A seismic mixed noise suppression reconstruction method

Abstract

The present application belongs to the technical field of seismic exploration data processing, and specifically discloses a method for suppressing and reconstructing seismic aliasing noise, comprising: obtaining common shot gather seismic data containing aliasing noise, and converting the common shot gather seismic data into non-shot gather domain data; obtaining structural continuity information of the non-shot gather domain data; using a first-stage network to suppress main aliasing noise in the non-shot gather domain data according to the structural continuity information, and obtaining first-stage reconstruction data; converting the first-stage reconstruction data into the common shot gather domain; using a second-stage network to suppress residual aliasing noise and reconstruct effective signals for the converted first-stage reconstruction data according to the structural continuity information, and obtaining target reconstruction data; and having the following advantages: through two-domain and two-stage collaborative processing and structural continuity guidance, aliasing noise is suppressed and effective signals are reconstructed in stages, complex structure damage is reduced, and data fidelity is improved.

Description

Technical Field

[0001] This invention relates to the field of seismic exploration data processing technology, and more specifically, to a method for suppressing and reconstructing seismic aliasing noise. Background Technology

[0002] As seismic exploration depth and acquisition scale continue to increase, sparse sampling or high-efficiency excitation methods are often employed to improve acquisition efficiency. This can easily lead to aliasing of different seismic wavefields in space and time, resulting in aliasing noise. This noise superimposes with the effective reflection signal, affecting velocity modeling, migration imaging, and reservoir interpretation. Existing methods mainly include traditional methods such as FK filtering, Tau-P transform, and Curvelet transform, as well as deep learning methods based on convolutional neural networks or Transformers. Traditional methods rely on artificial parameters and have limited adaptability to complex structures; existing deep learning methods mostly process data in a single data domain, failing to fully utilize the structural differences between different gather domains, which can easily lead to damage to the effective signal, retention of residual noise, and insufficient recovery of weak reflection structures. Summary of the Invention

[0003] The present invention aims to provide a method for suppressing and reconstructing seismic aliasing noise, in order to solve or improve the problem of the above-mentioned technical issues that it is difficult to protect the effective reflection signals of complex structural areas while taking into account the aliasing noise suppression effect and the accuracy of seismic data reconstruction.

[0004] In view of this, a first aspect of the present invention is to provide a method for suppressing and reconstructing earthquake aliasing noise.

[0005] A first aspect of the present invention provides a method for suppressing and reconstructing earthquake aliasing noise, comprising the following steps: Acquire common shot gathering seismic data containing aliasing noise and convert the common shot gathering seismic data into non-shot gathering domain data; acquire structural continuity information of the non-shot gathering domain data; use a first-stage network to suppress the main aliasing noise in the non-shot gathering domain data according to the structural continuity information to acquire first-stage reconstruction data; convert the first-stage reconstruction data to the common shot gathering domain; use a second-stage network to suppress residual aliasing noise and reconstruct effective signals on the converted first-stage reconstruction data according to the structural continuity information to acquire target reconstruction data.

[0006] The beneficial effects of this invention compared to the prior art are as follows: By constructing a two-stage network that sequentially and collaboratively processes non-shot domain and common shot domain, the main aliasing noise can be suppressed by utilizing the strong continuity of the phase axis in the non-shot domain, and the residual aliasing noise can be further suppressed by utilizing the seismic wave propagation characteristics in the common shot domain, thus improving the noise residue caused by single data domain processing. By introducing structural continuity information in two processing stages, the network is guided to identify and protect effective reflection events, thereby reducing the impairment of effective signals in complex structural regions. By reconstructing effective signals through network synchronization in the second stage, the problem of insufficient recovery of fracture phase axes and weak reflection structures can be improved, thereby enhancing the noise suppression accuracy and structural fidelity of seismic data.

[0007] Additional aspects and advantages of embodiments of the invention will become apparent in the following description or may be learned by practice of embodiments of the invention. Attached Figure Description

[0008] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 A schematic flowchart of the earthquake aliasing noise suppression and reconstruction method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the earthquake aliasing noise suppression and reconstruction system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0009] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0010] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0011] Please see Figures 1-3 The following describes a method, system, electronic device, and medium for suppressing and reconstructing earthquake aliasing noise according to some embodiments of the present invention.

[0012] An embodiment of the first aspect of the present invention provides a method for earthquake aliasing noise suppression and reconstruction. In some embodiments of the present invention, such as... Figure 1 As shown, the method includes: S101: Acquire common shot gathering seismic data containing aliasing noise, and convert the common shot gathering seismic data into non-shot gathering domain data.

[0013] In some implementations, before converting the original co-shot gather seismic data to the non-shot gather domain, the original co-shot gather seismic data undergoes amplitude equalization, outlier removal, and normalization sequentially. A training sample set and a test sample set are then constructed based on the processed seismic data. The training sample set is used to train the first-stage network and the second-stage network, while the test sample set is used to test the trained network to determine its aliasing noise suppression and reconstruction results for seismic data not used in training.

[0014] In some implementations, the non-shot gather domain may also be an Offset domain, Angle domain, Tau-P domain, FK domain, or Curvelet domain, or other data domains that can enhance the structural continuity of the effective seismic signal and make it present a different distribution from aliasing noise.

