An oil displacement rate prediction method based on reservoir displacement etching image

Abstract

The application discloses a method for predicting oil displacement rate based on a reservoir displacement etching image and belongs to the field of reservoir displacement etching oil displacement. In the feature coding stage of the network, a light deformable feature coding module is designed, feature extraction is performed on the water-oil image with rich shape and fuzzy edge through the light deformable feature coding module, the light deformable convolution module can not only improve the feature expression ability of the network by using the bias layer, but also effectively reduce the network parameter quantity through the depth separable convolution. In the feature fusion stage, the channel attention is decomposed through the synergistic coupling attention module, the features are aggregated along two spatial directions respectively, the feature map is coded into a direction and position perception feature map, the position information is embedded into the channel attention feature map, the robustness of the network model is improved, the segmentation precision of the etching image is effectively improved, the prediction error of the oil displacement rate is reduced, and therefore the prediction precision and efficiency of the oil displacement rate of the reservoir displacement etching image are improved.

Description

Technical Field

[0001] This invention belongs to the field of reservoir displacement etching for oil recovery, and relates to a method for predicting the oil recovery rate of reservoir displacement etching images based on image processing technology and pattern recognition. Background Technology

[0002] With the rapid development of the national economy, the demand for energy in domestic production and daily life is increasing day by day, and oil has become the energy source with the largest demand in my country. However, my country's oil fields are gradually entering the medium-to-high water-cut stage, making extraction extremely difficult. Using deionized water to drive residual oil is one of the most effective and environmentally friendly reservoir displacement methods. In recent years, in order to further improve energy utilization efficiency, clarify the current status of water-driven residual oil extraction, and enhance the international competitiveness of my country's petroleum resources, many scholars at home and abroad have been constantly exploring new methods to study the formation and distribution mechanism of water-driven residual oil. Among them, the most widely used technologies are CT (Computed Tomography) scanning experiments and microscopic simulation models.

[0003] CT scanning experiments utilize X-rays to construct three-dimensional shapes by analyzing the different absorption rates of rock samples and pores. By identifying rock layer properties, the characteristics of water-driven oil displacement can be further analyzed, achieving displacement experiments without altering the core state. However, CT scanning experiments typically face challenges such as low scanning resolution and difficulty in acquiring experimental subjects, leading to significant data errors. Therefore, researchers widely employ microscopic simulation models. This method combines a glass etching simulation model with a microscopic image acquisition system. Using numerical simulation testing, rocks are reconstructed on a 120mm × 120mm etched plate primarily composed of polydimethylsiloxane (PDMS) to simulate displacement in different rock layers. The water-driven oil displacement process is simulated using deionized water and mineral oil, achieving dynamic visualization of microscopic displacement. To better present the experimental results, the aqueous phase is typically stained blue with methylene blue during the displacement process, while oil red is dissolved in the oil phase, resulting in a red color. By analyzing the acquired images, the effects of various factors such as different fluid viscosities, different injection rates, and different initial water saturation on the waterflooding effect can be calculated. The distribution patterns of oil and water before and after waterflooding can be analyzed, which is of great significance for the study of the waterflooding mechanism in oil reservoirs.

[0004] Currently, researchers typically use manual labeling to separate oil and water pixels, calculating the oil displacement rate by marking oil-bearing and water-bearing regions. This method is highly subjective, time-consuming, and labor-intensive, causing significant challenges to subsequent research. Due to the influence of the imaging environment and chip material, the acquired displacement etching images often suffer from problems such as image blurring, low color contrast, and high noise. Furthermore, because water and oil often blend together during displacement, the water-oil boundary is unclear and exhibits a non-rigid state. Based on these characteristics, researchers usually use image segmentation techniques to count the number of pixels in the water and oil phases, and then calculate the oil displacement rate. Compared to manual labeling methods, image segmentation is more convenient and effective, providing technical support for analyzing reservoir displacement mechanisms.

[0005] While current deep learning can solve some of the challenges in image segmentation, its application to oil reservoir etching image segmentation still faces two problems. First, the fixed-geometric convolutional kernels used in conventional segmentation networks are ill-suited to the complex shapes and structures of etching images. The fixed receptive field of standard convolution significantly reduces the network's ability to model geometric transformations, resulting in poor segmentation of small targets such as water-oil interface regions. Although deformable convolutions have been applied in image segmentation, the massive computational cost and number of parameters make network fitting difficult. Second, current attention networks use channel and spatial attention to suppress noise, but they do not consider the feature relationships between channels and space, leading to an over-reliance on spatial or channel information and coarse segmentation results. Furthermore, most existing hybrid attention models use global average pooling, where each location of the feature map after convolution contains information from a local region of the original image. Therefore, these methods only consider local features and cannot model the long-range dependencies necessary for oil reservoir visualization tasks. Summary of the Invention

