VSP upper and lower wave field separation method based on low-rank regularization deep learning

By combining deep learning and iterative low-rank regularization method with singular value decomposition, the problem of uplink and downlink wave separation difficulties in VSP wavefield separation is solved, and high-quality separation effect and better geological explanatory are achieved, suitable for complex Geological environment.

CN119001854APending Publication Date: 2024-11-22UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411113670.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The prior art is difficult to effectively separate uplink and downlink waves in VSP wave field separation, especially in complex geological environments. Traditional methods such as low-rank approximation, sparse representation and spatial domain filtering have aliasing phenomenon, and deep learning methods lack explanation. and requires a lot of labeled data.

Method used

Using an iterative method based on low-rank regularization deep learning, combining deep learning and singular value decomposition, VSP wavefields are separated by iterative low-rank regularization deep learning model, and using deep convolutional autoencoder and iterative alternating optimization algorithm to generate High-quality uplink and downlink wave components.

Benefits of technology

It improves the accuracy and interpretability of wavefield separation, provides a new perspective on seismic data processing under complex geological conditions, significantly improves the separation effect, reduces the dependence of parameter adjustment, and improves the signal-to-noise ratio and local similarity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119001854A_ABST
    Figure CN119001854A_ABST
Patent Text Reader

Abstract

The invention discloses a VSP upper and lower wave field separation method based on low-rank regularization deep learning, is applied to the field of seismic data processing, and aims to solve the problems that existing upper and lower wave separation methods in vertical seismic section data driven by various models depend on pre-assumption, and various data driven methods need a large number of accurately separated signals as training samples, so that the processing efficiency is low. And the signals are often difficult to obtain. The invention provides an iterative low-rank regularization deep learning framework, an uplink wave is modeled by adopting low-rank hypothesis and solved by using singular value decomposition, and a downlink wave is directly modeled by utilizing strong data driving capability of deep learning. And the two types of waves are protected, so that the complete separation effect is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of seismic data processing, and in particular relates to a wave field separation technology. Background Art

[0002] VSP has become a key technology in subsurface exploration, providing finer resolution and depth information than surface seismic methods. In mild geological environments, within the close to medium offset range, the propagation wavefield and the reflection wavefield mainly exist in the downgoing wavefield and the upgoing wavefield, respectively. A key step in VSP data analysis is to effectively separate the wavefield into upgoing and downgoing components, a process that can significantly enhance the interpretation of the subsurface structure. The complex superposition and interaction of different waves in the VSP wavefield make separation challenging. In order to achieve the separation of upgoing and downgoing waves, it is usually necessary to accurately model one or both of the waves.

[0003] The methods for distinguishing upgoing waves from downgoing waves can be roughly divided into model-driven and data-driven methods. Common representatives of model-driven methods can be roughly divided into the following types: Low-rank Approximation (LRA), Sparse Representation (SR), and Spatial-domain Filtering (SDF). LRA assumes that the wavefield data of each event flattened into a regular wave is a low-rank matrix. A typical example is Singular Value Decomposition (SVD). However, this method of modeling upgoing waves or downgoing waves separately cannot protect the other wave well, limiting the effect of direct wavefield separation. SR methods, such as frequency-wavenumber filters (FK) and Radon transforms, model upgoing waves and downgoing waves through sparse representation in the transform domain. Although both upgoing waves and downgoing waves are modeled and protected by these methods, aliasing still exists due to the limitation of spatial sampling rate, which limits the effect of the separation process. In SDF, after aligning the desired signal, the filter treats the undesired signal as random noise, thereby reducing the undesired signal. The median filter method is a typical SDF method that can maintain the amplitude when separating VSP wavefield signals. But its effect is limited by time accuracy, waveform consistency, etc. However, manual experiments and adjustments of various parameters in these methods are crucial to the effectiveness of these methods.

[0004] In recent years, due to the remarkable success of deep learning (DL) in seismic data processing and interpretation, deep learning-based methods have been applied to wavefield separation of VSP data. The powerful feature representation ability and data adaptability of DL networks enable them to adopt an end-to-end training strategy to directly model VSP wavefield data. In the current situation, many deep learning networks are used for wavefield separation, such as Generative Adversarial Networks (GANs), Deep Neural Networks (DNNs), Regularized Residual Channel Attention Network (RRCAN), U-Network, etc. These methods mainly take VSP data as input, use downlink and uplink wavefields as labels, and directly obtain separation results end-to-end. However, this process lacks interpretability because it is purely data-driven and requires a large amount of labeled training data. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a VSP upper and lower wave field separation method based on Iterative Low-rank Regularization Deep Learning (ILRDL).

