel=64,dlatent_size=512 Conv3d in_channel=64,out_channel=64,kernel_size ApplyNoise AdaIN in_channel=64,dlatent_size=512 Phase 5 Upsample factor=2 ApplyNoise AdaIN in_channel=64,dlatent_size=512 Conv3d in_channel=64,out_channel=64,kernel_size ApplyNoise AdaIN in_channel=64,dlatent_size=512 Phase 6 Upsample factor=2 ApplyNoise AdaIN in_channel=64,dlatent_size=512 Conv3d in_channel=64,out_channel=64,kernel_size ApplyNoise AdaIN in_channel=32,dlatent_size=512 Tanh

[0051] The structure and parameter settings of the discriminator are shown in the following table:

[0052]

[0053]

[0054] Continued:

[0055] Number of layers Function parameter LeakyReLU Layer 6 Conv3d in_channel=64,out_channel=1,kernel_size= Linear in_features=512,out_features=1

[0056] Step S4: training a generative adversarial network model to obtain a random medium reconstruction network model; wherein the training set of the generator is the input random medium image, and the training set of the discriminator is the real image and the generated image;

[0057] Step S5: Take the random medium image as input, reconstruct the random medium generative adversarial network model obtained in step S4, and output the reconstructed image.

[0058] Technical Solution 2 based on Technical Solution 1: The generative adversarial network architecture uses the gradient penalty term of WGAN-GP for stable training and combines multi-scale information to optimize the naturalness of the pore structure; the specific loss function is:

[0059]

[0060] Among them, P data (x) is the probability distribution of the real sample, P z c represents the probability distribution corresponding to the samples generated by the generator G; Represents the mathematical expectation value of the discriminator D on the generated samples; Represents the mathematical expectation value of the discriminator D on the real sample; Indicates that in P data (x) and P z Random interpolation sampling is performed between c;

[0061] The loss function of the conditional information constraint part is:

[0062]

[0063] The Q classifier constrains the porosity distribution f(G(z,c)) of the generated samples to be consistent with the input condition information c through the loss of the above formula.

[0064] Technical Solution 3 based on Technical Solution 2: The neural network pruning technology adopts a layer-relevance propagation pruning framework (LRP), and the specific principle is:

[0065] The optimal subset selection of the parameter space is completed by the following formula, removing redundant network units whose contribution to the target task is lower than the threshold ∈:

[0066]

[0067] θ masked =θ⊙m,m∈{0,1} d

[0068] Where m is a binary mask vector, ∈ is a preset accuracy tolerance threshold; the essence of this optimization problem is to construct a parameter importance measurement function The accuracy of its mathematical representation directly determines the performance retention ability of the pruned model.

[0069] Technical Solution 4 based on Technical Solution 3: The core steps of the neural network pruning module in the generative adversarial network model are:

[0070] Step s1: propagate forward and activate records

[0071] Perform standard forward propagation and record the activation values of each layer and weight w ij ;

[0072] Step s2: Back-propagation of relevance scores

[0073] Initialization: Starting from the output layer, assuming the predicted category is c, initialize the output layer correlation Other output units are 0; back propagation layer by layer: according to the LRP-α1β0 rule, the output layer correlation is distributed to each unit of the previous layer. For the convolution layer, the sum of all channel correlations of the same filter is used as the total importance of the filter; aggregate importance: the correlation scores calculated for each reference sample are averaged to obtain the global importance score;

[0074] Step s3: Pruning and connection cleaning

[0075] Sorting and pruning: Arrange all units in ascending order of importance scores and remove the lowest r% of units; Clean invalid connections: Delete the input / output connections connected to the pruned units to ensure the integrity of the network structure;

[0076] Step s4: Fine-tuning

[0077] Use the pruned model to generate data.

[0078] Technical Solution 5 based on Technical Solution 4: The generative adversarial network architecture uses InfoGAN's prior information to guide data training:

[0079] A Q classifier is added to the discriminator to extract conditional information to ensure that the conditional information can constrain the generator. Similarly, when reconstructing 3D porous media images, to ensure that the reconstructed model has a similar pore structure to the input image, the porosity distribution of the 3D digital rock in the Z direction is extracted as conditional information to control the porosity of the synthetic model. The Wasserstein-1 distance based on InfoGAN is defined as the following problem with mutual information variational regularization and hyperparameters:

[0080]

[0081] Where c is the conditional parameter.