[0006] The main objective of this invention is to provide a method for predicting oil displacement rate based on reservoir displacement etching images. In the feature encoding stage of the network, a lightweight deformable feature encoding module is designed. This module extracts features from water-oil images with rich shapes and blurred edges. The lightweight deformable convolution module not only improves the network's feature representation ability using bias layers but also effectively reduces the number of network parameters through depthwise separable convolution. In the feature fusion stage, a collaboratively coupled attention module decomposes the channel attention, aggregating features along two spatial directions respectively. The feature map is encoded into a direction- and position-aware feature map, thereby embedding positional information into the channel attention feature map, improving the robustness of the network model, effectively increasing the segmentation accuracy of the etching image, reducing the oil displacement rate prediction error, and thus improving the accuracy and efficiency of oil displacement rate prediction from reservoir displacement etching images.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] This invention discloses a method for predicting oil displacement rate based on reservoir displacement etching images. The method uses the Naturalness Preserved Enhancement (NPE) algorithm to enhance the data of reservoir displacement etching images directly acquired from the equipment, improving data clarity and contrast. Cropping and rotation operations are used to perform data enhancement on the reservoir displacement etching image data, preprocessing the data to obtain a reservoir displacement etching image dataset. This dataset is then fed into a Lightweight Deformable Feather Encoder and Decoder Network (LDNet) model for etching image segmentation. In the model's feature encoding stage, a Lightweight Deformable Feature Encoding (LDFE) module is designed. This module adds an offset variable to the position of each sampling point in the convolutional kernel to extract features from reservoir displacement etching images with rich shapes and blurred edges. This approach not only leverages bias layers to improve the network's feature representation capabilities and acquire geometric information about the water-oil mixture, but also effectively reduces the number of network parameters and memory consumption through depthwise separable convolutions, facilitating subsequent engineering applications and deployments. In the feature fusion stage, a collaboratively coupled attention module is designed to decompose channel attention, aggregating features along two spatial directions. The feature maps are encoded as orientation- and position-aware feature maps, embedding positional information into the channel attention feature maps to obtain feature-rich global dependencies and long-range contextual information of the reservoir image, thereby improving the network's segmentation accuracy. A residual bottleneck module is used to fuse higher-level semantic information to improve the model's feature representation capabilities for etched images. Based on a lightweight deformable feature encoding module, a collaboratively coupled attention module, and an upsampling module, a reservoir displacement etched image segmentation model is constructed. This improves the training and inference speed of the reservoir displacement etched image segmentation model, reduces computational costs and memory requirements, and ultimately enhances the segmentation accuracy, efficiency, and robustness of the reservoir displacement etched image segmentation model. The oil reservoir displacement etching image segmentation model is used for image segmentation, which automatically marks water-bearing and oil-bearing regions with high accuracy and efficiency. The number of marked oil-bearing and water-bearing pixels is used to predict the oil displacement rate, thereby reducing the oil displacement rate prediction error and improving the accuracy, efficiency and robustness of oil displacement rate prediction in the etching image.

[0009] The present invention discloses a method for predicting the oil displacement rate of reservoir displacement etching images, comprising the following steps:

[0010] Step 1: The acquired reservoir displacement etching image data is cropped, and image enhancement and data enhancement operations are performed on the etching image data to achieve reservoir displacement etching image data preprocessing and obtain reservoir displacement etching image dataset.

[0011] The image enhancement process involves performing a logarithmic transformation on the image to adjust the detail information of the etched image, resulting in an enhanced etched image.

[0012] The data augmentation includes horizontal flipping, rotation, and adding noise.

[0013] To reduce the impact of uneven image illumination, the image enhancement preferably uses the NPE (naturalness preserved enhancement) algorithm for image preprocessing. The algorithm formula is shown in (1). It uses a brightness pass filter Q(·) to decompose the image into reflectance and illuminance, which determine the detail and naturalness of the image respectively. Then, a double logarithmic transformation U(·) is used to map the illumination to achieve a balance between detail and naturalness, thereby preserving the details of the oil reservoir displacement etching image under non-uniform illumination in the experiment, and maintaining the naturalness of the oil phase and water phase images.

[0014]

[0015] Step Two: In the feature encoding stage, a Lightweight Deformable Feature Encoding (LDFE) module is designed. This module adds an offset variable to the position of each sampling point in the convolutional kernel to extract features from oil reservoir displacement etching images with rich shapes and blurred edges. Compared to conventional convolution, the LDFE module not only improves the network's feature representation ability using bias layers but also effectively reduces the number of network parameters through depthwise separable convolution. In the feature fusion stage, a co-coupled attention module is designed to decompose channel attention, aggregating features along two spatial directions respectively. This encodes the feature maps into orientation- and position-aware feature maps, thus embedding positional information into the channel attention feature maps. A residual bottleneck module is used to fuse higher-level semantic information to improve the model's feature representation ability for etching images. The LDNet reservoir displacement etching image segmentation model is constructed based on a lightweight deformable feature encoding module, a collaboratively coupled attention module, and an upsampling module. This improves the training and inference speed of the reservoir displacement etching image segmentation model, reduces computational costs and memory requirements, and thus enhances the segmentation accuracy, efficiency, and robustness of the reservoir displacement etching image segmentation model.

[0016] Step 2.1: In the feature encoding stage, a lightweight deformable feature encoding module is designed. By adding an offset variable to the position of each sampling point in the convolution kernel, it is used to extract features from reservoir displacement and etching images with rich shapes and blurred edges. Compared with conventional convolution, the lightweight deformable convolution module can not only improve the feature representation ability of the network by utilizing bias layers, but also effectively reduce the number of network parameters through depthwise separable convolution.

[0017] The complex background of etched images and the non-rigid boundaries between the water and oil phases in oil reservoir etched images make segmentation difficult. Conventional models mostly use square convolutional kernels with fixed shapes for feature learning. However, the receptive field provided by convolutional kernels with fixed shapes is very limited, resulting in weak ability to extract contextual information from etched images, which significantly restricts the ability to extract multi-scale information from oil reservoir etched images with rich and complex shapes.

