Semi-supervised seismic wave impedance inversion method and system based on Swin Transform iterative gradient correction

Abstract

The invention discloses a semi-supervised seismic wave impedance inversion method and system based on Swin Transform iterative gradient correction, and belongs to the technical field of earthquakes in geophysical exploration. According to the method, the impedance gradient is iteratively corrected in the model space through the deep neural network, and complex nonlinear prior information is captured; and each anisotropic structure direction regularization item is established, and constraint is carried out along the direction of the stratum structure, so that the space consistency of the prediction result is ensured. Meanwhile, the USTNet network designed by the invention combines the advantages of the local feature capture capability of the convolutional neural network and the global modeling capability of Transform, and reduces the non-uniqueness of solutions. The network is used for inverting synthetic data and actual data, and the effectiveness of the method is verified.

Description

Technical Field

[0001] The invention belongs to the field of seismic technology in geophysical exploration, relates to a seismic wave impedance inversion method, and specifically to a semi-supervised seismic wave impedance inversion method and system based on Swin Transformer iterative gradient correction. Background Art

[0002] Post-stack seismic inversion is based on inversion theory and aims to map seismic data into wave impedance data for inferring the lithology and physical properties of underground geological bodies. Since the inversion theory was proposed, a variety of inversion methods have been developed, including sparse pulse inversion, generalized linear inversion, geostatistical methods, and stochastic optimization. However, due to the band-limited characteristics of seismic data, noise interference, and the linear assumption of the forward model, seismic inversion is essentially an ill-posed problem. Therefore, most methods usually introduce various regularization terms, reasonable initial impedance models, and prior knowledge (such as logging data, spatial lithology distribution, and statistical rock properties) as constraints to deal with the ill-posedness of the impedance inversion problem. These constraints are integrated into complex objective functions to obtain reasonable wave impedance solutions. Among them, total variation (TV) regularization is widely used in geophysical inversion. However, the traditional anisotropic TV regularization only constrains the inversion process by minimizing the L1 norm of the horizontal and vertical gradients in the model, without considering the local structural direction of the underground features. Therefore, traditional anisotropic TV regularization is more suitable for blocky structures, but it is difficult to preserve key geological features in more complex stratigraphic structures.

[0003] In recent years, deep learning has achieved remarkable success in fields such as image processing and natural language processing, which has promoted the development of deep learning seismic inversion methods. Deep learning seismic wave impedance inversion methods have achieved remarkable results mainly by establishing data-driven inversion models to map seismic data into impedance. However, in actual seismic inversion problems, this mapping process is usually pathological. Specifically, there is a large null space in the solution space, that is, there are multiple different wave impedance models that can produce almost the same seismic response. This means that some of the changing components in the predicted results will not have a significant impact on the observed values ​​of the seismic data. These components corresponding to the null space cannot be uniquely determined by the observed data, resulting in non-uniqueness and uncertainty of the inversion solution. In addition, most methods attempt to directly recover the impedance from the observed data through a single inversion, which may make the degrees of freedom of the null space appear in the solution. Especially when the data is sparse or noisy, the observed data itself may not be sufficient to fully recover the stable impedance distribution, resulting in increased uncertainty in the solution and reduced inversion accuracy. It is an innovative idea to transform the iterative optimization framework into an end-to-end inversion network. Some studies have constructed a new deep neural network by launching an alternating iterative optimization algorithm, introducing model prior information into the network structure for reflection coefficient and impedance inversion.

[0004] In addition, most deep learning impedance inversion methods use convolutional neural networks (CNNs) because CNNs have significant advantages in local feature extraction, robustness, and scalability. However, CNN methods also have certain limitations. The spatial correspondence between seismic data and impedance profiles is weak, which makes accurate spatial mapping more complicated. This spatial mismatch poses a major challenge to CNNs that rely on strong spatial correspondence, that is, the fixed filter weights and local receptive fields of CNNs are difficult to capture the changing spatiotemporal relationships in seismic data. In addition, the nonlinear relationship between seismic data and impedance requires the simultaneous capture of long-range dependencies and local correlations to cope with non-uniqueness in inversion. In recent years, the success of the Transformer model has opened up new ways to capture global relationships. For example, Fu proposed a joint deep model combining CNN and Transformer for impedance inversion; Ning designed a convolution-enhanced Transformer encoder for seismic inversion, processing time sampling points similar to words in natural language processing. However, these methods use a track-by-track processing approach, which makes it difficult to ensure the spatial continuity of the inversion results. Subsequently, the Visual Transformer (ViT) was introduced into computer vision tasks and gradually applied to the field of seismic inversion. For example, Li developed a ViT-based framework to estimate subsurface properties. However, the ViT method ignores the relationship between internal patches, which is crucial for seismic inversion with intrinsic locality and structural details. The Swin Transformer introduces a "window" mechanism to address this problem by limiting self-attention to a smaller area and using a sliding window to achieve hierarchical global feature capture. This improvement has made significant progress in many fields, but its potential in seismic inversion has not been fully explored. Therefore, it is urgent to develop new methods to break through the limitations of conventional methods and explore the potential performance of new tools to provide core technologies for accurate impedance inversion. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a semi-supervised seismic wave impedance inversion method based on SwinTransformer iterative gradient correction, so as to solve the technical problem that the direct mapping from data space to model space in the prior art leads to increased uncertainty of the solution.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention discloses a semi-supervised seismic wave impedance inversion method based on Swin Transformer iterative gradient correction, comprising:

