Physical and data combined super-resolution dynamic digital core intelligent reconstruction method
By combining low-resolution and high-resolution CT scanning and deep learning networks and combined with lattice Boltzmann method, the problem of dynamic CT scanning technology not being able to identify small pores and fluid changes in ultra-low permeability reservoirs is solved, achieving high-precision Dynamic digital core intelligent reconstruction.
Patent Information
- Application Number
- CN202411121714.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2024-11-26
AI Technical Summary
The existing dynamic CT scanning technology cannot effectively identify the pores and fluid changes in ultra-low permeability reservoirs, resulting in incomplete dynamic digital core model.
Using a super-resolution dynamic digital core intelligent reconstruction method combining physics and data, using J-net static convolutional neural network and U-net dynamic fine-tuning convolutional neural network for images through a combination of low-resolution and high-resolution CT scanning. Reconstruction and real-time calculation of permeability was performed in combination with the lattice Boltzmann method.
The reproduction of pixel-level details smaller than the original image is achieved, the core micro pore recognition capability and fluid change characteristics are improved, and the resolution and accuracy of dynamic digital core models are improved.
Smart Images

Figure CN119023529A_ABST
Abstract
Description
Technical Field
[0001] The invention provides a physical and data combined super-resolution dynamic digital core intelligent reconstruction method, which belongs to the oil field technical field Background Art
[0002] With the continuous improvement of the requirements for digital transformation of the petroleum industry, digital core technology will receive more attention. As an advanced technical means, digital core reconstructs the three-dimensional microstructure model of reservoir rocks through physical and mathematical methods, and studies the microscopic pore characteristics of the reservoir intuitively and accurately. With the further development of reservoirs, it is necessary to further deepen the study of the three-dimensional distribution characteristics of fluids in the core during displacement development. During the displacement process, the flow in the core changes dynamically, and the digital core model that depicts the dynamic changes of fluids in the core is called dynamic digital core. Usually, digital cores are obtained by CT scanning and other methods, and three-dimensional reconstruction is performed after data is obtained. Dynamic digital cores are systematically scanned in three dimensions at different displacement time points as the displacement experiment proceeds, giving the three-dimensional CT data of the core at different times during the displacement process. However, displacement CT currently faces the problems of representativeness and resolution. The size of the scanned sample is too small to represent the physical properties of the region, and it is not easy to carry out displacement experiments. The size of the selected scanning experiment sample is too large, and due to the limitations of current CT scanning technology, it is impossible to effectively identify fluid changes in smaller pores. How to effectively characterize the changes of fluid in pores during the development process and construct a relatively complete dynamic digital core model is the frontier and difficulty of current dynamic digital core research.
[0003] At present, dynamic digital core is mainly obtained by combining displacement and CT. During the displacement process, CT 3D scanning is performed, and then 3D reconstruction is performed using the 3D scanning data to obtain 3D digital core models at different times. The digital core models at different times are arranged into a sequence to obtain a dynamic digital core model. The disadvantages of this technology include:
[0004] 1) Existing dynamic CT scanning is limited by the resolution of the equipment and cannot be applied to ultra-low permeability reservoirs. Its resolution cannot identify smaller pores and can only identify pores larger than its resolution. For example, the resolution of existing general dynamic CT scanning equipment is about 1-10μm, while there are a large number of pores smaller than 1μm in ultra-low permeability cores. These pores smaller than 1μm cannot be effectively identified using existing dynamic CT scanning technology. This results in the inability of existing dynamic digital cores to establish a relatively complete three-dimensional dynamic digital core model.
[0005] 2) In addition to characterizing the pore structure, dynamic CT also needs to characterize the changes in fluid distribution during the dynamic displacement process, that is, the pore radius distribution characteristics of the fluid storage space, and the change characteristics of the pore radius distribution of the fluid storage space during the displacement process. Existing technologies are limited by resolution and cannot effectively identify the fluid change characteristics in smaller gaps. Summary of the invention
[0006] Existing dynamic CT scanning is limited by the resolution of the equipment and cannot be applied to ultra-low permeability reservoirs. Its resolution cannot resolve smaller pores, and it can only identify pores larger than its resolution, and cannot effectively identify the fluid change characteristics in smaller gaps. In order to solve the technical bottleneck of dynamic CT scanning equipment, the problem that the reconstructed digital core cannot meet the "resolution" and "representativeness" requirements, the present invention provides a super-resolution dynamic digital core intelligent reconstruction method that combines physics and data.
[0007] The specific technical solution is:
[0008] The super-resolution dynamic digital core intelligent reconstruction method based on physics and data, with the following specific steps:
[0009] Step 1: Sample preparation: Prepare the core sample into a core column with a height h and a center radius r for dynamic CT scanning.
[0010] Step 2: Sample preparation: First, clean, dry and weigh the core sample, evacuate the core at room temperature to saturate it, take out the sample after saturation and weigh it again to determine the saturated water content of the core, and add CT developer with a concentration of C.
[0011] Step 3: Place the vacuum-saturated core into the core holder and use the target displacement pressure difference P i Carry out displacement experiment, balance for more than n hours, every δ t time, record the displaced liquid flow rate Q(P i ,t i ),t i =1....t n , t is the displacement time, i is the different displacement pressure difference values.
[0012] Step 4: Data acquisition: In the rock physics experiment, the core CT 3D scanning is carried out in real time under the displacement state, and the image is reconstructed in 3D to obtain the data volume. : Phase 1 scanning, using low-resolution dynamic CT research: low-resolution sample scanning, dynamic CT scanning and periodic state CT scanning under the displacement state, and digital core 3D reconstruction, D L (P i ,t n ), the size of the data body is D LM *D LN *DLP , with a resolution of δ Lx , D LM , D LN , D LP are the length, width and height scale of the sample in low resolution scan pixels, respectively.
[0013] Step 5: Low-resolution dynamic CT scan image preprocessing: adjust the brightness and contrast of the original image, and then complete filtering and noise reduction to further segment the scanned image.
[0014] Step 6: Binarization of core CT images: Segment the preprocessed image region and divide the image into solid skeletons D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n ,S3) three types.
[0015] Step 7: Change displacement pressure P i P i+1 , repeat steps 2-6 to get P i+1 Low-resolution 3D reconstruction image D under displacement pressure difference L (P i+1 ,t n ) and the segmented image solid skeleton D LS (P i+1 ,t n , S1), displaced fluid D LS (P i+1 ,t n , S2), displacement fluid D LS (P i+1 ,t n , S3), S1 is the permeability of the solid skeleton, S2 is the permeability of the displaced fluid, and S3 is the permeability of the displacing fluid.