[0006] The technical solution adopted by the present invention is: a VSP upper and lower wave field separation method based on low-rank regularized deep learning, comprising:

[0007] S1. Obtain observation signal;

[0008] S2. Model the wave field separation problem as an iterative low-rank regularized deep learning model. By solving the iterative low-rank regularized deep learning model, the upgoing waves and downgoing waves can be separated from the observed signals.

[0009] Beneficial effects of the invention: The present invention proposes a model data-driven ILRDL method for upgoing and downgoing wavefield separation of VSP data, with the aim of obtaining high-quality separation results. In the model, the robust feature representation capabilities of DCAE and low-rank approximation are used to model the downgoing and upgoing wavefields, respectively. A new hybrid model-data-driven framework is established by combining the traditional SVD model with the deep learning model. In addition, an alternating minimization strategy is developed for the ILRDL method, which enables the generation of high-quality solutions. Experiments conducted on artificial and real data verify the effectiveness of the method. This method not only improves the accuracy of wavefield separation, but also provides a new perspective for seismic data processing under complex geological structure conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a comparison diagram of the SVD results of the original upgoing wave and the upgoing wave after τ transformation;

[0011] Figure 2 It is a schematic diagram of the ILRDL method of the present invention;

[0012] Figure 3 This is an example of artificial three-dimensional data;

[0013] Wherein, (a) is the VSP original data, (b) is the upgoing wave separation result obtained by the method of the present invention, (c) is the downgoing wave separation result obtained by the method of the present invention, (d) is the local similarity graph between the up and down waves constructed by the method of the present invention, (e) is the upgoing wave separation result obtained by the SVD method, (f) is the downgoing wave separation result obtained by the SVD method, (g) is the local similarity graph between the up and down waves constructed by the SVD method, (h) is the upgoing wave separation result obtained by median filtering, (i) is the downgoing wave separation result obtained by median filtering, (j) is the local similarity graph between the up and down waves constructed by median filtering, (k) is the upgoing wave separation result obtained by fk filtering, (l) is the downgoing wave separation result obtained by fk filtering, and (m) is the local similarity graph between the up and down waves constructed by fk filtering;

[0014] Figure 4 Spectral comparison of the results of separation for synthetic data;

[0015] Among them, (a) is the original synthetic VSP data, (b) is the upgoing wave field separated by the method of the present invention, (c) is the downgoing wave field separated by the method of the present invention, (d) is the upgoing wave field separated by the SVD method, (e) is the downgoing wave field separated by the SVD method, (f) is the upgoing wave field separated by the median filter, (g) is the downgoing wave field separated by the median filter, (h) is the upgoing wave field separated by the fk filter, and (i) is the downgoing wave field separated by the fk filter;

[0016] Figure 5 This is a single trace comparison diagram of the upgoing wave of the synthetic data;

[0017] Figure 6 This is an example of real data;

[0018] Wherein, (a) is the VSP original data, (b) is the upgoing wave separation result obtained by the method of the present invention, (c) is the downgoing wave separation result obtained by the method of the present invention, (d) is the local similarity graph between the up and down waves constructed by the method of the present invention, (e) is the upgoing wave separation result obtained by the SVD method, (f) is the downgoing wave separation result obtained by the SVD method, (g) is the local similarity graph between the up and down waves constructed by the SVD method, (h) is the upgoing wave separation result obtained by median filtering, (i) is the downgoing wave separation result obtained by median filtering, (j) is the local similarity graph between the up and down waves constructed by median filtering, (k) is the upgoing wave separation result obtained by fk filtering, (l) is the downgoing wave separation result obtained by fk filtering, and (m) is the local similarity graph between the up and down waves constructed by fk filtering;

[0019] Figure 7 Spectral comparison of the real data separation results;

[0020] Among them, (a) is the original actual VSP data, (b) is the upgoing wave field separated by the method of the present invention, (c) is the downgoing wave field separated by the method of the present invention, (d) is the upgoing wave field separated by the SVD method, (e) is the downgoing wave field separated by the SVD method, (f) is the upgoing wave field separated by the median filter, (g) is the downgoing wave field separated by the median filter, (h) is the upgoing wave field separated by the fk filter, and (i) is the downgoing wave field separated by the fk filter. DETAILED DESCRIPTION

[0021] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.

