Model training method and device for seismic data reconstruction, equipment and storage medium

Abstract

The present disclosure relates to a model training method, device and equipment for seismic data reconstruction and a storage medium. The method comprises: obtaining original seismic data, the original seismic data comprising a time sampling point number and a seismic trace number; determining to-be-reconstructed seismic data and a first feature representation of the to-be-reconstructed seismic data based on the original seismic data, the first feature representation being used to represent a missing area and a distribution feature of the to-be-reconstructed seismic data; performing a multi-scale structure feature extraction operation on the to-be-reconstructed seismic data to determine a second feature representation of the to-be-reconstructed seismic data, the second feature representation being used to represent a continuity of a structure feature of the to-be-reconstructed seismic data; processing the to-be-reconstructed seismic data, the first feature representation and the second feature representation by using a target model to determine reconstructed seismic data; and training the target model based on a difference between the reconstructed seismic data and the original seismic data.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a method, apparatus, device, and storage medium for model training of seismic data reconstruction. Background Technology

[0002] During seismic data acquisition, data gaps often occur due to factors such as complex terrain, equipment malfunctions, limitations of observation systems, and environmental interference. Therefore, how to reconstruct missing seismic data with high accuracy to restore complete data has become a critical problem that urgently needs to be solved in the field of seismic data processing. Summary of the Invention

[0003] To address the aforementioned technical issues, this disclosure provides a method, apparatus, device, and storage medium for model training of seismic data reconstruction.

[0004] A first aspect of this disclosure provides a model training method for seismic data reconstruction, comprising: acquiring raw seismic data, the raw seismic data including the number of time sampling points and the number of seismic traces; based on the raw seismic data, determining seismic data to be reconstructed and a first feature representation of the seismic data to be reconstructed, the first feature representation being used to characterize the missing regions and distribution characteristics of the seismic data to be reconstructed; performing a multi-scale structural feature extraction operation on the seismic data to be reconstructed to determine a second feature representation of the seismic data to be reconstructed, the second feature representation being used to characterize the continuity of the structural features of the seismic data to be reconstructed; processing the seismic data to be reconstructed, the first feature representation, and the second feature representation using a target model to determine reconstructed seismic data; and training the target model based on the differences between the reconstructed seismic data and the raw seismic data.

[0005] A second aspect of this disclosure provides an apparatus for training a model for seismic data reconstruction, comprising: an acquisition module for acquiring raw seismic data, the raw seismic data including the number of time sampling points and the number of seismic traces; a determination module for determining, based on the raw seismic data, seismic data to be reconstructed and a first feature representation of the seismic data to be reconstructed, the first feature representation being used to characterize the missing regions and distribution characteristics of the seismic data to be reconstructed; an execution module for performing a multi-scale structural feature extraction operation on the seismic data to be reconstructed to determine a second feature representation of the seismic data to be reconstructed, the second feature representation being used to characterize the continuity of the structural features of the seismic data to be reconstructed; a processing module for processing the seismic data to be reconstructed, the first feature representation, and the second feature representation using a target model to determine reconstructed seismic data; and a training module for training the target model based on the differences between the reconstructed seismic data and the raw seismic data.

[0006] A third aspect of this disclosure provides an electronic device, including: a processor; and a memory for storing executable instructions; wherein the processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the method of the first aspect described above.

[0007] A fourth aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the method of the first aspect described above.

[0008] Through the above scheme, the embodiments of this disclosure can accurately identify missing areas using the first feature representation, perceive the continuity of the stratigraphic structure using the second feature representation, and guide the target model to complete data recovery along the real geological direction, thereby effectively improving the reconstruction accuracy and geological rationality. Attached Figure Description

[0009] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0010] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a model training method for seismic data reconstruction provided in an embodiment of this disclosure; Figure 2 This is a visualization of the raw seismic data provided in the embodiments of this disclosure; Figure 3 This is a visualization of the seismic data to be reconstructed provided in the embodiments of this disclosure; Figure 4 This is a schematic diagram representing the first feature provided in an embodiment of this disclosure; Figure 5 This is a schematic diagram of multi-scale structural feature extraction provided in an embodiment of this disclosure; Figure 6 This is a schematic diagram of the target model structure provided in the embodiments of this disclosure; Figure 7 This is a schematic diagram of the overall process of a model training method for seismic data reconstruction provided in an embodiment of this disclosure; Figure 8 This is a schematic diagram of the structure of a device for model training for seismic data reconstruction provided in an embodiment of this disclosure; Figure 9This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0012] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0013] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0014] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0016] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] Traditional methods often result in missing seismic data during acquisition due to factors such as complex terrain, equipment malfunctions, limitations of observation systems, and environmental interference. Missing data reduces the accuracy of subsequent processing steps, including migration imaging, velocity modeling, reservoir prediction, and full waveform inversion.

