Seismic data denoising method and system based on generative adversarial network and storage medium
By combining generative adversarial networks with non-subsampled shear wave transform, the problem of difficulty in preserving high-frequency detail features of seismic data is solved, achieving better denoising effect and bandwidth expansion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2026-03-10
AI Technical Summary
Existing deep learning-based seismic data denoising methods struggle to effectively extract high-frequency detail features from seismic data under low signal-to-noise ratio conditions, resulting in poor denoising performance at the high-frequency end.
A seismic data denoising method based on generative adversarial networks is adopted, which combines transform domain multi-scale generative adversarial networks and non-subsampled shear wave transform. Through iterative learning of the MSFF-GAN generator and discriminator, high-frequency detail features of seismic data are extracted and preserved.
It improves the denoising effect of seismic data, especially the signal-to-noise ratio in the high-frequency band, widens the effective bandwidth, and enhances signal fidelity and denoising capability.
Smart Images
Figure CN121636901A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of seismic data processing, and particularly relates to a seismic data denoising method and system based on a generative adversarial network, a storage medium and equipment. BACKGROUND
[0002] In the field of seismic exploration, effective processing of noisy seismic data to obtain high-precision seismic exploration data can lay a foundation for subsequent interpretation work and maximize the effectiveness and value of seismic data. There are currently a variety of different denoising methods, such as traditional f-x prediction methods and f-k filtering methods for improving the signal-to-noise ratio of seismic data, wavelet transformation methods, curvelet transformation methods, singular value decomposition (SVD) methods, non-local mean-based denoising algorithms, and dictionary learning-based denoising methods. Deep learning is a data-driven multi-layer neural network learning algorithm that can combine deep learning and seismic exploration technology based on machine learning and deep learning methods to carry out research work in the field of seismic exploration.
[0003] Currently, Zhang Panlong et al. designed a U-Net deep neural network model for seismic data denoising, which achieved good denoising results; Mai Hao et al. proposed a deep residual seismic denoising model, which broke through the limitation of the number of network layers of the deep learning algorithm, learned seismic signal features more fully than previous shallow neural networks, and had better noise suppression effect; Fang Wenqian et al. proposed a CNN network based on a double residual structure, which allowed arbitrary combinations between two operators of different layer pairing operations to extract richer features and help improve denoising performance; Yu Kun improved the DnCNN by using a form of hollow convolution in its convolution kernel to increase the receptive field of seismic data, which can efficiently denoise and protect effective signals.
[0004] However, the existing deep learning-based seismic data denoising method has insufficient ability to mine high-frequency detail features of seismic data in low signal-to-noise ratio conditions, and it is difficult to broaden the effective frequency width of seismic data. Therefore, the denoising effect on the high-frequency end of seismic data is poor. SUMMARY
[0005] In view of the deficiencies of the prior art, the present application provides a seismic data denoising method and system based on a generative adversarial network, a storage medium and equipment, which combines non-subsampled shearlet transform and considers denoising as prediction of non-subsampled shearlet transform coefficients to preserve more rich high-frequency detail features based on a transform domain multi-scale generative adversarial network.
[0006] The present application is implemented by the following technical solutions:
[0007] Acquire synthetic data and combine it with actual earthquake data to construct a training set;
[0008] The training data in the training set is subjected to NSST transformation to obtain multiple frequency components;
[0009] The frequency components are input into the MSFF-GAN generator network for feature extraction to obtain denoising transform coefficients.
[0010] Perform NSST inverse transform on the denoising transform coefficients to obtain denoising training data;
[0011] Based on the training data, the denoised training data is identified by the MSFF-GAN discriminator network, and iterative learning and training are performed to obtain the generator denoising model;
[0012] The generator denoising model is used to denoise the seismic data to be processed.
