Core permeability prediction method based on time sequence decomposition and multi-dimensional feature interaction

Abstract

The invention provides a core permeability prediction method based on time sequence decomposition and multi-dimensional feature interaction, and belongs to the technical field of core permeability prediction. The method comprises the following steps: performing CT scanning on multiple sandstone samples to obtain high-resolution three-dimensional rock core original data; the method comprises the following steps: performing image preprocessing on original data, performing adaptive threshold segmentation, and dividing pores and matrixes to obtain three-dimensional binary image data; performing small window sliding sampling on the binary image data to obtain three-dimensional core small cubes, and calculating the permeability of each core cube as a label to obtain a final experimental data set; a core image sequence is decomposed from an experimental data set, fusion features are extracted, information is extracted from the fusion features from three dimensions of channel, height and width, features of different dimensions are combined, and the permeability is predicted through a full connection layer. According to the method, lightweight design is adopted, multi-dimensional information of the core image is fully utilized, prediction efficiency and precision are remarkably improved, and computing resource consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of core permeability prediction, and particularly to a core permeability prediction method based on time series decomposition and multi-dimensional feature interaction. Background Art

[0002] Permeability is an important parameter for measuring the fluid flow ability of rock reservoirs, and its accurate prediction is crucial for oil and gas resource development and management. Traditional methods usually adopt Darcy's law based on laboratory measurement of core samples, but this method requires a large number of samples, time and cost, and it is difficult to adapt to the scenarios of core fragmentation or limited samples. In addition, physical experiments are difficult to accurately reflect the complexity of the rock microstructure, resulting in limitations in its prediction ability.

[0003] With the progress of computer tomography (CT) technology, reconstructing three-dimensional core images through X-ray CT scanning provides a non-destructive method for calculating permeability. Through high-resolution three-dimensional CT scanning, information such as the pore structure and mineral composition of the core can be accurately obtained, thus providing rich image data for the analysis of geological parameters such as permeability.

[0004] With the development of artificial intelligence and computer vision technologies, deep learning has become an important tool for processing complex image data and feature modeling. In the problem of core permeability prediction, the key is to extract effective features from high-resolution three-dimensional core images and capture the complex geometric characteristics of the pore structure and its influence on fluid flow. However, core image data has multi-dimensional features, including space, height, width, and different image channels, etc. Traditional deep learning methods usually adopt two-dimensional convolutional networks, which fail to fully utilize this multi-dimensional information, limiting the mining of the complex spatial relationships in core images and resulting in insufficient prediction accuracy. Especially when dealing with complex pore structures and fine-grained information, the fluid flow characteristics within the core cannot be fully reflected.

[0005] The invention patent with the application number CN202410097446.4 discloses a three-dimensional core permeability prediction method based on a spatio-temporal graph neural network. It performs three-dimensional reconstruction on the X-ray μCT scanned core sample image sequence; filters the three-dimensional data noise of the core, enhances the contrast between the core entity and the pore-fracture, and performs automatic threshold segmentation to obtain binary three-dimensional data; cuts the three-dimensional data volume into small volume samples, calculates the core permeability using Darcy's formula as the data label; extracts the pore network for each small volume sample three-dimensional core and processes it into spatio-temporal graph characterization data; inputs the spatio-temporal graph characterization data into the spatio-temporal graph neural network model, uses the graph neural network with a multi-query attention mechanism and the gated logic unit module to extract the spatio-temporal features of the core, and predicts the permeability through a multi-layer perceptron. Although this method can reduce the interference of data redundancy components, due to the lack of a cross-dimensional information interaction strategy for complementary fusion and deep correlation mining of multi-perspective features of core data, the prediction accuracy of core permeability is poor, so further improvement is needed. Summary of the Invention

[0006] In view of this, the present invention provides a core permeability prediction method based on time series decomposition and multi-dimensional feature interaction to solve the problem of low accuracy of existing core permeability prediction methods.

[0007] The technical solution adopted by the present invention is as follows: The present invention provides a core permeability prediction method based on time series decomposition and multi-dimensional feature interaction, including: S1: Perform X-ray microtomography on multi-category sandstone core samples, and use a three-dimensional reconstruction algorithm to process the projection data collected by the scan to obtain a high-resolution three-dimensional core grayscale image; S2: Perform image preprocessing on the core grayscale image, enhance the contrast between the solid space and the pore-fracture, and perform adaptive threshold segmentation to divide the pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image; S3: Use the sliding window sampling method to block-process the three-dimensional core binary image to obtain multiple core cubes. At the same time, simulate fluid flow based on the lattice Boltzmann method and combine Darcy's law to calculate the permeability of each core cube as the corresponding data label to obtain the final experimental data set; S4: Perform spatio-temporal trend-residual decomposition on the core images in the final experimental data set, extract the global trend features and residual features, and fuse the global trend features and residual features to obtain a fused feature map; S5: Use a feature extractor with three parallel branches to separately extract the core image features in the channel dimension, height dimension, and width dimension from the fused feature map; among them, the feature extractor for each branch is composed of a stack of multiple layers of structured convolutional layers, SE attention modules, layer normalization layers, and activation functions; S6: Align the core image features in the channel dimension, height dimension, and width dimension, and perform concatenation and combination in the channel dimension in the order of channels, height, and width to obtain comprehensive features; S7: Use a deep non-linear transformation network composed of multiple layers of convolution and activation functions to perform multi-layer mapping on the comprehensive features, gradually reduce the dimensionality of the comprehensive features to the permeability scalar output, and predict the corresponding core permeability.

