Method, device, equipment and system for suppressing ground roll
By combining frequency-wavenumber domain filtering with unsupervised deep neural networks, comprehensive suppression of ground roll waves was achieved, solving the problem of noise residue in traditional methods and improving the processing efficiency and interpretation accuracy of seismic data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF PETROLEUM (BEIJING)
- Filing Date
- 2026-03-23
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies can easily affect effective signals when suppressing ground roll waves, leading to reduced efficiency and accuracy in seismic data processing, especially since residual noise in the low-frequency band is difficult to remove.
By combining frequency-wavenumber domain filtering preprocessing with unsupervised deep neural networks, and through frequency band segmentation and unsupervised learning, the kinematic characteristics of intermediate frequency signals are utilized to suppress ground roll waves. A self-supervised learning framework that does not require manual labeling is constructed to achieve comprehensive suppression of ground roll waves.
It improves the accuracy of roll wave suppression and the fidelity of effective signals, significantly enhances the signal-to-noise ratio and interpretation accuracy of seismic data, and is suitable for seismic exploration under complex geological conditions.
Smart Images
Figure CN122386393A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of seismic exploration technology, and in particular to a method, apparatus, equipment and system for suppressing ground roll waves. Background Technology
[0002] In the field of seismic exploration, because the frequency and apparent velocity differences between roll waves and effective signals are small in the low-frequency band, traditional suppression methods easily leave roll wave noise in the low-frequency region, causing the effective signal to be masked. Therefore, how to achieve comprehensive suppression of roll waves without relying on manually labeled data has become a key technical bottleneck for improving the efficiency of seismic data processing and interpretation accuracy.
[0003] Existing technologies mainly suppress ground roll waves through transform domain methods. For example, based on the apparent velocity difference between ground roll waves and effective signals in the frequency-wavenumber domain, low-speed noise is suppressed by setting a velocity threshold; or the signal is projected into a sparse domain using sparse representation theory to separate noise from the effective signal.
[0004] However, the above-mentioned technical solutions still affect the effective signal when performing roll wave suppression, causing the effective signal to be masked and reducing the accuracy of roll wave suppression. Summary of the Invention
[0005] The method, apparatus, equipment and system for suppressing ground roll waves provided in this application are used to solve the technical problem that existing ground roll wave suppression methods cannot accurately suppress ground roll waves, resulting in the masking of effective signals.
[0006] In a first aspect, this application provides a method for suppressing ground rolling waves, comprising:
[0007] The original seismic data were preprocessed by frequency-wavenumber domain filtering to obtain the initial suppression results;
[0008] The initial suppression results are subjected to frequency band segmentation processing to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data;
[0009] Based on the mid-frequency band sub-data and the pre-trained unsupervised deep neural network model, a low-mid-frequency prediction signal is obtained. The pre-trained unsupervised deep neural network model is trained based on the mid-frequency band sub-data and the low-mid-frequency band data.
[0010] The difference between the low-frequency prediction signal and the low-frequency band sub-data is calculated to determine the rolling wave noise to be processed;
[0011] The rolling wave noise to be processed is subtracted from the initial suppression result to complete the rolling wave suppression process.
[0012] Furthermore, the method also includes:
[0013] Determine the initial unsupervised deep neural network model;
[0014] Based on the mid-frequency band sub-data, the initial unsupervised deep neural network model is trained to obtain mid-frequency band training data;
[0015] Based on the mid-frequency band training data and the low-mid-frequency band sub-data, a loss function is constructed;
[0016] Based on the loss function, the initial unsupervised deep neural network model is trained to obtain a pre-trained unsupervised deep neural network model.
[0017] Further, based on the mid-frequency band sub-data and the pre-trained unsupervised deep neural network model, a low-mid-frequency prediction signal is obtained, including:
[0018] The mid-frequency band sub-data is subjected to feature extraction processing to generate mid-frequency band kinematic features;
[0019] Based on a preset unsupervised deep neural network model, the mid-frequency band kinematic features are subjected to nonlinear mapping processing to obtain a low-mid-frequency prediction signal.
[0020] Further, the initial suppression result is subjected to frequency band segmentation processing to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data, including:
[0021] Determine the low-pass filter and the band-pass filter;
[0022] The initial suppression result is filtered using the bandpass filter to obtain intermediate frequency sub-data.
[0023] The initial suppression result is filtered using the low-pass filter to obtain low-frequency sub-band data.
[0024] The mid-frequency sub-data and low-frequency sub-data are summed to obtain low-mid-frequency sub-data.
[0025] Further, the initial suppression result is subjected to frequency band segmentation processing to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data, including:
[0026] The initial suppression result is subjected to wavelet transform to obtain multiple wavelet coefficient subbands of different scales;
[0027] From the multiple wavelet coefficient sub-bands, wavelet coefficient sub-bands in the mid-frequency range are selected to obtain mid-frequency band sub-data;
[0028] From the multiple wavelet coefficient sub-bands, select the wavelet coefficient sub-band in the low-frequency range to obtain the low-frequency band sub-data;
[0029] The intermediate frequency band data is added to the low frequency band data to obtain the low-intermediate frequency band data.
[0030] Furthermore, before performing frequency-wavenumber domain filtering preprocessing on the raw seismic data to obtain the initial suppression result, the method further includes:
[0031] Collect raw seismic data;
[0032] The original seismic data is analyzed using the sliding window method to obtain the local energy distribution of the original seismic data;
[0033] Based on the local energy distribution of the original seismic data, determine the apparent velocity difference between the effective signal and the roll wave noise in the original seismic data;
[0034] Based on the apparent velocity difference, the apparent velocity threshold for frequency-wavenumber domain filtering is determined. The apparent velocity threshold is a velocity threshold parameter used to distinguish between the effective signal and the roll wave noise during the frequency-wavenumber domain filtering preprocessing of the original seismic data.
[0035] Furthermore, the pre-trained unsupervised deep neural network model is a U-Net network structure.