[0013] Optional,
[0014] The acquisition of synthetic data, combined with actual seismic data, and the construction of a training set include:
[0015] By using an enhancement scheme, a preset amount of synthesized data is obtained, including: noisy synthesized data and denoised synthesized data;
[0016] The actual earthquake data includes: noisy earthquake data and denoised earthquake data;
[0017] The noisy synthetic data, denoised synthetic data, actual earthquake data, and denoised earthquake data are integrated to form a training set.
[0018] Optional,
[0019] The step of inputting the frequency components into the MSFF-GAN generator network for feature extraction and obtaining denoising transform coefficients includes:
[0020] The frequency components are input to the shallow feature extraction module (SFE) to obtain shallow features;
[0021] The shallow features are input into the deep feature extraction module (DFE) to extract the deep features.
[0022] Optional,
[0023] The Deep Feature Extraction (DFE) module is cascaded with multiple Multi-Scale Feature Extraction (MSFE) modules. Deep features are extracted based on the MSFE modules, including:
[0024] Feature information of local long paths and short paths is extracted separately, and the feature information of the long paths and short paths is aggregated to obtain deep features.
[0025] Optional,
[0026] The shallow and deep features are fused to obtain the denoising transformation coefficients.
[0027] Optional,
[0028] The step of using the MSFF-GAN discriminator network to identify the denoised training data based on the training data, performing iterative learning and training to obtain a generator denoising model includes:
[0029] The denoised training data and the target earthquake data are input into the MSFF-GAN discriminator network for iterative learning and training.
[0030] The target seismic data includes: denoised synthetic data and denoised seismic data.
[0031] Optional,
[0032] The parameters of the MSFF-GAN generator network are fixed, and the parameters of the MSFF-GAN discriminator network are trained.
[0033] The parameters of the MSFF-GAN discriminator network are fixed, and the parameters of the MSFF-GAN generator network are trained.
[0034] A generator denoising model is obtained through iterative learning.
[0035] The present invention also provides a seismic data denoising system based on generative adversarial networks for implementing the aforementioned method, the system comprising:
[0036] The data synthesis module is used to acquire synthetic data and combine it with actual earthquake data to construct a training set;
[0037] The NSST transformation module is used to perform NSST transformation on the training data in the training set to obtain multiple frequency components.
[0038] The feature extraction module is used to input the frequency components into the MSFF-GAN generator network for feature extraction and to obtain the denoising transform coefficients.
[0039] The NSST inverse transform module is used to perform the NSST inverse transform on the denoising transform coefficients to obtain denoising training data;
[0040] The training module is used to identify the denoised training data through the MSFF-GAN discriminator network based on the training data, perform iterative learning training, and obtain the generator denoising model.
[0041] The denoising module is used to denoise the seismic data to be processed using the generator denoising model.
[0042] Optional,
[0043] The data synthesis module is also configured to:
[0044] By using an enhancement scheme, a preset amount of synthesized data is obtained, including: noisy synthesized data and denoised synthesized data;
[0045] The actual earthquake data includes: noisy earthquake data and denoised earthquake data;
[0046] The noisy synthetic data, denoised synthetic data, actual earthquake data, and denoised earthquake data are integrated to form a training set.
[0047] Optional,
[0048] The feature extraction module is also configured to:
[0049] The frequency components are input to the shallow feature extraction module (SFE) to obtain shallow features;
[0050] The shallow features are input into the deep feature extraction module (DFE) to extract the deep features.
[0051] Optional,
[0052] The system also includes:
[0053] The parameters of the MSFF-GAN generator network are fixed, and the parameters of the MSFF-GAN discriminator network are trained.
[0054] The parameters of the MSFF-GAN discriminator network are fixed, and the parameters of the MSFF-GAN generator network are trained.
[0055] A generator denoising model is obtained through iterative learning.
[0056] The present invention also provides a computer-readable storage medium storing one or more programs, which, when executed, can implement the aforementioned seismic data denoising method based on generative adversarial networks.
[0057] The present invention also provides a device, including a processor, a communication interface, a computer-readable storage medium, and a communication bus; wherein the processor, the communication interface, and the computer-readable storage medium communicate with each other through the communication bus;
[0058] The processor is used to execute programs stored in a computer-readable storage medium.