[0016] Step 8: Repeat steps 2-7 to complete all pressure difference conditions P i (i=0, 2, 4, 6) Low-resolution 3D reconstructed image and segmented image: D L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n, S2), displacement fluid D LS (P i ,t n ,S3).
[0017] Step 9: Second stage scanning: Repeat steps 2-8, but change the low-resolution scan in step 5 to a high-resolution dynamic CT study: The high-resolution dynamic CT study is to conduct partial core excavation on the basis of the low-resolution study and establish a high-resolution dynamic CT scanning experiment. Perform dynamic CT scanning and periodic CT scanning under displacement conditions, and perform three-dimensional reconstruction of digital cores. H (P i ,t n ), the size of the data body is D HM *D HN *D HP , with a resolution of δ Hx , ,D HM , D HN , D HP are the length, width and height scale of the sample's high-resolution scan pixels, respectively.
[0018] Step 10: Use the registration algorithm to convert D H (P i ,t n ) and the segmented image solid skeleton D HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid D HS (P i ,t n ,S3) and D HS (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n ,S3), perform position alignment, where i=0, 2, 4, 6.
[0019] Step 11: Based on step 10, L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,tn , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n ,S3),(i=0,2,4,6) cut and D HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid D HS (P i ,t n ,S3) and D HS (P i ,t n ) CT image samples in the same position space: D LH (P i ,t n ) and D LHS (P i ,t n , S1), displaced fluid D LHS (P i ,t n , S2), displacement fluid D LHS (P i ,t n , S3). Thus, an in-situ paired low-resolution and high-resolution data pair is formed: D LH (P i ,t n )~D HS (P i ,t n , S1), displaced fluid D LHS (P i ,t n ,S2)~D HS (P i ,t n , S2), displacement fluid D LHS (P i ,t n ,S3)~D HS (P i ,t n ,S3) and D LHS (P i ,t n )~D HS (P i ,t n ).
[0020] Step 12: Steps 2-11 are to build a training data set. Deep learning requires a large amount of core sample data, so the data set is expanded. The method of translation and division combined with three-dimensional rotation is used for data expansion. First, an inscribed cuboid (cuboid size L) is taken from the sample plunger area. A ×L B ×L C ), and then cut the cube from the cuboid step by step (the size is: is the stepping distance, V is the number of moves, and finally the sample plunger is rotated by an angle of χ. The above operation is repeated until the sample is rotated 360° in total, and then repeated for all samples to achieve data amplification.
[0021] Step 13: Data preprocessing is required before putting it into the model: This includes data standardization or normalization to reduce the dimensional differences between different features and increase the diversity of the data. Use the z-score normalization method: The numerical range is scaled to near 0, but the data distribution characteristics are not changed. The processed data conforms to the normal distribution, x is the data point, μ is the mean of the data set, and σ is the standard deviation of the data set.
[0022] Step 14: Data division after data preprocessing: In order to evaluate the performance of the model, the preprocessed data needs to be divided into a training set and a test set with a division ratio of 8:2. The training set is used to train the model, and the test set is used to finally evaluate the generalization ability of the model.
[0023] Step 15: Construct a J-net static convolutional neural network model, which mainly includes two parts: downsampling part and upsampling part. The downsampling part specifically includes convolution operation, activation function and pooling operation. In the convolution layer, the convolution kernel is used to extract data features. The convolution kernel is a weight matrix with a specific distribution. The original data is weighted with a specific distribution to achieve image feature extraction; the activation function introduces a nonlinear transformation to map the parameter features to the corresponding output end, which is a nonlinear mapping of the output result of the convolution layer; pooling is responsible for downsampling the output of the convolution layer to reduce the data space size and computational complexity. The upsampling part includes deconvolution, connection and convolution operations. Deconvolution is to restore the size of the feature map in the upsampling stage to the size of the input image; the connection operation is to connect the feature results obtained by each convolution operation in the downsampling stage with the corresponding deconvolution results in the upsampling stage to achieve multi-scale prediction.
[0024] Step 16: In the J-net static network, the generated network sets J_K layers of deconvolution layers, sets the first layer convolution kernel size to M×M, the step size to X, the input data size to a×b×c, and the output size to (B c ×N1)×a1×b1×c1,Bc Let Batchsize be, and N be the number of feature maps extracted by each layer of the convolutional network; set the size of the convolutional kernel in the second layer to M×M, the stride to X, and the input data size to (B c ×N1)×a1×b1×c1, and the output size to (B c ×N2)×a2×b2×c2... The size of the convolutional kernel in the Qth layer is M×M, the stride is set to X, and the input data size is (B c ×N Q-1 )×a Q-1 ×b Q-1 ×c Q-1 , and the output size is (B c ×N Q )×a Q ×b Q ×c Q , until the size of the convolutional kernel in the Kth layer is set to M×M, the stride is set to X, and the input data size is (B c ×N k-1 )×a k-1 ×b k-1 ×c k-1 , and the output size is (B c ×N k )×a k ×b k ×c k ; Select the SAME form for Padding in the convolutional operation to ensure that the size of the input image and the output result is the same, set the non-linear activation function, use the transposed convolution operation instead of the max pooling operation to "magnify" the feature map obtained from the previous layer, and reduce the number of channels to 1 / 2 of the original, and increase the resolution by N * times, set the size of the pooling window to S×S, and the size of the dropout ratio parameter to F.
[0025] Step 17: Set J_Q (Q < K) convolutional layers in the discriminator network. Set the size of the convolutional kernel in the first layer to M×M, the stride to X, and the input data size to (B c ×N Q )×a Q ×b Q ×c Q , and the output size is (B c ×N Q-1 )×a Q-1 ×b Q-1 ×c Q-1 ; Set the size of the convolutional kernel in the second layer to M×M, the stride to X, and the input data size to (B c ×N Q-1 )×a Q-1 ×b Q-1 ×c Q-1, the output size is (B c ×N Q-2 )×a Q-2 ×b Q-2 ×c Q-2 ; Set the third layer convolution kernel size to M×M, the step size to X, and the input data size to (B c ×N Q-2 )×a Q-2 ×bi-2×ci-2, the output size is (B c ×N i-3 )×a Q-3 ×b Q-3 ×c Q-3 , set the Qth layer convolution kernel size to M×M, the step size to X, and the input data size to (B c ×N2)×a2×b2×c2, (B c ×N)×a1×b1×c1. After each convolution operation, the number of channels of the feature map doubles. After two consecutive convolution layers, a maximum pooling layer is set. The deconvolution operation window size is S×S, and the dropout ratio parameter size is F. The results of the two layers are connected in series with a connecting line between upsampling and downsampling, so that the network can integrate features of different scales. In the J-net convolutional neural network, the number of generating network layers is greater than the number of discriminating network layers. On the basis of symmetry, it continues to expand to achieve the effect of improving resolution.
