A core permeability prediction method based on time series decomposition and multi-dimensional feature interaction

Abstract

The present invention provides a core permeability prediction method based on the interaction of time series decomposition and multi-dimensional features, which belongs to the field of core permeability prediction technology. The method performs CT scanning on multiple sandstone samples to obtain high-resolution three-dimensional core raw data; after image preprocessing, the raw data is adaptively thresholded and segmented to divide the pores and matrix to obtain three-dimensional binary image data; a small window sliding sampling is performed on the binary image data to obtain three-dimensional core small cubes, and then the permeability of each core cube is calculated as a label to obtain the final experimental data set; the core image sequence is decomposed from the experimental data set, fusion features are extracted, and information is extracted from the fusion features from the three dimensions of channel, height, and width, and features of different dimensions are merged to predict the permeability through a fully connected layer. The present invention adopts a lightweight design and makes full use of the multi-dimensional information of the core image, significantly improving the prediction efficiency and accuracy and reducing the consumption of computing resources.

Description

Technical Field

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

[0002] Permeability is a key parameter that measures the fluid flow capacity of rock reservoirs, and its accurate prediction is crucial for oil and gas resource development and management. Traditional methods typically employ Darcy's law based on laboratory measurements of core samples. However, this method requires a large number of samples, is time-consuming and expensive, and is not adaptable to scenarios with broken cores or limited sample sizes. Furthermore, physical experiments struggle to accurately reflect the complexity of rock microstructure, limiting their predictive capabilities.

[0003] Advances in computed tomography (CT) technology have enabled the reconstruction of three-dimensional core images using X-ray CT scans, providing a non-destructive method for permeability calculation. High-resolution three-dimensional CT scans accurately capture information such as the core's pore structure and mineral composition, 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 to capture the complex geometric characteristics of the pore structure and its impact on fluid flow. However, core image data has multidimensional characteristics, including space, height, width, and different image channels. Traditional deep learning methods usually use two-dimensional convolutional networks, which fail to fully utilize this multidimensional information, limit the mining of complex spatial relationships in core images, and lead to insufficient prediction accuracy. In particular, when dealing with complex pore structures and fine-grained information, it is impossible to fully reflect the fluid flow characteristics within the core.

[0005] Patent application number CN202410097446.4 discloses a three-dimensional core permeability prediction method based on a spatiotemporal graph neural network. The method performs three-dimensional reconstruction of a sequence of X-ray μCT scanned core sample images. It then filters the three-dimensional core data noise, enhances the contrast between the core solid and the pores, and performs automated threshold segmentation to obtain binary three-dimensional data. The three-dimensional data volume is then divided into small-volume samples, and the core permeability is calculated using the Darcy formula as the data label. For each small-volume three-dimensional core sample, the pore network is extracted and processed into spatiotemporal graph representation data. This spatiotemporal graph representation data is input into a spatiotemporal graph neural network model, where a graph neural network with a multi-query attention mechanism and a gated logic unit module are used to extract the core's spatiotemporal features. Permeability is then predicted using a multi-layer perceptron. While this method can reduce the interference of redundant data components, it lacks a cross-dimensional information interaction strategy to complement the multi-perspective feature fusion and deep correlation mining of the core data, resulting in poor prediction accuracy for core permeability. Therefore, 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 in the present invention is:

[0008] The present invention provides a core permeability prediction method based on time series decomposition and multi-dimensional feature interaction, comprising:

[0009] S1: Perform X-ray micro-tomography on multiple types of sandstone core samples and use a 3D reconstruction algorithm to process the collected projection data to obtain high-resolution 3D grayscale images of the core;

[0010] S2: Preprocess the core grayscale image to enhance the contrast between solid space and pores and fractures, and perform adaptive threshold segmentation to delineate pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image.

[0011] S3: The 3D core binary image is segmented using a sliding window sampling method to obtain multiple core cubes. The lattice Boltzmann method is used to simulate fluid flow and Darcy's law is used to calculate the permeability of each core cube as the corresponding data label to obtain the final experimental data set.

[0012] S4: Perform spatiotemporal 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;

[0013] S5: Using three parallel branch feature extractors to extract channel dimension, height dimension and width dimension core image features from the fused feature map respectively; wherein the feature extractor of each branch is composed of multiple layers of structured convolutional layers, SE attention modules, layer normalization layers and activation function stacks;

[0014] S6: performing channel alignment on the core image features in the channel dimension, height dimension, and width dimension, and performing cascade combination in the channel dimension in the order of channel, height, and width to obtain a comprehensive feature;

[0015] S7: A deep nonlinear transformation network composed of multi-layer convolution and activation functions is used to perform multi-layer mapping on the comprehensive features, gradually reducing the dimensionality of the comprehensive features to a permeability scalar output, and predicting the corresponding core permeability.