[0022] The present invention provides a VSP upper and lower wave field separation method based on low-rank regularized deep learning, and its implementation process includes the following contents:

[0023] 1. Modeling of 3D seismic outlier noise suppression

[0024] In the case of VSP recording wave field separation, the present invention assumes that the observed signal It is an upgoing wave Downward Wave and complex mixed noise The superposition of n 1 is the number of time samples, n 2 is the trace number. The recorded signal is defined as

[0025] Y=U+D+N (1)

[0026] The purpose of VSP wave field separation is to obtain the noise image observed in equation (1) Separate the unknown upgoing wave from and down wave Its specific purpose can be expressed as follows:

[0027]

[0028] in, represents the F-norm;

[0029] In fact, the characteristics of formula (2) are undetermined, which means that there is no unique solution. According to the basic principle of linear equation analysis, in order to determine the solution, it is necessary to add additional constraints in U and D. The present invention can summarize these additional restrictions in the following way:

[0030]

[0031] The next challenge is to accurately model U and D and choose appropriate additional constraints. and ψ(·), which is a non-trivial task.

[0032] 2. Iterative Low-Rank Regularization Deep Learning Model

[0033] To solve the aforementioned problems, the present invention adopts a hybrid approach that combines deep learning with singular value decomposition methods to iteratively separate upgoing and downgoing waves. This strategy consists of three different parts.

[0034] (1) Upward wave modeling

[0035] Upgoing waves have obvious directional characteristics in wave field form, but the wave field data of these regular waves usually do not have low-order characteristics. A special transformation is added to the signal to highlight the correlation of U, such as Figure 1 As shown, it can be well solved by low-rank approximation and expressed as:

[0036]

[0037] Among them, among them||·|| * is the matrix nuclear norm and τ° is the affine transformation that aligns the upgoing waves horizontally.

[0038] (2) Downward Wave Modeling

[0039] The process of separating the up and down waves in VSP data can be regarded as a point-to-point prediction process, similar to the denoising problem. Based on the success of unsupervised denoising using deep learning, the present invention selects a deep learning method to model the down wave data as a valuable signal. It can be expressed as follows:

[0040] ψ(D)=h θ (YU) (5)

[0041] h θ (·) is a mapping function developed through deep learning that connects the target data with the observed data. Essentially, this constitutes an unsupervised regression process that produces an estimate of the downlink wave.

[0042] (3) Completely separated model

[0043] According to the above discussion and equations (3), (4), and (5), the present invention can easily derive the following equation:

[0044]

[0045] Where γ is the control ||τ°U|| * is an important parameter for the sparsity level in . At the same time, h θ (·) represents a deep learning (DL) network that reduces the reconstruction error by using the backpropagation (BP) technique. The above equation is a combination of two different models, with the characteristics of data-driven and model-driven methods.

[0046] In addition, the Deep Convolutional Auto-Encoder (DCAE) is proficient in attenuating unwanted signals through an unsupervised mechanism, and the autoencoder can be used to obtain D from the input signal. Here, the present invention uses the DCAE model to complete the task, and this process can be expressed as:

[0047] h Θ (YU)=D Θ (E Θ (YU)) (7)

[0048] Where E Θ (·) represents the encoder structure, D Θ (·) represents the decoder structure, by which the upstream wave is separated from the downstream wave as much as possible.

[0049] Combining the information discussed above, the present invention can easily derive the following expression

[0050]

[0051] In the formula, F represents the downlink wave after filtering out a part of the uplink wave from the original signal, that is, F = YU. By reducing the system average error, ILRDL continuously optimizes the following unconstrained objective function:

[0052]

[0053] 3. Objective function solution

[0054] Considering the non-convex nature of the ILRDL objective function in (9), it is difficult to achieve global optimization of the objective. Therefore, the present invention introduces an iterative alternating optimization algorithm to deal with this problem, such as Figure 2 As shown. The algorithm can find a set of local optimal parameters to obtain a high-quality solution that meets the requirements. The present invention uses an alternating minimization method to decompose the problem into two simpler sub-problems related to Θ and U. This method consistently reduces the value of the objective function in each iteration, which is conducive to convergence to the local minimum. Table 1 summarizes the proposed ILRDL model and all its steps.