[0008] The impedance gradient that needs to be inverted is corrected through iterative gradient correction of deep neural network in the seismic wave impedance model space to obtain nonlinear prior information, and a semi-supervised training loss function framework based on iterative gradient correction is established;

[0009] Improve the anisotropic TV regularization and establish the anisotropic structural direction regularization;

[0010] Establish a gradient correction network USTNet and obtain a training data set for training the gradient correction network USTNet;

[0011] The training data set is used to train the gradient correction network USTNet to obtain the trained network USTNet, and the trained network USTNet is used to invert the test data.

[0012] Preferably, the iterative gradient correction method in the model space is:

[0013]

[0014] in, represents the gradient correction operator, and the parameters in the gradient correction operator are represented by Θ; the seismic data d and the output z of the previous iteration k-1 As the gradient correction operator Input.

[0015] Further preferably, a deep neural network is used to replace the gradient term in the traditional gradient descent method. The optimization method of the traditional gradient descent is shown as follows:

[0016]

[0017] Among them, z k represents the impedance of the kth iteration, σ represents the step size, and the brackets represent the objective function for z k-1 The derivative of L 2 Norm loss function and multiple regularization functions ψ; i is used to distinguish the number of regularizations and has no specific meaning, for example, ψ i represents the i-th regularization term.

[0018] The gradient term in the formula is learned through a deep neural network to complete adaptive gradient correction and establish an iterative gradient correction method in the model space.

[0019] More preferably, the semi-supervised training loss function framework based on iterative gradient correction is as follows:

[0020]

[0021] in, represents the forward model, ψ1 and ψ2 represent the supervised loss and spatial constraint loss respectively;

[0022] The supervised loss ψ1 is:

[0023]

[0024] Among them, z ★ is the logging data, which is used as the label data; M represents the mask matrix, M is set to 1 at the position corresponding to the label, and all other positions are set to zero, and the symbol ⊙ represents the matrix point.

[0025] Preferably, the anisotropic TV regularization is as follows:

[0026]

[0027] in, and They represent the partial derivatives along the horizontal and vertical axes, respectively; n and m represent the number of seismic traces and sampling points, respectively; i and j represent the i-th sampling point and the j-th seismic trace, respectively.

[0028] Then constrain the above formula along the direction of the stratum structure to establish the anisotropic structural direction regularization, as shown in the following formula:

[0029]

[0030] Wherein, θ represents the angle between the horizontal axis and the event axis.

[0031] Preferably, a gradient correction network USTNet is established, comprising three parts: an encoder, a bottleneck layer and a decoder;

[0032] The encoder has four layers, the first layer is an input mapping layer, which is composed of a convolution layer with a convolution kernel size of 3×3 and a LeakyReLU activation function; each layer from the second to the fourth layer of the encoder is composed of a Swin Transformer block, a maximum pooling, and a double-layer feature enhancer, wherein the Swin Transformer block is composed of a window attention and a shifted window attention, and the double-layer feature enhancer is composed of two convolution layers with a convolution kernel size of 3×3;

[0033] The bottleneck layer is composed of a Swin Transformer block alone;

[0034] The decoder has four layers. Each of the first three layers is composed of a bilinear interpolation upsampler and is jump-connected to the Swin Transformer block output in the corresponding encoder layer. The last layer of the decoder is the output mapping layer, which is composed of a convolution layer with a convolution kernel size of 3×3, a Linear layer, and a Tanh activation function. The input of this network contains the seismic data d and the output z of the previous iteration. k-1 , the output is the impedance gradient correction value of this iteration.

[0035] Preferably, the acquired training data set includes model data and actual data; wherein:

[0036] For the model data, the known impedance model is used for forward modeling to obtain seismic data. The real impedance model is low-pass filtered and smoothed to obtain the low-frequency impedance model. Part of the real impedance model is used as label data, and the seismic data and low-frequency impedance model are used together as training data sets.

