Seismic inversion method, system, device and medium based on intelligent hybrid learning

Abstract

This invention belongs to the field of seismic interpretation technology, specifically disclosing a seismic inversion method, system, equipment, and medium based on intelligent hybrid learning. The method includes: S1, acquiring raw well logging data and preprocessing it to obtain first well logging data; S2, expanding the first well logging data to obtain second well logging data; S3, generating first seismic data and constructing sample pairs between the first seismic data and the second well logging data to form a sample set; S4, constructing an artificial intelligence inversion network model and performing supervised and unsupervised learning training to obtain a trained artificial intelligence inversion network model; S5, using the trained artificial intelligence inversion network model to output predicted reservoir parameters. This invention can predict formation reservoir parameters, fully utilizes unlabeled data, improves the model's generalization ability and the accuracy of the inversion results, and finely depicts subsurface geological structures. This invention is applicable to the prediction of formation reservoir parameters.

Description

Technical Field

[0001] This invention belongs to the field of earthquake interpretation technology, specifically a seismic inversion method, system, device and medium based on intelligent hybrid learning. Background Technology

[0002] Seismic data inversion plays a crucial role in oil and gas exploration and geological research. Seismic inversion technology processes data on the propagation of seismic waves in the subsurface medium to obtain reservoir parameters of the formation, such as P-wave velocity, S-wave velocity, and density, thereby inferring subsurface structures and reservoir characteristics.

[0003] Seismic inversion technology can be traced back to the 1970s, when inversion methods were mainly based on linear theory, such as least squares and linear inversion. Subsequently, nonlinear inversion methods gradually emerged, such as simulated annealing and genetic algorithms, which can better handle complex geological models. In the 21st century, seismic inversion technology has gradually integrated multiple data sources and methods, including pre-stack and post-stack inversion, AVO analysis, etc., providing more possibilities for the accurate characterization of subsurface geological structures.

[0004] While traditional seismic inversion methods have achieved some success in practical applications, they still have some limitations. First, these methods typically rely on large amounts of high-quality labeled data, which is costly and time-consuming to acquire. Second, traditional inversion methods cannot meet the requirements for high-precision inversion when dealing with complex geological conditions. Furthermore, existing methods, to some extent, ignore the nonlinear characteristics of seismic wave propagation, leading to errors in the inversion results. Summary of the Invention

[0005] The purpose of this invention is to provide a seismic inversion method, system, device, and medium based on intelligent hybrid learning, so as to reduce the workload of manually labeled data, make full use of unlabeled data, improve the accuracy of seismic inversion, predict reservoir parameters of strata, and finely characterize underground geological structures.

[0006] To achieve the above objectives, the present invention employs the following technical methods:

[0007] A seismic inversion method based on intelligent hybrid learning includes the following steps performed sequentially:

[0008] S1. Obtain raw logging data, preprocess the raw logging data to obtain the first logging data;

[0009] S2. Expand the first logging data to obtain the second logging data;

[0010] S3. Based on the second well logging data, generate the first seismic data, construct sample pairs of the first seismic data and the second well logging data, and form a sample set;

[0011] S4. Construct an artificial intelligence inversion network model. Based on the sample set, conduct supervised learning training on the artificial intelligence inversion network model. Use real seismic data from the entire work area to conduct unsupervised learning training on the artificial intelligence inversion network model until the total loss function of the artificial intelligence inversion network model converges, and obtain the trained artificial intelligence inversion network model.

[0012] S5. Input the real seismic data of the entire work area into the trained artificial intelligence inversion network model, and output the predicted reservoir parameters.

[0013] As a limitation: the original logging data in step S1 includes P-wave velocity, S-wave velocity and density. The preprocessing of the original logging data is as follows: first, the original logging data is denoised, and then the denoised logging data is smoothed.

[0014] The formula for noise reduction is:

[0015]

[0016] Where X(i) is the original well logging data at location i, i.e., X(i) = {V p (i),V s (i),ρ(i)},V p (i) represents the P-wave velocity in the original well logging data at location i, V s (i) represents the shear wave velocity in the original logging data at location i, ρ(i) represents the density in the original logging data at location i, N is the length of the filter window, and X * (i) represents the denoised logging data at location i, i.e., X * (i)={V p * (i),V s * (i),ρ * (i)},V p * (i) represents the denoised longitudinal wave velocity at position i, V s * (i) represents the denoised transverse wave velocity at position i, ρ * (i) represents the density at position i after noise reduction;

[0017] The formula for smoothing is:

[0018]

[0019] in, The first logging data at location i, i.e. Let be the P-wave velocity in the first logging data at location i. Let be the shear wave velocity in the first logging data at location i. Let be the density in the first logging data at location i, and M be the length of the moving window.

[0020] As a limitation: in step S2, a variational autoencoder is used to expand the first logging data. The variational autoencoder includes an encoder and a decoder. The encoder and decoder have corresponding structures and are both neural network structures composed of multiple fully connected layers and have an inverse mapping relationship. The second logging data is generated by iteratively training the encoder and decoder.