[0016] Furthermore, the X-ray micro-tomography is performed on multiple types of sandstone core samples, and the projection data collected by the scanning is processed using a three-dimensional reconstruction algorithm to obtain a high-resolution three-dimensional core grayscale image, specifically including:

[0017] S11: Select multiple types of sandstone core samples, cut and polish them, and make them into standard cylindrical core samples;

[0018] S12: Perform multi-angle X-ray projection acquisition on multiple core samples and reconstruct the projection data using a filtered back-projection algorithm to obtain a high-resolution three-dimensional core grayscale image.

[0019] Furthermore, the core grayscale image is preprocessed to enhance the contrast between solid space and pores and fractures, and adaptive threshold segmentation is performed to divide pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image, specifically including:

[0020] S21: performing median filtering with a window size of 9*9*9 on the obtained 3D rock core grayscale image to remove noise, and using an interpolation method to fill in missing values ​​to obtain a smooth and complete rock core grayscale image;

[0021] S22: First, the smooth and complete core grayscale image is divided into multiple blocks, the local grayscale histogram is calculated and the grayscale value limit is set, the cumulative distribution function of each block is calculated and normalized, and then the normalized result is mapped by a mapping function to obtain a new grayscale value. Finally, each block is spliced ​​into a complete image to obtain an enhanced core grayscale image;

[0022] S23: Adaptive threshold segmentation is performed on the enhanced grayscale image. First, the enhanced grayscale image is divided into multiple image blocks. Then, a local threshold is calculated based on the grayscale characteristics of each image block. Image blocks with pixels above the local threshold are marked as solid matrix, and image blocks with pixels below or equal to the local threshold are marked as pores. Each image block is then spliced ​​together to obtain a complete three-dimensional core binary image.

[0023] Furthermore, the sliding window sampling method is used to block the three-dimensional core binary image to obtain multiple core cubes. At the same time, the lattice Boltzmann method is used to simulate fluid flow and the permeability of each core cube is calculated as the corresponding data label in combination with Darcy's law to obtain the final experimental data set, which specifically includes:

[0024] S31: Using an overlapping sliding window, the 3D core binary image is slid along the X / Y / Z spatial coordinate axes according to a specified step size to cut the image into several core cubes of specified scales;

[0025] S32: Use the lattice Boltzmann method to simulate fluid flow and combine it with Darcy's law to calculate the permeability of each core cube. This is used as the data label corresponding to the core cube to obtain a multi-sandstone dataset for training the deep learning model.

[0026] The permeability formula for each core cube is calculated based on the lattice Boltzmann method to simulate fluid flow and Darcy's law as follows:

[0027]

[0028] Where, It 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, It represents the pressure difference of a fluid between two points. Represents the distance the fluid flows in the porous medium;

[0029] S33: The multi-sandstone dataset is augmented by first removing data points that significantly deviate from the mean value, and then performing random sampling within each permeability value layer to obtain the final experimental dataset.

[0030] Furthermore, the core images in the final experimental data set are subjected to spatiotemporal trend-residual decomposition to extract global trend features and residual features, and the global trend features and residual features are fused to obtain a fused feature map, specifically including:

[0031] S41: Perform spatiotemporal trend-residual decomposition on the 3D core image sequence in the final experimental dataset. Sliding window smoothing is performed using a 3D average pooling layer to extract global trend features and capture macroscopic permeability patterns. Local detail features are separated as residual features through residual calculation, focusing on pore microstructure and local spatial variations. The spatiotemporal trend-residual decomposition formula is as follows:

[0032]

[0033] Where, is the global trend feature within the time window n, is the input image sequence of historical moments, is the residual feature at the current moment, n is the time window length of the sliding average, which controls the degree of smoothness;

[0034] S42: Global trend characteristics and residual features Perform data dimension replacement;

[0035] S43: Global trend characteristics Perform linear transformation of the channel dimension and activate the compressed time dimension. The formula is expressed as:

[0036]

[0037] in, is the learnable weight matrix, ReLU is the activation function, is the bias term, It is the global trend feature after channel conversion;

[0038] S44: Residual features Multi-scale convolution is performed, and convolution kernels of different sizes are used for convolution to extract pore structure characteristics. In the extraction process, both local pore throat and regional connectivity characteristics are taken into account. The pore structure characteristics extracted by convolution kernels of different sizes are then fused. The formula is expressed as:

[0039] + ;

[0040] in, Represents the residual feature after feature fusion, Indicates size The convolution kernel, Indicates size The convolution kernel, Indicates size The convolution kernel of

[0041] S45: Global trend characteristics and residual features Through skip connection fusion, a fused feature map that combines global information and local details is obtained:

[0042]

[0043] in, Represents the fused feature map.