[0015] Here, common-shot gather seismic data containing aliasing noise is acquired, and the seismic traces are rearranged according to their acquisition locations and gather relationships, converting the common-shot gather seismic data into target non-shot gather domain data. Specifically, the target non-shot gather domain can be a common centroid domain, a common reflection point domain, or a common receiver point domain. The gather conversion changes the arrangement relationship between the seismic traces, making the effective signals related to the same underground reflection structure exhibit strong in-phase axis continuity in the non-shot gather domain, while the aliasing noise exhibits dispersion or weak continuity due to the inconsistent arrangement relationship with the target gather, thereby increasing the difference in structural characteristics between the effective signal and the aliasing noise.

[0016] In this embodiment, the common shot gather seismic data and the transformed CMP domain data are represented as follows:

[0017] in, This represents preprocessed, co-shot gathering seismic data awaiting gather conversion. This represents the CMP field data obtained after gather conversion; Position variable representing the direction of the track number. The position variable represents the time direction; shot and cmp are used to identify the common shot set and the CMP field, respectively.

[0018] As can be seen from the above, unlike the direct single noise suppression in the common shot domain, this invention constructs non-shot domain data that is more conducive to characterizing the continuity of the reflection phase axis, which is used to extract structural continuity information and suppress the main aliasing noise, while retaining the data transformation relationship for residual noise processing when returning to the common shot domain.

[0019] S102, obtain structural continuity information of non-shot gather domain data.

[0020] Here, the non-shot gather domain data is decomposed into multiple scales, and structural attributes such as local dip angle, continuity, curvature, and directional consistency are extracted from the decomposed data at each scale. These structural attributes from different scales are then fused to form a multi-scale structural feature map, which is used as information on structural continuity. The decomposed data at different scales respectively characterize local reflection details and overall structural morphology, enabling the structural continuity information to simultaneously reflect the structural features of weak reflection phase axes, fault boundaries, and large-scale reflection layers.

[0021] As mentioned above, effective seismic reflection signals typically possess relatively stable extension directions and continuously changing structural morphologies, while aliasing noise after gather conversion exhibits characteristics such as directional dispersion and weak continuity. By comprehensively characterizing the direction, continuity, and curvature variations of the phase axis, reliable reflection structures in non-shot gather domain data can be identified. This differs from distinguishing signals from noise solely based on amplitude, global correlation, or abstract features within the network, as explicit prior information is obtained from the geometric structure of the seismic reflection phase axis.

[0022] Specifically, the steps for obtaining structural continuity information of non-shot gather domain data include: Multi-scale decomposition of non-shot gather domain data is performed using multiple different scale parameters to obtain scale decomposition data corresponding to each scale parameter.

[0023] Structural attributes are extracted from data decomposed at various scales, and structural continuity information is obtained by fusing these attributes.

[0024] To simultaneously acquire shallow details and deep structural information, CMP data is decomposed into fine-scale, meso-scale, and large-scale structures, based on the specific description above. This allows for the identification of local phase axes, fault boundaries, and large-scale reflection layers, respectively. Subsequently, a structural tensor is constructed using the gradients of seismic data at each scale along both spatial and temporal directions. The local principal directions are then obtained through eigenvalue decomposition of the structural tensor, leading to the calculation of dip properties and the resulting dip map.

[0025] Furthermore, continuity properties are calculated using structural tensor eigenvalues ​​to obtain a continuity map, where the effective reflective layer exhibits high continuity, while aliasing noise shows low continuity. Curvature features are calculated using the tilt field to obtain a curvature map, used to identify faults, folds, pinch-out bodies, and complex structures. Orientation distribution is statistically analyzed within local windows of 7×7, 11×11, and 15×15 to obtain orientation maps, where reflected waves exhibit consistent direction, while aliasing noise exhibits random direction. Thus, tilt, continuity, curvature, and orientation features are obtained, and structural continuity information is derived by fusing these four types of features.

[0026] Specifically, the steps of multi-scale decomposition of non-shot gather domain data using multiple different scale parameters include: Determine the Gaussian kernel corresponding to each scale parameter.

[0027] By convolving the non-shot domain data with Gaussian kernels, scale decomposition data corresponding to each scale parameter are obtained.

[0028] Regarding the specific description above, when performing multi-scale decomposition on CMP data, Gaussian kernels with different scale parameters are used to construct the scale space. For example, the scale parameters can be set to 1, 2, 4 and 8 in sequence.

[0029] The data at each scale are obtained according to the following formula:

[0030] in, Indicates the first The scaling decomposition data corresponding to each scaling parameter. Indicates scale index; The scale parameter is represented as Gaussian kernel, Indicates the first The scaling parameters of a Gaussian kernel; symbol This represents the convolution operation. This is an example of a four-level scale parameter setting.

[0031] As described above, for any given scale parameter, this scale parameter is used as the scale control variable for the corresponding Gaussian kernel. The weight distribution at each position within the Gaussian kernel is determined based on this scale parameter, thus obtaining the Gaussian kernel corresponding to that scale parameter. Gaussian kernels corresponding to different scale parameters have different weight distributions and ranges of influence, thereby ensuring a one-to-one correspondence between each Gaussian kernel and each scale parameter.

[0032] For each Gaussian kernel, using each sampling position in the non-shot domain data as the center, the data at adjacent sampling positions are weighted and summed using this Gaussian kernel, and the weighted result is used as the convolution result for the corresponding sampling position. By traversing each sampling position in the non-shot domain data along the trace number direction and the time direction, the scale decomposition data corresponding to this Gaussian kernel is obtained.

[0033] Convolution is performed using each Gaussian kernel in the manner described above to obtain multiple scale decomposition data that correspond one-to-one with each scale parameter. The sampling positions of each scale decomposition data and the non-shot domain data are kept in the trace number direction and time direction, so that structural attributes can be extracted and fused from the same sampling position later.