[0026] Step 18: Steps 13-17 are to train the J-net static network model, and then use the low-resolution dynamic CT scan data in step 8:D L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS ((P i ,t n ,S3),i=1...I) is input into the trained J-net network model to complete the reconstruction of the low-resolution dynamic CT model.
[0027] Step 19: Repeat step 18 and replace the input data with the high-resolution segmented image solid skeleton D completed in step 11 HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid DHS (P i ,t n ,S3), complete the high-resolution dynamic CT model reconstruction.
[0028] Step 20: Construct a U-net dynamic fine-tuning convolutional neural network for the reconstructed high-resolution dynamic CT model. The network structure is similar to that of J-net, including two parts: downsampling and upsampling. The specific network model design is as follows: The generated network is set with U_K * Deconvolution layer, set the first convolution kernel size to M * ×M * , the step size is set to X * , the input data size is (B * c ×N * 1)×a * 1×b * 1×c * 1, the output size is (B * c ×N * 2)×a * 2×b * 2×c * 2...The Qth layer convolution kernel size is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * Q-1 )×a * Q-1 ×b * Q-1 ×c * Q-1 , the output size is (B * c ×N * Q )×a * Q ×b * Q ×c * Q , until the Kth * The convolution kernel size is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * k-1 )×a * k-1 ×b *k-1 ×c * k-1 , the output size is (B * c ×N * k )×a k ×b k ×c k . B * c is Batchsize, N * The number of feature maps extracted by each layer of the convolutional network. The Padding of the convolution operation selects the SAME form to ensure that the input image and the output result are of the same size, set the nonlinear activation function, use the deconvolution operation instead of the maximum pooling operation to "enlarge" the feature map obtained in the previous layer, and the number of channels is reduced to 1 / 2 of the original, and the resolution is increased by N. *〃 times, the pooling window size is set to S * ×S * , the dropout ratio parameter size is F * .
[0029] Step 21: Determine if the network has U_K * The size of the first convolution kernel is M. * ×M * , the step size is set to X * , the input data size is (B * c ×N * Q )×a * Q ×b * Q ×c * Q , the output size is (B * c ×N * Q-1 )×a * Q-1 ×b * Q-1 ×c * Q-1 ; The size of the second convolution kernel is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * Q-1 )×a * Q-1 ×b * Q-1 ×c* Q-1 , the output size is (B * c ×N * Q-2 )×a * Q-2 ×b * Q-2 ×c * Q-2 , until the Kth * The convolution kernel size of the layer is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * 2)×a * 2×b * 2×c * 2. Output size is (B * c ×N * )×a * 1×b * 1×c * 1. After each convolution operation, the number of channels of the feature map becomes twice as much as before. After two consecutive convolution layers, a maximum pooling layer is set, and the deconvolution operation window size is S. * ×S * , the dropout ratio parameter size is F * .
[0030] Step 22: For the dynamically fine-tuned high-resolution dynamic CT reconstruction model, the lattice Boltzmann method (LBM) is used to calculate the model permeability in real time. LBM is a discrete format for solving the continuous Boltzmann equation. The distribution function of particles is updated through two basic steps: collision and streaming. According to the principle of mass conservation and Taylor expansion, the Boltzmann equation can be expressed as follows: where f(r, t) represents the number of particles whose velocity is between [ξ, ξ+dξ] in the volume unit [r, r+dr] at time t; a is the acceleration; ξ is the particle velocity; Ω is the collision function.
[0031]
[0032] Step 23: Verify the Lattice Boltzmann Method (LBM) to check the authenticity of the calculation results of the LBM. Conduct a simulation experiment of the Poiseuille flow in a circular tube to verify the correctness of the LBM. The Poiseuille flow velocity and the distance from the central axis r can be expressed by the formula Then the theoretical value of flow velocity is compared with the LBM simulation value to verify the effectiveness of the LBM method. is the pressure gradient; u r is the flow velocity at a distance r from the central axis; D is the diameter of the tube; μ is the dynamic viscosity coefficient of the fluid.
[0033] Step 24: In step 3, record the displaced liquid flow rate Q (P i ,t i ),t i =1…t n The displaced liquid flow rate Q output in the core reconstructed based on the U-net output by the LBM model LBM (P i ,t i ), the value calculated by the root mean square error method is used as the dynamic LOSS part, and the dynamic fine-tuning U-net convolutional network is reversely trained based on LOSS.
[0034] Step 25: For the fine-tuned U-net convolutional network parameter settings, refer to steps 16-17. Continue to propagate forward to obtain the prediction results after calculation and processing of each layer. Calculate the value of the loss function based on the application constraints and reconstructed prediction results. Calculate the gradient of the loss function to the network parameters through the back propagation algorithm. Update the network weights and biases according to the gradient. Set n epoch training. Each epoch training link includes m cycles until the model's performance on the validation set reaches the optimal level.
[0035] Step 26: When the difference LOSS value between the displaced liquid flow rate in the reconstructed core output by U-net based on the LBM model and the displaced liquid flow rate under different pressure differences recorded in step 3 is less than ε, the dynamic fine-tuning of the high-resolution dynamic reconstruction model is completed, otherwise return to the U-net dynamic training network to continue repeating steps 23-24 and continue training.
[0036] Step 27: Input the test set data into the network model, test the fully convolutional deep neural network model, and evaluate whether the recognition accuracy of the U-Net fully convolutional deep neural network model meets the preset requirements. Finally, complete the super-resolution dynamic digital core intelligent reconstruction based on the combination of physics and data, and output the network model and results.
[0037] The present invention has the following technical effects:
[0038] 1. The present invention proposes a dynamic CT reservoir micro-pore characterization method, which performs dynamic CT three-dimensional scanning of the core by changing the fluid filling pressure and the solution saturation concentration, and performs three-dimensional reconstruction of the dynamic CT image under the displacement state.