[0044] Furthermore, the feature extractor using three parallel branches extracts the core image features of the channel dimension, height dimension and width dimension from the fused feature map, specifically including:

[0045] S51: The core image features in the channel dimension, height dimension, and width dimension are extracted from the fused feature map through three parallel branch feature extractors. The feature extractor of each branch consists of multiple layers of structured convolutional layers, SE attention modules, layer normalization layers, and activation function stacks. The core formula is expressed as follows:

[0046]

[0047] Where, Represents the tensor of the input fused feature map after dimension permutation; Represents SE attention modules customized for different branches, including module, Modules and module; Represents layer normalization operation; ReLU represents activation function; Conv2D represents multi-layer structured convolutional layer; Represents the core image features extracted from each branch;

[0048] S52: Channel branches maintain the spatial arrangement order of fused feature maps , extract global structural features through the convolution layer, and then use the dimension arrangement parameter pos=(0, 1, 2, 3) The module adaptively weights the channel dimensions to quantify the importance of different channels;

[0049] S53: The height branch first replaces the dimension of the fusion feature map with , and then use the convolution layer to capture the layered structure and gradient changes in the vertical direction, and then use the dimension arrangement parameter pos=(0, 1, 3, 2) The module performs feature compression in the height dimension, dynamically adjusts the vertical feature weights, and quantifies the importance of the vertical layer structure;

[0050] S54: The width branch first replaces the dimension of the fusion feature map with , and then use the convolution layer to capture the layered structure and gradient changes in the horizontal direction, and then use the dimension arrangement parameter pos=(0, 3, 2, 1) The module performs feature compression in the width dimension to quantify the importance of horizontal pore connectivity;

[0051] 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 dimensionally aligned feature map. ;

[0052] S56: SE attention module first input feature map Global information compression is performed along the spatial dimension to generate a global description of each channel. The weight of each channel is then generated through two fully connected layers and a nonlinear activation function. Finally, the input feature map is rescaled using the generated channel weights, 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.

[0053] Furthermore, the core image features of the channel dimension, height dimension and width dimension are aligned by channel, and cascaded and combined in the channel dimension in the order of channel, height and width to obtain comprehensive features, specifically including:

[0054] S61: Use a 1×1 convolutional layer to map the core image features of the channel dimension, height dimension, and width dimension from each branch to the same feature dimension to ensure channel alignment;

[0055] S62: The channel-aligned branch core image features are cascaded and combined in the channel dimension in the order of channel, height, and width to achieve organic fusion at the feature level and obtain comprehensive features.

[0056] In summary, the beneficial effects of the present invention are as follows:

[0057] The present invention provides a core permeability prediction method based on the interaction of time series decomposition and multi-dimensional features. The method accurately extracts key feature information from high-resolution three-dimensional core images to construct a final experimental data set, and performs spatiotemporal trend-residual decomposition on the core images in the final experimental data set. 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 of the channel dimension, height dimension and width dimension in the feature graph obtained by feature fusion, and then the core image features are respectively extracted for feature cascade combination. By fusing multi-dimensional features, the accuracy of permeability prediction can be effectively improved. In particular, when processing complex core images, the method shows higher prediction performance than traditional methods, and avoids the high cost and long period of physical experiments, meeting the current demand for efficient and low-cost permeability prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work, and these are all within the scope of protection of the present invention.

[0059] Figure 1 Schematic diagram of the process of the core permeability prediction method based on time series decomposition and multi-dimensional feature interaction of the present invention;

[0060] Figure 2 This is a model architecture diagram of the core permeability prediction method based on time series decomposition and multi-dimensional feature interaction. DETAILED DESCRIPTION

[0061] 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. 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 of the present invention and the embodiments can be combined with each other and are all within the scope of protection of the present invention.

[0062] Example 1: Reference Figure 1 As shown, Figure 1 This is a flow chart of the core permeability prediction method based on time series decomposition and multi-dimensional feature interaction of the present invention, the method comprising:

[0063] S1: Perform X-ray micro-tomography on multiple types of sandstone core samples and use a 3D reconstruction algorithm to process the collected projection data to obtain high-resolution 3D grayscale images of the core;

[0064] S2: Preprocess the core grayscale image to enhance the contrast between solid space and pores and fractures, and perform adaptive threshold segmentation to delineate pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image.