[0018] To adapt to the characteristics of non-rigid aqueous and oil phases and reduce the parameter redundancy and training difficulties caused by increasing bias, a lightweight deformable feature encoding module is designed in the feature encoding stage. This module includes convolutional layers, bias layers, normalization layers, and activation layers. The output feature map after the input feature map x passes through the lightweight deformable feature encoding module is defined as follows:

[0019]

[0020] Where D represents the receptive field of the kernel during model training, m0 represents the target pixel, and m n This represents the position of the target pixel within the receptive field, w(·) represents the weights learned by the convolution, and x(·) and y(·) represent the input and output of the convolution. The learned offset {Δm} n |m n ∈D} makes the sampling positions irregular, m n Typically, the data is floating-point type. An offset variable is added to the position of each sampling point in the convolution kernel through a bias layer. This is used to extract features from oil reservoir displacement etching images with rich shapes and blurred edges, transforming the convolution into an irregular shape to adapt to the varying water and oil shapes. To ensure that the obtained bias values ​​are all integers, bilinear interpolation is used to determine the pixel value at the sampling position. The final pixel value y(m0) at the sampling position is defined as:

[0021] y(m0)=H(e i ,p j (3)

[0022] Among them, e i p represents the learned weights. jThe neighboring pixels of the convolution are represented by H(·), which represents the bilinear interpolation operation. For each layer of the lightweight deformable convolution, when the input of the convolutional layer is a feature map with N channels, the corresponding bias is 2N channels, each of which includes two offset maps in the x and y directions. After obtaining the pixel values ​​at all sampling locations, a new feature map is generated. Finally, depthwise convolution and pointwise convolution are used to decouple the spatial and channel correlations of the convolution, improving the feature representation capability of the lightweight deformable feature encoding module, significantly reducing redundant computations within the layer, and reducing the number of parameters and computational cost while ensuring the feature extraction capability of the convolutional layer, thereby improving the inference speed of the model. During training, the deformable convolution weights for generating the output feature map and the convolutional kernel weights for calculating the offsets are learned simultaneously. Its receptive field can be adaptively adjusted according to the size, shape, and dimensions of the water and oil phases, improving segmentation accuracy while minimizing the computational and memory overhead of the network.

[0023] The lightweight deformable feature encoding module adds a bias layer to the convolution operation, which adds an offset variable to the position of each sampling point in the convolution kernel. This is used to extract features from oil reservoir displacement etching images with rich shapes and blurred edges, transforming the convolution into an irregular shape to adapt to the varied water and oil shapes. Depthwise convolution and pointwise convolution are used to decouple the spatial and channel correlations of the convolution, reducing the number of parameters and computational cost while maintaining the feature extraction capability of the convolutional layer, thus improving the model's inference speed. Compared to conventional convolution, the lightweight deformable convolution module not only improves the network's feature representation capability using the bias layer but also effectively reduces the number of network parameters through depthwise separable convolution. Furthermore, to obtain more refined oil-water phase segmentation results, residual design is used after each lightweight deformable feature encoding module, thereby reducing the gradient vanishing phenomenon caused by excessively deep network layers and accelerating network convergence, improving the network's inference speed and thus increasing segmentation efficiency.

[0024] Step 2.2: In the feature fusion stage, a collaborative coupling attention module is designed to decompose the channel attention, aggregate features along two spatial directions respectively, and encode the feature map into a direction- and position-aware feature map, thereby embedding position information into the channel attention feature map.

[0025] For segmentation tasks of etched images with complex shapes and positions, the relationship between spatial and channel information is crucial for segmenting fine targets. In channel attention, global pooling is typically used to globally encode spatial information; however, this operation compresses global spatial information into channel information, making it difficult to preserve positional information. To enable the attention module to acquire spatial positional information, we first use a convolutional kernel of size (h,1) or (w,1) to perform average pooling on the input x, decomposing the channel attention into two parallel one-dimensional feature encodings. These encodings encode the channels along the horizontal and vertical coordinates, thereby coupling spatial positional information in different directions within the channel attention while capturing the relationship between positional information and channels. This allows us to obtain global dependencies and long-range contextual information, establishing and improving the model's global information extraction capability. Its horizontal output is:

[0026]

[0027] Where x is the given input, h is the height of the given feature map, and c is the number of channels of the given feature map.

[0028] Similarly, the output of the c-th channel with width w is:

[0029]

[0030] The transformations in the two directions described above are aggregated along the spatial directions to obtain a pair of direction-aware feature maps. These maps can capture long-range dependencies along one spatial direction and preserve precise location information along the other spatial direction, helping the network to more accurately locate targets of interest.

[0031] To better utilize the aforementioned modules to generate attention maps and fully leverage the captured location information, enabling a more complete fusion of channel and location information, two one-dimensional feature codes are used to aggregate the input features into two independent location maps along the vertical and horizontal directions, respectively. A two-layer shared 1×1 convolution is then used to transform the input features, resulting in feature maps f representing the horizontal and vertical spatial information of the convolution transformation function S1:

[0032] f=σ(S1([z h ,z w ])) (6)

[0033] Where [·] represents a concatenation operation along the spatial dimension. To improve the model's generalization ability, σ(·) is a nonlinear function. To reduce model complexity, an appropriate channel compression ratio r is often used to reduce the number of channels in the feature map. The feature map is split along the spatial dimension into two separate tensors. and Therefore, we get:

[0034]

[0035]

[0036] Where δ is the sigmoid activation function, and it is finally superimposed on the original image as a weight to obtain the cooperative attention:

[0037]

[0038] The collaboratively coupled attention module embeds location information into channel attention and fully integrates it. Horizontal and vertical attention are simultaneously applied to the input features, enabling the collaboratively coupled attention to more accurately locate the precise position of the object of interest. This helps the model achieve more accurate target localization, improves the segmentation accuracy of semantically rich reservoir displacement etching images, and thus more accurately predicts the oil displacement rate.

[0039] Step 2.3: Upsampling Module. This module decodes the feature-fused image back to its original size, thus completing image segmentation.

[0040] Step 2.4: Construct a reservoir displacement etching image segmentation model based on a lightweight deformable feature encoding module, a collaboratively coupled attention module, and an upsampling module. Train and predict reservoir displacement to improve the training and inference speed of the reservoir displacement etching image segmentation model, reduce computational costs and memory requirements, and thus improve the segmentation accuracy, segmentation efficiency, and robustness of the reservoir displacement etching image segmentation model.