[0059] Compared with the prior art, the present invention has the following advantages:
[0060] 1. The seismic data denoising method based on generative adversarial networks proposed in this invention integrates NSST into a multi-scale generative adversarial network in the transform domain. The denoising transform coefficients obtained by feature extraction after NSST decomposition are fed into a deep learning network, making it more effective in learning high-frequency information and better preserving high-frequency detail information, thereby improving the denoising effect of seismic data.
[0061] 2. Simultaneously, multi-scale features are integrated with generative adversarial networks. Multi-scale feature extraction blocks are designed in the generator network to extract local features and long and short path features at different scales in seismic data, thereby improving the feature extraction effect.
[0062] 3. By using the adversarial interaction between the generator network and the discriminator network, the fidelity of the extracted signal is ensured. Through iterative learning, the network parameters are gradually optimized to improve the denoising effect of the final denoising model.
[0063] 4. By combining synthetic data with real data, a more complete and effective training database can be constructed, thereby improving the learning effect of deep learning denoising networks.
[0064] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0066] Figure 1 A flowchart illustrating a seismic data denoising method based on generative adversarial networks is shown.
[0067] Figure 2 A schematic block diagram of a seismic data denoising system based on generative adversarial networks is shown.
[0068] Figure 3 A schematic diagram of the network architecture of the seismic data denoising method based on generative adversarial networks according to an embodiment of the present invention is shown.
[0069] Figure 4 This illustration shows a partial example diagram of a training set constructed based on synthetic data and actual seismic data according to an embodiment of the present invention.
[0070] Figure 5 A schematic diagram of the MSFE module structure in the MSFF-GAN generator network according to an embodiment of the present invention is shown.
[0071] Figure 6 A schematic diagram comparing the effects of embodiments of the present invention and conventional methods on synthetic data processing is shown;
[0072] Figure 7 This diagram illustrates a comparison of the denoising results of seismic data under processing using embodiments of the present invention and conventional methods.
[0073] Figure 8 This diagram illustrates a comparison of the high-frequency end of the denoising results of seismic data under processing using an embodiment of the present invention and a conventional method.
[0074] Figure 9 This is a schematic diagram of the structure of a device according to an embodiment of the present invention. Detailed Implementation
[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0076] See appendix Figure 1 The method of the present invention includes:
[0077] S1. Obtain synthetic data and combine it with actual earthquake data to construct a training set;
[0078] Among them, through the enhancement scheme, a preset amount of synthesized data is obtained, and the synthesized data includes: noisy synthesized data and denoised synthesized data;
[0079] The actual earthquake data includes: noisy earthquake data and denoised earthquake data;
[0080] The noisy synthetic data, denoised synthetic data, actual earthquake data, and denoised earthquake data are integrated to form a training set.
[0081] S2. Perform NSST transformation on the training data in the training set to obtain multiple frequency components.
[0082] S3. Input the frequency components into the MSFF-GAN generator network for feature extraction and obtain the denoising transform coefficients;
[0083] The frequency components are input to the shallow feature extraction module (SFE) to obtain shallow features.
[0084] The shallow features are input into the deep feature extraction (DFE) module to extract deep features.
[0085] The deep feature extraction module (DFE) is cascaded with multiple multi-scale feature extraction modules (MSFE). Deep features are extracted based on the MSFE modules, including:
[0086] Feature information of local long paths and short paths is extracted separately, and the feature information of the long paths and short paths is aggregated to obtain deep features.
[0087] In this process, the shallow and deep features are fused to obtain the denoising transformation coefficients.
[0088] S4. Perform NSST inverse transform on the denoising transform coefficients to obtain denoising training data.
[0089] S5. Based on the training data, the denoised training data is identified through the MSFF-GAN discriminator network, and iterative learning and training are performed to obtain the generator denoising model.
[0090] The denoised training data and the target earthquake data are input into the MSFF-GAN discriminator network for iterative learning and training.