[0034] Specifically, the steps for extracting structural attributes from data decomposed at various scales include: Based on the gradients along the channel number and time directions of the data decomposed at each scale, a structure tensor is constructed, and eigenvalue decomposition is performed on each structure tensor to obtain the local tilt angle and continuity in the structure properties.

[0035] Based on the changes in local tilt angle and the distribution of local tilt angle directions within a preset local window, the curvature and orientation consistency in the structural properties are obtained respectively.

[0036] Based on the specific description above, for each scale of data decomposition, the gradients along the channel number direction and the time direction at each sampling location are calculated. The square of the gradient along the channel number direction, the product of the gradient along the channel number direction and the gradient along the time direction, and the square of the gradient along the time direction are used as tensor components to construct a structure tensor corresponding to each sampling location. Eigenvalue decomposition is performed on each structure tensor to obtain two eigenvalues ​​and corresponding eigenvectors, and the local principal direction eigenvector is determined from the eigenvectors. The local dip angle is obtained by performing an arctangent operation based on the ratio between the component of the local principal direction eigenvector along the time direction and its component along the channel number direction.

[0037] Regarding the first The gradients in the channel direction and the time direction of the data are as follows:

[0038] The corresponding structure tensor is:

[0039] After performing eigenvalue decomposition on the structure tensor, the local tilt angle is determined according to the following formula:

[0040] in, express The gradient along the track number direction, express Gradient along the time direction, Represents partial differential operations; This represents the structure tensor composed of the gradients described above; and They represent the first The components of the local principal direction eigenvector at each scale in the channel direction and time direction; Indicates the first Local tilt angles at various scales This indicates the arctangent operation.

[0041] Continuity is obtained by considering the difference between the larger and smaller eigenvalues ​​in the structure tensor, and the normalization relationship formed by the sum of the two eigenvalues ​​and a preset stability term. The closer the continuity is to 1, the stronger the structural continuity at the corresponding sampling position; the closer the continuity is to 0, the weaker the structural continuity at the corresponding sampling position.

[0042] The continuity property is determined by the following formula:

[0043] in, Indicates the first Continuity at various scales; and They represent The larger and smaller eigenvalues; This represents a pre-defined positively stable term, used to avoid a denominator of zero. The closer it is to 1, the stronger the structural continuity at the corresponding position; The closer it is to 0, the weaker the structural continuity at the corresponding position.

[0044] The variations of local dip angle along the track number direction and the time direction are determined separately, and the variations in these two directions are combined to obtain curvature. For directional consistency, preset local windows of 7×7, 11×11, or 15×15 are established centered on each sampling location, and multiple local dip angles within these windows are obtained. The cosine components of each local dip angle are averaged to obtain the directional distribution of the local dip angles, and the cosine and sine components of each local dip angle are accumulated separately. The directional consistency is obtained based on the combined amplitude of the two accumulated results. Thus, local dip angle, continuity, curvature, and directional consistency are obtained from the data at each scale decomposition.

[0045] The curvature is determined by the following formula:

[0046] The distribution and consistency of local dip angle directions are determined by the following formulas:

[0047]

[0048] in, Indicates the first The curvature is obtained by combining the changes in local dip angle along the track number direction and the time direction at various scales; Indicates the first Distribution of local tilt angle directions within a preset local window at each scale; This indicates the number of sampling points included in the statistics within this local window; Indicates the index of the sampling point within the local window; Indicates the first Within this local window at the specified scale, the first... Local tilt angle of each sampling point; Indicates the first The directional consistency is determined by the combined amplitude of the cosine and sine components of the local tilt angle at each scale.

[0049] Specifically, the steps for obtaining structural continuity information by fusing structural properties include: By integrating local tilt angles, continuity, curvature, and directional consistency at various scales, a multi-scale structural feature map is obtained as information on structural continuity.

[0050] Regarding the specific description above, since the decomposition data at each scale all originate from the same non-shot gather domain data, the local dip angle, continuity, curvature, and directional consistency extracted from the decomposition data at each scale have corresponding trace number positions and time positions. Using the same trace number position and time position as the correspondence, the local dip angle, continuity, curvature, and directional consistency at different scales are fused to obtain the fused dip angle features, continuity features, curvature features, and directional features.

[0051] The four types of features after fusion are spliced ​​according to feature channels so that each track number position and time position corresponds to a structural feature composed of dip angle feature, continuity feature, curvature feature and direction feature. This forms a multi-scale structural feature map with corresponding sampling positions and containing four attribute channels to the non-shot domain data. The multi-scale structural feature map is used as the structural continuity information of the non-shot domain data.

[0052] Multi-scale structural feature maps are formed according to the following formula:

[0053]

[0054] in, and These represent the tilt angle feature, continuity feature, curvature feature, and orientation feature obtained after fusing the local tilt angle, continuity, curvature, and orientation consistency at each scale, respectively. This indicates splicing along the feature channels; Represents a multi-scale structural feature map; Represents the real number field; Indicates the number of sampling points in the time direction. This indicates the number of sampling points in the channel direction. The value 4 indicates that the multi-scale structural feature map contains four structural attribute channels.

[0055] In some implementations, in addition to using structural tensor analysis to extract structural properties, local tilt estimation, edge detection, multi-scale feature extraction networks, or pre-trained structure-aware networks can also be used to extract structural continuity features of non-shot domain data, and the extraction results can be used as prior structural information for the first-stage network and the second-stage network.

[0056] S103, using the first-stage network, suppresses the main aliasing noise in the non-shot cluster domain data based on structural continuity information to obtain the first-stage reconstruction data.