[0041] Step 3: Use the reservoir displacement etching image segmentation model to segment the image, and automatically mark the water-bearing and oil-bearing areas with high precision and efficiency. Use the number of marked oil-bearing and water-bearing pixels to predict the oil displacement rate, reduce the oil displacement rate prediction error, and thus improve the accuracy, efficiency and robustness of the oil displacement rate prediction of the etching image.

[0042] The oil displacement rate is mainly the proportion of oil-bearing pixels to all pixels in an image. This index is used to calculate the effects of various factors, such as different fluid viscosities, injection rates, and initial water saturation, on the water-drive oil effect. Analyzing the oil-water distribution patterns before and after water-drive is of significant research value for understanding the water-drive mechanism in oil reservoirs. The specific calculation method is as follows:

[0043]

[0044] Where S represents all pixel values ​​in the etched image, and O represents the set of pixels in the etched image that belong to oil. The smaller the value, the higher the oil removal rate during etching.

[0045] Oil displacement ratio error (OEDR) is used to measure the water-oil ratio error in practical applications. The calculation formula is as follows:

[0046]

[0047] Among them, F W and F O B represents the set of pixels belonging to the water and oil labels in the segmentation results, respectively. W and B O These represent the sets of pixels belonging to the water and oil labels in the annotation results, respectively. The closer the segmentation result is to the annotation result, the smaller the oil displacement rate error.

[0048] It also includes step four: based on the oil displacement rate prediction results of the reservoir displacement etching image, conduct reservoir displacement experiments, adjust and optimize the reservoir displacement experiment parameters, improve the accuracy of reservoir exploration location selection and oil displacement efficiency, and solve related engineering problems.

[0049] The adjustment and optimization of reservoir displacement experiment parameters include the addition of displacement reagents, pressure selection, and injection point selection.

[0050] Beneficial effects:

[0051] 1. Compared to manual reservoir displacement etching for oil displacement rate prediction, this invention discloses an oil displacement rate prediction method based on reservoir displacement etching images. Addressing the non-rigid structural characteristics of the oil and water phases in reservoir displacement etching images, and considering the influence of the shooting environment and chip material, the acquired displacement etching images often suffer from problems such as image blurring, uneven illumination, low color contrast, and high noise. This invention uses a lightweight deformable feature encoding module, a collaboratively coupled attention module, and an upsampling module to construct an oil displacement etching image segmentation model LDNet. Combined with prior knowledge of the reservoir displacement etching images, the image segmentation model is used for image segmentation. It can automatically mark water-bearing and oil-bearing regions with high accuracy and efficiency, and predict the oil displacement rate using the number of marked oil-bearing and water-bearing pixels. This invention innovatively applies theoretical knowledge from the fields of artificial intelligence and image segmentation to the segmentation of oil displacement etching images and the prediction of oil displacement rate in oil displacement etching, thereby avoiding the inaccuracy of oil displacement rate prediction in oil displacement etching, reducing the prediction error of oil displacement rate, and thus improving the accuracy, efficiency and robustness of oil displacement rate prediction in etched images.

[0052] 2. This invention discloses a method for predicting oil displacement rate based on reservoir displacement etching images. In the feature encoding stage, a lightweight deformable feature encoding module is designed. By adding an offset variable to the position of each sampling point in the convolution kernel, it is used to extract features from reservoir displacement etching images with rich shapes and blurred edges. Compared with conventional convolution, the lightweight deformable convolution module can not only improve the feature representation ability of the network by utilizing bias layers, but also effectively reduce the model size through depthwise separable convolution, which is conducive to efficiently completing the calculation of oil displacement rate.

[0053] 3. This invention discloses a method for predicting oil displacement rate based on reservoir displacement etching images. In the feature fusion stage, a collaboratively coupled attention module is designed to decompose channel attention, aggregating features along two spatial directions respectively, and encoding the feature map into a direction- and position-aware feature map, thereby embedding position information into the channel attention feature map. A residual bottleneck module is used to fuse higher-level semantic information to improve the model's feature representation ability for etching images. This enhances the model's representation ability, suppresses the influence of noise by focusing on important regions and channels, and improves the model's robustness to input noise, solving problems such as high noise in reservoir etching images, blurred water-oil fusion boundaries, and rough image segmentation edges caused by regional geometry.

[0054] 4. The present invention discloses an oil displacement rate prediction method based on reservoir displacement etching images. In the image preprocessing stage, the NPE algorithm is used for image enhancement to reduce the influence of uneven illumination, improve image segmentation accuracy, and thus improve the oil displacement rate prediction accuracy.

[0055] 5. The present invention discloses a method for predicting oil displacement rate based on reservoir displacement etching images. Based on achieving the above-mentioned beneficial effects 1, 2, 3, and 4, the method conducts reservoir displacement experiments based on the oil displacement rate prediction results of reservoir displacement etching images, adjusts and optimizes the reservoir displacement experiment parameters, and improves the accuracy of reservoir exploration location selection and oil displacement efficiency. Attached Figure Description

[0056] Figure 1 This is a flowchart of a method for predicting oil displacement rate based on reservoir displacement etching images disclosed in this invention.

[0057] Figure 2 It is a diagram of a lightweight, deformable codec network structure.

[0058] Figure 3 It is a lightweight, deformable feature encoding module.

[0059] Figure 4 This is a schematic diagram of the collaborative coupling attention module.

[0060] Figure 5 This is a graph showing the difference between the prediction results and the actual values ​​of the lightweight deformable coding module.

[0061] Figure 6 This is a visualization result of the feature map of collaboratively coupled attention.