[0037] For the actual data, the well logging records are interpolated and low-pass filtered to obtain the low-frequency impedance model, which is used together with the post-stack seismic data as the training data set;

[0038] Preferably, the network USTNet is trained using the acquired training data set, and the specific operations are as follows:

[0039] The training data is sliced ​​and the sliced ​​data is used as the network input; the low-frequency impedance background is used as the iteration initial value z0; then, the trained network is used to invert the corresponding data to be tested;

[0040] Among them, during the training of the gradient correction network USTNet, the loss function is as follows:

[0041]

[0042] Among them, r represents the number of iterations, λ, μ1, (μ x ,μ y ) are the data matching term, the supervised loss term, the parameters of the anisotropic structural direction regularization along the structural direction and perpendicular to the structural direction respectively;

[0043] The optimization is performed by using the adaptive moment estimation Adam algorithm, and a fixed number of iterations is set as the stopping condition of the iteration.

[0044] The present invention also discloses a system for implementing the above-mentioned semi-supervised seismic wave impedance inversion method based on Swin Transformer iterative gradient correction, comprising:

[0045] A framework building module is used to establish a semi-supervised training loss function framework based on iterative gradient correction according to the iterative gradient correction method in the model space;

[0046] Anisotropic structural orientation regularization building block, used to build anisotropic structural orientation regularization based on improved anisotropic TV regularization;

[0047] Gradient correction network USTNet building module, used to build the gradient correction network USTNet;

[0048] A training data set acquisition module, used to acquire a training data set for training a gradient correction network USTNet;

[0049] The network USTNet training module is used to train the gradient correction network USTNet using the training data set to obtain the trained network USTNet;

[0050] The data inversion module is used to invert the test data using the trained network USTNet.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The method disclosed in the present invention fully combines the advantages of iterative optimization methods and deep learning methods, and can realize seismic wave impedance inversion with high precision. First, the impedance gradient is iteratively corrected in the model space by a deep neural network, which can capture complex nonlinear prior information, and establish a semi-supervised training loss function framework based on iterative gradient correction, which includes data matching terms, supervised loss terms and various anisotropic structural direction regularization terms, to achieve high-precision seismic impedance inversion. At the same time, the present invention designs a USTNet network, which combines the advantages of the local feature capture ability of the convolutional neural network and the global modeling ability of the Transformer, and can reduce the non-uniqueness of the solution. After testing and verification, the present invention can obtain reliable inversion results. Then, the deep network USTNet is trained using the obtained training data set, and the trained deep network USTNet is used to invert the test data, thereby completing the overall inversion process. The present invention verifies the effectiveness of the present invention by using the deep network to invert synthetic data and actual data. The method fully combines the advantages of iterative optimization methods and deep learning methods, can reduce the non-uniqueness of the solution from the direct learning data space to the model space and avoid the lack of global modeling ability in the conventional deep learning method, and realizes high-precision inversion of seismic impedance.

[0053] Furthermore, inspired by the powerful learning ability of deep learning, the present invention adopts a gradient correction operator to replace the explicit calculation of the gradient term in the traditional gradient descent, where the parameters in the learning operator can be obtained through training. Therefore, the high dependence on the selection of regularization terms and the adjustment of hyperparameters in the traditional regularization method is avoided, the generalization ability of the model and the applicability to different data sets are improved, and the model is iteratively corrected in the model space, effectively reducing the overfitting or instability problems that may occur when the model is directly restored from the observed data. In addition, adding seismic data d to guide the update process in each iteration helps to reduce the risk of irreversible cumulative errors and ensure the physical rationality of the inversion solution.

[0054] Furthermore, in order to enhance the performance of deep neural networks, a semi-supervised training loss function framework based on iterative gradient correction is established, which effectively utilizes prior information to enhance the performance of the network, thereby improving the accuracy of inverting complex geophysical data, and also improving the generalization ability, making it robust when processing different data sets. In order to ensure that the model conforms to the characteristics of the stratigraphic structure, the present invention proposes an anisotropic structural direction regularization method, which changes the constraint direction of the traditional anisotropic TV regularization by rotating the matrix, and then constrains along the direction of the stratigraphic structure to obtain more accurate geological results.

[0055] Furthermore, by utilizing the Transformer's powerful global modeling capabilities and combining it with the UNet network, the USTNet network is designed to fully extract the multi-scale features of the data, including multi-scale global features and multi-scale local features. USTNet further reduces the uncertainty of inversion by introducing long-distance dependencies, while allowing the network to obtain higher-quality solutions in the early stages of training, thereby better guiding the network to approach the true solution, thereby improving the inversion accuracy and reliability of the results.

[0056] Furthermore, a training data set for training the deep network USTNet is obtained, and then the deep network USTNet is trained using the obtained training data set. During the training process, the loss function adopts a semi-supervised training loss function framework based on iterative gradient correction. Afterwards, the trained network is used to invert the corresponding test data.

