A neural network-based wave signal coupling noise suppression method

Abstract

The application discloses a wave signal coupling noise suppression method based on a neural network, is a self-encoder wave signal coupling noise suppression method based on adjacent position constraint and first arrival front sample guidance, and is applied to the field of wave signal data processing. In view of the problem that the prior art is prone to causing great damage to effective signals when suppressing coupling noise, the application designs a feature based on first arrival front noise sample weight representing coupling noise and the waveform similarity of adjacent spatial positions of wave signals, realizes separation of effective signals and strong coupling noise in an unsupervised mode under a robust principal component analysis framework based on a model and data joint driving strategy, and reduces damage to the effective signals.

Description

Technical Field

[0001] This invention relates to the field of wave signal data processing technology, and in particular to a method for suppressing coupled noise in wave signals based on neural networks. Background Technology

[0002] With the increasing demand for high-precision wavefield imaging and efficient data processing, various wave signal denoising methods and technologies are constantly being updated and iterated. High-quality and high-efficiency noise suppression methods have become the research focus in the field of wave signal processing.

[0003] Coupled noise is characterized by its wide distribution bandwidth, high overlap between its main frequency band and the effective signal spectrum, high energy intensity, and slow attenuation of strong amplitude. When coupled noise overlaps with the effective signal, it forms a complex interference waveform. This strong coupling noise causes the effective wave field energy to be masked, which not only significantly reduces the signal-to-noise ratio of the original data, but also interferes with the accuracy of subsequent key processing steps such as wave field separation and anisotropic inversion. It has become a technical bottleneck restricting the application of wave signal technology in complex scenarios.

[0004] Coupled noise, as a type of coherent noise, is often distinguished from the effective signal by its characteristics in the transform domain. These differences mainly manifest in frequency distribution, apparent velocity characteristics, and feature space structure. Based on this, a series of filtering and denoising methods have been proposed, including bandpass filtering, frequency-wavenumber domain filtering, and singular value decomposition filtering. These methods have achieved certain coupling noise suppression effects in practical wave signal processing, but they still have limitations and shortcomings such as low fidelity, high risk of effective signal damage, and low denoising efficiency. Therefore, it is necessary to study high-fidelity coupling noise suppression methods suitable for the characteristics of wave signals to effectively improve signal processing quality. Summary of the Invention

[0005] To address the problem of strong coupling noise interference in wave signal data processing, this invention proposes a wave signal coupling noise suppression method based on neural networks. By using a model-data hybrid driving method, unsupervised coupling noise suppression is achieved, effectively reducing manual intervention and improving processing efficiency.

[0006] The technical solution adopted in this invention is as follows:

[0007] A method for suppressing coupled noise in wave signals based on neural networks, comprising the following steps:

[0008] Step 1: Input the observation data of the wave signal to be coupled noise suppressed. Pick up observation data By calculating the time position corresponding to the first arrival wave of the wave signal at various spatial locations, and then calculating the noise weights at each spatial location before the arrival of the first arrival wave, a noise weight matrix is ​​constructed. ;

[0009] Step 2, randomly initialize strong coupling noise And set the convergence conditions for the separation iteration;

[0010] Step 3, from the observation data Remove current strong coupling noise To obtain the denoised data of the current iteration. And input it into the autoencoder under the current model parameters to obtain the clean reconstruction data of the current iteration. The model parameters of the autoencoder are updated based on reconstruction loss and waveform similarity loss.

[0011] Step 4, denoise the data from the current iteration. The autoencoder with updated input model parameters is used to obtain the denoised data for the next iteration based on its output. ;

[0012] And based on the denoised data of the current iteration Strongly coupled noise The update is performed to obtain the strongly coupled noise for the next iteration. ;

[0013] Step 5, based on observation data The latest noise reduction data and strong coupling noise Determine if the convergence condition for the separation iteration has been met. If not, proceed based on the current denoised data. and strong coupling noise Continue with steps 3 and 4; if so, then proceed based on the currently obtained clean reconstruction data. Output observation data The results of coupling noise suppression.

