Seismic data unmixing method and system based on cascaded deep neural network prior

Abstract

The application discloses a seismic data deblending method and system based on a cascaded deep neural network prior, adopts a random time jitter strategy to construct an aliasing matrix Γ, multiplies the aliasing matrix Γ by effective seismic data d to be separated to obtain aliasing seismic data b in a common receiver domain; solving d is converted into solving a regularized least square optimization problem, an effective signal is obtained by introducing a regularization term, a nonlinear estimation expression is converted into a new objective function by introducing an intermediate auxiliary variable u=d based on the thought of a HQS algorithm; based on the objective function, a denoising empty convolutional neural network is used to solve the noise suppression problem u (k+1) ; a closed-form sub-problem d (k+1) is solved; pseudo-deblended common shot gathers Γ T b and the aliasing matrix Γ are input into a cascaded deep neural network for end-to-end training, after cascaded deep neural network parameter information is obtained, pseudo-deblended common receiver gathers to be deblended are taken as input of the cascaded network, and deblended seismic data is obtained. Deblending of different noise types is realized.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, specifically relating to a method and system for unmixing seismic data based on cascaded deep neural network priors. Background Technology

[0002] Overlapping seismic data acquisition can excite more seismic sources within the same acquisition time, resulting in higher acquisition density and spatial sampling rate, and increasing the number of times the subsurface medium is covered, thus obtaining high-quality, comprehensive acquisition results. However, the overlapping signals from multiple seismic sources pose a significant challenge to the data processing and interpretation of conventional seismic data. To facilitate post-acquisition analysis, it is necessary to separate the signals excited by single seismic sources from the multi-source overlapping acquisition data.

[0003] Traditional physics-driven demixing methods include filtering-based and inversion-based methods. Filtering methods simply treat the separation problem as a noise attenuation problem, while inversion methods solve the separation problem of aliased data as a linear inverse problem, aiming to estimate the required unknown unaliased data. However, both methods often rely on manual parameter tuning, such as filter length and the threshold for sparsity-enhancing transformation. The quality of parameter selection often affects the effective signal separation results from seismic data.

[0004] In recent years, deep learning has been introduced into the field of seismic data unmixing. When dealing with large datasets or complex data features, complex classification or regression problems often pose significant challenges to traditional methods such as statistical learning, especially in high-dimensional data scenarios. As one of the most popular deep learning algorithms, neural networks can automatically learn the features and relationships hidden in large datasets, thereby matching seismic data in a non-linear manner. However, data-driven neural networks have strict requirements on the dataset. Combining data-driven methods with traditional physics-driven methods can enhance the network's understanding of data features and improve unmixing performance without changing the size of the existing dataset.

[0005] Disadvantages of existing technology:

[0006] 1. The unmixing performance of most current deep learning methods largely depends on the construction of training data. Collecting appropriate training data to avoid generalization problems is a challenging task, and the prior information of the data is often specified by humans based on experience.

[0007] 2. It is difficult to enhance the network's ability to understand data features and improve the unmixing effect without changing the size of the existing dataset. Summary of the Invention

[0008] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a seismic data unmixing method and system based on cascaded deep neural network priors. First, by introducing a neural network, prior information of the separation objective function is learned, thereby improving the separation effect. Then, using the idea of ​​iterative optimization, the unmixing process is decomposed into multiple alternating iterative sub-problems, which are iteratively expanded through cascaded deep neural networks to form an end-to-end unmixing framework, effectively solving the separation problem and adaptively learning data features to improve the universality of the algorithm. Finally, the sub-networks are replaced according to actual needs to achieve unmixing of different noise types.

[0009] The present invention adopts the following technical solution:

[0010] The seismic data unmixing method based on cascaded deep neural network priors includes the following steps:

[0011] S1. A random time jitter strategy is used to construct the aliasing matrix Γ, and the aliased seismic data b is obtained by multiplying it with the effective seismic data d to be separated in the common detection domain.

[0012] S2. Transform the solution of d into a regularized least squares optimization problem by introducing a regularization term. Obtain a valid signal

[0013] S3. Using the idea of ​​HQS algorithm, the nonlinear estimation expression obtained in step S2 is transformed into a new objective function by introducing an intermediate auxiliary variable u = d.

[0014] S4. Based on the objective function obtained in step S3, use a denoising dilated convolutional neural network to address the noise suppression problem. (k +1) Solve the problem;

[0015] S5. Solve the closed-form subproblem d based on step S4. (k+1) =(Γ T Γ+ρI) -1 (Γ T b+ρu (k) );

[0016] S6. Based on the results obtained in steps S4 and S5, the pseudo-solution mixed shot point collection Γ T b and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. After obtaining the parameter information of the cascaded deep neural network, the pseudo-demixed common receiver gather to be demixed is used as the input of the cascaded network to obtain the demixed seismic data.

[0017] Preferably, in step S1, the acquisition is performed with the detector fixed and without considering noise during the acquisition process. The aliased seismic data b specifically includes:

[0018] b=Γd

[0019] Where Γ is the aliasing matrix containing time jitter; d is the effective seismic data to be separated.

[0020] Preferably, in step S2, the valid signal for:

[0021]

[0022] Where λ is a non-negative Lagrange multiplier, and the first term The first item is for data fidelity, and the second item is for regularization.

[0023] Preferably, in step S3, the new objective function is:

[0024]

[0025] Where ρ is the non-negative penalty weight, λ is the non-negative Lagrange multiplier, and the first term... For data fidelity items, For regularization terms, This is an auxiliary variable term that is introduced.

[0026] More preferably, in the new objective function, the data fidelity term remains quadratic, the auxiliary variables are decoupled, and after further transformation, the constrained optimization problem is converted into two subproblems that are solved alternately and iteratively; for u (k+1) The solution process is equivalent to a typical noise suppression problem, i.e., fixing d. (k) For clean target data (k+1) The effective signal components to be solved; for d (k+1) Solving this problem is a closed-form problem; by designing a new fusion network framework, the prior is implicitly learned by using the network structure composed of two iterative subproblems as input instead of the regularization operator.

[0027] More preferably, the constrained optimization problem is transformed into two subproblems that are solved iteratively and alternately:

[0028]

[0029] Where k∈{1,2,···,n} represents the k-th iteration, denoiser(·) represents the denoising module, and Γ is the aliasing matrix. T b is the pseudo-solution mixed shot point set, ρ is the non-negative penalty weight, and u (k) The prior valid signal components to be solved after noise suppression.

[0030] Preferably, step S4 specifically includes:

[0031] In training the denoising dilated convolutional neural network, a common shot gather is used as an approximation of the common receiver gather to train the network. A aliasing matrix Γ is constructed using parameters with an aliasing degree of 1 and random time jitter within a range of 200 time sampling points. In a single aliasing acquisition, the time difference between the excitations of the two sources is greater than or equal to the signal length, and the signals generated by the two sources do not alias in time. The aliasing matrix Γ = [H1,…,H...]. n ],get Combining this with b = Γd, we get Γ H b represents the pseudo-unmixing result of the aliased seismic data, and the pseudo-unmixed common shot point gather Γ H Using b as input and clean aliased common shot point gathers as output, the prior network is trained end-to-end; then, pseudo-decoded aliased common receiver gathers containing aliasing noise are fed into the prior network to obtain clean data after suppressing aliasing noise.

[0032] Preferably, in step S5, the closed-form solution of the objective function is specifically as follows:

[0033]

[0034] Where k∈{1,2,···,n} represents the k-th iteration, denoiser(·) represents the denoising module, Γ is the aliasing matrix, ρ is the non-negative penalty weight, and u (k) Let d be the prior effective signal component to be solved after noise suppression. (k) b represents clean target data, δ represents aliased seismic data, and δ is a parameter controlling the gradient descent step size.

