Earthquake noise suppression method and system based on reversible decoupling deep learning

Abstract

The invention provides a seismic noise suppression method and system based on reversible decoupling deep learning. The method comprises the steps of decomposing seismic data with noise into high-frequency details and low-frequency structural components through wavelet transform; high-frequency details and low-frequency structural components are processed through reversible blocks, noise and low-resolution effective signal information are obtained, and each reversible block is composed of an affine coupling layer and a reversible convolution layer; after seismic data is successfully decoupled, a noise component is discarded, a new variable is randomly sampled from Laplacian distribution to recover an effective signal part in the noise component, and the effective signal part is recombined with a low-resolution structural component; and reversely inputting the combined components into the network, and reconstructing to obtain noiseless seismic data. According to the invention, all effective signals are kept, data can be restored in a lossless manner, and excessive suppression of the signals is avoided; the correlation between seismic data channels is kept; more effective information decoupling is achieved, and noise in seismic data is suppressed under the condition that effective signals are kept.

Description

Technical Field

[0001] This invention relates to the fields of signal pattern recognition and machine learning, and in particular to a method and system for suppressing seismic noise based on reversible decoupled deep learning. Background Technology

[0002] Seismic exploration is an important branch of geophysics, primarily involving the analysis of the propagation patterns of artificially generated seismic waves in different strata to obtain information about subsurface geological structures. Seismic data serves as a crucial foundation for geophysical exploration, and its quality directly impacts oil and gas resource exploration, seismic activity monitoring, and geological hazard risk assessment. However, seismic data is often affected by various factors, including environmental noise, equipment malfunctions, and complex terrain, resulting in complex and nonlinear noise components. This noise significantly affects data quality, thus influencing the assessment of subsurface structures and, consequently, major decisions regarding oil and gas exploration and disaster early warning.

[0003] Signal and noise in seismic data are often coupled, with the effective signal reflecting subsurface geological structure and rock physical properties, while noise is the interference introduced during acquisition. Effectively decoupling random noise from the effective signal in seismic data and recovering the effective signal from the seismic data is crucial for seismic data processing. Although traditional denoising methods, such as filtering, smoothing, and wavelet denoising, can suppress random noise to some extent, the complex distribution characteristics of noise in seismic data, especially its non-Gaussian distribution, often lead to excessive suppression of the effective signal. Therefore, how to effectively remove noise while preserving the effective signal to the greatest extent has always been a research hotspot in the field of seismic data processing.

[0004] In recent years, deep learning-based denoising methods have received widespread attention. Deep learning, with its powerful nonlinear feature extraction capabilities, can adaptively learn the differences between noise and signal from noisy data in a data-driven manner, demonstrating excellent performance in suppressing random noise. However, deep learning methods typically rely on assumptions about the noise distribution (such as prior noise assumptions). If the actual noise distribution does not conform to the assumptions, the denoising effect may be affected. Furthermore, the complexity of deep learning models and the large number of parameters can increase computational costs and lead to overfitting problems when the amount of data is limited.

[0005] Traditional denoising methods often oversuppress valid signals, while deep learning-based methods rely on assumptions about noise distribution. If the actual noise distribution deviates from these assumptions, the denoising effect may be affected. Existing deep learning models may suffer information loss due to the irreversibility of their network structure. In other words, while traditional denoising methods can suppress random noise to some extent, the complex distribution characteristics of noise in seismic data, especially non-Gaussian noise, often make it difficult to distinguish noise from valid signals, leading to excessive suppression of valid signals during denoising. Although deep learning models can learn the distribution characteristics of noise and signals, they still rely on assumptions about the noise distribution. If the actual distribution deviates from these assumptions, the denoising effect may be affected.

[0006] Invertible Neural Networks (INNs), as an emerging deep learning architecture, have demonstrated enormous potential. Through the reversibility of their modules, INNs ensure that data transformations at each layer are lossless, allowing for complete data reconstruction through inverse operations in both forward and backward propagation. This avoids the information loss caused by the irreversibility of traditional Convolutional Neural Networks (CNNs). INNs possess strong feature representation capabilities and efficient computational characteristics, making them particularly suitable for processing large-scale datasets. In the field of seismic data denoising, the reversibility and lossless nature of INNs provide a new solution for the efficient processing of seismic data. Based on this, several INN-based seismic data denoising and super-resolution algorithms have been proposed. Summary of the Invention

[0007] To address the problems in the existing technologies mentioned above, this application proposes a seismic noise suppression method and system based on reversible decoupling deep learning. This method utilizes a novel Invertible Disentangling Network (IDN). Unlike traditional reversible neural network models, where the latent variables follow a single distribution, IDN has two latent variables following two different distributions. This innovative design can better decouple the effective signal and noise signal in seismic data, thus improving the denoising effect.