[0057] Here, the structural attention weights are generated from the explicit structural properties of the seismic reflection phase axis, rather than being determined solely by the network's internal features. This guides the first-stage network to preserve complex structures and weak reflection signals while suppressing aliasing noise. Furthermore, the first-stage network employs a noise prediction method, avoiding direct fitting of complete, alias-free data. Instead, it focuses on learning the dispersed main aliasing noise in the non-shot domain, reducing structural ambiguity or amplitude distortion caused by directly reconstructing the effective signal. The first stage only suppresses the main noise components and retains non-shot domain features for subsequent co-shot domain processing, avoiding damage to the effective signal caused by excessive suppression in a single stage.

[0058] Specifically, the steps for obtaining the data in the first phase of reconstruction include: Feature mapping is performed on the multi-scale structural feature map to generate a structural attention weight map.

[0059] By fusing non-shot domain data and multi-scale structural feature maps, and using structural attention weight maps to weight and encode features, non-shot domain features are obtained.

[0060] The primary aliasing noise is predicted based on the characteristics of the non-shot domain, and the primary aliasing noise is subtracted from the non-shot domain data to obtain the first-stage reconstruction data.

[0061] Based on the specific description above, the structural continuity features are further mapped to obtain a structural attention module with values ​​ranging from 0 to 1, representing structural confidence. CMP data and structural continuity features are then input into SCMP-Net. SCMP-Net includes a feature fusion module, an encoder module, a structural continuity attention module, a bottleneck layer, a decoder module, and a main aliasing noise prediction head. A structural continuity attention module is added after each encoder to guide the network to focus on real seismic reflection event areas, enhance in-phase axes, and suppress noise in aliasing regions. After feature processing by the bottleneck layer and decoder, the main aliasing noise prediction head directly predicts the main aliasing noise N1 using a residual noise learning method, instead of directly predicting clean data. The first-stage reconstruction result D1 is obtained by subtracting the predicted noise from the input data, thus reducing the difficulty for the network to learn the main aliasing noise.

[0062] As described above, the multi-scale structural feature map is input into the convolutional mapping layer. Convolutional processing maps the structural features of each attribute channel and adjacent sampling positions in the multi-scale structural feature map. Then, Sigmoid activation is used to transform the mapping result to a value range of 0 to 1, generating a structural attention weight map. Each weight in the structural attention weight map corresponds to a specific sampling position in the multi-scale structural feature map, representing the structural confidence at that sampling position.

[0063] The structural attention weight map is generated according to the following formula:

[0064] in, This represents a structural attention weight map with values ​​ranging from 0 to 1. This indicates that the input features are subjected to convolutional mapping processing; This represents the Sigmoid activation function that transforms the convolution mapping result to the range of 0 to 1.

[0065] The non-shot gather domain data with a single data channel is concatenated with a multi-scale structural feature map containing four attribute channels according to the feature channels, forming fused data with five input channels. This fused data is then input into the first-stage network for layer-by-layer encoding to obtain the encoding features of different encoding layers. At the feature scale corresponding to each encoding layer, the weights in the structural attention weight map are multiplied element-wise with the encoding features to obtain weighted encoding features; wherein, the structural attention weights at the same sampling position are used to weight each feature channel corresponding to that position.

[0066] The dimensional relationship between the non-shot cluster data and the first-stage network fusion input is expressed as follows:

[0067] in, This represents the first-stage network fusion input obtained by splicing non-shot domain data and multi-scale structural feature maps; the value 1 indicates that the non-shot domain data contains one data channel, and the value 5 indicates that the fusion input contains one data channel and four structural attribute channels.

[0068] In some implementations, the encoder module of SCMP-Net includes four sequentially connected coding units, with a structural continuity attention module following each coding unit. For any coding unit, the coding features of that level are fused with multi-scale structural features of the corresponding scale, and then convolutional mapping and sigmoid activation are performed sequentially to obtain the structural attention weights for that level. These structural attention weights are then used to perform element-wise weighting on the corresponding coding features. The decoder module includes four decoding units corresponding to the four coding units, and the outputs of each coding unit are passed to the corresponding decoding unit through skip connections.

[0069] For the The level coding unit, structural continuity attention weights, and weighted coding features are determined according to the following formulas:

[0070]

[0071] in, Indicates the level of the coding unit; Indicates the first The coding features output by the level coding unit; Indicates the relationship with the first Level-coded features have multi-scale structural features corresponding to the feature scale; Indicates the first Attention weights for hierarchical structural continuity; This indicates element-wise multiplication; The first term represents the number of terms after structural continuity attention weighting. Level coding features.

[0072] The weighted encoded features of each layer are input into the bottleneck layer for deep feature extraction. The decoder then performs layer-by-layer decoding using the skip connection features passed from the encoder, and the decoder outputs these features as non-shot domain features. The main aliasing noise prediction head predicts the main aliasing noise based on these non-shot domain features, ensuring that the predicted main aliasing noise and the non-shot domain data have corresponding channel and time positions. The main aliasing noise is subtracted from the non-shot domain data according to the corresponding sampling positions to obtain the first-stage reconstructed data.

[0073] The data for the first phase of reconstruction is obtained using the following formula:

[0074] in, This represents the main aliasing noise predicted by the main aliasing noise prediction head of the first-stage network. This represents the first-stage reconstruction data obtained after subtracting the main aliasing noise from the non-shot domain data.

[0075] S104, convert the first-stage reconstruction data to the common gun domain.

