Coherent noise suppression method and device for seismic data
By training the generator and discriminator, a target denoising network is built to perform coherent noise suppression on actual seismic data, solving the problem of poor coherent noise suppression effect in the existing technology, and achieving more efficient noise suppression and effective signal recovery.
Patent Information
- Application Number
- CN202311111494.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-30
- Publication Date
- 2025-02-28
AI Technical Summary
The prior art is difficult to effectively suppress coherent noise in post-stack seismic data, and traditional algorithms are poor in universality. Relying on parameter settings, deep learning methods require a large amount of paired data and the learning process is unstable.
By using the trained denoising network model, the actual seismic data is subjected to coherent noise suppression. Specific methods include: training the generator through actual seismic data, artificial synthetic reflection system stratigraphic model and Gaussian random noise, and introducing discriminators for generation and adversarial learning to obtain the target generator; using the target generator to construct seismic response training samples, including noise-containing response samples and noise-free seismic response labels; building a target denoising network through these samples and applying them to the seismic data to be processed to obtain noise-free seismic response prediction data.
Effective coherent noise suppression on actual seismic data is achieved, effective reflected signals in post-stack seismic data are restored, and the universality and stability of the processing effect are improved.
Smart Images
Figure CN119535553A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of oil and gas geophysics, and in particular to a method and device for suppressing coherent noise of seismic data. Background Art
[0002] The sources of coherent noise in post-stack seismic data are complex and varied, including residual surface waves, multiple waves and other types of noise in the process of seismic signal acquisition, as well as offset artifacts caused by the processing process. Its specific noise intensity, distribution frequency band and morphological pattern are affected by many aspects such as acquisition environment, acquisition method and processing flow. It is difficult to establish a theoretical noise generation model through theoretical deduction, and the coherent noise manifestations of post-stack seismic data in different blocks are also very different.
[0003] Traditional seismic processing algorithms are mainly based on the assumptions of structure, continuity, sparsity or similarity of effective seismic signals, and use time domain structural guidance filtering, fx domain predictive filtering, dictionary learning, low-rank fitting and local similarity filtering algorithms to suppress coherent noise. Although these algorithms can remove coherent noise to a certain extent, they have problems in terms of method universality, and the processing effect on specific data is very dependent on the setting of algorithm parameters.
[0004] In this regard, a large number of deep learning-based seismic noise suppression algorithms have been proposed. Deep learning-based seismic noise suppression requires a large amount of paired seismic data as learning samples. Since the noise-free response corresponding to the actual noisy seismic data cannot be obtained, it is impossible to directly create sample labels for it for network learning. In this regard, the existing technology often adopts the following two ideas for sample synthesis: (1) Sample synthesis based on the denoising results of traditional methods and synthesis based on the forward response of artificially generated formation models. Both of these denoising network learning sample synthesis methods have certain problems. Therefore, a series of self-supervised learning and CycleGAN-based unpaired learning seismic denoising algorithms have emerged. However, self-supervised learning seismic denoising often assumes that it can only process random noise and cannot effectively suppress the coherent noise in seismic data. Denoising algorithms based on unpaired learning often have problems such as unstable learning process and the need to manually determine the training termination time. Therefore, it is still necessary to study noise suppression technology for coherent noise in post-stack seismic data. Summary of the invention
[0005] The present invention provides a method and device for suppressing coherent noise of seismic data, which are used to suppress coherent noise of actual seismic data using a denoising network model obtained through training.
[0006] In a first aspect, the present invention provides a method for suppressing coherent noise of seismic data, comprising:
[0007] The generator is trained through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise, and the discriminator is introduced to perform generative adversarial learning to obtain the target generator;
[0008] Using the target generator, based on the actual seismic data and the artificial synthetic reflection system formation model, construct seismic response training samples; the seismic response training samples include: noisy response samples and corresponding noise-free seismic response labels;
[0009] Constructing a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels;
[0010] The seismic data to be processed is obtained, and the seismic data to be processed is subjected to the target denoising network to obtain noise-free seismic response prediction data.
[0011] Optionally, the generator is trained by using the acquired actual seismic data, the artificial synthetic reflection system formation model and Gaussian random noise, and the discriminator is introduced to perform generative adversarial learning to obtain a target generator, including:
[0012] Determining the wavelet information of the actual seismic data by means of an autocorrelation function;
[0013] Convolving the wavelet information with the synthetic reflection coefficient in the artificial synthetic reflection coefficient formation model to synthesize a corresponding noise-free seismic response;
[0014] The generator is subjected to adversarial learning through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator.