[0036] Secondly, this application provides a device for suppressing rolling waves, comprising:
[0037] The initial suppression module for rolling waves is used to perform frequency-wavenumber domain filtering preprocessing on the raw seismic data to obtain the initial suppression result;
[0038] The frequency band segmentation module is used to perform frequency band segmentation processing on the initial suppression result to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data;
[0039] The signal prediction module is used to obtain a low-intermediate frequency predicted signal based on the intermediate frequency band sub-data and the pre-trained unsupervised deep neural network model, wherein the pre-trained unsupervised deep neural network model is trained based on the intermediate frequency band data and the low-intermediate frequency band data.
[0040] The rolling wave noise determination module is used to calculate the difference between the low-frequency prediction signal and the low-frequency sub-band data to determine the rolling wave noise to be processed.
[0041] The ground roll wave suppression module is used to subtract the ground roll wave noise to be processed from the initial suppression result to complete the ground roll wave suppression process.
[0042] Thirdly, this application provides a device for suppressing ground rolling waves, including: a memory and a processor;
[0043] The memory stores computer-executed instructions;
[0044] The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any of the first aspects.
[0045] Fourthly, this application provides a system for suppressing ground rolling waves, including a computer program that, when executed by a processor, implements the method described in any of the first aspects.
[0046] This application provides a method, apparatus, device, and system for suppressing roll waves. An initial suppression result is obtained by preprocessing the raw seismic data using frequency-wavenumber domain filtering. The initial suppression result is then segmented into mid-frequency (IF) and low-IF frequency (LIF) sub-data. A LIF prediction signal is obtained based on the IF sub-data and a pre-trained unsupervised deep neural network model trained on the IF and LIF sub-data. The difference between the LIF prediction signal and the LIF sub-data is calculated to determine the roll wave noise to be processed. The roll wave noise is subtracted from the initial suppression result to complete the roll wave suppression process, thus solving the problem of roll wave residue and improving the accuracy of roll wave suppression. Attached Figure Description
[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0048] Figure 1 A flowchart illustrating an embodiment of the method for suppressing rolling waves provided in this application;
[0049] Figure 2 A schematic flowchart of Embodiment 2 of the method for suppressing rolling waves provided in this application;
[0050] Figure 3 A flowchart illustrating Embodiment 3 of the method for suppressing rolling waves provided in this application;
[0051] Figure 4 A schematic diagram of the test results for the method of suppressing rolling waves provided in this application;
[0052] Figure 5 A schematic diagram of the structure of the device for suppressing rolling waves provided in this application;
[0053] Figure 6 A schematic diagram of the structure of the equipment for suppressing rolling waves provided in this application.
[0054] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0055] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0056] This application is applicable to seismic exploration scenarios with complex surface conditions on land (such as mountains, deserts, and urban areas), as well as deep target exploration of complex geological structures such as carbonate rocks and salt domes. The technical solution is based on an unsupervised deep neural network with a U-net structure, combined with frequency band segmentation and morphological similarity analysis of seismic data. It separates low, medium, and high frequency signals through frequency division processing, uses the kinematic characteristics of the medium frequency signal as input, and generates an estimate of the effective low-frequency signal through nonlinear mapping. Finally, residual rolling waves are suppressed through a subtraction operation. This method can be integrated into the denoising module of the seismic data processing workflow, forming a joint processing chain with traditional frequency-wavenumber domain filtering (FK filtering).
[0057] In existing technologies, while FK filtering can initially suppress ground roll waves, residual noise still contaminates the effective signal due to low-frequency apparent velocity overlap. Sparse transform methods rely on single-domain features and are difficult to adapt to multi-scale signals. Data-driven methods (such as interferometry) have high requirements for data integrity and cannot handle scattered ground roll waves. Deep learning methods rely on labeled data, which limits their generalization ability. In complex geological scenarios, these problems lead to the masking of effective signals and distortion of imaging profiles, thereby affecting the accuracy of geological interpretation.
[0058] To address the aforementioned technical challenges, this application combines traditional FK filtering preprocessing with an unsupervised deep neural network based on frequency band morphological similarity. Leveraging the kinematic positional consistency (such as phase axis dip and curvature) and nonlinear mapping capabilities across different frequency bands of seismic data, it achieves comprehensive suppression of roll waves. Specifically, FK filtering initially separates roll waves from the effective signal. Then, using the morphological features of the mid-frequency effective signal as prior information, an unsupervised learning framework is constructed to map the mid-frequency signal into a low-mid-frequency effective signal, thereby correcting low-frequency residual roll waves. This approach eliminates the need for manually labeled data and is suitable for seismic data processing under complex geological conditions.
[0059] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0060] Figure 1 This is a schematic flowchart of an embodiment of the method for suppressing rolling waves provided in this application. Figure 1 As shown, the method includes:
[0061] S101. Perform frequency-wavenumber domain filtering preprocessing on the original seismic data to obtain the initial suppression results.
[0062] Raw seismic data refers to the unprocessed, original time-space domain record acquired by geophones during seismic exploration, containing effective reflection signals and various noises. In seismic exploration, raw seismic data *d* consists of the effective signal *s* and total roll wave noise *n*, i.e. .
[0063] Frequency-wavenumber domain filtering (FK filtering) is a signal processing technique that transforms seismic data from the time-space domain to the frequency-wavenumber domain, filters the effective signal and noise (such as roll waves) in terms of frequency and apparent velocity (wavenumber), and then transforms it back to the time-space domain.
[0064] The initial suppression result refers to the intermediate data obtained after the original seismic data has been preprocessed by FK filtering. Some of the roll waves that are significantly different from the apparent velocity of the effective signal have been suppressed, but some roll wave components that are close to or overlap with the apparent velocity of the effective signal still remain.
[0065] This step aims to utilize the macroscopic difference in apparent velocity between roll waves and the effective reflected signal to perform a first-stage coarse suppression. Specifically, the FK filtering method can suppress the portion of the roll wave in the original seismic data d that has a significant apparent velocity difference from the effective signal s. After processing using the FK filtering method, the result will have residual rolling wave noise. The initial suppression result of the effective signal s is expressed by the following formula:
[0066]
[0067] in, This represents the initial suppression result, where d represents the original seismic data. This indicates that some of the rolling wave noise can be suppressed by the FK filtering method. This represents the remaining roll wave noise in the total roll wave noise n that cannot be suppressed by the FK filtering method. Wherein, .