[0008] The present invention provides a seismic noise suppression method based on reversible decoupling deep learning, which, in one embodiment, includes the following steps:

[0009] Step S1: Decompose the noisy seismic data into high-frequency detail components and low-frequency structure components using wavelet transform.

[0010] Step S2: Reversible blocks are used to process high-frequency detail components and low-frequency structural components to obtain noise information and low-resolution effective signal information. Each reversible block consists of an affine coupling layer and a reversible 1×1 convolutional layer. Each affine coupling layer consists of four reversible operations. The reversible 1×1 convolution realizes cross-channel information interaction.

[0011] Step S3: After successfully decoupling the seismic data into noise information and low-resolution effective signal information, discard the noise component, and randomly sample a new variable from the Laplace distribution to recover the effective signal part in the noise component, and recombine it with the low-resolution structural component.

[0012] Step S4: The combined components are inversely input into the network for processing to reconstruct noise-free seismic data; let it be a reversible transformation g, and the objective function of the forward process be...

[0013]

[0014] Where h(y) LR The low-resolution information learned by the reversibly decoupled network corresponds to the first channel of the latent representation obtained during the forward pass; M is the data size; ||·|2 is the 2-norm, x LR Low-resolution information for the effective signal;

[0015] The combined components are then inversely input into the network for processing to reconstruct noise-free seismic data; the objective function of the inverse process is:

[0016]

[0017] Where x represents noise-free data, and g -1 (g(y) LR ,z HF g(y) represents low-resolution information. LR and latent variable z HF The output after the reverse process is the recovered noise-free data, where N is the data size and ||·|1 is the 1-norm; the invertible transformation g is trained by minimizing the two objective functions of the forward and backward processes;

[0018] During data training, the overall loss function includes at least four parts: the loss function of the forward process, the reconstruction loss function of the backward process, the perceptual loss, and the spectral consistency loss. The weights of each part of the loss function are determined through training.

[0019] In one implementation, step S3 specifically involves the following steps:

[0020] Step S31: Let the original noise data be y, its corresponding noise-free data be x, and the noise in the data be n; then we have p(y) = p(x,n) = p(x)p(n|x); during network training, the scale of the input noise data is reduced, and the number of channels is increased; after training the previous channel to represent a low-resolution effective signal, high-frequency information and noise will be encoded in the remaining channels, as follows:

[0021] p(y)=p(x LR ,x HF ,n)=p(x LR )p(x HF ,n|x LR (2-3) Where, x HF High-frequency information of the effective signal;

[0022] Step S32: After successfully decoupling the seismic data into noise information and low-resolution effective signal information, discard the noise information portion of all potential representations.

[0023] Step S33: Randomly sample a new variable from the Laplace distribution, recover the effective signal portion of the noise component, and recombine it with the low-resolution structural component.

[0024] In one implementation, if the data is synthetic, the overall loss function is divided into four parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, and the spectral consistency loss. If the data is real, it also includes a domain adversarial transfer network to achieve feature alignment between the source and target domains, enabling the network to generalize in the target domain as well. The overall loss function is divided into five parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, the spectral consistency loss, and a domain loss function.

[0025] An earthquake noise suppression system based on reversible decoupling deep learning, according to one embodiment of the present invention, includes,

[0026] The wavelet transform module decomposes noisy seismic data into high-frequency detail components and low-frequency structure components using wavelet transform.

[0027] The reversible block module processes high-frequency detail components and low-frequency structural components to obtain noise information and low-resolution effective signal information. Each reversible block consists of an affine coupling layer and a reversible 1×1 convolutional layer. Each affine coupling layer consists of four reversible operations. The reversible 1×1 convolution enables cross-channel information interaction.

[0028] The recombination module successfully decouples the seismic data into noise information and low-resolution effective signal information. It discards the noise component and randomly samples a new variable from the Laplace distribution to recover the effective signal part in the noise component, and then recombines it with the low-resolution structural component.

[0029] The reconstruction module inversely inputs the combined components into the network for processing, reconstructing noise-free seismic data; let it be a reversible transformation g, and the objective function of the forward process be...