[0062] Figure 7 This is a comparison of the segmentation results of different segmentation networks on the etched image.

[0063] Figure 8 The error in calculating the oil displacement rate using different methods (unit: %). Detailed Implementation

[0064] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0065] Example 1:

[0066] like Figure 1 As shown in this embodiment, a method for predicting the oil displacement rate from reservoir displacement etching images includes the following steps:

[0067] Step 1: The acquired reservoir displacement etching image data is cropped, and image enhancement and data enhancement operations are performed on the etching image data to achieve reservoir displacement etching image data preprocessing and obtain reservoir displacement etching image dataset.

[0068] The image enhancement process involves performing a logarithmic transformation on the image to adjust the detail information of the etched image, resulting in an enhanced etched image.

[0069] The data augmentation includes horizontal flipping, rotation, and adding noise.

[0070] To reduce the impact of uneven image illumination, the image enhancement uses the NPE (naturalness-preserved enhancement) algorithm for image preprocessing. The algorithm formula is shown in (1). It uses a brightness-pass filter Q(·) to decompose the image into reflectance and illuminance, which determine the detail and naturalness of the image respectively. Then, a double logarithmic transformation U(·) is used to map the illumination to achieve a balance between detail and naturalness, thereby preserving the detail of the oil reservoir displacement etching image under non-uniform illumination in the experiment, and maintaining the naturalness of the oil phase and water phase images.

[0071]

[0072] Step Two: As Figure 2As shown, the main components of this invention are: In the feature encoding stage, a Lightweight Deformable Feature Encoding (LDFE) module is designed. This module adds an offset variable to the position of each sampling point in the convolutional kernel to extract features from oil reservoir displacement etching images with rich shapes and blurred edges. Compared to conventional convolution, the Lightweight Deformable Convolution module not only improves the network's feature representation ability using bias layers but also effectively reduces the number of network parameters through depthwise separable convolution. In the feature fusion stage, a Co-coupled Attention module is designed to decompose channel attention, aggregating features along two spatial directions respectively, encoding the feature map into a direction- and position-aware feature map, thereby embedding positional information into the channel attention feature map. A residual bottleneck module is used to fuse higher-level semantic information to improve the model's feature representation ability for etching images. The LDNet reservoir displacement etching image segmentation model is constructed based on a lightweight deformable feature encoding module, a collaboratively coupled attention module, and an upsampling module. This improves the training and inference speed of the reservoir displacement etching image segmentation model, reduces computational costs and memory requirements, and thus enhances the segmentation accuracy, efficiency, and robustness of the reservoir displacement etching image segmentation model.

[0073] Step 2.1: In the feature encoding stage, such as Figure 3 As shown, a lightweight deformable feature encoding module is designed. By adding an offset variable to the position of each sampling point in the convolution kernel, it is used to extract features from reservoir displacement and etching images with rich shapes and blurred edges. Compared with conventional convolution, the lightweight deformable convolution module can not only improve the feature representation ability of the network by utilizing bias layers, but also effectively reduce the number of network parameters through depthwise separable convolution.

[0074] The complex background of etched images and the non-rigid boundaries between the water and oil phases in oil reservoir etched images make segmentation difficult. Conventional models mostly use square convolutional kernels with fixed shapes for feature learning. However, the receptive field provided by convolutional kernels with fixed shapes is very limited, resulting in weak ability to extract contextual information from etched images, which significantly restricts the ability to extract multi-scale information from oil reservoir etched images with rich and complex shapes.

[0075] To adapt to the characteristics of non-rigid aqueous and oil phases and reduce the parameter redundancy and training difficulties caused by increasing bias, a lightweight deformable feature encoding module is designed in the feature encoding stage. This module includes convolutional layers, bias layers, normalization layers, and activation layers. The output feature map after the input feature map x passes through the lightweight deformable feature encoding module is defined as follows:

[0076]

[0077] Where D represents the receptive field of the kernel during model training, m0 represents the target pixel, and m n This represents the position of the target pixel within the receptive field, w(·) represents the weights learned by the convolution, and x(·) and y(·) represent the input and output of the convolution. The learned offset {Δm} n |m n ∈D} makes the sampling positions irregular, m n Typically, the data is floating-point type. An offset variable is added to the position of each sampling point in the convolution kernel through a bias layer. This is used to extract features from oil reservoir displacement etching images with rich shapes and blurred edges, transforming the convolution into an irregular shape to adapt to the varying water and oil shapes. To ensure that the obtained bias values ​​are all integers, bilinear interpolation is used to determine the pixel value at the sampling position. The final pixel value y(m0) at the sampling position is defined as:

[0078] y(m0)=H(e i ,p j (3)

[0079] Among them, e i p represents the learned weights. j The neighboring pixels of the convolution are represented by H(·), which represents the bilinear interpolation operation. For each layer of the lightweight deformable convolution, when the input of the convolutional layer is a feature map with N channels, the corresponding bias is 2N channels, each of which includes two offset maps in the x and y directions. After obtaining the pixel values ​​at all sampling locations, a new feature map is generated. Finally, depthwise convolution and pointwise convolution are used to decouple the spatial and channel correlations of the convolution, improving the feature representation capability of the lightweight deformable feature encoding module, significantly reducing redundant computations within the layer, and reducing the number of parameters and computational cost while ensuring the feature extraction capability of the convolutional layer, thereby improving the inference speed of the model. During training, the deformable convolution weights for generating the output feature map and the convolutional kernel weights for calculating the offsets are learned simultaneously. Its receptive field can be adaptively adjusted according to the size, shape, and dimensions of the water and oil phases, improving segmentation accuracy while minimizing the computational and memory overhead of the network.

