Carbon dioxide migration monitoring method based on CNN network combined with SVR micro-motion surface wave inversion

Abstract

The invention discloses a carbon dioxide migration monitoring method based on CNN network combined SVR micro-motion surface wave inversion, and the method comprises the steps: obtaining a theoretical underground shear wave velocity structure, employing a fast scalar transfer algorithm to forward simulate a frequency dispersion curve, and enabling the frequency dispersion curve to serve as a training set and a test set of the CNN network; constructing a CNN network, inputting the simulated frequency dispersion curve, and obtaining a feature vector with tag representativeness by using the CNN network, the feature vector being a frequency dispersion feature matrix; the SVR has certain difficulty for samples with large scale and multiple features, the CNN can reduce the randomness and non-stationarity of initial frequency dispersion data, and the SVR can map a nonlinear problem in an input space into a high-dimensional feature space and construct linear function discrimination, so that the global optimality of the algorithm is ensured, and the accuracy of the algorithm is improved. The method can obtain a high-precision and high-resolution shear wave velocity structure, and the calculation cost is low.

Description

Technical Field

[0001] The present invention relates to a method for monitoring carbon dioxide migration, and more specifically, the present invention mainly relates to a method for monitoring carbon dioxide migration based on CNN network combined with SVR micro-surface wave inversion. Background Art

[0002] Conventional seismic prediction technology for carbon dioxide storage and migration is mostly based on reflection wave seismic exploration. Although it has high resolution, the processing process is complicated and expensive. Micro-surface wave exploration can obtain the underground shear wave velocity structure, and it is easy to collect and relatively cheap. Therefore, in the seismic prediction technology for carbon dioxide storage and migration, the traditional linear surface wave inversion method has the disadvantage of falling into the local optimal solution and being unable to find the global optimal solution, while the nonlinear algorithm consumes a lot of computing costs and is not suitable for promotion and application. In this regard, machine learning will solve these two problems after establishing a suitable training set. The CNN network can share convolution kernels, is efficient in processing high-dimensional data, and can automatically extract the required features through convolution, but the application of the gradient descent algorithm will make the results converge to the local minimum. Therefore, it is necessary to further study and improve the seismic prediction technology for carbon dioxide storage and migration. Summary of the invention

[0003] One of the purposes of the present invention is to address the shortcomings and provide a monitoring method for carbon dioxide migration based on a CNN network combined with SVR micro-surface wave inversion, in order to solve the technical problems in the prior art that the traditional linear surface wave inversion method cannot be globally optimized and the nonlinear algorithm consumes a lot of computing costs.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] The present invention provides a method for monitoring carbon dioxide migration based on a CNN network combined with SVR micro-surface wave inversion, the method comprising the following steps:

[0006] Step A: Obtain the theoretical underground shear wave velocity structure and use the fast scalar transfer algorithm to simulate the dispersion curve as the training set and test set of the CNN network.

[0007] Step B: construct a CNN network, input the simulated dispersion curve, use the convolution layer and pooling layer of the CNN network to extract features and reduce the dimension of the dispersion curve data, and then use the Drop layer, smoothing layer and fully connected layer to generalize and converge the feature data to obtain a feature vector with representative labels, which is a dispersion feature matrix.

[0008] Step C: Input the feature vector into SVR for model training. SVR maps the low-dimensional data to a high-dimensional space, finds a linearly separable hyperplane in the high-dimensional space, and finally maps the hyperplane in the high-dimensional space back to the low-dimensional space. The underground shear wave velocity is obtained by the following formula:

[0009]

[0010] Where f(x) is the shear wave velocity, n is the number of dispersion points, and the value of i is 1, 2, ..., n, a i is the Lagrange coefficient introduced in the solution process and a i >0,y i is the sample data, is the kernel function, and b is the offset.

[0011] in,

[0012] Wherein, x is the eigenvector, T is the transposed dispersion characteristic matrix, and the value of i is 1, 2, ..., n;

[0013] Step D: Collect the micro-motion data before and after the injection of carbon dioxide in the work area and the simulated dispersion curve, respectively, and compare them with the corresponding underground shear wave velocity, and predict the carbon dioxide migration front according to the changes in the underground shear wave velocity before and after.

