Complex buried hill reservoir full-scale pore structure calculation method based on deep learning
Through the intelligent pore recognition model based on deep learning, the problem of multi-scale micropore structure characterization of complex latent mountain reservoirs is solved, efficient and accurate pore structure analysis is achieved, and work efficiency and quality are improved.
Patent Information
- Application Number
- CN202411823907.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-03-18
AI Technical Summary
The prior art is difficult to accurately and comprehensively reveal the multi-scale micropore structure of complex latent mountain reservoirs, and the traditional methods are large in workload and low in efficiency, and the results are greatly affected by personal experience, resulting in inaccurate results.
A deep learning-based method is adopted, combining image recognition and deep learning algorithms to establish an intelligent pore recognition model, and label the pore structure through multi-scale multi-resolution scanning and pre-processing, construct a deep convolutional neural network model, and perform multi-scale characterization of pore structure.
Multi-scale characterization of the pore structure of complex latent mountain reservoirs has been realized, which greatly improves the working efficiency and quality of digital core image analysis, and provides strong support for the fine characterization of complex latent mountain reservoirs.
Smart Images
Figure CN119649145A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of oil and gas exploration technology, and in particular to a method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning. Background Art
[0002] Complex buried-hill reservoirs have the characteristics of strong heterogeneity, dense reservoirs, and complex pore and fluid conditions, which makes its reservoir description and fluid identification difficult. Compared with conventional oil and gas reservoirs, the unique sedimentary environment and self-generation and self-reservoir characteristics of complex buried-hill reservoirs result in low porosity, ultra-low permeability, and multi-type and multi-scale pore structures. The fluid seepage law of complex buried-hill reservoirs is limited by the characteristics of rock pore structure. Therefore, accurate and comprehensive understanding of the pore structure of complex buried-hill reservoirs is the basis and key to accurate reserve assessment and efficient oil and gas exploitation.
[0003] At present, there are the following problems in the characterization of the microscopic pore structure of complex buried-hill reservoirs: single-scale imaging cannot accurately and comprehensively reveal the multi-scale microscopic pore structure of complex buried-hill reservoirs; the characterization of pore structure at a single scale depends on the selection of artificial regions of interest and resolution, which has poor objectivity and leads to inaccurate results of the overall pore structure analysis; the use of ImageJ, Avizo and other software for pore structure characterization, porosity and permeability simulation, and mineral registration is labor-intensive and inefficient, and the results are greatly affected by personal experience, resulting in inaccurate results; the pore structure characterization, porosity and permeability prediction, and mineral quantitative identification based on deep learning algorithms are mostly research, with less practical application. Summary of the invention
[0004] The purpose of the present invention is to provide a full-scale pore structure calculation method for complex buried hill reservoirs based on deep learning. It combines image recognition with deep learning algorithms to establish an intelligent pore recognition model, realize multi-scale characterization of the pore structure of complex buried hill reservoirs, greatly improve the work efficiency and work quality based on digital core image analysis of complex buried hill reservoirs, and provide strong support for the fine characterization of complex buried hill reservoirs.
[0005] To achieve the above object, the present invention adopts the following technical solution, including the following steps:
[0006] S1. Carry out multi-scale and multi-resolution scanning on complex buried hill reservoir samples, obtain images of complex buried hill reservoir cores at different scales, and pre-process the obtained complex buried hill reservoir cores at full scale;
[0007] S2, performing sample annotation on the preprocessed full-scale image, wherein the annotation target is the pores in the image, and performing polygon annotation on the pores in the image based on the DEXTR algorithm. After the annotation is completed, a multi-scale pore image dataset of a complex buried hill reservoir is obtained, and the multi-scale pore image dataset of the complex buried hill reservoir is divided into a training set and a validation set;
[0008] S3. Use deep learning methods to build a full-scale pore intelligent identification model for complex buried hill reservoirs;
[0009] S4. Use the training set and the validation set to train the complex buried hill reservoir full-scale pore intelligent identification model, use the training set to train the model, use the validation set to verify the model performance, use pixel accuracy, recall rate and mean intersection and union ratio to measure the performance of the model, determine whether the model meets the requirements, and select the model with high accuracy index of the validation set for the final model reasoning application;
[0010] S5. Use the full-scale pore intelligent identification model of complex buried-hill reservoirs to identify pores, input the multi-level and multi-scale images to be detected respectively, and obtain the pore radius distribution and pore volume distribution results of the complex buried-hill reservoir under multi-scale fusion.
[0011] Preferably, in step S1, the images of the complex buried hill reservoir core at different scales include three-dimensional images and two-dimensional scanning electron microscope images.