[0021] As a further limitation, the specific iterative training process of the encoder and decoder in step S2 is as follows:

[0022] First, the encoder and decoder parameters are randomly initialized. A pseudo-well curve is generated through forward propagation. Specifically, the first well logging data is input into the encoder, which maps the first well logging data to the distribution of the latent variable z. The calculation formula is as follows:

[0023]

[0024] in, The first logging data at all locations, i.e. For the P-wave velocity in the first logging data at all locations, The shear wave velocity in the first logging data at all locations. Let z be the density in the first logging data at all locations, μ and σ be the distribution parameters of the latent variable z, and z = μ + σε be the latent variable z sampled from the distribution of the latent variable z using the reparameterization technique. ε is a random variable sampled from the standard normal distribution.

[0025] The decoder maps the latent variable z back to the space of the first logging data, generating a pseudo-well curve, calculated using the following formula:

[0026]

[0027] Among them, z j For the j-th latent variable obtained from sampling, For the j-th latent variable z j The corresponding j-th pseudo-well curve, i.e., the second logging data. The P-wave velocity curve is shown in the second well logging data. The shear wave velocity curve in the second well logging data. The density curve in the second well logging data;

[0028] Secondly, the loss function of the variational autoencoder is calculated. The variational autoencoder loss function is derived from the first logging data. The reconstruction error consists of components, and the reconstruction error measures the pseudo-well curve. Compared with the first logging data The difference between them is calculated using the following formula:

[0029]

[0030] Where MSE is the total reconstruction error. The reconstruction error between the pseudo-well curve of P-wave velocity and the P-wave velocity curve in the first logging data is denoted as . The reconstruction error between the pseudo-well curve of shear wave velocity and the shear wave velocity curve in the first logging data is denoted as . The reconstruction error between the density pseudo-well curve and the density curve in the first logging data, |||| 2 Euclidean distance symbol;

[0031] Then, the gradient of the variational autoencoder loss function with respect to the encoder and decoder parameters is calculated using the backpropagation algorithm;

[0032] Finally, an optimization algorithm is selected to update the encoder parameters based on the gradient of the encoder parameters, and the decoder parameters are updated based on the gradient of the decoder parameters. This process is repeated for multiple iterations until the maximum number of iterations is reached or the variational autoencoder loss function converges, generating a series of pseudo-well curves. That is, the second logging data Where K is the number of pseudo-well curves generated.

[0033] As a limitation: Step S3 specifically involves: based on the second well logging data, using pre-calculated seismic wavelets, and employing the traditional Zoeppritz equation forward modeling method, generating the first seismic data. The first seismic data and the corresponding second well logging data form a sample pair. Using the forward modeling label construction method, a sample set consisting of multiple forward modeling samples is obtained.

[0034] As a limitation: the AI ​​inversion network model in step S4 includes convolutional layers, pooling layers, and fully connected layers;

[0035] Supervised learning and training, specifically:

[0036] Using the first earthquake data in the sample set As input to the artificial intelligence inversion network model, Let J be the j-th first seismic data in the sample set, and K be the number of first seismic data in the sample set, and let K be the corresponding second well logging data in the sample set. As the output of the artificial intelligence inversion network model The P-wave velocity in the second logging data, The shear wave velocity in the second logging data. The density in the second logging data is K, and the number of second logging data is K;

[0037] Train the artificial intelligence inversion network model and calculate the loss function L of the supervised learning training part. inv L inv The calculation formula is:

[0038]

[0039] in, To verify the j-th longitudinal wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th shear wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th density curve output at step t when the artificial intelligence inversion network model is used, || || is the regularization function;

[0040] Unsupervised learning training, specifically:

[0041] Based on real seismic data from the entire work area As input to the artificial intelligence inversion network model, N is the number of real seismic data points in the entire work area. Using the nth real seismic data in the entire work area, reservoir parameters are predicted through an artificial intelligence inversion network model. To predict the nth longitudinal wave velocity, To predict the nth transverse wave velocity, To obtain the predicted nth density, and based on the predicted reservoir parameters, the traditional Zoeppritz equation forward modeling method is used to generate the second seismic data. For the nth second earthquake data generated by forward modeling, let the second earthquake data Ψ pre The error between the actual seismic data Ψ and the data from the entire work area is used as the loss function L for the unsupervised learning training part. seis By minimizing the loss function L of the unsupervised learning training part seis Training is performed, and the loss function L for the unsupervised learning training part is used. seis The calculation formula is:

[0042]

[0043] By iteratively training the AI ​​inversion network model until the total loss function L of the AI ​​inversion network model converges, a well-trained AI inversion network model is obtained.

[0044] The formula for calculating the total loss function L of the artificial intelligence inversion network model is as follows:

[0045] L=αL inv +βL seis ,

[0046] Where α is the weight coefficient of the loss function for the supervised learning training part, and β is the weight coefficient of the loss function for the unsupervised learning training part.

[0047] This invention also provides a seismic inversion system based on intelligent hybrid learning, comprising:

[0048] The data preprocessing module acquires the raw logging data, preprocesses the raw logging data, and obtains the first logging data.

[0049] The data augmentation module augments the first logging data to obtain the second logging data.

