A core 3D reconstruction method based on KAN convolutional network

Abstract

The present invention discloses a core three-dimensional reconstruction method based on a KAN convolutional network, comprising the following steps: S1: pre-processing a core three-dimensional image to form a sample set for model training; S2: using a 3D model to train the sample set; S3: extracting features from the results of the preliminary upsampling in sequence, and using a Tanh activation function for nonlinear transformation; S4: the discriminator of the 3D model performs true and false judgment on the generated core image through multiple three-dimensional convolutional layers; S5: performing model training, using a random gradient descent method for back propagation, and jointly optimizing and updating the network parameters of the generator and the discriminator; S6: after the training is completed, using the generator to generate a new core three-dimensional image, and using porosity and permeability as evaluation indicators to evaluate the reconstruction effect of the model. The present invention can effectively reconstruct the three-dimensional structure inside the core, and provides an efficient auxiliary means for studying the flow of fluids inside the core pores.

Description

Technical Field

[0001] The invention relates to the technical field of petroleum exploration and development, and in particular to a core three-dimensional reconstruction method based on a KAN convolutional network. Background Art

[0002] The core pore structure has a huge impact on the flow of fluid in the core pores. Observing the pore structure inside the core and analyzing the seepage law of the fluid are of great significance to the exploration and development and cost reduction and efficiency improvement of the oil and gas industry.

[0003] Traditional core 3D reconstruction methods are mainly based on physical experimental methods and numerical reconstruction methods. The former uses high-end optical and scanning instruments to obtain two-dimensional slice images of the core and construct digital cores, while the latter constructs digital cores through mathematical modeling and other methods, which can basically reflect the topological structure and geometric characteristics of the real core pore space. However, this type of method still has many problems. For example, in some cases where core samples are difficult to obtain, traditional 3D reconstruction methods cannot reconstruct the 3D core structure with a small amount of core image information. Summary of the invention

[0004] The present invention aims to provide a core three-dimensional reconstruction method based on a KAN convolutional network to solve the technical problem of how to effectively reconstruct the three-dimensional structure inside the core.

[0005] The present invention is implemented by adopting the following technical scheme: a core three-dimensional reconstruction method based on KAN convolutional network, comprising the following steps:

[0006] S1: Preprocess the core 3D image to form a sample set for model training;

[0007] S2: Use the 3D model to train the sample set. The generator of the 3D model takes random noise as input, performs preliminary upsampling on the input through the deconvolution layer, and performs batch normalization on the upsampling result. Finally, nonlinear features are introduced through the ReLU activation function.

[0008] S3: Extract features from the preliminary upsampling results in turn, and use the Tanh activation function for nonlinear transformation to finally generate a three-dimensional core image;

[0009] S4: The discriminator of the 3D model judges the authenticity of the generated core image through multiple 3D convolutional layers;

[0010] S5: Perform model training, use stochastic gradient descent for back propagation, and jointly optimize and update the network parameters of the generator and discriminator;

[0011] S6: After training, the generator is used to generate new core 3D images, and the porosity and permeability are used as evaluation indicators to evaluate the reconstruction effect of the model.

[0012] Furthermore, step S1 includes the following sub-steps:

[0013] S11: Read the core image files and check the file formats one by one;

[0014] S12: setting the cutting size, defining the pixel size of the small cube in the width, height and depth directions respectively, and cutting out a plurality of cube sub-regions for training from the core image according to the size;

[0015] S13: Save each sub-region sample in HDF5 format.

[0016] Furthermore, step S2 includes the following sub-steps:

[0017] S21: training the sample set by a KAN-DCGAN 3D model, wherein the KAN-DCGAN 3D model uses a KAN convolutional network and a three-dimensional generative adversarial network to generate a three-dimensional core image by adversarial learning between a generator and a discriminator;

[0018] S22: The generator takes random noise as input, where the random noise follows a normal distribution;

[0019] S23: The generator first uses a deconvolution layer to upsample the input noise, then normalizes the output of the deconvolution layer, and finally uses the ReLU activation function to introduce nonlinearity to help the generator capture complex features;

[0020] S24: Generator composite convolution operation and normalization to refine the image generation process;

[0021] S25: In the final stage of the generator, the feature map is converted into the final output image using a deconvolution layer and a nonlinear transformation is performed through the Tanh activation function to ensure that the pixel value range of the generated image is between [-1, 1].