[0057] Furthermore, during the network training process, the adaptive moment estimation (Adam) algorithm can be used for optimization. For the convenience of optimization, a fixed number of iterations is set as the stop condition of the iteration. In addition, the impedance low-frequency background is set to z0. Because PyTorch can automatically calculate gradients and optimize parameters, all source codes are implemented through PyTorch. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Iterative gradient correction process;

[0059] Figure 2 A brief diagram of the stratigraphic structure;

[0060] Figure 3 Deep network USTNet; (a) USTNet overall structure, (b) Swin Transformer block;

[0061] Figure 4 Model training data and true model. The data at the dotted line position is assumed to be logging data. (a) Observed data, (b) impedance low-frequency background, (c) true impedance model.

[0062] Figure 5 The inversion results of the test data set are shown in 1. The data at the dotted line position is assumed to be logging data; among them, (a) the inversion result of the UNet network iteration once, (b) the inversion result of the UNet network iteration twice, (c) the inversion result of the present invention iteration once, (d) the inversion result of the present invention iteration twice;

[0063] Figure 6 The inversion results of the present invention in the noisy test data set are shown in 2. The data at the dotted line position is assumed to be the logging data; among them, (a) the inversion result of the UNet network iteration twice, (b) the inversion result of the present invention iteration once; (c) the inversion result of the present invention iteration once with the addition of ASO constraint; (d) the inversion result of the present invention iteration once with the addition of TV constraint;

[0064] Figure 7 The inversion results of actual data are shown; among them, (a) actual data impedance low-frequency model, (b) actual observed data, (c) the inversion result of the present invention after one iteration, (d) the inversion result of the present invention after two iterations, and (e) the inversion result of the present invention after two iterations with TV regularization added. DETAILED DESCRIPTION

[0065] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0066] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0067] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0068] The present invention proposes a semi-supervised seismic wave impedance inversion method based on Swin Transformer iterative gradient correction. The method establishes a semi-supervised training loss function framework based on iterative gradient correction, and designs a high-performance deep network USTNet. Specifically, the method iteratively corrects the impedance gradient in the model space through a deep neural network to capture complex nonlinear prior information; at the same time, various anisotropic structural direction regularization terms are established to constrain along the direction of the stratigraphic structure to obtain prediction results that are more in line with geological characteristics. In addition, the designed USTNet network combines the advantages of the local feature capture capability of the convolutional neural network and the global modeling capability of the Transformer to reduce the non-uniqueness of the solution. Finally, the network is used to invert synthetic data and actual data to verify the effectiveness of the present invention.

[0069] The semi-supervised seismic wave impedance inversion method based on Swin Transformer iterative gradient correction of the present invention comprises the following steps:

[0070] 1) Establish an iterative gradient correction method in the model space and establish a semi-supervised training loss function framework based on iterative gradient correction;

[0071] 2) Improve the anisotropic TV regularization and establish anisotropic structural direction regularization;

[0072] 3) Establish the gradient correction network USTNet;

[0073] 4) Obtain a training data set for training the gradient correction network USTNet;

[0074] 5) Use the training data set to train the gradient correction network USTNet to obtain the trained network USTNet;

[0075] 6) Use the trained network USTNet to invert the test data.

[0076] The model space described here refers to the model parameter seismic wave impedance that needs to be inverted; the gradient of the impedance that needs to be inverted is corrected. The semi-supervised training loss function framework based on iterative gradient correction includes a spatial constraint. The conventional method for this spatial constraint is to use anisotropic TV regularization, but the anisotropic TV does not work well in complex geology, so it needs to be improved.

[0077] 1. Specific implementation process

[0078] 1. Seismic impedance inversion framework

[0079] Based on the traditional convolution model, seismic records can be modeled using the following mathematical framework:

[0080] d=Wr+n=WDz+n, (1)

[0081] d∈R m×n Represents the observed data, W∈R m×m is the convolution matrix composed of seismic wavelets, r∈R m×n Represents the reflection coefficient, n∈R m×n represents random noise. z∈R (m+1)×n is the seismic impedance to be determined, D∈R m×(m+1) represents the vertical difference operator. m and n represent the number of sampling points and the total number of seismic traces, respectively. The calculation process of the reflection coefficient r can be expressed as

[0082]

[0083] (i,j)∈{1,2,…,m}×{1,2,…,n}, i represents the i-th sampling point, and j represents the j-th seismic trace. The seismic impedance inversion framework aims to find the optimal solution to minimize the loss function between the observed data and the calculated data, which can be expressed as follows

[0084]

[0085] Due to the band-limited characteristics of seismic wavelets and noise interference, the inverse problem in the above formula often exhibits significant ill-posedness, especially in terms of the non-uniqueness of the solution. In order to solve this problem, a regularization constraint term is introduced. Therefore, the optimization model is described as

[0086]

[0087] p represents the total number of regularization constraints corresponding to the variable z, ψ irepresents the i-th regularization term. Currently, there are a variety of algorithms to solve the above optimization problem; however, these optimization methods often face the challenges of limited adaptability, the risk of falling into local minima, and achieving high-resolution inversion results from low signal-to-noise ratio seismic data. In order to improve the inversion accuracy, the present invention combines the iterative optimization strategy with a semi-supervised deep learning framework, and uses a neural network to determine the best update for each iteration.