[0065] S3: The 3D core binary image is segmented using a sliding window sampling method to obtain multiple core cubes. The lattice Boltzmann method is used to simulate fluid flow and Darcy's law is used to calculate the permeability of each core cube as the corresponding data label to obtain the final experimental data set.

[0066] S4: Perform spatiotemporal 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;

[0067] S5: Using three parallel branch feature extractors to extract channel dimension, height dimension and width dimension core image features from the fused feature map respectively; wherein the feature extractor of each branch is composed of multiple layers of structured convolutional layers, SE attention modules, layer normalization layers and activation function stacks;

[0068] S6: performing channel alignment on the core image features in the channel dimension, height dimension, and width dimension, and performing cascade combination in the channel dimension in the order of channel, height, and width to obtain a comprehensive feature;

[0069] S7: A deep nonlinear transformation network composed of multi-layer convolution and activation functions is used to perform multi-layer mapping on the comprehensive features, gradually reducing the dimensionality of the comprehensive features to a permeability scalar output, and predicting the corresponding core permeability.

[0070] Furthermore, in an embodiment of the present invention, in step S1, X-ray micro-tomography is performed on multiple types of sandstone core samples, and the projection data collected by the scanning is processed using a three-dimensional reconstruction algorithm to obtain a high-resolution three-dimensional core grayscale image, which specifically includes:

[0071] S11: Sandstone core samples from various categories were selected, cut, and polished into standard cylindrical cores. This standard cylindrical shape ensures consistent experimental conditions, making results comparable across different laboratories or at different time points. This standardization helps improve the accuracy and reliability of scientific research.

[0072] S12: Perform multi-angle X-ray projection acquisition on multiple core samples and reconstruct the projection data using a filtered back-projection algorithm to obtain a high-resolution three-dimensional core grayscale image.

[0073] Filtered back-projection (FBP) is a widely used reconstruction algorithm in imaging technologies such as computed tomography (CT). This algorithm is based on the Radon transform of projection data and achieves image reconstruction through two main steps: filtering and back-projection. Specifically, the algorithm first projects rays onto the object from different angles to collect a series of projection data. These projection data are then filtered to eliminate artifacts and noise. Finally, the filtered projection data are back-projected back into the object space. This means that the filtered projection data are evenly distributed according to the path of the ray through the object at each projection angle, resulting in a preliminary reconstructed image. Through multiple back-projections and stacking, the original image can be inferred. The filtered back-projection algorithm has the advantages of high speed and relatively simple implementation. It can reconstruct an image of the object's internal structure from the projection data in a short time.

[0074] Furthermore, in the embodiment of the present invention, in step S2, image preprocessing is performed on the core grayscale image to enhance the contrast between solid space and pores and fractures, and adaptive threshold segmentation is performed to delineate pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image, which specifically includes:

[0075] S21: Perform median filtering with a window size of 9*9*9 on the obtained three-dimensional rock core grayscale image to denoise it, and use an interpolation method to fill in missing values ​​to obtain a smooth and complete rock core grayscale image.

[0076] The denoising processing formula in the embodiment of the present invention is as follows:

[0077]

[0078] Where, represents the new voxel value after median filtering, represents the voxel value of the original image, is the offset of the neighborhood voxel relative to the central voxel, and the offset range is , N is 9, and the median function Median() calculates the median of the selected neighborhood voxel set.

[0079] The interpolation formula of the embodiment of the present invention is as follows:

[0080]

[0081] Where, is the value of the point to be interpolated, are the coordinates of the center point of the interpolation, represents the weighted sum of all relevant voxel values ​​and their corresponding deviations, is the coefficient calculated from the surrounding voxel values, is the deviation of the interpolated point from the center point.

[0082] S22: First, the smooth and complete core grayscale image is divided into multiple blocks, the local grayscale histogram is calculated and the grayscale value limit is set, the cumulative distribution function of each block is calculated and normalized, and then the normalized result is mapped by the mapping function to obtain a new grayscale value. Finally, each block is spliced ​​into a complete image to obtain the enhanced core grayscale image.

[0083] The formula for adaptive histogram equalization in the embodiment of the present invention is as follows:

[0084]

[0085] Where, Indicates that the equalized three-dimensional image is at position The gray value of Indicates that the input image is at position The gray value of It is the grayscale histogram in the local area, indicating that the grayscale values ​​in the specified local window are Frequency of occurrence, Respectively represent the height, width, and depth considered when calculating the histogram, 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 the appropriate grayscale range.

[0086] Specifically, the adaptive threshold segmentation formula is as follows:

[0087]

[0088] Indicates that the image is at position The calculated adaptive threshold, Represents the original grayscale image at position The gray value of Represents surrounding pixels The relevant local neighborhood, represents the total number of pixels in the local area, It is an offset constant that controls the adaptive threshold and has a value range of 0.1 to 0.3.

