Intelligent Analysis Method and System for Pore Structure in Core MicroCT Images

Abstract

This invention relates to the field of oil and gas reservoir development technology, and discloses an intelligent analysis method and system for pore structure in core micro-CT images. The method includes: acquiring three-dimensional image data from core micro-CT scans; performing noise filtering, grayscale correction, and artifact removal preprocessing; identifying pores, throats, and skeletons using a three-dimensional deep learning semantic segmentation network containing a hybrid attention module; extracting the pore network model using the maximum sphere algorithm combined with adaptive pore throat segmentation factors; calculating structural parameters such as porosity, coordination number, and tortuosity; analyzing connectivity and seepage channel characteristics; performing two-phase flow simulations with multiple displacement modes to predict displacement efficiency; identifying the residual oil occurrence state and outputting an analysis report. This invention achieves end-to-end intelligent analysis, improving pore segmentation accuracy and the reliability of displacement efficiency prediction.

Description

Technical Field

[0001] This invention relates to the field of oil and gas reservoir development and digital core analysis technology, specifically to a method and system for intelligent analysis of pore structure in core micro-CT images. Background Technology

[0002] Accurate characterization of reservoir core pore structure is a crucial foundation for evaluating reservoir properties and predicting development outcomes. Micro-CT scanning technology, due to its non-destructive testing and three-dimensional imaging capabilities, has become an important tool for studying core pore structure. However, existing micro-CT image analysis methods still face the following technical bottlenecks:

[0003] Traditional core pore space segmentation mainly relies on thresholding methods, such as the Otsu's method. While these methods are simple in principle and computationally efficient, they often struggle to accurately distinguish between micron-sized pores and the rock matrix when processing micro-CT images with complex grayscale distributions and blurred pore and skeleton boundaries, resulting in limited segmentation accuracy. Chinese invention CN119579835A discloses a method for selecting reasonable pore throat segmentation factors in a sandstone / conglomerate reservoir pore network model. This method uses the Otsu's method to perform binary segmentation of the three-dimensional grayscale image, then extracts the pore network model using the maximum sphere algorithm, and selects reasonable pore throat segmentation factors by comparing theoretical and measured permeability. Although this method provides a framework for selecting pore throat segmentation factors, its image segmentation still relies on traditional thresholding methods. For core images with complex pore boundaries and obvious grayscale transition regions, segmentation accuracy is difficult to guarantee, and it cannot achieve refined semantic differentiation of pores, throats, and the skeleton.

[0004] In constructing pore network models, existing methods generally employ fixed pore-throat segmentation factors to delineate pores and throats. CN119579835A points out that the selection of pore-throat segmentation factors significantly impacts the accuracy of pore network models; however, its ergonomic experimental comparison method is computationally expensive, and the selection of pore-throat segmentation factors is independent of the image segmentation process, lacking a feedback optimization mechanism for network extraction parameters based on segmentation accuracy. Furthermore, existing pore network analysis methods primarily focus on calculating static parameters such as porosity and pore-throat ratio, lacking a systematic approach to connectivity evaluation and seepage channel characteristic analysis.

[0005] In evaluating displacement efficiency, existing methods typically rely on pore network models to simulate flow under a single displacement mechanism, lacking the ability to compare and analyze multiple displacement mechanisms. This makes it difficult to provide microscopic mechanistic support for optimizing the design of enhanced oil recovery schemes. Furthermore, current residual oil distribution analysis mainly depends on manual observation and classification statistics, lacking intelligent identification methods based on morphological features. This prevents the quantitative characterization of residual oil occurrence and the automated analysis of its spatial distribution patterns.

[0006] In summary, existing methods for analyzing pore structure in core micro-CT images suffer from low segmentation accuracy, lack of synergistic parameter optimization, and incomplete evaluation of displacement efficiency, making it difficult to meet the needs of refined reservoir development analysis. Therefore, there is an urgent need for an intelligent analysis method and system capable of achieving high-precision pore space segmentation, adaptive pore network extraction, and comprehensive displacement efficiency evaluation. Summary of the Invention

[0007] To address the problems of low segmentation accuracy, lack of collaborative mechanism for parameter optimization, and incomplete evaluation of displacement efficiency in existing core micro-CT image pore structure analysis technologies, this invention provides an intelligent analysis method and system for core micro-CT image pore structure.

[0008] The first aspect of this invention provides an intelligent analysis method for pore structure in core micro-CT images, comprising the following steps: an image acquisition step, acquiring three-dimensional image data of a core sample via micro-CT scanning; an image preprocessing step, performing noise filtering, grayscale correction, and artifact removal on the three-dimensional image data of the core sample via micro-CT scanning to generate preprocessed three-dimensional image data; a pore space segmentation step, inputting the preprocessed three-dimensional image data into a three-dimensional deep learning semantic segmentation network for processing, the three-dimensional deep learning semantic segmentation network including an encoder path, a decoder path, and a hybrid attention module connecting the encoder path and the decoder path, outputting three-dimensional semantic segmentation result data, the three-dimensional semantic segmentation result data including pore voxel labels, throat voxel labels, and skeleton voxel labels; and a pore network extraction step, based on the three-dimensional semantic segmentation result data, extracting an equivalent pore network model using a combination of the maximum sphere algorithm and the median transformation method, wherein the pore network extraction step employs an adaptive pore-throat segmentation factor. The process includes: 1) Dynamically optimizing the adaptive pore-throat segmentation factor based on the spatial distribution characteristics of throat voxel labels in the 3D semantic segmentation results; 2) Calculating pore structure parameters based on an equivalent pore network model, including porosity, pore-throat ratio, coordination number, tortuosity, and pore size distribution parameters; 3) Calculating connectivity evaluation indices based on the equivalent pore network model and pore structure quantitative parameters, including connectivity index and seepage channel characteristic parameters; 4) Performing two-phase flow simulation based on the equivalent pore network model and connectivity evaluation indices, obtaining two-phase flow simulation results data, including sweep efficiency and oil displacement efficiency prediction results under different displacement methods; and 5) Identifying the residual oil occurrence state type based on the two-phase flow simulation results data and generating residual oil occurrence state distribution data, outputting a core pore structure quantitative analysis report and displacement efficiency evaluation conclusions.