[0015] Optionally, the generator is subjected to adversarial learning through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator, including:
[0016] Generate, by a generator, coherent noise consistent with the noise style of the actual seismic data;
[0017] combining the coherent noise with the noise-free seismic response according to an additive model to obtain a noisy response;
[0018] Based on the noisy response and the actual seismic data, the generator is subjected to adversarial learning in combination with the discriminator to obtain the target generator.
[0019] Optionally, based on the noisy response and the actual seismic data, adversarial learning is performed on the generator in combination with the discriminator to obtain the target generator, including:
[0020] Determining a training error between the noisy response and the actual seismic data using the discriminator;
[0021] According to the training error, the generator is adjusted to obtain optimal network parameters;
[0022] The target generator is generated using the optimal network parameters.
[0023] In a second aspect, the present invention provides a coherent noise suppression device for seismic data, comprising:
[0024] The target generator construction module is used to train the generator through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise, and introduce the discriminator to perform generative adversarial learning to obtain the target generator;
[0025] A training sample construction module is used to construct a seismic response training sample using the target generator based on the actual seismic data and the artificial synthetic reflection system formation model; the seismic response training sample includes: a noisy response sample and a corresponding noise-free seismic response label;
[0026] A target denoising network construction module is used to construct a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels;
[0027] The prediction module is used to obtain the seismic data to be processed, and apply the seismic data to the target denoising network to obtain noise-free seismic response prediction data.
[0028] Optionally, the target generator construction module includes:
[0029] A wavelet information determination submodule, used to determine the wavelet information of the actual seismic data by means of an autocorrelation function;
[0030] A synthesis submodule, used for convolving the wavelet information with the synthetic reflection coefficient in the artificial synthetic reflection coefficient formation model to synthesize the corresponding noise-free seismic response;
[0031] A submodule is constructed, which is used to perform adversarial learning on the generator through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator.
[0032] Optionally, the construction submodule includes:
[0033] A coherent noise generating unit, used for generating coherent noise consistent with the noise style of the actual seismic data through a generator;
[0034] A combining unit, used for combining the coherent noise with the noise-free seismic response according to an additive model to obtain a noisy response;
[0035] An adversarial learning unit is used to perform adversarial learning on the generator based on the noisy response and the actual seismic data in combination with the discriminator to obtain the target generator.
[0036] Optionally, the adversarial learning unit includes:
[0037] a training error determination subunit, configured to determine a training error between the noisy response and the actual seismic data using the discriminator;
[0038] An optimal network parameter determination subunit, used to adjust the generator according to the training error to obtain the optimal network parameter;
[0039] A generating subunit is used to generate the target generator by adopting the optimal network parameters.
[0040] In a third aspect, the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps in the method provided in the first aspect are performed.
[0041] In a fourth aspect, the present application provides a storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps in the method provided in the first aspect are performed.
[0042] It can be seen from the above technical solutions that the present invention has the following advantages:
[0043] The present invention provides a method and device for suppressing coherent noise of seismic data. The method comprises: training a generator through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise obtained, and introducing a discriminator to perform generative adversarial learning to obtain a target generator; using the target generator, based on the actual seismic data and the artificial synthetic reflection system formation model, constructing seismic response training samples; the seismic response training samples include: noisy response samples and corresponding noise-free seismic response labels; constructing a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels; obtaining seismic data to be processed, and assigning the seismic data to be processed to the target denoising network to obtain noise-free seismic response prediction data. By learning the noise adversarial generative modeling for the actual seismic data, a corresponding target generator is established, and then the artificial synthetic seismic response training samples for denoising network learning are generated by the target generator, thereby obtaining the target denoising network, and finally the target denoising network model is used to perform coherent noise suppression processing on the seismic data to be processed. Therefore, the coherent noise suppression can be performed on the actual seismic data to be processed more effectively, and the effective reflection signals in the post-stack seismic data can be better restored. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0045] Figure 1 A flowchart of a first embodiment of a method for suppressing coherent noise of seismic data according to the present invention;
[0046] Figure 2 This is a flow chart of a second embodiment of a method for suppressing coherent noise of seismic data according to the present invention;