[0008] Further, perform X-ray microtomography on multi-category sandstone core samples, and use a three-dimensional reconstruction algorithm to process the projection data collected by scanning to obtain high-resolution three-dimensional core grayscale images, specifically including: S11: Select multi-category sandstone core samples, perform cutting and polishing treatments to make standard cylindrical multi-core samples; S12: Perform X-ray projection acquisitions of the multi-core samples from multiple angles, and use the filtered back-projection algorithm to reconstruct the projection data to obtain high-resolution three-dimensional core grayscale images.

[0009] Further, perform image preprocessing on the core grayscale images, enhance the contrast of the solid space and pore fissures, and perform adaptive threshold segmentation to divide pores and matrices in the three-dimensional core grayscale images to obtain three-dimensional core binary images, specifically including: S21: Perform median filtering denoising with a window size of 9*9*9 on the obtained three-dimensional core grayscale images, and use an interpolation method to fill in missing values to obtain smooth and complete core grayscale images; S22: First divide the smooth and complete core grayscale image into multiple blocks, calculate the local grayscale histogram and set the grayscale value limit, calculate the cumulative distribution function of each block and perform normalization processing, then map the normalized result through a mapping function to obtain new grayscale values, and finally splice each block into a complete image to obtain an enhanced core grayscale image; S23: Perform adaptive threshold segmentation on the enhanced grayscale image. First, divide the enhanced grayscale image into multiple image blocks, then calculate the local threshold according to the grayscale characteristics of each image block, mark the image blocks with pixels higher than the local threshold as solid matrices, mark the image blocks with pixels lower than or equal to the local threshold as pores, and then splice each image block to obtain a complete three-dimensional core binary image.

[0010] Furthermore, the three-dimensional core binary image is processed in blocks by using the sliding window sampling method to obtain a plurality of core cubes. At the same time, the fluid flow is simulated according to the lattice Boltzmann method and combined with Darcy's law to calculate the permeability of each core cube as the corresponding data label, and the final experimental data set is obtained, specifically including: S31: Use an overlapping sliding window to slide the three-dimensional core binary image along the X / Y / Z spatial coordinate axes at a specified step size, and cut the image into several core cubes of a specified scale; S32: Use the lattice Boltzmann method to simulate fluid flow and combine with Darcy's law to calculate the permeability of each core cube as the data label corresponding to the core cube, so as to obtain a multi-sandstone data set for training a deep learning model; Among them, the formula for calculating the permeability of each core cube by simulating fluid flow according to the lattice Boltzmann method and combining with Darcy's law is as follows: In the formula, represents the volume of fluid passing through the cross-section per unit time, represents the characteristic permeability coefficient of the porous medium, represents the cross-sectional area through which the fluid passes, represents the pressure difference of the fluid between two points, represents the distance that the fluid flows in the porous medium; S33: Perform data augmentation processing on the multi-sandstone data set. First, remove the data points that deviate significantly from the average value, and then perform random sampling within each layer of permeability values to obtain the final experimental data set.

[0011] Furthermore, the spatiotemporal trend-residual decomposition is performed on the core images in the final experimental data set, the global trend features and residual features are extracted, and the global trend features and residual features are feature-fused to obtain a fused feature map, specifically including: S41: Perform spatiotemporal trend-residual decomposition on the three-dimensional core image sequence in the final experimental data set. The global trend features are extracted through sliding window smoothing processing by a three-dimensional average pooling layer to capture the macroscopic seepage law, and the local detail features are separated by residual calculation as residual features to focus on the pore microstructure and local spatial changes. The spatiotemporal trend-residual decomposition formula is as follows: In the formula, is the global trend feature within the time window n, is the input image sequence at the historical moment, is the residual feature at the current moment, and n is the length of the time window for moving average, which controls the smoothing degree; S42: For the global trend feature and the residual features Perform data dimension permutation; S43: For the global trend features Perform a linear transformation on the channel dimension and activate and compress the time dimension. The formula is expressed as: where, is the learnable weight matrix, ReLU is the activation function, is the bias term, is the global trend feature after channel conversion; S44: For the residual features Perform multi-scale convolution, use convolution kernels of different sizes for convolution, extract pore structure features, and take into account local pore throats and regional connectivity features during the extraction process. Then, fuse the pore structure features extracted by convolution kernels of different sizes. The formula is expressed as: + ; where, represents the residual feature after feature fusion, represents the convolution kernel of size represents the convolution kernel of size represents the convolution kernel of size ; S45: Fuse the global trend feature and the residual feature through skip connection to obtain a fused feature map that combines global information and local details: where, represents the fused feature map.