[0068] Based on the above parameters, the total roll wave noise can be expressed as: The raw seismic data can be represented as .
[0069] While traditional FK filtering methods can suppress some ground roll waves, direct application often leads to incomplete suppression or damage to the effective signal due to energy overlap between the ground roll waves and the effective signal in the FK domain. This step, as preprocessing, does not aim to completely suppress ground roll waves, but rather to significantly reduce the difficulty of subsequent processing. By removing "easily separable" noise, conditions are created for subsequent fine separation based on deep learning. This "staged processing" strategy is one of the key concepts of this invention, avoiding the limitations of a single method attempting to solve all problems.
[0070] Prior to this step, to ensure that the FK filtering preprocessing step can achieve the optimal initial suppression effect for seismic data under different work areas and acquisition parameters, and to avoid effective signal damage or insufficient noise suppression due to the use of a fixed threshold, this method preferably includes an adaptive filter parameter determination process before step S101. This process aims to automatically and objectively set the key parameter upon which FK filtering depends—the apparent velocity threshold. Its specific implementation steps are as follows:
[0071] First, raw seismic data was acquired to obtain the target seismic gathers to be processed. Then, the sliding window method was used to analyze the raw seismic data. This method divides the entire seismic data volume into multiple local windows in the time-space domain, and performs a two-dimensional Fourier transform on the data within each window to obtain the energy distribution of each local window in the frequency-wavenumber domain. This analysis reveals the local correlation characteristics between different frequency components and different apparent velocity (corresponding wavenumber) components in the data.
[0072] Based on the local energy distribution obtained from the above analysis, the apparent velocity difference range between the effective signal and the roll wave noise is identified and quantified. Specifically, in the frequency-wavenumber domain, the effective reflected signal typically manifests as an energy cluster concentrated in the higher apparent velocity region, while the roll wave noise manifests as a significant energy band concentrated in the lower apparent velocity region (especially at the low frequency end). By statistically analyzing the distribution results of multiple local windows (such as energy peak detection and cluster analysis), the lower limit of the apparent velocity of the main energy distribution of the effective signal and the upper limit of the apparent velocity of the main energy distribution of the roll wave noise can be quantitatively determined, thereby clarifying the difference range between the two in terms of apparent velocity.
[0073] Finally, based on the identified and quantified range of apparent velocity differences, the apparent velocity threshold for the FK filter is adaptively determined. This threshold is typically set within a safe transition band between the lower limit of the effective signal apparent velocity and the upper limit of the ground roll noise apparent velocity, for example, by taking a weighted average of the two or optimizing it according to the signal-to-noise ratio target. This apparent velocity threshold constitutes the core design parameter of the FK filter, essentially defining a velocity boundary in the frequency-wavenumber domain: during filtering, the components with the main energy located in the low apparent velocity region defined by this threshold (usually corresponding to ground roll waves) are suppressed or attenuated, while the components with the main energy located in the high apparent velocity region (usually corresponding to the effective reflected signal) are preserved.
[0074] This step abandons the crude approach of relying on experience to set fixed filter parameters. Instead, it automatically extracts the features of the data itself through a data-driven method, enabling the FK filter preprocessing to intelligently adapt to the variations in apparent velocity of roll waves caused by differences in surface structure and excitation and reception conditions in different exploration areas. This improves the robustness and generalization ability of the preprocessing step, providing a more stable and reasonable "initial suppression result" for subsequent unsupervised deep learning steps, thus laying a more reliable foundation for the final effect of the entire methodology.
[0075] The steps in this application preprocess the data by using frequency-wavenumber domain filtering. By utilizing the macroscopic difference in apparent velocity between the effective signal and the roll wave, the most easily separable noise in the original seismic data is quickly and initially suppressed, which significantly reduces the data complexity of subsequent processing and creates more favorable initial conditions for fine separation based on deep learning.
[0076] S102. Perform frequency band segmentation processing on the initial suppression results to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data.
[0077] Among them, frequency band segmentation refers to the process of using digital filters to decompose time-space domain seismic data into different subbands according to frequency components.
[0078] Intermediate frequency band sub-data refers to the sub-data separated from the initial suppression results, which mainly contains the effective reflected signal in the intermediate frequency band.
[0079] Low-to-intermediate frequency (LIFM) sub-data refers to sub-data composed of LIFM and IFFM sub-data superimposed on each other. It includes both effective low-frequency and IFFM signals as well as the main residual rolling noise.
[0080] Since the residual rolling wave noise mainly exists in the low-frequency portion of the initial suppression result, a multi-pass filter can be used in this step to divide the initial suppression result into low-frequency band suppression results. Mid-frequency band suppression results High-frequency band suppression results ;
[0081] Specifically, in this step, the low-pass filter is first determined. bandpass filter and high-pass filter ;
[0082] According to the high-pass filter The initial suppression results are filtered to obtain high-frequency band data. The initial suppression result is filtered using a bandpass filter to obtain the intermediate frequency (IF) sub-data. The initial suppression result is filtered using a low-pass filter to obtain low-frequency sub-band data. .
[0083] Among them, the high-frequency band data is Intermediate frequency band sub-data is Low-frequency band data is .
[0084] Based on the sub-data after frequency division, the initial suppression result obtained by S101 can be expressed as the sum of low-frequency band sub-data, mid-frequency band sub-data, and high-frequency band data, as shown in the following formula:
[0085]
[0086] On the other hand, the effective signal s is also filtered by a low-pass filter. bandpass filter and high-pass filter Segmented into low-frequency band effective signals Effective signal in the intermediate frequency band and high-frequency band effective signal At this point, the effective signal s can be expressed as .
[0087] Generally, roll waves are present in the initial suppression results of the low-frequency band. The initial suppression results of the mid-frequency band only contain the effective mid-frequency band signal. The initial suppression results of the high-frequency band only contain the effective high-frequency band signal. Based on the above discussion, the following formula is given:
[0088]
[0089]
[0090]
[0091] Finally, the intermediate frequency band data and the low frequency band data are summed to obtain the low-intermediate frequency band data.