[0047] Figure 3 This is a flowchart of a third embodiment of a method for suppressing coherent noise of seismic data according to the present invention;
[0048] Figure 4 It is a general schematic diagram of Embodiment 3 of a method for suppressing coherent noise of seismic data of the present invention;
[0049] Figure 5 is a structural diagram of a generator according to an embodiment of the present invention;
[0050] Figure 6 is a structural diagram of a discriminator according to an embodiment of the present invention;
[0051] Figure 7 A denoising network structure diagram of an embodiment of the present invention;
[0052] Figure 8 A schematic diagram of noisy post-stack seismic data in Example 4 of a method for suppressing coherent noise in seismic data of the present invention;
[0053] Fig. 9 The sample data corresponding to the noisy post-stack seismic data generated by the generator model;
[0054] Fig.10 This is a schematic diagram of the denoising results corresponding to the noisy post-stack seismic data obtained through the denoising network;
[0055] Fig.11 Schematic diagram of noise removal obtained through the denoising network;
[0056] Fig.12 Schematic diagram of denoising results obtained by the DnSYN method;
[0057] Fig.13 Schematic diagram of noise removal obtained by DnSYN method;
[0058] Fig.14 Schematic diagram of denoising results obtained by the CycleGAN method;
[0059] Fig.15 Schematic diagram of noise removal obtained by CycleGAN method;
[0060] Fig.16 Post-stack seismic data for a certain area in Mouta River;
[0061] Fig.17 This is a schematic diagram of the denoising results corresponding to the post-stack seismic data of a certain area in Mouta River obtained through the denoising network;
[0062] Fig.18 This is a schematic diagram of noise removal corresponding to the post-stack seismic data of a certain area in Mouta River obtained through the denoising network;
[0063] Fig.19 The present invention is a structural block diagram of an embodiment of a coherent noise suppression device for seismic data according to the present invention. DETAILED DESCRIPTION
[0064] The embodiment of the present invention provides a method and device for suppressing coherent noise of seismic data, which is used to suppress coherent noise of actual seismic data using a trained denoising network model.
[0065] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0066] For example, see Figure 1 , Figure 1 The flowchart of the first embodiment of a method for suppressing coherent noise of seismic data of the present invention includes:
[0067] S101, training the generator through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise, and introducing the discriminator to perform generative adversarial learning to obtain a target generator;
[0068] S102, using the target generator to construct a seismic response training sample based on the actual seismic data and the artificial synthetic reflection system formation model; the seismic response training sample includes: a noisy response sample and a corresponding noise-free seismic response label;
[0069] S103, constructing a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels;
[0070] S104, obtaining the seismic data to be processed, and applying the seismic data to the target denoising network to obtain noise-free seismic response prediction data.
[0071] A method for suppressing coherent noise of seismic data provided in an embodiment of the present invention includes: training a generator through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise obtained, and introducing a discriminator to perform generative adversarial learning to obtain a target generator; using the target generator, based on the actual seismic data and the artificial synthetic reflection system formation model, constructing seismic response training samples; the seismic response training samples include: noisy response samples and corresponding noise-free seismic response labels; constructing a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels; obtaining seismic data to be processed, and assigning the seismic data to be processed to the target denoising network to obtain noise-free seismic response prediction data. Through noise adversarial generative modeling learning for actual seismic data, a corresponding target generator is established, and then the artificial synthetic seismic response training samples for denoising network learning are generated through the target generator, thereby obtaining a target denoising network, and finally using the target denoising network model to suppress coherent noise on the seismic data to be processed. Thereby, coherent noise suppression can be performed on the actual seismic data to be processed more effectively, and effective reflection signals in post-stack seismic data can be better restored.
[0072] For example 2, please refer to Figure 2 , Figure 2 This is a flow chart of a second embodiment of a method for suppressing coherent noise of seismic data according to the present invention, the method comprising:
[0073] Step S201, determining the wavelet information of the actual seismic data by an autocorrelation function method;
[0074] Step S202, convolving the wavelet information with the synthetic reflection coefficient in the artificial synthetic reflection coefficient formation model to synthesize a corresponding noise-free seismic response;
[0075] Step S203, performing adversarial learning on the generator through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator;
[0076] Step S204, using the target generator to construct seismic response training samples based on the actual seismic data and the artificial synthetic reflection system formation model; the seismic response training samples include: noisy response samples and corresponding noise-free seismic response labels;
[0077] Step S205, constructing a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels;
[0078] Step S206, obtaining the seismic data to be processed, and applying the seismic data to the target denoising network to obtain noise-free seismic response prediction data.