[0035] Preferably, in step S6, the aliasing matrix Γ does not change with network iteration.

[0036] Secondly, embodiments of the present invention provide a seismic data unmixing system based on cascaded deep neural network priors, comprising:

[0037] The data module uses a random time jitter strategy to construct an aliasing matrix Γ, which is then multiplied with the effective seismic data d to be separated in the common detection domain to obtain the aliased seismic data b.

[0038] The regularization module transforms the solution of d into a regularized least squares optimization problem by introducing a regularization term. Obtain a valid signal

[0039] The function module, using the idea of ​​HQS algorithm, transforms the nonlinear estimation expression obtained by the regularization module into a new objective function by introducing an intermediate auxiliary variable u = d.

[0040] The solution module, based on the objective function obtained from the function module, utilizes a denoising dilated convolutional neural network to address the noise suppression problem u. (k+1) Solve the problem;

[0041] The calculation module solves the closed-form subproblem d based on the solution module. (k+1) =(Γ T Γ+ρI) -1 (Γ T b+ρu (k) );

[0042] The output module, based on the results obtained from the solution module and the calculation module, outputs the pseudo-solution mixed shot point gather Γ. T b and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. After obtaining the parameter information of the cascaded deep neural network, the pseudo-demixed common receiver gather to be demixed is used as the input of the cascaded network to obtain the demixed seismic data.

[0043] Compared with the prior art, the present invention has at least the following beneficial effects:

[0044] This invention presents a seismic data unmixing method based on cascaded deep neural network priors. It combines inversion-type problems with deep neural network priors, reducing the reliance on large datasets in supervised learning. The unmixing objective is treated as a noise energy minimization problem. By introducing intermediate variables, the unmixing process is decomposed into two alternating iterative sub-problems. The first sub-problem is transformed into extracting the coherent components and suppressing the incoherent components from the cascaded seismic data using convolutional neural network priors. The second sub-problem uses a gradient update strategy to approximate the solution for the effective signal data. This method is end-to-end trainable and incorporates a learning mechanism at each level, enabling adaptive learning of data features and improving the algorithm's universality. Furthermore, the sub-networks can be replaced according to actual needs to achieve unmixing of different noise types.

[0045] Furthermore, the seismic data aliasing process generates incoherent wavefields. When the coherence between the effective signal and interference is weak, the effective signal in the aliased seismic data can be separated using demixing methods. The key to seismic data aliasing acquisition lies in designing an aliasing scheme that allows for the acquisition of as much information as possible within the same timeframe, while simultaneously enabling effective separation of the aliased signals through algorithms. Therefore, the design of incoherence characteristics should be considered in the design of seismic aliasing acquisition strategies. Introducing randomness into the source excitation sequence during the aliasing acquisition process can enhance the incoherence characteristics. This invention introduces random time jitter during source excitation, generating incoherent wavefields that retain complete temporal and spatial distances, which is crucial for the separation of effective signals from aliased seismic data.

[0046] Furthermore, based on the physical meaning of the aliasing matrix Γ, the process of solving for d is underdetermined. Therefore, it is transformed into solving a regularized least squares optimization problem by introducing a regularization term. To obtain a definite solution The expression for solving this problem now contains two terms, the first term... This is a data fidelity term to ensure the unmixing result obtained by the algorithm. The second term is consistent with the clean, unaliased data d; it is a regularization term. It provides prior information on clean, unaliased data; it transforms the solution of d into solving a regularized least squares optimization problem and introduces a regularization term. To obtain a definite solution.

[0047] Furthermore, this invention aims to use neural networks as learning operators to learn prior information, rather than manually specifying it based on experience. Data-trained neural networks can represent data features, containing implicit prior information about the data. To integrate data-driven deep networks with traditional inversion techniques, this invention utilizes the HQS (Half-quadratic splitting) algorithm, introducing an intermediate auxiliary variable u = d to transform the nonlinear estimate into a new objective function, enabling solutions over a larger data domain. In the improved objective function, the data fidelity term remains quadratic, the auxiliary variable is decoupled, and the original objective function can be obtained by jointly minimizing the auxiliary variable and the original objective. In the new objective function, the data fidelity term remains quadratic, the auxiliary variable is decoupled, and the original objective function can be obtained by jointly minimizing the auxiliary variable. The original objective is obtained; after further transformation, the constrained optimization problem is transformed into two subproblems that are solved alternately and iteratively; where u (k+1) The solution process is equivalent to a typical noise suppression problem, i.e., fixing d. (k) For clean target data u (k+1) The effective signal components to be solved; and for d (k+1)The solution can be viewed as a closed-loop problem. By designing a new fusion network framework, the network structure composed of two iterative subproblems is used as input to replace the regularization operator to implicitly learn the prior. Furthermore, the noise suppression problem is further solved using a neural network, transforming it into a noise suppression problem and a closed-loop problem. The subproblem containing the prior is transformed into a noise suppression operator problem in the iterative algorithm. This invention designs a new fusion network framework, using the network structure composed of two iterative subproblems as input to replace the regularization operator to implicitly learn the prior, enabling end-to-end training. Then, a suitable dataset is constructed to train the network. Since the acquired aliased common shot gathers and non-aliased common receiver gathers have similar geological characteristics, common shot gathers can be used as approximations of common receiver gathers during network training. Different aliasing matrices Γ are constructed by selecting different random time jitter ranges and different aliasing degrees to generate aliased seismic data. The pseudo-demixing results of the aliased seismic data are solved through pseudo-demixing operations, and the pseudo-demixed common shot gathers Γ are then used to train the network. T b is used as the input to the prior network, and the aliased common shot gather is used as the output. Three different noise suppression network structures are selected to train the network end-to-end. Then, the pseudo-solution aliased common receiver gather containing aliasing noise is fed into the prior network to obtain clean data after suppressing aliasing noise.

[0048] Furthermore, regarding the closed-form solution of the subproblem d... (k+1) =(Γ T Γ+ρI) -1 (Γ T b+ρu (k) This involves the Γ T The inverse process of Γ + ρI. In unmixing problems, Γ is often very large. T Γ also has a very large rank, which is often undesirable in inversion problems, leading to extremely high computational costs and time consumption. At the same time, the introduction of the learnable parameter ρ also potentially affects the solution of Γ during the iterative process. T The inverse of Γ+ρI is required. In this invention, the inverse process is transformed into a gradient descent optimization problem, and an approximation of the optimal solution is sought using the iterative classical conjugate gradient algorithm: Through gradient descent, d can be updated in one step. (k+1) It converges to a local optimum.

[0049] Furthermore, based on steps S4 and S5, the pseudo-solution mixed shot point collection Γ Tb and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. The aliasing matrix Γ remains unchanged during network iterations. After obtaining the parameters of the cascaded deep neural network, the pseudo-unaliased co-receiver gathers to be dealiased are used as input to the cascaded network to obtain the dealiased seismic data. Since aliased co-shot gather data and clean, unaliased co-receiver gather data have similar structural characteristics in the time domain diagram, when probing the same stratigraphic structure, co-shot gathers and co-receiver gathers exhibit the same characteristics in characterizing the stratigraphic structure, i.e., in wave morphology. Therefore, the dealiasing method proposed in this invention is applicable not only to co-shot gathers but also to co-receiver gathers. It can adaptively learn data characteristics, improving the algorithm's universality. Furthermore, sub-networks can be replaced according to actual needs to achieve dealiasing for different noise types.

[0050] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0051] In summary, this invention can effectively solve the separation problem when acquiring seismic data in aliased form, and it adaptively learns data features to improve the universality of the method. At the same time, the sub-network can be replaced according to actual needs to achieve demixing of different noise types.