[0022] Furthermore, step S3 includes the following sub-steps:

[0023] S31: The result of the preliminary upsampling is passed through multiple ReLUKANConv 3D layers in sequence for feature extraction, and the input dimension is [c, h, w, l], where c is the number of channels, h, w, and l are the height, width, and depth of the image respectively;

[0024] S32: Initialize the convolution kernel weights of base_conv and relukan_conv;

[0025] S33: The input is preliminarily convolved through base_conv to obtain the basic feature basis;

[0026] S34: The input passes through the feature generation module relukan_conv for phase mapping, phase activation and convolution calculation to further extract multi-scale features;

[0027] S35: Add the feature y generated by the relukan_conv layer to the basic feature basis calculated by the base_conv layer, and apply the regularization layer layer_norm to the addition result to obtain a normalized output , the specific calculation formula is as follows:

[0028] ;

[0029] in, and Characteristics The mean and variance of and are learnable scaling and translation parameters, is a small constant;

[0030] S36: Normalize the output Nonlinear activation is performed through the activation function SiLU to generate the final output features.

[0031] Further, step S4 includes the following sub-steps:

[0032] S41: The input single-channel 3D image is expanded to 16 channels using 3D convolution, and then LeakyReLU is used for activation;

[0033] S42: Three layers of 3D convolution are used consecutively, each followed by BatchNorm3d and LeakyReLU activation;

[0034] S43: Use the last layer of 3D convolution to reduce the number of channels to 1, and use the Sigmoid activation function to limit the output to [0, 1], indicating the true or false judgment score of the sample.

[0035] Further, step S5 includes the following sub-steps:

[0036] S51: Update the network parameters of the discriminator;

[0037] S52: Update the network parameters of the generator;

[0038] S53: Repeat steps S51 and S52 until the parameters of the generator and the discriminator converge; in continuous iteration and optimization, the samples generated by the generator gradually approach the distribution of real samples, and the discriminator more accurately identifies real and generated samples.

[0039] Further, step S51 includes the following sub-steps:

[0040] S511: Input random noise and real samples , the random noise Input the generator and get the generated sample ;

[0041] S512: Calculate the real sample Loss ;

[0042] S513: Calculate and generate samples Loss ;

[0043] S514: Calculate the total loss of the discriminator for ;

[0044] S515: Update the discriminator parameters using the Adam optimizer , specifically:

[0045] ;

[0046] in, is the learning rate, is the loss gradient of the discriminator.

[0047] Further, step S52 includes the following sub-steps:

[0048] S521: Random noise Input generator to get generated samples ;

[0049] S522: Input the false sample into the discriminator to obtain the output , represents the score of the discriminator on the generated samples;

[0050] S523: Setting the target label of the generator to a range of randomly uniformly distributed values;

[0051] S524: Computing generator loss using binary cross entropy loss , specifically:

[0052] ;in, is the target label;

[0053] S525: Loss to the generator Perform backpropagation and update the generator parameters using the Adam optimizer , specifically:

[0054] ;

[0055] in, is the learning rate, is the loss gradient of the generator.

[0056] Further, step S6 includes the following sub-steps:

[0057] S61: Generate three-dimensional image samples of new cores using the trained generator;

[0058] S62: Calculate the porosity and permeability of the core structure based on the generated three-dimensional image. The porosity represents the ratio of the pore volume to the total volume in the core image. The specific calculation method is:

[0059] ;

[0060] in, represents the pore volume in a three-dimensional image, is the total volume of the core image;

[0061] Permeability measures the ability of fluid flow in the generated core, and the specific calculation expression is:

[0062] ;

[0063] in, is the fluid flow rate, is the fluid viscosity, is the core length, is the flow area, is the pressure difference;

[0064] S63: The porosity and permeability of the generated images are compared with the corresponding indicators of the real core samples to evaluate the effectiveness and accuracy of the model in reconstructing the core structure.