[0055] (1) Update D

[0056] When U is fixed, D can be updated as follows

[0057]

[0058] In (10), the BP algorithm is used to refine the weight parameters in the DL network to ensure the desired performance. Subsequently, this improved model is applied to ensure high-quality downlink wave components, and the expression is as follows:

[0059] D (l) =D Θ (E Θ (F (l) )) (11)

[0060] Once the parameter Θ is established, the present invention can deduce D from (11) (l) The solution of (l) By F (l) Update, F (l) =F (l-1) -U (l-1) is the input of the DL network. It should be noted that F (0) is the original observed signal Y.

[0061] (2) Update U

[0062] Then, the present invention calculates U by implementing the following method. When Θ is fixed, a new optimization problem is obtained from formula (9):

[0063]

[0064] Among them, λ is a parameter in the process of solving U;

[0065] Equation (12) corresponds to the sparse optimization problem of the matrix. Its purpose is to solve the undetermined linear equation (4). The solution method is as follows:

[0066] [Q,S,V]=SVD(τ°(YD (l) )) (13)

[0067] Where Q and V represent the orthogonal matrices after SVD, and S represents the singular value matrix. Here, a soft threshold operation is used to ensure high-quality continuous wave components:

[0068]

[0069] in is the matrix transpose, · represents matrix multiplication, and shrink(·) is a function representing the threshold operation. Specifically, the function shrink(·) takes the following form:

[0070]

[0071] where θ is the input of the function shrink(·), represents the threshold parameter, and max(·,·) takes the maximum value of the two parameters. After applying the threshold, an inverse transformation needs to be performed

[0072]

[0073] where τ -1 This process enables the present invention to obtain the separated downlink wave signal.

[0074] Table 1 Proposed ILRDL algorithm flow

[0075]

[0076] The architecture of the DCAE network plays a key role in the feature representation and overall performance of the ILRDL model. Referring to the traditional DCAE configuration, the network consists of two parts: the encoder and the decoder, which learn the mapping relationship from the original data to the feature space and the mapping relationship from the feature space to the target data respectively. The network has 11 layers, using a 3×3 kernel convolution layer and a rectified linear activation function, and a 2×2 maximum pooling layer. For clarity, the network details are summarized in Table 2.

[0077] In order to obtain enough unsupervised training samples, the present invention designs a strategy to split the original VSP data into overlapping and non-overlapping blocks of the same size, so as to effectively collect a sufficient number of samples for training. In the actual application of the model, the slice size of the real data is 400×400 dimensions, while the slice size of the synthetic data is 200×200 dimensions.

[0078] Table 2 Network parameters of DCAE model in ILEDL

[0079]

[0080] In order to evaluate the performance of the ILRDL model and the comparison algorithm of the present invention on two data sets (real data and synthetic data), the present invention introduces the concept of local similarity. Local similarity can measure whether the upgoing and downgoing waves of VSP data are well separated. Specifically, the value range of local similarity is 0 to 1. Smaller values ​​are separated almost completely, that is, there is no signal entanglement between the separated upgoing and downgoing waves. In addition, another commonly used indicator for evaluating the signal strength compared to Gaussian noise in a specific seismic image is the signal-to-noise ratio (SNR), which is defined as follows:

[0081]

[0082] where Y and Represent the true value data and separation results respectively.

[0083] The comparison algorithm used in this embodiment is described as follows:

[0084] In the context of VSP wavefield separation, various techniques are used to separate upgoing and downgoing waves, thereby improving the quality and interpretability of seismic data. For comparison, the present invention selected three commonly used reference methods with published proven results.

[0085] FK filtering is a widely used method in seismic data processing that separates wavefields based on their velocity. By converting the data into the frequency-wavenumber domain, FK filtering allows the identification and removal of unwanted wavefield components based on their different propagation characteristics.