[0014] Furthermore, in step 1, the noise weight matrix Each matrix element is used to characterize the observed data. Weight of the k-th spatial location signal Its expression is ,in, This represents the arrival time of the first wave at the k-th spatial location. This represents the Frobenius norm.

[0015] Furthermore, when updating the autoencoder model parameters based on reconstruction loss and waveform similarity loss, the objective function is set as follows: ,in, Indicates the losses incurred during reconstruction. Indicates waveform similarity loss. The weights representing waveform similarity loss This represents the model parameters of the convolutional autoencoder.

[0016] Furthermore, the autoencoder employs a convolutional autoencoder.

[0017] Furthermore, the model structure of the convolutional autoencoder includes an encoding part and a decoding part. The encoding part includes, in sequence, a first stacked network and a second stacked network, wherein the stacking units of the first stacked network include, in sequence, a convolutional layer, an activation function, and a max pooling layer; the stacking units of the second stacked network include, in sequence, a convolutional layer and an activation function; the decoding part includes, in sequence, a third stacked network, a convolutional layer, and a linear layer, wherein the stacking units of the third stacked network include, in sequence, a transposed convolutional layer and an activation function.

[0018] Furthermore, reconstruction losses Specifically:

[0019]

[0020] in, , These represent the outputs of the encoding and decoding parts of the convolutional autoencoder, respectively. This represents the model parameters of the convolutional autoencoder.

[0021] Furthermore, waveform similarity loss Specifically:

[0022]

[0023] in, and Denoising data The number of rows and columns, , These represent the denoised data. The data element in the i-th row and j-th column, and the data element in the i-th row and (j+1)-th column. , These represent the denoised data. The mean of the j-th column and the (j+1)-th column.

[0024] Furthermore, in step 4, the strongly coupled noise is... When performing the update, the optimization objective function is set as follows:

[0025]

[0026] in, Represents the balance parameters. This represents the element-wise multiplication operation, also known as the Hadaman product.

[0027] Furthermore, in step 4, a soft threshold shrinkage operator is used. Strongly coupled noise Update:

[0028]

[0029] in, , These represent the strongly coupled noise before and after the update, respectively. This represents an element-wise multiplication operation;

[0030] Soft threshold shrinkage operator expression for:

[0031]

[0032] Among them, threshold , Indicates strong coupling noise The column vectors in the corresponding two-dimensional matrix.

[0033] Furthermore, the convergence condition for the separation iteration is set as follows:

[0034]

[0035] in, , This represents the latest denoised data and strongly coupled noise. This indicates the preset error limit.

[0036] Furthermore, in step 2, strongly coupled noise The initial value is set to all zeros.

[0037] The technical solution provided by this invention brings at least the following beneficial effects:

[0038] The proposed method, within the framework of robust principal component analysis, constructs a model for separating the effective signal from strongly coupled noise by introducing signal waveform similarity constraints at adjacent spatial locations. A convolutional autoencoder is used to model the effective signal, and an alternating optimization algorithm iteratively solves for the optimal solution. This method starts with the column sparsity characteristics of the initial-to-initial coupled noise and the waveform similarity characteristics of the fluctuating signal at adjacent spatial locations. It employs a separable sparsity penalty to separate strongly coupled noise and guides the convolutional autoencoder to gradually learn the features of the effective signal in an unsupervised manner, ultimately separating the effective signal from the strongly coupled noise. This method integrates a dual-driven mechanism of model and data, improving the suppression of coupled noise while ensuring the quality of the effective signal, increasing processing efficiency, and saving labor costs. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart illustrating a method for suppressing coupled noise in a wave signal based on a neural network, provided in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the structure of the convolutional autoencoder used in an embodiment of the present invention;

[0042] Figure 3 This is the original noisy schematic diagram;

[0043] Figure 4 This is a denoised result image from an embodiment of the present invention;

[0044] Figure 5 This is a noise diagram suppressed according to an embodiment of the present invention. Detailed Implementation

[0045] 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 described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using 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 application, but merely represents selected embodiments of the present invention.