[0065] The beneficial effect of the present invention is that: the present invention combines the generative adversarial network model and introduces a generator with a ReLUKANConv3D layer to achieve effective reconstruction of the three-dimensional structure of the core. This method can more effectively capture the complex pore structure inside the core, improve the authenticity and detail retention of the generated image, and thus achieve better reconstruction effects on key indicators such as porosity and permeability. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0067] Figure 1 It is a flow chart of the present invention;

[0068] Figure 2 A schematic diagram of the network structure. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, 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 part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0070] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0071] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0072] See also Figure 1 , Figure 2 , a core 3D reconstruction method based on KAN convolutional network, comprising the following steps:

[0073] S1: Preprocess the core 3D image, extract small cubes from the core image at regular intervals as training data, and form a sample set for model training;

[0074] S2: The KAN-DCGAN 3D model is used to train the sample set. The generator G of this model takes random noise as input, first performs preliminary upsampling of the input through the deconvolution layer, then performs batch normalization on the upsampling result, and finally introduces nonlinear features through the ReLU activation function;

[0075] S3: The result of the preliminary upsampling is sequentially passed through multiple ReLUKANConv3D layers for feature extraction, and then the feature map is converted into an output image through a layer of deconvolution, and a nonlinear transformation is performed using the Tanh activation function to finally generate a three-dimensional core image;

[0076] S4: The discriminator D of the model judges the authenticity of the generated core image through multiple three-dimensional convolutional layers;

[0077] S5: Perform model training, use stochastic gradient descent method for back propagation, and jointly optimize and update the network parameters of the generator G and the discriminator D;

[0078] S6: After training, the generator G is used to generate new core 3D images, and the porosity and permeability are used as evaluation indicators to evaluate the reconstruction effect of the model.

[0079] In this embodiment, step S1 includes the following sub-steps:

[0080] S11: read the core image files and check the file formats one by one, and keep the image files in .tiff or .tif format;

[0081] S12: Set cutting size , respectively define the pixel size of the small cube in the width, height and depth directions, and cut out multiple cube sub-regions for training from the core image according to the size;

[0082] S13: Save each sub-region sample in HDF5 format and use gzip compression to optimize storage efficiency.

[0083] In this embodiment, step S2 includes the following sub-steps:

[0084] S21: The core of the KAN convolutional network is to implement the convolution kernel operation through a set of unary nonlinear functions to improve the nonlinear expression ability of feature extraction. The specific definition is as follows:

[0085] S211: Input 3D image data ,in is the number of channels, , and Represent the depth, height and width of the image respectively, and set the convolution kernel size to ;

[0086] S212: For each pixel position , perform the following convolution operation:

[0087]

[0088] Among them, each Each is a univariate nonlinear learnable function with independent trainable parameters that processes the input data.

[0089] S213: Use the following nonlinear function form to define each :

[0090]

[0091]

[0092]

[0093] in, and are trainable weights, is a nonlinear term, usually using a spline function or other nonlinear basis functions. Spline(x) is a cubic spline function. It is the activation function, and SiLU function is usually used.

[0094] S214: Output the generated 3D feature map .

[0095] S22: The core of the KAN-DCGAN3D model is the KAN convolutional network and the 3D generative adversarial network, which generates 3D core images through adversarial learning between the generator and the discriminator. The specific process is as follows:

[0096] S221: The generator receives a random noise , the vector follows a normal distribution and is initially upsampled by a deconvolution layer, batch normalization, and ReLU activation function; then, multiple ReLUKANConv3D layers are used to gradually generate the complex three-dimensional structural features of the core by stacking these layers; finally, a deconvolution layer is used to convert the feature map into the final core image, and the Tanh activation function is used to limit the pixel values ​​to the range of [-1, 1];

[0097] S222: The discriminator receives a real or generated three-dimensional core image, and extracts high-order features of the input image layer by layer through multiple three-dimensional convolutional layers, and finally outputs a probability value indicating the probability that the input image is real or generated;

[0098] S223: Using the binary cross entropy loss function, the generator and discriminator are trained based on the real labels and generated labels respectively. The generator tries to maximize the error rate of the discriminator, while the discriminator tries to correctly distinguish between real and generated images. The model parameters are updated using the stochastic gradient descent method, and the generator and discriminator are gradually optimized through back propagation until the generator can generate realistic core images;

[0099] S224: The trained KAN-DCGAN3D model can be used to generate high-quality 3D core reconstruction images.