[0050] The sample set construction module generates first seismic data based on the second well logging data, constructs sample pairs of the first seismic data and the second well logging data, and forms a sample set;

[0051] The AI ​​inversion network model construction module constructs an AI inversion network model, performs supervised learning training on the AI ​​inversion network model based on a sample set, and performs unsupervised learning training on the AI ​​inversion network model using real seismic data from the entire work area until the total loss function of the AI ​​inversion network model converges, thus obtaining a trained AI inversion network model.

[0052] The reservoir parameter prediction module inputs real seismic data from the entire work area into a trained artificial intelligence inversion network model and outputs predicted reservoir parameters.

[0053] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program in the memory to execute the above-described seismic inversion method based on intelligent hybrid learning.

[0054] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the above-described seismic inversion method based on intelligent hybrid learning.

[0055] The beneficial effects achieved by this invention, due to the adoption of the above-described solution, compared with the prior art, are as follows:

[0056] This invention provides a seismic inversion method, system, equipment, and medium based on intelligent hybrid learning. By expanding well logging data and using the expanded data as existing labeled data for supervised learning training of an artificial intelligence (AI) inversion network model, and by using a large amount of unlabeled real seismic data for unsupervised learning training of the AI ​​inversion network model, the AI ​​inversion network model becomes familiar with the data characteristics of real seismic data, predicts reservoir parameters of strata, reduces the workload of manual data labeling, fully utilizes unlabeled data, improves the model's generalization ability and the accuracy of inversion results, and finely depicts underground geological structures. By constructing an AI inversion network model, it can better capture the nonlinear characteristics in seismic data, improving the accuracy of inversion results. Using the traditional Zoeppritz equation forward modeling method as physical constraints ensures that the inversion results conform to geophysical laws. Combined with the AI ​​inversion network model, it can leverage the advantages of data-driven approaches to capture more complex geological features, achieving complementary advantages and improving the accuracy and reliability of seismic inversion.

[0057] This invention is applicable to the prediction of formation and reservoir parameters. Attached Figure Description

[0058] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0059] Figure 1 This is a flowchart of the seismic inversion method based on intelligent hybrid learning in Embodiment 1 of the present invention;

[0060] Figure 2 This is a schematic diagram of the artificial intelligence inversion network model in Embodiment 1 of the present invention;

[0061] Figure 3 The figures show the loss curve and accuracy change curve during the training of the artificial intelligence inversion network model in Embodiment 1 of the present invention.

[0062] Figure 4 This is a comparison diagram of the reservoir predicted by the artificial intelligence inversion network model after 400 training cycles and 1000 training cycles, the original seismic data, and the reservoir vertically obtained by SSI inversion in Embodiment 1 of the present invention.

[0063] Figure 5 This is a comparison image of the reservoir predicted by the trained artificial intelligence inversion network model and the reservoir plane obtained by SSI inversion in Embodiment 1 of the present invention.

[0064] Figure 6 This is a structural block diagram of the earthquake inversion system based on intelligent hybrid learning, as described in Embodiment 2 of the present invention.

[0065] Figure 7 This is a schematic diagram of the structure of the electronic device in Embodiment 3 of the present invention. Detailed Implementation

[0066] The present invention will be further described below with reference to the embodiments. However, those skilled in the art should understand that the present invention is not limited to the following embodiments. Any improvements and equivalent changes made based on the specific embodiments of the present invention are within the scope of protection of the claims of the present invention.

[0067] Example 1: Seismic Inversion Method Based on Intelligent Hybrid Learning

[0068] A seismic inversion method based on intelligent hybrid learning, such as Figure 1 As shown, the steps are performed sequentially:

[0069] S1. Obtain raw logging data, which includes P-wave velocity, S-wave velocity, and density. Perform noise reduction on the raw logging data, and then smooth the noise-reduced logging data to obtain the first logging data.

[0070] The formula for noise reduction is:

[0071]

[0072] Where X(i) is the original well logging data at location i, i.e., X(i) = {V p (i),V s (i),ρ(i)},V p (i) represents the P-wave velocity in the original well logging data at location i, V s (i) represents the shear wave velocity in the original logging data at location i, ρ(i) represents the density in the original logging data at location i, and X * (i) represents the denoised logging data at location i, i.e., X * (i)={V p * (i),V s * (i),ρ * (i)},V p * (i) represents the denoised longitudinal wave velocity at position i, V s * (i) represents the denoised transverse wave velocity at position i, ρ *(i) represents the denoised density at position i, and N is the length of the filtering window. The choice of the filtering window length N is crucial for the denoising effect. If N is too small, it may not effectively remove noise; if N is too large, it may over-smooth the data, causing the loss of some useful details. The value of N is determined by the characteristics of the P-wave velocity curve, S-wave velocity curve, and density curve in the original well logging data, as well as the characteristics of the noise. Taking the P-wave velocity curve as an example, when the noise frequency is high and the P-wave velocity curve changes relatively gently, a relatively small N value is chosen; when the noise is more complex and the P-wave velocity fluctuates greatly, a relatively large N value is chosen. The S-wave velocity curve and density curve can be denoised using the same method.