[0052] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0053] Figure 1 This is a flowchart of the present invention;

[0054] Figure 2 The diagram shows the cascaded deep neural network framework of this invention. The entire network contains k sub-network modules, naturally incorporating the iterative process of the optimization problem. Each sub-network module corresponds to an update step, and each sub-network module contains two training sub-blocks: a neural network module (Network-Block, NB) and an inversion correlation solving module (Closed Block, CB), which update and solve the two sub-problems respectively.

[0055] Figure 3 This is a schematic diagram of the earthquake data velocity model of the present invention, in which different colors represent different velocities and characterize different inter-layer impedance characteristics;

[0056] Figure 4 This is a schematic diagram of the CNND network structure of the present invention;

[0057] Figure 5 This is a schematic diagram of the DnCNN network structure of the present invention;

[0058] Figure 6 This is a schematic diagram of the denoising dilated convolutional neural network (CNNL) structure of the present invention;

[0059] Figure 7 This is a schematic diagram of the simulated data aliasing common shot gather for training cascaded deep neural networks according to the present invention. It has similar structural features to the non-aliased common receiver gather data in the time domain plot.

[0060] Figure 8 This is a schematic diagram of the actual data aliasing common shot point gather used in this invention for training cascaded deep neural networks;

[0061] Figure 9 The real co-detector point gathers used in this invention are derived from open-source datasets for testing.

[0062] Figure 10 The real co-detector point gathers used in this invention are derived from open-source datasets for testing.

[0063] Figure 11 Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot gather data, the first simulated aliased data demixing result is obtained using a two-layer network structure. Among them, (a) is the first data CNND result (26.87dB), (b) is the difference between the first data CNND results, (c) is the first data CNNL result (20.57dB), (d) is the difference between the first data CNNL results, (e) is the first data DnCNN result (29.16dB), and (f) is the difference between the first data DnCNN results.

[0064] Figure 12 Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot point gather data, the demixing result of the 48th simulated aliased data is obtained using a two-layer network structure. Among them, (a) is the CNND result of the 48th data (23.23dB), (b) is the difference of the CNND result of the 48th data, (c) is the CNNL result of the 48th data (18.49dB), (d) is the difference of the CNNL result of the 48th data, (e) is the DnCNN result of the 48th data (26.49dB), and (f) is the difference of the DnCNN result of the 48th data.

[0065] Figure 13 The diagram shows the average amplitude spectrum of the unmixed results of the 1st and 48th simulated aliased data under a two-layer network structure. (a) is the average amplitude spectrum of the unmixed results of the 1st simulated data, and (b) is the average amplitude spectrum of the unmixed results of the 48th simulated data.

[0066] Figure 14 The comparison of the 16th channel of the demixing results of the 1st and 48th simulated aliased data under a two-layer network structure is shown in (a) and (b).

[0067] Figure 15 Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot gather data, the first simulated aliased data demixing result is obtained using a three-layer network structure. Among them, (a) is the first data CNND result (27.60dB), (b) is the difference of the first data CNND result, (c) is the first data CNNL result (22.88dB), (d) is the difference of the first data CNNL result, (e) is the first data DnCNN result (33.96dB), and (f) is the difference of the first data DnCNN result.

[0068] Figure 16 Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot point gather data, the demixing result of the 48th simulated aliased data is obtained using a three-layer network structure. Among them, (a) is the CNND result of the 48th data (23.92dB), (b) is the difference of the CNND result of the 48th data, (c) is the CNNL result of the 48th data (20.34dB), (d) is the difference of the CNNL result of the 48th data, (e) is the DnCNN result of the 48th data (32.94dB), and (f) is the difference of the DnCNN result of the 48th data.

[0069] Figure 17 The diagram shows the average amplitude spectrum of the unmixed results of the 1st and 48th simulated aliased data under a three-layer network structure. (a) is the average amplitude spectrum of the unmixed results of the 1st simulated data, and (b) is the average amplitude spectrum of the unmixed results of the 48th simulated data.

[0070] Figure 18 The comparison of the 16th channel of the demixing results of the 1st and 48th simulated aliased data under a three-layer network structure is shown in (a) and (b).

[0071] Figure 19Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot point gather data, the first actual aliased data demixing result is obtained using a two-layer network structure. Among them, (a) is the first data CNND result (18.59dB), (b) is the difference between the first data CNND results, (c) is the first data CNNL result (14.68dB), (d) is the difference between the first data CNNL results, (e) is the first data DnCNN result (20.19dB), and (f) is the difference between the first data DnCNN results.

[0072] Figure 20 Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot point gather data, the demixing result of the 48th actual aliased data is obtained using a two-layer network structure. Among them, (a) is the CNND result of the 48th data (14.76dB), (b) is the difference of the CNND result of the 48th data, (c) is the CNNL result of the 48th data (12.52dB), (d) is the difference of the CNNL result of the 48th data, (e) is the DnCNN result of the 48th data (16.54dB), and (f) is the difference of the DnCNN result of the 48th data.

[0073] Figure 21 The diagram shows the average amplitude spectrum of the unmixed results of the 1st and 48th actual aliased data under a two-layer network structure. (a) is the average amplitude spectrum of the unmixed result of the 1st actual data, and (b) is the average amplitude spectrum of the unmixed result of the 48th actual data.

[0074] Figure 22 The comparison of the 16th channel of the demixing results of the 1st and 48th actual aliased data under a two-layer network structure is shown in (a) and (b).

[0075] Figure 23Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot point gather data, the first actual aliased data demixing result is obtained using a three-layer network structure. Among them, (a) is the first data CNND result (19.27dB), (b) is the difference of the first data CNND result, (c) is the first data CNNL result (17.45dB), (d) is the difference of the first data CNNL result, (e) is the first data DnCNN result (22.60dB), and (f) is the difference of the first data DnCNN result.

[0076] Figure 24 Based on the cascaded neural network proposed in this invention, CNND, CNNL, and DnCNN are embedded as prior substructures of the network. After training on aliased common shot point gather data, the demixing result of the 48th actual aliased data is obtained using a three-layer network structure. Among them, (a) is the CNND result of the 48th data (15.70dB), (b) is the difference of the CNND result of the 48th data, (c) is the CNNL result of the 48th data (13.84dB), (d) is the difference of the CNNL result of the 48th data, (e) is the DnCNN result of the 48th data (19.03dB), and (f) is the difference of the DnCNN result of the 48th data.

[0077] Figure 25 The diagram shows the average amplitude spectrum of the unmixed results of the 1st and 48th actual aliased data under a three-layer network structure. (a) is the average amplitude spectrum of the unmixed result of the 1st actual data, and (b) is the average amplitude spectrum of the unmixed result of the 48th actual data.

[0078] Figure 26 The comparison of the 16th channel of the demixing results of the 1st and 48th actual aliased data under a three-layer network structure is shown in (a) and (b).

[0079] Figure 27 A schematic diagram of a computer device provided in an embodiment of the present invention;

[0080] Figure 28 This is a block diagram of a chip provided according to an embodiment of the present invention. Detailed Implementation

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] This invention provides a seismic data unmixing method based on cascaded deep neural network priors. Starting from the aliased acquisition of seismic data, it analyzes and processes the aliasing patterns using common receiver gather data. The unmixing inverse problem is modeled as a regularized least squares optimization problem, and a supervised unmixing method utilizing deep neural network priors is proposed. The regularized least squares optimization problem is decomposed into two subproblems using the Half-quadratic splitting (HQS) optimization algorithm. The subproblem containing the regularization term can be viewed as a Gaussian noise suppression problem, solved using a deep denoising network; the quadratic term subproblem is approximated by gradient updates. This iterative algorithm is expanded into a cascaded structure and trained end-to-end using a training dataset. This invention effectively solves the separation problem during aliased seismic data acquisition, adaptively learns data features to improve the algorithm's universality, and allows for sub-network replacement to achieve unmixing of different noise types based on actual needs.