[0014] As a preferred embodiment, a further technical solution is: the vector regression model for SVR model training is:

[0015] f(x)=w T x n +b

[0016] Where f(x) is the shear wave velocity, x n is the dispersion characteristic matrix, n is the number of dispersion points, w T is the weight vector and b is the offset.

[0017] The distance formula of the above SVR is:

[0018] |f(x)-y i |=|w T x+by i |<ε

[0019] Where f(x) is the shear wave velocity, y i is the sample data, w T is the weight vector, b is the offset, x is the dispersion feature matrix, b is the offset, and ε is the deviation between the sample data and the predicted result.

[0020] A further technical solution is: the SVR is optimized by the following formula to obtain the underground shear wave velocity formula:

[0021]

[0022] Where C is the regularization coefficient, ξ i and i * Different slack variables are introduced for the upper and lower bounds of the current sample point interval, w is the weight vector, n is the number of dispersion points, i takes the value of 1, 2, ..., n, f(x i ) is the shear wave velocity corresponding to i, ε is the deviation between the sample data and the predicted result, y i is the sample data.

[0023] A further technical solution is: the theoretical underground shear wave velocity structure is to apply the refined layering theory, divide the underground strata into layers of equal or unequal thickness, and randomly set the strata shear wave velocity.

[0024] Compared with the prior art, one of the beneficial effects of the present invention is that since SVR has certain difficulties for samples with large scale and many features, CNN can reduce the randomness and non-stationarity of the initial dispersion data, which helps to improve the SVR inversion accuracy. In addition, SVR can map the nonlinear problem in the input space to the high-dimensional feature space and construct a linear function discriminant, and finally transform it into a convex optimization problem, which ensures the global optimality of the algorithm, so that the method can obtain high-precision and high-resolution shear wave velocity structure, thereby improving the prediction accuracy of carbon dioxide storage and migration, and the calculation cost is low.

[0025] At the same time, the refined stratification can not only reduce the inversion of stratum thickness and uncertainty, but also be more consistent with the actual underground situation, improve the efficiency of surface wave inversion, and enhance the resolution of surface wave inversion. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the CNN network model in the present invention.

[0027] Figure 2 It is a schematic diagram of micro-motion data in the present invention.

[0028] Figure 3 It is the dispersion curve diagram in the present invention.

[0029] Figure 4 This is a cross-sectional diagram of the shear wave velocity in the present invention. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings.

[0031] First, as mentioned in the background of the present invention, the traditional linear surface wave inversion method will fall into the local optimal solution and cannot find the global optimal solution, and the nonlinear algorithm will consume a large amount of computing costs. Machine learning will solve these two problems after establishing a suitable training set. The CNN network can share convolution kernels, which is efficient when processing high-dimensional data and can automatically extract the required features through convolution, but the result will converge to the local minimum due to the application of the gradient descent algorithm. SVR has certain difficulties for samples with large scale and many features, and CNN can reduce the randomness and non-stationarity of initial dispersion data, and its result helps to improve the SVR inversion accuracy. In addition, SVR can map the nonlinear problem in the input space to the high-dimensional feature space and construct a linear function to discriminate, which is finally converted into a convex optimization problem, ensuring the global optimality of the algorithm. The refinement of the stratification can not only reduce the inversion of the stratum thickness to reduce uncertainty but also be more in line with the underground real situation, improve the efficiency of surface wave inversion, and enhance the resolution of surface wave inversion.

[0032] Therefore, in a preferred embodiment of the present invention, the monitoring method of carbon dioxide migration based on CNN network combined with SVR micro-surface wave inversion is performed according to the following steps:

[0033] Step 1: Apply the refined stratification theory to stratify the underground strata into equal or unequal thickness strata, and randomly set the strata shear wave velocity to obtain a large number of theoretical underground shear wave velocity structures, which are the labels in network training; then use the conventional fast scalar transfer algorithm to forward simulate the dispersion curves, which will be used as the training set and test set in the following steps.