[0012] Furthermore, the feature extractor with three parallel branches extracts core image features in the channel dimension, height dimension, and width dimension from the fused feature map respectively, specifically including: S51: Use the feature extractor with three parallel branches to extract core image features in the channel dimension, height dimension, and width dimension from the fused feature map respectively; where, the feature extractor of each branch is composed of a stack of multiple structured convolutional layers, SE attention modules, layer normalization layers, and activation functions. The core formula is expressed as follows: In the formula, represents the tensor after dimension permutation of the input fused feature map; represents the SE attention module customized for different branches, including​​ Module, module, and module; The layer normalization operation; ReLU represents the activation function; Conv2D represents the multi - layer structured convolutional layer; represents the core image features extracted by each branch; S52: The channel branch maintains the spatial arrangement order of the fused feature map , extracts global structure features through a convolutional layer, and then uses the module with the dimension permutation parameter pos=(0, 1, 2, 3) to adaptively weight the channel dimension and quantify the importance of different channels; S53: The height branch first permutes the dimensions of the fused feature map to , then captures the layered structure and gradient changes in the vertical direction through a convolutional layer, and then uses the module with the dimension permutation parameter pos=(0, 1, 3, 2) to compress features in the height dimension, dynamically adjust the vertical - direction feature weights, and quantify the importance of the vertical layered structure; S54: The width branch first permutes the dimensions of the fused feature map to , then captures the layered structure and gradient changes in the horizontal direction through a convolutional layer, and then uses the module with the dimension permutation parameter pos=(0, 3, 2, 1) to compress features in the width dimension and quantify the importance of the horizontal pore connectivity; S55: Each branch gradually downsamples through a convolutional layer with a stride of 2, reducing the spatial resolution while extracting features, and finally outputs a feature map with aligned dimensions ; S56: The SE attention module first performs global information compression on the input feature map along the spatial dimension to generate a global description for each channel, then generates the weight for each channel through two fully - connected layers and a non - linear activation function, and finally uses the generated channel weights to readjust the input feature map and multiply the feature map of each channel by the corresponding scaling factor to obtain the core image features in the channel dimension, height dimension, and width dimension.

[0013] Furthermore, aligning the core image features in the channel dimension, height dimension, and width dimension and cascading and combining them in the channel dimension in the order of channel, height, and width to obtain the comprehensive features specifically includes: S61: Use a 1×1 convolutional layer to map the core image features in the channel dimension, height dimension, and width dimension from each branch to the same feature dimension to ensure channel alignment; S62: Cascade and combine the core image features of each branch with aligned channels in the channel dimension in the order of channel, height, and width to achieve an organic fusion at the feature level and obtain comprehensive features.

[0014] In summary, the beneficial effects of the present invention are as follows: The present invention provides a core permeability prediction method based on time series decomposition and multi-dimensional feature interaction. The method accurately extracts key feature information from high-resolution three-dimensional core images to construct a final experimental dataset, and performs spatio-temporal trend-residual decomposition on the core images in the final experimental dataset. The global trend features and residual features obtained by the decomposition are subjected to feature fusion. A feature extractor with three parallel branches is used to extract the core image features in the channel dimension, height dimension, and width dimension respectively from the feature map obtained by feature fusion for feature cascade combination. By fusing multi-dimensional features, the accuracy of permeability prediction can be effectively improved. Especially when dealing with complex core images, it shows higher prediction performance than traditional methods, and avoids the high cost and long cycle of physical experiments, meeting the current demand for efficient and low-cost permeability prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to these drawings, and all of these are within the protection scope of the present invention.

[0016] Figure 1 It is a schematic flow chart of the core permeability prediction method based on time series decomposition and multi-dimensional feature interaction of the present invention; Figure 2 It is a model architecture diagram of the core permeability prediction method based on time series decomposition and multi-dimensional feature interaction; Figure 3 It is a schematic diagram of the mean square error (RMSE) of the model experiment; Figure 4 It is a schematic diagram of the coefficient of determination (R2) of the model experiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. If there is no conflict, the various features in the present invention and its embodiments can be combined with each other, and all are within the protection scope of the present invention.

[0018] Example 1: Refer to Figure 1 as shown in Figure 1 is the flow chart of the core permeability prediction method based on time series decomposition and multi-dimensional feature interaction of the present invention. The method includes: S1: Perform X-ray microtomography on multi-category sandstone core samples, and use a three-dimensional reconstruction algorithm to process the projection data collected by the scan to obtain a high-resolution three-dimensional core grayscale image; S2: Perform image preprocessing on the core grayscale image to enhance the contrast of the solid space and pore fissures, and perform adaptive threshold segmentation to divide pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image; S3: Use the sliding window sampling method to block-process the three-dimensional core binary image to obtain multiple core cubes. At the same time, simulate fluid flow based on the lattice Boltzmann method and calculate the permeability of each core cube in combination with Darcy's law as the corresponding data label to obtain the final experimental data set; S4: Perform spatio-temporal trend-residual decomposition on the core images in the final experimental data set, extract global trend features and residual features, and fuse the global trend features and residual features to obtain a fused feature map; S5: Use a feature extractor with three parallel branches to extract core image features in the channel dimension, height dimension, and width dimension from the fused feature map respectively; among them, the feature extractor of each branch is composed of a stack of multi-layer structured convolutional layers, SE attention modules, layer normalization layers, and activation functions; S6: Align the core image features in the channel dimension, height dimension, and width dimension, and perform cascade combination in the channel dimension in the order of channel, height, and width to obtain a comprehensive feature; S7: Use a deep non-linear transformation network composed of multi-layer convolution and activation functions to perform multi-layer mapping on the comprehensive feature, gradually reduce the dimension of the comprehensive feature to the permeability scalar output, and predict the corresponding core permeability.