[0009] A second aspect of this invention provides an intelligent analysis system for the pore structure of core micro-CT images, comprising: an image acquisition module for acquiring three-dimensional image data of core samples obtained by micro-CT scanning; an image preprocessing module for performing noise filtering, grayscale correction, and artifact removal processing on the three-dimensional image data of the core samples by micro-CT scanning, generating preprocessed three-dimensional image data; and a pore space segmentation module for processing the preprocessed three-dimensional image data using a three-dimensional deep learning semantic segmentation network, wherein the three-dimensional deep learning semantic segmentation network includes an encoder path, a decoder path, and a hybrid attention module, and outputs three-dimensional semantic segmentation result data including pore voxel labels, throat voxel labels, and skeleton voxel labels; pore space segmentation module. The network extraction module extracts an equivalent pore network model based on the 3D semantic segmentation results using the maximum sphere algorithm and the median transformation method, and uses an adaptive pore-throat segmentation factor to divide pores and throats. The parameter calculation module calculates quantitative pore structure parameters such as porosity, pore-throat ratio, coordination number, tortuosity, and pore size distribution based on the equivalent pore network model. The connectivity analysis module calculates connectivity evaluation indicators such as connectivity index and seepage channel characteristic parameters. The displacement simulation module performs two-phase flow simulation and outputs displacement efficiency prediction results including sweep efficiency and oil displacement efficiency. The residual oil analysis module identifies the residual oil occurrence state type and generates residual oil occurrence state distribution data and displacement efficiency evaluation conclusions.

[0010] The beneficial effects of this invention include: First, it employs a three-dimensional deep learning semantic segmentation network for pore space segmentation, enhancing the ability to identify micron-level pore boundaries through a hybrid attention module, resulting in significantly improved segmentation accuracy compared to traditional threshold segmentation methods; Second, it proposes an adaptive pore-throat segmentation factor optimization mechanism, dynamically adjusting the pore-throat segmentation factor based on the spatial distribution characteristics of throat voxels in the semantic segmentation results, achieving synergistic optimization of segmentation accuracy and network extraction parameters; Third, it constructs an end-to-end intelligent analysis framework from image acquisition to displacement efficiency evaluation, realizing integrated processing of quantitative characterization of pore structure, connectivity analysis, and displacement efficiency prediction; Fourth, it uses multi-scale morphological features to achieve intelligent identification and quantitative evaluation of residual oil occurrence status, providing data support for the study of the microscopic mechanisms of enhanced oil recovery schemes. Attached Figure Description

[0011] Figure 1 This is a flowchart of the intelligent analysis method for pore structure in core micro-CT images according to the present invention.

[0012] Figure 2 This is an architecture diagram of the intelligent analysis system for pore structure in core micro-CT images of the present invention. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0014] See Figure 1 As shown, the intelligent analysis method for pore structure of core micro-CT images provided by this invention includes the following steps:

[0015] Step S1: Image acquisition step.

[0016] In one embodiment of the present invention, the image acquisition step is used to acquire three-dimensional image data of a core sample via micro-CT scanning. Specifically, the core sample to be analyzed is placed on the sample stage of the micro-CT scanning device. A cone-shaped X-ray beam emitted by an X-ray source penetrates the core sample, and the transmitted X-ray signal is received by a detector. During the scanning process, the sample stage rotates 360° to acquire a sequence of projected images at different angles. Subsequently, a filtered back-projection algorithm is used to perform three-dimensional reconstruction of the projected image sequence to generate three-dimensional image data of the core sample via micro-CT scanning.

[0017] Preferably, the scanning resolution of the micro-CT scanning device is set to 0.5 μm to 10 μm, with the specific resolution selected based on the pore size characteristics of the core sample. For sandstone samples with larger pore sizes, a resolution of 5 μm to 10 μm can be used to obtain a larger field of view; for dense sandstone or carbonate rock samples with smaller pore sizes, a resolution of 0.5 μm to 2 μm is used to ensure effective identification of micron-sized pores. In this embodiment, for conglomerate reservoir samples, a voxel resolution of 5 μm is selected for scanning to obtain three-dimensional image data with a voxel size of 400×400×400, corresponding to a physical size of 2 mm×2 mm×2 mm.

[0018] Furthermore, the scanning voltage is set according to the mineral composition and density characteristics of the core sample, generally within the range of 80kV to 160kV. For sandstone samples mainly composed of quartz and feldspar, the scanning voltage is set to 100kV to 120kV; for samples containing high-density minerals such as pyrite and barite, the scanning voltage is appropriately increased to 140kV to 160kV to obtain sufficient penetration. The scanning current is set according to the image signal-to-noise ratio requirements, typically between 50μA and 200μA. Higher scanning current is beneficial for improving the image signal-to-noise ratio but increases the scanning time. In this embodiment, the scanning voltage is set to 120kV and the scanning current is set to 100μA.

[0019] In one embodiment of the present invention, the three-dimensional image data from micro-CT scanning is stored in a 16-bit grayscale format, with each voxel having a grayscale value ranging from 0 to 65535. Lower grayscale values ​​correspond to pore space, while higher grayscale values ​​correspond to the rock skeleton. The grayscale value distribution characteristics reflect the mineral composition and pore development degree of the core sample, providing a data basis for subsequent image preprocessing and pore space segmentation.

[0020] Step S2: Image preprocessing step.

[0021] In one embodiment of the present invention, the image preprocessing step performs noise filtering, grayscale correction, and artifact removal on the micro-CT scan three-dimensional image data of the core sample to generate preprocessed three-dimensional image data. Image preprocessing is a key step in ensuring the accuracy of subsequent pore space segmentation. Its purpose is to eliminate various noises and artifacts introduced during the scanning process and enhance the grayscale contrast between pores and the skeleton.

[0022] Noise filtering is implemented using a three-dimensional nonlocal mean filtering algorithm. Compared to traditional Gaussian filtering, nonlocal mean filtering can effectively remove random noise while preserving edge details in the image. The basic principle of nonlocal mean filtering is to estimate the denoised value of the current voxel using a weighted average of similar regions in the image, where the weights are determined by the similarity of the voxel's neighborhood. Specifically, for a voxel located at position (…), the noise is filtered using a three-dimensional nonlocal mean filtering algorithm. The voxels, their filtered values Calculated by the following formula: ,in: voxels The filtered grayscale value at the specified location is dimensionless. voxels The original grayscale value at that location is dimensionless. For voxels The central search window area; voxels Relative to voxels The weighting coefficients are dimensionless. This is the normalization factor, equal to the sum of all weight coefficients, and is dimensionless. Weight coefficients The weight is calculated based on the similarity between the neighborhoods of two voxels; the higher the similarity, the greater the weight. The weight coefficient is determined by the following formula: ,in: For voxels It is a vector composed of the gray values ​​of all voxels within the local neighborhood of the center. For voxels It is a vector composed of the gray values ​​of all voxels within the local neighborhood of the center. Represents Euclidean norm operations; This is the filter intensity parameter, with the same unit as the grayscale value. This parameter controls the degree of filtering, and its value ranges from 0.5 to 2 times the standard deviation of the image grayscale. In this embodiment, the search window size is set to 21×21×21 voxels, the local neighborhood size is set to 7×7×7 voxels, and the filter intensity parameter... Set to 1.2 times the standard deviation of image grayscale.