[0030]

[0031] Where g(y) LR The low-resolution information learned by the reversibly decoupled network corresponds to the first channel of the latent representation obtained during the forward pass; M is the data size; ||·|2 is the 2-norm, x LR Low-resolution information for the effective signal;

[0032] The combined components are then inversely input into the network for processing to reconstruct noise-free seismic data; the objective function of the inverse process is:

[0033]

[0034] Where x represents noise-free data, and g -1 (g(y) LR ,z HF g(y) represents low-resolution information. LR and latent variable z HF The output after the reverse process is the recovered noise-free data, where N is the data size and ||·|1 is the 1-norm; the invertible transformation g is trained by minimizing the two objective functions of the forward and backward processes;

[0035] During data training, the overall loss function includes at least four parts: the loss function of the forward process, the reconstruction loss function of the backward process, the perceptual loss, and the spectral consistency loss. The weights of each part of the loss function are determined through training.

[0036] In one implementation, the reassembly module includes:

[0037] Let the original noisy data be y, and its corresponding noise-free data be x, with n representing the noise in the data; then we have p(y) = p(x,n) = p(x)p(n|x); during network training, the scale of the input noise data is reduced, and the number of channels increases; after training the previous channel to represent a low-resolution effective signal, the high-frequency information and noise of the effective signal will be encoded in the remaining channels, as follows:

[0038] p(y)=p(x LR ,xHF ,n)=p(x LR )p(x HF ,n|x LR (2-6) Where, x HF High-frequency information of the effective signal;

[0039] After successfully decoupling seismic data into noise information and low-resolution effective signal information, the noise-related parts of all potential representations are discarded.

[0040] A new variable is randomly sampled from the Laplace distribution, the effective signal portion of the noise component is recovered, and it is recombined with the low-resolution structural component.

[0041] In one implementation, if the data is synthetic, the overall loss function is divided into four parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, and the spectral consistency loss. If the data is real, it also includes a domain adversarial transfer network to achieve feature alignment between the source and target domains, enabling the network to generalize in the target domain as well. The overall loss function is divided into five parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, the spectral consistency loss, and a domain loss function.

[0042] The above-mentioned technical features can be combined in various suitable ways or replaced by equivalent technical features, as long as the purpose of the present invention can be achieved.

[0043] The present invention provides a seismic noise suppression method and system based on reversible decoupling deep learning, which has at least the following advantages compared with the prior art:

[0044] This invention utilizes a reversible neural network to process seismic data. Its reversibility ensures that all valid signals are preserved during forward and backward propagation, and data can be restored without loss, thus avoiding excessive signal suppression. The introduction of a self-attention mechanism allows the model to focus on relevant parts of the input, maintaining the correlation between seismic data traces. Reversible 1×1 convolution, as a channel interaction mechanism, achieves more effective decoupling of noise and valid information. Compared with traditional methods, this invention can effectively decouple valid and noise signals in seismic data, suppressing noise in seismic data while preserving valid signals. Attached Figure Description

[0045] The invention will now be described in more detail with reference to embodiments and the accompanying drawings.

[0046] Figure 1 A flowchart of the method of the present invention is shown;

[0047] Figure 2The synthetic seismic data of this invention is shown;

[0048] Figure 3 The noise data of the synthesized seismic data of this invention is shown;

[0049] Figure 4 The denoising results of the synthesized seismic data of this invention are shown;

[0050] Figure 5 The actual seismic data of this invention is displayed;

[0051] Figure 6 The noise data of the actual seismic data of this invention is displayed;

[0052] Figure 7 The denoising results of actual seismic data presented in this invention are shown. Detailed Implementation

[0053] The invention will now be further described with reference to the accompanying drawings.

[0054] This invention provides a method and system for suppressing seismic noise based on reversible decoupling deep learning.

[0055] In one embodiment, this invention provides a seismic noise suppression method based on reversible decoupled deep learning. This invention utilizes the adaptive learning capability of reversible neural networks to remove complex noise without requiring strong prior assumptions about noise. Furthermore, this invention leverages the lossless nature of reversible neural networks to ensure the complete preservation of key geological information during the denoising process, improving the accuracy of subsequent geological interpretation. This method provides an efficient and high-fidelity denoising approach for seismic data processing, offering strong technical support for fields such as oil and gas exploration and geological disaster early warning.