[0092] Specifically, in the preprocessing results of the FK filtering method (initial suppression results), residual ground roll waves are still retained in the low-frequency range. That is, the residual ground roll waves contaminate the effective low-frequency signal. Because the residual ground roll waves overlap with the effective signal in both the spatiotemporal and frequency domains, it is difficult to completely attenuate and suppress them without damaging the effective signal. However, through frequency division processing, the mid-frequency sub-bands obtained from the initial suppression results do not contain residual ground roll waves and only contain the effective mid-frequency signal.
[0093] Furthermore, while the effective signals in the mid-frequency band and the low-frequency band differ significantly in frequency and amplitude, they share inter-band morphological similarity, meaning their kinematic characteristics, such as spatial location, are similar. To further enhance this inter-band morphological similarity, this application adds the effective signals in the low-frequency band and mid-frequency band to obtain the low-mid-frequency band effective signal. The formula is as follows:
[0094]
[0095] The effective signal in the low-to-intermediate frequency band (LIB) contains the effective signal in the intermediate frequency band (IF), thus exhibiting better inter-band morphological similarity between the two. Therefore, the final low-to-intermediate frequency band sub-data is obtained. for:
[0096]
[0097] That is, the low-to-intermediate frequency band sub-data includes the effective signal of the low-to-intermediate frequency band and the remaining ground roll wave.
[0098] Another implementation of this step includes:
[0099] The initial suppression result is subjected to wavelet transform to obtain multiple wavelet coefficient sub-bands of different scales; from the multiple wavelet coefficient sub-bands, the wavelet coefficient sub-band in the mid-frequency range is selected to obtain the mid-frequency band sub-data; from the multiple wavelet coefficient sub-bands, the wavelet coefficient sub-band in the low-frequency range is selected to obtain the low-frequency band sub-data; the mid-frequency band sub-data and the low-frequency band sub-data are added together to obtain the low-mid-frequency band sub-data.
[0100] Specifically, the initial suppression results are first subjected to wavelet transform. Wavelet transform decomposes seismic data in the time-space domain into a series of sub-bands of different scales, each sub-band corresponding to a different frequency range and temporal (or spatial) local characteristics, thus providing a more flexible and adaptable frequency band analysis capability for non-stationary signals than traditional fixed filters. Subsequently, from the multiple wavelet coefficient sub-bands obtained by the transform, those wavelet coefficient sub-bands corresponding to the mid-frequency range are selected according to a preset frequency division standard. By performing inverse wavelet transform (i.e., reconstruction) on these selected mid-frequency wavelet coefficient sub-bands, the mid-frequency band sub-data can be obtained, which mainly carries the effective reflection signal in the mid-frequency band. Similarly, wavelet coefficient sub-bands corresponding to the low-frequency range are selected and reconstructed to obtain low-frequency band sub-data, which contains the effective low-frequency signal and the main residual roll wave noise. Finally, the reconstructed mid-frequency band sub-data and the low-frequency band sub-data are superimposed to generate the low-mid-frequency band sub-data required for subsequent steps. This method leverages the advantages of wavelet transform in multi-resolution analysis, enabling more precise matching of the local time-frequency characteristics of seismic signals. Especially when dealing with seismic wavefields with non-stationary or transiently changing characteristics, it may achieve better frequency band separation than fixed filters, thus providing higher quality input and target data for subsequent unsupervised learning tasks.
[0101] The frequency band segmentation process in this application forms the core foundation of the unsupervised learning framework of this invention. By separating the pre-suppressed seismic data into sub-data of different frequency bands, and cleverly constructing a learning task pair with relatively clean mid-frequency band sub-data as input and noisy low-mid-frequency band sub-data as the optimization target, this step successfully transforms the challenging problem of "recovering clean signals from noisy data," which typically requires supervised labeling, into an unsupervised learning problem of "predicting cross-frequency band signals based on morphological similarity." This provides a clear and geophysically compliant optimization direction for subsequent neural network construction, ensuring that while suppressing noise, it can maximize the protection of the kinematic and dynamic characteristics of the effective reflected signals, achieving high-fidelity noise suppression.
[0102] S103. Based on the intermediate frequency band data and the pre-trained unsupervised deep neural network model, the low-intermediate frequency prediction signal is obtained.
[0103] Among them, the pre-trained unsupervised deep neural network model refers to a neural network model that has been trained using mid-frequency band sub-data as input and low-mid-frequency band sub-data as optimization target, thereby learning the mapping relationship from "mid-frequency effective signal" to "low-mid-frequency effective signal". In this embodiment, the U-Net structure is preferred.
[0104] Low-mid frequency (LIF) prediction signal refers to the signal output by a pre-trained neural network after forward propagation calculation, with the input of mid-frequency band sub-data. Its goal is to approximate the real effective low-mid frequency band signal.
[0105] In this step, feature extraction is first performed on the input mid-frequency band sub-data. Feature extraction refers to the pre-trained unsupervised deep neural network model (especially the encoder part) automatically learning and abstracting a high-level, discriminative data representation from the input data through multi-layer convolution and nonlinear activation operations. These learned features essentially capture the key kinematic characteristics of the effective mid-frequency band signal, such as the local tilt angle, curvature, continuity of the phase axis, and their spatial distribution patterns and structural relationships. These features characterize the propagation laws and spatial morphology of the effective reflected signal in underground structures.
[0106] Subsequently, the pre-trained unsupervised deep neural network enters the nonlinear mapping processing stage. This stage mainly corresponds to the decoder part of the network, which receives the previously extracted kinematic feature representations and, through a series of upsampling, deconvolution, and other operations, combines multi-scale detail information provided by the skip connections from the encoder to "transform" and "reconstruct" the kinematic features of these mid-frequency band signals. This "transformation" is a complex nonlinear function mapping, the goal of which is to generate a predicted signal that is spatially aligned with the input mid-frequency band signal, but extends to lower frequencies in terms of frequency components, and is consistent with it in terms of amplitude characteristics. Ultimately, the network output is the low-mid-frequency predicted signal, which is essentially the optimal estimate of the real low-mid-frequency band effective signal made by the network based on the learned mapping rule of "inferring the corresponding low-mid-frequency effective signal from the pure mid-frequency signal".