[0023] Gray-level correction is used to eliminate gray-level inhomogeneity caused by X-ray beam hardening. Beam hardening occurs because low-energy photons are preferentially absorbed when multi-energy X-rays penetrate a sample, resulting in a lower gray-level value in the image's central region compared to the edge regions. This invention employs a polynomial fitting method for gray-level correction. First, a homogeneous region is selected as a reference in the filtered image to establish a polynomial mapping relationship between position coordinates and gray-level deviation. Then, gray-level compensation is applied to the entire image. The gray-level correction formula is expressed as: ,in: voxels The value after grayscale correction is dimensionless. voxels Radial distance to the image center, in voxels; For the first The coefficients of the order polynomial are obtained by fitting the gray-level distribution of the reference region using the least squares method. The order of the polynomial is typically 2 to 4. In this embodiment, a second-order polynomial is used for grayscale correction, and the polynomial coefficients are obtained by fitting the grayscale distribution of the homogeneous region of the rock skeleton.

[0024] Artifact removal primarily targets the elimination of annular and stripe artifacts. Annular artifacts are caused by inconsistent detector pixel responses, manifesting as concentric annular grayscale anomalies centered at the rotation center. This invention employs a polar coordinate transformation combined with median filtering for annular artifact removal: first, the image is transformed from Cartesian coordinates to polar coordinates, where annular artifacts appear as vertical stripes; then, one-dimensional median filtering is performed along the angular direction to extract the artifact components; finally, the artifact components are subtracted from the original image and transformed back to Cartesian coordinates. Stripe artifacts are caused by fluctuations in X-ray source power or local high-density regions in the sample. This invention uses Fourier transform domain filtering to eliminate them, specifically identifying and suppressing the high-frequency narrowband signals corresponding to the stripe artifacts in the Fourier frequency domain.

[0025] After the above preprocessing, the grayscale values ​​of the generated preprocessed 3D image data are normalized to the range of 0 to 1, facilitating input processing by the subsequent 3D deep learning semantic segmentation network. The normalization formula is: ,in: The normalized voxel gray value, ranging from 0 to 1, is dimensionless. This represents the minimum grayscale value of the image after grayscale correction. This represents the maximum grayscale value of the image after grayscale correction.

[0026] Step S3: Pore space segmentation step.

[0027] In one embodiment of the present invention, the pore space segmentation step inputs the preprocessed 3D image data into a 3D deep learning semantic segmentation network for processing, and outputs 3D semantic segmentation result data. Unlike traditional threshold segmentation methods that can only distinguish between pores and skeletons, the 3D deep learning semantic segmentation network of the present invention can simultaneously identify three semantic labels: pores, throats, and skeletons, providing more refined boundary information for subsequent pore network extraction.

[0028] The 3D deep learning semantic segmentation network employs an encoder-decoder structure. The encoder path extracts multi-scale features, while the decoder path restores spatial resolution and outputs the segmentation result. The encoder path contains four downsampling stages, each consisting of a 3D convolutional layer, a batch normalization layer, an activation function layer, and a max-pooling layer. The 3D convolutional layer has a 3×3×3 kernel size and uses a zero-padding strategy to maintain the feature map size. The activation function is LeakyReLU. The max-pooling layer has a 2×2×2 kernel size with a stride of 2, used to reduce the spatial resolution of the feature map. The number of feature channels in the four downsampling stages are set to 32, 64, 128, and 256, respectively.

[0029] The decoder path also includes four upsampling stages, each consisting of a 3D transposed convolutional layer, skip connections, a 3D convolutional layer, and an activation function layer. The 3D transposed convolutional layer is used to increase the spatial resolution of the feature maps by 2 times. Skip connections concatenate the feature maps of the corresponding scale in the encoder path with the feature maps in the decoder path along the channel dimension to fuse low-level detailed features with high-level semantic features.

[0030] One of the core innovations of this invention lies in the design of the hybrid attention module. The hybrid attention module connects the encoder path and the decoder path, and includes two parallel branches: a spatial attention branch and a channel attention branch. The spatial attention branch learns the importance weights for different spatial locations, enhancing attention to the pore boundary region; the channel attention branch learns the importance weights for different feature channels, highlighting feature responses relevant to pore recognition.

[0031] The calculation process of the spatial attention branch is as follows: First, global average pooling and global max pooling are performed along the channel dimension of the input feature map to obtain two single-channel spatial descriptors; then, the two spatial descriptors are concatenated along the channel dimension, passed through a 3D convolutional layer with a kernel size of 7×7×7 and a sigmoid activation function, and the spatial attention weight map is output; finally, the spatial attention weight map is multiplied with the input feature map voxel by voxel to obtain the spatially enhanced feature map. The formula for calculating the spatial attention weights is: ,in: The feature map is after spatial attention enhancement; Input feature map; This represents a global average pooling operation along the channel dimension; This represents the global max pooling operation along the channel dimension; This indicates a channel-level concatenation operation; This represents a 3D convolution operation with a kernel size of 7×7×7; This represents the Sigmoid activation function; This indicates an element-wise multiplication operation.

[0032] The calculation process of the channel attention branch is as follows: First, global average pooling is performed on the input feature map to compress the spatial information of each channel into a single value, resulting in a channel description vector. Then, the channel description vector is passed sequentially through two fully connected layers. The first fully connected layer reduces the number of channels to 1 / 16 of the original number, while the second fully connected layer restores the original number of channels. A ReLU activation function is used between the two fully connected layers. Finally, a Sigmoid activation function is applied to the output vector to obtain the channel attention weights. These weights are multiplied by each channel of the input feature map to obtain the channel-enhanced feature map. The formula for calculating the channel attention weights is: ,in: The feature map is after channel attention enhancement; This represents a global average pooling operation, with an output dimension of... , Number of channels; Here is the weight matrix of the first fully connected layer, with dimension 1. ; This is the weight matrix of the second fully connected layer, with dimension 1. ; Represents the ReLU activation function; This indicates a channel-dimension broadcast multiplication operation.

[0033] The hybrid attention module weights and fuses the outputs of the spatial attention branch and the channel attention branch, with the fusion weights adaptively adjusted through learnable parameters. (Hybrid attention feature map) The calculation formula is: ,in: The fusion weights for the spatial attention branch are initially set to 0.5, with a range of 0 to 1, and are dimensionless. They are automatically learned and optimized during network training. The fusion weight for the channel attention branch is initially set to 0.5, with a value ranging from 0 to 1, and is dimensionless. Through the design of the hybrid attention module, the 3D deep learning semantic segmentation network can adaptively focus on the spatial features of the pore boundary region and the channel features related to pore recognition, thereby improving the segmentation accuracy of micron-sized pores.