[0089] S23: Adaptive threshold segmentation is performed on the enhanced grayscale image. First, the enhanced grayscale image is divided into multiple image blocks. Then, a local threshold is calculated based on the grayscale characteristics of each image block. Image blocks with pixels above the local threshold are marked as solid matrix, and image blocks with pixels below or equal to the local threshold are marked as pores. Each image block is then spliced ​​together to obtain a complete three-dimensional core binary image.

[0090] Furthermore, in the embodiment of the present invention, in step S3, a sliding window sampling method is used to block the three-dimensional core binary image to obtain multiple core cubes. At the same time, the fluid flow is simulated according to the lattice Boltzmann method and the permeability of each core cube is calculated as the corresponding data label in combination with Darcy's law to obtain the final experimental data set, which specifically includes:

[0091] S31: Using an overlapping sliding window, the three-dimensional core binary image is slid along the X / Y / Z spatial coordinate axes according to a specified step size to cut the image into several core cubes of specified scales.

[0092] S32: Use the lattice Boltzmann method (LBM method) to simulate fluid flow and combine it with 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 deep learning models.

[0093] The permeability formula for each core cube is calculated based on the lattice Boltzmann method to simulate fluid flow and Darcy's law as follows:

[0094]

[0095] Where, It 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, It represents the pressure difference of a fluid between two points. Represents the distance a fluid flows in a porous medium.

[0096] S33: The multi-sandstone dataset is augmented by first removing data points that significantly deviate from the mean value, and then performing random sampling within each permeability value layer to obtain the final experimental dataset.

[0097] Furthermore, in the embodiment of the present invention, in step S4, the core image in the final experimental data set is subjected to spatiotemporal trend-residual decomposition to extract global trend features and residual features, and the global trend features and residual features are fused to obtain a fused feature map, which specifically includes:

[0098] S41: Perform spatiotemporal trend-residual decomposition on the 3D core image sequence in the final experimental dataset. Sliding window smoothing is performed using a 3D average pooling layer to extract global trend features and capture macroscopic permeability patterns. Local detail features are separated as residual features through residual calculation, focusing on pore microstructure and local spatial variations. The spatiotemporal trend-residual decomposition formula is as follows:

[0099]

[0100] Where, is the global trend feature within the time window n, is the input image sequence of historical moments, is the residual feature at the current moment, n is the time window length of the sliding average, which controls the degree of smoothness;

[0101] S42: Global trend characteristics and residual features Perform data dimension replacement;

[0102] S43: Global trend characteristics Perform linear transformation of the channel dimension and activate the compressed time dimension. The formula is expressed as:

[0103]

[0104] in, is the learnable weight matrix, ReLU is the activation function, is the bias term, It is the global trend feature after channel conversion;

[0105] S44: Residual features Multi-scale convolution is performed, and convolution kernels of different sizes are used for convolution to extract pore structure characteristics. In the extraction process, both local pore throat and regional connectivity characteristics are taken into account. The pore structure characteristics extracted by convolution kernels of different sizes are then fused. The formula is expressed as:

[0106] + ;

[0107] in, Represents the residual feature after feature fusion, Indicates size The convolution kernel, Indicates size The convolution kernel, Indicates size The convolution kernel of

[0108] S45: Global trend characteristics and residual features Through skip connection fusion, a fused feature map that integrates global information and local details is obtained to form the encoder output:

[0109]

[0110] in, Represents the fused feature map.

[0111] Furthermore, in an embodiment of the present invention, in step S5, three parallel branch feature extractors are used to extract core image features in channel dimension, height dimension, and width dimension from the fused feature map, specifically including:

[0112] S51: The core image features in the channel dimension, height dimension, and width dimension are extracted from the fused feature map through three parallel branch feature extractors. The feature extractor of each branch consists of multiple layers of structured convolutional layers, SE attention modules, layer normalization layers, and activation function stacks. The core formula is expressed as follows:

[0113]

[0114] Where, Represents the tensor of the input fused feature map after dimension permutation; Represents SE attention modules customized for different branches, including module, Modules and module; Represents layer normalization operation; ReLU represents activation function; Conv2D represents multi-layer structured convolutional layer; Represents the core image features extracted from each branch;

[0115] S52: Channel branches maintain the spatial arrangement order of fused feature maps , extract global structural features through the convolution layer, and then use the dimension arrangement parameter pos=(0, 1, 2, 3) The module adaptively weights the channel dimensions to quantify the importance of different channels;