[0086] Median filtering is a nonlinear digital filtering technique that is particularly useful in removing short-term spikes or anomalies without significantly distorting the signal. In VSP wavefield separation, median filtering can be employed to reduce random noise, thereby helping to isolate the main signal from the others. This approach is beneficial because it preserves the edges of the waveform while removing outliers.

[0087] SVD is a mathematical technique that decomposes a matrix into three other matrices and is useful in signal processing and noise reduction. In VSP wavefield separation, SVD can distinguish between signal and noise by analyzing the singular values ​​that represent the energy of the data. Singular value decomposition can significantly improve the signal-to-noise ratio of seismic data by retaining high singular value (signal) components and discarding low singular value (noise) components.

[0088] When implementing the ILRDL method, the present invention needs to set the parameters of the model and the numerical algorithm. There are two types of parameters, one is the parameters of the U subproblem, and the other is the parameters of the DCAE training algorithm θ.

[0089] (1) U subproblem parameters:

[0090] In this sub-problem, the key is to control the threshold parameter value in equation (12) to obtain a higher quality uplink wave. The present invention mainly focuses on the ratio of γ to λ, which is set to γ / λ=2 in artificial synthetic data and γ / λ=5 in real data. In addition, the threshold parameter in the shrinkage function is also very important. It is worth mentioning that the energy of the signal will change during the iteration process, so it is invalid to use a hard threshold here.

[0091] (2) Parameters of the Θ subproblem:

[0092] The DCAE part of the ILRDL model is trained using an adaptive moment estimation optimizer with a learning rate set to 1×10 -5 The number of training epochs for both synthetic VSP data and real data is 100, and the batch size is set to 360.

[0093] To verify the performance of the proposed ILRDL model, experiments are carried out on simulated VSP data.

[0094] The artificial data is provided by OpenVSPdata3. The data consists of multiple high-energy upgoing and downgoing waves and a small number of low-energy waves. The comprehensive experimental results are shown in Figure 3 As shown, the middle two columns intuitively show the comparison of the two separation results of the proposed ILRDL model and the comparison algorithm. The results show that the method of the present invention provides a relatively complete distinction between upgoing and downgoing waves compared to the higher residual signals in other methods. Specifically, the basic SVD technique is ineffective in distinguishing upgoing and downgoing waves. Similarly, median filtering results in obvious energy retention in stronger signals, and some axes are lost. The FK filtering method produces truncation artifacts in both upgoing and downgoing waves, which in turn impairs the quality of separation.

[0095] Through Figure 3 This is further verified by comparing the local similarity metrics in (d), (g), (j) and (m). Figure 3 In (g), the lower right quadrant of the downlink wave processed by the SVD method is shown as two obvious yellow lines, indicating poor separation performance. The median filter shows a wide yellow area across the main downlink and uplink components. Due to the presence of truncation artifacts and residual signals, some areas of the FK filtering method appear bright yellow. In addition, the spectral comparison of the separation results is shown in Figure 2. Figure 4 As shown, it further illustrates the effectiveness of the separation results of the method of the present invention.

[0096] The comparison of the signal-to-noise ratio of the upgoing and downgoing waves and the mean and variance of the local similarity are shown in Table 3, which quantitatively demonstrates the effectiveness of the proposed method. Figure 5The single trace waveform of the upgoing wave in the separation results of different methods. The comparison shows that the separation results of this method are most consistent with the ground truth. As expected, the above comparison shows that the proposed ILRDL method produces better results than the baseline method.

[0097] Table 3 Comparison of synthetic data performance based on signal-to-noise ratio and local similarity indicators

[0098]

[0099] In order to further verify the effectiveness of this method, the present invention applies the ILRDL method to complex VSP real seismic data.

[0100] like Figure 6 As shown in (a), real earthquakes show obvious interweaving of upgoing and downgoing waves, and the presence of Gaussian noise poses a considerable challenge to the wavefield separation process.

[0101] Figure 6 It is the separation result based on the entire real experimental data. Due to the complex algorithm and innovative method, the ILRDL method of the present invention has achieved a good separation effect. In order to better demonstrate the separation performance obtained by different methods, the present invention provides a local similarity measure for the separation results of each method. Similar to the results of the synthetic data experiment, the results of the ILRDL method are mainly blue areas (indicating effective separation), which are in sharp contrast to the sharp yellow spots or continuous yellow blocks that are prevalent in the results of other methods. In addition, the method of the present invention effectively eliminates the stripe noise and Gaussian noise in the signal during the separation process, further improving the image quality. Although the FK filtering method also presents a large area of ​​blue, it cannot remove stripes and Gaussian noise, resulting in artifacts.