[0019] Furthermore, in the embodiment of the present invention, in step S1, perform X-ray microtomography on multi-category sandstone core samples, and use a three-dimensional reconstruction algorithm to process the projection data collected by the scan to obtain a high-resolution three-dimensional core grayscale image, specifically including: S11: Select multi-category sandstone core samples, perform cutting and polishing treatments to make standard cylindrical multi-core samples. The standard cylindrical samples ensure the consistency of experimental conditions, making the experimental results at different laboratories or different time points comparable. This standardization helps to improve the accuracy and reliability of scientific research.

[0020] S12: Collect X-ray projections of the multi-core samples from multiple angles, and use the filtered back-projection algorithm to reconstruct the projection data to obtain a high-resolution three-dimensional core gray-scale image.

[0021] Among them, filtered back-projection (FBP) is a reconstruction algorithm widely used in imaging technologies such as computed tomography (CT). The filtered back-projection algorithm is based on the Radon transform of the projection data and realizes image reconstruction through two main steps: filtering and back-projection. Specifically, the algorithm first projects rays through the object at different angles to collect a series of projection data. Then, these projection data are filtered to eliminate artifacts and noise. Finally, the filtered projection data are back-projected into the object space, that is, distributed evenly along the paths of the rays passing through the object at each projection angle, thereby obtaining a preliminary reconstructed image. By performing multiple back-projections and superpositions, the original image can be inferred. The filtered back-projection algorithm has the advantages of fast speed and relatively simple implementation. It can reconstruct the internal structure image of the object from the projection data in a relatively short time.

[0022] Furthermore, in the embodiments of the present invention, in step S2, image preprocessing is performed on the core gray-scale image to enhance the contrast between the solid space and the pore-fracture, and adaptive threshold segmentation is performed to divide the pores and the matrix in the three-dimensional core gray-scale image to obtain a three-dimensional core binary image, which specifically includes: S21: Perform median filtering denoising with a window size of 9*9*9 on the obtained three-dimensional core gray-scale image, and use the interpolation method to fill in the missing values to obtain a smooth and complete core gray-scale image.

[0023] Among them, the denoising processing formula in the embodiments of the present invention is as follows: In the formula, represents the new voxel value after median filtering processing, represents the voxel value of the original image, is the offset of the neighboring voxel relative to the central voxel, and the range of the offset is , N is taken as 9, and the median function Median() calculates the median of the selected set of neighboring voxel values.

[0024] The interpolation formula in the embodiments of the present invention is as follows: In the formula, is the value of the point to be interpolated, is the coordinate of the central point of interpolation, represents the weighted sum of all relevant voxel values and their corresponding deviations, is a coefficient calculated from the surrounding voxel values, is the deviation between the interpolation point and the center point.

[0025] S22: First, divide the smooth and complete core gray-scale image into multiple blocks, calculate the local gray-scale histogram and set the gray-scale value limit, calculate the cumulative distribution function of each block and perform normalization processing, then map the normalized result through a mapping function to obtain new gray-scale values, and finally splice each block into a complete image to obtain the enhanced core gray-scale image.

[0026] Among them, the formula for the adaptive histogram equalization in the embodiments of the present invention is as follows: In the formula, represents the gray-scale value of the equalized three-dimensional image at position ; represents the gray-scale value of the input image at position ; is the gray-scale histogram within the local area, indicating the frequency of occurrence of each gray-scale value within the specified local window, respectively represent the height, width, and depth considered when calculating the histogram, and are used to normalize the cumulative distribution function, is the total number of gray levels, and is used to map the normalized cumulative distribution function to an appropriate gray-scale range.

[0027] Specifically, the adaptive threshold segmentation formula is as follows: represents the adaptive threshold calculated for the image at position ; represents the gray-scale value of the original gray-scale image at position ; represents the local neighborhood related to the pixel ; represents the total number of pixels within the local area, is an offset constant that controls the adaptive threshold, and its value range is 0.1 to 0.3.

[0028] S23: Perform adaptive threshold segmentation on the enhanced gray-scale image. First, divide the enhanced gray-scale image into multiple image blocks, then calculate the local threshold according to the gray-scale characteristics of each image block, mark the image blocks with pixels higher than the local threshold as solid matrix, mark the image blocks with pixels lower than or equal to the local threshold as pores, and then splice each image block to obtain the complete three-dimensional core binary image.

