Intelligent seismic velocity modeling system based on attention mechanism and well-seismic data constraint

Abstract

The application provides an intelligent seismic velocity modeling method based on an attention mechanism and well-seismic data constraint, comprising the following steps: constructing a target function of the seismic velocity modeling method, and obtaining logging data, migration data, migration velocity data, well-seismic depth error data and weight parameters of a TripleANet; constructing a well circumference data random sampler to randomly sample the migration data, the migration velocity data, the logging data and the well-seismic depth error data; calculating a predicted velocity data block, an error constraint term loss of the logging data and a loss of a well-seismic error constraint term; calculating a gradient of the weight parameters of the TripleANet and updating the weight parameters of the TripleANet; and calculating a predicted velocity model. The intelligent seismic velocity modeling method based on the attention mechanism and the well-seismic data constraint improves the seismic velocity modeling efficiency and can further improve the representation learning capability of three-dimensional data.

Description

Technical Field

[0001] This invention relates to the field of seismic exploration technology, and in particular to an intelligent seismic velocity modeling system based on attention mechanisms and well-seismic data constraints. Background Technology

[0002] Seismic imaging is a process that indirectly samples the subsurface medium using artificially generated seismic waves, and then uses the recorded seismic data to image the sampled strata. Accurate seismic propagation velocity is crucial to seismic imaging technology. With the development of seismic imaging technology, these techniques have placed higher demands on seismic velocity models. Traditional seismic velocity modeling methods, such as tomographic inversion, can construct smooth, low-frequency seismic velocity models, but are not suitable for wave equation-based seismic imaging techniques such as reverse time migration, which require higher imaging accuracy. Although full-waveform inversion techniques can be used to invert low- and mid-frequency information from seismic velocity models, full-waveform inversion faces challenges such as strong nonlinearity, multiple solutions, and high computational cost. Therefore, how to construct seismic velocity modeling techniques for high-precision seismic imaging has received increasing attention and research.

[0003] In recent years, deep learning has achieved great success in image recognition and natural language processing, which has promoted the development of intelligent seismic velocity modeling methods based on deep learning. Intelligent seismic velocity modeling has evolved from a purely data-driven, end-to-end approach to a dual-driven learning paradigm of data and physical models. However, purely data-driven methods require a large number of training samples, and the trained neural networks suffer from poor generalization when applied to new exploration areas. The dual-driven approach uses neural networks to assist in physical-driven seismic velocity modeling, but it suffers from high computational cost and unclear physical operating mechanisms.

[0004] Chinese patent application CN202410193117.X discloses a model generation method, a well-seismic combined velocity modeling method, apparatus, and equipment. The method includes: constructing a texture transfer network model, the texture transfer network comprising a generator, a discriminator, and an autoencoder; the generator and discriminator forming a conditional generative adversarial network, and the autoencoder reconstructing logging velocities; acquiring sample data, inputting the sample data into the generator, and training the texture transfer network model based on the conditional generative adversarial network formed by the generator and discriminator; the sample data including initial velocity and depth profiles; inputting the logging velocities into the autoencoder to optimize the parameters of the texture transfer network model, obtaining a trained texture transfer network model; the trained texture transfer network model outputting a velocity modeling image based on the input initial velocity and depth profiles, wherein the velocity in the velocity modeling image matches the reflection surface morphology. The aforementioned patents only utilize well logging information, which has the problem of insufficient training samples. In addition, generative adversarial networks are mainly used to estimate the probability distribution of samples and generate new samples. Generative adversarial networks have poor training stability and low efficiency when the amount of sample data is small.

[0005] Chinese patent application CN201910216730.8 discloses a multi-information fusion seismic velocity modeling method. The method includes: employing a strategy of setting virtual wells to increase training well information; filling the velocity model with virtual structural wells and real wells along the structural dip angle; then, based on the Gaussian function, calculating the weighting coefficients of each well logging data, and summing the weighted values ​​to obtain a well logging velocity interpolation model; further fusing the well logging velocity interpolation model with prior migration velocities, and statistically analyzing the ratio of the result to the prior migration velocities, adjusting the result based on the ratio to obtain a mid-deep velocity model; determining the fusion top surface and fusion region of the near-surface model based on the ray coverage of the near-surface model; obtaining a unified full velocity field by fusing near-surface velocities and mid-deep velocities in the fusion region, which is then output as the final result, and deriving the relative error between the model and the well logging velocities. The aforementioned patent has shortcomings such as high dependence on the work area for the selection criteria of virtual well settings, large errors in the weighting coefficients obtained by the Gaussian function, and the selection of too many prior parameters in the overall process.