[0107] This application utilizes a pre-trained unsupervised deep neural network model to achieve accurate and adaptive computation of the physical concept of "signal prediction based on frequency band similarity." Through end-to-end deep learning, the network implicitly models the complex nonlinear mapping relationship between the mid-frequency effective signal and the low-mid-frequency effective signal, encompassing comprehensive adjustments to waveform, amplitude, and phase. Compared to traditional methods based on fixed transform bases or linear assumptions, this data-driven nonlinear mapping capability is more powerful and flexible, better handling complex wavefield interference and local variations in actual seismic data. This enables finer and more thorough separation of residual rolling waves while maximizing the preservation of low-frequency effective signals with similar kinematic characteristics to the mid-frequency signal, significantly improving the fidelity of the suppression process and the signal-to-noise ratio of the final signal.
[0108] S104. Calculate the difference between the low-frequency prediction signal and the low-frequency sub-band data to determine the rolling wave noise to be processed.
[0109] Among them, the unprocessed ground rolling noise refers to the residual ground rolling noise component that exists in the low-to-mid frequency band sub-data, estimated by difference calculation.
[0110] In this step, the low-to-intermediate frequency (LIFM) prediction signal generated by the unsupervised deep neural network in step S103 is directly compared with the LIFM sub-data obtained through frequency band segmentation in step S102. Here, the LIFM sub-data is a mixture containing the real low-frequency effective signal, the intermediate frequency (IF) effective signal, and residual ground rolling noise. The LIFM prediction signal is the prediction result estimated by the neural network based on the pure IF signal, theoretically containing only the effective signal component (i.e., the sum of the low-frequency and IF effective signals). By subtracting the latter from the former, i.e., subtracting the "estimated value of the pure signal component" from the "mixed observation of signal and noise," the mathematical residual corresponds to the portion of the data that cannot be explained by the effective signal prediction model. Within the physical and algorithmic framework of this application, this residual mainly corresponds to the ground rolling noise that needs to be suppressed. Therefore, this simple subtraction operation essentially completes a precise "blind source separation" of signal and noise, clearly separating the effective signal component learned by the neural network from the original mixed data, thereby obtaining an explicit estimate of the residual ground rolling noise.
[0111] This application achieves quantitative estimation of noise components through a single subtraction operation, making the entire method simple and stable. Secondly, this step transforms the "signal reconstruction capability" learned by the neural network in unsupervised learning, implicit in the parameters, into a clear and operable noise estimate, making subsequent suppression operations direct and reliable. Finally, since this noise estimation is based on the high-fidelity signal prediction results from the preceding steps, it minimizes the introduction of biases or damage to the effective signal during noise estimation, ensuring the purity and reliability of the final suppression result and laying a solid foundation for obtaining a high-quality effective signal.
[0112] S105. Subtract the rolling wave noise to be processed from the initial suppression result to complete the rolling wave suppression process.
[0113] Specifically, this step subtracts the roll wave noise to be processed, as determined in step S104, from the initial suppression result obtained in step S101. The initial suppression result is data preprocessed by FK filtering, which has removed most of the noise that differs significantly from the apparent velocity of the effective signal, but still contains residual roll waves that are difficult to separate using linear methods. The roll wave noise to be processed is a quantitative estimate of these residual noise components. Through this subtraction operation, it is equivalent to further and more precisely removing residual interference components from the initially purified data, thus finally obtaining the suppressed seismic data, which is mainly composed of effective reflection signals. This process completes a full closed loop from "noisy observation data" to "purified signal estimation".
[0114] The roll wave suppression method provided in this application first uses FK filtering for preliminary denoising to reduce problem complexity. Then, leveraging the inherent frequency band self-similarity of seismic signals, it takes a clean mid-frequency band signal as input and a noisy low-mid-frequency mixed signal as the learning target. This successfully constructs a self-supervised learning task that does not require manual labeling of "clean" data, eliminating reliance on unrealistic training data and improving the practicality and universality of the roll wave suppression method. The trained deep neural network can achieve a precise nonlinear mapping from the mid-frequency signal to the effective low-mid-frequency signal. Through difference calculation and signal reconstruction, it ultimately achieves more thorough and higher-fidelity suppression of roll waves, effectively solving the technical problems of excessive residual noise in traditional methods and poor applicability of supervised deep learning methods. This significantly improves the signal-to-noise ratio and overall quality of onshore seismic data processing.
[0115] Figure 2 This is a schematic flowchart of Embodiment 2 of the method for suppressing rolling waves provided in this application. Figure 2 As shown, in Figure 1 Based on the examples, the process of constructing a pre-trained unsupervised neural network model is further illustrated. This method includes:
[0116] S201. Determine the initial unsupervised deep neural network model.
[0117] In this context, the initial unsupervised deep neural network model refers to a neural network model whose network structure has been completed before training begins, but whose internal parameters (such as convolutional kernel weights, biases, etc.) have not yet been learned from data and are typically initialized randomly or according to specific rules. This embodiment preferably uses a U-Net network with an encoder-decoder structure and skip connections as the basic model architecture.
[0118] This step marks the beginning of model training. First, based on the characteristics of the seismic signal mapping task to be solved, a suitable deep neural network architecture is selected and built. This architecture needs to possess powerful feature extraction and signal reconstruction capabilities. After determining the network structure, all its weights and bias parameters are initialized to prepare for the subsequent data-driven optimization process.
[0119] S202. Based on the mid-frequency band sub-data, train the initial unsupervised deep neural network model to obtain mid-frequency band training data.
[0120] The mid-frequency band training data refers to the actual set of input samples used to train the neural network model. In this method, it specifically refers to the mid-frequency band sub-data obtained from actual seismic data through the aforementioned steps.
[0121] This step aims to prepare the input data required for model training. Specifically, using the method in step S102 of Example 1, one or more batches of actual seismic data are processed to generate corresponding mid-frequency band sub-data. These mid-frequency band sub-data constitute the input part of the training dataset. As the "source signals" for network learning, their purity (mainly containing effective reflection signals) is the basis for the network to learn the correct mapping relationship.