[0082] like Figure 1 As shown in FIG, the working process of the neural network pruning module in the generative adversarial network model is as follows:

[0083] In the first stage, the network performs standard forward propagation and collects activation values of each layer; in the second stage, the network output score f(x) is back-propagated to each layer according to the LRP propagation rule; in the third stage, the current model is pruned by removing units that are not related to the correlation measure R generated by LRP, and further fine-tuning is optional; LRP is based on the inter-layer conservation criterion to ensure that the propagation quantity (such as the correlation of the predicted category) is conserved between neurons in adjacent layers.

[0084] Let Ri(l) be the correlation of neuron i in layer l, and Rj(l+1) be the correlation of neuron j in layer l+1. If we impose a stricter definition on conservation, we can further constrain the local redistribution of correlation in the bottom layer, and then define for The share of neurons that are redistributed to the lower layer i. This conservation property always satisfies:

[0085]

[0086] The summation range is all neurons in the previous layer l i When correlation is used as a pruning criterion, this property ensures that its magnitude is conserved layer by layer, regardless of changes in the hidden layer dimension or the number of neurons during iterative pruning. In layer l, the nodes i The global importance of Direct quantification.

[0087] In this example, the Shale shale dataset was selected as the benchmark test set (TI). This example also selected two methods, the Deep Convolutional Generative Adversarial Network (DCGAN) and the Single Normal Equation Simulation (SNESIM), as comparison benchmarks. The DCGAN effectively improves the instability of GAN training by introducing convolutional networks into the GAN framework and shares the same multi-image joint training paradigm as ELGAN. The SNESIM, a classic method for reconstructing traditional random media, has been widely used in the field of random media modeling.

[0088] After training, this embodiment obtains the corresponding random medium image reconstruction network model.

[0089] Reference Figure 3 , showing the outer surface and cross-sectional views (X=32, Y=32, Z=32) of this embodiment and other reconstruction methods. In the intuitive diagram, the skeleton is represented by black, and the pore space is represented by gray. The pore space in the Shale image is a clear crack running through the cube. By comparison, it can be observed that the pore space generated by the SNESIM method is irregular in shape and does not generate cracks, which is significantly different from the TI. The reconstruction results generated by DCGAN and ELGAN show clear crack spaces and are closer to the TI.

[0090] The table below lists the total porosity, connected porosity, and isolated porosity averages for the input data and the 10 reconstructions. Comparing the porosities in the table, we can see that the ELGAN data is much closer to the input data.

[0091] type TI SNESIM DCGAN ELGAN Total porosity (%) 21.33 28.63 23.32 20.51 Connected porosity (%) 21.25 28.32 23.15 20.46 Isolated porosity (%) 0.08 0.31 0.17 0.05

[0092] Furthermore, to compare the resource consumption of different reconstruction methods, we compared the average time and resource consumption of each method during training on the dataset. The following table lists the GPU / CPU utilization, memory usage, GPU memory consumption, and reconstruction time for ten reconstructions using each method.

[0093] SNESIM DCGAN ELGAN MaximalCPUusage 61％ 45％ 33％ MaximalGPUusage 0 99％ 99％ Maximalmemoryusage 60％ 37％ 29％ MaximalGPUgraphicmemory 0 9178 8556 Time taken for the initial 1012 27242 24832 Timefortheother9 9294 2 2

[0094] Compared to traditional SNESIM methods, DCGAN and ELGAN take longer to reconstruct initially, but their subsequent nine reconstructions are significantly faster due to the reuse of the model and parameters established during the initial training phase. In contrast, traditional SNESIM cannot use pre-trained parameters in subsequent reconstructions, so resource consumption increases significantly with the number of reconstructions. Shale data is characterized by a distinct crack pore in the model. This allows ELGAN training to focus on the crack shape, reducing the resources used for the larger solid skeleton on either side. As a result, ELGAN demonstrates lower resource usage.