[0018] Currently, commonly used seismic data reconstruction methods mainly include interpolation methods, transform domain reconstruction methods, and deep learning-based reconstruction methods. Traditional interpolation methods rely on data regularity and are difficult to adapt to complex geological structures. While deep learning-based methods can improve reconstruction accuracy, they mostly rely solely on missing amplitude data for recovery, failing to fully utilize the distribution information of missing areas and the continuity characteristics of seismic phase axes. This results in reconstruction results with structural distortion, missing high-frequency information, and insufficient geological continuity.

[0019] Therefore, there is an urgent need to propose an intelligent reconstruction method that can make full use of missing pattern information and seismic structural features to improve the accuracy of missing seismic data recovery.

[0020] Based on this, embodiments of this disclosure provide a scheme for training a model for seismic data reconstruction. The scheme includes: acquiring raw seismic data, which includes the number of time sampling points and seismic traces; determining, based on the raw seismic data, seismic data to be reconstructed and a first feature representation of the seismic data to be reconstructed, the first feature representation being used to characterize the missing regions and distribution characteristics of the seismic data to be reconstructed; performing multi-scale structural feature extraction on the seismic data to be reconstructed to determine a second feature representation of the seismic data to be reconstructed, the second feature representation being used to characterize the continuity of the structural features of the seismic data to be reconstructed; processing the seismic data to be reconstructed, the first feature representation, and the second feature representation using a target model to determine the reconstructed seismic data; and training the target model based on the differences between the reconstructed seismic data and the original seismic data.

[0021] Through the above scheme, the embodiments of this disclosure can accurately identify missing areas using the first feature representation, perceive the continuity of the stratigraphic structure using the second feature representation, and guide the target model to complete data recovery along the real geological direction, thereby effectively improving the reconstruction accuracy and geological rationality.

[0022] The method will be described below with reference to specific embodiments.

[0023] Figure 1 This is a flowchart of a model training method for seismic data reconstruction provided in an embodiment of this disclosure. The method can be executed by a model training device, which can be implemented in software and / or hardware. The model training device can be configured in an electronic device, such as a server or terminal, wherein the terminal specifically includes a mobile phone, computer or tablet computer, etc.

[0024] like Figure 1 As shown, the model training method provided in this embodiment includes the following steps: S110. Electronic equipment acquires raw seismic data, which includes the number of time sampling points and the number of seismic traces.

[0025] In some embodiments, the raw seismic data D (also referred to as the raw seismic dataset) may refer to the complete seismic records acquired by seismic acquisition equipment as label data (also referred to as the ground truth) for training the target model.

[0026] In some embodiments, the raw seismic data D can be presented in any suitable form. As an example, the raw seismic data D can be represented as a two-dimensional matrix, for example, where the rows of the matrix correspond to the number of time sampling points, and the columns correspond to the number of seismic traces. In some embodiments, the raw seismic data D can be represented as:

[0027] Where H represents the number of time sampling points and W represents the number of seismic traces.

[0028] like Figure 2 As shown, Figure 2 This is a visualization of the raw seismic data provided in this embodiment of the disclosure, wherein the vertical axis represents the number of time sampling points and the horizontal axis represents the number of seismic traces.

[0029] In some embodiments, after acquiring the raw seismic data D, the electronic device can use it as a supervision label for subsequent training. During the model training phase, the electronic device can compare the output reconstruction result with the raw seismic data D, calculate the difference between the two, and use this as a basis for adjusting the model parameters.

[0030] S120. Based on the original seismic data, the electronic device determines the seismic data to be reconstructed and the first feature representation of the seismic data to be reconstructed. The first feature representation is used to characterize the missing areas and distribution characteristics of the seismic data to be reconstructed.

[0031] In some embodiments, seismic data to be reconstructed It can be incomplete seismic data with missing regions obtained by performing missing processing on the original seismic data D, which is used to simulate data missing scenarios caused by factors such as complex terrain, equipment failure or observation system limitations during actual acquisition.