[0100] S23: Initialize the weights of the generator G. For the convolutional layer, the weights are set to normal distribution random numbers with a mean of 0 and a standard deviation of 0.02. For the batch normalization layer, the weights are initialized to normal distribution random numbers with a mean of 1 and a standard deviation of 0.02.

[0101] S24: Generator G generates random noise As input, It follows a normal distribution (mean 0, standard deviation 1).

[0102] S25: The generator G first uses a deconvolution layer to filter the noise of the input Upsampling is performed, and then the output of the deconvolution layer is normalized. Finally, the ReLU activation function is used to introduce nonlinearity to help the generator capture complex features.

[0103] S26: The core of the generator G is the ReLUKANConv3D layer, each of which refines the image generation process through composite convolution operations and normalization.

[0104] S27: In the final stage of the generator, the feature map is converted into the final output image using a deconvolution layer and a nonlinear transformation is performed through the Tanh activation function to ensure that the pixel value range of the generated image is between [-1, 1].

[0105] In this embodiment, step S3 includes the following sub-steps:

[0106] S31: ReLUKANConv3D is a 3D convolutional layer based on Kolmogorov-Arnold convolution with the following submodules:

[0107] S311: Initialize the convolution kernel weights of base_conv and relukan_conv so that they are evenly distributed within a reasonable range;

[0108] S312: Perform preliminary feature extraction on the input through the base_conv layer to obtain the basic feature map. Use the SiLU activation function to improve the nonlinear expression ability and extract the basic structural features in the input data;

[0109] S313: After the input is further extracted with multi-scale features by the relukan_conv module, it is added to the basic feature map generated by the base_conv layer. The feature extraction process of the relukan_conv module includes phase mapping, phase activation, and phase constraint output operations: First, the activation interval of each channel of the input is determined by phase mapping, and the multi-scale features are captured using learnable phase lower and upper bound parameters; then, the input features are activated within the phase interval to enhance the nonlinear expression capability; finally, the result of the phase activation is combined with the scaling factor to form a phase constraint output, and a convolution operation is performed to capture more local features;

[0110] S314: The multi-scale features extracted by relukan_conv are fused with the basic features of base_conv, processed by the regularization layer to ensure the stability of the features, and finally the final output features are generated through the SiLU activation function.

[0111] S32: relukan_conv is a convolution module for 3D feature extraction. It captures the multi-scale information of input features through phase mapping and phase activation mechanism, and enhances the nonlinear expression ability. It has the following steps:

[0112] S321: The relukan_conv module first performs phase mapping on the input features. For each channel, two learnable phase parameters, the lower and upper bounds, are defined, and the initial range is determined by linear initialization. The purpose of phase mapping is to determine the activation interval of each channel in order to better capture the multi-scale information of the feature within a specific range;

[0113] S322: performing a phase activation operation on the input feature within the interval determined by the phase mapping, and capturing local nonlinear features by calculating the distance between the input feature and the lower bound and the upper bound of the phase;

[0114] S323: The phase activation result is combined with the scaling factor to generate a phase-constrained output to control the activation strength. The phase-constrained output is then squared and flattened, and then convolved to extract multi-scale features.

[0115] S33: Random Noise After processing in step S25, the input of the ReLUKANConv3D layer is generated , whose dimensions are [c,h,w,l], where c is the number of channels, h, w, and l are the height, width, and depth of the image respectively;

[0116] S34: Initialize the convolution kernel weights of base_conv and relukan_conv. The specific process is as follows:

[0117] S341: Calculate the initialization range of convolution kernel weights , the calculation formula used is:

[0118]

[0119] in, Represents the number of input channels multiplied by the size of the convolution kernel. The three-dimensional convolution kernel is This limit defines the random range of initial weights. , where the weights will be evenly distributed within this range.

[0120] S342: Assign each weight of the base_conv layer and the relukan_conv layer to a value in the interval A uniformly distributed random value in .

[0121] S343: Set the nonlinear parameters to linear to ensure that the initialization range of the weights adapts to the linear activation function, and takes into account the stability of the convolution layer signal in the early training and the uniformity of feature transfer.