[0034] The final layer of the decoder path is a classification layer, which uses a 1×1×1 3D convolution to map the number of feature channels to 3, corresponding to the three categories of pores, throats, and skeletons. After the classification layer, a Softmax activation function is used to normalize the three-channel output of each voxel into a probability distribution. Each voxel in the 3D semantic segmentation result data is assigned the category label with the highest probability.

[0035] Preferably, the training of the 3D deep learning semantic segmentation network uses a weighted combination of the cross-entropy loss function and the Dice loss function as the optimization objective. The cross-entropy loss function is used to optimize the voxel-by-voxel classification accuracy, and the Dice loss function is used to optimize the overlap of segmented regions. The combined loss function is expressed as: ,in: The total loss value is dimensionless. The cross-entropy loss value is dimensionless. The Dice loss value is dimensionless. is the weight coefficient for cross-entropy loss, ranging from 0 to 1; in this embodiment, it is set to 0.5. The network is trained using the Adam optimizer, with an initial learning rate set to... The batch size was set to 4, and the number of training iterations was set to 200 epochs.

[0036] In one embodiment of this invention, the training dataset for the 3D deep learning semantic segmentation network includes 50 sets of labeled core micro-CT images. The labeling was performed by technicians with a petrology background, using a semi-automatic labeling tool. The labeling results underwent cross-validation by multiple researchers to ensure labeling quality. Data augmentation strategies, including random rotation, random flipping, and random cropping, were employed during training to increase the diversity of the training data and improve the network's generalization ability. On the test dataset, the 3D deep learning semantic segmentation network achieved an average cross-union ratio (CUP) of 0.89, with a CUP of 0.91 for pores, 0.85 for throats, and 0.92 for skeletons, significantly outperforming the segmentation accuracy of traditional thresholding methods.

[0037] Step S4: Pore network extraction step.

[0038] In one embodiment of the present invention, the pore network extraction step is based on three-dimensional semantic segmentation results data, and uses a combination of the maximum sphere algorithm and the median transformation method to extract an equivalent pore network model. The core task of pore network extraction is to discretize the continuous pore space into a topological network structure composed of pore nodes and throat connections, providing a computational model for subsequent pore structure parameter calculation, connectivity analysis, and displacement simulation.

[0039] The basic principle of the maximum sphere algorithm is to fill the pore space with a maximum inscribed sphere, where the center of the sphere represents the location of the pore or throat, and the radius of the sphere represents the equivalent size of the pore or throat. The specific implementation process includes the following sub-steps: First, a binary representation of the pore space is extracted based on the pore voxel labels and throat voxel labels in the 3D semantic segmentation result data, where pore and throat voxels are labeled as 1, and skeleton voxels are labeled as 0. Then, the Euclidean distance transform is calculated on the binary pore space to obtain the distance value from each pore voxel to the nearest skeleton voxel; this distance value is the radius of the maximum inscribed sphere centered at that voxel. Next, a non-maximum suppression algorithm is used to extract local maxima points in the distance transform map as candidate locations for the maximum sphere center.

[0040] Preferably, the selection of the largest ball's center employs a hierarchical sorting strategy. First, the balls are sorted from largest to smallest radius, starting with the ball with the largest radius: if the center of the current ball is not covered by a selected ball, the ball is retained; if the center of the current ball is covered by a selected ball, the ball is deleted. This process is repeated until all candidate balls have been processed, ultimately obtaining a set of mutually exclusive largest balls. In this embodiment, balls with radii greater than a preset threshold are marked as pore nodes, and balls with radii less than the preset threshold are marked as throat nodes.

[0041] The mid-axis transformation method is used to extract the skeleton structure of the pore space, as a supplement to the maximum sphere algorithm. The result of the mid-axis transformation is the topological skeleton line of the pore space, where each point on the skeleton line is equidistant from the pore boundary. The mid-axis transformation is implemented through iterative erosion and skeleton refinement algorithms: first, a three-dimensional morphological erosion operation is performed on the pore space; then, the skeleton points generated during the erosion process are retained; the above process is repeated until the pore space is completely eroded, finally obtaining the mid-axis skeleton of the pore space. The bifurcation points of the mid-axis skeleton correspond to the positions of pore nodes, and the skeleton segments between the bifurcation points correspond to the positions of throats. The results of the mid-axis transformation are fused with the results of the maximum sphere algorithm, and the connection relationship between pore nodes and throats is established through spatial position matching to construct a complete equivalent pore network model.

[0042] One of the core innovations of this invention lies in the design of an adaptive pore throat segmentation factor. Traditional methods use a fixed pore throat segmentation factor. The boundary between pores and throats is defined by a pore-throat segmentation factor, which ranges from 0 to 1 and represents the proportion of length occupied by the throat between two adjacent pore nodes. The invention described above selects the optimal value by iterating through multiple pore-throat segmentation factor values ​​and comparing theoretical and measured permeability. However, this method is computationally expensive and independent of the image segmentation process.

[0043] This invention proposes dynamically optimizing the pore-throat segmentation factor based on the spatial distribution characteristics of throat voxel labels in the 3D semantic segmentation results data. Specifically, for two adjacent pore nodes... and The voxels connected along these lines were labeled as pore voxels or throat voxels by a 3D deep learning semantic segmentation network. The proportion of throat voxels on these lines to the total number of voxels was statistically analyzed. This ratio is used as the local pore throat segmentation factor for that pore pair. Global adaptive pore throat segmentation factor. The result is obtained by weighted averaging of local factors for all pore pairs: ,in: The adaptive pore throat segmentation factor has a value range of 0 to 1 and is dimensionless. It is the set of all pore node pairs in the equivalent pore network model; Pore ​​nodes and pore nodes The equivalent throat radius of the connecting line, in μm; Pore ​​nodes and pore nodes The proportion of throat voxels to total voxels along the connecting line, ranging from 0 to 1, is dimensionless. Through the above adaptive mechanism, the value of the pore-throat segmentation factor can reflect the actual spatial distribution characteristics of the throat region in the 3D semantic segmentation result, achieving synergistic optimization of image segmentation accuracy and pore network extraction parameters.

[0044] In this embodiment, the equivalent pore network model extracted based on the above method contains approximately 5000 to 6000 pore nodes and 9000 to 11000 throat connections. The equivalent radii of the pore nodes range from 5 μm to 200 μm, and the equivalent radii of the throats range from 2 μm to 80 μm. The calculated adaptive pore-throat segmentation factor is 0.28, which is consistent with the optimal value of 0.2 to 0.4 selected through experimental comparison in CN119579835A, verifying the effectiveness of the method of the present invention.

[0045] Step S5: Calculation steps for pore structure parameters.