[0076] Here, based on the gather conversion relationship between the common shot gather domain and the non-shot gather domain, the seismic traces in the first-stage reconstruction data are reverse-arranged, restoring each seismic trace to its corresponding common shot gather position, thus obtaining the converted first-stage reconstruction data. This conversion preserves the time sampling information and amplitude information of the first-stage reconstruction data and re-establishes the data arrangement relationship of seismic traces based on the source location.

[0077] The reverse gather transformation of the first-stage reconstructed data can be represented as:

[0078] in, This represents the reverse set transformation corresponding to the set rearrangement from the shared artillery set domain to the non-artillery set domain. This represents the first-stage reconstruction data after conversion to the common artillery domain.

[0079] As mentioned above, the same effective signal and aliasing noise have different continuity and distribution characteristics in different gather domains. The first stage utilizes the characteristics of continuous reflection phase axes and dispersed aliasing noise in the non-shot gather domain to suppress the main aliasing noise. After converting the reconstructed data from the first stage back to the common shot gather domain, the seismic wave propagation characteristics can be re-characterized under the common shot gather arrangement, so that the aliasing noise that was not completely suppressed in the first stage and the effective signal affected by it show different characteristics.

[0080] Unlike other methods that convert the non-shot domain processing results back to the common shot domain and use them directly as the final result, this invention uses domain conversion as a connection between two processing stages, so that the first-stage reconstructed data after conversion can continue to be used for residual aliasing noise suppression and effective signal reconstruction.

[0081] S105, using the second-stage network, residual aliasing noise is suppressed and effective signal is reconstructed on the converted first-stage reconstructed data based on structural continuity information to obtain the target reconstructed data.

[0082] Here, the difference data reflects the data components separated from the original co-shot gather seismic data in the first stage, providing a reference for identifying residual aliasing noise and potentially weakened effective signals. Non-shot gather domain features contain continuous structural information about the reflection phase axes, while co-shot gather domain features characterize the propagation characteristics of seismic waves under co-shot gather arrangement. The correlation between the two types of domain features is established through cross-attention calculation, and structural continuity information is used to limit the reconstruction direction, enabling the second-stage network to compensate for noise and structural errors left over from the first stage.

[0083] Unlike repeatedly performing the same noise suppression operation in a single data domain, this invention takes residual noise suppression and effective signal recovery as processing objectives to achieve collaborative reconstruction between two data domains and two processing stages.

[0084] Specifically, the steps for acquiring target reconstruction data include: Obtain the difference data between the common shot set seismic data and the converted first-stage reconstruction data.

[0085] The differential data, the transformed first-stage reconstruction data, non-shot cluster domain features, and structural continuity information are input into the second-stage network. The second-stage network suppresses residual aliasing noise and reconstructs the effective signal to obtain target reconstruction data.

[0086] Based on the specific description above, a reverse lookup transformation is performed on the first-stage reconstruction data to obtain the transformed first-stage reconstruction data. D 1,shot And based on the original co-shot collection seismic data and D 1,shot Determine the differences between them. R shot Subsequently, D 1,shot , R shot and structural continuity information F struct The dual-domain residual compensation reconstruction network DRCR-Net is input, and the non-shot domain features obtained from the first-stage network are input into its cross-domain feature interaction unit. DRCR-Net jointly encodes the transformed first-stage reconstruction data and difference data through a feature fusion layer, and the dual-domain feature encoding unit extracts common-shot domain features. Then, the cross-domain feature interaction unit uses non-shot domain features to guide common-shot domain recovery. On this basis, the residual compensation and structure recovery unit suppresses residual aliasing noise and recovers fractured phase axes, thin-layer structures, and weak reflection events. Finally, the target reconstruction data is obtained from the output layer. D 2.

[0087] As described above, the converted first-stage reconstruction data and the original co-shot gathering seismic data are mapped according to shot number, trace number, and time sampling location. At each corresponding sampling location, the converted first-stage reconstruction data is subtracted from the co-shot gathering seismic data to obtain the difference data. This difference data has the same co-shot gathering arrangement as the converted first-stage reconstruction data and is used to provide the second-stage network with the difference characteristics between the data before and after the first-stage processing.

[0088] The difference data is obtained using the following formula:

[0089] The joint input of the feature fusion layer in the second-stage network is represented as follows:

[0090] in, R shot This represents the difference between the original common shot gather seismic data and the transformed first-stage reconstruction data at the corresponding sampling locations; X 2 represents the joint input of the feature fusion layer in the second-stage network, and the square brackets indicate that the data within the brackets are used as common input; the non-shot domain features obtained by the first-stage network. F cmp Enter the cross-domain feature interaction unit of the second-stage network separately.

[0091] The second-stage network jointly encodes each input through a feature fusion layer and establishes the correlation between common shot set features and non-shot set features through a dual-domain residual compensation reconstruction layer. Guided by structural continuity information, it identifies residual aliasing noise and suppressed effective signals retained in the transformed first-stage reconstruction data. The residual compensation and structural restoration unit suppresses residual aliasing noise and recovers fractured phase axes, thin-layer structures, and weak reflection events. Finally, the reconstruction output head obtains target reconstruction data in the form of common shot set arrangement.

[0092] Specifically, the steps of suppressing residual aliasing noise and reconstructing the effective signal through a second-stage network to obtain the target reconstructed data include: The first-stage reconstructed data and the differential data after transformation are jointly encoded to obtain the common-shot domain features.

[0093] Cross-attention calculation is performed on the common-shot domain features and non-shot domain features. Based on the cross-attention calculation results and structural continuity information, residual aliasing noise in the first-stage reconstructed data after conversion is suppressed and effective signals are reconstructed to obtain target reconstruction data.