[0083] While traditional methods offer good separation results, they often require manual parameter tuning, such as adjusting filter length and the threshold for sparsity-enhanced transformation. The quality of these parameters significantly impacts the effective signal separation from seismic data. Furthermore, supervised learning deep networks, which have emerged in recent years, aim to establish a relationship between input sample data and the desired output or true values. The quality and availability of training data are crucial to the effectiveness of the training process and the accuracy of predictions. The quality of the labeled data used during training should be as close to the true values ​​as possible, and supervised learning also has certain requirements on the number of samples in the dataset, which is often difficult to achieve in the field of seismic data processing. For example, when training on data from aliased seismic acquisitions, ideal single-source excitation data as the true values ​​is unavailable. Therefore, collecting appropriate training data to avoid generalization problems is a challenging task for most current deep learning methods. Moreover, prior information about the data is often manually assigned based on experience, making it difficult to enhance the network's understanding of data features and improve unmixing performance without changing the size of the existing dataset. Therefore, combining data-driven methods with traditional physical-driven methods can enhance the network's understanding of data features and improve unmixing performance without changing the size of the existing dataset.

[0084] This invention combines inversion-type problems with the prior knowledge of deep neural networks, reducing the dependence on large datasets in supervised learning. It treats the unmixing objective as a noise energy minimization problem, and by introducing intermediate variables, decomposes the unmixing process into two alternating iterative sub-problems. The first sub-problem is transformed into extracting the coherent components and suppressing the incoherent components from the cascaded seismic data using convolutional neural network priors. The second sub-problem uses a gradient update strategy to approximate the solution for the effective signal data. This is end-to-end trainable, and a learning mechanism is introduced at each level to achieve the optimal unmixing objective.

[0085] This invention, based on the inversion concept, transforms the unmixing problem of cascaded seismic data into an optimization objective problem. Using the HQS (Half-quadratic splitting) algorithm, intermediate variables are introduced into the optimization objective problem, transforming it into two iteratively solved sub-problems. Based on the unrolling concept, a neural network is introduced as a learning operator to learn prior information from the seismic data. Then, the two iteratively solved sub-problems are embedded into the network framework, constructing a cascaded deep neural network that adaptively solves the two sub-problems in the unmixing objective to obtain the optimal unmixing objective.

[0086] Please see Figure 1 This invention discloses a seismic data unmixing method based on cascaded deep neural network priors, comprising the following steps:

[0087] S1. With the detector fixed and the noise of the acquisition process not considered, a random time jitter strategy is used to construct the aliasing matrix Γ, and the aliased seismic data b is obtained by multiplying it with the effective seismic data d to be separated in the common detection domain.

[0088] The random time jitter strategy is adopted as follows:

[0089] Seismic data aliasing is a process that generates incoherent wavefields. When the coherence between the effective signal and interference is weak, the effective signal in the aliased seismic data can be separated using demixing methods. The key to successful seismic data aliasing acquisition lies in designing an aliasing scheme that allows for the acquisition of as much information as possible within the same timeframe, while simultaneously enabling effective separation of the aliased signals through algorithms. Therefore, the design of seismic aliasing acquisition strategies should incorporate consideration of incoherence characteristics. Introducing randomness into the source excitation sequence during aliasing acquisition can enhance incoherence. This invention introduces random time jitter during source excitation, generating incoherent wavefields that retain complete temporal and spatial distances, which is crucial for separating the effective signal from the aliased seismic data.

[0090] After adopting the random time jitter strategy, the aliased seismic data b is specifically as follows:

[0091] In multi-source acquisition, if we use d to represent the seismic data acquired by a single source and b(t,r) to represent the signal recorded by the r-th detector at time t during multi-source acquisition, the acquired seismic data can be represented as:

[0092]

[0093] Where t∈{1,...,N-1}Δt represents different sampling times, Δb represents the sampling time interval (Δb is not fixed but random), and N represents the total number of sampling points recorded by the detector. s Let d(t+τ) be the total number of seismic sources that the r-th detector can effectively record. i ,s i r) represents the earthquake originating from the i-th single source s i Random time delay or jitter τ i The signal at time t is excited and acquired by the r-th detector.

[0094] Furthermore, the aliasing matrix Γ is constructed to represent the aliased seismic data b as follows:

[0095] The aliasing matrix Γ = [N1,…,H] n ], where H i This represents the set of columns in the aliasing matrix, used to partition H. i Represented as:

[0096]

[0097] in, For a size of N s ×N s unit array, 0 i1 For a size of N0×N s A zero matrix of size 0 i2 For size (NN) s -N0)×N s The zero matrix is ​​N0, which is the source excitation delay, and varies with the source excitation sequence at different locations and the known random time jitter.

[0098] The common receiving domain is selected as the effective domain for the main separation of effective signals. Within the common receiving domain, a random time-dithered aliasing acquisition technique is employed with a fixed detector. Rewritten as:

[0099] b=Γd

[0100] Where b represents the aliased seismic data; Γ represents the aliasing matrix including time jitter; and d represents the effective seismic data to be separated. Based on the physical meaning of the aliasing matrix, the solution to the above equation is underdetermined.

[0101] S2. Based on the physical meaning of the aliasing matrix Γ, the process of solving for d is underdetermined. Therefore, it is transformed into solving a regularized least squares optimization problem by introducing a regularization term. To obtain a definite solution The expression for solving this problem now contains two terms, the first term... This is a data fidelity term to ensure the unmixing result obtained by the algorithm. The second term is consistent with the clean, unaliased data d; it is a regularization term. It provides prior information about clean, unaliased data;

[0102] The solution of d is transformed into solving a regularized least squares optimization problem, and a regularization term is introduced. To obtain a definite solution, as follows:

[0103]

[0104] Where λ is a non-negative Lagrange multiplier, and the first term This is a data fidelity term to ensure the unmixing result obtained by the algorithm. The first term is consistent with the clean, unaliased data d. The second term is a regularization term that provides prior information about the clean, unaliased data, such as its sparsity properties in certain transform domains.

[0105] S3. Using the idea of ​​HQS algorithm, the nonlinear estimation expression obtained in step S2 is transformed into a new objective function by introducing an intermediate auxiliary variable u = d.

[0106] In the new objective function, the data fidelity term remains quadratic, and the auxiliary variables are decoupled. The original objective function can be achieved by jointly minimizing the auxiliary variables. The original objective is obtained; after further transformation, the constrained optimization problem is transformed into two subproblems that are solved alternately and iteratively; where u (k+1) The solution process is equivalent to a typical noise suppression problem, i.e., fixing d. (k) For clean target data u (k+1) The effective signal components to be solved; and for d (k+1) The solution can be viewed as a closed-form problem; by designing a new fusion network framework, the prior is implicitly learned by using the network structure composed of two iterative subproblems as input instead of the regularization operator.

[0107] Using the HQS algorithm, intermediate auxiliary variables are introduced as follows:

[0108] This invention aims to use neural networks as learning operators to learn prior information, rather than manually specifying it based on experience. Data-trained neural networks can represent data features, containing implicit prior information. To integrate data-driven deep networks with traditional inversion techniques, this invention utilizes the HQS (Half-quadratic splitting) algorithm, introducing an intermediate auxiliary variable u = d to transform the nonlinear estimate into a new objective function, enabling solutions over a larger data domain. In the improved objective function, the data fidelity term remains quadratic, the auxiliary variable is decoupled, and the original objective function can be obtained by jointly minimizing the auxiliary variable and the original objective. Therefore, the equation... Represented as:

[0109]

[0110] Where ρ is the non-negative penalty weight.

[0111] The constrained optimization problem is transformed into two subproblems that are solved iteratively and alternately as follows:

[0112]

[0113] Where k∈{1,2,···,n} represents the k-th iteration. This can be viewed as a typical noise suppression problem, i.e., assuming a fixed d (k) For clean target data u (k+1) The effective signal components to be solved are denoted as .