[0039] 2. The present invention proposes a J-net network model, which is an improvement over the traditional U-net model. Unlike the symmetrical structure of U-net, J-net introduces an additional convolutional layer in its decoder part, so that the output image can be scaled up and a clearer and more delicate image can be generated. This image processing super-resolution method using deep learning allows details smaller than the pixel level of the original image to be reproduced.
[0040] 3. The present invention proposes a dynamic and static combined intelligent optimization method based on LBM, which calculates the parameters under the constraints in real time, compares them with the given constraints, and reversely optimizes the neural network. LBM has higher parallel efficiency than the microscopic two-phase seepage simulation based on finite elements and higher accuracy than the ball-and-stick model. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 is a flow chart of the present invention;
[0042] Figure 2 It is a schematic diagram of the J-net network structure of the present invention. DETAILED DESCRIPTION
[0043] The following is a clear and complete description of the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, but not all of them:
[0044] like Figure 1 As shown, the present invention comprises the following steps:
[0045] The first step is to create a training dataset:
[0046] S1. Prepare the core sample into a core column with a height of 2 cm and a center radius of 2 mm for dynamic CT scanning.
[0047] S2. The core sample is then cleaned, dried and weighed, and the core is vacuum-saturated at room temperature. After saturation, the sample is taken out and weighed again to determine the saturated water volume of the core, and a CT developing agent with a concentration of 60% KI and a radioactive compound is added.
[0048] S3. Place the vacuum-saturated core into the core holder and conduct displacement experiments using the target displacement pressure difference P0 (i=0). Balance for more than 4 hours and record the outlet displacement liquid flow rate Q(P0,t0) every 20 minutes. i =0, 1, 2…t n
[0049] S4. Data collection: In rock physics experiments, real-time CT 3D scanning of the core is performed under displacement conditions, and 3D image reconstruction is performed to obtain the data volume. The first stage of scanning uses low-resolution dynamic CT research: low-resolution sample scanning, dynamic CT scanning and periodic CT scanning under displacement conditions, and 3D digital core reconstruction. L (P0,t n ), the size of the data volume is 2000 pixels × 2000 pixels × 2000 pixels, and the resolution is 3.0 μm.
[0050] S5. Preprocess the low-resolution dynamic CT scan image: adjust the brightness and contrast of the original image, and then complete filtering and noise reduction to further segment the scanned image. The filtering and noise reduction method uses the median filtering method, which replaces the pixel value in the image with the median of its adjacent pixel values to achieve noise reduction. After multiple filtering, the comprehensive output result.
[0051] S6, then binarize the reconstructed core CT image: segment the pre-processed image region and divide the image into solid skeletons D LS (P0,t n , S1), displaced fluid D LS (P0,t n , S2), displacement fluid D LS (P0,t n , S3) three types, use the threshold method to count the gray values of different regions, use multiple thresholds to distinguish different regions, here set the solid skeleton D LS = 1, the displaced fluid D LS is 0.
[0052] S7, change the displacement pressure from 0MPA to 2MPA, 4MPA, 6MPA, repeat S2-S6, complete all pressure difference conditions P i 3D reconstructed image and segmented image under (i=0, 2, 4, 6MPA): D L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n ,S3),i=0,2,4,6MPA.
[0053] S8, second stage scanning: repeat steps S2-S7, change the low-resolution scanning in S5 to high-resolution dynamic CT research: high-resolution dynamic CT research is to conduct partial core excavation on the basis of low-resolution research, and establish high-resolution dynamic CT scanning experiment. Perform dynamic CT scanning and periodic CT scanning under displacement state, and perform digital core three-dimensional reconstruction, D H (P i ,t n ), the size of the data volume is 2000 pixels × 2000 pixels × 2000 pixels, and the resolution is 0.25 μm.
[0054] S9, using the registration algorithm, the low-resolution D H (P i ,t n ) and the segmented image solid skeleton D HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid D HS (P i ,t n ,S3) and D HS (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n ,S3), perform position alignment, where i = 0, 2, 4, 6MPA.
[0055] S10. In establishing a high-resolution dynamic CT scanning experiment, a registration algorithm (scale-invariant feature transformation algorithm) is used to perform position registration between high-resolution and low-resolution images.
[0056] S11. Based on S10, a CT image sample in the same spatial position as the positioned high-resolution image is cut from the low-resolution image, thereby forming an in-situ paired low-resolution and high-resolution dynamic CT data pair.
[0057] S12. Deep learning requires a large amount of core sample data, so the data set is expanded. The method of translation and cutting combined with three-dimensional rotation is used for data expansion. First, an inscribed cuboid is taken from the sample plunger area (the cuboid size is 700×700×1024), and then a cube (700+(4*102))×(700+(4*102))×(1024+(4*102))( is the stepping distance of 102 pixels, V is the number of moves), and finally the sample plunger is rotated 30 degrees. The above operation is repeated until the sample is rotated 360° in total, and then repeated for all samples to achieve data amplification.
[0058] Then, we will carry out super-resolution dynamic core intelligent reconstruction based on the joint driving of physics and data:
[0059] S13. After the dataset is established, the dataset is preprocessed: This includes data standardization or normalization to reduce the dimensional differences between different features and increase the diversity of the data, using the z-score normalization method.
[0060] S14. Then divide the data into training set and test set with a division ratio of 8:2. The training set is used to train the model, and the test set is used to finally evaluate the generalization ability of the model.
[0061] S15. Construct a J-net static convolutional neural network model, which mainly includes two parts: downsampling part and upsampling part. The downsampling part specifically includes convolution operation, activation function and pooling operation. In the convolution layer, the convolution kernel is used to extract data features. The convolution kernel is a weight matrix with a specific distribution. The original data is weighted with a specific distribution to achieve image feature extraction; the activation function introduces a nonlinear transformation to map the parameter features to the corresponding output end, which is a nonlinear mapping of the output result of the convolution layer; pooling is responsible for downsampling the output of the convolution layer to reduce the data space size and computational complexity. The upsampling part includes deconvolution, connection and convolution operations. Deconvolution is to restore the size of the feature map in the upsampling stage to the size of the input image; the connection operation is to connect the feature results obtained by each convolution operation in the downsampling stage with the corresponding deconvolution results in the upsampling stage to achieve multi-scale prediction.