[0046] This invention provides a neural network-based method for suppressing coupled noise in wave signals. It is a method for suppressing coupled noise in wave signals using an autoencoder guided by neighbor constraints and first-arrival samples. Addressing the problem that existing technologies often cause significant damage to the effective signal when suppressing coupled noise, this invention designs a feature based on the weights of first-arrival noise samples to characterize the waveform similarity between adjacent spatial positions of coupled noise and the wave signal. Within a robust principal component analysis framework, using first-arrival noise samples, the weighted column sparse weights of pure noise are obtained to characterize the relative strength relationship between different spatial positions of strongly coupled noise. Based on the sparsity of the weighted column of first-arrival noise and the waveform similarity features of adjacent positions in the wave signal, a convolutional autoencoder is constrained. An alternating minimization algorithm iteratively separates the wave signal and strongly coupled noise in an unsupervised manner. The innovation of this invention lies in using the waveform similarity features of adjacent positions in the wave signal to model the effective signal. Based on a model-data joint driving strategy, it achieves the separation of the effective signal and strongly coupled noise in an unsupervised manner, reducing damage to the effective signal. This invention is applicable to wave signal data collected under different geological conditions, and can suppress strongly coupled noise in different wave signal data, thereby significantly improving the signal-to-noise ratio of the effective signal. While improving data quality, this invention also significantly improves the quality and efficiency of suppressing coupled noise in wave signal data, providing technical support for high-precision imaging of complex geological structures, and offering a feasible solution for widespread application in actual work areas.

[0047] In one embodiment, the processing steps of the method proposed in this invention mainly include: (1) calculating the initial arrival noise weight matrix (constructing the noise weight matrix) (2) Establish a separation model for separating the signal and noise; (3) Iteratively solve the model to obtain an effective fluctuating signal, thereby suppressing the coupled noise in the fluctuating signal. In constructing the noise weight matrix... The process includes picking up the initial arrival position of the wave signal and calculating the noise weights in the signal records at each spatial location before the arrival of the initial wave, in order to obtain the noise weight matrix. When establishing a model to separate signal and noise, under the framework of robust principal component analysis, a separation model for separating effective signal and strongly coupled noise is constructed by constraining the similarity of signal waveforms at adjacent positions, and a convolutional autoencoder is used for partial solution. The separation model used in this embodiment is shown in formula (1):

[0048] (1)

[0049] in, This indicates strong coupling noise in a wave signal. This represents the model parameters of the convolutional autoencoder. , These represent the outputs of the encoding and decoding parts of the convolutional autoencoder, respectively. , These represent the encoder and decoder in a convolutional autoencoder, respectively. This represents the denoised data. Indicates the losses incurred during reconstruction. This represents the input observation data. , For the two balance parameters set, This represents element-wise multiplication. This represents the Frobenius norm.

[0050] Since the separation model shown in formula (1) is a non-convex function model equation, in this embodiment of the invention, it is divided into two parts, and an iterative alternating optimization algorithm is used to solve it using an unsupervised convolutional autoencoder to obtain an effective fluctuation signal.

[0051] In one embodiment, such as Figure 1 As shown in the figure, the method for suppressing coupled noise of wave signals based on neural networks provided in this embodiment of the invention includes the following steps:

[0052] Step 1, Input observation data Pick up observation data By calculating the time position corresponding to the first arrival wave of the wave signal at various spatial locations, and then calculating the noise weights at each spatial location before the arrival of the first arrival wave, a noise weight matrix is ​​constructed. ;

[0053] Step 2, randomly initialize strong coupling noise And set the convergence conditions for the separation iteration;

[0054] Step 3, from the observation data Remove current strong coupling noise To obtain the denoised data of the current iteration. And input it into the autoencoder under the current model parameters to obtain the clean reconstruction data of the current iteration. ;