[0006] The existing technologies described above are significantly different from this invention. A search reveals no literature in the XY category, indicating the innovativeness of this invention. Since existing technologies lack solutions to the technical problem we aim to address, we have invented a novel intelligent seismic velocity modeling method based on an attention mechanism and well-seismic data constraints. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent seismic velocity modeling method based on attention mechanism and well-seismic data constraints, which improves the migration velocity model by utilizing existing 3D migration data, 3D migration velocity, well logging data and well seismic error data through deep learning, and realizes intelligent modeling of 3D velocity model.

[0008] The objective of this invention can be achieved through the following technical measures: an intelligent seismic velocity modeling method based on attention mechanism and well-seismic data constraints, comprising:

[0009] Step 1: Construct the objective function for the seismic velocity modeling method;

[0010] Step 2: Obtain well logging data, migration data, migration velocity data, well-seismic depth error data, and weight parameters of TripleANet;

[0011] Step 3: Construct a wellbore data random sampler to randomly sample migration data, migration velocity data, logging data, and well seismic depth error data;

[0012] Step 4: Calculate the predicted velocity data block, the loss of the error constraint term of the logging data, and the loss of the well-seismic error constraint term;

[0013] Step 5: Calculate the gradient of the TripleANet weight parameters and update the TripleANet weight parameters;

[0014] Step 6: Calculate the predicted velocity model.

[0015] The objective of this invention can also be achieved through the following technical measures:

[0016] In step 1, the objective function of the intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints is constructed, specifically including:

[0017] Step 1.1: Set TripleANet in Step 1 to the function f(:w) to predict velocity data. The calculation formula is in and It belongs to offset data and offset speed data The offset data blocks and offset velocity blocks are defined, where t < i, w < c, h < z represents the size of the data blocks, and the actual and predicted velocity models are defined as follows: Real speed data w are the learnable weight parameters of TripleANet;

[0018] S1.2 The formula for calculating the error constraint between well logging data blocks and predicted velocity data is as follows:

[0019]

[0020] Where ||·||1 is the L1 norm, M j For the label mask, ⊙ represents element-wise multiplication, and y... j For predicting speed data, For well logging data blocks;

[0021] S1.3 The formula for calculating the loss of the well vibration error constraint term is:

[0022]

[0023] Where ||·|| is the L2 norm, This is a block of wellbore seismic depth error data; the predicted wellbore seismic depth error is... The calculation formula is:

[0024]

[0025] Where d represents the depth interval size, and n represents the number of depth intervals in the target layer. This represents the prediction velocity corresponding to each depth interval;

[0026] S1.4, The objective function is:

[0027]

[0028] S1.5, The formula for updating the weight parameters is:

[0029]

[0030] Where η is the iteration update step size, w k+1 These are the network parameters for the (k+1)th iteration.

[0031] In step 1.1, this pertains to well logging data. The logging data block is represented as and use a label mask The value is 1 if the logging data block is not empty, and 0 otherwise; this belongs to the well-seismic depth error data. The well depth error data block is represented as Where m represents the number of well seismic depth error data.

[0032] Step 3, the specific steps for constructing the well perimeter data random sampler include:

[0033] Step 3.1: Set the logging data block to the starting position of the logging data as (wp) i wp c wp zIf the position sampling interval is [wp], then the position sampling interval can be set to [wp]. i -t,wp i ],[wp c -w,wp c ],[wp z -h,wp z ], where wp i -t>0,wp c -w>0,wp z -h>0,wp i <i,wp c <c,wp z <z;

[0034] Step 3.2: Set the location random variable p t ~U(wp) i -t,wp i ),p w ~U(wp) c -w,wp c ), p h ~U(wp) z -h,wp z ), where p t ~U(wp) i -t, wp i ) represents a random variable p t Follow the interval [wp] i -t, wp i Uniform distribution on ], p w ~U(wp) c -w,wp c ) represents a random variable p w Follow the interval [wp] c -w,wp c Uniform distribution on ], p h ~U(wp) z -h,wp z ) represents a random variable p h Follow the interval [wp] z -h,wp z Uniform distribution on ];

[0035] Step 3.3: Based on Step 3.2, data positions can be randomly sampled, with the starting position represented as (p). t p w p h The endpoint position is represented as (p) t +t, p w +w, p h +h);

[0036] Step 3.4: Based on the starting and ending positions in 3.3, sample and segment the logging data, migration data, migration velocity data, and well-seismic depth error data to obtain sonic logging data blocks, migration data blocks, migration velocity data blocks, and well-seismic depth error data blocks.