[0122] S203. Construct a loss function based on the mid-frequency band training data and the low-mid-frequency band sub-data.
[0123] In machine learning, the loss function is a function used to quantify the difference between the model's predicted output and the desired target. It provides the correct optimization direction for the pre-defined unsupervised deep neural network, ensuring that the pre-defined unsupervised deep neural network is trained to obtain optimal unsupervised deep neural network parameters.
[0124] This step is crucial for constructing the unsupervised learning task, defining the objective of model optimization. The mid-frequency band training data obtained in step S202 is used as the input to the neural network. Simultaneously, the low-mid-frequency band sub-data (obtained from step S102 in Example 1), generated by pairing with the mid-frequency band training data, is used as the "target" or "pseudo-label" that the model aims to approximate during training. Based on this input-target pair, a specific loss function is constructed. The core of this loss function typically employs forms such as the L1 norm, directly measuring the overall difference between the predicted signal output by the neural network and the low-mid-frequency band sub-data. Furthermore, to improve the model's stability and generalization ability, a regularization term is usually added to the loss function.
[0125] Since the low-to-mid frequency band sub-data contains both valid signals and noise, and the input mid-frequency band data has kinematic similarity to the valid signal components therein, minimizing this loss function will drive the network to learn how to extract and reconstruct components from the input that match the valid signals in the target data, rather than fitting irrelevant noise.
[0126] For example, in order to correct the effective signal in the intermediate frequency band The frequency and phase, and the effective signal in the intermediate frequency band. The correct mapping is the estimated value of the effective signal in the low-to-mid frequency band. This application proposes the following loss function, as shown in the formula below:
[0127]
[0128] in, Represents the regularization coefficient. Effective signal in the mid-frequency band. While there are differences in frequency and amplitude between the effective low- and mid-frequency band signals and the initial suppression results, their kinematic positions are similar. Furthermore, when the input data and target seismic data have similar kinematic positions, a pre-trained unsupervised deep neural network can correct the amplitude and frequency of the input data to match the amplitude and frequency of the target seismic data. Under the constraint of the loss function, the unsupervised deep neural network can correctly map the effective mid-frequency band signal to the predicted effective low- and mid-frequency band signal.
[0129] S204. Based on the loss function, train the initial unsupervised deep neural network model to obtain a pre-trained unsupervised deep neural network model.
[0130] This step involves performing model training and optimization. Specifically, based on the loss function constructed in S203, the initial unsupervised deep neural network model determined in step S201 is iteratively trained using the backpropagation algorithm and a gradient descent optimizer (such as Adam).
[0131] In each iteration, the mid-frequency band training data is input into the network to obtain the predicted output. The loss value between the predicted output and the corresponding low-mid-frequency band sub-data is calculated, and then all parameters in the network are adjusted in reverse based on the loss value.
[0132] Through numerous iterations, the network parameters are gradually optimized, causing the loss function value to continuously decrease. When the training process reaches the preset stopping condition (such as reaching the maximum number of iterations), a pre-trained unsupervised deep neural network model is obtained, which is capable of predicting effective low-mid frequency signals from mid-frequency signals.
[0133] This application's embodiments elaborate on how to utilize physically correlated mid-frequency and low-mid-frequency sub-data pairs generated from actual seismic data to construct a specific unsupervised loss function and train a deep neural network, fully revealing the technical path for obtaining a pre-trained model. This training mechanism not only enables the direct application of deep learning methods in actual exploration scenarios, solving the fundamental problem of limited applicability caused by supervised methods relying on synthetic or traditional methods to generate labels, but also learns signal conversion laws that are more complex and adaptable to changes in actual data through a data-driven approach than fixed algorithms. This provides a reliable core model guarantee for ultimately achieving high-fidelity, high-precision roll wave suppression.
[0134] Figure 3 A schematic diagram of the structure of the pre-trained unsupervised deep neural network model provided in this application. (See diagram below.) Figure 3As shown in the figure, this embodiment elaborates on the specific architecture and data flow of the pre-trained unsupervised deep neural network model. The model preferably adopts the U-Net network structure, which specifically uses an "encoder-decoder" structure to learn and extract seismic data features.
[0135] The encoder-decoder architecture is a deep neural network architecture commonly used for tasks such as image segmentation and signal reconstruction. The encoder part progressively compresses the input data through downsampling to extract high-level abstract features; the decoder part progressively restores the data size through upsampling and uses the features extracted by the encoder for accurate reconstruction.
[0136] like Figure 3 As shown, the pre-trained unsupervised deep neural network model (referred to as network F in this embodiment) adopts the U-Net architecture, the core of which lies in the symmetrical encoder-decoder path and the skip connections connecting the two.
[0137] The left half of the network is the encoding process, which is responsible for extracting multi-level feature representations from the input data. The input data (i.e., mid-frequency band sub-data) first passes through a series of feature extraction modules consisting of convolutional layers and activation functions. Specifically, each module may contain multiple consecutive convolutional layers, and each convolutional layer is typically followed by a ReLU activation function to introduce non-linearity. Subsequently, the feature map is downsampled through max pooling layers, halving its spatial dimensions (e.g., time-channel dimension) while increasing the number of feature channels to capture more abstract features. This combination of "convolution-activation-pooling" is repeated multiple times in the encoding path, forming a hierarchical feature pyramid, thereby progressively extracting and compressing the kinematic and morphological features of the input signal from local details to global structure.
[0138] For example, seismic data feature information is extracted using operations such as convolution and max pooling. Applying three 3×3 convolutional layers and one 2×2 max pooling layer to the input seismic data halves the data size. During encoding, a ReLU activation function is used after each convolutional layer to enhance the network's nonlinear mapping capability.