[0091] The target seismic data includes: denoised synthetic data and denoised seismic data.
[0092] Specifically, the parameters of the MSFF-GAN generator network are fixed, and the parameters of the MSFF-GAN discriminator network are trained.
[0093] The parameters of the MSFF-GAN discriminator network are fixed, and the parameters of the MSFF-GAN generator network are trained.
[0094] A generator denoising model is obtained through iterative learning.
[0095] S6. The seismic data to be processed is denoised using the generator denoising model.
[0096] Specifically,
[0097] 1. Acquire synthetic data and combine it with actual earthquake data to construct a training set.
[0098] 1. Obtain synthetic data.
[0099] The enhancement scheme obtains a preset amount of synthetic data.
[0100] In this embodiment, a synthetic data enhancement scheme can be used to construct 375 data points from different wavelet main frequencies (15Hz-60Hz) and different co-directional axes. A synthetic tree enhancement scheme can be used to perform enhancement processing such as rotation, flipping, translation, dithering, scaling, and noise perturbation on the 375 data points to construct 2625 data points as synthetic data.
[0101] The synthesized data includes: noisy synthesized data and denoised synthesized data.
[0102] The enhancement scheme obtains completely noise-free denoised synthetic data, and then the noise-containing synthetic data is obtained by adding actual collected noise or Gaussian random noise through noise perturbation processing.
[0103] The amount of composite data can be set according to actual needs.
[0104] 2. Construct the training set.
[0105] A complete and effective training set is constructed based on synthetic data and actual earthquake data.
[0106] The actual earthquake data includes both noisy and denoised earthquake data.
[0107] Denoising earthquake data is obtained by denoising actual earthquake data using commercial software.
[0108] Noisy synthetic data, denoised synthetic data, actual earthquake data, and denoised earthquake data are integrated into training data to construct a training set.
[0109] In subsequent learning and training processes, the training set allows the network to effectively remove noise and retain effective signals with similar characteristics through synthetic data; while the actual earthquake data enables the model to approximate real earthquake signals, ensuring the fidelity of the denoising results.
[0110] II. NSST Transformation.
[0111] The training data Inoise in the training set is decomposed into multiple frequency components D by performing multi-level NSST decomposition. j,l Where j represents the decomposition level, j=1 represents a low-frequency coefficient, reflecting the main characteristics of the signal, and an increase in value represents a higher frequency coefficient; l represents the decomposition direction. NSST decomposition is completed using formula (1):
[0112] D j,l =NSST{I noise}= noise ,ψ j,l,k >, (1)
[0113] Where, ψ j,l,k To form a localized waveform set for the shear wave at different scales j, directions l, and positions k, the expression is given by formula (2):
[0114] ψ j,l,k (x)=|detA| j / 2 ψ(B l A j xk):j,l∈Z,k∈Z 2 (2)
[0115] Where, ψ∈L 2 (R 2 Let L be an integrable space, R be the set of real numbers, and R' ... 2 Let A be a two-dimensional real space; det is the determinant of the matrix; A and B are both 2×2 invertible matrices, A is the scaling matrix, B is the shearing matrix, and detB = 1.
[0116] III. Feature extraction to obtain denoising transformation coefficients.
[0117] Different frequency components obtained through NSST decomposition are input into the MSFF-GAN generator network for feature extraction.
[0118] 1. Shallow feature extraction.
[0119] The different frequency component data S obtained after NSST transformation noise ={D j,l The input is fed into the MSFF-GAN generator network, where shallow features are extracted using the shallow feature extraction (SFE) module. According to formula (3), the shallow feature extraction module SFE uses two convolutional layers to extract shallow features from data of different frequency components S. noise Extracting shallow features:
[0120] M0 = H sFE1 (H SFE2 (S noise (3)
[0121] Among them, H SFE1 and H SFE2 These represent the convolution operations of the two layers in the SFE module; M0 is the extracted shallow features.
[0122] 2. Deep feature extraction.