[0037] In step 4, based on the weight parameters of TripleANet obtained in step 2 and the offset data block and offset velocity data block obtained in step 3, the predicted velocity data block is calculated according to step 1.1.

[0038] In step 4, based on the logging data block obtained in step 3 and the predicted velocity data block obtained in step 4, the error constraint term loss of the logging data is calculated according to step 1.2.

[0039] In step 4, based on the well-seismic depth error data block obtained in step 3 and the predicted velocity data block obtained in step 4, the loss of the well-seismic error constraint term is calculated according to step 1.3.

[0040] In step 5, based on the logging data error and well vibration error constraint term loss obtained in step 4, the gradient of the TripleANet weight parameters is calculated according to steps 1.4 and 1.5, and the weight parameters of TripleANet are updated.

[0041] The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints also includes repeating steps 2 to 5 after step 5 to iteratively update the weight parameters of TripleANet, and stopping the calculation when the number of iterations or the error meets the exit condition.

[0042] In step 6, the predicted velocity model is calculated based on the offset data, offset velocity data and weight parameters of TripleANet obtained in step 2.

[0043] The objective of this invention can also be achieved through the following technical measures: an intelligent seismic velocity modeling system based on attention mechanism and well seismic data constraints, wherein the intelligent seismic velocity modeling system based on attention mechanism and well seismic data constraints uses an intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints to model seismic velocity.

[0044] The intelligent seismic velocity modeling method based on attention mechanisms and well-seismic data constraints in this invention considers the use of deep learning to construct a seismic velocity modeling learning paradigm from acquired and processed sonic logging data and migration data. This method improves the accuracy of existing seismic velocity models based on readily available data and avoids additional data acquisition and generation, thereby enhancing the efficiency of seismic velocity modeling. Compared to autoencoder networks (ANet) built on convolutional neural networks, this invention utilizes axis self-attention and an autoencoder architecture to build a TripleANet, which further improves the representation learning capability of 3D data.

[0045] The present invention discloses an intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints. It applies TripleANet and well seismic data constraints to seismic velocity modeling, uses the TripleANet network to learn and extract the representation of the input data, and uses well seismic data information to design sonic logging data error and well seismic depth error constraint terms, thereby forming an intelligent seismic velocity modeling learning paradigm based on well seismic data.

[0046] The intelligent seismic velocity modeling method based on attention mechanism and well-seismic data constraints described in this invention effectively utilizes the feature learning capability of the TripleANet network to construct a mapping relationship between migration data and well logging data, and uses a well perimeter data random sampler to randomly sample and segment the data for calculation, effectively reducing the computational burden and thus improving computational efficiency.

[0047] The intelligent seismic velocity modeling method based on attention mechanism and well-seismic data constraints described in this invention effectively utilizes well logging data and well-seismic depth error data, and effectively constrains the mapping solution space between migration data and well logging data, thereby improving the seismic velocity prediction performance. Attached Figure Description

[0048] Figure 1 This is a block diagram of an intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints as described in this invention.

[0049] Figure 2 This is a flowchart of an intelligent seismic velocity modeling and calculation process based on an attention mechanism and well seismic data constraints, as described in this invention.

[0050] Figure 3 This is a diagram of the TripleANet structure described in this invention;

[0051] Figure 4 This is a schematic diagram illustrating the visualization of SEAM data and the prediction velocity model in a specific implementation method.

[0052] Figure 5This is a comparison chart of the actual velocity, ANet predicted velocity, and TripleANet predicted velocity curves at well positions (200, 385) and (350, 200) in a specific implementation.