[0046] In one embodiment of the present invention, the pore structure parameter calculation step is based on an equivalent pore network model to calculate quantitative pore structure parameters, including porosity, pore throat ratio, coordination number, tortuosity, and pore size distribution parameters. These parameters quantitatively characterize the pore structure features of the core from different perspectives, providing a basis for reservoir property evaluation and development scheme design.

[0047] Porosity Porosity is defined as the ratio of pore space volume to the total core volume, reflecting the core's storage capacity. Based on 3D semantic segmentation data and an equivalent pore network model, the formula for calculating porosity is: ,in: Porosity is expressed as a percentage, ranging from 0 to 40%. The total volume of the pore space is equal to the sum of the volumes of the pore voxels and throat voxels in the 3D semantic segmentation result data, in μm. 3 ; The total volume of the core sample is equal to the total number of primes in the 3D image data multiplied by the volume of a single voxel, in μm. 3 In this embodiment, the calculated porosity of the core sample is 8.66%.

[0048] The pore-throat ratio is defined as the ratio of the equivalent radius of a pore to the equivalent radius of the throat connecting that pore, reflecting the relative size relationship between the pore and throat dimensions. A larger pore-throat ratio indicates a stronger confinement effect of the throat on fluid flow. Each pore node in the equivalent pore network model... throat ratio Calculated by the following formula: ,in: Pore ​​nodes The pore-throat ratio is dimensionless. Pore ​​nodes The equivalent radius, in μm; For pore nodes The average equivalent radius of all connected throats, in μm. The global pore-throat ratio is the arithmetic mean of the pore-throat ratios of all pore nodes. In this embodiment, the average pore-throat ratio of the core sample is 3.5.

[0049] Coordination number The coordination number is defined as the number of throats connected to each pore node, reflecting the connectivity of the pore space. A higher coordination number indicates more connections between the pore node and surrounding pores, resulting in richer fluid flow paths. Pore nodes in the equivalent pore network model... coordination number Equal to the number of throats connected to that node. Global average coordination number. Calculated by the following formula: ,in: The mean coordination number is dimensionless and ranges from 1 to 20. This represents the total number of throats in the equivalent porous network model. This represents the total number of pore nodes in the equivalent pore network model. Multiplying by a factor of 2 is because each throat connects two pore nodes. In this embodiment, the average coordination number of the core sample is 3.7.

[0050] Degree of curvature Tortivity is defined as the ratio of the actual flow path length to the straight-line distance, reflecting the degree of tortuosity in the pore space. A greater tortuosity indicates greater fluid flow resistance. Tortivity is calculated based on the shortest path search in an equivalent pore network model. For flow from the inlet to the outlet face, Dijkstra's algorithm is first used to search for the shortest topological path from the inlet pore nodes to the outlet pore nodes; the path length is defined as the sum of the throat lengths traversed. Tortivity is calculated using the following formula: ,in: The degree of tortuosity is dimensionless and ranges from 1 to 10. This represents the shortest topological path length from the inlet face to the outlet face, in μm. The distance is the straight-line distance from the inlet to the outlet, in μm. In this embodiment, the tortuosity of the core sample along the three principal axes is 1.58, 1.62 and 1.55, respectively, with an average tortuosity of 1.58.

[0051] The pore size distribution parameters were obtained by calculating the equivalent radius distribution of all pore nodes in the statistical equivalent pore network model. The pore size distribution was fitted using a log-normal distribution, with fitting parameters including the geometric mean radius. and geometric standard deviation The probability density function of the aperture distribution is: ,in: Pore ​​radius The probability density, in μm⁻ 1 ; The equivalent radius of the pores is expressed in μm. The geometric mean radius is expressed in μm. The geometric standard deviation is dimensionless. In this embodiment, the geometric mean pore radius of the core sample is 28.5 μm, and the geometric standard deviation is 1.85.

[0052] Step S6: Connectivity analysis step.

[0053] In one embodiment of the present invention, the connectivity analysis step calculates connectivity evaluation indices based on an equivalent pore network model and quantitative parameters of pore structure. These indices include a connectivity index and characteristic parameters of seepage channels. Connectivity is a key factor affecting reservoir permeability; high connectivity means that fluids can flow smoothly within the pore space, which is beneficial for oil and gas extraction.

[0054] Connectivity Index The connectivity index is defined as the ratio of the volume of connected pores to the total pore volume, reflecting the overall connectivity of the pore space. A connected pore refers to the set of pore nodes that can connect from the inlet face to the outlet face. The calculation process of the connectivity index is as follows: First, a breadth-first search algorithm is used to traverse the equivalent pore network model starting from all pore nodes at the inlet face, marking all reachable pore nodes; then, it is determined whether the pore nodes at the outlet face are marked as reachable. If they are reachable, the pore node at that inlet face belongs to the connected pore; the above process is repeated to traverse all inlet face pore nodes, and the total volume of connected pores is counted. The connectivity index is calculated by the following formula: ,in: is the connectivity index, dimensionless, and ranges from 0 to 1; The total volume of the interconnected pores, in μm. 3 ; The total volume of the pore space, in μm. 3 The closer the connectivity index is to 1, the better the connectivity of the pore space. In this embodiment, the connectivity index of the core sample is 0.95, indicating that the vast majority of the pore space participates in effective seepage.

[0055] The characteristic parameters of seepage channels are used to describe the geometric and topological features of the main seepage paths. This invention defines a seepage channel as a set of paths connecting the inlet and outlet surfaces whose flow contribution rate exceeds a preset threshold. The identification process of seepage channels is as follows: First, a single-phase flow simulation is performed based on an equivalent pore network model to calculate the flow distribution of each throat; then, the throats are sorted from largest to smallest flow rate, and the set of throats with a cumulative flow contribution rate reaching 80% is identified as the main seepage channels. The characteristic parameters of seepage channels include:

[0056] Number of seepage channels The number of seepage channels is defined as the number of independent connected paths from the inlet to the outlet, reflecting the diversity of seepage paths. The number of seepage channels is calculated using the edge-connected component algorithm in graph theory.

[0057] Average length of seepage channel , defined as the arithmetic mean of the lengths of all seepage channels, in μm. The length of a seepage channel is equal to the sum of the lengths of the throats it passes through.

[0058] Average radius of seepage channels , defined as the harmonic average of the equivalent radii of the throat through which the seepage channel passes, is expressed in μm. The harmonic average is used because the flow resistance of the throat is inversely proportional to the fourth power of the radius, and the harmonic average better reflects the actual impact of the flow resistance.

[0059] In this embodiment, 12 main seepage channels were identified in the core sample, with an average length of 2.85 mm and an average radius of 15.3 μm.

[0060] Step S7: Displacement simulation step.