[0116] S53: The height branch first replaces the dimension of the fusion feature map with , and then use the convolution layer to capture the layered structure and gradient changes in the vertical direction, and then use the dimension arrangement parameter pos=(0, 1, 3, 2) The module performs feature compression in the height dimension, dynamically adjusts the vertical feature weights, and quantifies the importance of the vertical layer structure;

[0117] S54: The width branch first replaces the dimension of the fusion feature map with , and then use the convolution layer to capture the layered structure and gradient changes in the horizontal direction, and then use the dimension arrangement parameter pos=(0, 3, 2, 1) The module performs feature compression in the width dimension to quantify the importance of horizontal pore connectivity;

[0118] 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 dimensionally aligned feature map. ;

[0119] S56: SE attention module first input feature map Global information compression is performed along the spatial dimension to generate a global description of each channel. The weight of each channel is then generated through two fully connected layers and a nonlinear activation function. Finally, the input feature map is rescaled using the generated channel weights, 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.

[0120] 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, the input feature map is globally compressed along the spatial dimension to generate a global description of each channel. Then, the weight of each channel is generated through two fully connected layers and a nonlinear activation function. Finally, the generated channel weight is used to rescale the input feature map. The feature map of each channel is multiplied by the corresponding scaling factor, thereby enhancing important features and suppressing unimportant features.

[0121] The Squeeze step formula in the SE module is defined as follows:

[0122]

[0123] Where, Indicates channel The global information compression value, H and W represent the spatial dimensions of the feature map, Represents the input feature map Middle Channels in The value of the position;

[0124] The Excitation step formula in the SE module is defined as follows:

[0125]

[0126] Where, represents the weight coefficient of channel c, is the Sigmoid function, and is the learnable weight matrix, is the ReLU activation function, Indicates channel The global information compression value of

[0127] The Scale step formula in the SE module is defined as follows:

[0128]

[0129] Where, represents the output after attention, Indicates channel The weight coefficient ranges from 0 to 1 It is the feature map Channels in The value of the position

[0130] Furthermore, in step S6 of the embodiment of the present invention, the core image features of the channel dimension, height dimension, and width dimension are aligned by channel, and cascaded and combined in the channel dimension in the order of channel, height, and width to obtain comprehensive features, specifically including:

[0131] S61: Use a 1×1 convolutional layer to map the core image features of the channel dimension, height dimension, and width dimension from each branch to the same feature dimension to ensure channel alignment;

[0132] S62: The channel-aligned branch core image features are cascaded and combined in the channel dimension in the order of channel, height, and width to achieve organic fusion at the feature level and obtain comprehensive features.

[0133] Specifically, the present invention uses a deep nonlinear transformation network composed of multi-layer convolution and activation functions to process the cascaded comprehensive features, extracts high-level features through multi-layer mapping, and uses residual connections to enhance the expression ability of deep features. Finally, the fused features are gradually reduced to a permeability scalar output to predict the core permeability.

[0134] Specifically, in the embodiment of the present invention, refer to Figure 2 The model architecture shown in the diagram can also be used to construct a full-process prediction network model based on the aforementioned core permeability prediction method flow, including an encoder module and a decoder module (which can be further divided into a feature interaction module and a feature fusion module). After the input data is processed by the encoder and decoder to form fused features, the fused features are gradually reduced in dimension to a permeability scalar output through a fully connected layer, completing the end-to-end mapping from 3D images to physical parameters. The encoder corresponds to step S4 of the method, the feature interaction module in the decoder corresponds to step S5, and the feature fusion module corresponds to step S6.

[0135] Specifically, in order to evaluate the performance of the core permeability prediction model of this method, our model was compared with five models (Model1~Model5), in which the neural network architecture was implemented using the pytorch framework.

[0136] This experiment uses five indicators to measure model capability and performance, including the model's coefficient of determination (R 2 ), mean square error (RMSE); number of parameters (Params), in megabytes (M); number of floating-point operations (Flops), in megabytes (M); training time (TT), in minutes per round (min / epoch). All results are averaged over four experiments on the same device. RMSE and R 2 The calculation formula is as follows.

[0137]

[0138]

[0139] 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, It is the average value of the true value of n samples. RMSE reflects the average deviation between the model prediction value and the true value. The smaller the RMSE, the better the prediction. 2 Reflects the performance of model prediction, R 2 The larger the value, the better the model.

[0140] To analyze the permeability prediction performance of the spatiotemporal multidimensional collaborative fusion model, we conducted the following experiments based on the aforementioned experimental indicators. To verify the model's effectiveness, we conducted four experiments with different models and obtained the average values ​​of the corresponding RMSE and R². The experimental results are shown in Table 1.