[0114] The network structure consisting of two iterative subproblems is used as input to implicitly learn prior information, specifically as follows:

[0115] To match the size of the common receiver point gather data, a random pruning scheme was adopted for the aliased common shot point data. In addition, various data augmentation strategies were employed, including random left-right flipping, random pruning, and the addition of weak-energy Gaussian noise, in CNND (such as...). Figure 4 ), DnCNN (such as Figure 5 ) and CNNL (such as Figure 6 Based on the selection of the three methods, the pseudo-solution mixed shot point collection Γ Tb and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. The aliasing matrix Γ does not change with network iterations. In this example, the simulated data network is trained for 100 epochs, using mean absolute loss as the loss function and a learning rate of 0.001, and the Adam optimization algorithm is used to train the network. In the actual data example, mean absolute loss is used as the loss function, a learning rate of 0.0008 is selected, and the Adam optimization algorithm is used to train the network.

[0116] The noise suppression problem is further solved using neural networks, transforming it into a noise suppression problem and a closed-form solution problem, as follows:

[0117]

[0118] Here, denoiser(·) represents the denoising module. The subproblem containing the prior is transformed into a noise suppression operator problem in the iterative algorithm. This invention designs a novel fusion network framework that uses the network structure composed of two iterative subproblems as input to replace the regularization operator to implicitly learn the prior, enabling end-to-end training.

[0119] S4. Based on step S3, utilize the powerful learning ability of neural networks to address the noise suppression problem. (k+1) Solve the problem;

[0120] Please see Figure 2 Based on the cascaded deep neural network framework, a suitable dataset is constructed. In this example, a layered model is used for aliased acquisition of numerical simulation data to synthesize the simulation data. Figure 3 The velocity model used in the simulation is shown, with a vertical depth of 2000 meters. To make the simulation data more consistent with actual geological conditions, the first 1524 meters of the model containing velocity anomalies were specifically selected for simulation. A total of 128 geophones and 128 seismic sources were horizontally positioned during the simulation. The spacing between seismic sources was 12 meters, and the spacing between geophones was also chosen to be equal at 12-meter intervals. All geophones and seismic sources were located on the Earth's surface. Figure 6 Showing based on Figure 2 The mixed shot point gathers obtained by the velocity model simulation in the middle are sampled at a time interval of 0.002 seconds, with 750 sample points extracted. There are 128 single source excitation data with an inter-channel interval of 12 meters.

[0121] Since the acquired aliased common shot gathers and unaliased common receiver gathers have similar geological characteristics, common shot gathers can be used as approximations of common receiver gathers during network training. Different aliasing matrices Γ are constructed by selecting different random time jitter ranges and different degrees of aliasing to generate aliased seismic data. The pseudo-demixing results of the aliased seismic data are solved through a pseudo-demixing operation, and the pseudo-demixed common shot gathers Γ are then used to... T b is used as the input to the prior network, and the aliased common shot gather is used as the output. Three different noise suppression network structures are selected to train the network end-to-end. Then, the pseudo-solution aliased common receiver gather containing aliasing noise is fed into the prior network to obtain clean data after suppressing aliasing noise.

[0122] The simulated data network was trained for 100 epochs, using mean absolute loss as the loss function and a learning rate of 0.001, and the Adam optimization algorithm was used to train the network. In the actual data example, mean absolute loss was used as the loss function, a learning rate of 0.0008 was selected, and the Adam optimization algorithm was used to train the network.

[0123] Addressing the problem of noise suppression using the powerful learning capabilities of neural networks. (k+1) The solution is as follows:

[0124] This invention selects three noise suppression network structures to verify their effectiveness in suppressing aliasing noise in simulated and real data, and embeds these three networks into the cascaded network framework proposed in this invention, as follows:

[0125] (1) Simple convolutional neural network: This network contains only a few simple convolutional network layers and has a relatively simple structure. In order to introduce nonlinear terms into the network structure, ReLU is added to the network to improve its ability to deal with noise suppression problems. The cascaded neural network method with simple convolutional neural network structure as noise reduction module is called CNNL.

[0126] (2) DnCNN (Denoising convolutional neural networks) can accelerate the training process by using residual learning and batch normalization. It shows excellent performance in Gaussian denoising under unknown energy levels. It can use the residual learning strategy to implicitly remove potential clean images in the hidden layer, achieve a high level of noise estimation, and thus achieve separation between noise and effective clean image data. The cascaded neural network method that uses DnCNN as the denoising substructure is called DnCNN.

[0127] (3) Denoising dilated convolutional neural networks introduce dilated convolutions of different sizes to expand the receptive field in order to seek a more favorable result for demixing, and are called denoising dilated convolutional neural networks. Dilated convolution introduces a dilation rate parameter in traditional convolution, which allows the convolution kernel to skip some pixels in local areas for calculation, thereby expanding the receptive field of the convolution kernel to a certain extent. This allows the convolution operation to obtain a wider range of information and extract richer features. The cascaded neural network method that embeds denoising dilated convolutional neural networks is called the CNND method.

[0128] Constructing a suitable dataset and using common shot point gathers as approximate substitutes for common receiver point gathers during network training is specifically as follows:

[0129] Because overlapping common shot gather data and clean, non-overlapping common receiver gather data have similar structural characteristics in the time domain plot, when probing the same stratigraphic structure, common shot gathers and common receiver gathers exhibit the same characteristics in characterizing the stratigraphic structure, i.e., in wave morphology. The degree of overlap refers to the extent to which the (n+1)th source overlaps with the previous source during the acquisition process.

[0130] Considering both the efficiency of the aliasing acquisition process and the difficulty of dealiasing, after designing and testing experiments, parameters were selected to construct the aliasing matrix Γ, with an aliasing degree of 1 and a random time jitter range within 200 time sampling points (which can be considered as [0, 0.4] s in simulated data and [0, 0.8] s in actual data). Figure 7 The aliased common shot point gather b obtained from the simulation and Figure 8 The aliased common shot point gathers from the open-source dataset are used for pseudo-demixing.

[0131] Specifically, it is defined as the proportion of the overlapping portion of the signals from the (n+1)th source and the nth source to the length of the (n+1)th source signal. If the time difference between the excitation of the two sources in a single acquisition is greater than or equal to the length of the signal, it can be considered that the signals generated by the two sources do not overlap in time, and the degree of overlap is 0.

[0132] The aliasing matrix Γ = [H1,…,H n ], where H i This represents the set of columns in the aliased matrix. For a size of N s ×N s unit array, 0 i1 For a size of N0×N s A zero matrix of size 0 i2 For a size of (NN)s -N0)×N s The zero matrix.

[0133] Accordingly, it is possible to obtain but

[0134]

[0135] Combining the above equation with b = Γd, we have:

[0136]

[0137] Γ H b represents the pseudo-demixing result of aliased seismic data. The part on the right side of the formula is the aliasing noise, which is incoherent in the data organization domain outside the common shot domain. Pseudo-demixing makes the resulting data have the same number of traces as the un-aliased data, and the aliasing interference is converted into incoherent noise components.

[0138] To match the size of the common-receiver point gather data, a random pruning scheme is adopted for the aliased common-shot point data. In addition, various data augmentation strategies are employed, including random left-right flipping, random pruning, and the addition of weak-energy Gaussian noise. Based on the selection of three methods—CNND, CNNL, and DnCNN—this invention uses the pseudo-resolved aliased common-shot point gather Γ H Using b as input and clean aliased common shot point gathers as output, the prior network is trained end-to-end.

[0139] Then, the pseudo-decoded mixed detector gather containing aliasing noise is fed into the prior network to obtain clean data after suppressing aliasing noise.