[0029] Furthermore, in the embodiment of the present invention, in step S3, the sliding window sampling method is used to block-process the three-dimensional core binary image to obtain a plurality of core cubes. At the same time, according to the lattice Boltzmann method to simulate fluid flow and combined with Darcy's law, the permeability of each core cube is calculated as the corresponding data label, and the final experimental data set is obtained, specifically including: S31: Using an overlapping sliding window, slide the three-dimensional core binary image along the X / Y / Z spatial coordinate axes at a specified step length, and cut the image into several core cubes of a specified scale.

[0030] S32: Use the lattice Boltzmann method (i.e., the LBM method) to simulate fluid flow and combine with Darcy's law to calculate the permeability of each core cube as the data label corresponding to the core cube, so as to obtain a multi-sandstone data set for training the deep learning model.

[0031] Among them, the formula for calculating the permeability of each core cube according to the lattice Boltzmann method to simulate fluid flow and combined with Darcy's law is as follows: In the formula, represents the fluid volume passing through the cross-section per unit time, represents the characteristic permeability coefficient of the porous medium, represents the cross-sectional area through which the fluid passes, represents the pressure difference between two points of the fluid, represents the distance that the fluid flows in the porous medium.

[0032] S33: Perform data augmentation processing on the multi-sandstone data set. First, remove the data points that deviate significantly from the average value, and then perform random sampling within each layer of permeability values to obtain the final experimental data set.

[0033] Furthermore, in the embodiment of the present invention, in step S4, the spatio-temporal trend-residual decomposition is performed on the core images in the final experimental data set, the global trend features and residual features are extracted, and the global trend features and residual features are feature-fused to obtain a fused feature map, specifically including: S41: Perform spatio-temporal trend-residual decomposition on the three-dimensional core image sequence in the final experimental data set. The global trend features are extracted through sliding window smoothing processing by a three-dimensional average pooling layer to capture the macroscopic seepage law, and the local detail features are separated by residual calculation as residual features to focus on the pore microstructure and local spatial changes. The spatio-temporal trend-residual decomposition formula is as follows: In the formula, is the global trend feature within the time window n, is the input image sequence at the historical moment, is the residual feature at the current moment, n is the time window length of the moving average, which controls the smoothing degree; S42: For the global trend feature and the residual feature perform data dimension permutation; S43: Perform a linear transformation on the global trend feature in the channel dimension, and activate and compress the time dimension. The formula is expressed as: where, is the learnable weight matrix, ReLU is the activation function, is the bias term, is the global trend feature after channel conversion; S44: Perform multi-scale convolution on the residual feature using convolution kernels of different sizes for convolution, extract the pore structure features, and take into account the local pore throats and regional connectivity features during the extraction process. Then, fuse the pore structure features extracted by convolution kernels of different sizes. The formula is expressed as: + ; where, represents the residual feature after feature fusion, represents the convolution kernel of size , represents the convolution kernel of size ; represents the convolution kernel of size ; S45: Fuse the global trend feature and the residual feature through skip connection to obtain a fused feature map that combines global information and local details, forming the encoder output: where, represents the fused feature map.

[0034] Furthermore, in the embodiment of the present invention, in step S5, a feature extractor with three parallel branches is used to extract the core image features in the channel dimension, height dimension, and width dimension from the fused feature map, specifically including: S51: Use a feature extractor with three parallel branches to extract the core image features in the channel dimension, height dimension, and width dimension from the fused feature map respectively; among them, the feature extractor of each branch is composed of a stack of multiple structured convolutional layers, SE attention modules, layer normalization layers, and activation functions. The core formula is expressed as follows: In the formula, represents the tensor after the dimension permutation of the input fusion feature map; represents the SE attention module customized for different branches, including module, module and module; represents the layer normalization operation; ReLU represents the activation function; Conv2D represents the multi-layer structured convolutional layer; represents the core image features extracted by each branch; S52: The channel branch maintains the spatial arrangement order of the fusion feature map , extracts the global structure features through the convolutional layer, and then uses the module with the dimension arrangement parameter pos=(0, 1, 2, 3) to adaptively weight the channel dimension and quantify the importance of different channels; S53: The height branch first permutes the dimension of the fusion feature map to , then captures the layered structure and gradient changes in the vertical direction through the convolutional layer, and then uses the module with the dimension arrangement parameter pos=(0, 1, 3, 2) to compress the features in the height dimension, dynamically adjust the feature weights in the vertical direction, and quantify the importance of the vertical layered structure; S54: The width branch first permutes the dimension of the fusion feature map to , then captures the layered structure and gradient changes in the horizontal direction through the convolutional layer, and then uses the module with the dimension arrangement parameter pos=(0, 3, 2, 1) to compress the features in the width dimension and quantify the importance of the horizontal pore connectivity; S55: Each branch gradually downsamples through the convolutional layer with a stride of 2, reduces the spatial resolution while extracting features, and finally outputs the feature maps with aligned dimensions ; S56: The SE attention module first performs global information compression on the input feature map along the spatial dimension to generate the global description of each channel, then generates the weights of each channel through two fully connected layers and the non-linear activation function, and finally uses the generated channel weights to readjust the input feature map and multiply the feature map of each channel by the corresponding scaling factor to obtain the core image features in the channel dimension, height dimension and width dimension.