[0079] For example 3, please refer to Figure 3 and Figure 4 , Figure 3 FIG. 1 is a flow chart of a third embodiment of a method for suppressing coherent noise of seismic data according to the present invention. Figure 4 FIG. 1 is a general schematic diagram of a third embodiment of a method for suppressing coherent noise of seismic data according to the present invention, wherein the method comprises:
[0080] Step S301, determining the wavelet information of the actual seismic data by using an autocorrelation function method;
[0081] Step S302, convolving the wavelet information with the synthetic reflection coefficient in the artificial synthetic reflection coefficient formation model to synthesize a corresponding noise-free seismic response;
[0082] Step S303, generating coherent noise consistent with the noise style of the actual seismic data through a generator;
[0083] Step S304, combining the coherent noise with the noise-free seismic response according to an additive model to obtain a noisy response;
[0084] Step S305, based on the noisy response and the actual seismic data, the generator is subjected to adversarial learning in combination with the discriminator to obtain the target generator;
[0085] In an optional embodiment, based on the noisy response and the actual seismic data, adversarial learning is performed on the generator in combination with the discriminator to obtain the target generator, including:
[0086] Determining a training error between the noisy response and the actual seismic data using the discriminator;
[0087] According to the training error, the generator is adjusted to obtain optimal network parameters;
[0088] The target generator is generated using the optimal network parameters.
[0089] In the embodiment of the present invention, a generator is used to refer to the distribution of useful signals in the frequency domain in actual seismic data, and combined with the synthetic inverse coefficient convolution in the artificial synthetic reflection coefficient formation model to obtain a noise-free seismic response. Then, the coherent noise generated by the generator, which is consistent with the noise style of the actual seismic data, is combined with the noise-free seismic response according to the additive model to obtain a noisy response. Then, a discriminator is introduced to perform generative adversarial learning, requiring that the sample data and coherent noise generated by the generator are consistent with the style of the actual seismic data. Through the coherent noise adversarial generative learning process, a target generator is established.
[0090] See also Figure 5 , Figure 5 : is a structural diagram of a generator of an embodiment of the present invention. The generator takes the actual seismic response, synthetic reflection coefficient and Gaussian random noise as input, and generates a noise-free seismic response and coherent noise consistent with the actual seismic response and subsequent seismic data to be processed through a series of calculations, namely:
[0091] X syn , N gen =G(N gaus , Y field , R syn );
[0092] Among them, Y field is the actual earthquake response, X syn is the noise-free seismic response, N gen is the coherent noise, N guas is Gaussian random noise and G is the generator.
[0093] The wavelet information of the actual seismic response is estimated using a method based on the autocorrelation function, and the noise-free seismic response is synthesized based on it, namely:
[0094]
[0095] in, is the ith channel of the actual earthquake response, cor is the autocorrelation function, is the main frequency of the ith channel of the actual earthquake response, R syn is the composite reflection coefficient.
[0096] In order to eliminate the sidelobe effect of the autocorrelation functor wave number estimate, the wavelet estimate will be multiplied by a cosine window set according to the main frequency of the ith channel of the actual seismic response. Finally, the average wavelet of the wavelet estimation results on each channel will be used as the final wavelet estimate, and the corresponding noise-free seismic response will be synthesized by convolution with the synthetic reflection coefficient.
[0097] For the generation of coherent noise, the generator in the embodiment of the present invention will first use G NsC The module performs bandpass filtering on the Gaussian random noise from the fx domain according to the fx spectrum of the actual seismic response, and converts the Gaussian random noise into colored noise with a frequency band consistent with the noise in the actual seismic response, that is:
[0098]
[0099] in, and They are the fx forward transform and inverse transform respectively. The above formula roughly estimates the band distribution of the logarithmic amplitude spectrum of the wavelet by longitudinally averaging the fx domain logarithmic amplitude spectrum of the actual seismic response. Through the band distribution of the logarithmic amplitude spectrum of the wavelet, a rough noise logarithmic amplitude band distribution can be obtained. On this basis, the corresponding fx domain bandpass filter can be constructed for Gaussian random noise, and finally it can be converted into a colored noise image with specific frequency band distribution characteristics through the inverse fx transform. On this basis, the generator uses a series of convolutions and step convolutions to form the encoder module G ENC , mapping the colored noise into feature maps at four scales, namely:
[0100] n (1) , n (2) , n (3) , n (4) =G ENC (N color );
[0101] In order to ensure that the spatial distribution of the generated noise amplitude is consistent with the energy distribution of the noise-free seismic response, the generator will also obtain the instantaneous amplitude of the noise-free seismic response based on the Hilbert transform, and then construct the local noise intensity adjustment module G AMP , this module uses a series of maxpool pooling downsampling and then upsampling to fuse into amplitude control feature maps at four scales, namely:
[0102] w (1) , w (2) , w (3) , w (4) =G AMP (A ins (X syn ));
[0103] Among them, A ins Represents the instantaneous amplitude calculation based on the Hilbert transform, which is embedded in the implementation of the network.