[0055] The model parameters of the autoencoder are updated based on reconstruction loss and waveform similarity loss;

[0056] Step 4, denoise the data from the current iteration. The autoencoder with updated input model parameters is used to obtain the denoised data for the next iteration based on its output. ;

[0057] And based on the denoised data of the current iteration Strongly coupled noise The update is performed to obtain the strongly coupled noise for the next iteration. ;

[0058] Step 5, based on observation data The latest noise reduction data and strong coupling noise Determine if the convergence condition for the separation iteration has been met. If not, proceed based on the current denoised data. and strong coupling noise Continue with steps 3 and 4; if so, then proceed based on the currently obtained clean reconstruction data. Output observation data The results of coupling noise suppression.

[0059] In one embodiment, the present invention first picks up the time position corresponding to the first arrival wave of the wave signal at each spatial location in the observed data Y, and calculates the noise weight at each spatial location before the arrival of the first arrival wave. It is assumed that the arrival time of the first arrival wave at the k-th spatial location is t. k Then the k-th spatial location signal Y(1:t) of the observed data Y k The weights of k) for:

[0060] (2)

[0061] After obtaining the weights for all spatial locations, the noise weight matrix W for the entire spatial location can be obtained:

[0062] (3)

[0063] Where M and N are the number of rows and columns of the observed data Y. For ease of calculation, the noise weight matrix W can also be calculated using the minimum weight. Scaling.

[0064] For example, if the observation data Y is a vertical seismic profile VSP signal, then the time position t corresponding to the first arrival wave of the VSP in each seismic trace of data Y is picked. k The noise weights in each seismic trace before the arrival of the first arrival wave are calculated to construct the noise weight matrix W.

[0065] In this embodiment of the invention, the separation model is obtained based on robust principal component analysis and a convolutional autoencoder. The principle of robust principal component analysis is to decompose the data matrix X into a low-rank matrix L and a sparse matrix S, i.e.:

[0066] (4)

[0067] Specifically, maintaining the low rank of a low-rank matrix L requires minimizing the kernel norm of L, i.e., the sum of the singularities of the matrix. Similarly, maintaining the sparsity of a sparse matrix S requires minimizing the 1-norm of S, i.e., the sum of the absolute values ​​of its terms. Therefore, the optimization problem to be solved can be described as follows:

[0068] (5)

[0069] By solving the optimization problem shown in formula (5), the low-rank nature of matrix L and the sparsity of matrix S are simultaneously satisfied. Based on the robust principal component analysis framework, a convolutional autoencoder is used instead of solving the low-rank matrix L. The input of the convolutional autoencoder is the data matrix X, and the reconstructed effective signal is obtained through the encoding and decoding networks of the convolutional autoencoder. Therefore, according to S=X- The residuals represented by matrix S are obtained, which include noise and outliers.

[0070] The residual S contains noise and outliers that are difficult to reconstruct, therefore the optimization problem can be further expressed as:

[0071] (6)

[0072] Wherein, λ is a balance parameter that adjusts the sparsity in S.

[0073] Based on the robust principal component analysis and convolutional autoencoder principles described above, this invention proposes to obtain the reconstructed effective signal by adding a convolutional autoencoder constrained by the waveform similarity of the wave signal between adjacent spatial locations. Meanwhile, the sparsity constraint of the noise column is guaranteed, and the separation model equation is obtained.

[0074] In one embodiment, the specific implementation process of the method proposed in this invention includes:

[0075] (1) Actual data, i.e., observation data It can be regarded as a fluctuation signal With strong coupling noise The linear superposition, that is: Where R represents the real number field, and These represent the number of time samples and the number of spatial samples, respectively, and the effective fluctuation signal. It can also be called denoised data.

[0076] From the observed noisy observation data Unknown fluctuation signal of the lieutenant general With strong coupling noise During separation, it is necessary to... and To ensure the separation effect, additional constraints are added. These constraints can be expressed as follows:

[0077] (7)

[0078] in, For the two balance parameters set, Here are the constraints on D. For the constraints on C, This represents the square of the Frobenius norm.