[0140] Using based Figure 2 The common receiver point gathers obtained from the velocity model simulation (e.g.) Figure 9 (as shown) and real co-detection point data from open-source datasets (such as...) Figure 10 The pseudo-unmixed co-detector point gathers (as shown) were used to verify the network's effectiveness. The simulated co-detector point data had a sampling time interval of 0.002 seconds, with 750 sampling points, including 128 channels from a single source excitation, and an inter-channel interval of 12 meters. The real co-detector point data had 512 sampling points per channel per data slice, each containing 128 channels, with an inter-channel interval of 12.5 meters and a sampling time interval of 0.004 seconds per channel.

[0141] S5. Solve the closed-form subproblem d based on step S4. (k+1) =(Γ T Γ+ρI) -1 (Γ T b+ρu (k) );

[0142] Because it involves Γ T The inverse of Γ + ρI incurs extremely high computational costs and time consumption. The introduction of the learnable parameter ρ also potentially impacts the solution of Γ during the iterative process. T The inverse of Γ+ρI is required. In this invention, the inverse process is transformed into a gradient descent optimization problem, and the iterative classical conjugate gradient algorithm is used to find an approximation of the optimal solution.

[0143] Specifically, for closed-form problems involving the objective function:

[0144] Closed-form solution of subproblem d (k+1) =(Γ T Γ+ρI) -1 (Γ T b+ρu (k) This involves the Γ T The inverse process of Γ + ρI. In unmixing problems, Γ is often very large. T Γ also has a very large rank, which is often undesirable in inversion problems, leading to extremely high computational costs and time consumption. At the same time, the introduction of the learnable parameter ρ also potentially affects the solution of Γ during the iterative process. T The inverse of Γ+ρI is required. In this invention, the inverse process is transformed into a gradient descent optimization problem, and an approximation of the optimal solution is sought using the iterative classical conjugate gradient algorithm:

[0145]

[0146] in, δ is a parameter that controls the step size of gradient descent. Gradient descent allows us to update d in a single step. (k+1) It converges to a local optimum.

[0147] Furthermore, the above equation can be rewritten as:

[0148] d (k+1) =d (k) -δ[Γ T (Γd (k) -b)+0.5ρ(d (k) -u (k) )]

[0149] Therefore, the objective function can be rewritten as:

[0150]

[0151] S6. Based on steps S4 and S5, the pseudo-solution mixed shot point collection Γ Tb and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. The aliasing matrix Γ does not change with network iteration. After obtaining the parameter information of the cascaded deep neural network, the pseudo-demixed common receiver gather to be demixed is used as the input of the cascaded network to obtain the demixed seismic data.

[0152] In another embodiment of the present invention, a seismic data unmixing system based on cascaded deep neural network priors is provided. This system can be used to implement the above-mentioned seismic data unmixing method based on cascaded deep neural network priors. Specifically, the seismic data unmixing system based on cascaded deep neural network priors includes a data module, a regularization module, a function module, a solution module, a calculation module, and an output module.

[0153] The data module uses a random time jitter strategy to construct an aliasing matrix Γ, which is multiplied with the effective seismic data d to be separated in the common detection domain to obtain the aliased seismic data b.

[0154] The regularization module transforms the solution of d into a regularized least squares optimization problem by introducing a regularization term. Obtain a valid signal

[0155] The function module, using the idea of ​​HQS algorithm, transforms the nonlinear estimation expression obtained by the regularization module into a new objective function by introducing an intermediate auxiliary variable u = d.

[0156] The solution module, based on the objective function obtained from the function module, utilizes a denoising dilated convolutional neural network to address the noise suppression problem u. (k+1) Solve the problem;

[0157] The calculation module solves the closed-form subproblem d based on the solution module. (k+1) =(Γ T Γ+ρI) -1 (Γ T b+ρu (k) );

[0158] The output module, based on the results obtained from the solution module and the calculation module, outputs the pseudo-solution mixed shot point gather Γ. T b and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. After obtaining the parameter information of the cascaded deep neural network, the pseudo-demixed common receiver gather to be demixed is used as the input of the cascaded network to obtain the demixed seismic data.

[0159] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment can be used for the operation of a seismic data unmixing method based on cascaded deep neural network priors, including:

[0160] A random time jitter strategy is used to construct the aliasing matrix Γ, which is then multiplied with the effective seismic data d to be separated in the common detection domain to obtain the aliased seismic data b. Solving for d is transformed into solving a regularized least squares optimization problem by introducing a regularization term. Obtain a valid signal Utilizing the HQS algorithm, the nonlinear estimation expression is transformed into a new objective function by introducing an intermediate auxiliary variable u = d. Based on this objective function, a denoising dilated convolutional neural network is used to address the noise suppression problem u. (k+1) Solve the closed-form subproblem d. (k+1) ; to set up pseudo-solution mixed cannon points Γ T b and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. After obtaining the parameter information of the cascaded deep neural network, the pseudo-demixed common receiver gather to be demixed is used as the input of the cascaded network to obtain the demixed seismic data.

[0161] Please see Figure 27The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the fluid composition calculation method in the reservoir stimulation wellbore of this embodiment. To avoid repetition, details are omitted here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the seismic data unmixing system based on cascaded deep neural network priors of this embodiment. To avoid repetition, details are omitted here.

[0162] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 27 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0163] The processor 61 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0164] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or RAM of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0165] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0166] Please see Figure 28 The terminal device is a chip. In this embodiment, the chip 600 includes a processor 622, which may be one or more, and a memory 632 for storing computer programs executable by the processor 622. The computer program stored in the memory 632 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processor 622 may be configured to execute the computer program to perform the aforementioned seismic data unmixing method based on cascaded deep neural network priors.

[0167] Additionally, chip 600 may also include a power supply component 626 and a communication component 650. The power supply component 626 can be configured to perform power management of chip 600, and the communication component 650 can be configured to enable communication of chip 600, such as wired or wireless communication. Furthermore, chip 600 may also include an input / output (I / O) interface 658. Chip 600 can operate on an operating system stored in memory 632.

[0168] In another embodiment of the present invention, a storage medium is also provided, specifically a computer-readable storage medium (memory). This computer-readable storage medium is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0169] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the seismic data unmixing method based on cascaded deep neural network priors in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps:

[0170] A random time jitter strategy is used to construct the aliasing matrix Γ, which is then multiplied with the effective seismic data d to be separated in the common detection domain to obtain the aliased seismic data b. Solving for d is transformed into solving a regularized least squares optimization problem by introducing a regularization term. Obtain a valid signal Utilizing the HQS algorithm, the nonlinear estimation expression is transformed into a new objective function by introducing an intermediate auxiliary variable u = d. Based on this objective function, a denoising dilated convolutional neural network is used to address the noise suppression problem u. (k+1) Solve the closed-form subproblem d. (k+1) ; to set up pseudo-solution mixed cannon points Γ T b and the aliasing matrix Γ are fed into a cascaded deep neural network for end-to-end training. After obtaining the parameter information of the cascaded deep neural network, the pseudo-demixed common receiver gather to be demixed is used as the input of the cascaded network to obtain the demixed seismic data.