[0073] The formula for smoothing is:

[0074]

[0075] in, The first logging data at location i, i.e. Let be the P-wave velocity in the first logging data at location i. Let be the shear wave velocity in the first logging data at location i. Let M be the density in the first well logging data at location i, and M be the length of the moving window. The choice of the moving window length M is closely related to the data resolution and the required smoothing level. If M is small, the smoothing effect will be relatively weak, preserving more data details but failing to effectively remove some small-scale fluctuations. If M is large, a stronger smoothing effect can be produced, better eliminating high-frequency noise and small fluctuations in the data, but over-smoothing the data can obscure some important geological features. In practical applications, the moving window length M needs to be reasonably selected based on the resolution of the P-wave velocity curve, S-wave velocity curve, and density curve, the scale of geological features, and the specific requirements for the smoothing effect. Taking the P-wave velocity curve as an example, for P-wave velocity curves with high resolution and complex geological features, a smaller M value can be tried for initial smoothing, and then the M value can be gradually adjusted based on the analysis results. For P-wave velocity curves with low resolution and relatively simple geological features, a larger M value can be appropriately selected to quickly obtain a smoother result. The same method can be used to select the M value for smoothing of S-wave velocity curves and density curves.

[0076] S2. The first logging data is expanded using a variational autoencoder. The variational autoencoder includes an encoder and a decoder. The encoder and decoder have corresponding structures and are both neural network structures composed of multiple fully connected layers and have an inverse mapping relationship. The second logging data is generated by iteratively training the encoder and decoder.

[0077] The specific iterative training process for the encoder and decoder is as follows:

[0078] First, the encoder and decoder parameters are randomly initialized. The encoder parameters are the weights and biases in the encoder neural network, and the decoder parameters are the weights and biases in the decoder neural network. A pseudo-well curve is generated through forward propagation. Specifically, the first well logging data is input into the encoder, which maps the first well logging data to the distribution of the latent variable z. The calculation formula is as follows:

[0079]

[0080] in, The first logging data at all locations, i.e. For the P-wave velocity in the first logging data at all locations, The shear wave velocity in the first logging data at all locations. Let z be the density in the first logging data at all locations, μ and σ be the distribution parameters of the latent variable z, and z = μ + σε be the latent variable z sampled from the distribution of the latent variable z using the reparameterization technique. ε is a random variable sampled from the standard normal distribution.

[0081] The decoder maps the latent variable z back to the space of the first logging data, generating a pseudo-well curve, calculated using the following formula:

[0082]

[0083] Among them, z j For the j-th latent variable obtained from sampling, For the j-th latent variable z j The corresponding j-th pseudo-well curve, i.e., the second logging data. The P-wave velocity curve is shown in the second well logging data. The shear wave velocity curve in the second well logging data. The density curve in the second well logging data;

[0084] Secondly, the loss function of the variational autoencoder is calculated. The variational autoencoder loss function is derived from the first logging data. The reconstruction error consists of components, and the reconstruction error measures the pseudo-well curve. Compared with the first logging data The difference between them is calculated using the following formula:

[0085]

[0086] Where MSE is the total reconstruction error. The reconstruction error between the pseudo-well curve of P-wave velocity and the P-wave velocity curve in the first logging data is denoted as . The reconstruction error between the pseudo-well curve of shear wave velocity and the shear wave velocity curve in the first logging data is denoted as . The reconstruction error between the density pseudo-well curve and the density curve in the first logging data, |||| 2 Euclidean distance symbol;

[0087] Then, the gradient of the variational autoencoder loss function with respect to the encoder and decoder parameters is calculated using the backpropagation algorithm;

[0088] Finally, an optimization algorithm is selected to update the encoder parameters based on the gradient of the encoder parameters, and the decoder parameters are updated based on the gradient of the decoder parameters. In this embodiment, the Adam algorithm is used to update the encoder and decoder parameters. The parameter values ​​are adjusted according to the gradient direction and magnitude and the algorithm rules to reduce the variational autoencoder loss function value. The above process is repeated for multiple iterations until the maximum number of iterations is reached or the variational autoencoder loss function converges, generating a series of pseudo-well curves. That is, the second logging data Where K represents the number of pseudo-well curves generated. Expanding the first well logging data overcomes the problem of insufficient well logging sample labels in practical applications, providing sufficient sample data for subsequent seismic inversion based on an artificial intelligence inversion network model.

[0089] S3, based on second logging data The first set of seismic data was generated using pre-calculated seismic wavelets and the traditional Zoeppritz equation forward modeling method. For the j-th first seismic data in the sample set, the first seismic data and the corresponding second well logging data form a sample pair, i.e. Using a forward-modeling label construction method, a sample set consisting of multiple forward-modeled samples is obtained.

[0090] S4. Construct an artificial intelligence inversion network model, such as Figure 2 As shown, the artificial intelligence inversion network model includes convolutional layers (conv), pooling layers (pooling), and fully connected layers (FN). Convolutional layers extract spatial features from seismic data, pooling layers extract multi-scale features and reduce parameter size, and fully connected layers combine features. Based on the sample set... Supervised learning training is performed on the AI ​​inversion network model, and unsupervised learning training is performed on the AI ​​inversion network model using real seismic data from the entire work area until the total loss function of the AI ​​inversion network model converges, thus obtaining the trained AI inversion network model.