[0079] 1-a) Modeling of strongly coupled noise:

[0080] Since strongly coupled noise only appears in certain spatial locations, it is considered to have column sparsity. Norms characterize the column sparsity of strongly coupled noise:

[0081] (8)

[0082] (9)

[0083] in, It is strongly coupled noise. The line, number Column elements, and Strong coupling noise The number of rows and columns.

[0084] The obtained noise weight matrix Combining these equations into equation (9), we obtain the strongly coupled noise model equation:

[0085] (10)

[0086] in, This represents the Hadaman product, which is an element-wise multiplication operation.

[0087] 1-b) Wave signal modeling:

[0088] The wave signal obtained using a convolutional autoencoder can be represented as:

[0089] (11)

[0090] in, This represents clean data constructed using a convolutional autoencoder, i.e., reconstructed clean data.

[0091] Due to noise reduction data Substituting equations (10) and (11) into equation (7), the constraint clause can be expressed as:

[0092] (12)

[0093] For noise-reducing data Although the wave signals received by adjacent detectors are delayed in time, the wave signal waveforms have a certain similarity. In order to preserve the physical characteristic of the similarity between the waveforms of adjacent detectors, this embodiment of the invention adds a waveform similarity loss. for:

[0094] (13)

[0095] in, and Denoising data The number of rows and columns, , These represent the denoised data. The data element in the i-th row and j-th column, and the data element in the i-th row and (j+1)-th column. , These represent the denoised data. The mean of the j-th column and the (j+1)-th column.

[0096] Equation (12) can then be rewritten as:

[0097] (14)

[0098] (2) Solve equation (14) to obtain the coupling noise suppression result based on the solution result. However, the solution of this equation belongs to the solution of a non-convex optimization problem. In this embodiment of the invention, it is decomposed into two parts, and then the iterative alternating minimization algorithm is used to find the local optimal solution of each part alternately.

[0099] The first part of equation (14) is an optimization problem of the strongly coupled noise C:

[0100] (15)

[0101] It can be shrunk by a soft threshold operator Equation (15) yields the updated strongly coupled noise, which can be expressed as follows: The specific update process can be represented as follows:

[0102] (16)

[0103] (17)

[0104] in, , Represent the strongly coupled noise before and after the update, and the threshold, respectively. c is a column vector of strongly coupled noise. During computation, the soft-threshold shrinkage operator... The input parameters are .

[0105] The second part of equation (14) is the optimization problem of the denoised data D:

[0106] (18)

[0107] By training the convolutional autoencoder, equation (18) can be used as the loss function for backpropagation to obtain the network parameters of the trained model. And after completing the solution, the new denoised data is obtained, which can be represented as:

[0108] (19)

[0109] in, , These represent the denoised data before and after the update, respectively. In this embodiment of the invention, the data is first analyzed from the observation data. Remove current strong coupling noise We obtain the denoised data for the current iteration, denoted as... ,in Indicates the number of iterations; denoise the data. Input a convolutional autoencoder, based on its output Backpropagation is performed using the loss function shown in Equation (18) to update the model parameters of the convolutional autoencoder. Then... Feed the updated convolutional autoencoder to obtain the current clean reconstruction data. Simultaneously, the current reconstructed clean data will be used. As denoised data for the next iteration .

[0110] In this embodiment of the invention, after obtaining the denoised data At that time, it can be based on The updated strongly coupled noise is obtained. To further improve the accuracy of the processing, in this embodiment of the invention, when updating the strongly coupled noise based on the soft threshold shrinkage operator, in order to obtain the updated value of the strongly coupled noise in the next iteration... can As The value is then obtained according to formula (16). The value of, i.e. .

[0111] In one embodiment, the iteration termination condition (iteration convergence condition) of the iterative alternating minimization algorithm is set as follows:

[0112] (20)

[0113] in, It is an artificially set error limit. This indicates the separation error.