[0032] In some embodiments, electronic devices may perform missing data processing based on a preset missing data strategy to construct the seismic data to be reconstructed. As an example, the preset missing data strategy can employ at least one of random missing data and regular missing data. Random missing data can refer to randomly deleting a portion of the seismic trace data, such as randomly deleting 50% of the seismic trace data, to simulate a randomly distributed missing data scenario. Regular missing data can refer to deleting seismic trace data at fixed intervals, such as continuously deleting multiple traces at fixed intervals, to simulate a continuous missing data scenario caused by mountain obstacles, equipment failure, etc.

[0033] like Figure 3 As shown, Figure 3 This is a visualization of the seismic data to be reconstructed provided in the embodiments of this disclosure, used to demonstrate the effect of the original seismic data after processing with random and regular missing data.

[0034] In some embodiments, the first feature representation M can be related to the seismic data to be reconstructed. The corresponding feature representation can be any appropriate form, examples of which may include, but are not limited to, encoding matrices, feature matrices, and eigenvectors, to indicate the seismic data to be reconstructed. Check if any data is missing from each location.

[0035] As an example, the first feature representation M can refer to the seismic data to be reconstructed. A corresponding encoding matrix (e.g., a missing pattern encoding matrix) is used to explicitly describe the missing data at various locations in the seismic data to be reconstructed. This encoding matrix can be correlated with the seismic data to be reconstructed. They have the same spatial dimensions, where the element value at each location can be used to indicate whether the data at that location is missing.

[0036] like Figure 4 As shown, Figure 4 This is a schematic diagram of the first feature representation provided in an embodiment of this disclosure. Taking a binary matrix as an example, in this binary matrix, the first value (e.g.) This can indicate that the data at the corresponding position is valid data. The second value (e.g.) The symbol () can indicate that the data at the corresponding location is missing data. In this way, embodiments of this disclosure can characterize the location and distribution characteristics of missing data regions.

[0037] In some embodiments, electronic devices can reconstruct seismic data. Together with the first feature representation M, they serve as input data for training the subsequent target model. As an example, electronic devices can use the seismic data to be reconstructed. The first feature representation M is concatenated with the first feature representation M along the channel dimension to form dual-channel input data, and the concatenated dual-channel data is used as the input data for subsequent training of the target model.

[0038] In this way, the embodiments of this disclosure enable the target model to accurately identify the location and distribution characteristics of the missing regions, thereby improving the targeting of data reconstruction.

[0039] S130. The electronic device performs a multi-scale structural feature extraction operation on the seismic data to be reconstructed to determine the second feature representation of the seismic data to be reconstructed. The second feature representation is used to characterize the continuity of the structural features of the seismic data to be reconstructed.

[0040] In some embodiments, multi-scale structural feature extraction can refer to using feature extraction operators with multiple different receptive fields to extract seismic data to be reconstructed. Perform multi-scale feature extraction to capture the seismic data to be reconstructed. Geological structural features at different scales, such as the direction of phase axis extension, stratigraphic continuity trend, and local tectonic boundaries in seismic data.

[0041] As an example, feature extraction operators may include, but are not limited to, convolution kernels, filters, wavelet transform operators, etc. In some embodiments, multiple feature extraction operators with different receptive fields may have different receptive field sizes, enabling each operator to capture geological structural features at a corresponding scale.

[0042] For ease of understanding, the following section uses a convolution kernel as an example to introduce this disclosure.

[0043] Because seismic data contains phase axes with different spatial distribution scales, single-scale feature extraction methods are insufficient to simultaneously capture local details and macroscopic structures. In some embodiments, electronic devices can utilize a set of convolution kernels at different scales to extract features from the seismic data to be reconstructed. m Perform convolution operations to extract geological structural features at different scales.

[0044] As an example, a set of convolution kernels at different scales can include small-scale convolution kernels (e.g., 3×3 convolution kernels), medium-scale convolution kernels (e.g., 5×5 convolution kernels), and large-scale convolution kernels (e.g., 7×7 convolution kernels). Small-scale convolution kernels can be used to extract local detailed structural features (e.g., thin-layer reflections, local amplitude variations), medium-scale convolution kernels are used to extract medium-range stratigraphic distribution features (e.g., the continuity of a set of phase axes), and large-scale convolution kernels are used to extract large-scale geological structural features (e.g., overall stratigraphic strike, large fault zone boundaries).

[0045] It is understood that the number and size of the convolutional kernels described above are merely illustrative examples, and this disclosure is not intended to limit the specific number and size of the convolutional kernels.