[0080] The lightweight deformable feature encoding module adds a bias layer to the convolution operation, which adds an offset variable to the position of each sampling point in the convolution kernel. This is used to extract features from oil reservoir displacement etching images with rich shapes and blurred edges, transforming the convolution into an irregular shape to adapt to the varied water and oil shapes. Depthwise convolution and pointwise convolution are used to decouple the spatial and channel correlations of the convolution, reducing the number of parameters and computational cost while maintaining the feature extraction capability of the convolutional layer, thus improving the model's inference speed. Compared to conventional convolution, the lightweight deformable convolution module not only improves the network's feature representation capability using the bias layer but also effectively reduces the number of network parameters through depthwise separable convolution. Furthermore, to obtain more refined oil-water phase segmentation results, residual design is used after each lightweight deformable feature encoding module, thereby reducing the gradient vanishing phenomenon caused by excessively deep network layers and accelerating network convergence, improving the network's inference speed and thus increasing segmentation efficiency.

[0081] When the output feature map x has a size of H×W×N, the computational cost of a regular convolutional kernel is K. 2 The computational complexity of deformable convolution is K ×M×N×H×W. 2 ×M×H×W×2+K 2 The computational cost of lightweight deformable convolution is K ×M×N×H×W. 2 ×M×H×W×2+H×W×M×N. Compared to deformable convolution, the computational cost of the lightweight deformable feature encoding module can be reduced to 0.6 times that of deformable convolution, where M is the number of input feature channels, N is the number of output feature channels, the convolution kernel size is K×K, and H and W are the height and width of the input feature, respectively. Furthermore, to obtain more refined oil-water phase segmentation results, residual design is used after each lightweight deformable feature encoding module, thereby reducing the gradient vanishing phenomenon caused by excessively deep network layers and accelerating network convergence.

[0082] To verify that the Lightweight Deformable Feature Encoding Module (LDFE) can learn flexible positional and shape information compared to standard convolutions, U-Net was used as the backbone network, and the standard convolutions in U-Net were replaced with the Lightweight Deformable Feature Encoding Module (LDFE) module to verify the advantages of the Lightweight Deformable Feature Encoding Module. Figure 4The figure shows the differences in the segmentation results of U-Net, U-Net+DC (conventional convolution), and U-Net+LDFE on the etched image library of low-permeability reservoirs. The red and blue foregrounds represent differences, while the black background represents identical pixels. As shown, both deformable convolution and the lightweight deformable feature encoding module help U-Net acquire more edge information, but the lightweight deformable feature encoding module captures a larger receptive field through point convolution and group convolution and obtains more accurate contour prediction using residual structures. The feature map provided by U-Net+LDFE contains more detailed information, and its result is closest to the labeled data.

[0083] Step 2.2: As Figure 5 As shown, in the feature fusion stage, a collaborative coupling attention module is designed to decompose the channel attention, aggregate features along two spatial directions respectively, and encode the feature map into a direction- and position-aware feature map, thereby embedding position information into the channel attention feature map.

[0084] For segmentation tasks of etched images with complex shapes and positions, the relationship between spatial and channel information is crucial for segmenting fine targets. In channel attention, global pooling is typically used to globally encode spatial information; however, this operation compresses global spatial information into channel information, making it difficult to preserve positional information. To enable the attention module to acquire spatial positional information, we first use a convolutional kernel of size (h,1) or (w,1) to perform average pooling on the input x, decomposing the channel attention into two parallel one-dimensional feature encodings. These encodings encode the channels along the horizontal and vertical coordinates, thereby coupling spatial positional information in different directions within the channel attention while capturing the relationship between positional information and channels. This allows us to obtain global dependencies and long-range contextual information, establishing and improving the model's global information extraction capability. Its horizontal output is:

[0085]

[0086] Where x is the given input, h is the height of the given feature map, and c is the number of channels of the given feature map.

[0087] Similarly, the output of the c-th channel with width w is:

[0088]

[0089] The transformations in the two directions described above are aggregated along the spatial directions to obtain a pair of direction-aware feature maps. These maps can capture long-range dependencies along one spatial direction and preserve precise location information along the other spatial direction, helping the network to more accurately locate targets of interest.

[0090] To better utilize the aforementioned modules to generate attention maps and fully leverage the captured location information, enabling a more complete fusion of channel and location information, two one-dimensional feature codes are used to aggregate the input features into two independent location maps along the vertical and horizontal directions, respectively. A two-layer shared 1×1 convolution is then used to transform the input features, resulting in feature maps f representing the horizontal and vertical spatial information of the convolution transformation function S1:

[0091] f=σ(S1([z h ,z w ])) (6)

[0092] Where [·] represents a concatenation operation along the spatial dimension. To improve the model's generalization ability, σ(·) is a nonlinear function. To reduce model complexity, an appropriate channel compression ratio r is often used to reduce the number of channels in the feature map. The feature map is split along the spatial dimension into two separate tensors. and Therefore, we get:

[0093]

[0094]

[0095] Where δ is the sigmoid activation function, and it is finally superimposed on the original image as a weight to obtain the cooperative attention:

[0096]

[0097] The collaboratively coupled attention module embeds location information into channel attention and fully integrates it. Horizontal and vertical attention are simultaneously applied to the input features, enabling the collaboratively coupled attention to more accurately locate the precise position of the object of interest. This helps the model achieve more accurate target localization, improves the segmentation accuracy of semantically rich reservoir displacement etching images, and thus more accurately predicts the oil displacement rate.