[0061] In one embodiment of the present invention, the displacement simulation step performs two-phase flow simulation based on an equivalent pore network model and connectivity evaluation index to obtain two-phase flow simulation results data. These results data include displacement efficiency predictions for sweep efficiency and oil displacement efficiency under different displacement methods. Two-phase flow simulation is a key technical means to predict reservoir development effectiveness, capable of revealing the distribution and variation patterns of fluids during the displacement process at a microscale.

[0062] The two-phase flow simulation employs a quasi-steady-state pore network simulation method. Basic assumptions include: the fluid is an incompressible Newtonian fluid; the flow process is laminar; gravity effects are neglected; the pressure distribution within the pores is uniform; and the flow is driven by both capillary forces and pressure gradients. The simulation process uses an intrusion percolation algorithm, where the fluid gradually intrudes into the pore space from the inlet surface, with the intrusion sequence determined by the capillary inlet pressure.

[0063] Capillary inlet pressure Calculated using the Young-Laplace equation: ,in: This represents the capillary inlet pressure, in Pa. This represents the interfacial tension between oil and water, expressed in N / m, with a typical range of 0.02 to 0.05 N / m. The contact angle is expressed in degrees. The contact angle of water-wet rocks is less than 90°, while the contact angle of oil-wet rocks is greater than 90°. The equivalent radius of the throat is expressed in meters (m). In this embodiment, the interfacial tension is set to 0.032 N / m, and the contact angle is set to 30° (under wet conditions).

[0064] Flow rate of fluid in the throat Calculated using the Hagen-Poiseuille equation: ,in: Throat flow rate, unit: m³ / s 3 / s; The equivalent radius of the throat is expressed in meters. This represents the pressure difference between the two ends of the throat, expressed in Pa. The viscosity of the fluid is expressed in Pa·s. This represents the length of the throat, measured in meters (m).

[0065] This invention simulates and compares three displacement methods: water flooding, polymer flooding, and surfactant flooding. In the water flooding simulation, the viscosity of the displacing phase is set to 1 mPa·s, and the viscosity of the displaced phase (crude oil) is set to 10 mPa·s. The polymer flooding simulation simulates the rheological properties of the polymer solution by increasing the viscosity of the displacing phase to 5 mPa·s. The surfactant flooding simulation simulates the interfacial tension-reducing effect of surfactants by lowering the interfacial tension to 0.005 N / m.

[0066] Sweep efficiency Defined as the ratio of the pore volume affected by the displaced phase to the total pore volume: ,in: The ripple efficiency is expressed as a percentage, ranging from 0 to 100%. The pore volume affected by the displaced phase, in μm. 3 ; The total volume of the pore space, in μm. 3 .

[0067] Oil displacement efficiency Defined as the ratio of the volume of the displaced phase to the initial volume of the displaced phase: ,in: Oil displacement efficiency is expressed as a percentage, ranging from 0 to 100%. The volume of the oil phase displaced is expressed in μm. 3 ; The initial volume of the oil phase is expressed in μm. 3 .

[0068] In this embodiment, simulation results show that: under water-driven conditions, the sweep efficiency is 68% and the oil displacement efficiency is 52%; under polymer-driven conditions, the sweep efficiency increases to 78% and the oil displacement efficiency increases to 61%; under surfactant-driven conditions, the sweep efficiency is 72% and the oil displacement efficiency is 58%. Polymer-driven systems improve the mobility ratio and expand the sweep range by increasing the viscosity of the displacing phase; surfactant-driven systems reduce capillary resistance and improve micro-displacement efficiency by lowering interfacial tension.

[0069] This invention also constructs a predictive model for displacement efficiency and pore structure parameters, establishing a multiple regression relationship between sweep efficiency, oil displacement efficiency, porosity, pore-throat ratio, coordination number, tortuosity, and connectivity index. The predictive model was fitted with 50 sets of training data, and the predictive coefficient of determination for sweep efficiency was determined. The predictive coefficient of determination for oil displacement efficiency reached 0.88. Reaching 0.85 provides a quantitative reference for the optimized design of enhanced oil recovery schemes.

[0070] Step S8: Residual oil analysis step.

[0071] In one embodiment of the present invention, the residual oil analysis step identifies the occurrence state type of residual oil based on two-phase flow simulation data and generates residual oil occurrence state distribution data, outputting a quantitative analysis report of core pore structure and a displacement efficiency evaluation conclusion. Residual oil analysis is a key step in studying the microscopic mechanism of enhanced oil recovery. By identifying the occurrence form and spatial distribution of residual oil, a theoretical basis can be provided for the optimization of subsequent displacement schemes.

[0072] The identification of residual oil occurrence state type is based on the shape factor classification method. Shape factor The regularity of the shape of residual oil clumps is quantitatively characterized by the following formula: ,

[0073] in: is the shape factor, dimensionless, and its value ranges from 0 to 1; The surface area of ​​the residual oil clumps, in μm. 2 ; The volume of the residual oil clumps is expressed in μm. 3 The smaller the shape factor, the more complex and irregular the shape of the residual oil clumps.

[0074] Based on the range of shape factor values, this invention classifies the residual oil occurrence state into three categories:

[0075] Network-like residual oil: shape factor Less than 0.1, the residual oil agglomerates exhibit a complex network distribution, occupying multiple interconnected pores and possessing a relatively large volume. This network-like residual oil is primarily distributed in the macroporous region and is the main type of residual oil in the later stages of waterflooding.

[0076] Porous residual oil: shape factor Between 0.1 and 0.3, residual oil agglomerates exhibit an irregular porous distribution, occupying several adjacent pores. Porous residual oil is a medium-sized residual oil type that can be effectively utilized by improving the mobility ratio or reducing interfacial tension.

[0077] Isolated residual oil: shape factor For values ​​greater than 0.3, residual oil agglomerates are distributed in approximately spherical or ellipsoidal shapes, existing in isolation within individual pores. Isolated residual oil is mainly distributed in small pores or pore corners and is the most difficult type of residual oil to utilize.

[0078] This invention further calculates the volume percentage and spatial distribution characteristics of various types of residual oil. The volume percentage reflects the contribution of residual oil in different occurrence states to the total residual oil saturation. The spatial distribution characteristics are characterized by the centroid coordinates and distribution density map of residual oil agglomerates.

[0079] In this embodiment, the residual oil saturation after water flooding is 48%, of which network-like residual oil accounts for 58%, porous residual oil accounts for 32%, and isolated residual oil accounts for 10%. After polymer flooding, the residual oil saturation decreases to 39%, the proportion of network-like residual oil decreases to 42%, the proportion of porous residual oil increases to 35%, and the proportion of isolated residual oil increases to 23%. This indicates that polymer flooding effectively utilizes the network-like residual oil in the macropores, but some of the network-like residual oil is separated into porous and isolated residual oil.