[0114] In one embodiment, the model structure of the convolutional autoencoder used in this embodiment includes an encoder and a decoder; wherein, the encoder sequentially includes: a first stacked network and a second stacked network; the stacking units of the first stacked network sequentially include a convolutional layer, an activation function, and a max-pooling layer; the stacking units of the second stacked network sequentially include a convolutional layer and an activation function; the decoder sequentially includes: a third stacked network, a convolutional layer, and a linear layer, wherein the stacking units of the third stacked network sequentially include a transposed convolutional layer and an activation function. Figure 2 As shown, in this embodiment, the first stacked network has three stacking units, and the convolutional kernels of its convolutional layers are all 3×3, with channel numbers of 32, 64, and 128 respectively; the pooling kernels of the max pooling layers are all 2×2. The second stacked network has two stacking units, and the convolutional kernels of the convolutional layers in this stacked network are all 3×3, with channel numbers of 256 respectively. The third stacked network has three stacking units, and the convolutional kernels of the transposed convolutional layers are all 3×3, with channel numbers of 128, 64, and 32 respectively; the convolutional kernel of the convolutional layer after the third stacked network is 3×3, with channel number 1.

[0115] In one embodiment, the specific steps for performing iterative solution based on the set separation model and iteration termination condition include:

[0116] Step S1: Input the observed data Y and set the equilibrium parameters. and and setting error limits ;

[0117] Step S2: Pick the initial arrival position of the observation data Y and calculate the noise weight matrix based on the initial arrival. ;

[0118] Step S3, Initialize strong coupling noise (It can be randomly initialized, or it can be initialized as a matrix of all zeros), initialize the model parameters of the convolutional autoencoder. and the number of initial iterations ;

[0119] Step S4, Update ;

[0120] Step S5, will As the input to the convolutional autoencoder, its output is used to update the model parameters of the convolutional autoencoder according to formula (18), resulting in... According to the loss function shown in formula (18), the model parameters are updated by gradient descent, and the updated model parameters are denoted as... ;

[0121] Step S6: [The sentence is incomplete and requires more context to be translated accurately.] Input the updated convolutional autoencoder with its parameters, and assign its output to the corresponding value. Rebuild clean data ,Right now ;

[0122] Step S7, according to formula (16) ,Right now ;

[0123] Step S8: Determine whether convergence has occurred according to formula (20). When the separation error satisfies... When the iteration ends, the currently obtained reconstructed clean data is displayed. The effective fluctuation signal obtained from the solution is output; otherwise, the iteration number is set to... Increment by 1 and continue executing steps S4-S7.

[0124] The method proposed in this invention was experimentally verified by combining specific wave signal data with vertical seismic profile signal data. The input observation data is as follows: Figure 3 As shown, the method proposed in the embodiments of the present invention is used to... Figure 3 The denoising results of the observation data in the example are as follows Figure 4 As shown, the suppressed strong coupling noise is as follows Figure 5 As shown in the figure, it can be seen that the method proposed in this embodiment of the invention can remove strong coupling noise while effectively preserving the valid signal.

[0125] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0126] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include at least one of those features.