[0035] Among them, the SE attention module of the present invention can be divided into three steps: Squeeze (global information embedding), Excitation (adaptive recalibration), and Scale (reweighting). First, global information compression is performed on the input feature map along the spatial dimension to generate a global description for each channel. Then, the weights for each channel are generated through two fully connected layers and a non-linear activation function. Finally, the input feature map is readjusted using the generated channel weights, and the feature map of each channel is multiplied by the corresponding scaling factor to enhance important features and suppress unimportant features. The formula definition of the Squeeze step in the SE module is as follows: In the formula, represents the global information compression value of channel , H and W represent the spatial dimensions of the feature map, represents the input feature map in the th channel at the position; The formula definition of the Excitation step in the SE module is as follows: In the formula, represents the weight coefficient of channel c, is the Sigmoid function, and are learnable weight matrices, is the ReLU activation function, represents the global information compression value of channel ; The formula definition of the Scale step in the SE module is as follows: In the formula, represents the output after attention, represents the weight coefficient of channel , with a range between 0 and 1 is the value of the th channel of the feature map at the position Furthermore, in step S6 of the embodiment of the present invention, the core image features in the channel dimension, height dimension, and width dimension are aligned in channels and cascaded and combined in the channel dimension in the order of channels, height, and width to obtain a comprehensive feature, which specifically includes: S61: Use a 1×1 convolutional layer to map the core image features in the channel dimension, height dimension, and width dimension from each branch to the same feature dimension to ensure channel alignment; S62: Cascade and combine the core image features with aligned channels in the channel dimension in the order of channel, height, and width to achieve an organic fusion at the feature level and obtain comprehensive features.

[0036] Specifically, the present invention uses a deep non-linear transformation network composed of multiple convolutional layers and activation functions to process the cascaded comprehensive features, extracts high-level features through multiple mappings, and uses residual connections to enhance the expression ability of deep features. Finally, the fused features are gradually reduced in dimension to the permeability scalar output to predict the core permeability.

[0037] Specifically, in the embodiments of the present invention, referring to Figure 2 the shown model architecture diagram, a full-process prediction network model including an encoder module and a decoder module (which can be further divided into a feature interaction module and a feature fusion module) can also be constructed based on the above core permeability prediction method process. After the input data is processed by the encoder and the decoder to form fused features, the fused features are gradually reduced in dimension to the permeability scalar output through a fully connected layer to complete the end-to-end mapping from a three-dimensional image to a physical parameter. Among them, the encoder corresponds to the processing process of step S4 in the method, the feature interaction module in the decoder corresponds to the processing process of step S5 in the method, and the feature fusion module corresponds to the processing process of step S6 in the method.

[0038] Specifically, referring to Figure 3 and Figure 4 shown, in order to evaluate the performance of the core permeability prediction model of this method, our model is compared with 5 models in a comparative experiment, and the architecture of the neural network is implemented through the pytorch framework.

[0039] Our is the prediction network model constructed above in the present invention; Model1: The model from the paper "[1] A J T, B C Q, C Y S, et al. Surrogate permeability modelling of low-permeable rocks using convolutional neural networks - ScienceDirect[J]. Computer Methods in Applied Mechanics and Engineering, 366[2025-04-17]. DOI:10.1016 / j.cma.2020.113103."; Model2 and Model3: models from the paper "[2]Fei J, Yaotian G, Takeshi T, etal. Upscaling Permeability Using Multiscale X‐Ray‐CT Images With Digital RockModeling and Deep Learning Techniques[J]. Water Resources Research, 2023, 59(3):"; Model4: model from the paper: "[3]Nabipour I, Raoof A, Cnudde V, et al. A computationally efficient modeling of flow in complex porous media by coupling multiscale digital rock physics and deep learning: Improving the tradeoff between resolution and field-of-view[J]. Advances in Water Resources, 2024, 188104695-."; Model5: model from the paper: "[4]Mingliang Liu, Rasool Ahmad, Wei Cai, and Tapan Mukerji. Hierarchical homogenization with deep-learning-based surrogate model for rapid estimation of effective permeability from digital rocks. Journal of Geophysical Research: Solid Earth, 128(2):e2022JB025378, 2023."; In this experiment, five metrics are used to measure the model's capabilities and performance, including the coefficient of determination (R 2 ), root mean square error (RMSE); number of parameters (Params), in megabytes (M); floating point operations (Flops), in megabytes (M); and training time (T.T), in minutes per epoch (min / epoch). All results are averaged over four experiments on the same device. Among them, RMSE and R 2The calculation formula is as follows.

[0040] Where n is the number of samples, is the true value of the i-th sample, is the predicted value of the i-th sample, is the average value of the true values of n samples. RMSE reflects the average deviation between the predicted value and the true value of the model. The smaller the RMSE, the better the prediction. R 2 reflects the performance of the model prediction. The larger R 2 is, the better the model.

[0041] In order to analyze the permeability prediction effect of the spatio-temporal multi-dimensional collaborative fusion model, the following experiments were carried out in the present invention. In order to verify the effectiveness of the model, four experiments were conducted on different models to obtain the average values of the corresponding RMSE and R2. The experimental results are shown in Table 1.