[0088] 2. Semi-supervised deep learning impedance inversion based on gradient correction

[0089] Through the forward operator Describing Impedance To earthquake data The mapping relationship of , formula (4) can be expressed as follows

[0090]

[0091] Indicates that based on L 2 The loss function of the norm.

[0092] The traditional deep learning method can be expressed as finding the pseudo-inverse operator The process is as follows

[0093]

[0094] is a pseudo-inverse operator with parameter θ, learning the data space To Model Space However, due to the large null space, directly learning from arrive In order to solve this problem, inspired by the classic gradient descent technique in iterative optimization strategy, a method is proposed in the model space. Iterative method for internal operations. This method gradually optimizes the current model by stepping through gradient corrections to make it close to a physically reasonable solution, thereby effectively reducing the impact of the null space. Assuming that the target functional on the right side of equation (5) is Fréchet differentiable and (strictly) convex, a simple gradient descent scheme can be used to find the minimum.

[0095]

[0096] As can be seen from formula (7), prior information is usually explicitly encoded into the objective function or regularization term, which limits the type of prior information that can be incorporated and makes the selection of regularization parameters challenging. In addition, too many regularization terms will introduce a large number of hyperparameters, and the sensitivity of different regularization terms to data types further increases the complexity of hyperparameter selection, ultimately reducing the generalization ability of the model. Such a complex regularization setting not only increases the difficulty of model training, but also may cause inconsistent performance of the model on different data sets, thereby limiting its applicability in actual data. In order to solve the above problems, the present invention estimates the update value in each iteration through a neural network to capture complex nonlinear prior information and automatically learn the update step size. By converting the direct mapping method into a progressive gradient correction in the model space, the present invention effectively alleviates the overfitting or instability problems that may occur when directly restoring the model from the observed data. Therefore, the present invention converts the pseudo-inverse operator Replace update operator The gradient correction scheme can be reformulated as follows

[0097]

[0098] In this iterative framework, seismic data d is introduced in each iteration to guide the updating process, which helps to reduce the risk of irreversible cumulative errors and ensure the physical rationality of the solution, such as Figure 1 As shown. It is worth noting that during the training phase, although the neural network can capture complex nonlinear relationships, the present invention does not completely abandon the use of explicit regularization. On the contrary, the present invention integrates the regularization term into the loss function, making it part of the network learning process. Subsequently, the objective function expression for network training is as follows:

[0099]

[0100] λ is the data loss parameter, r is the number of iterations, ψ1 and ψ2 are logging constraints (supervised loss) and spatial structure constraints, respectively, and the impedance low-frequency background is set to z0. This strategy improves the accuracy of the network in inversion of complex geophysical data by more effectively utilizing prior information, while enhancing its generalization ability, making it more robust when processing diverse data sets. The logging constraint ψ1 can be written as follows

[0101]

[0102] z ★is the well logging data, as the label data, M represents the mask matrix, M is set to 1 at the position corresponding to the label, and all other positions are set to zero, and the symbol ⊙ represents the matrix dot product. In order to improve the performance of the model, it is crucial to incorporate the spatial structure constraint ψ2. This constraint can effectively capture the spatial characteristics of the data, thereby enhancing the predictive ability of the model and improving its accuracy and stability in complex geological environments.

[0103] 3. Anisotropic Structural Orientation Regularization (ASO)

[0104] For the spatial structure constraint ψ2, the present invention introduces anisotropic structural direction regularization to further enhance the inversion capability of the network. This regularization increases the sensitivity of the network to directional changes in the data, thereby significantly improving the accuracy and robustness of the inversion in complex geological environments. In order to effectively integrate the spatial structure, the present invention starts from the perspective of the coordinate system. The standard Cartesian coordinate system consists of a horizontal axis (denoted as x) and a vertical axis (denoted as y). In the inversion task, anisotropic TV regularization is usually applied along the x and y directions, by calculating the derivative and then applying L 1 Norm to achieve smoothing effect. The specific expression is as follows

[0105]

[0106] and denote the partial derivatives along the horizontal and vertical axes, respectively, and n and m denote the number of seismic traces and sampling points, respectively. Similarly, we impose constraints on the structural direction of the geological layer. Figure 2 As shown, the simplified stratigraphic structure shows three different types of geological structures, each with a different dip. The dip is defined as the angle between the horizontal axis and the reflection axis (isochronous axis), which can be estimated by the plane wave destruction (PWD) method. The traditional anisotropic TV regularization smoothes the model along the Cartesian coordinate system, whose direction is represented by the solid line. However, for structures II and III, especially when dealing with steep dips, this approach is not applicable because such constraints introduce significant artifacts.