[0046] Furthermore, electronic devices can fuse geological structural features at different scales to determine the seismic data to be reconstructed. The second feature representation S (also known as the structural feature map) is used. In some embodiments, fusion may refer to stitching together convolution results at different scales along the channel dimension, and then performing feature compression through convolutional layers to determine the seismic data to be reconstructed. The second characteristic representation S can be expressed as:

[0047] Among them, among them, , Indicates the feature map size. Indicates the number of channels.

[0048] In some embodiments, the second feature representation S may refer to the seismic data to be reconstructed. Structural continuity feature maps are used to characterize the seismic data to be reconstructed. The continuity of structural features. This structural continuity feature map is compared with the seismic data to be reconstructed. They can have the same spatial dimensions, where the eigenvalues ​​at each location are used to characterize the continuity and extension direction of the seismic phase axis at that location.

[0049] like Figure 5 As shown, Figure 5 This is a schematic diagram illustrating multi-scale structural feature extraction provided in an embodiment of this disclosure. The electronic device can extract seismic data to be reconstructed. Perform multi-scale feature extraction, for example, extracting the seismic data to be reconstructed using 3×3, 5×5, and 7×7 convolutions respectively. The local, mesoscale, and large-scale structural features are further analyzed. The electronic device can then concatenate these three types of feature maps through channels and perform feature fusion via 1×1 convolution, ultimately outputting a dimension of H×W×C. s The structural feature diagram S (i.e., the second feature representation S).

[0050] In this way, by fusing convolution results at different scales, the second feature representation S can be extracted, enabling the subsequent network to fully utilize the prior information of structural continuity of the seismic data during the reconstruction process, and guide the network to complete the restoration of the missing area along the direction of the real geological structure.

[0051] It should be noted that the second feature representation S and the first feature representation M differ in function and purpose. Specifically, the first feature representation M can be used to indicate the seismic data to be reconstructed. The location and distribution characteristics of the missing regions serve to guide the target model to focus on these areas, ensuring that the model prioritizes these locations during feature extraction and reconstruction. The second feature representation, S, can be used to characterize the seismic data D to be reconstructed. m The structural continuity feature serves to provide the target model with prior structural information on the extension direction of the seismic phase axis and the trend of stratigraphic continuity, guiding the target model to reconstruct along the direction of the real geological structure when recovering missing data, thereby helping to achieve seismic data reconstruction driven by multi-dimensional information collaboration.

[0052] S140. The electronic device uses the target model to process the seismic data to be reconstructed, the first feature representation, and the second feature representation to determine the reconstructed seismic data.

[0053] In some embodiments, the target model may include an encoder and a decoder. In some embodiments, electronic devices may utilize the encoder of the target model to process the seismic data to be reconstructed. The first feature representation M and the second feature representation S are used to obtain the third feature representation F. As an example, electronic devices can reconstruct seismic data... The first feature representation M and the second feature representation S are concatenated along the channel dimension to form a multi-channel input tensor.

[0054] As discussed above, the seismic data to be reconstructed By concatenating the first feature representation M with the data along the channel dimension to form dual-channel input data, and using the concatenated dual-channel data as input data for subsequent target model training, the target model can accurately identify the location and distribution characteristics of missing areas, thereby improving the relevance of seismic data reconstruction.

[0055] The second feature representation S can be used to characterize the seismic data D to be reconstructed. m The structural continuity feature serves to provide the model with prior structural information on the extension direction of the seismic phase axis and the trend of stratigraphic continuity, guiding the target model to reconstruct along the direction of the real geological structure when recovering missing data, thereby helping to achieve seismic data reconstruction driven by multi-dimensional information collaboration.

[0056] Furthermore, the electronic device can input this multi-channel input tensor to the encoder of the target model. The encoder can perform feature extraction and downsampling on this multi-channel input tensor to output a third feature representation F (also known as multi-scale deep semantic information). The third feature representation F can be expressed as:

[0057] in, , Indicates the feature map size. Indicates the number of channels.

[0058] In some embodiments, the third feature representation F can be used to characterize the seismic data to be reconstructed. The first feature represents the deep semantic information of M and the second feature represents the deep semantic information of S.

[0059] Specifically, the encoder may include multiple downsampling modules connected in sequence, each downsampling module comprising a convolutional layer and a pooling layer. During the encoding process, the encoder can sequentially perform multiple downsampling operations on the multi-channel input tensor through multiple convolutional layers and pooling layers, extracting amplitude features, missing pattern features, and structural continuity features around the missing regions layer by layer, in order to gradually obtain shallow features. Mid-layer characteristics and deep semantic features The final output is the third feature representation F. In some embodiments, the electronic device can store the encoded features obtained after each downsampling (e.g., , and This facilitates skip connections and multi-scale structure fusion in the subsequent decoding stage.