[0042] Table 1 Comparison table of model experimental results Therefore, in this embodiment, by proposing a core permeability prediction method based on time series decomposition and multi-dimensional feature interaction, making full use of the multi-dimensional information in the core image, and through its efficient feature fusion ability, integrating spatial information, pore structure and other key physical features in different dimensions, thereby comprehensively capturing the fluid flow characteristics inside the core, overcoming the limitations of traditional methods in dealing with complex and multi-dimensional core data. In addition, compared with the traditional method relying on physical experiments, this model can meet the prediction requirements under different geological conditions, is more suitable for large-scale rock permeability prediction, can process massive data and perform batch prediction efficiently, meeting the actual market demand for efficient and accurate prediction. At the same time, the memory occupancy and calculation speed of the model have been optimized, which can significantly reduce the consumption of computing resources while ensuring high-precision prediction, improve the prediction efficiency, and ensure the feasibility and real-time performance in large-scale application scenarios.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A core permeability prediction method based on time series decomposition and multi-dimensional feature interaction, characterized in that Including: S1: Conduct X-ray microtomography on multi-category sandstone core samples, and use a three-dimensional reconstruction algorithm to process the projection data collected by the scan to obtain a high-resolution three-dimensional core grayscale image; S2: Perform image preprocessing on the core grayscale image to enhance the contrast between the solid space and pore-fracture, and perform adaptive threshold segmentation to divide pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image; S3: Use the sliding window sampling method to block-process the three-dimensional core binary image to obtain multiple core cubes. At the same time, simulate fluid flow based on the lattice Boltzmann method and calculate the permeability of each core cube in combination with Darcy's law as the corresponding data label to obtain the final experimental dataset; S4: Perform spatio-temporal trend-residual decomposition on the core images in the final experimental dataset, extract global trend features and residual features, and fuse the global trend features and residual features to obtain a fused feature map; S5: Use a feature extractor with three parallel branches to extract core image features in the channel dimension, height dimension, and width dimension from the fused feature map respectively; among them, the feature extractor of each branch is composed of a stack of multi-layer structured convolutional layers, SE attention modules, layer normalization layers, and activation functions; S6: Align the core image features in the channel dimension, height dimension, and width dimension, and perform cascade combination in the channel dimension in the order of channel, height, and width to obtain a comprehensive feature; S7: Use a deep non-linear transformation network composed of multi-layer convolution and activation functions to perform multi-layer mapping on the comprehensive feature, gradually reduce the comprehensive feature to the permeability scalar output, and predict the corresponding core permeability.

2. The core permeability prediction method according to claim 1, characterized in that The conduct of X-ray microtomography on multi-category sandstone core samples and the use of a three-dimensional reconstruction algorithm to process the projection data collected by the scan to obtain a high-resolution three-dimensional core grayscale image specifically include: S11: Select multi-category sandstone core samples, perform cutting and polishing to make standard cylindrical multi-core samples; S12: Conduct X-ray projection acquisition of the multi-core samples from multiple angles, and use the filtered back-projection algorithm to reconstruct the projection data to obtain a high-resolution three-dimensional core grayscale image.

3. The core permeability prediction method according to claim 1, characterized in that The performance of image preprocessing on the core grayscale image to enhance the contrast between the solid space and pore-fracture, and perform adaptive threshold segmentation to divide pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image specifically includes: S21: Perform median filtering denoising with a window size of 9*9*9 on the obtained three-dimensional core grayscale image, and use an interpolation method to fill in missing values to obtain a smooth and complete core grayscale image; S22: First divide the smooth and complete core grayscale image into multiple blocks, calculate the local grayscale histogram and set the grayscale value limit, calculate the cumulative distribution function of each block and perform normalization processing, then map the normalization result through a mapping function to obtain a new grayscale value, and finally splice each block into a complete image to obtain an enhanced core grayscale image; S23: Perform adaptive threshold segmentation on the enhanced grayscale image. First, divide the enhanced grayscale image into multiple image blocks, then calculate the local threshold according to the grayscale characteristics of each image block. Mark the image blocks with pixels higher than the local threshold as solid matrices, and mark the image blocks with pixels lower than or equal to the local threshold as pores. Then, splice each image block to obtain a complete three-dimensional core binary image.

4. The core permeability prediction method according to claim 1, characterized in that The three-dimensional core binary image is processed by block division using the sliding window sampling method to obtain multiple core cubes. At the same time, the fluid flow is simulated based on the lattice Boltzmann method and combined with Darcy's law to calculate the permeability of each core cube as the corresponding data label, resulting in the final experimental dataset, specifically including: S31: Use an overlapping sliding window to slide the three-dimensional core binary image along the X / Y / Z spatial coordinate axes at a specified step size, and cut the image into several core cubes of a specified scale; S32: Use the lattice Boltzmann method to simulate fluid flow and combine Darcy's law to calculate the permeability of each core cube as the data label corresponding to the core cube, thereby obtaining a multi-sandstone dataset for training the deep learning model; Among them, the formula for calculating the permeability of each core cube based on the lattice Boltzmann method to simulate fluid flow and combined with Darcy's law is as follows: wherein, represents the volume of fluid passing through the cross-section per unit time, represents the characteristic permeability coefficient of the porous medium, represents the area of the cross-section through which the fluid passes, represents the pressure difference of the fluid between two points, represents the distance that the fluid flows in the porous medium; S33: Perform data augmentation processing on the multi-sandstone dataset. First, remove the data points that deviate significantly from the average value, and then perform random sampling within each layer of permeability values to obtain the final experimental dataset.