[0107] To address this limitation, a more appropriate smoothing strategy should be adopted for Structures II and III. Instead of smoothing along the Cartesian coordinate system, smoothing should be performed along the local coordinate system defined by the dashed line, which is aligned with the geological structure direction, including the directions parallel and perpendicular to the strata. This method ensures that the regularization follows the natural direction of the geological features, thus generating more geologically realistic results. By introducing the rotation matrix, the anisotropic structural regularization can be expressed as follows

[0108]

[0109] By observing formula (12), we can see that anisotropic TV regularization is essentially a special case of the ASO constraint when the tilt angle is zero (scenario I).

[0110] Finally, the network training loss function can be expressed as

[0111] λ、μ1、(μ x ,μ y ) are the data matching term, the supervised loss term (logging constraint), and the parameters of anisotropic structural direction regularization along the structural direction and perpendicular to the structural direction, respectively.

[0112] 4. USTNet network structure

[0113] like Figure 3 As shown in (a), the designed network USTNet combines the multi-scale feature extraction capability of UNet with the ability of SwinTransformer to capture global context. The key to this network is to introduce the ST block in the encoder to capture global features at different scales through downsampling, while avoiding the introduction of redundant information in the upsampling process. Unlike the use of ST blocks in the decoder, the present invention retains spatial continuity and solves potential information loss problems through jump connections. The ST block in the encoder adopts a "shift window" mechanism to enhance the model's ability to model long-distance dependencies. In addition, unlike traditional multi-head attention that only captures the relationship between adjacent blocks, ST applies local self-attention within non-overlapping windows, thereby improving computational efficiency. However, the window division of ST may introduce artifacts or discontinuities, such as boundary effects, in the block division process. To alleviate these problems, ASO regularization can be used to effectively eliminate the impact of such artifacts and ensure the continuity and accuracy of feature representation. As shown Figure 3 As shown in (b), the network also integrates a convolution-enhanced MLP (CeMLP) module in the ST block to better capture pixel-level features, which is particularly important for high-resolution tasks such as seismic inversion. This combination of global context capture and local feature extraction distinguishes USTNet from traditional convolutional models and Transformer-based models. Because PyTorch can automatically calculate gradients and optimize parameters, all source code is implemented using PyTorch.

[0114] 2. Experimental Effect Analysis

[0115] 1. Synthetic signal calculation example

[0116] In the present invention, the Marmousi II model is used to synthesize acoustic impedance and seismic data to verify the effectiveness of the proposed method. The channel signals of the Marmousi II benchmark model are downsampled to a total of 1361 channels with an interval of 10. Each channel signal contains 551 time sampling points with a time interval of 1 millisecond. A Ricker waveform with a main frequency of 30 Hz is used to generate 2D synthetic seismic data from 2D Marmousi II impedance data. The channels located at CDP 200, 500, 750 and 1100 are then designated as logging labels, whose true impedance is known. The remaining 1357 signals are regarded as unlabeled data, whose true impedance is unknown, and are used to evaluate the inversion results. The low-frequency impedance model is obtained by low-pass filtering the true impedance model. Figure 4 2D synthetic seismic data, low-frequency background, and true impedance profiles are shown, with the dotted lines indicating well logging labels. During training, only four well logging labels, 2D low-frequency impedance profiles, and 2D seismic profiles were used.

[0117] The importance of iterative updates is analyzed by comparing the single and double iteration inversion results of UNet and the proposed network USTNet, where the TV constraint is incorporated into the training loss function of all networks in this test. Figure 5 As shown, Figure 5 Figure (a) shows the inversion result of the UNet network after one iteration, Figure (b) shows the inversion result of the UNet network after two iterations, Figure (c) shows the inversion result of the present invention after one iteration, and Figure (d) shows the inversion result of the present invention after two iterations. It can be found through visual observation that the resolution of the result of UNet iteration twice is higher than that of the result of one iteration, and the inversion result is close to the true solution. This shows that the iterative gradient correction strategy proposed in the present invention can improve the inversion accuracy. However, by observing Figures (c) and (d), it can be found that the network USTNet designed by the present invention is not much different from the one after two iterations visually, and one iteration can be better than UNet iteration twice, which shows the high efficiency of the USTNet network designed by the present invention.