[0091] Supervised learning and training, specifically:

[0092] sample set For labeled well logging data, in sample set The first earthquake data in China As input to the artificial intelligence inversion network model, the sample set The corresponding second logging data As the output of the artificial intelligence inversion network model;

[0093] Train the artificial intelligence inversion network model and calculate the loss function L of the supervised learning training part. inv L inv The calculation formula is:

[0094]

[0095] in, To verify the j-th longitudinal wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th shear wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th density curve output at step t when the artificial intelligence inversion network model is used, || || is the regularization function;

[0096] Unsupervised learning training, specifically:

[0097] Based on real seismic data from the entire work area As input to the artificial intelligence inversion network model, N is the number of real seismic data points in the entire work area. Using the nth real seismic data in the entire work area, reservoir parameters are predicted through an artificial intelligence inversion network model. To predict the nth longitudinal wave velocity, To predict the nth transverse wave velocity, To obtain the predicted nth density, and based on the predicted reservoir parameters, the traditional Zoeppritz equation forward modeling method is used to generate the second seismic data. For the nth second earthquake data generated by forward modeling, let the second earthquake data Ψ pre The error between the actual seismic data Ψ and the data from the entire work area is used as the loss function L for the unsupervised learning training part. seis By minimizing the loss function L of the unsupervised learning training part seis Training is performed, and the loss function L for the unsupervised learning training part is used. seis The calculation formula is:

[0098]

[0099] By iteratively training the AI ​​inversion network model until the total loss function L of the AI ​​inversion network model converges, a well-trained AI inversion network model is obtained; the loss curve and accuracy change curve during the training process are shown in the figure. Figure 3 As shown, the loss is decreasing and the accuracy is increasing.

[0100] The formula for calculating the total loss function L of the artificial intelligence inversion network model is as follows:

[0101] L=αL inv +βL seis ,

[0102] Where α is the weight coefficient of the loss function for the supervised learning training part, and β is the weight coefficient of the loss function for the unsupervised learning training part.

[0103] S5. Transfer the actual seismic data of the entire work area. Input is fed into a trained AI inversion network model, which outputs predicted reservoir parameters. V′ p,n For the nth P-wave velocity obtained in the final prediction, V′ s,n For the nth predicted transverse wave velocity, ρ′ n This represents the nth density obtained from the final prediction.

[0104] This embodiment utilizes reservoir and original seismic data predicted by an AI inversion network model trained 400 times and 1000 times, as well as vertical reservoir comparisons obtained through SSI inversion. Figure 4 As shown, the reservoir predicted by the artificial intelligence inversion network model in this embodiment has higher vertical resolution, better matches the parameters of the actual well logging, and is more regular. The comparison between the reservoir predicted by the trained artificial intelligence inversion network model and the reservoir obtained by SSI inversion in the plane is shown in the figure. Figure 5 As shown, in the planar analysis process, this embodiment utilizes a trained artificial intelligence inversion network model to predict reservoir well passage with high consistency and greater regularity; the method of this embodiment improves the accuracy of the inversion results and can finely characterize the underground geological structure.

[0105] Example 2: Seismic Inversion System Based on Intelligent Hybrid Learning

[0106] A seismic inversion system based on intelligent hybrid learning, such as Figure 6 As shown, it includes:

[0107] The data preprocessing module acquires raw logging data, which includes P-wave velocity, S-wave velocity, and density. The raw logging data is then denoised, and the denoised logging data is smoothed to obtain the first logging data.

[0108] The formula for noise reduction is:

[0109]

[0110] Where X(i) is the original well logging data at location i, i.e., X(i) = {V p (i),V s (i),ρ(i)},V p (i) represents the P-wave velocity in the original well logging data at location i, V s (i) represents the shear wave velocity in the original logging data at location i, ρ(i) represents the density in the original logging data at location i, and X * (i) represents the denoised logging data at location i, i.e., X * (i)={V p * (i),V s * (i),ρ * (i)},V p * (i) represents the denoised longitudinal wave velocity at position i, V s * (i) represents the denoised transverse wave velocity at position i, ρ * (i) represents the density at position i after denoising, and N is the length of the filtering window.

[0111] The formula for smoothing is:

[0112]

[0113] in, The first logging data at location i, i.e. Let be the P-wave velocity in the first logging data at location i. Let be the shear wave velocity in the first logging data at location i. Let be the density in the first logging data at location i, and M be the length of the moving window.

[0114] The data augmentation module uses a variational autoencoder to augment the first logging data. The variational autoencoder includes an encoder and a decoder. The encoder and decoder have corresponding structures, both of which are neural network structures composed of multiple fully connected layers and have an inverse mapping relationship. The second logging data is generated by iteratively training the encoder and decoder.