[0127] Any process or method description described in this specification can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0128] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0129] Note that, unless otherwise explicitly stated, all features disclosed in this specification can be replaced by alternative features for achieving the same, equivalent, or similar purpose. Therefore, unless explicitly stated otherwise, each disclosed feature is merely one example of a set of equivalent or similar features. Where used, "further," "preferably," "even further," and "more preferably" are simply starting points for describing another embodiment based on the foregoing embodiments, the combination of which with the foregoing embodiments constitutes the complete configuration of another embodiment. Any combination of several "further," "preferably," "even further," or "more preferably" settings following the same embodiment constitutes yet another embodiment.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0131] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for suppressing coupled noise in wave signals based on neural networks, characterized in that, Includes the following steps: Step 1: Input the observation data of the wave signal to be coupled noise suppressed. Pick up observation data By calculating the time position corresponding to the first arrival wave of the wave signal at various spatial locations, and then calculating the noise weights at each spatial location before the arrival of the first arrival wave, a noise weight matrix is ​​constructed. ; Step 2, randomly initialize strong coupling noise And set the convergence conditions for the separation iteration; Step 3, from the observation data Remove current strong coupling noise To obtain the denoised data of the current iteration. And input it into the autoencoder under the current model parameters to obtain the clean reconstruction data of the current iteration. The model parameters of the autoencoder are updated based on reconstruction loss and waveform similarity loss. Step 4, denoise the data from the current iteration. The autoencoder with updated input model parameters is used to obtain the denoised data for the next iteration based on its output. ; And based on the denoised data of the current iteration and noise weight matrix Strongly coupled noise The update is performed to obtain the strongly coupled noise for the next iteration. ; Step 5, based on observation data The latest noise reduction data and strong coupling noise Determine if the convergence condition for the separation iteration has been met. If not, proceed based on the current denoised data. and strong coupling noise Continue with steps 3 and 4; if so, then proceed based on the currently obtained clean reconstruction data. Output observation data The results of coupling noise suppression; When updating the model parameters of the autoencoder based on reconstruction loss and waveform similarity loss, the training objective function is set as follows: ,in, Indicates the losses incurred during reconstruction. Indicates waveform similarity loss. The weights representing waveform similarity loss Represents the model parameters of the convolutional autoencoder; Reconstruction losses Specifically: ； in, , These represent the outputs of the encoding and decoding parts of the convolutional autoencoder, respectively. This represents the model parameters of the convolutional autoencoder; Waveform similarity loss Specifically: ； in, and Denoising data The number of rows and columns, , These represent the denoised data. The data element in the i-th row and j-th column, and the data element in the i-th row and (j+1)-th column. , These represent the denoised data. The mean of the j-th column and the (j+1)-th column.

2. The method for suppressing coupled noise of wave signals based on neural networks as described in claim 1, characterized in that, In step 1, the noise weight matrix Each matrix element is used to characterize the observed data. Weight of the k-th spatial location signal Its expression is ,in, This represents the arrival time of the first wave at the k-th spatial location. This represents the Frobenius norm.

3. The method for suppressing coupled noise of wave signals based on neural networks as described in claim 1, characterized in that, The autoencoder uses a convolutional autoencoder.

4. The method for suppressing coupled noise of wave signals based on neural networks as described in claim 3, characterized in that, The model structure of a convolutional autoencoder includes an encoding part and a decoding part. The encoding part consists of a first stacked network and a second stacked network, wherein the stacking units of the first stacked network consist of a convolutional layer, an activation function, and a max pooling layer, respectively; the stacking units of the second stacked network consist of a convolutional layer and an activation function, respectively. The decoding part consists of a third stacked network, a convolutional layer, and a linear layer, wherein the stacking units of the third stacked network consist of a transposed convolutional layer and an activation function, respectively.

5. The method for suppressing coupled noise of wave signals based on neural networks as described in claim 1, characterized in that, In step 4, strong coupling noise is addressed. When performing the update, the optimization objective function is set as follows: ； in, Represents the balance parameters. This indicates an element-wise multiplication operation.

6. The method for suppressing coupled noise of wave signals based on neural networks as described in claim 1, characterized in that, In step 4, a soft threshold shrinkage operator is used. Strongly coupled noise Update: ； in, , These represent the strongly coupled noise before and after the update, respectively. This represents an element-wise multiplication operation; Soft threshold shrinkage operator expression for: ； Among them, threshold , Indicates strong coupling noise The column vectors in the corresponding two-dimensional matrix.

7. The method for suppressing coupled noise of wave signals based on neural networks as described in claim 1, characterized in that, The convergence condition for the separation iteration is set as follows: ； in, , This represents the latest denoised data and strongly coupled noise. This indicates the preset error limit.

Images