[0034] Step 2: Construct a CNN network (i.e., convolutional neural network), such as Figure 1 As shown in the figure, the dispersion curve data of the training set is input, and the convolution layer and pooling layer are used to extract features and reduce the dimension of the dispersion curve data. Then, the Drop layer, smoothing layer and fully connected layer are used to generalize and converge the feature data to obtain a feature vector that is more representative of the label.

[0035] Step 3: Input the feature vector output from step 2 into SVR for model training. In nonlinear regression problems, SVR (Support Vector Regression Machine) usually maps low-dimensional data to high-dimensional space, finds a linearly separable hyperplane in the high-dimensional space, and finally maps the hyperplane in the high-dimensional space back to the low-dimensional space.

[0036] Its regression model is:

[0037] f(x)=w T x n +b (1)

[0038] Where f(x) is the predicted result, i.e., the shear wave velocity, x nis the dispersion characteristic matrix, n is the number of dispersion points, w T is the weight vector and b is the offset.

[0039] The distance formula is:

[0040] The distance formula of the above SVR is:

[0041] |f(x)-y i |=|w T x+by i |<ε (2)

[0042] Where f(x) is the shear wave velocity, y i is the sample data, w T is the weight vector, b is the offset, x is the dispersion feature matrix, b is the offset, and ε is the deviation between the sample data and the predicted result. Thus, an interval band with a width of 2ε is constructed so that as many sample points as possible fall into it. In this way, the prediction is considered correct and no loss is calculated.

[0043] In order to reduce the regression error, different slack variables ξ are introduced at the upper and lower bounds of the sample point interval. i and i * , so that points not in the regression band can also satisfy the constraints, the SVR optimization problem can be formalized as:

[0044]

[0045] Where C is the regularization coefficient, which is also the penalty coefficient of the error term in the SVR model and is generally greater than zero; i and i * Different slack variables are introduced for the upper and lower bounds of the current sample point interval, w is the weight vector, n is the number of dispersion points, i takes the value of 1, 2, ..., n, f(x i ) is the shear wave velocity corresponding to i, ε is the deviation between the sample data and the predicted result, y i is the sample data.

[0046] The above equations (3) and (4) are the real optimization problems of SVR. In order to solve the SVR optimization problem, the Lagrangian function and kernel function are introduced to map low dimensions to high dimensions and convert nonlinearity into linearity. The RBF kernel function is used for nonlinear mapping to obtain the final decision function, as shown in equation (5).

[0047]

[0048] Where f(x) is the shear wave velocity, n is the number of dispersion points, and the value of i is 1, 2, ..., n, a iis the Lagrange coefficient introduced in the solution process and a i >0,y i is the sample data, is the kernel function, and b is the offset.

[0049] in, Where x is the above eigenvector, T is the transpose of the above dispersion characteristic matrix, and the value of i is 1, 2,…, n.

[0050] In this step, the large number of training sets (dispersion curve data) and labels (shear wave velocity structure) generated in step 1 are used to find the best mapping relationship between the output shear wave velocity and the input dispersion curve through the CNN-SVR network prediction and optimization in steps 2 and 3, and complete the CNN-SVR network training. Then, the test set data and the corresponding label data can be used to test the prediction accuracy of the current network model.

[0051] Step 4: Use the collected micro-vibration data before and after the injection of carbon dioxide in the work area. The aforementioned micro-vibration data is a weak vibration signal on the earth's surface. In this step, it is actual data collected directly from the field. Figure 2 The dispersion curves are extracted from the micromotion data of the three stations using the spatial autocorrelation method ( Figure 3 The CNN-SVR network formed by step 2 and step 3 is used to calculate the shear wave velocity structure of the formation before and after carbon dioxide injection ( Figure 4 As shown in Figure 2, the leading edge of the shear wave velocity change in the target layer is the carbon dioxide migration front. Figure 4 The left side in the middle shows the background shear wave velocity, and the right side shows the shear wave velocity after carbon dioxide injection.