[0056] In one embodiment, the reversible neural network consists of multiple reversible blocks, which are affine coupling layers generalized from the RealNVP model. In NICE, for the forward mapping, the reversible block converts the input feature u... i Divided into two parts by channel: and Then comes an additive affine transformation, denoted as:

[0057]

[0058] Where φ is an arbitrary function. Its inverse process is expressed as:

[0059]

[0060] To enhance transformation capabilities, RealNVP improves upon NICE. For forward mapping, the invertible block transforms the input features u... i Divided into two parts by channel: and Then φ1 and φ2 will and Convert to And φ3 and φ4 will and Convert to The process is as follows:

[0061]

[0062]

[0063] In this code, `split(·)` splits the data by channel, `concat(·)` merges the data by channel, `×` multiplies the data element-wise, and `φ1`, `φ2`, `φ3`, and `φ4` are arbitrary functions or neural networks. Backmaps are performed by reversing this process.

[0064]

[0065] Where / represents element-wise division.

[0066] The NICE and RealNVP methods disrupt the coupling order through interleaving and randomization. The Glow model, which follows the RealNVP framework, introduces invertible 1×1 convolutions to replace the sorting layers. It uses invertible 1×1 convolutions to disrupt the coupling order and achieves better results. The feature extraction methods used in invertible networks mainly include compression and wavelet transform. Compression methods arrange and reorganize all pixels within the feature map with minimal information loss. Wavelet transform uses row and column transformations to decompose the input into a low-frequency component LL, which can be viewed as a thumbnail of the original image, and three high-frequency components LH, HL, and HH, corresponding to the horizontal, vertical, and diagonal directions, respectively.

[0067] In one embodiment, self-attention is an attention mechanism that associates different positions within a single sequence to compute a representation of the same sequence. The purpose of self-attention is to capture the correlations between different positions in a sequence, enabling the model to focus on important information regardless of its position in the input sequence. Self-attention dynamically captures the dependencies between elements at different positions in the sequence and generates new sequence representations based on these dependencies. It allows the model to focus on relevant parts of the input sequence, thus more effectively capturing contextual information. The core idea of ​​self-attention is to dynamically adjust the representation of each element by generating attention weights through computational relationships between each element and other elements in the sequence, thereby enhancing the extraction of important features. Self-attention is implemented through the following steps:

[0068] 1. Generate Query, Key, and Value matrices. For input data X, first generate the query, key, and value respectively through three different convolutional layers:

[0069] Q = XW Q K = XW K V = XW V (3-13)

[0070] Among them, W Q W K W V It is the weight matrix obtained through learning.

[0071] 2. Calculate the attention score by calculating the dot product of the query and the key to obtain the attention weight:

[0072] Score(Q,K)=QK T (3-14)

[0073]

[0074] Where d k It is the dimension of the key vector. To avoid the dot product result being too large and causing gradient problems, it will be scaled.

[0075] 3. Normalization:

[0076]

[0077] To convert attention scores into weights, a Softmax function is used for normalization. The Softmax function ensures that the sum of all output weights is 1, allowing the model to learn the importance of each element pair.

[0078] 4. Weighted summation:

[0079]

[0080] Finally, the normalized attention weights are used to perform a weighted summation of the value vector V to generate the final output sequence.

[0081] In one embodiment, perceptual loss and spectral consistency loss are used. Perceptual loss is a commonly used loss function in deep learning-based image style transfer methods. Compared to the traditional Mean Square Error (MSE) loss function, perceptual loss focuses more on the perceived quality of the image, which is more in line with human perception of image quality. Perceptual loss calculates the difference between two images using a pre-trained neural network. Typically, pre-trained convolutional neural networks are used; these networks have been trained on large-scale datasets and can extract high-level features of images. The calculation of perceptual loss usually involves passing the input and target images separately through the pre-trained neural network to obtain their feature representations in the network. These feature representations are then used as input to the loss function to calculate the Euclidean distance or Manhattan distance between them. The formula for calculating perceptual loss is as follows:

[0082]

[0083] Where x is the input image, y is the target image, and F i (·) represents the i-th feature extractor in the pre-trained neural network, where N is the number of feature layers.