[0115] The specific iterative training process for the encoder and decoder is as follows:

[0116] First, the encoder and decoder parameters are randomly initialized. The encoder parameters are the weights and biases in the encoder neural network, and the decoder parameters are the weights and biases in the decoder neural network. A pseudo-well curve is generated through forward propagation. Specifically, the first well logging data is input into the encoder, which maps the first well logging data to the distribution of the latent variable z. The calculation formula is as follows:

[0117]

[0118] in, The first logging data at all locations, i.e. For the P-wave velocity in the first logging data at all locations, The shear wave velocity in the first logging data at all locations. Let z be the density in the first logging data at all locations, μ and σ be the distribution parameters of the latent variable z, and z = μ + σε be the latent variable z sampled from the distribution of the latent variable z using the reparameterization technique. ε is a random variable sampled from the standard normal distribution.

[0119] The decoder maps the latent variable z back to the space of the first logging data, generating a pseudo-well curve, calculated using the following formula:

[0120]

[0121] Among them, z j For the j-th latent variable obtained from sampling, For the j-th latent variable z j The corresponding j-th pseudo-well curve, i.e., the second logging data. The P-wave velocity curve is shown in the second well logging data. The shear wave velocity curve in the second well logging data. The density curve in the second well logging data;

[0122] Secondly, the loss function of the variational autoencoder is calculated. The variational autoencoder loss function is derived from the first logging data. The reconstruction error consists of components, and the reconstruction error measures the pseudo-well curve. Compared with the first logging data The difference between them is calculated using the following formula:

[0123]

[0124] Where MSE is the total reconstruction error. The reconstruction error between the pseudo-well curve of P-wave velocity and the P-wave velocity curve in the first logging data is denoted as . The reconstruction error between the pseudo-well curve of shear wave velocity and the shear wave velocity curve in the first logging data is denoted as . The reconstruction error between the density pseudo-well curve and the density curve in the first logging data, || || 2 Euclidean distance symbol;

[0125] Then, the gradient of the variational autoencoder loss function with respect to the encoder and decoder parameters is calculated using the backpropagation algorithm;

[0126] Finally, an optimization algorithm is selected to update the encoder parameters based on the gradient of the encoder parameters, and the decoder parameters are updated based on the gradient of the decoder parameters. In this embodiment, the Adam algorithm is used to update the encoder and decoder parameters. The parameter values ​​are adjusted according to the gradient direction and magnitude and the algorithm rules to reduce the variational autoencoder loss function value. The above process is repeated for multiple iterations until the maximum number of iterations is reached or the variational autoencoder loss function converges, generating a series of pseudo-well curves. That is, the second logging data Where K is the number of pseudo-well curves generated.

[0127] The sample set construction module is based on the second logging data. The first set of seismic data was generated using pre-calculated seismic wavelets and the traditional Zoeppritz equation forward modeling method. For the j-th first seismic data in the sample set, the first seismic data and the corresponding second well logging data form a sample pair, i.e. Using a forward-modeling label construction method, a sample set consisting of multiple forward-modeled samples is obtained.

[0128] The AI ​​inversion network model construction module builds an AI inversion network model, which includes convolutional layers (conv), pooling layers (pooling), and fully connected layers (FN). Convolutional layers extract spatial features from seismic data, pooling layers extract multi-scale features and reduce parameter size, and fully connected layers combine features. This is based on a sample set. Supervised learning training is performed on the AI ​​inversion network model, and unsupervised learning training is performed on the AI ​​inversion network model using real seismic data from the entire work area until the total loss function of the AI ​​inversion network model converges, thus obtaining the trained AI inversion network model.

[0129] Supervised learning and training, specifically:

[0130] sample set For labeled well logging data, in sample set The first earthquake data in China As input to the artificial intelligence inversion network model, the sample set The corresponding second logging data As the output of the artificial intelligence inversion network model;

[0131] Train the artificial intelligence inversion network model and calculate the loss function L of the supervised learning training part. inv L inv The calculation formula is:

[0132]

[0133] in, To verify the j-th longitudinal wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th shear wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th density curve output at step t when the artificial intelligence inversion network model is used, || || is the regularization function;

[0134] Unsupervised learning training, specifically:

[0135] Based on real seismic data from the entire work area As input to the artificial intelligence inversion network model, N is the number of real seismic data points in the entire work area. Using the nth real seismic data in the entire work area, reservoir parameters are predicted through an artificial intelligence inversion network model. To predict the nth longitudinal wave velocity, To predict the nth transverse wave velocity, To obtain the predicted nth density, and based on the predicted reservoir parameters, the traditional Zoeppritz equation forward modeling method is used to generate the second seismic data. For the nth second earthquake data generated by forward modeling, let the second earthquake data Ψ pre The error between the actual seismic data Ψ and the data from the entire work area is used as the loss function L for the unsupervised learning training part. seis By minimizing the loss function L of the unsupervised learning training part seis Training is performed, and the loss function L for the unsupervised learning training part is used. seis The calculation formula is:

[0136]

[0137] By iteratively training the AI ​​inversion network model until the total loss function L of the AI ​​inversion network model converges, a well-trained AI inversion network model is obtained.

[0138] The formula for calculating the total loss function L of the artificial intelligence inversion network model is as follows:

[0139] L=αL inv +βL seis ,

[0140] Where α is the weight coefficient of the loss function for the supervised learning training part, and β is the weight coefficient of the loss function for the unsupervised learning training part.