[0122] S35: Input The basic feature basis is obtained by performing preliminary convolution calculation through base_conv. The specific process is as follows:

[0123] S351: Input tensor Perform non-offline activation. Use the activation function SiLU to act on , mapping it to an activated tensor:

[0124]

[0125] Among them, the SiLU function introduces nonlinearity, which improves the network's ability to express complex features;

[0126] S352: After activation by base_conv Perform convolution operations to extract basic features of the input . Use the convolution kernel weights predefined in step S34 For each pixel position Perform the following convolution calculation:

[0127]

[0128] in For the input tensor The number of channels, is the convolution kernel size, is a 3D tensor , used for subsequent multi-scale feature combination and reconstruction steps.

[0129] S36: Input Phase mapping, phase activation and convolution calculation are performed through the feature generation module relukan_conv to further extract multi-scale features. The specific process is as follows:

[0130] S361: Input Each channel Phase mapping is performed to determine the activated phase interval. Specifically, for channel Defining the lower phase limit and phase high limit Two learnable phase parameters, which are initialized by linear functions, are expressed as follows:

[0131]

[0132]

[0133] in is the number of phase activations, Control parameters for multi-scale features.

[0134] S362: Based on input The phase interval is phase activated to capture the local nonlinearity of the input features. The specific calculation expression is:

[0135]

[0136]

[0137] in, and Respectively represent the distance between the input feature and the lower bound and upper bound of the phase.

[0138] S363: Combine phase activation output to generate phase constraint output , the calculation formula is:

[0139]

[0140] in, is a scaling factor that controls the phase activation strength;

[0141] S364: Output the phase constraint Perform a square operation, that is , to enhance the nonlinear expression ability of the features. Flatten and perform convolution operations to generate the final multi-scale features , the specific expression is as follows:

[0142]

[0143] in, As a trainable nonlinear function, SiLU is used for activation and local nonlinear feature capture.

[0144] S37: The features generated by the relukan_conv layer The basic features calculated with the base_conv layer Add and apply the regularization layer layer_norm to the addition result to obtain the normalized output , the specific calculation formula is as follows:

[0145]

[0146] in, and Characteristics The mean and variance of and are learnable scaling and translation parameters, is a small constant.

[0147] S38: Normalize the output The activation function SiLU is used for nonlinear activation to generate the final output features. .

[0148] In this embodiment, step S4 includes the following sub-steps:

[0149] S41: The input single-channel 3D image is expanded to 16 channels using 3D convolution. The convolution kernel is , the stride is , then use LeakyReLU activation;

[0150] S42: Three layers of 3D convolution are used in succession, with the number of channels increased to 32, 64, and 128 in each layer. The convolution kernel is , the stride is , each convolution layer is followed by BatchNorm3d and LeakyReLU activation;

[0151] S43: Use the last layer of 3D convolution to reduce the number of channels to 1, and the convolution kernel is , the stride is , and the output is limited to [0, 1] through the Sigmoid activation function, indicating the true or false judgment score of the sample.

[0152] In this embodiment, step S5 includes the following sub-steps:

[0153] S51: Update the network parameters of the discriminator D:

[0154] S511: Input random noise and real samples , random noise Input generator G to get generated samples ;

[0155] S512: Calculate the real sample Loss , specific steps:

[0156] S5121: Real sample Set the label range to random uniform distribution The value of , through slight perturbation, avoids the discriminator from being too dependent on the real sample label and improves the robustness of the model;

[0157] S5122: The real sample In the input discriminator D, the binary cross entropy loss is used to calculate the loss, specifically:

[0158]

[0159] in Represents the perturbation label value set for each real sample, is the output of the discriminator for the real sample; S5123: for the loss Perform backpropagation to update the discriminator parameters.

[0160] S513: Calculate and generate samples Loss , specific steps:

[0161] S5131: Generate sample Set the label range to random uniform distribution The value of

[0162] S5132: Sample will be generated In the input discriminator D, the binary cross entropy loss is used to calculate the loss, specifically:

[0163]

[0164] in, represents the perturbation label value set for each generated sample, is the output of the discriminator for the generated sample;

[0165] S5123: For loss Perform backpropagation to update the discriminator parameters.