[0123] The shallow features M0 are fed into the deep feature extraction (DFE) module, which is cascaded with multiple multi-scale feature extraction (MSFE) modules. The MSFE module consists of two parts, used to extract features of local short paths and long paths respectively. Assume the input to the first part is M. d-1 The output is O p1 Extraction is performed layer by layer according to formula (4):
[0124]
[0125] Where the subscript d represents the level of the MSFE module, and are the 1×1, 3×3, 5×5 and 7×7 convolution functions of the first part of the MSFE module, respectively; σ represents the ReLU activation function.
[0126] For the first MSFE module, M d-1 M0 is the output of the shallow feature extraction module (SFE); for the MSFE module cascaded in the middle (d-th layer), M... d-1 This is the output of the MSFE at the next higher level (level d-1); O p1 This is the output of the first part of the MSFE module at level d.
[0127] The first part of the MSFE processing results O p1 The feature map is 128-dimensional. A slicing operation S extracts 64-dimensional features, which are then compared with the input M of the multi-scale feature extraction module. d-1 Connect to the channel dimension to obtain the shortest path information R. Obtain the shortest path information R according to formula (5):
[0128] R = C(S(O) p1 ,64),M d-1 (5)
[0129] Here, C and S represent the join operation and the slice operation, respectively.
[0130] From O p1 The purpose of extracting 64-dimensional features is to combine the current multi-scale information with the previous information, which can be regarded as the retained short path information.
[0131] At the same time, O p1 The remaining 64-dimensional feature map is used as the input to the second part of the MSFE module to extract long path information. The long path information O is obtained according to formula (6). P2 :
[0132]
[0133] in, These are the 1×1, 3×3, 5×5, and 7×7 convolution functions from the second part of the MSFE module.
[0134] The short-path and long-path information are aggregated, and the deep feature M is obtained according to formula (7). d :
[0135] M d =R+O p2 (7)
[0136] 3. Feature fusion to obtain denoising transformation coefficients.
[0137] The output of each cascaded MSFE module is M d The input is concatenated, and the output information is fused through a 1×1 convolutional layer. The fused feature M is obtained according to formula (8). OP :
[0138] M OP =H GPP ([M1,M2,...,M X (8)
[0139] Where M's subscripts 1, 2, ..., C represent different levels of MSFE modules; H GPP It is the fusion function of a 1×1 convolutional layer.
[0140] After the features are fused, the denoising transform coefficients S are obtained through global residual learning. OP According to formula (9), the fused feature M OP Add the shallow feature M0 to obtain the denoising transform coefficient S. OP :
[0141] S OP =M OP +M0 (9)
[0142] Fourth, obtain denoised training data through NSST inverse transform.
[0143] The obtained feature map S OP Perform inverse NSST transform to obtain the denoised training data I generated by the generator. out p ut The relevant calculation formula is as follows:
[0144]
[0145] 5. The MSFF-GAN discriminator network is used for identification, and the generator denoising model is obtained through iterative learning.
[0146] The denoised training data I obtained from the MSFF-GAN generator network out put With target seismic data I clear Simultaneously, the input is given to the MSFF-GAN discriminator network for learning and training. The MSFF-GAN discriminator network D and the MSFF-GAN generator network G are alternately optimized:
[0147]
[0148] Where, N -1 Represents the inverse NSST transform; θ D For the MSFF-GAN discriminator network parameters; θ G These are the network parameters for the MSFF-GAN generator.
[0149] Target seismic data I clear The training set consists of denoised synthetic data and denoised real data. The target seismic data, Iclear, is used to evaluate the denoising results of the MSFF-GAN generator network and optimize the generator denoising model.
[0150] Since the MSFF-GAN generator network and the MSFF-GAN discriminator network have opposite objectives, the training process of the two networks in the entire framework can be viewed as a minimax game. Both the MSFF-GAN generator network G and the MSFF-GAN discriminator network D are deep neural networks. The training model first randomizes the convolution parameters, and then uses an iterative method to train the MSFF-GAN generator network G and the MSFF-GAN discriminator network D respectively.