[0052] In addition to the above, it should be noted that "one embodiment", "another embodiment", "embodiment", etc. mentioned in this specification refer to the specific features, structures or characteristics described in conjunction with the embodiment included in at least one embodiment generally described in this application. The same expression appearing in multiple places in the specification does not necessarily refer to the same embodiment. Further, when describing a specific feature, structure or characteristic in conjunction with any embodiment, it is claimed that the realization of such feature, structure or characteristic in conjunction with other embodiments also falls within the scope of the present invention.

[0053] Although the present invention is described herein with reference to a number of illustrative embodiments of the present invention, it will be appreciated that those skilled in the art may devise many other modifications and implementations that fall within the scope and spirit of the principles disclosed herein. More specifically, within the scope of the present disclosure, drawings, and claims, a variety of variations and improvements may be made to the components and / or layout of the subject combination layout. In addition to variations and improvements made to the components and / or layout, other uses will also be apparent to those skilled in the art.

Claims

1. A monitoring method for carbon dioxide migration based on CNN network combined with SVR micro-surface wave inversion, characterized in that The method includes: The theoretical underground shear wave velocity structure is obtained, and the dispersion curve is simulated forward using the fast scalar transfer algorithm as the training set and test set of the CNN network; Constructing a CNN network, inputting the simulated dispersion curve, using the convolution layer and pooling layer of the CNN network to perform feature extraction and data dimension reduction on the dispersion curve data, and then using the Drop layer, smoothing layer and fully connected layer to generalize and converge the feature data to obtain a feature vector with representative labels, wherein the feature vector is a dispersion feature matrix; The feature vector is input into SVR for model training. SVR maps the low-dimensional data to a high-dimensional space, finds a linearly separable hyperplane in the high-dimensional space, and finally maps the hyperplane in the high-dimensional space back to the low-dimensional space. The underground shear wave velocity is obtained by the following formula: Where f(x) is the shear wave velocity, n is the number of dispersion points, and the value of i is 1, 2, ..., n, a i is the Lagrange coefficient introduced in the solution process and a i >0,y i is the sample data, is the kernel function, b is the offset; in, In the formula, x is the eigenvector, T is the transpose of the dispersion characteristic matrix, and the value of i is 1, 2, ..., n. By collecting the micro-motion data before and after the injection of carbon dioxide in the work area and the dispersion curve extracted from the micro-motion data, respectively, and corresponding to the underground shear wave velocity, the carbon dioxide migration front is predicted according to the changes in the underground shear wave velocities before and after.

2. The method for monitoring carbon dioxide migration based on CNN network combined with SVR micro-surface wave inversion according to claim 1 is characterized by: The vector regression model used by the SVR for model training is: f(x)=w T x n +b Where f(x) is the shear wave velocity, x n is the dispersion characteristic matrix, n is the number of dispersion points, w T is the weight vector, b is the offset; The distance formula of the SVR is: |f(x)-y i |=|w T x+by i |<ε Where f(x) is the shear wave velocity, y i is the sample data, w T is the weight vector, b is the offset, x is the dispersion feature matrix, b is the offset, and ε is the deviation between the sample data and the predicted result.

3. The method for monitoring carbon dioxide migration based on CNN network combined with SVR micro-surface wave inversion according to claim 1 is characterized by: The SVR is optimized by the following formula to obtain the underground shear wave velocity: Where C is the regularization coefficient, ξ i and i * Different slack variables are introduced for the upper and lower bounds of the current sample point interval, w is the weight vector, n is the number of dispersion points, i takes the value of 1, 2, ..., n, f(x i ) is the shear wave velocity corresponding to i, ε is the deviation between the sample data and the predicted result, y i is the sample data.

4. The method for monitoring carbon dioxide migration based on CNN network combined with SVR micro-surface wave inversion according to claim 1 is characterized by: The theoretical underground shear wave velocity structure is obtained by applying the refined layering theory, dividing the underground strata into layers of equal or unequal thickness, and randomly setting the strata shear wave velocity.

Images