[0166] S514: Calculate the total loss of the discriminator for ;

[0167] S515: Update the discriminator parameters using the Adam optimizer , specifically:

[0168]

[0169] in, is the learning rate, is the loss gradient of the discriminator.

[0170] S52: Update the network parameters of the generator G:

[0171] S521: Random noise Input generator G to get generated samples ;

[0172] S522: Input the false sample into the discriminator D to obtain the output , represents the score of the discriminator on the generated samples;

[0173] S523: Set the target label of the generator Set the range to random uniform distribution The value of

[0174] S524: Calculate the generator loss using binary cross entropy loss , specifically:

[0175] ;in, is the target label;

[0176] S525: Loss to the generator Perform backpropagation and update the generator parameters using the Adam optimizer , specifically:

[0177]

[0178] in, is the learning rate, is the loss gradient of the generator.

[0179] S53: Repeat steps S51 and S52 until the generator parameters and the discriminator parameters Convergence. In continuous iteration and optimization, the samples generated by the generator gradually approach the distribution of real samples, and the discriminator more accurately identifies real and generated samples.

[0180] In this embodiment, step S6 includes the following sub-steps:

[0181] S61: Generate a three-dimensional image sample of a new core using the trained generator;

[0182] S62: Calculate the porosity and permeability of the core structure based on the generated three-dimensional image. Porosity represents the ratio of the pore volume to the total volume in the core image, and the specific calculation method is:

[0183]

[0184] in, represents the pore volume in a three-dimensional image, is the total volume of the core image.

[0185] Permeability measures the ability of fluid flow in the generated core, and the specific calculation expression is:

[0186]

[0187] in, is the fluid flow rate, is the fluid viscosity, is the core length, is the flow area, is the pressure difference;

[0188] S63: The porosity and permeability of the generated images are compared with the corresponding indicators of the real core samples to evaluate the effectiveness and accuracy of the model in reconstructing the core structure.

[0189] The present invention combines the generative adversarial network model and introduces a generator with the ReLUKANConv3D layer to achieve effective reconstruction of the three-dimensional structure of the core. This method can more effectively capture the complex pore structure inside the core, improve the authenticity and detail retention of the generated image, and thus achieve better reconstruction effects on key indicators such as porosity and permeability.

[0190] For the aforementioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily required by the present application.

[0191] The above embodiments describe the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the changes and modifications made by those skilled in the art shall be within the scope of protection of the appended claims of the present invention without departing from the spirit and scope of the present invention.

Claims

1. A core 3D reconstruction method based on KAN convolutional network, characterized in that: The steps include: S1: Preprocess the core 3D image to form a sample set for model training; S2: Use the 3D model to train the sample set. The generator of the 3D model takes random noise as input, performs preliminary upsampling on the input through the deconvolution layer, and performs batch normalization on the upsampling result. Finally, nonlinear features are introduced through the ReLU activation function. Step S2 includes the following sub-steps: S21: training the sample set by a KAN-DCGAN 3D model, wherein the KAN-DCGAN 3D model uses a KAN convolutional network and a three-dimensional generative adversarial network to generate a three-dimensional core image by adversarial learning between a generator and a discriminator; S22: The generator takes random noise as input, where the random noise follows a normal distribution; S23: The generator first uses a deconvolution layer to upsample the input noise, then normalizes the output of the deconvolution layer, and finally uses the ReLU activation function to introduce nonlinearity to help the generator capture complex features; S24: Generator composite convolution operation and normalization to refine the image generation process; S25: In the final stage of the generator, the feature map is converted into the final output image using a deconvolution layer, and a nonlinear transformation is performed through the Tanh activation function to ensure that the pixel value range of the generated image is between [-1, 1]; S3: Extract features from the preliminary upsampling results in sequence, and use the Tanh activation function for nonlinear transformation to finally generate a three-dimensional core image; step S3 includes the following sub-steps: S31: The result of the preliminary upsampling is passed through multiple ReLUKANConv 3D layers in sequence for feature extraction, and the input dimension is [c, h, w, l], where c is the number of channels, h, w, and l are the height, width, and depth of the image respectively; S32: Initialize the convolution kernel weights of base_conv and relukan_conv; S33: The input is preliminarily convolved through base_conv to obtain the basic feature basis; S34: The input passes through the feature generation module relukan_conv for phase mapping, phase activation and convolution calculation to further extract multi-scale features; S35: The features generated by the relukan_conv layer The basic features calculated with the base_conv layer Add and apply the regularization layer layer_norm to the addition result to obtain the normalized output , the specific calculation formula is as follows: ； in, and Characteristics The mean and variance of and are learnable scaling and translation parameters, is a small constant; S36: Normalize the output Nonlinear activation is performed through the activation function SiLU to generate the final output features; S4: The discriminator of the 3D model judges the authenticity of the generated core image through multiple 3D convolutional layers; S5: Perform model training, use stochastic gradient descent for back propagation, and jointly optimize and update the network parameters of the generator and discriminator; S6: After training, the generator is used to generate new core 3D images, and the porosity and permeability are used as evaluation indicators to evaluate the reconstruction effect of the model.