[0151] Each iteration consists of two steps: fixing the MSFF-GAN generator network parameters and training only the MSFF-GAN discriminator network parameters; and fixing the MSFF-GAN discriminator network parameters and training only the MSFF-GAN generator network parameters. These two steps are executed alternately throughout the training process, as follows:
[0152] For m noisy training data sets {z 1 ,z 2 ,...z m} and noise-free training data set {x 1 ,x 2 ,...x m First, the parameters θ of the MSFF-GAN generator network G are set. G Fixed, train the MSFF-GAN discriminator network D, and use the MSFF-GAN generator network G to generate a denoised dataset. The MSFF-GAN discriminator network parameters θ are updated by maximizing the following loss function. DUpdate the MSFF-GAN discriminator network parameters θ according to formulas (12) and (13). D :
[0153]
[0154] in, η represents the gradient operation; η represents the learning rate, which ranges from 0 to 1 and is determined by testing with actual data; V is the loss function.
[0155] The parameters θ of the MSFF-GAN discriminator network D D Fix the parameters θ of the MSFF-GAN generator network G during training. G Update the MSFF-GAN generator network parameters θ according to formulas (14) and (15). G :
[0156]
[0157] The MSFF-GAN discriminator network is used to distinguish target seismic data. clear and denoised training data I output Using the ReLU activation function can avoid max pooling throughout the network.
[0158] In this embodiment, a total of 8 convolutional layers are set, and the number of filter kernels is increased from 64 to 512. The final 512 feature maps are generated by two complete connected layers and a ReLU activation function.
[0159] A generator denoising model is obtained through iterative learning.
[0160] VI. Use a generator denoising model to denoise the seismic data to be processed.
[0161] Generator Denoising Model It can be used for seismic data to be processed. data The denoising result I of the seismic data to be processed is obtained according to formula (16). result :
[0162]
[0163] Among them, S data For the seismic data to be processed, I data The NSST decomposition results.
[0164] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0165] See appendix Figure 2The diagram illustrates the structure of a seismic data denoising system based on generative adversarial networks for implementing the above method, including: a data synthesis module, an NSST transform module, a feature extraction module, an inverse NSST transform module, a training module, and a denoising module.
[0166] The data synthesis module is used to acquire synthetic data and combine it with actual earthquake data to construct a training set;
[0167] The NSST transformation module is used to perform NSST transformation on the training data in the training set to obtain multiple frequency components.
[0168] The feature extraction module is used to input the frequency components into the MSFF-GAN generator network for feature extraction and to obtain the denoising transform coefficients.
[0169] The NSST inverse transform module is used to perform NSST inverse transform on the denoising transform coefficients to obtain denoising training data;
[0170] The training module is used to identify the denoised training data through the MSFF-GAN discriminator network based on the training data, perform iterative learning training, and obtain a generator denoising model.
[0171] The denoising module is used to denoise the seismic data to be processed using the generator denoising model.
[0172] Figure 3 A schematic diagram of the network architecture of the seismic data denoising method based on generative adversarial networks according to an embodiment of the present invention is shown. The training data in the training set is decomposed using NSST to obtain one low-frequency component and multiple high-frequency components. The low-frequency component and high-frequency components are input into the MSFF-GAN generator network G for feature extraction: input to the SFE module to extract shallow features; input the shallow features to the DFE module, and through multiple MSFE modules cascaded with the DFE module, extract deep features; feature fusion of the shallow and deep features is performed to obtain denoising transform coefficients; the denoising transform coefficients are reconstructed using NSST to obtain denoised training data.
[0173] The denoised training data and the target seismic data are simultaneously input into the MSFF-GAN discriminator network D, and multiple convolutional layers are used to distinguish between real and fake data.
[0174] Figure 4 This illustration shows a partial example diagram of a training set constructed based on synthetic data and actual seismic data according to an embodiment of the present invention.