[0012] Preferably, the preprocessing includes equalization processing, Gaussian filtering and median filtering; firstly, histogram equalization processing is used to make the light and dark contrast of the input image clear, then Gaussian filtering is used to suppress high-frequency noise in the image, and then median filtering is used to remove noise while retaining image edge information;
[0013] The two-dimensional Gaussian function is:
[0014]
[0015] Where x, y are the template coordinates of the image pixels, and σ is the variance, which controls the width of the Gaussian function.
[0016] Then the filtering result Y(x,y) of each point (x,y) in the Gaussian filtered image is calculated as:
[0017]
[0018] Where: X(x,y) is the grayscale value corresponding to pixel x,y, i,j is the pixel value separated from the domain.
[0019] The filtering result Y(x,y) of each point (x,y) in the median filter image is calculated as follows:
[0020] Y(x,y)=medianX(x+i,y+j)i∈[-K,K],j∈[-K,K]
[0021] Where: K is the filter kernel size.
[0022] Preferably, in step S2, the DEXTR algorithm automatically identifies the pore range in the annotation system, and simultaneously selects the leftmost, rightmost, topmost and bottommost edge points of the pore. The annotation system converts the points into onehot format and inputs them together with the original image into a pre-trained segmentation model to predict the pore boundary through the pre-trained segmentation model.
[0023] Preferably, in step S2, the complex buried hill reservoir multi-scale pore image dataset is divided into a training set and a validation set in a ratio of 80%:20%.
[0024] Preferably, in step S3, the deep learning method includes extracting image features using a deep convolutional neural network deeplabv3+, combining dilated convolution and dilated spatial pyramid pooling techniques to capture multi-scale information, and utilizing an encoding-decoding structure to improve the model's ability to segment objects of different sizes in an image;
[0025] The calculation formula of the dilated convolution is:
[0026] [y[i]=\sum_{k}x[i+r\cdot k]\cdot w[k]]
[0027] Among them, x is the input feature map, y is the output feature map, w is the convolution kernel, r is the dilation rate, and k is the index of the convolution kernel. By adjusting the dilation rate r, the receptive field size of the dilated convolution can be changed;
[0028] The global average pooling layer in the atrous spatial pyramid pooling module performs a global average operation on the input feature map to obtain the overall image features. The calculation formula is as follows:
[0029] [z_i=\frac{1}{H\times W}\sum_{j=1}^{H}\sum_{k=1}^{W}x_{ijk}]
[0030] Among them, x is the input feature map, z is the feature map after global average pooling, and H and W are the height and width of the feature map respectively.
[0031] Preferably, in step S4, assuming that the model outputs a total of N categories, the set of pixels marked as the i-th category is recorded as Gi, the set of pixels predicted to be the i-th category is recorded as Fi, and |·| represents the number of elements in the set;
[0032] Pixel accuracy, calculate the ratio of correctly classified pixels to all pixels in the image. The calculation formula is as follows:
[0033]
[0034] Average recall rate, calculate the proportion of pixels in each category in the image that are correctly classified, and then find the average value. The calculation formula is as follows:
[0035]
[0036] Average intersection-to-union ratio: Calculate the ratio of the intersection and union between each type of prediction result and the labeled value in the image, and then find the average value. The calculation formula is as follows:
[0037]
[0038] The model with high accuracy index on the validation set is selected for the final model inference application.
[0039] Preferably, in step S4, pixel accuracy is the percentage of correctly labeled pixels to total pixels. The higher the PixelAccuracy, the more accurate the model is in segmenting the image. Recall is the proportion of samples whose predicted value is 1 and whose true value is also 1 to all samples whose true value is 1. IoU is calculated as the ratio of the intersection to the union between the predicted area and the actual area. Mean IoU is the average of the IoUs of all categories to balance the imbalance between different categories. The higher the Mean IoU, the more accurate the model is in segmenting different categories.
[0040] Preferably, the pore radius distribution result of the complex buried hill reservoir under multi-scale fusion obtained in step S5 is a pie chart.
[0041] The beneficial effects of the present invention are: combining image recognition with deep learning algorithms, establishing an intelligent pore recognition model, realizing multi-scale characterization of the pore structure of complex buried hill reservoirs, greatly improving the work efficiency and work quality based on digital core image analysis of complex buried hill reservoirs, and providing strong support for the fine characterization of complex buried hill reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 This is a flow chart of the method for calculating the full-scale pore structure of complex buried hill reservoirs based on deep learning in the present invention.