[0060] It is understood that the number of downsampling times and the number of convolutional layers of the encoder can be configured according to the actual application scenario and computing resources, and this disclosure does not limit this.

[0061] In some embodiments, the electronic device may determine the structure-guided weight matrix based on the second feature representation S. Structure-guided weight matrix It can be used to characterize the degree of structural continuity at various locations in the seismic data to be reconstructed.

[0062] Specifically, the electronic device can first perform a feature transformation on the second feature representation S. The formula for this feature transformation can be expressed as:

[0063] As an example, this feature transformation can be implemented using a 1×1 convolutional layer to change the number of channels in the second feature representation S to be consistent with the number of channels in the third feature representation F, so that the structural features and deep semantic features maintain the same dimension. The structural features after feature transformation can characterize structural information in the seismic data to be reconstructed, such as the direction of phase axis extension, stratigraphic continuity trend, location of local fault boundaries, and tectonic change regions.

[0064] Furthermore, the electronic device can perform normalization on the transformed second feature representation to determine the structure-guided weight matrix. Structure-guided weight matrix It can be represented as:

[0065] in, .

[0066] As an example, this normalization process can be implemented using the Sigmoid function, which maps the element values ​​in the transformed second feature representation to... The interval is used to generate the structure-guided weight matrix. Among them, the structure-guided weight matrix Each element value in the array is used to characterize the degree of structural continuity at the corresponding location in the seismic data to be reconstructed. The closer the element value is to 1, the higher the structural continuity at that location, and the closer it is to 0, the weaker the structural continuity at that location.

[0067] In this way, the embodiments of this disclosure can transform the second feature representation S into a weight matrix with clear physical meaning (i.e., structural continuity) through feature transformation and normalization, providing a quantifiable guiding basis for the subsequent enhancement processing of the third feature representation F.

[0068] In some embodiments, the electronic device may be based on a structure-guided weight matrix. The third feature representation F is enhanced to determine the structure-guided enhancement features. This structure guides the enhancement features. It can be represented as:

[0069] in, This indicates element-wise multiplication.

[0070] As an example, electronic devices can guide the weight matrix of the structure. A weighted operation is performed on the third feature representation F to determine the structurally guided enhancement feature. For example, electronic devices can use element-wise multiplication operations to guide the structure through a weight matrix. Each element value in the algorithm acts on the corresponding feature vector in the third feature representation F, thereby enhancing features in regions of high structural continuity and suppressing features in regions of low structural continuity. This structure-guided feature enhancement... It retains the deep semantic information of the third feature representation F, and incorporates the prior information of structural continuity represented by the second feature representation S.

[0071] In this way, the embodiments of this disclosure can effectively enhance the quality information around the missing area in the seismic data to be reconstructed, providing a more reliable feature basis for the directional recovery of the missing area in the subsequent decoding stage.

[0072] In some embodiments, the electronic device may utilize the decoder of the target model to process structure-guided enhancement features. To determine the reconstructed seismic data .

[0073] Specifically, the electronic device can acquire multiple levels of encoded features generated by the encoder. These multiple levels of encoded features can include shallow features output by the encoder at different downsampling stages. Mid-layer characteristics and deep features .

[0074] Furthermore, electronic devices can guide and enhance features. Perform a fusion operation with encoded features from multiple different levels to determine the fused features. As an example, electronic devices can... , , and By concatenating the data along the channel dimension, a fused feature is obtained. This fused feature can be represented as:

[0075] in, Indicates shallow structural features. Indicates the structural features of the middle layer. It represents deep semantic features.

[0076] In some embodiments, the electronic device can perform fusion features Perform convolution operations to determine the final reconstructed features. As an example, electronic devices can use convolutional layers to compress the stitched fused features to obtain the final reconstructed features, which can be represented as:

[0077] Furthermore, the electronic device can utilize the decoder of the target model to process the final reconstructed features. To determine the reconstructed seismic data As an example, electronic devices can ultimately reconstruct features. The input is fed to the decoder. Further, the decoder can progressively recover the spatial resolution of the feature map through upsampling operations, and during the upsampling process, it performs skip connections with the encoded features of the corresponding level from the encoder, fusing shallow detail features with deep semantic features, ultimately outputting reconstructed seismic data with the same size as the original seismic data. The reconstructed seismic data It can be represented as:

[0078] In some embodiments, during the upsampling recovery process of the decoder, for the seismic data to be reconstructed... For missing regions in the data, the decoder can prioritize amplitude prediction by referencing the extension direction of the phase axis, the continuity of neighboring reflection events, and the distribution of local tectonic features. For the seismic data to be reconstructed... Within the effective area, the decoder can preserve the original seismic information.