[0175] Figure 5This diagram illustrates the structure of the MSFE module in the MSFF-GAN generator network according to an embodiment of the present invention. The diagram shows the structure of the MSFE module at layer d, M... d-1 The output of the (d-1)th layer MSFE module, after passing through the 1×1, 3×3, 5×5, and 7×7 convolutional functions and concatenation layers in the first part of the MSFE module, outputs O. p1 For O p1 Perform connection and slicing operations to extract the 64-dimensional features, and then combine them with M. d-1 Connect to the channel dimension to obtain the shortest path information.
[0176] O p1 The remaining 64-dimensional feature maps are used as input to the second part. After passing through the 1×1, 3×3, 5×5, and 7×7 convolutional functions and concatenation layers in the second part, the long path information O is output. P2 The short-path and long-path information is aggregated using a 1×1 convolutional layer and the ReLU activation function to output deep feature M. d The output M d It will be used as the input to the MSFE module at level d+1.
[0177] Figure 6 This diagram illustrates a comparison of the processing effects of embodiments of the present invention and conventional methods on synthetic data. The left side shows the denoising effect using a conventional method (denoising method of a large-scale seismic data processing software). Analysis revealed that it is difficult to achieve satisfactory results when the signal-to-noise ratio (SNR) of the denoised data is below -12dB. The right side shows the denoising effect using the method of this embodiment, where a relatively satisfactory result can still be obtained when the SNR of the denoised data is below -20dB.
[0178] Figure 7 A schematic diagram comparing the denoising results of seismic data processed by the embodiment of the present invention with those of conventional methods is shown. In actual seismic data processing, compared with conventional methods, the effective frequency band processed by the method of this embodiment is widened by 7Hz, and the high-frequency energy is significantly improved.
[0179] Figure 8 The diagram illustrates a high-frequency comparison of the denoising results of seismic data under processing using an embodiment of the present invention and conventional methods. In the high-frequency range of 60-120Hz, the signal-to-noise ratio is significantly improved.
[0180] Furthermore, embodiments of the present invention also provide a seismic data denoising device based on generative adversarial networks, comprising:
[0181] The data synthesis module acquires synthetic data and combines it with actual earthquake data to construct a training set. The synthetic data includes noisy synthetic data and denoised synthetic data, and the actual earthquake data includes noisy earthquake data and denoised earthquake data. The noisy synthetic data, denoised synthetic data, actual earthquake data, and denoised earthquake data are integrated to form the training set.
[0182] The NSST transformation module performs NSST transformation on the training data in the training set to obtain multiple frequency components; wherein, the frequency components are input into the MSFF-GAN generator network for feature extraction to obtain denoising transformation coefficients.
[0183] The NSST inverse transform module performs the NSST inverse transform on the denoising transform coefficients to obtain denoised training data.
[0184] The training module, based on the training data, uses the MSFF-GAN discriminator network to discriminate the denoised training data, performs iterative learning training, and obtains a generator denoising model; wherein the parameters of the MSFF-GAN generator network are fixed, the parameters of the MSFF-GAN discriminator network are trained, and the parameters of the MSFF-GAN generator network are fixed, and the generator denoising model is obtained through iterative learning.
[0185] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing one or more programs, which, when executed, can implement the aforementioned seismic data denoising method based on generative adversarial networks.
[0186] like Figure 9 As shown in the illustration, this embodiment of the invention also provides a device, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus.
[0187] The memory is a computer-readable storage medium used to store one or more programs.
[0188] The processor is configured to execute a program stored in a computer-readable storage medium.
[0189] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus.
[0190] 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.
Claims
1. A seismic data denoising method based on a generative adversarial network, characterized in that, The method comprises the following steps: obtaining synthetic data, combining actual seismic data to build a training set; performing NSST transformation on the training data in the training set to obtain a plurality of frequency components; inputting the frequency components into an MSFF-GAN generator network for feature extraction to obtain denoising transformation coefficients; performing NSST inverse transformation on the denoising transformation coefficients to obtain denoising training data; based on the training data, discriminating the denoising training data through an MSFF-GAN discriminator network, performing iterative learning training to obtain a generator denoising model; performing denoising processing on the to-be-processed seismic data through the generator denoising model.