[0043] Figure 2 It is a micrometer CT three-dimensional scanning image of the sample in the embodiment;
[0044] Figure 3 It is a MAPS two-dimensional scanning image of the sample in the embodiment;
[0045] Figure 4This is a FIB 3D scanning image of the sample in the embodiment;
[0046] Figure 5 This is a schematic diagram of the sample before labeling in the embodiment;
[0047] Figure 6 It is a schematic diagram of the labeled samples in the embodiment;
[0048] Figure 7 Input label graph for model training in the embodiment;
[0049] Figure 8 Graph showing the pore radius distribution calculated in the examples. DETAILED DESCRIPTION
[0050] The invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement it according to the description.
[0051] It should be understood that terms such as “having”, “including” and “comprising” used herein do not exclude the existence or addition of one or other elements or combinations thereof.
[0052] Example
[0053] like Figure 1-8 As shown, the method 1 for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning of the present invention comprises the following steps:
[0054] S1. Multi-scale and multi-resolution scanning is performed on complex buried hill reservoir samples to obtain images of complex buried hill reservoir cores at different scales, wherein the images of complex buried hill reservoir cores at different scales include three-dimensional images and two-dimensional scanning electron microscope images.
[0055] like Figure 2 This is a three-dimensional micro-CT scanning image of the sample, with a resolution of 2.65 μm and a characterization unit volume size of 1000×1000×1000 (pixels); Figure 3 This is a large-field 2D scanning image of the sample MAPS, with a resolution of 2.5 μm and a characterization unit volume size of 3000 × 3000 (pixels); Figure 4 This is a FIB 3D scanning image of the sample with a resolution of 50nm and a characterization unit cell size of 700×700×1000 (pixels).
[0056] The obtained complex buried hill reservoir core is preprocessed at full scale; the preprocessing includes equalization processing, Gaussian filtering and median filtering; firstly, histogram equalization processing is used to make the light and dark contrast of the input image clear, then Gaussian filtering is used to suppress high-frequency noise in the image, and then median filtering is used to remove noise while retaining image edge information;
[0057] The two-dimensional Gaussian function is:
[0058]
[0059] Where x, y are the template coordinates of the image pixels, and σ is the variance, which controls the width of the Gaussian function.
[0060] Then the filtering result Y(x,y) of each point (x,y) in the Gaussian filtered image is calculated as:
[0061]
[0062] Where: X(x,y) is the grayscale value corresponding to pixel x,y, i,j is the pixel value separated from the domain.
[0063] The filtering result Y(x,y) of each point (x,y) in the median filter image is calculated as follows:
[0064] Y(x,y)=medianX(x+i,y+j)i∈[-K,K],j∈[-K,K]
[0065] Where: K is the filter kernel size.
[0066] S2, performing sample annotation on the preprocessed full-scale image, wherein the annotation target is the pores in the image, and performing polygon annotation on the pores in the image based on the DEXTR algorithm. After the annotation is completed, a multi-scale pore image dataset of a complex buried hill reservoir is obtained, and the multi-scale pore image dataset of the complex buried hill reservoir is divided into a training set and a validation set;
[0067] In step S2, the DEXTR algorithm automatically identifies the pore range in the annotation system and selects the leftmost, rightmost, topmost and bottommost four edge points of the pore, such as Figure 6 The annotation system converts the points into onehot format and inputs them into the pre-trained segmentation model together with the original image. The pore boundary is predicted by the pre-trained segmentation model, such as Figure 7 As a preferred embodiment, the complex buried hill reservoir multi-scale pore image dataset is divided into a training set and a validation set in a ratio of 80%:20%.
[0068] S3. Use deep learning methods to build a full-scale pore intelligent recognition model for complex buried hill reservoirs; the deep learning method includes using deep convolutional neural network deeplabv3+ to extract image features, combining dilated convolution and dilated spatial pyramid pooling techniques to capture multi-scale information, and using an encoding-decoding structure to improve the model's ability to segment objects of different sizes in the image;
[0069] The calculation formula of the dilated convolution is:
[0070] [y[i]=\sum_{k}x[i+r\cdot k]\cdot w[k]]
[0071] Among them, x is the input feature map, y is the output feature map, w is the convolution kernel, r is the dilation rate, and k is the index of the convolution kernel. By adjusting the dilation rate r, the receptive field size of the dilated convolution can be changed;
[0072] The global average pooling layer in the atrous spatial pyramid pooling module performs a global average operation on the input feature map to obtain the overall image features. The calculation formula is as follows:
[0073] [z_i=\frac{1}{H\times W}\sum_{j=1}^{H}\sum_{k=1}^{W}x_{ijk}]
[0074] Among them, x is the input feature map, z is the feature map after global average pooling, and H and W are the height and width of the feature map respectively.