5. The core permeability prediction method according to claim 1, characterized in that Perform spatio-temporal trend-residual decomposition on the core images in the final experimental dataset, extract the global trend features and residual features, and fuse the global trend features and residual features to obtain a fused feature map, specifically including: S41: Perform spatio-temporal trend-residual decomposition on the three-dimensional core image sequence in the final experimental dataset. Extract the global trend features through sliding window smoothing processing by a three-dimensional average pooling layer to capture the macroscopic seepage law, and separate the local detail features as residual features through residual calculation to focus on the pore microstructure and local spatial changes. The spatio-temporal trend-residual decomposition formula is as follows: In the formula, is the global trend feature within the time window n, is the input image sequence at historical moments, is the residual feature at the current moment, and n is the time window length of the moving average, which controls the smoothness; S42: Perform data dimension permutation on the global trend feature and the residual feature ; S43: Perform a linear transformation on the global trend features in the channel dimension and activate the compression of the time dimension, which is expressed by the formula: Among them, is a learnable weight matrix, ReLU is an activation function, is a bias term, is the global trend feature after channel conversion; S44: For the residual features perform multi-scale convolution, use convolution kernels of different sizes for convolution, extract pore structure features, and take into account local pore throats and regional connectivity features during the extraction process. Then, fuse the pore structure features extracted by convolution kernels of different sizes. The formula is expressed as: + ； Among them, represents the residual feature after feature fusion, represents a convolution kernel of size ; represents a convolution kernel of size ; represents a convolution kernel of size ; S45: Fuse the global trend feature with the residual feature through skip connection to obtain a fused feature map that integrates global information and local details: Among them, represents the fused feature map.

6. The core permeability prediction method according to claim 1, wherein Use a feature extractor with three parallel branches to extract the core image features in the channel dimension, height dimension, and width dimension from the fused feature map, specifically including: S51: Extract the core image features in the channel dimension, height dimension, and width dimension from the fused feature map through a feature extractor with three parallel branches; among them, the feature extractor for each branch is composed of a stack of multi-layer structured convolutional layers, SE attention modules, layer normalization layers, and activation functions. The core formula is expressed as follows: In the formula, represents the tensor after the dimensional permutation of the input fusion feature map; represents the SE attention module customized for different branches, including module, module and module; represents the layer normalization operation; ReLU represents the activation function; Conv2D represents the multi-layer structured convolutional layer; represents the core image features extracted by each branch; S52: The channel branches maintain the spatial arrangement order of the fused feature maps , extract global structural features through the convolutional layer, and then use the module with the dimension permutation parameter pos=(0, 1, 2, 3) to adaptively weight the channel dimension and quantify the importance of different channels; S53: The height branch first permutes the dimensions of the fused feature map to , then captures the layered structure and gradient changes in the vertical direction through a convolutional layer, and then uses a module with dimension permutation parameters pos=(0, 1, 3, 2) to perform feature compression in the height dimension, dynamically adjust the feature weights in the vertical direction, and quantify the importance of the vertical layered structure; S54: The width branch first permutes the dimensions of the fused feature map to , then captures the layered structure and gradient changes in the horizontal direction through a convolutional layer, and then uses a module with dimension permutation parameter pos=(0, 3, 2, 1) to perform feature compression in the width dimension and quantify the importance of horizontal pore connectivity; S55: Each branch is gradually downsampled through a convolutional layer with a stride of 2, reducing the spatial resolution while extracting features, and finally outputting a feature map with aligned dimensions. ; S56: The SE attention module first performs global information compression on the input feature map along the spatial dimension to generate a global description for each channel, then generates the weights for each channel through two fully connected layers and a non-linear activation function, and finally uses the generated channel weights to readjust the input feature map and multiply the feature map of each channel by the corresponding scaling factor to obtain the core image features in the channel dimension, height dimension, and width dimension. Along the spatial dimension, global information compression is performed on the input feature map to generate a global description for each channel. Then, the weights for each channel are generated through two fully connected layers and a non-linear activation function. Finally, the generated channel weights are used to readjust the input feature map, and the feature map of each channel is multiplied by the corresponding scaling factor to obtain the core image features in the channel dimension, height dimension, and width dimension.

7. The core permeability prediction method according to claim 1, characterized in that Align the core image features in the channel dimension, height dimension, and width dimension, and perform cascade combination in the channel dimension in the order of channel, height, and width to obtain a comprehensive feature, specifically including: S61: Use a 1×1 convolutional layer to map the core image features in the channel dimension, height dimension, and width dimension from each branch to the same feature dimension to ensure channel alignment; S62: Cascading and combining the channel-aligned core image features of each branch in the order of channel, height, and width in the channel dimension to achieve an organic fusion at the feature level and obtain comprehensive features.

Images