[0084] Spectral consistency loss is a loss function introduced in signal processing, image reconstruction, or neural network training to ensure that the spectrum of the reconstructed signal is consistent with that of the original signal. Spectral consistency loss aims to measure the difference between the reconstructed data and the original data in the frequency domain (such as Fourier transform, power spectrum, or wavelet domain), thereby helping the model to accurately preserve frequency features while maintaining its temporal structure and avoiding spectral distortion or loss of high-frequency information in the reconstruction result.

[0085] The formula for calculating the spectrum consistency loss is as follows:

[0086]

[0087] Where x is the original data, To reconstruct the data, F(·) is the Fourier transform, and ||·||1 is the 1-norm.

[0088] In one embodiment, the method flow of the present invention is as follows: Figure 1 As shown, the specific implementation steps of this invention are as follows:

[0089] (1) First, the noisy seismic data is decomposed into high-frequency detail components and low-frequency structure components by wavelet transform.

[0090] (2) The high-frequency detail components and low-frequency structural components are then processed by the reversible blocks to obtain noise information and low-resolution effective signal information. Each reversible block consists of an affine coupling layer and a reversible 1×1 convolutional layer.

[0091] 1. Each affine coupling layer consists of four operations: φ1, φ2, φ3, and φ4. These operations need to be reversible. They extract key features from seismic data, enabling the differentiation between noise and valid signal information.

[0092] 2. Introduce reversible 1×1 convolution to achieve cross-channel information interaction and improve the decoupling accuracy of components.

[0093] (3) After successfully decoupling the seismic data into noise information and low-resolution effective signal information, the noise component is discarded, and a new variable is randomly sampled from the Laplace distribution to recover the effective signal part of the noise component. This is then recombined with the low-resolution structural component. The specific steps are as follows:

[0094] 1. Let the original noisy data be y, its corresponding noise-free data be x, and the noise in the data be n. Then we have: p(y) = p(x,n) = p(x)p(n|x). During network training, the scale of the input noise data is reduced, and the number of channels increases. According to sampling theory, during downsampling, high-frequency information, including noise, is removed, thereby achieving the effect of noise suppression. Since reversible decoupled networks are lossless, after training the previous channel to represent a low-resolution effective signal, the high-frequency information and noise of the effective signal will be encoded in the remaining channels. The above process is as follows:

[0095] p(y)=p(x LR ,x HF ,n)=p(x LR )p(x HF ,n|x LR (3-20)

[0096] Where, x LR For low-resolution information of the effective signal, x HF This refers to the high-frequency information of the effective signal.

[0097] 2. After successfully decoupling the seismic data into noise information and low-resolution effective signal information, it is difficult to decouple the high-frequency information of the effective signal from the noise. In order to completely remove the noise, all parts of the potential representation related to noise information are discarded.

[0098] 3. The high-frequency information of the effective signal is also discarded. In order to recover this part, it is known that this part follows a Laplace distribution. A new variable is randomly sampled from the Laplace distribution.

[0099] (4) The combined components are input into the network in reverse for processing, and noise-free seismic data is reconstructed.

[0100] 1. Let this invertible transformation be g, and the objective function in the forward process be...

[0101]

[0102] Where g(y) LR This represents the low-resolution information learned by the reversibly decoupled network, corresponding to the first channel in the latent representation obtained during the forward pass. M is the data size. ‖·‖2 is the 2-norm.

[0103] 2. The combined components are then input back into the network for processing to reconstruct noise-free seismic data. The objective function of the reverse process is:

[0104]

[0105] Where x represents noise-free data, and g -1 (g(y) LR ,z HF g(y) represents low-resolution information. LR and latent variable z HF The output after the reverse process is the recovered noise-free data, where N is the data size and ||·|1 is the 1-norm. The invertible transformation g is trained by minimizing two objective functions for the forward and backward processes.

[0106] 3. For the training process of synthetic data, the overall loss function is divided into four parts: the loss function of the forward process, the reconstruction loss function of the backward process, the perceptual loss, and the spectral consistency loss. The weights of each part of the loss function are determined to be the most suitable values ​​through training.

[0107] 4. Compared to synthetic data, the application of real-world data adds a domain adversarial transfer network to achieve feature alignment between the source domain (synthetic data) and the target domain (real-world data), giving the network good generalization ability in the target domain as well. The overall loss function is divided into five parts, with an additional domain loss function compared to synthetic data.