[0094] Regarding the specific description above, in DRCR-Net, the first-stage reconstructed data after transformation is processed through a feature fusion layer. D 1,shot and difference data R shot Joint encoding is performed, and the confocal set domain features are extracted by the dual-domain feature encoding unit. F shot ; then, F shot As query features, the non-shot domain features obtained from the first-stage network F cmp Cross-attention is performed using key and value features to guide data recovery of the confocal set domain by leveraging structural knowledge of the CMP domain. Based on this, the cross-attention calculation results are integrated with structural continuity information. F struct The common input structural constraint recovery module recovers fracture phase axes, thin-layer structures, and weak reflection events, and ultimately outputs target reconstruction data. D 2.

[0095] As described above, the transformed first-stage reconstruction data and differential data are combined according to the corresponding shot number, trace number, and time sampling location, and then input into the feature fusion layer of the second-stage network. The feature fusion layer jointly encodes the two types of data to simultaneously extract the seismic event features retained by the transformed first-stage reconstruction data and the first-stage processing differences represented by the differential data, thereby obtaining the common shot set domain features.

[0096] The features of the common-gun set domain are used as query features in the cross-attention calculation, while the features of the non-gun set domain obtained in the first-stage network are used as both key and value features. The correlation between each query feature and each key feature is calculated, and the correlation is converted into attention weights through Softmax processing. Then, the corresponding value features are weighted and aggregated using each attention weight to obtain the cross-attention calculation results corresponding to the common-gun set domain features. Thus, the structural features formed in the non-gun set domain are introduced into the common-gun set domain data recovery process.

[0097] In cross-attention calculation, the query feature, key feature, and value feature are as follows:

[0098] Cross-attention weights are determined according to the following formula:

[0099] in, F shot This represents the common-shot domain feature obtained by jointly encoding the first-stage reconstructed data and the difference data after transformation. F cmp This represents the non-shot set domain features obtained by the network in the first stage; Q att , K att and V att Let represent the query matrix, key matrix, and value matrix in the cross-attention calculation, respectively; Softmax(·) represents converting the product of the query matrix and the transposed key matrix into attention weights; Attn represents the cross-attention weight matrix. This weight matrix is ​​then used to... V att After weighted convergence, the cross-attention calculation result is obtained. The target reconstruction data output by the second-stage network is denoted as... D 2.

[0100] The cross-attention calculation results and structural continuity information are jointly input into the residual compensation and structural restoration unit. The structural continuity information guides the restoration location and direction of the earthquake reflection structure. Based on the cross-attention calculation results, the residual aliasing noise retained in the first-stage reconstruction data after conversion is suppressed. At the same time, the fracture phase axis, thin layer structure and weak reflection event are restored. Then, the restored features are converted into target reconstruction data in the form of a common shot set arrangement through the reconstruction output head.

[0101] In some implementations, the first-stage network and the second-stage network can be implemented using U-Net, ResU-Net, SwinTransformer, VisionTransformer, ConvNeXt, Mamba networks, or a combination of the above networks. The first-stage network is used to predict the main aliasing noise under the guidance of structural continuity information, and the second-stage network is used to perform cross-domain feature interaction, residual noise compensation, and effective signal reconstruction.

[0102] In some implementations, the dual-domain collaborative processing of the non-gun set domain and the common gun set domain can be extended to multi-domain collaborative processing. By adding auxiliary constraints from the Tau-P domain, FK domain, or frequency domain, the features of the corresponding data domains, along with the features of the non-gun set domain and the common gun set domain, can participate in network optimization to achieve multi-domain joint constraints.

[0103] In any of the above embodiments, the first-stage network and the second-stage network are obtained through joint training, specifically including: Based on the sample pairs consisting of data with aliasing noise and corresponding data without aliasing noise, the first-stage network and the second-stage network are updated using non-shot domain reconstruction loss, common-shot domain reconstruction loss, structural continuity constraint loss, and dual-domain consistency constraint loss.

[0104] Furthermore, sample pairs are formed by selecting data with and without aliasing noise that correspond to each other in terms of acquisition location, gather arrangement, and time sampling location. The data with aliasing noise in the sample pairs is input into the first-stage network and the second-stage network to obtain the first-stage reconstruction data and the target reconstruction data in sequence. In addition, according to the gather rearrangement method used for the data with aliasing noise, the corresponding data without aliasing noise is transformed to the non-shot gather domain to obtain the non-shot gather domain reference data.

[0105] The non-shot domain reconstruction loss is determined based on the reconstruction difference between the first-stage reconstructed data and the non-shot domain reference data to constrain the main aliasing noise suppression results of the first-stage network. The common-shot domain reconstruction loss is determined based on the reconstruction difference between the target reconstructed data and the non-aliasing noise data in the sample pair to constrain the residual aliasing noise suppression and effective signal reconstruction results of the second-stage network. The above reconstruction differences can be measured using L1 loss, L2 loss, or a combination of both.

[0106] In some implementations, the non-shot domain reconstruction loss, co-shot domain reconstruction loss, structural continuity constraint loss, and dual-domain consistency constraint loss may also employ one or more of SSIM loss, spectral loss, perceptual loss, or adversarial loss, and be combined with L1 loss or L2 loss to constrain network training from at least one aspect of data amplitude, structural similarity, or spectral characteristics.