[0098] To verify the effectiveness of collaborative attention, it was compared with the most popular channel attention mechanism SENet and the hybrid attention mechanism CBAM. The visualization results are as follows: Figure 6As shown in the diagram, red indicates areas of high attention, and blue indicates areas of low attention. It can be seen that adding a hybrid attention module helps the network achieve better extraction of the water and oil phases compared to a single-channel attention module. Cooperative attention fully couples channel information with spatial information, resulting in stronger perception of target areas, less susceptibility to noise interference, and more accurate extraction of small targets compared to hybrid attention. Unlike attention modules that only focus on channel or spatial importance, the cooperative attention module embeds positional information into channel attention and fully integrates it. Attention along both horizontal and vertical directions is simultaneously applied to the input features, enabling cooperative attention to more accurately locate the precise position of the object of interest, thus helping the model achieve more accurate target localization.

[0099] Step 2.3: Upsampling Module. This module decodes the feature-fused image back to its original size, thus completing image segmentation.

[0100] Step 2.4: Construct a reservoir displacement etching image segmentation model based on a lightweight deformable feature encoding module, a collaboratively coupled attention module, and an upsampling module. Train and predict reservoir displacement to improve the training and inference speed of the reservoir displacement etching image segmentation model, reduce computational costs and memory requirements, and thus improve the segmentation accuracy, segmentation efficiency, and robustness of the reservoir displacement etching image segmentation model.

[0101] To further demonstrate the superiority of the proposed network, mainstream segmentation networks were selected as comparison methods, including: U-Net, U-Net++, CE-Net, DefED-Net, Attention U-Net, H-DenseUNet, and Rednet. Segmentation accuracy (Acc), mean intersection-over-union ratio (mIOU), and recall were used to evaluate the segmentation results.

[0102] As shown in Table 1, the proposed LDNet exhibits the best performance, with a model size of 16.63M and a segmentation accuracy of 89.94%. The base network U-Net uses standard convolutions, which cannot effectively extract edge details from images, resulting in the lowest segmentation accuracy. UNet++ redesigns the skip connection method compared to U-Net, thus improving the accuracy to 82.52%. CE-Net and Rednet both improved accuracy to 86.82% and 84.54% respectively by designing new convolution operators, but both resulted in a large number of parameters. DefED-Net and LDNet both use deformable convolutions, which are more effective at extracting edges from etched images. LDNet uses point convolutions and group convolutions separately for deformable convolutions and incorporates a residual network to more effectively extract multi-scale information from images. It also employs collaborative coupling attention to fuse contextual information and establish long-distance dependencies, thus achieving higher model accuracy. Furthermore, since deformable convolutions are computationally expensive, combining point convolutions and group convolutions can significantly reduce the computational cost and model size. Compared to DefED-Net, LDNet reduces the computational cost of the model by 56.27% and the model size by 84.99%.

[0103] Table 1. Experimental results of different methods on etched images.

[0104]

[0105] like Figure 7 As shown, the visualization results of different methods are presented on a partial etching image dataset. The basic U-Net network performs poorly in segmentation of etched images. UNet++ captures more detailed information than U-Net, but struggles to segment blurred boundary regions. H-DenseUNet uses a dense connection strategy and captures more detailed information than UNet++, but its segmentation of water and oil edges is poor. CE-Net and Rednet both improve the traditional convolution operator in the encoding stage, achieving ideal segmentation of blurred boundaries, but they both lack fine water and oil information. Attention U-Net enhances the segmentation of small targets by adding attention gating information, but its segmentation of water and oil boundary positions is poor. DefED-Net enhances edge details by using deformable convolution, but its segmentation of small targets in water and oil is poor, and its computational and parameter count is large. Compared with the above methods, LDNet provides more accurate segmentation boundaries and fine target extraction, thus achieving better feature encoding and feature fusion, and realizing high-precision segmentation of etched images.

[0106] Step 3: Use the reservoir displacement etching image segmentation model to segment the image, and automatically mark the water-bearing and oil-bearing areas with high precision and efficiency. Use the number of marked oil-bearing and water-bearing pixels to predict the oil displacement rate, reduce the oil displacement rate prediction error, and thus improve the accuracy, efficiency and robustness of the oil displacement rate prediction of the etching image.

[0107] The oil displacement rate is mainly the proportion of oil-bearing pixels to all pixels in an image. This index allows for the calculation of the effects of various factors, including different fluid viscosities, injection rates, and initial water saturation, on the water-drive oil effect. Analyzing the oil-water distribution patterns before and after water-drive is of significant research value for understanding the water-drive mechanism in oil reservoirs. The specific calculation method is as follows:

[0108]

[0109] Where S represents all pixel values ​​in the etched image, and O represents the set of pixels in the etched image that belong to oil. The smaller the value, the higher the oil removal rate during etching.

[0110] Oil displacement ratio error (OEDR) is used to measure the water-oil ratio error in practical applications. The calculation formula is as follows:

[0111]

[0112] Among them, F W and F O B represents the set of pixels belonging to the water and oil labels in the segmentation results, respectively. W and B O These represent the sets of pixels belonging to the water and oil labels in the annotation results, respectively. The closer the segmentation result is to the annotation result, the smaller the oil displacement rate error. Figure 8 The oil displacement rate error of eight methods is shown. The oil displacement rate error is minimized based on a lightweight deformable encoder-decoder network, demonstrating the effectiveness and superiority of the network for different etched images.

[0113] Step 4: Based on the oil displacement rate prediction results from the reservoir displacement etching images, conduct reservoir displacement experiments, adjust and optimize the reservoir displacement experiment parameters, improve the accuracy of reservoir exploration location selection and oil displacement efficiency, and solve related engineering problems.

[0114] The adjustment and optimization of reservoir displacement experiment parameters include the addition of displacement reagents, pressure selection, and injection point selection.