[0080] Finally, this invention outputs a quantitative analysis report of the core pore structure and an evaluation conclusion on displacement efficiency. The quantitative analysis report includes pore structure parameters such as porosity, pore-throat ratio, coordination number, tortuosity, pore size distribution, and connectivity index, along with their statistical distribution charts. The displacement efficiency evaluation conclusion includes comparisons of sweep efficiency and oil displacement efficiency under different displacement methods, as well as analysis results of the residual oil occurrence state distribution, providing a scientific basis for the study of the microscopic mechanisms and optimization design of enhanced oil recovery schemes.

[0081] See Figure 2 As shown, the intelligent analysis system for pore structure of core micro-CT images provided by the present invention includes an image acquisition module 1, an image preprocessing module 2, a pore space segmentation module 3, a pore network extraction module 4, a parameter calculation module 5, a connectivity analysis module 6, a displacement simulation module 7, and a residual oil analysis module 8.

[0082] Image acquisition module 1 is used to acquire three-dimensional image data of core samples via micro-CT scanning. In one embodiment of the present invention, the image acquisition module is communicatively connected to a micro-CT scanning device, receives the projected image sequence output by the micro-CT scanning device, and calls a three-dimensional reconstruction algorithm to generate three-dimensional image data of the core sample via micro-CT scanning. The scanning parameter configuration function of the image acquisition module allows users to set the scanning resolution, scanning voltage, and scanning current according to the characteristics of the core sample. The scanning resolution is adjustable from 0.5μm to 10μm, and the scanning voltage is adjustable from 80kV to 160kV.

[0083] Image preprocessing module 2 is used to perform noise filtering, grayscale correction, and artifact removal on the micro-CT scan three-dimensional image data of the core sample, generating preprocessed three-dimensional image data. In one embodiment of the present invention, the image preprocessing module integrates a three-dimensional nonlocal mean filtering algorithm, a polynomial grayscale correction algorithm, and a polar coordinate transformation annular artifact removal algorithm. The parameters of each algorithm can be user-defined or automatically optimized using default settings. Image preprocessing module 1 normalizes the processed image and outputs normalized preprocessed three-dimensional image data.

[0084] The pore space segmentation module 3 is used to process the preprocessed 3D image data using a 3D deep learning semantic segmentation network, outputting 3D semantic segmentation result data containing pore voxel labels, throat voxel labels, and skeleton voxel labels. In one embodiment of the present invention, the pore space segmentation module has a built-in trained 3D deep learning semantic segmentation network model. The 3D deep learning semantic segmentation network includes an encoder path, a decoder path, and a hybrid attention module connecting the encoder path and the decoder path. The hybrid attention module includes a spatial attention branch and a channel attention branch. The specific working principle of the pore space segmentation module 3 is consistent with step S3 in the method embodiment.

[0085] The pore network extraction module 4 is used to extract an equivalent pore network model based on the 3D semantic segmentation result data using the maximum sphere algorithm and the median transformation method, and to divide pores and throats using an adaptive pore-throat segmentation factor. In one embodiment of the present invention, the pore network extraction module includes a maximum sphere algorithm calculation unit, a median transformation calculation unit, and an adaptive factor optimization unit. The maximum sphere algorithm calculation unit extracts a set of maximum spheres based on distance transformation and non-maximum suppression. The median transformation calculation unit extracts the median skeleton of the pore space based on morphological erosion and skeleton refinement. The adaptive factor optimization unit calculates the adaptive pore-throat segmentation factor according to the spatial distribution characteristics of the throat voxel labels in the 3D semantic segmentation result data, and the calculation principle is consistent with step S4 in the method embodiment.

[0086] The parameter calculation module 5 is used to calculate quantitative parameters of the pore structure, including porosity, pore throat ratio, coordination number, tortuosity, and pore size distribution, based on an equivalent pore network model. In one embodiment of the present invention, the parameter calculation module includes a porosity calculation unit, a pore throat ratio calculation unit, a coordination number calculation unit, a tortuosity calculation unit, and a pore size distribution fitting unit. The calculation formulas of each calculation unit are consistent with those described in step S5 of the method embodiment. The parameter calculation module 5 outputs the quantitative parameters of the pore structure and their statistical distribution charts.

[0087] The connectivity analysis module 6 is used to calculate connectivity evaluation indicators based on connectivity indices and seepage channel characteristic parameters. In one embodiment of the present invention, the connectivity analysis module 6 includes a connectivity index calculation unit and a seepage channel analysis unit. The connectivity index calculation unit uses a breadth-first search algorithm to count the volume of connected pores and calculate the connectivity index. The seepage channel analysis unit identifies the main seepage channels based on single-phase flow simulation and counts characteristic parameters such as the number, average length, and average radius of seepage channels. The calculation principle is consistent with step S6 in the method embodiment.

[0088] The displacement simulation module 7 is used to perform two-phase flow simulation and output displacement efficiency prediction results including sweep efficiency and oil displacement efficiency. In one embodiment of the present invention, the displacement simulation module adopts a quasi-steady-state pore network simulation method, supporting simulation comparisons of various displacement methods such as water flooding, polymer flooding, and surfactant flooding. The fluid parameter configuration interface of the displacement simulation module allows users to set parameters such as interfacial tension, contact angle, displacement phase viscosity, and displaced phase viscosity. The displacement simulation module 7 outputs displacement efficiency prediction results for sweep efficiency and oil displacement efficiency under different displacement methods, and the calculation principle is consistent with step S7 in the method embodiment.

[0089] The residual oil analysis module 8 is used to identify the residual oil occurrence state type and generate residual oil occurrence state distribution data and displacement efficiency evaluation conclusions. In one embodiment of the present invention, the residual oil analysis module includes a shape factor calculation unit, an occurrence state classification unit, and a report generation unit. The shape factor calculation unit calculates the shape factor based on the surface area and volume of the residual oil agglomerates. The occurrence state classification unit classifies the residual oil into three occurrence types—network, porous, and isolated—according to the shape factor values. The report generation unit summarizes the quantitative parameters of pore structure, the connectivity evaluation index, the displacement efficiency prediction results, and the residual oil occurrence state distribution data to generate a core pore structure quantitative analysis report and displacement efficiency evaluation conclusions. The output format supports multiple formats such as PDF, Word, and Excel.

[0090] The system of this invention transmits data between its various modules via a data bus, forming a complete data processing pipeline. The output data of image acquisition module 1 is transmitted to image preprocessing module 2; the preprocessed 3D image data is transmitted to pore space segmentation module 3; the 3D semantic segmentation results are transmitted to pore network extraction module 4; the equivalent pore network model data is transmitted to parameter calculation module 5, connectivity analysis module 6, and displacement simulation module 7, respectively; and the two-phase flow simulation results are transmitted to residual oil analysis module 8 for comprehensive analysis. The entire system achieves end-to-end automated processing from micro-CT image acquisition to displacement efficiency evaluation.