[0107] Local dip angles, continuity, curvature, and orientation consistency are extracted from the first-stage reconstructed data and non-shot gather domain reference data, respectively. Structural continuity constraint losses are determined based on the differences between corresponding structural attributes to reduce in-phase axis breakage or structural displacement during the suppression of primary aliasing noise. Dual-domain consistency constraint losses are used to constrain the structural and amplitude consistency between the first-stage reconstructed data and the target reconstructed data in different data domains.

[0108] Weight coefficients are set for the four types of losses, and the weighted losses are summed to obtain the total loss. Backpropagation is then performed based on the total loss to jointly update the network parameters of the first-stage and second-stage networks. The training errors generated by the common-shot domain reconstruction loss and the dual-domain consistency constraint loss are propagated to the first-stage network through the second-stage network, enabling the two stages to optimize collaboratively during the same training process.

[0109] The weighted total loss for the four types of losses is determined by the following formula:

[0110] in, This represents the total loss used for jointly updating the first-phase and second-phase networks; This indicates the reconstruction losses outside the artillery cluster area. This indicates the losses incurred in rebuilding the artillery cluster area. This represents the structural continuity constraint loss. This represents the loss due to the two-domain consistency constraint. and These represent the weighting coefficients for the four types of losses mentioned above.

[0111] In any of the above embodiments, the step of obtaining the bi-domain consistency constraint loss includes: The target reconstruction data is converted to a non-gun cluster domain.

[0112] Based on the structural and amplitude differences between the converted target reconstruction data and the first-stage reconstruction data, the dual-domain consistency constraint loss is determined.

[0113] In this embodiment, to ensure structural and amplitude consistency among the processing results of different data domains, the present invention proposes a dual-domain consistency constraint optimization method. During network training, non-shot domain reconstruction loss, common-shot domain reconstruction loss, structural continuity constraint loss, and dual-domain consistency constraint loss are constructed simultaneously, thereby reducing information distortion caused by traditional single-domain processing.

[0114] Furthermore, the target reconstruction data is the common shot gather domain data output by the second-stage network. According to the gather rearrangement relationship used when converting common shot gather seismic data to non-shot gather domain data, the seismic traces in the target reconstruction data are rearranged, and the target reconstruction data is converted to the same non-shot gather domain as the first-stage reconstruction data, so that the sampling positions of the two correspond to each other in the trace number direction and time direction.

[0115] Using the same scale parameters and structural attribute extraction methods, local dip angle, continuity, curvature, and directional consistency were extracted from both the transformed target reconstruction data and the first-stage reconstruction data. The differences between the two sets of structural attributes were determined according to their corresponding track number and time positions, and the differences in structural attributes at each sampling location were summarized to obtain the structural differences. These structural differences characterize the changes in the in-phase axis direction, continuity, and local structural changes in the target reconstruction data relative to the first-stage reconstruction data after dual-domain processing.

[0116] In some implementations, the first-stage network and the second-stage network can be trained using supervised learning, self-supervised learning, semi-supervised learning, or transfer learning, in addition to supervised learning. After training, the co-shot gathering seismic data to be processed is input into the trained first-stage network and the second-stage network, and the aforementioned domain transformation, structural continuity feature extraction, main aliasing noise suppression, and residual noise compensation and reconstruction processes are performed to output the target reconstruction data.

[0117] Based on the corresponding track number and time positions, the seismic amplitude difference between the converted target reconstruction data and the first-stage reconstruction data is calculated, and the amplitude differences at each sampling location are summarized to obtain the amplitude difference. Corresponding constraint weights are assigned to the structural differences and amplitude differences respectively. The weighted structural differences and amplitude differences are combined to determine the dual-domain consistency constraint loss. The larger the structural difference or amplitude difference, the greater the dual-domain consistency constraint loss.

[0118] This invention provides a seismic aliasing noise suppression and reconstruction method. By using structural continuity information to guide the first-stage network to suppress the main aliasing noise in the non-shot cluster domain, and using the second-stage network to further suppress residual aliasing noise and reconstruct effective signals in the common shot cluster domain, this method can fully utilize the characteristic advantages of different data domains, reduce signal damage from complex structures, weak reflections, and fractured phase axes, and maintain structural and amplitude consistency between the processing results of different domains. This improves the noise suppression accuracy, structural fidelity, and subsequent imaging quality of seismic data.

[0119] In some implementations, the common shot gather seismic data to be processed can be 2D seismic data, 3D seismic data, marine seismic data, or terrestrial seismic data. The target reconstruction data obtained by this invention can be used for velocity modeling, migration imaging, full waveform inversion, or reservoir prediction.

[0120] A second aspect of the present invention provides an earthquake aliasing noise suppression and reconstruction system, such as... Figure 2 As shown, the system includes a data conversion module, a structural information acquisition module, a first-stage reconstruction module, a domain recovery module, and a second-stage reconstruction module.

[0121] The data conversion module acquires common-shot gathering seismic data containing aliasing noise and converts it into non-shot gathering domain data. The structural information acquisition module acquires structural continuity information from the non-shot gathering domain data. The first-stage reconstruction module uses a first-stage network to suppress the main aliasing noise in the non-shot gathering domain data based on the structural continuity information to obtain first-stage reconstruction data. The domain recovery module converts the first-stage reconstruction data to the common-shot gathering domain. The second-stage reconstruction module uses a second-stage network to suppress residual aliasing noise and reconstruct effective signals from the converted first-stage reconstruction data based on the structural continuity information to obtain the target reconstruction data. The specific processing procedures of each module can be referred to the aforementioned method embodiment, and will not be repeated here.