[0062] S16. First train the J-net static convolutional neural network model and adjust the parameters. Set the generated network to have a total of 7 deconvolution layers, set the first layer convolution kernel size to 3×3, the step size to 2, the input data size to 800×1×1×1, and the output size to (64×16)×5×5×5, where N is the number of feature maps extracted by each layer of the convolutional network; set the second layer convolution kernel size to 3×3, the step size to 2, the input data size to (64×16)×5×5×5, and the output size to (64×8)×11×11×11; set the third layer convolution kernel size to 3×3, the step size to 2, the input data size to (64×8)×11×11×11, and the output size to (64×4)×23×23×23; set the fourth layer convolution kernel size to 3×3, the step size to 2, the input data size The convolution kernel size of the fifth layer is set to 3×3, the step size is set to 2, the input data size is (64×2)47×47×47, and the output size is (64×1)×95×95×95; the convolution kernel size of the sixth layer is set to 3×3, the step size is set to 2, the input data size is (64×1)×95×95×95, and the output size is 32×191×191×191; the convolution kernel size of the seventh layer is set to 3×3, the step size is set to 2, the input data size is 32×191×191×191, and the output size is 1×385×385×385; the activation function of the first layer is ReLU, dropout ratio is 0.1, the second layer activation function is ReLU, the dropout ratio is 0.2, the third layer activation function is ReLU, the dropout ratio is 0.2, the fourth layer activation function is ReLU, the dropout ratio is 0.3, the fifth layer activation function is ReLU, the dropout ratio is 0.4, the sixth layer activation function is ReLU, the dropout ratio is 0.4 and the seventh layer activation function is sigmoid, the dropout ratio is 0.5, the ReLU activation function equation is as follows: The sigmoid function equation is as follows:
[0063] S17, set the discriminant network, with a total of 6 convolutional layers. The first layer sets the convolution kernel size to 3×3, the step size to 2, the input data size to 1×191×191×191, and the output size to (64×1)×95×95×95; the second layer sets the convolution kernel size to 3×3, the step size to 3, the input data size to (64×1)×95×95×95, and the output size to (64×2)×47×47×47; the third layer sets the convolution kernel size to 3×3, the step size to 2, the input data size to (64×2)×47×47×47, and the output size to (64×4)×23 ×23×23; set the fourth layer convolution kernel size to 3×3, the step size to 2, the input data size to (64×4)×23×23×23, and the output size to (64×8)×11×11×11; set the fifth layer convolution kernel size to 3×3, the step size to 2, the input data size to (64×8)×11×11×11, and the output size to (64×16)×5×5×5; set the sixth layer convolution kernel size to 3×3, the step size to X, the input data to (64×16)×5×5×5, and the output to 800×1×1×1. The first layer's activation function is LeakyReLU, and the dropout ratio is 0.5. The second layer's activation function is LeakyReLU, and the dropout ratio is 0.3. The third layer's activation function is LeakyReLU, and the dropout ratio is 0.3. The fourth layer's activation function is LeakyReLU, and the dropout ratio is 0.2. The fifth layer's activation function is LeakyReLU, and the dropout ratio is 0.2. The sixth layer's activation function is sigmoid, and the dropout ratio is 0.1. The LeakyReLU function equation is:
[0064] The network structure is "J" shaped, such as Figure 2 shown.
[0065] S18. After training the static J-net network model, the low-resolution dynamic CT scan data: D L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n,S3),(i=0, 2, 4, 6MPA) is input into the trained J-net network model to complete the reconstruction of the low-resolution dynamic CT model;
[0066] S19, replace the input data with a high-resolution segmented image solid skeleton D HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid D HS (P i ,t n ,S3), complete the reconstruction of high-resolution dynamic CT model;
[0067] S20. Construct a U-net dynamic fine-tuning convolutional neural network for the reconstructed high-resolution dynamic CT model. The U-net generation network can be set to six layers, the input data size is 800×1×1×1, the output size is (64×16)×5×5×5, and N is the number of feature maps extracted by each layer of the convolutional network; set the second layer convolution kernel size to 3×3, the step size to 2, the input data size is (64×16)×5×5×5, and the output size is (64×8)×11×11×11; set the third layer convolution kernel size to 3×3, the step size to 2, the input data size is (64×8)×11×11×11, and the output size is (64×8)×11×11×11. The output size is (64×4)×23×23×23; the fourth layer convolution kernel size is set to 3×3, the step size is set to 2, the input data size is (64×4)×23×23×23, and the output size is (64×2)×47×47×47; the fifth layer convolution kernel size is set to 3×3, the step size is set to 2, the input data size is (64×2)47×47×47, and the output size is (64×1)×95×95×95; the sixth layer convolution kernel size is set to 3×3, the step size is set to 2, the input data size is (64×1)×95×95×95, and the output size is 32×191×191×191. The first layer's activation function is LeakyReLU, and the dropout ratio is 0.1; the second layer's activation function is LeakyReLU, and the dropout ratio is 0.2; the third layer's activation function is LeakyReLU, and the dropout ratio is 0.2; the fourth layer's activation function is LeakyReLU, and the dropout ratio is 0.3; the fifth layer's activation function is LeakyReLU, and the dropout ratio is 0.4; the sixth layer's activation function is Tanh, and the dropout ratio is 0.4. The Tanh function equation is:
[0068] S21, U-net discriminant network, a total of 6 convolutional layers. The first layer sets the convolution kernel size to 3×3, the step size to 2, the input data size to 1×191×191×191, and the output size to (64×1)×95×95×95; the second layer sets the convolution kernel size to 3×3, the step size to 3, the input data size to (64×1)×95×95×95, and the output size to (64×2)×47×47×47; the third layer sets the convolution kernel size to 3×3, the step size to 2, the input data size to (64×2)×47×47×47, and the output size to (64×4)×23 ×23×23; set the fourth layer convolution kernel size to 3×3, the step size to 2, the input data size to (64×4)×23×23×23, and the output size to (64×8)×11×11×11; set the fifth layer convolution kernel size to 3×3, the step size to 2, the input data size to (64×8)×11×11×11, and the output size to (64×16)×5×5×5; set the sixth layer convolution kernel size to 3×3, the step size to X, the input data to (64×16)×5×5×5, and the output to 800×1×1×1. The first layer's activation function is LeakyReLU, and the dropout ratio is 0.5; the second layer's activation function is LeakyReLU, and the dropout ratio is 0.3; the third layer's activation function is LeakyReLU, and the dropoutratio is 0.3; the fourth layer's activation function is LeakyReLU, and the dropoutratio is 0.2; the fifth layer's activation function is LeakyReLU, and the dropoutratio is 0.2; the sixth layer's activation function is sigmoid, and the dropoutratio is 0.1.