[0141] Table 1 Comparison of model test results

[0142]

[0143] Therefore, this embodiment proposes a core permeability prediction method based on the interaction of time series decomposition and multi-dimensional features, making full use of the multi-dimensional information in the core image. Through its efficient feature fusion capability, it integrates spatial information, pore structure and other key physical features of different dimensions, thereby comprehensively capturing the fluid flow characteristics inside the core, overcoming the limitations of traditional methods in processing complex, multi-dimensional core data. In addition, compared with traditional methods that rely on physical experiments, this model can meet the prediction needs under different geological conditions, is more suitable for large-scale rock permeability prediction, can process massive data and perform batch predictions efficiently, and meets the market's actual demand for efficient and accurate predictions. At the same time, the model's memory usage and computing speed have been optimized, which can significantly reduce the consumption of computing resources while ensuring high-precision predictions, improve prediction efficiency, and ensure feasibility and real-time performance in large-scale application scenarios.

[0144] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to 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: include: S1: Perform X-ray micro-tomography on multiple types of sandstone core samples and use a 3D reconstruction algorithm to process the collected projection data to obtain high-resolution 3D grayscale images of the core; S2: Preprocess the core grayscale image to enhance the contrast between solid space and pores and fractures, and perform adaptive threshold segmentation to delineate pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image. S3: The 3D core binary image is segmented using a sliding window sampling method to obtain multiple core cubes. The lattice Boltzmann method is used to simulate fluid flow and Darcy's law is used to calculate the permeability of each core cube as the corresponding data label to obtain the final experimental data set. S4: Perform spatiotemporal 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: Using three parallel branch feature extractors to extract channel dimension, height dimension and width dimension core image features from the fused feature map respectively; wherein the feature extractor of each branch is composed of multiple layers of structured convolutional layers, SE attention modules, layer normalization layers and activation function stacks; S6: performing channel alignment on the core image features in the channel dimension, height dimension, and width dimension, and performing cascade combination in the channel dimension in the order of channel, height, and width to obtain a comprehensive feature; S7: A deep nonlinear transformation network composed of multi-layer convolution and activation functions is used to perform multi-layer mapping on the comprehensive features, gradually reducing the dimensionality of the comprehensive features to a permeability scalar output, and predicting the corresponding core permeability.

2. The core permeability prediction method according to claim 1, characterized in that: The X-ray micro-tomography is performed on multiple types of sandstone core samples, and the projection data collected by the scanning is processed using a three-dimensional reconstruction algorithm to obtain a high-resolution three-dimensional core grayscale image, which specifically includes: S11: Select multiple types of sandstone core samples, cut and polish them, and make them into standard cylindrical core samples; S12: Perform multi-angle X-ray projection acquisition on multiple core samples and reconstruct the projection data using a filtered back-projection algorithm to obtain a high-resolution three-dimensional core grayscale image.

3. The core permeability prediction method according to claim 1, characterized in that: The image preprocessing of the core grayscale image is performed to enhance the solid space and pore contrast, and adaptive threshold segmentation is performed to divide the pores and matrix in the three-dimensional core grayscale image to obtain a three-dimensional core binary image, specifically including: S21: performing median filtering with a window size of 9*9*9 on the obtained 3D rock core grayscale image to remove noise, and using an interpolation method to fill in missing values ​​to obtain a smooth and complete rock core grayscale image; S22: First, the smooth and complete core grayscale image is divided into multiple blocks, the local grayscale histogram is calculated and the grayscale value limit is set, the cumulative distribution function of each block is calculated and normalized, and then the normalized result is mapped by a mapping function to obtain a new grayscale value. Finally, each block is spliced ​​into a complete image to obtain an enhanced core grayscale image; S23: Adaptive threshold segmentation is performed on the enhanced grayscale image. First, the enhanced grayscale image is divided into multiple image blocks. Then, a local threshold is calculated based on the grayscale characteristics of each image block. Image blocks with pixels above the local threshold are marked as solid matrix, and image blocks with pixels below or equal to the local threshold are marked as pores. Each image block is then spliced ​​together to obtain a complete three-dimensional core binary image.