[0122] An embodiment of a third aspect of the present invention provides an electronic device. For example... Figure 3 As shown, the electronic device 3 includes a processor 301 and a memory 302. The memory 302 stores a computer program 303. When the processor 301 executes the computer program 303, it implements the earthquake aliasing noise suppression and reconstruction method described in any of the above embodiments. The electronic device 3 can be a desktop computer, a laptop computer, a workstation, or a cloud server, etc. The processor 301 can be a central processing unit, a digital signal processor, an application-specific integrated circuit, or a field-programmable gate array, etc.; the memory 302 can be the internal memory or an external storage device of the electronic device 3.

[0123] An embodiment of the fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the earthquake aliasing noise suppression and reconstruction method described in any of the above embodiments.

[0124] The modules in the above system can be implemented using software, hardware, or a combination of both. Each module can be set up independently or integrated into one or more processing units. The division method is only used to illustrate the corresponding functions and does not constitute a limitation on the scope of protection of this invention.

[0125] The above embodiments are only used to illustrate the technical solutions of the present invention. Any modifications or equivalent substitutions made by those skilled in the art to the above embodiments without departing from the technical concept of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for suppressing and reconstructing earthquake aliasing noise, characterized in that, Includes the following steps: Acquire common shot gathering seismic data containing aliasing noise, and convert the common shot gathering seismic data into non-shot gathering domain data; Obtain the structural continuity information of the non-shot cluster domain data; Using the first-stage network, the main aliasing noise in the non-shot cluster domain data is suppressed based on the structural continuity information to obtain the first-stage reconstruction data; Transform the first-stage reconstruction data into the common gun set domain; Using the second-stage network, residual aliasing noise is suppressed and effective signal is reconstructed on the converted first-stage reconstructed data based on the structural continuity information to obtain the target reconstructed data.

2. The earthquake aliasing noise suppression and reconstruction method according to claim 1, characterized in that, The step of obtaining the structural continuity information of the non-shot set domain data includes: The non-shot gather domain data is decomposed into multiple scales using multiple different scale parameters to obtain scale decomposition data corresponding to each scale parameter. Structural attributes are extracted from data decomposed at various scales, and structural continuity information is obtained by fusing these structural attributes.

3. The earthquake aliasing noise suppression and reconstruction method according to claim 2, characterized in that, The step of performing multi-scale decomposition on the non-shot gather domain data using multiple different scale parameters includes: Determine the Gaussian kernel corresponding to each scale parameter; The non-shot domain data is convolved using Gaussian kernels to obtain scale decomposition data corresponding to each scale parameter.

4. The earthquake aliasing noise suppression and reconstruction method according to claim 2, characterized in that, The steps for extracting structural attributes from data decomposed at various scales include: Based on the gradients along the channel number and time directions of the decomposed data at each scale, a structure tensor is constructed, and eigenvalue decomposition is performed on each structure tensor to obtain the local tilt angle and continuity in the structure properties. Based on the changes in local tilt angle and the distribution of local tilt angle directions within a preset local window, the curvature and directional consistency of the structural attributes are obtained respectively.

5. The earthquake aliasing noise suppression and reconstruction method according to claim 4, characterized in that, The step of fusing the structural properties to obtain structural continuity information includes: By integrating local tilt angles, continuity, curvature, and directional consistency at various scales, a multi-scale structural feature map is obtained as information on structural continuity.

6. The earthquake aliasing noise suppression and reconstruction method according to claim 5, characterized in that, The steps for obtaining the reconstruction data in the first stage include: Feature mapping is performed on the multi-scale structural feature map to generate a structural attention weight map; By fusing non-gun set domain data and multi-scale structural feature maps, and using structural attention weight maps to weight and encode features, non-gun set domain features are obtained. The primary aliasing noise is predicted based on the characteristics of the non-shot domain, and the primary aliasing noise is subtracted from the non-shot domain data to obtain the first-stage reconstruction data.

7. The earthquake aliasing noise suppression and reconstruction method according to claim 6, characterized in that, The steps for acquiring the target reconstruction data include: Obtain the difference data between the common shot gathering seismic data and the transformed first-stage reconstruction data; The difference data, the converted first-stage reconstruction data, non-shot cluster domain features, and structural continuity information are input into the second-stage network. The second-stage network suppresses residual aliasing noise and reconstructs the effective signal to obtain target reconstruction data.

8. The earthquake aliasing noise suppression and reconstruction method according to claim 7, characterized in that, The steps described above for suppressing residual aliasing noise and reconstructing effective signals through a second-stage network to obtain target reconstructed data include: The first-stage reconstructed data and the differential data after transformation are jointly encoded to obtain the confocal set domain features; Cross-attention calculation is performed on the common-shot domain features and non-shot domain features. Based on the cross-attention calculation results and structural continuity information, residual aliasing noise in the first-stage reconstructed data after conversion is suppressed and effective signals are reconstructed to obtain target reconstruction data.

9. The earthquake aliasing noise suppression and reconstruction method according to any one of claims 1-8, characterized in that, The first-stage network and the second-stage network are obtained through joint training, specifically including: Based on the sample pairs consisting of data with aliasing noise and corresponding data without aliasing noise, the first-stage network and the second-stage network are updated using non-shot domain reconstruction loss, common-shot domain reconstruction loss, structural continuity constraint loss, and dual-domain consistency constraint loss.

10. The earthquake aliasing noise suppression and reconstruction method according to claim 9, characterized in that, The steps for obtaining the dual-domain consistency constraint loss include: The target reconstruction data is converted to a non-shot cluster domain; The dual-domain consistency constraint loss is determined based on the structural and amplitude differences between the converted target reconstruction data and the first-stage reconstruction data.

Images