[0104] The generator uses a simplified AdaIN mechanism to implement the use of w (i) Control (i) The calculation of local intensity, specifically, the simplified AdaIN calculation process is as follows:
[0105]
[0106] in, Representative according to n (i) The expectation and variance of a single feature map of a single instance are standardized, namely:
[0107]
[0108] AdaIN will use a learnable parameter vector a to map w (i) Perform weighted combination according to the channel and use softmax operation to determine the pair The amplitude gain scaling degree of the converted feature map m on this scale is obtained (i) The above AdaIN operation will be combined with a series of convolutions and transposed convolutions to form the decoder module G DEC , so that it can be (1) , w (2) , w (3) , w (4) Under the control of n (1) , n (2) , n (3) , n (4) , and finally reconstructed into coherent noise with a specific style. CSC , G NSC and A ins The operations such as averaging, autocorrelation, convolution, fx transformation, etc. involved in the modules can be implemented by operations supported by TensorFlow or PyTorch, so they can be directly embedded in the network's computing process. Therefore, these modules are implemented as special network layers that perform specified operations according to given logic. The computational operations corresponding to these special network layers directly embed the domain knowledge of seismic processing into the network computing process, making it possible to combine the domain knowledge of seismic processing for noise generation learning. While reducing the difficulty of generator learning, it also improves the rationality of using it for coherent noise generation.
[0109] At the same time, in order to ensure that the aforementioned generator can effectively generate coherent noise that is consistent with the style of the seismic data to be processed (and the actual seismic data), another discriminator needs to be introduced to perform adversarial noise generation learning. The specific noise generation adversarial learning process first starts with the pre-training of the discriminator. Its learning goal is to be able to judge whether the noisy seismic response is from the actual seismic data to be processed or the noisy response sample synthesized by the generator through the style pattern of the input noisy seismic response.
[0110] See also Figure 6 , Figure 6 is a diagram of the discriminator structure of an embodiment of the present invention. In a specific implementation, the embodiment of the present invention analyzes the characteristics of the problem of coherent noise adversarial generative learning and adopts the following method: Figure 6 The network structure in is used to realize the discriminator. Compared with the conventional discriminator, Figure 6 The discriminator in has the following main features:
[0111] 1. In order to avoid the discriminator's identification process being affected by the difference in clean signals in the input noisy seismic response, when identifying the input noisy seismic response, a series of Gaussian filters with different parameters are first used to perform preliminary denoising, and the standardized denoised residual image is obtained as a feature for further identification. Although these Gaussian filters cannot perfectly eliminate the noise in the noisy seismic response, they can effectively reduce the energy proportion of the effective signal and reduce the impact of the effective signal difference between the actual seismic data and the noisy response on the discriminator. The discriminator will then use a similar multi-scale encoder-decoder structure to infer the source of the noisy seismic response based on the coherent noise.
[0112] 2. Unlike the classic discriminator, which often gives a unique overall identification result in the form of 0 to 1 numerical value for the entire input image, Figure 6 The discriminator in will output a discriminative feature map D(Y) of the same size as the input noisy seismic response, representing the discriminative result of each pixel in the noisy seismic response. This design is necessary for adversarial noise generation modeling: outputting only the overall discriminative result will make adversarial learning only require the overall style of the noisy response synthesized by the generator to be consistent with the seismic data to be processed and the actual seismic data; while outputting the pixel-level discriminative result further requires that the style near each pixel of the noisy response is consistent with the seismic data to be processed and the actual seismic data. This will further improve the realism of the coherent noise in the generated results.