[0053] Figure 6 A partial comparison diagram of SEAM velocity prediction in a specific implementation;

[0054] Figure 7 This is a comparison chart of the offset data and well seismic error between the actual velocity and the offset velocity in a specific implementation method.

[0055] Figure 8 This is a comparison chart of the deviation data of the actual velocity and the predicted velocity in a specific implementation method, showing the well seismic error.

[0056] Figure 9 This is a comparison chart of the offset data of the actual velocity and the predicted velocity under well-seismic error constraints in a specific implementation method.

[0057] Figure 10 This is a comparison chart of well vibration errors in Implementation Method 1;

[0058] Figure 11 This is a flowchart of a specific embodiment of the intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints of the present invention. Detailed Implementation

[0059] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0060] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0061] like Figure 11 As shown, Figure 11 This is a flowchart of the intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints of the present invention. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints includes:

[0062] Step 1: Construct the objective function for the seismic velocity modeling method;

[0063] Step 2: Obtain well logging data, migration data, migration velocity data, well-seismic depth error data, and weight parameters of TripleANet;

[0064] Step 3: Construct a wellbore data random sampler to randomly sample migration data, migration velocity data, logging data, and well seismic depth error data;

[0065] Step 4: Calculate the predicted velocity data block based on the weight parameters of TripleANet obtained in Step 2 and the offset data and offset velocity data obtained in Step 3.

[0066] Step 5: Based on the logging data block obtained in Step 3 and the predicted velocity data obtained in Step 4, calculate the loss of the error constraint term of the logging data;

[0067] Step 6: Calculate the loss of the well seismic error constraint term based on the well seismic depth error data block obtained in Step 3 and the predicted velocity data obtained in Step 4.

[0068] Step 7: Based on the logging data error and well vibration error constraint term loss obtained in Steps 5 and 6, calculate the gradient of the TripleANet weight parameters and update the TripleANet weight parameters.

[0069] Step 8: Repeat steps 2-7 to iteratively update the weight parameters of TripleANet. Stop the calculation when the number of iterations or the error meets the exit condition.

[0070] Step 9: Calculate the predicted velocity model based on the offset data, offset velocity data, and weight parameters of TripleANet obtained in Step 2.

[0071] This invention presents an intelligent seismic velocity modeling method based on attention mechanisms and well-seismic data constraints. It applies TripleANet and well-seismic data constraints to seismic velocity modeling, using the TripleANet network to learn and extract representations from the input data, and employing well-seismic data information to design constraints for sonic logging data errors and well-seismic depth errors, thereby improving the accuracy of velocity modeling. This method has not been reported in any published literature, either domestically or internationally.

[0072] This invention proposes a method for calculating the error constraint between well logging data blocks and predicted velocity data. This method can calculate the error in the network's predicted velocity based on the well logging data blocks, and then use this error to modify the neural network parameters, thereby improving the final predicted velocity model.

[0073] This invention proposes a method for calculating the loss of a well-seismic error constraint term. This method can accurately calculate the error between well logging data and seismic data, and use this error to construct a loss function. During velocity modeling, this loss function is used to treat well data as a priori constraints, thereby improving the accuracy of velocity modeling.

[0074] The following are several specific embodiments of the application of the present invention.

[0075] Example 1

[0076] In a specific embodiment 1 of the present invention, the present invention is applied. Figure 1 A flowchart of an intelligent seismic velocity modeling method based on attention mechanism and well-seismic data constraints; Figure 2 A flowchart illustrating the intelligent seismic velocity modeling and calculation process based on attention mechanisms and well-seismic data constraints; Figure 3 The structure diagram of TripleANet is shown below. This intelligent seismic velocity modeling method based on attention mechanism and well-seismic data constraints includes the following steps:

[0077] S1. Construct the objective function of an intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints;

[0078] The specific implementation method for constructing the objective function includes the following steps:

[0079] S1.1 Setting TripleANet in step S1 to the function f(:w) to predict velocity data. The calculation formula is in and It belongs to offset data and offset speed data The offset data blocks and offset velocity blocks are defined, where t < i, w < c, h < z represents the size of the data blocks, and the actual and predicted velocity models are defined as follows: Real speed data Additionally, w are the learnable weight parameters of TripleANet. This pertains to well logging data. The logging data block is represented as and use a label mask The value is 1 if the logging data block is not empty, and 0 otherwise. This belongs to the well-seismic depth error data. The well depth error data block is represented as Where m represents the number of well seismic depth error data.