[0069] S22. For the dynamically fine-tuned high-resolution dynamic CT reconstruction model, the lattice Boltzmann method (LBM) is used to calculate the model permeability in real time. The following equation is solved continuously, where f(r, t) represents the number of particles with velocities between [ξ, ξ+dξ] in the volume unit [r, r+dr] at time t; a is the acceleration; ξ is the particle velocity; Ω is the collision function.
[0070]
[0071] S23. Verify the lattice Boltzmann method (LBM), calculate the dynamic permeability of the fine-tuned high-resolution reconstruction model, and verify the authenticity of the calculation results of the lattice Boltzmann method. Conduct a circular pipe Poiseuille flow simulation experiment to verify the correctness of the LBM. The Poiseuille flow velocity and the distance from the central axis r can be expressed by the formula Then the theoretical flow velocity value is compared with the LBM simulation value, and the comparison result can be output in a chart.
[0072] S24: In step S3, the flow rate Q (P) of the displaced liquid at the outlet under different pressure differences is recorded. i ,t i ),t i =1…t n The displaced liquid flow rate Q output in the core reconstructed based on the U-net output by the LBM model LBM (P i ,t i ), the root mean square error method is used to calculate the difference between the two as the dynamic LOSS part, and the dynamic fine-tuning U-net convolutional network is reversely trained based on the LOSS value.
[0073] S25. For the fine-tuned U-net convolutional network parameter setting, refer to the J-net network model parameters in step 1. Continue to propagate forward and obtain the prediction results after calculation and processing at each layer. Set the model to process 4 data samples at a time (batch). The generator network is trained once for each batch, with an initial learning rate of 0.0006 and a batch decay strategy. The discriminator network is trained once for every 5 batches, with a learning rate of 0.0002.
[0074] The loss function is calculated based on the constraints and the reconstructed prediction results (the loss function uses the mean square error (MSE)), and the gradient of the loss function to the network parameters is calculated through the back propagation algorithm (such as the gradient descent algorithm, the Adam optimizer is used to update the parameters, and the parameter value is set to 0.5). The weights and biases of the network are updated according to the gradient, and 100 epoch training is set. Each epoch training link includes 50 cycles until the performance of the model on the validation set reaches the optimal level; the root mean square error calculation equation is:
[0075] S26. When the difference LOSS value between the displaced liquid flow rate in the reconstructed core output by U-net based on the LBM model and the displaced liquid flow rate under different pressure differences recorded in step S3 is less than 0.01, the dynamic fine-tuning of the high-resolution dynamic reconstruction model is completed, otherwise, the U-net training network is returned to continue to repeat steps 24-26 for back-propagation update.
[0076] S27. Input the test set data into the convolutional deep neural network model for testing. Comparison results show that after using the J-net network model on the low-resolution digital core, the solid skeleton recognition rate of the digital core image is improved by 18.63%, and the reconstruction accuracy of the super-resolution digital core model exceeds 80%, which better meets the super-resolution dynamic digital core intelligent reconstruction under the constraint conditions, and finally outputs the network model.
[0077] The present invention proposes a core resolution enhancement method based on J-net deep learning. J-net is a fully convolutional neural network model based on an encoding-decoding architecture, which can achieve mapping from signal to image. The model amplifies the machine-generated data on the generator network structure based on the U-net symmetric structure, improves the output image resolution, generates more delicate and clear images, can better identify the tiny pores in the core, and achieve super-resolution network reconstruction.
[0078] The present invention proposes a dynamic and static joint intelligent optimization method based on LBM. The method is a discrete format for continuously solving the Boltzmann equation. It updates the distribution function of particles through two basic steps, collision and flow. It can have good numerical stability and intuitive physical expression for reconstructing network permeability, and can be directly used for CT scanning data without simplifying complex spatial structures. It has higher authenticity and effectively identifies the characteristics of fluid changes in smaller gaps.
[0079] The present invention proposes a pore reconstruction method for dynamic CT of low-permeability reservoirs. The method is to increase the CT signal of tight sandstone pore fluid by 40-50 times by optimizing the saturated solution concentration and adjusting the injection gas pressure; dynamic CT three-dimensional scanning is performed under displacement conditions to obtain experimental data at different times under different displacement pressure differences. The experimental data is dynamically reconstructed at low-high resolution, and then driven jointly by physics and data to obtain a super-resolution dynamic digital core intelligent reconstruction network, which realizes high-precision identification of low-permeability reservoir pores and completes quantitative analysis at a low cost.