[0113] Specifically, with the help of the designed generator and discriminator, adversarial noise generation learning will first start with the learning of the discriminator. The process will use the following loss function:
[0114]
[0115] in, It is a typical generative adversarial loss, which requires that when the noisy response generated by the generator is identified by the current discriminator, every pixel of the output identification feature map is 1; The generated noise-free response label and the coherent noise are required to be orthogonal to each other. Using this loss function to learn the generator requires that the style of each generated noisy response is as consistent as possible with the actual seismic data, so that the current discriminator can no longer distinguish between the two, and the generated noise-free response label and the coherent noise have a detailed difference, and the denoising model can be used to suppress the coherent noise.
[0116] The specific adversarial noise generation learning process will repeatedly alternate the learning of the discriminator and the generator. In each round of learning, the generator will force the noisy response it generates to be as consistent as possible with the style of the actual seismic data under the scale of the current discriminator. Next, the current generator will be used to distinguish the subtle differences between the generated noisy response and the actual seismic data through the learning of the discriminator. The updated discriminator will be used to assist the next round of generator learning to eliminate the difference between the noisy response generated by the generator at the current stage and the actual seismic data. And so on and so forth, iterating continuously, and eventually the discriminator will be unable to distinguish the difference between the generated noisy response and the actual seismic data. At this point, the style of the noisy response generated by the generator will be highly consistent with the actual seismic data. However, since the noise-free seismic response corresponding to the noisy response is known, the coherent noise suppression network can be learned based on this.
[0117] Step S306, using the target generator, based on the actual seismic data and the artificial synthetic reflection system formation model, constructing a seismic response training sample; the seismic response training sample includes: a noisy response sample and a corresponding noise-free seismic response label;
[0118] In the embodiment of the present invention, a large number of noisy response samples consistent with the style of actual seismic data are generated by a target generator, and corresponding noise-free seismic response labels are annotated.
[0119] Step S307, constructing a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels;
[0120] In the embodiment of the present invention, the denoising network D is learned according to the supervised end-to-end mapping learning method of the noise-free response label. The denoising network D finally obtained can effectively suppress the coherent noise in the actual seismic data, and produces good results when applied to the actual seismic data for coherent noise suppression processing.
[0121] Step S308, obtaining the seismic data to be processed, and applying the seismic data to the target denoising network to obtain noise-free seismic response prediction data.
[0122] In an embodiment of the present invention, after the corresponding generator is established through the aforementioned coherent noise adversarial generation learning process, a large number of noisy responses consistent with the style of the seismic data to be processed and the actual seismic data are generated by the generator, and the corresponding noise-free response labels are known. On this basis, the denoising network can be learned according to the supervised end-to-end mapping learning method of the noise-free seismic response type. The final denoising network will be able to effectively suppress the coherent noise in the actual seismic data, and can be applied to the actual seismic data for the final coherent noise suppression process.
[0123] In the specific implementation, after completing the adversarial noise modeling, the generator at this time will be able to refer to the actual seismic data, and generate a large number of noise-free seismic responses and coherent noises with similar signal frequency bands and noise styles according to the artificially synthesized reflection coefficient formation model and Gaussian random noise. Then, according to the additive model, a large number of noisy responses consistent with the style of the actual seismic data are synthesized to construct training samples. For these training samples, since the noise-free seismic response labels corresponding to the noisy response samples are known, the denoiser model can be directly learned in the form of supervised end-to-end mapping from the noisy response sample input to the noise-free seismic response output. The specific learning process will be carried out in collaboration with labeled artificially synthesized samples and actual seismic data. The loss function used for denoiser learning is as follows:
[0124]
[0125] in, The denoiser is required to completely remove the coherent noise in the noisy response data and restore it to a noise-free seismic response. 2 (X, Y) is the square of the linear correlation coefficient between images X and Y. 2 (F(Y field ), Y field -F(Y field )) requires that when the denoiser processes the actual seismic data, the denoising result F(Y field ) and noise removal Y field -F(Y field ) are mutually orthogonal.