[0080] S1.2 The formula for calculating the error constraint between well logging data blocks and predicted velocity data is as follows:

[0081]

[0082] Where ||·||1 is the L1 norm, M j For the label mask, ⊙ represents element-wise multiplication, and y... j For predicting speed data, This is a block of well logging data.

[0083] S1.3 The formula for calculating the loss of the well vibration error constraint term is:

[0084]

[0085] Where ||·|| is the L2 norm, This is a block of wellbore seismic depth error data; the predicted wellbore seismic depth error is... The calculation formula is:

[0086]

[0087] Where d represents the depth interval size, and n represents the number of depth intervals in the target layer. This represents the prediction speed corresponding to each depth interval.

[0088] S1.4, The objective function is:

[0089]

[0090] S1.5, The formula for updating the weight parameters is:

[0091]

[0092] Where η is the iteration update step size, w k+1 These are the network parameters for the (k+1)th iteration.

[0093] S2. Obtain well logging data, migration data, migration velocity data, well-seismic depth error data, and weight parameters of TripleANet;

[0094] S3. Construct a wellbore data random sampler to randomly sample offset data, offset velocity data, logging data and well-seismic depth error data to obtain logging data blocks, offset data blocks, offset velocity data blocks and well-seismic depth error data blocks;

[0095] The specific implementation method for constructing a well perimeter data random sampler includes the following steps:

[0096] S3.1, Set the logging data block to the starting position of the logging data as (wp) i wp c wp z If the position sampling interval is [wp], then the position sampling interval can be set to [wp]. i -t,wp i ],[wp c -w,wpc ],[wp z -h,wp z ], where wp i -t>0,wp c -w>0,wp z -h>0,wp i <i,wp c <c,wp z <z.

[0097] S3.2, Set the position random variable p t ~U(wp) i -t,wp i ),p w ~U(wp) c -w,wp c ),p h ~U(wp) z -h,wp z ), where p t ~U(wp) i -t, wp i ) represents a random variable p t Follow the interval [wp] i -t, wp i Uniform distribution on ], p w ~U(wp) c -w,wp c ) represents a random variable p w Follow the interval [wp] c -w,wp c Uniform distribution on ], p h ~U(wp) z -h,wp z ) represents a random variable p h Follow the interval [wp] z -h,wp z A uniform distribution on [the surface].

[0098] S3.3. According to S3.2, the data location can be randomly sampled, and the starting position is represented as (p). t p w p h The endpoint position is represented as (p) t +t, p w +w, p h +h).

[0099] S3.4 Based on the starting and ending positions in S3.3, the logging data, migration data, migration velocity data, and well-vibration depth error data can be sampled and segmented to obtain sonic logging data blocks, migration data blocks, migration velocity data blocks, and well-vibration depth error data blocks.

[0100] S4. Based on the weight parameters of TripleANet obtained in step S2 and the offset data block and offset velocity data block obtained in step S3, calculate the predicted velocity data block.

[0101] S5. Calculate the error constraint term loss of the logging data based on the logging data block obtained in step S3 and the predicted velocity data block obtained in step S4.

[0102] S6. Calculate the loss of the well vibration error constraint term based on the well vibration depth error data block obtained in step S3 and the predicted velocity data block obtained in step S4.

[0103] S7. Calculate the gradient of the TripleANet weight parameters and update the TripleANet weight parameters based on the logging data error and well vibration error constraint term loss obtained in steps S5 and S6.

[0104] S8. Repeat steps S2-S7 to iteratively update the weight parameters of TripleANet. Stop the calculation when the number of iterations or the error meets the exit condition.

[0105] S9. Based on the offset data, offset velocity data and weight parameters of TripleANet obtained in step S2, calculate the predicted velocity model.

[0106] Example 2

[0107] In a specific embodiment 2 of the application of the present invention, the appendix Figure 4 This consists of SEAM simulated migration data (501*501*400 pixels), a SEAM simulated seismic velocity model, and a migration velocity model. Seismic velocity models were predicted using ANet and TripleANet, respectively. Here, 'ad' represents the migration data, the actual seismic velocity model, the ANet predicted velocity model, and the TripleANet predicted velocity model, respectively.