[0108] Specifically, a validation set of data was simulated to verify the effectiveness of the present invention in suppressing noise in seismic data. Figure 2 For the synthesized simulation data, Figure 2 The left figure shows the data profile (first data point), with the vertical axis representing time and the horizontal axis representing the number of data points. Figure 2 The right figure shows a pre-stack seismic data gather extracted from a certain point in the middle. The vertical axis represents time, and the horizontal axis represents angle. Figure 3 This is the noise data after adding Gaussian noise. Figure 4 The result is shown after denoising using a reversible network. Therefore, this invention can effectively suppress noise and extract valid signals from synthetic data.

[0109] In one embodiment, the actual data used in this invention is xline*inline*t*a = 500*500*50*10 actual pre-stack seismic data, where xline represents the number of gathers along the connecting line, inline represents the number of gathers along the main line direction, t represents the time window length of the target stratum, and a represents the number of AVO gathers of a single pre-stack seismic data. Figure 5 This is actual earthquake data. Figure 5 The left image shows a slice of the actual data (first data), with the vertical and horizontal axes representing xline and inline, respectively. Figure 5 The right figure shows pre-stack seismic data extracted from a specific point in the middle. The vertical axis represents time, and the horizontal axis represents angle. Since noise is not very noticeable in actual seismic data, Gaussian noise was manually added. Figure 6 This is the noise data after adding Gaussian noise. Figure 7 The image shows the result after denoising using a reversible network. This demonstrates that even with complex real-world data, this invention can effectively suppress seismic data noise and extract valid signals.

[0110] Specifically, this invention combines reversible neural networks, reversible 1×1 convolutions, perceptual loss, and spectral consistency loss. The reversible neural network processes seismic data, ensuring that all valid signals are preserved during forward and backward propagation and enabling lossless data reconstruction, thus avoiding excessive signal suppression. The introduction of a self-attention mechanism allows the model to focus on relevant parts of the input, maintaining the correlation between seismic data traces. The reversible 1×1 convolution, as a channel interaction mechanism, achieves more effective decoupling of noise and valid information. Tests using synthetic and real data demonstrate that, compared to traditional methods, this invention effectively decouples valid and noise signals in seismic data, suppressing noise while preserving valid signals.

[0111] Compared with traditional methods, this invention combines reversible neural networks, reversible 1×1 convolution, perceptual loss, and spectral consistency loss. After decoupling the effective signal and noise signal using a reversible neural network, it completely removes noise by discarding noise-related latent representations and then randomly samples a new variable from a Laplace distribution to recover the high-frequency information of the effective signal, ultimately achieving the goal of seismic data denoising. The network employs two different distributions of latent variables, enabling more efficient decoupling of the effective signal and noise signal in seismic data and improving the denoising effect. Tests using synthetic and real data demonstrate that this invention can effectively decouple the effective signal and noise signal in seismic data, suppressing noise while preserving the effective signal.

[0112] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A seismic noise suppression method based on reversible decoupling deep learning, characterized in that, Includes the following steps, Step S1: Decompose the noisy seismic data into high-frequency detail components and low-frequency structure components using wavelet transform. Step S2: Reversible blocks are used to process high-frequency detail components and low-frequency structural components to obtain noise information and low-resolution effective signal information. Each reversible block consists of an affine coupling layer and a reversible 1×1 convolutional layer. Each affine coupling layer consists of four reversible operations. The reversible 1×1 convolution realizes cross-channel information interaction. Step S3: After successfully decoupling the seismic data into noise information and low-resolution effective signal information, discard the noise component, and randomly sample a new variable from the Laplace distribution to recover the high-frequency effective signal part in the noise component, and recombine it with the low-resolution structural component. Step S4: The combined components are input into the network in reverse for processing, and noise-free seismic data is reconstructed. Let it be an invertible transformation g, and the objective function of the forward process be... Where g(y) LR The low-resolution information learned by the reversibly decoupled network corresponds to the first channel of the latent representation obtained during the forward pass; M is the data size; ||·|2 is the 2-norm, x LR Low-resolution information for the effective signal; The combined components are then inversely input into the network for processing to reconstruct noise-free seismic data; the objective function of the inverse process is: Where x represents noise-free data, and g -1 (g(y) LR ,z HF g(y) represents low-resolution information. LR and latent variable z HF The output after the reverse process is the recovered noise-free data, where N is the data size and ||·|1 is the 1-norm; the invertible transformation g is trained by minimizing the two objective functions of the forward and backward processes; During data training, the overall loss function includes at least four parts: the loss function of the forward process, the reconstruction loss function of the backward process, the perceptual loss, and the spectral consistency loss. The weights of each part of the loss function are determined through training.