[0075] S4. Use the training set and validation set to train the full-scale pore intelligent identification model for the complex buried hill reservoir. Use the training set for model training and the validation set for model performance verification. Use pixel accuracy, recall rate and mean intersection and union ratio to measure the performance of the model to determine whether the model meets the requirements. Select the model with high accuracy index of the validation set for the final model inference application. Calculate the error between model prediction and annotation at each step during training iteration, and use the gradient descent method to update the model parameters. In order to improve the generalization performance of the model, the input data is augmented using random truncation, random rotation, random flipping, and random color change. The training hyperparameters are as follows: the model input size is 384x384, the gradient descent optimizer is the Adam adaptive optimizer, the initial learning rate is set to 0.001, the total number of training iteration epochs is 100, and the L2 regularization parameter is set to 1.0×10 -5 The model training code is implemented using the Tensorflow deep learning framework.
[0076] As a preferred embodiment, assuming that the model outputs a total of N categories, the pixel set marked as the i-th category is recorded as Gi, the pixel set predicted as the i-th category is recorded as Fi, and |·| represents the number of elements in the set;
[0077] Pixel accuracy, calculate the ratio of correctly classified pixels to all pixels in the image. The calculation formula is as follows:
[0078]
[0079] Average recall rate, calculate the proportion of pixels in each category in the image that are correctly classified, and then find the average value. The calculation formula is as follows:
[0080]
[0081] Average intersection-to-union ratio: Calculate the ratio of the intersection and union between each type of prediction result and the labeled value in the image, and then find the average value. The calculation formula is as follows:
[0082]
[0083] The model with high accuracy index on the validation set is selected for the final model inference application.
[0084] Preferably, in step S4, pixel accuracy is the percentage of correctly labeled pixels to total pixels. The higher the PixelAccuracy, the more accurate the model is in segmenting the image. Recall is the proportion of samples whose predicted value is 1 and whose true value is also 1 to all samples whose true value is 1. IoU is calculated as the ratio of the intersection to the union between the predicted area and the actual area. Mean IoU is the average of the IoUs of all categories to balance the imbalance between different categories. The higher the Mean IoU, the more accurate the model is in segmenting different categories.
[0085] S5. Use the full-scale pore intelligent identification model of complex buried-hill reservoirs to identify pores, input the multi-level and multi-scale images to be detected, and obtain the pore radius distribution and pore volume distribution results of complex buried-hill reservoirs under multi-scale fusion. The pore radius distribution results of complex buried-hill reservoirs under multi-scale fusion are pie charts. Figure 3 The full-scale pore radius distribution range of the 3D micron CT core model is shown in Figure 8 Among them, A is the pore radius distribution of 50 nanometers to 100 nanometers, accounting for 6.4%; B is the pore radius distribution of 100 nanometers to 500 nanometers, accounting for 35.0%; C is the pore radius distribution of 500 nanometers to 20 microns, accounting for 57.4%; D is the pore radius distribution of 20 microns to 50 microns, accounting for 1.2%.
[0086] In summary, the present invention provides a full-scale pore structure calculation method 1 for complex buried hill reservoirs based on deep learning. It combines image recognition and deep learning algorithms to establish an intelligent pore recognition model, realize multi-scale characterization of the pore structure of complex buried hill reservoirs, greatly improve the work efficiency and work quality based on digital core image analysis of complex buried hill reservoirs, and provide strong support for the fine characterization of complex buried hill reservoirs.
[0087] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.
Claims
1. A method for calculating the full-scale pore structure of complex buried-hill reservoirs based on deep learning, characterized in that: The following steps are involved: S1. Carry out multi-scale and multi-resolution scanning on complex buried hill reservoir samples, obtain images of complex buried hill reservoir cores at different scales, and pre-process the obtained complex buried hill reservoir cores at full scale; S2, performing sample annotation on the preprocessed full-scale image, wherein the annotation target is the pores in the image, and performing polygon annotation on the pores in the image based on the DEXTR algorithm. After the annotation is completed, a multi-scale pore image dataset of a complex buried hill reservoir is obtained, and the multi-scale pore image dataset of the complex buried hill reservoir is divided into a training set and a validation set; S3. Use deep learning methods to build a full-scale pore intelligent identification model for complex buried hill reservoirs; S4. Use the training set and the validation set to train the complex buried hill reservoir full-scale pore intelligent identification model, use the training set to train the model, use the validation set to verify the model performance, use pixel accuracy, recall rate and mean intersection and union ratio to measure the performance of the model, determine whether the model meets the requirements, and select the model with high accuracy index of the validation set for the final model reasoning application; S5. Use the full-scale pore intelligent identification model of complex buried-hill reservoirs to identify pores, input the multi-level and multi-scale images to be detected respectively, and obtain the pore radius distribution and pore volume distribution results of the complex buried-hill reservoir under multi-scale fusion.
2. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 1, characterized in that: In step S1, the images of the complex buried hill reservoir core at different scales include three-dimensional images and two-dimensional scanning electron microscope images.
3. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 1 or 2, characterized in that: The preprocessing includes equalization processing, Gaussian filtering and median filtering; firstly, histogram equalization processing is used to make the light and dark contrast of the input image clear, then Gaussian filtering is used to suppress high-frequency noise in the image, and then median filtering is used to remove noise while retaining image edge information; The two-dimensional Gaussian function is: Where x, y are the template coordinates of the image pixels, and σ is the variance, which controls the width of the Gaussian function. Then the filtering result Y(x,y) of each point (x,y) in the Gaussian filtered image is calculated as: Where: X(x,y) is the grayscale value corresponding to pixel x,y, i,j is the pixel value separated from the domain. The filtering result Y(x,y) of each point (x,y) in the median filter image is calculated as follows: Y(x,y)=medianX(x+i,y+j)i∈[-K,K],j∈[-K,K] Where: K is the filter kernel size.
4. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 1, characterized in that: In step S2, the DEXTR algorithm automatically identifies the pore range in the annotation system and selects the leftmost, rightmost, topmost and bottommost edge points of the pore. The annotation system converts the points into onehot format and inputs them into the pre-trained segmentation model together with the original image to predict the pore boundary through the pre-trained segmentation model.
5. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 4 is characterized in that: In step S2, the complex buried hill reservoir multi-scale pore image dataset is divided into a training set and a validation set in a ratio of 80%:20%.
6. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 1, characterized in that: In step S3, the deep learning method includes extracting image features using a deep convolutional neural network deeplabv3+, combining dilated convolution and dilated spatial pyramid pooling techniques to capture multi-scale information, and using an encoding-decoding structure to improve the model's ability to segment objects of different sizes in the image; The calculation formula of the dilated convolution is: [y[i]=\sum_{k}x[i+r\cdot k]\cdot w[k]] Among them, x is the input feature map, y is the output feature map, w is the convolution kernel, r is the dilation rate, and k is the index of the convolution kernel. By adjusting the dilation rate r, the receptive field size of the dilated convolution can be changed; The global average pooling layer in the atrous spatial pyramid pooling module performs a global average operation on the input feature map to obtain the overall image features. The calculation formula is as follows: [z_i=\frac{1}{H\times W}\sum_{j=1}^{H}\sum_{k=1}^{W}x_{ijk}] Among them, x is the input feature map, z is the feature map after global average pooling, and H and W are the height and width of the feature map respectively.
7. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 1, characterized in that: In step S4, assuming that the model outputs a total of N categories, the set of pixels marked as the i-th category is denoted as Gi, the set of pixels predicted as the i-th category is denoted as Fi, and |·| represents the number of elements in the set; Pixel accuracy, calculate the ratio of correctly classified pixels to all pixels in the image. The calculation formula is as follows: Average recall rate, calculate the proportion of pixels in each category in the image that are correctly classified, and then find the average value. The calculation formula is as follows: Average intersection-to-union ratio: Calculate the ratio of the intersection and union between each type of prediction result and the labeled value in the image, and then find the average value. The calculation formula is as follows: The model with high accuracy index on the validation set is selected for the final model inference application.
8. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 7, characterized in that: In step S4, pixel accuracy is the percentage of correctly labeled pixels to total pixels. The higher the PixelAccuracy, the more accurate the model is in segmenting the image. Recall is the proportion of samples whose predicted value is 1 and whose true value is also 1 to all samples whose true value is 1. IoU calculates the ratio of the intersection and union between the predicted area and the actual area, while MeanIoU takes the average of the IoUs of all categories to balance the imbalance between different categories. The higher the Mean IoU, the more accurate the model's segmentation of different categories.
9. The method for calculating the full-scale pore structure of a complex buried hill reservoir based on deep learning according to claim 1, characterized in that: The pore radius distribution result of the complex buried hill reservoir under multi-scale fusion obtained in step S5 is a pie chart.