[0171] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0172] Utilize based on Figure 2 The common receiver point gathers obtained from the velocity model simulation (e.g.) Figure 9 ) and real co-detection point data from open-source datasets (such as Figure 10 The pseudo-unmixed common-detector point gathers obtained were used to verify the effectiveness of the network. To demonstrate the effectiveness of the method, given that the effective signals used to construct the dataset are known, the signal-to-noise ratio (SNR) was introduced to objectively evaluate the unmixing results. The result obtained after demixing, where d is the true label value. The signal-to-noise ratio expression is:

[0173]

[0174] This example discusses the impact of different numbers of layers and different sub-network structures on the unmixing accuracy of the method proposed in this invention. The time-domain plot of the unmixing result obtained using a two-layer cascaded network for the first simulation data is shown in Figure 11. Figure 11 (a) shows the results obtained based on the CNND method, demonstrating that the cascaded network proposed in this invention can achieve a beneficial suppression effect on aliasing noise. Figure 11 (c) The CNNL results still contain residual noise components, and the difference results... Figure 11(d) reveals a large number of effective signal components. Compared to the previous two methods, the cascaded DnCNN achieves superior demixing results, such as... Figure 11 As shown in (e), the grayscale image has almost no strong aliasing noise components, which is reflected in the difference. Figure 11 (f) shows that only a small portion of the noise components were not completely suppressed, with almost no significant loss of effective signal. The signal-to-noise ratios (SNRs) of the two-layer cascaded CNND, CNNL, and DnCNN are 26.87 dB, 20.57 dB, and 29.16 dB, respectively. Based on the two-layer cascade, the demixing result of the 48th data point is as follows: Figure 12 As shown, the signal-to-noise ratios obtained by using two-layer cascades with three different subnetwork structures are 23.23 dB, 18.49 dB, and 26.49 dB, respectively. Among them, as shown... Figure 12 (b) Differences in CNND results Figure 12 (d) The differences in the CNNL results all contain relatively obvious effective signal components, which presents a significant disadvantage in the demixing problem. Consistent with the first data, the results obtained by DnCNN after two cascaded layers still achieve relatively good signal fidelity results in the difference profile. Figure 12 It is difficult to find effective target components in (f). The average amplitude spectrum comparison diagram is shown below. Figure 13 As shown in the figure, the black lines represent the clean, valid signals that are not aliased, the red lines represent the results obtained after the pseudo-demixing operation of the aliased data, the blue and green lines represent the results of the two-stage cascaded CNND and CNNL methods, respectively, and the yellow lines represent the results obtained by the two-stage cascaded DnCNN network. Figure 13 (a) and Figure 13 (b) Comparison of the mean amplitude spectrum for the first and 48th simulation data, respectively. From Figure 13 As shown in (a), the two-layer cascaded neural network method based on the three sub-network structure priors can achieve approximate regression of the effective signal. Compared with other priors, CNNL, when used as a prior, does not completely suppress the low-frequency components of 0-5Hz, which is evident in... Figure 13 (b) A similar situation can also be observed. This actually corresponds to the results in the time-domain plot. Due to its relatively large energy, the aliasing noise exhibits a bar-like shape in the time domain, which is reflected as a low-frequency component in the frequency spectrum. Then, the present invention analyzes and compares the 16th time-domain waveform in the demixing result, and the results are as follows. Figure 14 As shown, comparing the time-domain differences reveals that all three methods achieve good demixing results, restore the effective signal to a high quality, and significantly reduce the noise interference generated by the pseudo-demixing operation.

[0175] Furthermore, the time-domain plot of the demixing result of the first simulation data using the three-layer cascaded network is shown in Figure 15. Figure 15(a) shows the demixing results obtained by cascading three layers of CNND. Figure 15 (c) and Figure 15 (e) The results obtained from three-layer cascaded CNNL and DnCNN networks are shown respectively. The signal-to-noise ratios (SNRs) of the three frameworks are 27.6 dB, 22.88 dB, and 33.96 dB, respectively, which fully verifies the effectiveness of the proposed framework in the demixing task. Comparing the two-layer cascaded case, taking CNNL as the sub-network structure as an example, the difference can be analyzed... Figure 15 As shown in (b), using a higher number of cascaded layers can more effectively improve the network's performance in the demixing task, better suppressing aliasing noise components while better protecting the valid signal. In the other two sub-network structures, increasing the number of cascaded layers yields better results. The result of demixing the 48th aliased data using a three-layer cascaded network is shown below. Figure 16 As shown. For data with more complex features, more cascaded layers result in a higher signal-to-noise ratio compared to fewer cascades. Similar to the first data set, when using CNND as the sub-network structure, a three-layer cascade yields a more beneficial result than a two-layer cascade. A comparison of the difference plots shows that... Figure 16 The difference in the three-layer cascaded DnCNN results shown in (f) contains the least effective signal damage component, which also proves that the selection of the noise suppression network is also of great significance to obtaining high-quality unmixing results under the unmixing framework proposed in this invention. Figure 17 The average amplitude spectrum of the demixed results of the first and 48th simulation data is shown. It can be seen that the amplitude spectrum of the demixed result almost completely overlaps with the effective signal. Figure 18 (a) By comparing the time-domain waveform of channel 16 of the first simulation data, it can be seen that the energy of some aliasing noise has a strong amplitude, which will cause great interference to the effective signal. Figure 18 (b) shows the comparison of the time-domain waveform of the 16th channel of the 48th data. It can be seen that the method proposed in this invention can significantly suppress the aliasing noise component, and the result obtained is almost completely consistent with the effective signal.

[0176] Furthermore, the effectiveness of the two-layer and three-layer cascaded networks was validated using actual co-detection point data from open-source datasets. In network training on real data, mean absolute loss was used as the loss function, a learning rate of 0.0008 was selected, and the Adam optimization algorithm was employed for network training. Following the simulated data processing steps, the two-layer cascaded network was used, and its performance on the first set of real data was analyzed. Figure 19 In (a), embedding CNND into the network achieves a result of 18.59 dB, a significant improvement compared to the 11.19 dB obtained by processing it alone. The results for CNNL and DnCNN are shown below. Figure 19 As shown in (c) and 19(e), relatively high levels of demixing results can also be achieved, at 14.68 dB and 20.19 dB respectively. The difference results of the three methods show that all three methods can effectively separate aliased signals; however, the CNNL result still suffers from insufficient noise suppression compared to the other two, and all three methods exhibit some degree of signal impairment. Similarly, comparing the two methods of using CNND and CNNL in cascaded and standalone denoising modules, we can draw the same conclusion as in the simulation data: introducing CNND with a larger receptive field and shallower layers into the cascaded network can effectively improve the network's performance in demixing. Meanwhile, the 48th actual dataset was also used to verify the two-layer cascaded network structure, such as... Figure 20 As shown. When faced with real-world data with relatively more complex characteristics, the three methods achieved signal-to-noise ratios of 14.76dB, 12.52dB, and 16.54dB respectively, which is sufficient to prove the effectiveness of the proposed method. The difference plot shows some loss of effective signal energy, which requires further improvement of the network. Simultaneously, this example performs average amplitude spectrum and single-channel time-domain analysis on real-world data, such as... Figure 21 and Figure 22 As shown in the average amplitude spectrum, it is quite evident in the low-frequency range of 0-5Hz that, compared to the results of CNNL cascade, CNND is more effective at suppressing low-frequency noise components, while DnCNN cascade achieves results that are even more desirable than the former two. Figure 22 As can be seen, the pseudo-demixing operation introduces high-amplitude noise components into the single-channel signal. After processing by three cascaded networks with different substructures, the interference noise energy of abnormal patterns can be suppressed.