2. The method of claim 1, wherein, The method comprises the following steps: obtaining a preset number of synthetic data through an enhancement scheme, wherein the synthetic data comprises noisy synthetic data and denoising synthetic data; the actual seismic data comprises noisy seismic data and denoising seismic data; integrating the noisy synthetic data, the denoising synthetic data, the actual seismic data and the denoising seismic data to form a training set.
3. The method of claim 1, wherein, The method comprises the following steps: inputting the frequency components into a shallow feature extraction module SFE to obtain shallow features; inputting the shallow features into a deep feature extraction module DFE to extract deep features.
4. The method of claim 3, wherein, The deep feature extraction module DFE is cascaded with a plurality of multi-scale feature extraction modules MSFE, and the deep features are extracted based on the multi-scale feature extraction modules MSFE, comprising: extracting feature information of local long and short paths respectively, aggregating the feature information of the long and short paths to obtain deep features.
5. The method of claim 4, wherein, The method further comprises the following steps: performing feature fusion on the shallow features and the deep features to obtain denoising transformation coefficients.
6. The method according to any one of claims 1 to 5, characterized in that, The method further comprises the following steps: inputting the denoising training data and target seismic data into the MSFF-GAN discriminator network for iterative learning training; the target seismic data comprises denoising synthetic data and denoising seismic data.
7. The method according to any one of claims 1 to 5, characterized in that, The method further comprises the following steps: fixing the parameters of the MSFF-GAN generator network and training the parameters of the MSFF-GAN discriminator network; fixing the parameters of the MSFF-GAN discriminator network and training the parameters of the MSFF-GAN generator network; obtaining a generator denoising model through iterative learning.
8. A seismic data denoising system based on a generative adversarial network, characterized in that, The system comprises: a data synthesis module for obtaining synthetic data, combining actual seismic data to build a training set; an NSST transformation module for performing NSST transformation on the training data in the training set to obtain a plurality of frequency components; a feature extraction module for inputting the frequency components into an MSFF-GAN generator network for feature extraction to obtain denoising transformation coefficients; an NSST inverse transformation module for performing NSST inverse transformation on the denoising transformation coefficients to obtain denoising training data; The training module is configured to perform iterative learning training on the denoising training data by using the MSFF-GAN discriminator network based on the training data, and obtain a generator denoising model. The denoising module is configured to perform denoising processing on the to-be-processed seismic data by using the generator denoising model.
9. The system of claim 8, wherein, The data synthesis module is further configured to: obtain a preset number of synthetic data by using an enhancement scheme, wherein the synthetic data includes noisy synthetic data and denoised synthetic data; the actual seismic data includes noisy seismic data and denoised seismic data; integrate the noisy synthetic data, the denoised synthetic data, the actual seismic data and the denoised seismic data to form a training set.
10. The system of claim 8, wherein, The feature extraction module is further configured to: input the frequency components into a shallow feature extraction module SFE to obtain shallow features; input the shallow features into a deep feature extraction module DFE to extract deep features.
11. The system according to any one of claims 8 to 10, characterized in that, The system further includes: fix the parameters of the MSFF-GAN generator network, and train the parameters of the MSFF-GAN discriminator network; fix the parameters of the MSFF-GAN discriminator network, and train the parameters of the MSFF-GAN generator network; obtain a generator denoising model by using iterative learning. 12.A computer readable storage medium storing one or more programs, wherein when the one or more programs are executed, the method of claim 1-7 is implemented. A processor, a communication interface and a computer readable storage medium are in electronic communication with each other through a communication bus; and 13. An electronic device comprising a processor, a communication interface, the computer readable storage medium of claim 12 and a communication bus; wherein, the processor is configured to execute the program stored in the computer readable storage medium.