[0126] By using a generator to construct a seismic response training sample based on an artificial synthetic reflection system formation model at the actual seismic data level, the corresponding denoising network can be learned under the constraints of the above loss function. Since the generator can effectively imitate the noise process in the actual seismic data and eliminate the synthetic coherent noise in the noisy seismic response through the generated anti-noise modeling learning, it can also effectively suppress the coherent noise with a consistent pattern in the seismic data to be processed. Therefore, the obtained denoising network can be directly applied to the seismic data to be processed for coherent noise suppression. In the subsequent application effect demonstration, the present invention uses the following Figure 7 The denoising network is constructed using the network structure shown in . In fact, other end-to-end image denoising networks can also be used to suppress coherent noise in post-stack seismic data using the above method.
[0127] Embodiment 4, with Figure 8 Taking the noisy post-stack seismic data of a certain area in as an example, the coherent noise therein is subjected to adversarial generative modeling by using the methods proposed in the first to third embodiments of the present invention. The sample data generated by the generator model finally obtained is as follows: Fig. 9 As shown. Fig. 9 The synthetic noisy data and Figure 8 It can be seen that the two have a high degree of style consistency. On this basis, the denoising network is used to Figure 8 To suppress the coherent noise, the corresponding denoising results and removed noise are as follows: Fig.10 and Fig.11 The result of this method, namely the DnANG denoising result, can be compared with the existing methods, such as Fig.12 Schematic diagram of denoising results using DnSYN (using artificially synthesized noise-free responses + traditional methods to extract noise to make samples for denoising network learning), and Fig.13 The noise removal diagram shown, and Fig.14 Schematic diagram of denoising results using CycleGAN (using CycleGAN unpaired image transfer learning for denoising network learning), and Fig.15 Schematic diagram of CycleGAN removing noise. By comparison, we can see that: (1) DnSYN cannot completely remove Figure 8 (2) The denoising result of CycleGAN is too smooth, which obviously damages the effective signal; (3) DnANG can effectively suppress the coherent noise while maintaining Figure 8 The effective signal details in the data achieved the best processing effect.
[0128] To verify the versatility of the technical method of the present invention, we further applied it to Fig.16 The denoising results are as follows: Fig.17 As shown, the noise is removed as Fig.18 As shown in Figure 2, although the coherent noise in the post-stack seismic data in this area is Figure 8 Although there are obvious differences in the data, they can still be imitated through the coherent noise adversarial generative modeling in the present invention. The resulting denoising network can effectively suppress the coherent noise without damaging the effective signal.
[0129] For example 5, please refer to Fig.19 , Fig.19 The structural block diagram of an embodiment of a coherent noise suppression device for seismic data of the present invention comprises:
[0130] The target generator construction module 501 is used to train the generator through the acquired actual seismic data, the artificial synthetic reflection system formation model and Gaussian random noise, and introduce the discriminator to perform generative adversarial learning to obtain the target generator;
[0131] The training sample construction module 502 is used to construct a seismic response training sample based on the actual seismic data and the artificial synthetic reflection system formation model by using the target generator; the seismic response training sample includes: a noisy response sample and a corresponding noise-free seismic response label;
[0132] A target denoising network construction module 503 is used to construct a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels;
[0133] The prediction module 504 is used to obtain the seismic data to be processed, and apply the seismic data to the target denoising network to obtain noise-free seismic response prediction data.
[0134] In an optional embodiment, the target generator construction module 501 includes:
[0135] A wavelet information determination submodule, used to determine the wavelet information of the actual seismic data by means of an autocorrelation function;
[0136] A synthesis submodule, used for convolving the wavelet information with the synthetic reflection coefficient in the artificial synthetic reflection coefficient formation model to synthesize the corresponding noise-free seismic response;
[0137] A submodule is constructed, which is used to perform adversarial learning on the generator through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator.
[0138] In an optional embodiment, the construction submodule includes:
[0139] A coherent noise generating unit, used for generating coherent noise consistent with the noise style of the actual seismic data through a generator;
[0140] A combining unit, used for combining the coherent noise with the noise-free seismic response according to an additive model to obtain a noisy response;
[0141] An adversarial learning unit is used to perform adversarial learning on the generator based on the noisy response and the actual seismic data in combination with the discriminator to obtain the target generator.
[0142] In an optional embodiment, the adversarial learning unit includes:
[0143] a training error determination subunit, configured to determine a training error between the noisy response and the actual seismic data using the discriminator;
[0144] An optimal network parameter determination subunit, used to adjust the generator according to the training error to obtain the optimal network parameter;
[0145] A generating subunit is used to generate the target generator by adopting the optimal network parameters.
[0146] An embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a coherent noise suppression method for seismic data as described in any of the above embodiments.