2. The seismic noise suppression method based on reversible decoupling deep learning according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: Let the original noise data be y, its corresponding noise-free data be x, and the noise in the data be n; then we have p(y) = p(x,n) = p(x)p(n|x); during network training, the scale of the input noise data is reduced, and the number of channels is increased; after training the previous channel to represent a low-resolution effective signal, the high-frequency information and noise of the effective signal will be encoded in the remaining channels, as follows: p(y)=p(x LR ,x HF ,n)=p(x LR )p(x HF ,n|x Lr (1-3) where x HF High-frequency information of the effective signal; Step S32: After successfully decoupling the seismic data into noise information and low-resolution effective signal information, discard the noise information portion in all potential representations; Step S33: Randomly sample a new variable from the Laplace distribution, recover the effective signal portion of the noise component, and recombine it with the low-resolution structural component.

3. The seismic noise suppression method based on reversible decoupling deep learning according to claim 1, characterized in that, If the data is synthetic, the overall loss function is divided into four parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, and the spectral consistency loss. If the data is real, it also includes a domain adversarial transfer network to achieve feature alignment between the source and target domains, enabling the network to generalize in the target domain as well. The overall loss function is divided into five parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, the spectral consistency loss, and a domain loss function.

4. A seismic noise suppression system based on reversible decoupling deep learning, characterized in that, include, The wavelet transform module decomposes noisy seismic data into high-frequency detail components and low-frequency structure components using wavelet transform. The reversible block module processes high-frequency detail components and low-frequency structural components to obtain noise information and low-resolution effective signal information. Each reversible block consists of an affine coupling layer and a reversible 1×1 convolutional layer. Each affine coupling layer consists of four reversible operations. The reversible 1×1 convolution enables cross-channel information interaction. The recombination module successfully decouples the seismic data into noise information and low-resolution effective signal information. It discards the noise component and randomly samples a new variable from the Laplace distribution to recover the effective signal part in the noise component, and then recombines it with the low-resolution structural component. The reconstruction module inversely inputs the combined components into the network for processing, and reconstructs noise-free seismic data. Let it be an invertible transformation g, and the objective function of the forward process be... Where g(y) LR The low-resolution information learned by the reversibly decoupled network corresponds to the first channel of the latent representation obtained during the forward pass; M is the data size; ||·|2 is the 2-norm, x LR Low-resolution information for the effective signal; The combined components are then inversely input into the network for processing to reconstruct noise-free seismic data; the objective function of the inverse process is: Where x represents noise-free data, and g -1 (g(y) LR ,z HF g(y) represents low-resolution information. LR and latent variable z HF The output after the reverse process is the recovered noise-free data, where N is the data size and ||·|1 is the 1-norm; the invertible transformation g is trained by minimizing the two objective functions of the forward and backward processes; During data training, the overall loss function includes at least four parts: the loss function of the forward process, the reconstruction loss function of the backward process, the perceptual loss, and the spectral consistency loss. The weights of each part of the loss function are determined through training.

5. The seismic noise suppression system based on reversible decoupling deep learning according to claim 4, characterized in that, The reconfiguration module includes: Let the original noisy data be y, and its corresponding noise-free data be x, with n representing the noise in the data; then we have p(y) = p(x,n) = p(x)p(n|x); during network training, the scale of the input noise data is reduced, and the number of channels increases; after training the previous channel to represent a low-resolution effective signal, the high-frequency information and noise of the effective signal will be encoded in the remaining channels, as follows: p(y)=p(x LR ,x HF ,n)=p(x LR )p(x HF ,n|x LR (1-6) where x HF High-frequency information of the effective signal; After successfully decoupling seismic data into noise information and low-resolution effective signal information, the noise-related parts of all potential representations are discarded. A new variable is randomly sampled from the Laplace distribution, the effective signal portion of the noise component is recovered, and it is recombined with the low-resolution structural component.

6. The seismic noise suppression system based on reversible decoupling deep learning according to claim 4, characterized in that, If the data is synthetic, the overall loss function is divided into four parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, and the spectral consistency loss. If the data is real, it also includes a domain adversarial transfer network to achieve feature alignment between the source and target domains, enabling the network to generalize in the target domain as well. The overall loss function is divided into five parts: the forward process loss function, the backward process reconstruction loss function, the perception loss, the spectral consistency loss, and a domain loss function.

Images