[0139] Secondly, the right half of the network is the decoding path, which is responsible for reconstructing the target signal based on the features extracted by the encoder. The decoding path is symmetrical to the encoding path and mainly consists of upsampling layers and convolutional layers. The upsampling layers gradually restore the feature map size to the original input size. After upsampling, skip connections are used to concatenate the feature maps of layers at the same scale in the encoding path with the feature map of the current decoding layer. This key design allows the decoder to simultaneously utilize the semantic information provided by deep features (such as the overall shape of the signal) and the fine spatial information preserved by shallow features (such as the precise position and edges of the phase axis) when reconstructing the signal, thereby achieving high-precision reconstruction of the output signal. The ReLU activation function is also used in the convolutional layers of the decoding path. Only in the last convolutional layer of the network, based on the characteristics of the target data, is Tanh selected as the activation function to constrain the final output value within a suitable dynamic range.
[0140] For example, upsampling layers can increase the size of the feature maps, ultimately restoring them to the original image size. During decoding, the last convolutional layer uses the Tanh function, while other convolutional layers also use ReLU as the activation function. U-Net also connects the feature maps between the encoder and decoder through Concat connections (also known as skip connections), which better preserves the detailed information of the seismic data, facilitating the extraction of data features from the input data and mapping them to the correct output data.
[0141] Specifically, the pre-trained unsupervised deep neural network model U-net can be called... The input data of the pre-trained unsupervised deep neural network is denoted as x, and the output data is denoted as y. In this embodiment, after the parameters of the deep neural network are obtained through training... The nonlinear mapping relationship of a pre-trained unsupervised deep neural network is as follows:
[0142]
[0143] Frequency segmentation of the initial suppression result yields the effective mid-frequency signal. In the training process of pre-trained unsupervised deep neural networks, the effective signal in the mid-frequency band... Yes, it is available and used as input data. After the parameters of the pre-trained unsupervised deep neural network are trained, the output data of the pre-trained unsupervised deep neural network is an estimate of the effective signal in the low and mid-frequency bands. The formula is as follows:
[0144]
[0145] At this point, the residual roll wave estimated by the pre-trained unsupervised deep neural network is:
[0146]
[0147] In addition, the loss function proposed in Example 2 can ensure that the parameters of the pre-trained unsupervised deep neural network are optimized in the correct direction. For the specific loss function construction process, please refer to Example 2.
[0148] The pre-trained unsupervised deep neural network structure proposed in this embodiment provides an efficient and reliable implementation platform for the unsupervised learning task of this invention. Its encoder-decoder structure excels at capturing the complex mapping relationship between input and output data, making it suitable for the conversion task in this application, which requires maintaining precise spatial correspondence from "intermediate frequency signal" to "low intermediate frequency signal." The introduction of skip connections is crucial, effectively solving the problem of lost detail information in deep networks. It ensures that the network can retain and utilize key kinematic details of the effective reflected signal (such as the continuity of tilt angle and curvature) during learning and prediction, thereby achieving high-fidelity reconstruction of the effective signal features while suppressing noise. The powerful feature learning and reconstruction capabilities of this network structure are one of the core technological foundations for the successful implementation of unsupervised signal prediction based on frequency band similarity and ultimately achieving high-precision rolling wave suppression in this invention.
[0149] Figure 4 A schematic diagram illustrating the test results of the method for suppressing rolling waves provided in this application. Figure 4 The above, Figure 4 A shows the original data, in which surface waves mainly exist within the triangular region enclosed by straight lines. Surface waves within this triangular region exhibit strong capabilities, characterized by high amplitude, low frequency, and low apparent velocity, severely interfering with the effective signal.
[0150] Figure 4 b shows all the effective signals separated by the pre-trained unsupervised deep neural network method proposed in this application.
[0151] Figure 5 This is a schematic diagram of the structure of the earth rolling wave suppression device provided in this application. Figure 5 As shown, the ground rolling wave suppression device 50 provided in this embodiment includes:
[0152] The initial suppression module 501 for rolling waves is used to perform frequency-wavenumber domain filtering preprocessing on the raw seismic data to obtain the initial suppression result.
[0153] The frequency band segmentation module 502 is used to perform frequency band segmentation processing on the initial suppression result to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data.
[0154] The signal prediction module 503 is used to obtain the low-intermediate frequency prediction signal based on the intermediate frequency band sub-data and the pre-trained unsupervised deep neural network model. The pre-trained unsupervised deep neural network model is trained based on the intermediate frequency band data and the low-intermediate frequency band data.
[0155] The roll wave noise determination module 504 is used to perform difference calculation on the low-intermediate frequency prediction signal and the low-intermediate frequency band sub-data to determine the roll wave noise to be processed.
[0156] The ground roll wave suppression module 505 is used to subtract the ground roll wave noise to be processed from the initial suppression result to complete the ground roll wave suppression process.
[0157] In one possible implementation, the wave-suppressing device 50 is further specifically used for:
[0158] Determine the initial unsupervised deep neural network model;
[0159] Based on the mid-frequency band sub-data, the initial unsupervised deep neural network model is trained to obtain mid-frequency band training data;
[0160] A loss function is constructed based on the mid-frequency band training data and the low-mid-frequency band sub-data;
[0161] Based on the loss function, the initial unsupervised deep neural network model is trained to obtain a pre-trained unsupervised deep neural network model.
[0162] In one possible implementation, the signal prediction module 503 is further specifically used for:
[0163] Feature extraction processing is performed on the mid-frequency band sub-data to generate mid-frequency band kinematic features;
[0164] Based on a pre-defined unsupervised deep neural network model, the mid-frequency kinematic features are nonlinearly mapped to obtain the low-mid-frequency prediction signal.
[0165] In one possible implementation, the frequency band segmentation module 502 is further specifically used for:
[0166] Determine the low-pass filter and the band-pass filter;
[0167] The initial suppression result is filtered using a bandpass filter to obtain intermediate frequency sub-band data;
[0168] The initial suppression result is filtered using a low-pass filter to obtain low-frequency sub-band data;
[0169] The mid-frequency and low-frequency sub-data are summed to obtain the low-mid-frequency sub-data.
[0170] In one possible implementation, the frequency band segmentation module 502 is further specifically used for:
[0171] Wavelet transform is applied to the initial suppression result to obtain multiple wavelet coefficient sub-bands at different scales;
[0172] From multiple wavelet coefficient sub-bands, wavelet coefficient sub-bands in the mid-frequency range are selected to obtain mid-frequency band sub-data;
[0173] From multiple wavelet coefficient sub-bands, select the wavelet coefficient sub-band in the low-frequency range to obtain the low-frequency band sub-data;
[0174] The intermediate frequency (IF) band data is added to the low frequency (LFM) band data to obtain the low-IF band data.