[0141] The reservoir parameter prediction module uses real seismic data from the entire work area. Input is fed into a trained AI inversion network model, which outputs predicted reservoir parameters. V′ p,n For the nth P-wave velocity obtained in the final prediction, V′ s,n For the nth predicted transverse wave velocity, ρ′ n This represents the nth density obtained from the final prediction.

[0142] Example 3: An electronic device

[0143] The electronic device of this embodiment includes a memory and a processor. The memory stores a computer program, and the processor calls the computer program in the memory to execute the seismic inversion method based on intelligent hybrid learning in Embodiment 1. Figure 7 This is a schematic diagram of the structure of the electronic device provided in this embodiment. The electronic device can be a terminal device or a server. The terminal device can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, personal digital assistants (PDAs), portable Android devices (PADs), portable media players (PMPs), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of this embodiment.

[0144] like Figure 7 As shown, electronic devices may include processing units, such as central processing units (CPUs) and graphics processors (GPUs), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or loaded from storage devices into random access memory (RAM). RAM also stores various programs and data required for the operation of the electronic device. The processing unit, ROM, and RAM are interconnected via a bus. Input devices, output devices, communication devices, and storage devices are also connected to the bus via I / O interfaces.

[0145] Typically, the following devices can be connected to an I / O interface: input devices such as touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices such as liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices such as magnetic tapes, hard drives, etc.; and communication devices. Communication devices allow electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0146] Example 4: A computer-readable medium

[0147] The computer-readable storage medium of this embodiment stores a computer program, which, when executed by a processor, is used to implement the seismic inversion method based on intelligent hybrid learning of Embodiment 1. The computer-readable storage medium of this embodiment may be included in an electronic device; alternatively, it may exist independently and not be assembled into an electronic device.

[0148] The computer-readable storage medium of this embodiment may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit them. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A method of seismic inversion based on smart hybrid learning, characterized in that, This includes the following steps performed sequentially: S1. Obtain raw logging data, preprocess the raw logging data to obtain the first logging data; S2. Expand the first logging data to obtain the second logging data; S3. Based on the second well logging data, generate the first seismic data, construct sample pairs of the first seismic data and the second well logging data, and form a sample set; S4. Construct an artificial intelligence inversion network model. Based on the sample set, conduct supervised learning training on the artificial intelligence inversion network model. Use real seismic data from the entire work area to conduct unsupervised learning training on the artificial intelligence inversion network model until the total loss function of the artificial intelligence inversion network model converges, and obtain the trained artificial intelligence inversion network model. S5. Input the real seismic data of the entire work area into the trained artificial intelligence inversion network model, and output the predicted reservoir parameters.

2. The smart hybrid learning based seismic inversion method of claim 1, wherein, The raw logging data in step S1 includes P-wave velocity, S-wave velocity and density. The preprocessing of the raw logging data is as follows: first, the raw logging data is denoised, and then the denoised logging data is smoothed. The formula for noise reduction is: Where X(i) is the original well logging data at location i, i.e., X(i) = {V p (i),V s (i),ρ(i)},V p (i) represents the P-wave velocity in the original well logging data at location i, V s (i) represents the shear wave velocity in the original logging data at location i, ρ(i) represents the density in the original logging data at location i, N is the length of the filter window, and X * (i) represents the denoised logging data at location i, i.e., X * (i)={V p * (i),V s * (i),ρ * (i)},V p * (i) represents the denoised longitudinal wave velocity at position i, V s * (i) represents the denoised transverse wave velocity at position i, ρ * (i) represents the density at position i after noise reduction; The formula for smoothing is: wherein, is the first log data at position i, i.e. is the P-wave velocity in the first log data at position i, is the S-wave velocity in the first log data at position i, is the density in the first log data at position i, and M is the length of the moving window.

3. The smart hybrid learning based seismic inversion method of claim 1 or 2, wherein, In step S2, a variational autoencoder is used to augment the first logging data. The variational autoencoder includes an encoder and a decoder. The encoder and decoder have corresponding structures, both of which are neural network structures composed of multiple fully connected layers and have an inverse mapping relationship. The second logging data is generated by iteratively training the encoder and decoder.