2. A core three-dimensional reconstruction method based on a KAN convolutional network as claimed in claim 1, characterized in that: Step S1 includes the following sub-steps: S11: Read the core image files and check the file formats one by one; S12: setting the cutting size, defining the pixel size of the small cube in the width, height and depth directions respectively, and cutting out a plurality of cube sub-regions for training from the core image according to the size; S13: Save each sub-region sample in HDF5 format.

3. The core three-dimensional reconstruction method based on KAN convolutional network according to claim 1, characterized in that: Step S4 includes the following sub-steps: S41: The input single-channel 3D image is expanded to 16 channels using 3D convolution, and then LeakyReLU is used for activation; S42: Three layers of 3D convolution are used consecutively, each followed by BatchNorm3d and LeakyReLU activation; S43: Use the last layer of 3D convolution to reduce the number of channels to 1, and use the Sigmoid activation function to limit the output to [0, 1], indicating the true or false judgment score of the sample.

4. A core three-dimensional reconstruction method based on a KAN convolutional network as claimed in claim 3, characterized in that: Step S5 includes the following sub-steps: S51: Update the network parameters of the discriminator; S52: Update the network parameters of the generator; S53: Repeat steps S51 and S52 until the parameters of the generator and the discriminator converge; in continuous iteration and optimization, the samples generated by the generator gradually approach the distribution of real samples, and the discriminator more accurately identifies real and generated samples.

5. A core three-dimensional reconstruction method based on a KAN convolutional network as claimed in claim 4, characterized in that: Step S51 includes the following sub-steps: S511: Input random noise and real samples , random noise Input the generator and get the generated sample ; S512: Calculate the real sample Loss ; S513: Calculate and generate samples Loss ; S514: Calculate the total loss of the discriminator for ; S515: Update the discriminator parameters using the Adam optimizer , specifically: ； in, is the learning rate, is the loss gradient of the discriminator.

6. A core three-dimensional reconstruction method based on a KAN convolutional network as claimed in claim 4, characterized in that: Step S52 includes the following sub-steps: S521: Random noise Input generator to get generated samples ; S522: Input the false sample into the discriminator to obtain the output , represents the score of the discriminator on the generated samples; S523: Setting the target label of the generator to a range of randomly uniformly distributed values; S524: Computing generator loss using binary cross entropy loss , specifically: ； S525: Loss to the generator Perform backpropagation and update the generator parameters using the Adam optimizer , specifically: ； in, is the learning rate, is the loss gradient of the generator.

7. The method for three-dimensional reconstruction of cores based on a KAN convolutional network according to claim 4, characterized in that: Step S6 includes the following sub-steps: S61: Generate three-dimensional image samples of new cores using the trained generator; S62: Calculate the porosity and permeability of the core structure based on the generated three-dimensional image. The porosity represents the ratio of the pore volume to the total volume in the core image. The specific calculation method is: ； in, represents the pore volume in a three-dimensional image, is the total volume of the core image; Permeability measures the ability of fluid flow in the generated core, and the specific calculation expression is: ； in, is the fluid flow rate, is the fluid viscosity, is the core length, is the flow area, is the pressure difference; S63: The porosity and permeability of the generated images are compared with the corresponding indicators of the real core samples to evaluate the effectiveness and accuracy of the model in reconstructing the core structure.

Images