[0118] The inversion results were quantitatively analyzed, and the mean square error between each inversion result and the true inversion solution was calculated, which was 4.85×10 -3 3.00×10 -3 , 2.46×10 -3 , 2.31×10 -3 Through comparative analysis, it is found that the USTNet inversion result has a smaller mean square error than the true solution, which indicates that the inversion accuracy of the method of the present invention is very high. Figure 4 15dB noise is added to (a) to verify the sensitivity of the method of the present invention to noise. The inversion results of this test are shown in Figure 6As shown, (a) is the inversion result of the UNet network after two iterations, (b) is the inversion result of the present invention after one iteration; (c) is the inversion result of the present invention after one iteration with the addition of ASO constraints; (d) is the inversion result of the present invention after one iteration with the addition of TV constraints. Through visual observation, it can be found that the addition of ASO regularization and TV regularization to the USTNet network helps to suppress the discontinuous segmentation lines (artifacts) generated by the Swin Transformer, thereby enhancing the lateral continuity and overall inversion smoothness of the inversion results. However, the inversion results of these two regularizations are very close to the true solution and have higher resolution. This may be because the two-dimensional deep network also captures the spatial structural information of the data, making the difference between the two regularization methods relatively small. In addition, the estimation of the dip angle will also affect the results, especially in noisy seismic data, where it becomes challenging to accurately capture the underground structure. The inversion results were quantitatively analyzed, and the peak signal-to-noise ratios between the inversion results and the true solution were calculated, which were 25.2901, 26.0672, 28.8770, and 29.1262, respectively. Through comparative analysis, it is found that after USTNet adds ASO regularization and TV regularization, the inversion result has a larger peak signal-to-noise ratio than the true solution, which shows that the method of the present invention has good noise resistance.

[0119] 2. Calculation examples based on actual seismic data

[0120] A section of a post-stack three-dimensional data volume of an oil field was selected to test the method of the present invention. The section size is 216×110, the sampling rate is 1ms, and there are four logging curves in the section. We selected one of the wells, Well2, as a blind well for verification, and used the other three wells to interpolate and pass through a low-pass filter to obtain the impedance low-frequency background as the initial value of the iteration. It has been shown in the synthetic record test that based on two-dimensional deep learning, TV regularization and ASO regularization have similar performance, and ASO regularization is easily affected by noise. Therefore, in the actual seismic data example, we use TV regularization as a simplification of ASO regularization to test spatial structure constraints. Figure 7The inversion results of actual data are shown, where (a) is the impedance low-frequency model of actual data, (b) is the actual observed post-stack data, (c) is the inversion result of the present invention after one iteration, (d) is the inversion result of the present invention after two iterations, and (e) is the inversion result of the present invention after two iterations with TV regularization added. Visual observation shows that compared with the inversion result of one iteration, the inversion result of two iterations has more detailed information and higher resolution. The main reason is that the actual test data is small, and the limited training data limits the ability of Swin Transformer to capture comprehensive seismic features on a global scale. After two iterations, it can effectively learn a wider range of features, thereby improving the inversion results. At the same time, we observe that even without applying any spatial constraints, the inversion results do not exhibit the artifacts seen in the synthetic data test. This can be attributed to the smaller block size used when processing the actual data, which enables the smaller window of Swin Transformer to effectively capture long-range dependencies. Finally, in order to verify the reliability of the inversion results, the impedance logging curve is inserted into Figure 7 Through observation, it is found that the wellside inversion results of Well 2 blind well are well matched with the logging curve, which shows that the inversion credibility of the method adopted by the present invention is high.

[0121] In summary, the present invention establishes a semi-supervised training loss function framework based on iterative gradient correction, and designs a high-performance deep network USTNet. The invention uses a deep neural network to iteratively correct impedance gradients in the model space to capture complex nonlinear prior information; at the same time, various anisotropic structural direction regularization terms are established to constrain along the direction of the stratum structure to ensure the spatial consistency of the prediction results. In addition, the designed USTNet network combines the advantages of the local feature capture capability of the convolutional neural network and the global modeling capability of the Transformer to reduce the non-uniqueness of the solution. Finally, the network is used to invert synthetic data and actual data to verify the effectiveness of the present invention.

[0122] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction, characterized in that: include: The impedance gradient is corrected iteratively in the seismic wave impedance model space through a deep neural network to obtain nonlinear prior information, and a semi-supervised training loss function framework based on iterative gradient correction is established; Improve the anisotropic TV regularization and establish the anisotropic structural direction regularization; Establish a gradient correction network USTNet and obtain a training data set for training the gradient correction network USTNet; The training data set is used to train the gradient correction network USTNet to obtain the trained network USTNet, and the trained network USTNet is used to invert the test data.

2. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 1 is characterized in that: The iterative gradient correction method in the seismic impedance model space is: in, represents the gradient correction operator, and the parameters in the gradient correction operator are represented by Θ; the seismic data d and the output z of the previous iteration k-1 As the gradient correction operator Input.

3. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 2 is characterized in that: Using deep neural networks to replace the gradient term in the traditional gradient descent method, the optimization method of traditional gradient descent is as follows: Among them, z k represents the impedance of the kth iteration, σ represents the step size, and the brackets represent the objective function for z k-1 The derivative of L 2 Norm loss function and multiple regularization functions ψ; The gradient term in the formula is learned through a deep neural network to complete adaptive gradient correction and establish an iterative gradient correction method in the model space.

4. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 3 is characterized in that: The framework of the semi-supervised training loss function based on iterative gradient correction is as follows: in, represents the forward model, ψ1 and ψ2 represent the supervised loss and spatial constraint loss respectively; The supervised loss ψ1 is: Among them, z ＊ is the logging data, which is used as the label data; M represents the mask matrix, M is set to 1 at the position corresponding to the label, and all other positions are set to zero, and the symbol ⊙ represents the matrix point.

5. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 1, characterized in that: The anisotropic TV regularization is shown as follows: in, and Represent the partial derivatives along the horizontal axis and the vertical axis, respectively; n and m represent the number of seismic traces and the number of sampling points, respectively; i and j represent the i-th sampling point and the j-th seismic trace, respectively; Then constrain the above formula along the direction of the stratum structure to establish the anisotropic structural direction regularization, as shown in the following formula: Wherein, θ represents the angle between the horizontal axis and the event axis.

6. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 1, characterized in that: Establish a gradient correction network USTNet, which includes three parts: encoder, bottleneck layer and decoder; The encoder has four layers, the first layer is an input mapping layer, which is composed of a convolution layer with a convolution kernel size of 3×3 and a LeakyReLU activation function; each layer from the second to the fourth layer of the encoder is composed of a Swin Transformer block, a maximum pooling, and a double-layer feature enhancer, wherein the Swin Transformer block is composed of a window attention and a shifted window attention, and the double-layer feature enhancer is composed of two convolution layers with a convolution kernel size of 3×3; The bottleneck layer is composed of a Swin Transformer block alone; The decoder has four layers. Each of the first three layers consists of a bilinear interpolation upsampler and is jump-connected to the Swin Transformer block output in the corresponding encoder layer. The last layer of the decoder is the output mapping layer, which consists of a convolutional layer with a convolution kernel size of 3×3, a Linear layer, and a Tanh activation function. The input of this network contains the seismic data d and the output z of the previous iteration. k-1 , the output is the impedance gradient correction value of this iteration.

7. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 1, characterized in that: The acquired training data set includes model data and actual data; among which: For the model data, the known impedance model is used for forward modeling to obtain seismic data. The real impedance model is low-pass filtered and smoothed to obtain the low-frequency impedance model. Part of the real impedance model is used as label data, and then used together with the seismic data and the low-frequency impedance model as the training data set; For the actual data, the well logging records are interpolated and low-pass filtered to obtain the low-frequency impedance model, which is then used together with the post-stack seismic data as a training data set.

8. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 1, characterized in that: Use the acquired training data set to train the network USTNet. The specific operations are as follows: The training data is sliced ​​and the sliced ​​data is used as the network input; the low-frequency impedance background is used as the iteration initial value z0; then, the trained network is used to invert the corresponding test data.

9. The semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction according to claim 8, characterized in that: During the training of the gradient correction network USTNet, the loss function is as follows: Among them, r represents the number of iterations, λ, μ1, (μ x ,μ y ) are the data matching term, the supervised loss term, the anisotropic structural direction regularization parameters along the structural direction and perpendicular to the structural direction respectively; The optimization is performed by using the adaptive moment estimation Adam algorithm, and a fixed number of iterations is set as the stopping condition of the iteration.

10. A system for implementing the semi-supervised seismic impedance inversion method based on Swin Transformer iterative gradient correction as described in any one of claims 1 to 9, characterized in that: include: A framework building module is used to establish a semi-supervised training loss function framework based on iterative gradient correction according to the iterative gradient correction method in the model space; Anisotropic structural orientation regularization building block, used to build anisotropic structural orientation regularization based on improved anisotropic TV regularization; Gradient correction network USTNet building module, used to build the gradient correction network USTNet; A training data set acquisition module, used to acquire a training data set for training a gradient correction network USTNet; The network USTNet training module is used to train the gradient correction network USTNet using the training data set to obtain the trained network USTNet; The data inversion module is used to invert the test data using the trained network USTNet.

Images