[0095] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art will appreciate that the present invention is not limited to the foregoing embodiments and that various modifications and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such modifications and improvements are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A random medium reconstruction method based on generative adversarial network (ELGAN), characterized by: The following steps are included: Step S1, obtaining three-dimensional random medium image data; Step S2, performing binarization or segmentation preprocessing on the image in step 1; Step S3: Construct a deep learning network model based on a generative adversarial network architecture, including a generator, a discriminator, and a neural network pruning module; the generator is constructed based on InfoGAN, the discriminator is constructed based on a deep convolutional neural network, and the neural network pruning module is constructed based on LRP technology; Step S4: training a generative adversarial network model to obtain a random medium reconstruction network model; wherein the training set of the generator is the input random medium image, and the training set of the discriminator is the real image and the generated image; Step S5: take the random medium image as input, reconstruct the generative adversarial network model of random medium obtained in step 4, and output the reconstructed image.

2. The random medium reconstruction method based on generative adversarial network (ELGAN) according to claim 1, characterized in that: The data preprocessing includes: The dataset was trimmed to 64*64*64 pixels and 64 images were used as input data. The porosity distribution of the 64 images was extracted.

3. The random medium reconstruction method based on generative adversarial network (ELGAN) according to claim 1, characterized in that: The neural network pruning technology adopts the layer-relevance propagation pruning framework (LRP), and the specific principle is: The optimal subset selection of the parameter space is completed by the following formula, removing redundant network units whose contribution to the target task is lower than the threshold ∈: i masked =θ⊙m,m∈{0,1} d Where m is a binary mask vector and ∈ is the preset accuracy tolerance threshold.

4. The random medium reconstruction method based on generative adversarial network (ELGAN) according to claim 1, characterized in that: The specific operation steps of the neural network pruning module include: Step s1, forward propagation and activation recording: perform standard forward propagation and record the activation values of each layer and weight w ij ; Step s2, back-propagation relevance score: starting from the output layer, assuming the predicted category is c, initialize the output layer relevance Other output units are 0; back propagation layer by layer: according to the LRP-α1β0 rule, the output layer correlation is distributed to each unit of the previous layer. For the convolution layer, the sum of all channel correlations of the same filter is used as the total importance of the filter; aggregate importance: the correlation scores calculated for each reference sample are averaged to obtain the global importance score; Step s3, pruning and connection cleaning: (1) Sorting and pruning: Arrange all units in ascending order of importance scores and remove the units with the lowest ranking of r%; (2) Clean invalid connections: Delete the input / output connections connected to the pruned units to ensure the integrity of the network structure; Step s4, fine-tuning: Use the pruned model to generate data.

5. The random medium reconstruction method based on generative adversarial network (ELGAN) according to claim 1, characterized in that: The generative adversarial network architecture uses the prior information of InfoGAN to guide data training: A Q classifier is added to the discriminator to extract conditional information to ensure that the conditional information can constrain the generator; Similarly, when reconstructing 3D porous media images, to ensure that the reconstructed model has a similar pore structure to the input image, the porosity distribution of the 3D digital rock in the Z direction is extracted as conditional information to control the porosity of the synthetic model. The Wasserstein-1 distance based on InfoGAN is defined as the following problem with mutual information variational regularization and hyperparameters: Where c is a conditional parameter; V(D,G) is the adversarial loss term of the Wasserstein-1 distance; λ is a hyperparameter used to balance the adversarial loss; z is a noise vector, and c is a structured latent variable; where I(C; G(z,c)) forces the generator to use the latent variable c to generate controllable and diverse samples.

6. The random medium reconstruction method based on generative adversarial network (ELGAN) according to claim 1, characterized in that: The generative adversarial network architecture uses the gradient penalty term of WGAN-GP to stabilize training and combines multi-scale information to optimize the naturalness of the pore structure; the specific loss function is: Among them, P data (x) is the probability distribution of the real sample, P z c represents the probability distribution corresponding to the samples generated by the generator G; Represents the mathematical expectation value of the discriminator D on the generated samples; Represents the mathematical expectation value of the discriminator D on the real sample; Indicates that in P data (x) and P z Random interpolation sampling is performed between c; The loss function of the conditional information constraint part is: Where G(z,c) is the sample generated by the generator G through the input noise z and the latent variable c; f(·) is the function for calculating the porosity sequence; f(G(z,c)) is the reconstructed porosity sequence; The Q classifier constrains the porosity distribution f(G(z,c)) of the generated samples to be consistent with the input condition information c through the loss of the above formula.

Images