[0177] Furthermore, the time-domain plot of the demixing results of the 1st and 48th actual data points from the three-layer cascaded network is obtained by... Figure 23 and Figure 24 As shown, in the case of a three-layer cascade, a demixing result with a signal-to-noise ratio of 17.45dB can be obtained based on CNNL. Figure 23 As shown in (c), the contour has been largely recovered from the valid signal, but the network has generated interference components. Compare the differences. Figure 23 (d) It can be seen that the damage to the effective signal components still exists. However, compared to Figure 19 (d) shows the difference between the two cascaded layers. Figure 23 The results shown in (c) demonstrate a significant improvement. Similarly, Figure 23 (a) shows the demixing result obtained by applying CNND to the first data in the case of three-level cascade, which is more effective in suppressing interference components than CNNL at the same level of cascade. Figure 23(e) shows the demixing results of DnCNN in the three-layer cascaded configuration, achieving a signal-to-noise ratio of 22.6 dB, resulting in better effective signal protection and aliasing suppression. Similar to the first dataset, the demixing experiment on the 48th actual dataset also confirms that the cascaded DnCNN network performs optimally. Compared to the first dataset results, the demixing difference graphs for the three cascaded network structures show more loss of effective signal components, consistent with the previous simulated data experiments. When the data becomes more complex, the network's ability to extract features is limited. Then, Figure 25 The average amplitude spectra of the demixing results for the first and 48th data points in the three-layer cascade case are compared. It can be seen that the cascaded networks embedding the three network substructures can all achieve the recovery of the effective signal components. Compared to the results of CNND and DnCNN, the CNNL results show differences from the effective signal in both high and low frequency components, which corresponds to the signal-to-noise ratio results mentioned earlier. The time-domain waveform plotted using the 16th channel is used for comparison. Figure 26 As shown, the method proposed in this invention can effectively suppress high-amplitude interference present in the initial pseudo-unmixing results. Figure 26 As can be seen in (b), compared to CNNL and CNND, the three-layer cascade of DnCNN can achieve the best demixing effect.

[0178] The above simulation and actual seismic data unmixing examples illustrate that the seismic data unmixing method based on cascaded deep neural network priors proposed in this invention can be applied to the unmixing of aliased seismic data. This avoids manually specifying the prior information of the data based on experience, improves the network's ability to understand data features without changing the size of the existing dataset, and enhances the unmixing effect.

[0179] In summary, the present invention provides a seismic data unmixing method and system based on cascaded deep neural network priors, which avoids manually specifying prior information of the data based on experience, thus increasing the workload; and can enhance the network's ability to understand data features and improve the unmixing effect without changing the size of the existing dataset.

[0180] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A seismic data unmixing method based on cascaded deep neural network priors, characterized in that, Includes the following steps: S1. Constructing the aliasing matrix using a random time jitter strategy. Through the effective seismic data to be separated Multiplying in the common-detector domain yields aliased seismic data. ; S2, Solve The problem is transformed into solving a regularized least squares optimization problem by introducing a regularization term. Obtain a valid signal ; S3. Utilizing the HQS algorithm, by introducing intermediate auxiliary variables. The nonlinear estimation expression obtained in step S2 is transformed into a new objective function, which is: in, Non-negative penalty weights For non-negative Lagrange multipliers, the first term For data fidelity items, For regularization terms, For the introduction of auxiliary variable terms; In the new objective function, the data fidelity term remains quadratic, the auxiliary variables are decoupled, and after further transformation, the constrained optimization problem is converted into two subproblems that are solved alternately and iteratively; The solution process is equivalent to a typical noise suppression problem, i.e., fixed... For clean target data The effective signal components to be solved; for Solving this problem is a closed-form problem; by designing a new fusion network framework, the prior is implicitly learned by using the network structure composed of two iterative subproblems as input instead of the regularization operator. The constrained optimization problem is transformed into two subproblems that are solved iteratively and alternately as follows: in, Indicates the first k In the next iteration, denoiser(·) represents the denoising module. It is an aliased matrix. For the pseudo-solution of the mixed communist cannon point collection, for, Non-negative penalty weights The prior valid signal components to be solved after noise suppression; S4. Based on the objective function obtained in step S3, use a denoising dilated convolutional neural network to address the noise suppression problem. Solve the problem; S5. Solve the closed-form subproblem based on step S4. ; S6. Based on the results obtained in steps S4 and S5, the pseudo-solution mixed shot point gather is... and aliasing matrix The data is fed into a cascaded deep neural network for end-to-end training. After obtaining the parameters of the cascaded deep neural network, the pseudo-demixed co-detector gathers to be demixed are used as the input to the cascaded network to obtain the demixed seismic data.

2. The seismic data unmixing method based on cascaded deep neural network priors according to claim 1, characterized in that, In step S1, seismic data is acquired with the detector fixed and without considering noise during the acquisition process. Specifically: in, This is an aliasing matrix that includes time jitter; The data is the effective seismic data to be separated.

3. The seismic data unmixing method based on cascaded deep neural network priors according to claim 1, characterized in that, In step S2, the valid signal for: in, For non-negative Lagrange multipliers, the first term The first item is for data fidelity, and the second item is for regularization.

4. The seismic data unmixing method based on cascaded deep neural network priors according to claim 1, characterized in that, Step S4 is as follows: In training the denoising dilated convolutional neural network, a common shot gather is used as an approximation of the common receiver gather to train the network. A parameter selection method is employed to construct the aliasing matrix, where the aliasing degree is 1 and the random time jitter range is within 200 time sampling points. In a single aliasing acquisition, the time difference between the excitation of two seismic sources is greater than or equal to the signal length. The signals generated by the two sources do not alias in time, and the aliasing matrix... ,get ,and Combined, we obtain To resolve the pseudo-unmixing results of aliased seismic data, the pseudo-unmixed common shot point gathers are... As input, a clean aliased common shot point collection is used as output to train the prior network end-to-end. Then, the pseudo-decoded mixed detector gather containing aliasing noise is fed into the prior network to obtain clean data after suppressing aliasing noise.

5. The seismic data unmixing method based on cascaded deep neural network priors according to claim 1, characterized in that, In step S5, the closed-form solution of the objective function is specifically as follows: in, Indicates the first k In the next iteration, denoiser(·) represents the denoising module. It is an aliased matrix. Non-negative penalty weights The prior valid signal components to be solved after noise suppression. For clean target data, For aliased seismic data, These are parameters used to control the step size of gradient descent.

6. The seismic data unmixing method based on cascaded deep neural network priors according to claim 1, characterized in that, In step S6, the aliasing matrix It does not change with network iterations.

7. A seismic data unmixing system based on cascaded deep neural network priors, characterized in that, include: The data module employs a random time jitter strategy to construct the aliasing matrix. Through the effective seismic data to be separated Multiplying in the common-detector domain yields aliased seismic data. ; The regularization module will solve the problem. The problem is transformed into solving a regularized least squares optimization problem by introducing a regularization term. Obtain a valid signal ; The function module utilizes the HQS algorithm concept by introducing intermediate auxiliary variables. The nonlinear estimation expression obtained from the regularization module is transformed into a new objective function, which is: in, Non-negative penalty weights For non-negative Lagrange multipliers, the first term For data fidelity items, For regularization terms, For the introduction of auxiliary variable terms; In the new objective function, the data fidelity term remains quadratic, the auxiliary variables are decoupled, and after further transformation, the constrained optimization problem is converted into two subproblems that are solved alternately and iteratively; The solution process is equivalent to a typical noise suppression problem, i.e., fixed... For clean target data The effective signal components to be solved; for Solving this problem is a closed-form problem; by designing a new fusion network framework, the prior is implicitly learned by using the network structure composed of two iterative subproblems as input instead of the regularization operator. The constrained optimization problem is transformed into two subproblems that are solved iteratively and alternately as follows: in, Indicates the first k In the next iteration, denoiser(·) represents the denoising module. It is an aliased matrix. For the pseudo-solution of the mixed communist cannon point collection, for, Non-negative penalty weights The prior valid signal components to be solved after noise suppression; The solution module, based on the objective function obtained from the function module, utilizes a denoising dilated convolutional neural network to address the noise suppression problem. Solve the problem; The computation module solves closed-form subproblems based on the solution module. ; The output module, based on the results obtained from the solution module and the calculation module, outputs the pseudo-solution mixed shot point gather. and aliasing matrix The data is fed into a cascaded deep neural network for end-to-end training. After obtaining the parameters of the cascaded deep neural network, the pseudo-demixed co-detector gathers to be demixed are used as the input to the cascaded network to obtain the demixed seismic data.

Images