4. The core permeability prediction method according to claim 1, characterized in that: The sliding window sampling method is used to block the three-dimensional core binary image to obtain multiple core cubes. At the same time, the lattice Boltzmann method is used to simulate fluid flow and the permeability of each core cube is calculated in combination with Darcy's law as the corresponding data label to obtain the final experimental data set, which specifically includes: S31: Using an overlapping sliding window, the 3D core binary image is slid along the X / Y / Z spatial coordinate axes according to a specified step size to cut the image into several core cubes of specified scales; S32: Use the lattice Boltzmann method to simulate fluid flow and combine it with Darcy's law to calculate the permeability of each core cube. This is used as the data label corresponding to the core cube to obtain a multi-sandstone dataset for training the deep learning model. The permeability formula for each core cube is calculated based on the lattice Boltzmann method to simulate fluid flow and Darcy's law as follows: Where, It 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, It represents the pressure difference of a fluid between two points. Represents the distance the fluid flows in the porous medium; S33: The multi-sandstone dataset is augmented by first removing data points that significantly deviate from the mean value, and then performing random sampling within each permeability value layer to obtain the final experimental dataset.

5. The core permeability prediction method according to claim 1, characterized in that: The core images in the final experimental data set are subjected to spatiotemporal trend-residual decomposition to extract global trend features and residual features, and the global trend features and residual features are fused to obtain a fusion feature map, specifically including: S41: Perform spatiotemporal trend-residual decomposition on the 3D core image sequence in the final experimental dataset. Sliding window smoothing is performed using a 3D average pooling layer to extract global trend features and capture macroscopic permeability patterns. Local detail features are separated as residual features through residual calculation, focusing on pore microstructure and local spatial variations. The spatiotemporal trend-residual decomposition formula is as follows: Where, is the global trend feature within the time window n, is the input image sequence of historical moments, is the residual feature at the current moment, n is the time window length of the sliding average, which controls the degree of smoothness; S42: Global trend characteristics and residual features Perform data dimension replacement; S43: Global trend characteristics Perform linear transformation of the channel dimension and activate the compressed time dimension. The formula is expressed as: in, is the learnable weight matrix, ReLU is the activation function, is the bias term, It is the global trend feature after channel conversion; S44: Residual features Multi-scale convolution is performed, and convolution kernels of different sizes are used for convolution to extract pore structure characteristics. In the extraction process, both local pore throat and regional connectivity characteristics are taken into account. The pore structure characteristics extracted by convolution kernels of different sizes are then fused. The formula is expressed as: + ； in, Represents the residual feature after feature fusion, Indicates size The convolution kernel, Indicates size The convolution kernel, Indicates size The convolution kernel of S45: Global trend characteristics and residual features Through skip connection fusion, a fused feature map that combines global information and local details is obtained: in, Represents the fused feature map.

6. The core permeability prediction method according to claim 1, characterized in that: The feature extractor using three parallel branches extracts the core image features of the channel dimension, height dimension and width dimension from the fused feature map, specifically including: S51: The core image features in the channel dimension, height dimension, and width dimension are extracted from the fused feature map through three parallel branch feature extractors. The feature extractor of each branch consists of multiple layers of structured convolutional layers, SE attention modules, layer normalization layers, and activation function stacks. The core formula is expressed as follows: Where, Represents the tensor of the input fused feature map after dimension permutation; Represents SE attention modules customized for different branches, including module, Modules and module; Represents layer normalization operation; ReLU represents activation function; Conv2D represents multi-layer structured convolutional layer; Represents the core image features extracted from each branch; S52: Channel branches maintain the spatial arrangement order of fused feature maps , extract global structural features through the convolution layer, and then use the dimension arrangement parameter pos=(0, 1, 2, 3) The module adaptively weights the channel dimensions to quantify the importance of different channels; S53: The height branch first replaces the dimension of the fusion feature map into , and then use the convolution layer to capture the layered structure and gradient changes in the vertical direction, and then use the dimension arrangement parameter pos=(0, 1, 3, 2) The module performs feature compression in the height dimension, dynamically adjusts the vertical feature weights, and quantifies the importance of the vertical layer structure; S54: The width branch first replaces the dimension of the fusion feature map with , and then use the convolution layer to capture the horizontal layer structure and gradient changes, and then use the dimension arrangement parameter pos=(0, 3, 2, 1) The module performs feature compression in the width dimension to 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 dimensionally aligned feature map. ; S56: SE attention module first input feature map Global information compression is performed along the spatial dimension to generate a global description of each channel. The weight of each channel is then generated through two fully connected layers and a nonlinear activation function. Finally, the input feature map is rescaled using the generated channel weights, 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: The core image features of the channel dimension, height dimension and width dimension are aligned, and cascaded and combined in the channel dimension in the order of channel, height and width to obtain comprehensive features, specifically including: S61: Use a 1×1 convolutional layer to map the core image features of the channel dimension, height dimension, and width dimension from each branch to the same feature dimension to ensure channel alignment; S62: The channel-aligned branch core image features are cascaded and combined in the channel dimension in the order of channel, height, and width to achieve organic fusion at the feature level and obtain comprehensive features.

Images