[0175] In one possible implementation, the initial suppression module 501 for the rolling wave is further specifically used for:
[0176] Collect raw seismic data;
[0177] The local energy distribution of the original seismic data is obtained by analyzing the original seismic data using the sliding window method.
[0178] Based on the local energy distribution of the original seismic data, the apparent velocity difference between the effective signal and the roll wave noise in the original seismic data is determined.
[0179] Based on the apparent velocity difference, the apparent velocity threshold for frequency-wavenumber domain filtering is determined. The apparent velocity threshold is a velocity threshold parameter used to distinguish between effective signals and roll wave noise during the frequency-wavenumber domain filtering preprocessing of the original seismic data.
[0180] In one possible implementation, the pre-trained unsupervised deep neural network model is a U-Net network structure.
[0181] The earth rolling wave suppression device provided in this embodiment can perform the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.
[0182] Figure 6 This is a schematic diagram of the structure of the equipment for suppressing rolling waves provided in this application. Figure 6 As shown, the rolling wave suppression device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the device 60 also includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus 604.
[0183] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.
[0184] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.
[0185] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.
[0186] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.
[0187] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.
[0188] This application also provides a system for suppressing ground rolling waves, including a computer program that, when executed by a processor, implements the above-described method.
[0189] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0190] If a function is implemented as 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 this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0191] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0192] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.
Claims
1. A method for suppressing ground rolling waves, characterized in that, include: The original seismic data were preprocessed by frequency-wavenumber domain filtering to obtain the initial suppression results; The initial suppression results are subjected to frequency band segmentation processing to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data; Based on the mid-frequency band sub-data and the pre-trained unsupervised deep neural network model, a low-mid-frequency prediction signal is obtained. The pre-trained unsupervised deep neural network model is trained based on the mid-frequency band data and the low-mid-frequency band data. The difference between the low-frequency prediction signal and the low-frequency band sub-data is calculated to determine the rolling wave noise to be processed; The rolling wave noise to be processed is subtracted from the initial suppression result to complete the rolling wave suppression process.
2. The pressing method according to claim 1, characterized in that, The method further includes: Determine the initial unsupervised deep neural network model; Based on the mid-frequency band sub-data, the initial unsupervised deep neural network model is trained to obtain mid-frequency band training data; Based on the mid-frequency band training data and the low-mid-frequency band sub-data, a loss function is constructed; Based on the loss function, the initial unsupervised deep neural network model is trained to obtain a pre-trained unsupervised deep neural network model.
3. The pressing method according to claim 1, characterized in that, Based on the mid-frequency band sub-data and the pre-trained unsupervised deep neural network model, the low-mid-frequency predicted signal is obtained, including: The mid-frequency band sub-data is subjected to feature extraction processing to generate mid-frequency band kinematic features; Based on a preset unsupervised deep neural network model, the mid-frequency band kinematic features are subjected to nonlinear mapping processing to obtain a low-mid-frequency prediction signal.
4. The pressing method according to any one of claims 1 to 3, characterized in that, The initial suppression result is subjected to frequency band segmentation processing to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data, including: Determine the low-pass filter and the band-pass filter; The initial suppression result is filtered using the bandpass filter to obtain intermediate frequency sub-data. The initial suppression result is filtered using the low-pass filter to obtain low-frequency sub-band data. The mid-frequency sub-data and low-frequency sub-data are summed to obtain low-mid-frequency sub-data.
5. The pressing method according to any one of claims 1 to 3, characterized in that, The initial suppression result is subjected to frequency band segmentation processing to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data, including: The initial suppression result is subjected to wavelet transform to obtain multiple wavelet coefficient subbands of different scales; From the multiple wavelet coefficient sub-bands, wavelet coefficient sub-bands in the mid-frequency range are selected to obtain mid-frequency band sub-data; From the multiple wavelet coefficient sub-bands, select the wavelet coefficient sub-band in the low-frequency range to obtain the low-frequency band sub-data; The intermediate frequency band data is added to the low frequency band data to obtain the low-intermediate frequency band data.
6. The pressing method according to any one of claims 1 to 3, characterized in that, Before performing frequency-wavenumber domain filtering preprocessing on the raw seismic data to obtain the initial suppression results, the method further includes: Collect raw seismic data; The original seismic data is analyzed using the sliding window method to obtain the local energy distribution of the original seismic data; Based on the local energy distribution of the original seismic data, determine the apparent velocity difference between the effective signal and the roll wave noise in the original seismic data; Based on the apparent velocity difference, the apparent velocity threshold for frequency-wavenumber domain filtering is determined. The apparent velocity threshold is a velocity threshold parameter used to distinguish between the effective signal and the roll wave noise during the frequency-wavenumber domain filtering preprocessing of the original seismic data.
7. The pressing method according to any one of claims 1 to 3, characterized in that, The pre-trained unsupervised deep neural network model is a U-Net network structure.
8. A device for suppressing rolling waves, characterized in that, include: The initial suppression module for rolling waves is used to perform frequency-wavenumber domain filtering preprocessing on the raw seismic data to obtain the initial suppression results. The frequency band segmentation module is used to perform frequency band segmentation processing on the initial suppression result to obtain intermediate frequency band sub-data and low intermediate frequency band sub-data; The signal prediction module is used to obtain a low-intermediate frequency predicted signal based on the intermediate frequency band sub-data and the pre-trained unsupervised deep neural network model, wherein the pre-trained unsupervised deep neural network model is trained based on the intermediate frequency band data and the low-intermediate frequency band data. The rolling wave noise determination module is used to calculate the difference between the low-frequency prediction signal and the low-frequency band sub-data to determine the rolling wave noise to be processed. The rolling wave suppression module is used to subtract the rolling wave noise to be processed from the initial suppression result to complete the rolling wave suppression process.
9. A device for suppressing rolling waves, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.
10. A system for suppressing rolling waves, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-7.