4. The smart hybrid learning based seismic inversion method of claim 3, wherein, The specific iterative training process of the encoder and decoder in step S2 is as follows: First, the encoder and decoder parameters are randomly initialized. A pseudo-well curve is generated through forward propagation. Specifically, the first well logging data is input into the encoder, which maps the first well logging data to the distribution of the latent variable z. The calculation formula is as follows: in, The first logging data at all locations, i.e. For the P-wave velocity in the first logging data at all locations, The shear wave velocity in the first logging data at all locations. Let z be the density in the first logging data at all locations, μ and σ be the distribution parameters of the latent variable z, and z = μ + σ·ε be the latent variable z sampled from the distribution of the latent variable z using the reparameterization technique. ε is a random variable sampled from the standard normal distribution. The decoder maps the latent variable z back to the space of the first logging data, generating a pseudo-well curve, calculated using the following formula: Among them, z j For the j-th latent variable obtained from sampling, For the j-th latent variable z j The corresponding j-th pseudo-well curve, i.e., the second logging data. The P-wave velocity curve is shown in the second well logging data. The shear wave velocity curve in the second logging data. The density curve in the second well logging data; Secondly, the loss function of the variational autoencoder is calculated, the variational autoencoder loss function is composed of a reconstruction error of the first well logging data , the reconstruction error measures the difference between the pseudo well curve and the first well logging data , and the calculation formula is: Where MSE is the total reconstruction error. The reconstruction error between the pseudo-well curve of P-wave velocity and the P-wave velocity curve in the first logging data is denoted as . The reconstruction error between the pseudo-well curve of shear wave velocity and the shear wave velocity curve in the first logging data is denoted as . The reconstruction error between the density pseudo-well curve and the density curve in the first logging data, || || 2 Euclidean distance symbol; Then, the gradient of the variational autoencoder loss function with respect to the encoder and decoder parameters is calculated using the backpropagation algorithm; Finally, an optimization algorithm is selected to update the encoder parameters based on the gradient of the encoder parameters, and the decoder parameters are updated based on the gradient of the decoder parameters. This process is repeated for multiple iterations until the maximum number of iterations is reached or the variational autoencoder loss function converges, generating a series of pseudo-well curves. That is, the second logging data Where K is the number of pseudo-well curves generated.

5. The smart hybrid learning based seismic inversion method of claim 1 or 2, wherein, Step S3 specifically involves: based on the second well logging data, using pre-calculated seismic wavelets, and employing the traditional Zoeppritz equation forward modeling method, generating the first seismic data. The first seismic data and the corresponding second well logging data form a sample pair. Using the forward modeling label construction method, a sample set composed of multiple forward modeling samples is obtained.

6. The smart hybrid learning based seismic inversion method of claim 1 or 2, wherein, The AI ​​inversion network model in step S4 includes convolutional layers, pooling layers, and fully connected layers; Supervised learning and training, specifically: The first earthquake data in the sample set As input to the artificial intelligence inversion network model, Let J be the j-th first seismic data in the sample set, and K be the number of first seismic data in the sample set, and let K be the corresponding second well logging data in the sample set. As the output of the artificial intelligence inversion network model The P-wave velocity in the second logging data, The shear wave velocity in the second logging data. The density in the second logging data is K, and the number of second logging data is K; The artificial intelligence inversion network model is trained, and a loss function L of a supervised learning training part is calculated inv , L inv The calculation formula is: in, To verify the j-th longitudinal wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th shear wave velocity curve output at step t when using the artificial intelligence inversion network model, To verify the j-th density curve output at step t when the artificial intelligence inversion network model is used, || || is the regularization function; Unsupervised learning training, specifically: Based on real seismic data from the entire work area As input to the artificial intelligence inversion network model, N is the number of real seismic data points in the entire work area. Using the nth real seismic data in the entire work area, reservoir parameters are predicted through an artificial intelligence inversion network model. To predict the nth longitudinal wave velocity, To predict the nth transverse wave velocity, To obtain the predicted nth density, and based on the predicted reservoir parameters, the traditional Zoeppritz equation forward modeling method is used to generate the second seismic data. For the nth second earthquake data generated by forward modeling, let the second earthquake data Ψ pre The error between the actual seismic data Ψ and the data from the entire work area is used as the loss function L for the unsupervised learning training part. seis By minimizing the loss function L of the unsupervised learning training part seis Training is performed, and the loss function L for the unsupervised learning training part is used. seis The calculation formula is: By iteratively training the AI ​​inversion network model until the total loss function L of the AI ​​inversion network model converges, a well-trained AI inversion network model is obtained. The formula for calculating the total loss function L of the artificial intelligence inversion network model is as follows: L = aL + βL inv L = aL + βL seis , Where α is the weight coefficient of the loss function for the supervised learning training part, and β is the weight coefficient of the loss function for the unsupervised learning training part.

7. A seismic inversion system based on intelligent hybrid learning, characterized in that, include: The data preprocessing module acquires the raw logging data, preprocesses the raw logging data, and obtains the first logging data. The data augmentation module augments the first logging data to obtain the second logging data. The sample set construction module generates first seismic data based on the second well logging data, constructs sample pairs of the first seismic data and the second well logging data, and forms a sample set; The AI ​​inversion network model construction module constructs an AI inversion network model, performs supervised learning training on the AI ​​inversion network model based on a sample set, and performs unsupervised learning training on the AI ​​inversion network model using real seismic data from the entire work area until the total loss function of the AI ​​inversion network model converges, thus obtaining a trained AI inversion network model. The reservoir parameter prediction module inputs real seismic data from the entire work area into a trained artificial intelligence inversion network model and outputs predicted reservoir parameters.

8. An electronic device, characterized in that, It includes a memory and a processor. The memory stores a computer program, and the processor calls the computer program in the memory to execute the seismic inversion method based on intelligent hybrid learning as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the seismic inversion method based on intelligent hybrid learning as described in any one of claims 1-6.

Images