[0115] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the oil displacement rate from reservoir displacement etching images, characterized in that: Includes the following steps, Step 1: The acquired reservoir displacement etching image data is cropped, and image enhancement and data enhancement operations are performed on the etching image data to achieve reservoir displacement etching image data preprocessing and obtain reservoir displacement etching image dataset. Step 2: In the feature encoding stage, a lightweight deformable feature encoding module LDFE is designed, which adds an offset variable to the position of each sampling point in the convolution kernel; In the feature fusion stage, a collaborative coupling attention module is designed to decompose the channel attention, aggregate features along two spatial directions respectively, and encode the feature map into a direction- and position-aware feature map, thereby embedding position information into the channel attention feature map; The residual bottleneck module is used to fuse higher-level semantic information to improve the model's ability to represent features of etched images; the reservoir displacement etched image segmentation model LDNet is constructed based on the lightweight deformable feature encoding module, the collaboratively coupled attention module, and the upsampling module. Step 3: Use the reservoir displacement etching image segmentation model to segment the image, automatically mark the water-bearing and oil-bearing regions, and predict the oil displacement rate using the number of marked oil-bearing and water-bearing pixels.

2. The method for predicting the oil displacement rate of reservoir displacement etching images as described in claim 1, characterized in that: It also includes step four, which involves conducting reservoir displacement experiments based on the predicted oil displacement rate from the reservoir displacement etching images, and adjusting and optimizing the reservoir displacement experiment parameters.

3. The method for predicting the oil displacement rate of a reservoir displacement etching image as described in claim 1 or 2, characterized in that: The image enhancement process performs a logarithmic transformation on the image to adjust the detail information of the etched image, resulting in an enhanced etched image. The data augmentation includes horizontal flipping, rotation, and adding noise; The image enhancement uses the NPE algorithm for image preprocessing, and the algorithm formula is shown in (1). It uses a brightness pass filter. The image is decomposed into reflectance and illuminance, and then a double logarithmic transform is applied. ; (1)。 4. The method for predicting the oil displacement rate of a reservoir displacement etching image as described in claim 3, characterized in that: Step 2.1: In the feature encoding stage, a lightweight deformable feature encoding module is designed, which includes convolutional layers, bias layers, normalization layers, and activation layers; input feature map. The output feature map after passing through the lightweight deformable feature encoding module is defined as follows: Among them, among them, This represents the receptive field of the kernel during model training. Represents the target pixel. This indicates the position where the target pixel falls within the receptive field. Represents the weights learned by the convolution. and Represents the input and output of the convolution; learns the offset. This causes the sampling location to become irregular. Typically, the data is floating-point type; an offset variable is added to the position of each sampling point in the convolution kernel through a bias layer; to ensure that the obtained bias values ​​are all integers, bilinear interpolation is used to determine the pixel value at the sampling position, and the final pixel value at the sampling position is... Defined as: in, This represents the learned weights. Represents the neighboring pixels of the convolution. Represents a bilinear interpolation operation; for each layer of a lightweight deformable convolution, when the input to the convolutional layer is a bilinear interpolation operation; When mapping the features of each channel, the corresponding bias is: Each of the channels includes [number] channels. and Two offset mappings in the direction; after obtaining the pixel values ​​at all sampling locations, a new feature map is generated; finally, depthwise convolution and pointwise convolution are used to decouple the spatial and channel correlation of the convolution; during training, deformable convolution weights for generating the output feature map and convolution kernel weights for calculating the offset are learned simultaneously. The lightweight deformable feature encoding module decouples the spatial and channel correlations of convolution by adding a bias layer to the convolution operation and using depthwise convolution and pointwise convolution; residual design is used after each lightweight deformable feature encoding module; Step 2.2: In the feature fusion stage, a collaborative coupling attention module is designed to decompose the channel attention, aggregate features along two spatial directions respectively, and encode the feature map into a direction- and position-aware feature map, thereby embedding position information into the channel attention feature map; First, use the size... or The convolution kernel for the input Average pooling is performed to decompose the channel attention into two parallel one-dimensional feature codes, which encode the channels along the horizontal and vertical coordinates. The horizontal output is: in, Given the input, Given the height of the feature map, The number of channels for a given feature map; Similarly, the width is The The output of each channel is: The transformations in the two directions mentioned above are aggregated along the spatial direction to obtain a pair of direction-aware feature maps; The input features are aggregated into two independent positional maps by using two one-dimensional feature encoders along the vertical and horizontal directions, respectively, and two layers of shared... Convolution is transformed to obtain the convolution transformation function. Feature maps of horizontal and vertical spatial information for: in This indicates a concatenation operation along a spatial dimension; It is a nonlinear function; to reduce the complexity of the model, an appropriate channel compression ratio is often used. To reduce the number of channels in the feature map; the feature map is split into two separate tensors along the spatial dimension. and Therefore, we get: , , The collaboratively coupled attention module embeds positional information into channel attention and performs full fusion; horizontal and vertical attention are simultaneously applied to the input features. Step 2.3: Upsampling module; This module decodes the image after feature fusion back to the original image size, thereby completing image segmentation; Step 2.4: Construct a reservoir displacement etching image segmentation model based on a lightweight deformable feature encoding module, a collaboratively coupled attention module, and an upsampling module, and train and predict reservoir displacement.

5. The method for predicting the oil displacement rate of a reservoir displacement etching image as described in claim 4, characterized in that: The oil removal rate is the proportion of oil-containing pixels to all pixels in the image, and its specific calculation method is as follows: （10） in, This represents all pixel values ​​of the etched image. This represents the set of pixels in the etched image that belong to the oil. The oil displacement ratio error (OEDR) is used to measure the water-oil ratio error in practical applications. The calculation formula is as follows: （11） in, and These represent the sets of pixels belonging to the water and oil labels in the segmentation results, respectively. and These represent the sets of pixels belonging to the water and oil labels in the annotation results, respectively.

Images