[0102] Figure 7 The spectral comparison diagram of the separation results. It is difficult for the singular value decomposition method to distinguish the upgoing wave from the downgoing wave. The downgoing wave of the median filter still has a lot of energy in the upgoing wave. The FK filter method has a clear truncation boundary and retains almost all the Gaussian noise. The results discussed above lead the present invention to conclude that the ILRDL method has more advantages in VSP signal separation than these baseline methods.

[0103] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, the present invention may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A VSP upper and lower wave field separation method based on low-rank regularized deep learning, characterized in that: include: S1. Obtain observation signal; S2. Model the wave field separation problem as an iterative low-rank regularized deep learning model. By solving the iterative low-rank regularized deep learning model, the upgoing waves and the downgoing waves can be separated from the observed signals.

2. According to the VSP upper and lower wave field separation method based on low-rank regularized deep learning according to claim 1, it is characterized in that: The observation signal of step S1 is defined as: Y=U+D+N Where Y is the observed signal, D is a downward wave, U is the upgoing wave, N is complex mixed noise, represents a set of real numbers, n1 is the number of time samples, and n2 is the number of traces.

3. The VSP upper and lower wave field separation method based on low-rank regularized deep learning according to claim 2 is characterized in that: The wavefield separation problem is modeled as an iterative low-rank regularized deep learning model. The specific process is as follows: The purpose of VSP wave field separation is to separate the unknown upgoing wave U and downgoing wave D from the observed Y in Y = U + D + N. The purpose of VSP wave field separation is expressed as: in, represents the F-norm; By adding additional constraints to U and D, It is expressed as: The upgoing wave is modeled as: Among them, among them||·|| * is the matrix nuclear norm, τ is the affine transformation that aligns the upgoing waves horizontally; The descending wave is modeled as: ψ(D)=h θ (Y-U) h θ (·) is the mapping function developed by deep learning; According to the modeled upgoing and downgoing waves, we get: stY=D+U Where γ is the control ||τ°U|| * An important parameter for the sparsity level in ; h θ (·) The DCAE model is specifically adopted and expressed as: h Θ (Y-U)=D Θ (E Θ (Y-U)) Among them, E Θ (·) represents the encoder structure, D Θ (·) indicates the decoder structure; Thus It is expressed as: stY=D+U stY-FU=0 In the formula, F represents the downlink wave after filtering out a part of the uplink wave from the original signal; By reducing the average system error, ILRDL continuously optimizes the following unconstrained objective function: Among them, λ is a parameter.

4. The VSP upper and lower wave field separation method based on low-rank regularized deep learning according to claim 3 is characterized in that: h θ (·) Reduce reconstruction error by using back-propagation techniques to optimize.

5. The VSP upper and lower wave field separation method based on low-rank regularized deep learning according to claim 4 is characterized in that: Using the alternating minimization method, Decompose into two sub-problems related to Θ and U; specifically: (1) Update D When U is fixed, D is updated as follows According to formula (10), the downlink wave component is obtained as follows: D (l) =D Θ (E Θ (F (l) )) (11) Among them, F (l) =F (l-1) -U (l-1) is the input of the DL network, F (0) is the original observed signal Y. (2) Update U When Θ is fixed, Transformed into a new optimization problem: Solving equation (12), we get: [Q,S,V]=SVD(τ°(YD (l) )) (13) Where Q and V represent the orthogonal matrices after SVD, and S represents the singular value matrix; Use a soft thresholding operation to ensure high quality of continuous wave components: in is the matrix transpose, · represents matrix multiplication, shrink(·) is a function representing the threshold operation; By performing an inverse transform, the separated downlink signal is obtained: where τ -1 is the inverse affine transformation.

6. The VSP upper and lower wave field separation method based on low-rank regularized deep learning according to claim 5 is characterized in that: The function shrink(·) takes the following form: where θ is the input of the function shrink(·), represents the threshold parameter, and max(·,·) means taking the maximum value of the two parameters.