[0091] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for intelligent analysis of pore structure in core micro-CT images, characterized in that, Includes the following steps: The image acquisition step acquires three-dimensional image data of the core sample via micro-CT scanning; the image preprocessing step performs noise filtering, grayscale correction, and artifact removal on the three-dimensional image data of the core sample via micro-CT scanning to generate preprocessed three-dimensional image data. The pore space segmentation step involves inputting the preprocessed 3D image data into a 3D deep learning semantic segmentation network for processing. The 3D deep learning semantic segmentation network includes an encoder path, a decoder path, and a hybrid attention module connecting the encoder path and the decoder path. It outputs 3D semantic segmentation result data, which includes pore voxel labels, throat voxel labels, and skeleton voxel labels. The porosity network extraction step involves extracting an equivalent porosity network model based on the three-dimensional semantic segmentation result data using a combination of the maximum sphere algorithm and the median transformation method. The porosity network extraction step employs an adaptive pore throat segmentation factor, which is dynamically optimized based on the spatial distribution characteristics of the throat voxel labels in the three-dimensional semantic segmentation result data. The steps include: calculating pore structure parameters based on the equivalent pore network model, including porosity, pore-throat ratio, coordination number, tortuosity, and pore size distribution parameters; calculating connectivity evaluation indices based on the equivalent pore network model and the quantitative pore structure parameters, including connectivity index and seepage channel characteristic parameters; and performing two-phase flow simulation based on the equivalent pore network model and the connectivity evaluation indices to obtain two-phase flow simulation results data, including sweep efficiency and oil displacement efficiency prediction results under different displacement methods. For two adjacent pore nodes and The voxels on the connecting lines are labeled as pore voxels or throat voxels by a 3D deep learning semantic segmentation network, and the proportion of throat voxels on the connecting lines to the total number of voxels is counted. This ratio is used as the local pore throat segmentation factor for the pore pair, and the global adaptive pore throat segmentation factor. The result is obtained by weighted averaging of local factors for all pore pairs: ,in: The adaptive pore throat segmentation factor has a value range of 0 to 1 and is dimensionless. It is the set of all pore node pairs in the equivalent pore network model; Pore ​​nodes and pore nodes The equivalent throat radius of the connecting line, in μm; Pore ​​nodes and pore nodes The proportion of the larynx voxels in the total voxels is the value of the voxels in the line connecting the lines, ranging from 0 to 1, and is dimensionless.

2. The intelligent analysis method for pore structure of core micro-CT images according to claim 1, characterized in that, The voxel resolution of the micro-CT scan in the image acquisition step is 0.5μm to 10μm. The noise filtering process adopts a three-dimensional nonlocal mean filtering algorithm, and the grayscale correction process adopts a polynomial fitting method to eliminate the X-ray beam hardening effect.

3. The intelligent analysis method for pore structure of core micro-CT images according to claim 1, characterized in that, The hybrid attention module includes a spatial attention branch and a channel attention branch. The spatial attention branch is used to learn the importance weights of different spatial locations, and the channel attention branch is used to learn the importance weights of different feature channels. The outputs of the spatial attention branch and the channel attention branch are weighted and fused through learnable parameters.

4. The intelligent analysis method for pore structure of core micro-CT images according to claim 1, characterized in that, The training of the three-dimensional deep learning semantic segmentation network uses a weighted combination of cross-entropy loss function and Dice loss function as the optimization objective. The encoder path contains four downsampling stages, and the decoder path contains four upsampling stages. Skip connections are set between the downsampling stages and the upsampling stages.

5. The intelligent analysis method for pore structure of core micro-CT images according to claim 1, characterized in that, The tortuosity is calculated based on Dijkstra's algorithm to search for the shortest topological path from the inlet face pore node to the outlet face pore node, and the connectivity index is calculated based on the breadth-first search algorithm to calculate the ratio of the pore volume connected from the inlet face to the outlet face to the total pore volume.

6. The intelligent analysis method for pore structure of core micro-CT images according to claim 1, characterized in that, The two-phase flow simulation employs a quasi-steady-state pore network simulation method. The capillary inlet pressure is calculated using the Young-Laplace equation, and the fluid flow rate in the throat is calculated using the Hagen-Poiseuille equation. The different displacement methods include water drive, polymer drive, and surfactant drive.

7. The intelligent analysis method for pore structure of core micro-CT images according to claim 1, characterized in that, It also includes a residual oil analysis step, which identifies the occurrence state type of residual oil based on the two-phase flow simulation results data and generates residual oil occurrence state distribution data, outputs a quantitative analysis report of core pore structure and a displacement efficiency evaluation conclusion. The identification of the occurrence state type of residual oil is based on the shape factor classification. The shape factor is calculated according to the ratio of the surface area to the volume of residual oil agglomerates. The occurrence state types include network residual oil, porous residual oil and isolated residual oil.

8. The intelligent analysis method for pore structure of core micro-CT images according to claim 7, characterized in that, When the shape factor is less than 0.1, it is classified as network-like residual oil; when the shape factor is between 0.1 and 0.3, it is classified as porous residual oil; when the shape factor is greater than 0.3, it is classified as isolated residual oil. The residual oil occurrence state distribution data includes the volume percentage and spatial distribution characteristics of each type of residual oil.

9. A core micro-CT image pore structure intelligent analysis system, used to implement the core micro-CT image pore structure intelligent analysis method as described in claim 8, characterized in that, include: The image acquisition module is used to acquire three-dimensional image data of core samples via micro-CT scanning. The image preprocessing module is used to perform noise filtering, grayscale correction and artifact removal on the micro-CT scan three-dimensional image data of the core sample to generate preprocessed three-dimensional image data. The system includes the following modules: a pore space segmentation module, which processes the preprocessed 3D image data using a 3D deep learning semantic segmentation network (including encoder path, decoder path, and hybrid attention module), and outputs 3D semantic segmentation results containing pore voxel labels, throat voxel labels, and skeleton voxel labels; a pore network extraction module, which extracts an equivalent pore network model based on the 3D semantic segmentation results using the maximum sphere algorithm and the median transformation method, and uses an adaptive pore-throat segmentation factor to divide pores and throats; a parameter calculation module, which calculates quantitative pore structure parameters such as porosity, pore-throat ratio, coordination number, tortuosity, and pore size distribution based on the equivalent pore network model; a connectivity analysis module, which calculates connectivity evaluation indicators such as connectivity index and seepage channel characteristic parameters; a displacement simulation module, which performs two-phase flow simulation and outputs displacement efficiency prediction results including sweep efficiency and oil displacement efficiency; and a residual oil analysis module, which identifies the residual oil occurrence state type and generates residual oil occurrence state distribution data and displacement efficiency evaluation conclusions.

Images