[0108] Appendix Figure 5 This is a comparison chart of the actual velocity, ANet predicted velocity, and TripleANet predicted velocity curves at well locations (200, 385) and (350, 200). Figure 5 The image shows a comparison between the predicted seismic velocity model obtained using the ANet and TripleANet methods at the well location, the actual velocity curve, and the migration velocity curve. The comparison shows that the method of this implementation is more effective.

[0109] Based on the results of the earthquake velocity prediction model, the formula for calculating the model's mean absolute error is as follows:

[0110]

[0111] in For real and predicted velocity models, real velocity data Predicted speed data y j .

[0112] The formula for calculating the mean squared error of the model is as follows:

[0113]

[0114] in For real and predicted velocity models, real velocity data Predicted speed data y j .

[0115] The formula for calculating model structural similarity is as follows:

[0116]

[0117] Where μ Y , σ Y These are earthquake velocity models. The mean and standard deviation.

[0118] Table 1 Comparison of Earthquake Velocity Model Prediction Performance

[0119]

[0120] As can be seen from Table 1, the TripleANet method of this implementation is superior to the ANet method in terms of MAE, MSE and SSIM evaluation metrics.

[0121] Appendix Figure 6 This is a comparison chart of local velocity predictions using SEAM. The chart shows the local velocity predictions using SEAM. Here, 'ad' represents the actual velocity, the migration velocity, the TripleANet predicted velocity without wellbore error constraints, and the TripleANet predicted velocity with wellbore error constraints, respectively. Figure 6 It can be seen that, compared to the offset velocity, the predicted velocity exhibits a higher frequency component, richer layered structure, and better matches the actual velocity. Furthermore, the offset velocity is lower than the actual velocity at depths less than 500 meters. Predicted velocities without wellbore error constraints improve the prediction performance in shallow layers, while predicted velocities with wellbore error constraints further improve the prediction performance in shallow layers. This demonstrates that the method in this embodiment effectively improves the accuracy of shallow layer predicted velocities by introducing wellbore error constraints.

[0122] Appendix Figure 7-9This is a comparison chart of wellbore seismic errors in SEAM migration data. The predicted velocity migration data matches the actual velocity migration data more closely in overall structure compared to the actual velocity migration data. Furthermore, the introduction of wellbore seismic error constraints further reduces the wellbore seismic error between the predicted velocity migration data and the actual velocity migration data.

[0123] Figure 10 Well-seismic error comparison chart. The blue, orange, and green lines represent the well-seismic error variations with depth for the migration velocity, prediction velocity, and well-seismic error-constrained prediction velocity, respectively. (Attached) Figure 10 The graph compares the wellbore error calculated using the offset velocity, the predicted velocity without wellbore error constraints, and the predicted velocity with wellbore error constraints. It can be seen that the wellbore error decreases compared to the offset velocity, and the wellbore error decreases further after introducing wellbore error constraints. This indicates that the TripleANet proposed in this implementation method and the wellbore error constraints effectively improve the velocity prediction accuracy and effectively reduce wellbore error.

[0124] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0125] Except for the technical features described in the specification, all other technologies are known to those skilled in the art.

Claims

1. A smart seismic velocity modeling method based on attention mechanism and well-seismic data constraints, characterized in that, This intelligent seismic velocity modeling method, based on attention mechanisms and well-seismic data constraints, includes: Step 1: Construct the objective function for the seismic velocity modeling method; Step 2: Obtain well logging data, migration data, migration velocity data, well-seismic depth error data, and weight parameters of TripleANet; Step 3: Construct a wellbore data random sampler to randomly sample migration data, migration velocity data, logging data, and well seismic depth error data; Step 4: Calculate the predicted velocity data block, the loss of the error constraint term of the logging data, and the loss of the well-seismic error constraint term; Step 5: Calculate the gradient of the TripleANet weight parameters and update the TripleANet weight parameters; Step 6: Calculate the predicted velocity model.

2. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 1, characterized in that, In step 1, the objective function of the intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints is constructed, specifically including: Step 1.1: Set TripleANet in Step 1 to the function f(:w) to predict velocity data. The calculation formula is in and It belongs to offset data and offset speed data The offset data blocks and offset velocity blocks are defined, where t < i, w < c, h < z represents the size of the data blocks, and the actual and predicted velocity models are defined as follows: Real speed data w are the learnable weight parameters of TripleANet; S1.2 The formula for calculating the error constraint between well logging data blocks and predicted velocity data is as follows: Where ||·||1 is the L1 norm, M j For the label mask, ⊙ represents element-wise multiplication, and y... j For predicting speed data, For well logging data blocks; S1.3 The formula for calculating the loss of the well vibration error constraint term is: Where ||·|| is the L2 norm, This is a block of wellbore seismic depth error data; the predicted wellbore seismic depth error is... The calculation formula is: Where d represents the depth interval size, and n represents the number of depth intervals in the target layer. This represents the prediction velocity corresponding to each depth interval; S1.4, The objective function is: S1.5, The formula for updating the weight parameters is: Where η is the iteration update step size, w k+1 These are the network parameters for the (k+1)th iteration.

3. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 2, characterized in that, In step 1.1, this pertains to well logging data. The logging data block is represented as and use a label mask The value is 1 if the logging data block is not empty, and 0 otherwise; this belongs to the well-seismic depth error data. The well depth error data block is represented as Where m represents the number of well seismic depth error data.

4. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 2, characterized in that, Step 3, the specific steps for constructing the well perimeter data random sampler include: Step 3.1: Set the logging data block to the starting position of the logging data as (wp) i wp c wp z If the position sampling interval is [wp], then the position sampling interval can be set to [wp]. i -t, wp i ], [wp c -w, wp c ], [wp z -h,wp z ], where wp i -t>0, wp c -w > 0, wp c -h > 0, wp i <i,wp c <c,wp z <z; Step 3.2: Set the location random variable p t ~U(wp) i -t, wp i ), p w ~U(wp) c -w, wp c ), p h ~U(wp) z -h,wp z ), where p t ~U(wp) i -t, wp i ) represents a random variable p t Follow the interval [wp] i -t, wp i Uniform distribution on ], p w ~U(wp) c -w,wp c ) represents a random variable p w Follow the interval [wp] c -w,wp c Uniform distribution on ], p h ~U(wp) z -h,wp z ) represents a random variable p h Follow the interval [wp] z -h,wp z Uniform distribution on ]; Step 3.3: Based on Step 3.2, data positions can be randomly sampled, with the starting position represented as (p). t p w p h The endpoint position is represented as (p) t +t, p w +w, p h +h); Step 3.4: Based on the starting and ending positions in 3.3, sample and segment the logging data, migration data, migration velocity data, and well-seismic depth error data to obtain sonic logging data blocks, migration data blocks, migration velocity data blocks, and well-seismic depth error data blocks.

5. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 4, characterized in that, In step 4, based on the weight parameters of TripleANet obtained in step 2 and the offset data block and offset velocity data block obtained in step 3, the predicted velocity data block is calculated according to step 1.

1.

6. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 5, characterized in that, In step 4, based on the logging data block obtained in step 3 and the predicted velocity data block obtained in step 4, the error constraint term loss of the logging data is calculated according to step 1.

2.

7. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 5, characterized in that, In step 4, based on the well-seismic depth error data block obtained in step 3 and the predicted velocity data block obtained in step 4, the loss of the well-seismic error constraint term is calculated according to step 1.

3.

8. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 1, characterized in that, In step 5, based on the logging data error and well vibration error constraint term loss obtained in step 4, the gradient of the TripleANet weight parameters is calculated according to steps 1.4 and 1.5, and the weight parameters of TripleANet are updated.

9. The intelligent seismic velocity modeling method based on attention mechanism and well-seismic data constraints according to claim 1, characterized in that, The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints also includes repeating steps 2 to 5 after step 5 to iteratively update the weight parameters of TripleANet, and stopping the calculation when the number of iterations or the error meets the exit condition.

10. The intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints according to claim 1, characterized in that, In step 6, the predicted velocity model is calculated based on the offset data, offset velocity data and weight parameters of TripleANet obtained in step 2.

11. An intelligent seismic velocity modeling system based on attention mechanism and well-seismic data constraints, characterized in that, The intelligent seismic velocity modeling system based on attention mechanism and well seismic data constraints uses the intelligent seismic velocity modeling method based on attention mechanism and well seismic data constraints as described in any one of claims 1-10 to perform seismic velocity modeling.

Images