Claims
1. A super-resolution dynamic digital core intelligent reconstruction method combining physics and data, characterized in that: Including the following steps: Step 1: Sample preparation: Prepare the core sample into a core column with a height of h and a center radius of r; Step 2: Sample preparation: First, clean, dry, and weigh the core sample. Evacuate and saturate the core under normal temperature. After saturation, take out the sample and weigh it again to determine the saturated water volume of the core. Add a CT imaging agent with a concentration of C; Step 3: Place the vacuum-saturated core into the core holder and use the target displacement pressure difference P i Carry out displacement experiment, balance for more than n hours, every δ t time, record the displaced liquid flow rate Q(P i ,t i ),t i =1....t n , t is the displacement time, i is the different displacement pressure difference values; Step 4: Data acquisition: In the rock physics experiment, real-time CT 3D scanning of the core is performed under displacement conditions, and 3D image reconstruction is performed to obtain the data volume; in the first stage of scanning, low-resolution dynamic CT research is used: low-resolution sample scanning, dynamic CT scanning and periodic CT scanning are performed under displacement conditions, and digital core 3D reconstruction is performed, D L (P i ,t n ), the size of the data body is D LM *D LN *D LP , with a resolution of δ Lx , D LM , D LN , D LP are the length, width and height scale of the low-resolution scanned pixels of the sample, respectively; Step 5: Preprocessing of low-resolution dynamic CT scan images: Adjust the brightness and contrast of the original images, and then complete filtering and noise reduction; Step 6: Binarization of core CT images: Segment the preprocessed image region and divide the image into solid skeletons D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n , S3), S1 is the permeability of the solid skeleton, S2 is the permeability of the displaced fluid, and S3 is the permeability of the displacing fluid; Step 7: Change displacement pressure P i P i+2 , repeat steps 2-6 to get P i+2 Low-resolution 3D reconstruction image D under displacement pressure difference L (P i+2 ,t n ) and the segmented image solid skeleton D LS (P i+2 ,t n , S1), displaced fluid D LS (P i+2 ,t n , S2), displacement fluid D LS (P i+2 ,t n ,S3); Step 8: Repeat steps 2-7 to complete all pressure difference conditions P i (i=0, 2, 4, 6) Low-resolution 3D reconstructed image and segmented image: D L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n ,S3); Step 9: Second stage scanning: Repeat steps 2-8, but change the low-resolution scan in step 5 to a high-resolution dynamic CT study: The high-resolution dynamic CT study is to conduct partial core excavation on the basis of the low-resolution study and establish a high-resolution dynamic CT scanning experiment; perform dynamic CT scanning and periodic CT scanning under displacement conditions, and perform digital core 3D reconstruction. H (P i ,t n ), the size of the data body is D HM *D HN *D HP , with a resolution of δ Hx , D HM , D HN , D HP are the length, width and height scale of the high-resolution scanned pixels of the sample, respectively; Step 10: Use the registration algorithm to convert D H (P i ,t n ) and the segmented image solid skeleton D HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid D HS (P i ,t n ,S3) and D HS (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n ,S3), perform position registration, where i = 0, 2, 4, 6; Step 11: Based on step 10, L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS (P i ,t n , S3) cut and D HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid D HS (P i ,t n ,S3) and D HS (P i ,t n ) CT image samples in the same position space: D LH (P i ,t n ) and D LHS (P i ,t n , S1), displaced fluid D LHS (P i ,t n , S2), displacement fluid D LHS (P i ,t n , S3), thus forming an in-situ paired low-resolution and high-resolution data pair: D LH (P i ,t n )~D HS (P i ,t n , S1), displaced fluid D LHS (P i ,t n ,S2)~D HS (P i ,t n , S2), displacement fluid D LHS (P i ,t n ,S3)~D HS (P i ,t n ,S3) and D LHS (P i ,t n )~D HS (P i ,t n ); Step 12: Amplify the data set; use the method of translation and division combined with three-dimensional rotation to amplify the data. First, take an inscribed cuboid from the sample plunger area. The cuboid size L A ×L B ×L C , and then step and cut the cube from the cuboid is the stepping distance, V is the number of moves, and finally the sample plunger is rotated by an angle of x. The above operation is repeated until the sample is rotated 360° in total, and then the operation is repeated for all samples to achieve data amplification; Step 13: Data preprocessing needs to be carried out before putting it into the model to reduce the dimensionality difference between different features and increase the diversity of data; Step 14: After data preprocessing, perform data division: Divide the preprocessed data into a training set and a test set, with a division ratio of 8:2; the training set is used to train the model, and the test set is used to finally evaluate the generalization ability of the model; Step 15: Construct a J-net static convolutional neural network model, including two parts: a downsampling part and an upsampling part; Step 16: In the J-net static network, the generation network sets J_K deconvolution layers…; Step 17: The discriminant network sets J_Q convolutional layers, Q < K, where K is the number of deconvolution layers of the generation network and Q is the number of convolutional layers of the discriminant network; Step 18: Convert the low-resolution dynamic CT scan data from step 8 to: L (P i ,t n ) and the segmented image solid skeleton D LS (P i ,t n , S1), displaced fluid D LS (P i ,t n , S2), displacement fluid D LS ((P i ,t n ,S3),i=0,2,4,6) are input into the trained J-net network model to complete the reconstruction of the low-resolution dynamic CT model; Step 19: Repeat step 18 and replace the input data with the high-resolution segmented image solid skeleton D completed in step 11 HS (P i ,t n , S1), displaced fluid D HS (P i ,t n , S2), displacement fluid D HS (P i ,t n ,S3), complete the reconstruction of high-resolution dynamic CT model; Step 20: Construct a U-net dynamic fine-tuning convolutional neural network for the reconstructed high-resolution dynamic CT model, including two parts: a downsampling part and an upsampling part; Step 21: The generator network and the discriminator network are set with U_K * layer convolutional layer, K * is the number of deconvolution layers of the generative network and the number of convolution layers of the discriminative network; Step 22: For the dynamically fine-tuned high-resolution dynamic CT reconstruction model, use the lattice Boltzmann method (LBM) to calculate the permeability of the model in real time; Step 23: Verify the lattice Boltzmann method (LBM) to check the authenticity of the calculation results of the lattice Boltzmann method; Step 24: In step 3, record the displaced liquid flow rate Q (P i ,t i ),t i =1…t n ; The displaced liquid flow Q output in the reconstructed core based on the U-net output by the LBM model LBM (P i ,t i ), the difference between the two is calculated by the root mean square error method as the dynamic LOSS part, and the dynamic fine-tuning U-net convolutional network is reversely trained based on LOSS; Step 25: Set the parameters of the fine-tuned U-net convolutional network with reference to steps 16-17. Continue to propagate forward through the calculations and processes of each layer to obtain the prediction results. Calculate the value of the loss function according to the application constraints and the reconstructed prediction results. Calculate the gradient of the loss function with respect to the network parameters through the backpropagation algorithm. Update the weights and biases of the network according to the gradient. Set n epochs for training, and each epoch training session includes m loops until the performance of the model on the validation set reaches the optimal; Step 26: When the difference LOSS value between the displaced liquid flow rate in the reconstructed core output by the U-net based on the LBM model and the displaced liquid flow rate recorded in step 3 under different pressure differences is less than ε, complete the dynamic fine-tuning of the high-resolution dynamic reconstruction model. Otherwise, return to the U-net dynamic training network and continue to repeat steps 23-24 for continuous training; Step 27: Input the test set data into the network model to test the fully convolutional deep neural network model, and evaluate whether the recognition accuracy of the U-Net fully convolutional deep neural network model meets the preset requirements; finally, complete the intelligent reconstruction of the super-resolution dynamic digital core based on the combination of physics and data, and output the network model and results.
2. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: Step 13 uses the z-score normalization method: The numerical range is scaled to near 0, but the data distribution characteristics are not changed. The processed data conforms to the normal distribution, x is the data point, μ is the mean of the data set, and σ is the standard deviation of the data set.
3. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: The two parts of the J-net static convolutional neural network model in step 15 are respectively: The downsampling part specifically includes convolution operation, activation function and pooling operation. In the convolution layer, the convolution kernel is used to extract data features. The convolution kernel is a weight matrix with a specific distribution. The original data is weighted with a specific distribution to achieve image feature extraction. The activation function introduces a nonlinear transformation to map the parameter features to the corresponding output end, which is a nonlinear mapping of the output results of the convolutional layer; Pooling is responsible for downsampling the output of the convolutional layer to reduce the data space size and computational complexity; The upsampling part includes deconvolution, concatenation and convolution operations. Deconvolution is to restore the size of the feature map in the upsampling stage to the size of the input image. The connection operation is to connect the feature results obtained by each convolution operation in the downsampling stage with the corresponding deconvolution results in the upsampling stage to achieve multi-scale prediction.
4. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: Step 16 is as follows: set the size of the first convolution kernel to M×M, the step size to X, the input data size to a×b×c, and the output size to (B c ×N1)×a1×b1×c1,B c is Batchsize, N is the number of feature maps extracted by each layer of the convolutional network; the second layer convolution kernel size is set to M×M, the step size is set to X, and the input data size is (B c ×N1)×a1×b1×c1, the output size is (B c ×N2)×a2×b2×c2...The Qth layer convolution kernel size is M×M, the step size is set to X, and the input data size is (B c ×N Q-1 )×a Q-1 ×b Q-1 ×c Q-1 , the output size is (B c ×N Q )×a Q ×b Q ×c Q , until the Kth layer convolution kernel size is set to M×M, the step size is set to X, and the input data size is (B c ×N k-1 )×a k-1 ×b k-1 ×c k-1 , the output size is (B c ×N k )×a k ×b k ×c k ; The Padding of the convolution operation selects the SAME form to ensure that the input image and the output result have the same size, set the nonlinear activation function, use the deconvolution operation instead of the maximum pooling operation to "amplify" the feature map obtained in the previous layer, and the number of channels is reduced to 1 / 2 of the original, and the resolution is increased by N * times, the pooling window size is set to S×S, and the dropout ratio parameter size is F.
5. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: Step 17 is as follows: The first layer sets the convolution kernel size to M×M, the step size to X, and the input data size to (B c ×N Q )×a Q ×b Q ×c Q , the output size is (B c ×N Q-1 )×a Q-1 ×b Q-1 ×c Q-1 ; Set the second layer convolution kernel size to M×M, the step size to X, and the input data size to (B c ×N Q-1 )×a Q-1 ×b Q-1 ×c Q-1 , the output size is (B c ×N Q-2 )×a Q-2 ×b Q-2 ×c Q-2 ; Set the third layer convolution kernel size to M×M, the step size to X, and the input data size to (B c ×N Q-2 )×a Q-2 ×bi-2×ci-2, the output size is (B c ×N i-3 )×a Q-3 ×b Q-3 ×c Q-3 , set the Qth layer convolution kernel size to M×M, the step size to X, and the input data size to (B c ×N2)×a2×b2×c2, (B c ×N)×a1×b1×c1; after each convolution operation, the number of channels of the feature map becomes twice as much as before, a maximum pooling layer is set after two consecutive convolution layers, the deconvolution operation window size is S×S, and the dropout ratio parameter size is F; the results of the two layers are connected in series with a connecting line between upsampling and downsampling, so that the network can fuse features of different scales. The number of generating network layers in the J-net convolutional neural network is greater than the number of discriminative network layers, and it continues to expand on the basis of symmetry.
6. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: The specific network model design in step 20 is as follows: The generated network setting has U_K * Deconvolution layer, set the first convolution kernel size to M * ×M * , the step size is set to X * , the input data size is (B * c ×N * 1)×a * 1×b * 1×c * 1, the output size is (B * c ×N * 2)×a * 2×b * 2×c * 2...The Qth layer convolution kernel size is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * Q-1 )×a * Q-1 ×b * Q-1 ×c * Q-1 , the output size is (B * c ×N * Q )×a * Q ×b * Q ×c * Q , until the Kth * The convolution kernel size is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * k-1 )×a * k-1 ×b * k-1 ×c * k-1 , the output size is (B * c ×N * k )×a k ×b k ×c k ; B * c is Batchsize, N * The number of feature maps extracted by each layer of the convolutional network; the padding of the convolution operation selects the SAME form to ensure that the input image and the output result are of the same size, set the nonlinear activation function, use the deconvolution operation instead of the maximum pooling operation to "enlarge" the feature map obtained in the previous layer, and the number of channels is reduced to 1 / 2 of the original, and the resolution is increased by N *〃 times, the pooling window size is set to S * ×S * , the dropout ratio parameter size is F * .
7. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: Step 21 is specifically as follows: The size of the first convolution kernel is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * Q )×a * Q ×b * Q ×c * Q , the output size is (B * c ×N * Q-1 )×a * Q-1 ×b * Q-1 ×c * Q-1 ; The size of the second convolution kernel is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * Q-1 )×a * Q-1 ×b * Q-1 ×c * Q-1 , the output size is (B * c ×N * Q-2 )×a * Q-2 ×b * Q-2 ×c * Q-2 , until the Kth * The convolution kernel size of the layer is M * ×M * , the step size is set to X * , the input data size is (B * c ×N * 2)×a * 2×b * 2×c * 2. Output size is (B * c ×N * )×a * 1×b * 1×c * 1. After each convolution operation, the number of channels of the feature map becomes twice as much as before. After two consecutive convolution layers, a maximum pooling layer is set, and the deconvolution operation window size is S. * ×S * , the dropout ratio parameter size is F * .
8. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: Step 22: The specific method is: The Boltzmann equation can be expressed as follows: Where f(r, t) represents the number of particles with velocities between [ξ, ξ+dξ] in the volume unit [r, r+dr] at time t; a is the acceleration; ξ is the particle velocity; and Ω is the collision function.
9. The physical and data combined super-resolution dynamic digital core intelligent reconstruction method according to claim 1 is characterized in that: Step 23: The specific method is to conduct a circular tube Poiseuille flow simulation experiment to verify the correctness of LBM. The Poiseuille flow velocity and the distance from the central axis r are expressed by the formula Indicates that is the pressure gradient; u r is the flow velocity at a distance r from the central axis; D is the diameter of the tube; μ is the dynamic viscosity coefficient of the fluid. Then the theoretical flow velocity values were compared with the LBM simulation values to verify the effectiveness of the LBM method.