[0147] An embodiment of the present invention further provides a computer storage medium having a computer program stored thereon, and when the computer program is executed by the processor, the steps of a coherent noise suppression simulation method for seismic data as described in any of the above embodiments are implemented.
[0148] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0149] In the several embodiments provided in the present application, it should be understood that the methods, devices, electronic devices and storage media disclosed in the present invention can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.
[0150] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0151] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0152] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a readable storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.
[0153] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for suppressing coherent noise of seismic data, characterized in that: include: The generator is trained through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise, and the discriminator is introduced to perform generative adversarial learning to obtain the target generator; Using the target generator, based on the actual seismic data and the artificial synthetic reflection system formation model, construct a seismic response training sample; The seismic response training samples include: noisy response samples and corresponding noise-free seismic response labels; Constructing a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels; The seismic data to be processed is obtained, and the seismic data to be processed is subjected to the target denoising network to obtain noise-free seismic response prediction data.
2. The method for suppressing coherent noise of seismic data according to claim 1, characterized in that: The generator is trained through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise, and the discriminator is introduced to perform generative adversarial learning to obtain the target generator, including: Determining the wavelet information of the actual seismic data by means of an autocorrelation function; Convolving the wavelet information with the synthetic reflection coefficient in the artificial synthetic reflection coefficient formation model to synthesize a corresponding noise-free seismic response; The generator is subjected to adversarial learning through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator.
3. The method for suppressing coherent noise of seismic data according to claim 2, characterized in that: The generator is subjected to adversarial learning through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator, including: Generate, by a generator, coherent noise consistent with the noise style of the actual seismic data; combining the coherent noise with the noise-free seismic response according to an additive model to obtain a noisy response; Based on the noisy response and the actual seismic data, the generator is subjected to adversarial learning in combination with the discriminator to obtain the target generator.
4. The method for suppressing coherent noise of seismic data according to claim 3, characterized in that: Based on the noisy response and the actual seismic data, adversarial learning is performed on the generator in combination with the discriminator to obtain the target generator, including: Determining a training error between the noisy response and the actual seismic data using the discriminator; According to the training error, the generator is adjusted to obtain optimal network parameters; The target generator is generated using the optimal network parameters.
5. A coherent noise suppression device for seismic data, characterized in that: include: The target generator construction module is used to train the generator through the actual seismic data, artificial synthetic reflection system formation model and Gaussian random noise, and introduce the discriminator to perform generative adversarial learning to obtain the target generator; A training sample construction module, used to construct a seismic response training sample based on the actual seismic data and the artificial synthetic reflection system formation model by using the target generator; The seismic response training samples include: noisy response samples and corresponding noise-free seismic response labels; A target denoising network construction module is used to construct a target denoising network through the noisy response samples and the corresponding noise-free seismic response labels; The prediction module is used to obtain the seismic data to be processed, and apply the seismic data to the target denoising network to obtain noise-free seismic response prediction data.
6. The coherent noise suppression device for seismic data according to claim 5, characterized in that: The target generator building blocks include: A wavelet information determination submodule, used to determine the wavelet information of the actual seismic data by means of an autocorrelation function; A synthesis submodule, used for convolving the wavelet information with the synthetic reflection coefficient in the artificial synthetic reflection coefficient formation model to synthesize the corresponding noise-free seismic response; A submodule is constructed, which is used to perform adversarial learning on the generator through the actual seismic data machine frequency band distribution, the noise-free seismic response and the discriminator to obtain the target generator.
7. The coherent noise suppression device for seismic data according to claim 6, characterized in that: The construction submodule includes: A coherent noise generating unit, used for generating coherent noise consistent with the noise style of the actual seismic data through a generator; A combining unit, used for combining the coherent noise with the noise-free seismic response according to an additive model to obtain a noisy response; An adversarial learning unit is used to perform adversarial learning on the generator based on the noisy response and the actual seismic data in combination with the discriminator to obtain the target generator.
8. The coherent noise suppression device for seismic data according to claim 7, characterized in that: The adversarial learning unit comprises: a training error determination subunit, configured to determine a training error between the noisy response and the actual seismic data using the discriminator; An optimal network parameter determination subunit, used to adjust the generator according to the training error to obtain the optimal network parameter; A generating subunit is used to generate the target generator by adopting the optimal network parameters.
9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 4 is executed.
10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is executed.