[0079] In this manner, embodiments of the present disclosure enable the reconstructed seismic data Without introducing additional errors in the effective area, only the missing area is restored in a targeted manner, thereby improving the geological rationality and fidelity of the reconstruction results.

[0080] S150: Electronic equipment trains the target model based on the difference between the reconstructed seismic data and the original seismic data.

[0081] In some embodiments, the electronic device may be based on reconstructed seismic data. The difference between the data and the original seismic data D is used to determine the joint loss. Furthermore, the electronic device can adjust the parameters of the target model based on this joint loss to train the target model.

[0082] In some embodiments, the joint loss may include the missing region reconstruction loss. Such missing area reconstruction loss It can be used to characterize reconstructed seismic data. Reconstruction error in the missing region compared to the original seismic data D. Reconstruction loss in this missing region. It can be represented as:

[0083] in, D represents the reconstructed seismic data, while D represents the original seismic data.

[0084] Furthermore, since the effective region accounts for a large proportion, the target model tends to prioritize learning known regions during training, while ignoring the missing regions that truly need to be recovered. Therefore, embodiments of this disclosure can define a weight matrix W, which can be expressed as: , Where M represents the first feature and λ is a weight coefficient greater than 1. In this weight matrix W, the weight of the effective region is 1, and the weight of the missing region is λ (λ>1).

[0085] By assigning higher weights to the missing regions using the weight matrix W, the target model can maintain the data of the effective regions unchanged during training, while focusing on the recovery of the missing regions.

[0086] In addition, due to By retaining only the location of the missing region, the target model can focus on amplitude recovery, event continuity recovery, and detail reconstruction of the missing region during training, thereby effectively improving the model's learning efficiency and reconstruction accuracy for the missing region.

[0087] In some embodiments, the joint loss may include spectrum preservation loss. Such spectrum retention loss It can be used to characterize reconstructed seismic data. Spectral differences between the original seismic data D and the original seismic data. This spectral retention loss. It can be represented as:

[0088] in, This indicates the Fourier transform operation. D represents the reconstructed seismic data, while D represents the original seismic data.

[0089] It should be noted that seismic data typically contains rich frequency information. While traditional mean square error loss can recover amplitude, it easily leads to the attenuation of high-frequency components, thus affecting the identification capability of thin-layer reflections and the quality of subsequent migration imaging. Therefore, this disclosure introduces a spectrum preservation loss, which calculates the spectral difference between the reconstructed and original seismic data in the frequency domain by performing Fourier transforms on both. This spectrum preservation loss constraint maintains the consistency of the frequency distribution between the reconstructed results and the original data, effectively reducing the loss of high-frequency information, improving the identification capability of thin-layer reflections, and enhancing the quality of subsequent migration imaging.

[0090] In some embodiments, the joint loss may include structural continuity loss. Such structural continuity loss It can be used to characterize reconstructed seismic data. The structural continuity difference between the original seismic data D and the original seismic data. This structural continuity loss. It can be represented as:

[0091] in, This is a gradient extraction operator used to extract the horizontal and vertical gradients of the input data. D represents the reconstructed seismic data, while D represents the original seismic data.

[0092] It should be noted that one of the most important characteristics of seismic data is the continuity of phase axes, which reflects the spatial distribution and structural morphology of underground strata. To ensure that the reconstruction results conform to the actual geological structure, this embodiment introduces a structural continuity loss, employing gradient extraction operators to extract the horizontal and vertical gradients of the reconstructed and original seismic data respectively, and comparing the extraction results. Through the constraint of this structural continuity loss, the network can maintain the continuity of phase axes, the integrity of reflection interfaces, the clarity of structural boundaries, and the consistency of fault locations during training, thereby enhancing the geological rationality of the reconstruction results and reducing the risk of structural distortion.

[0093] In this way, the embodiments of this disclosure can utilize the inherent structural continuity characteristics of seismic data to constrain and restore missing regions. Compared with traditional reconstruction methods that rely solely on amplitude information for prediction, the embodiments of this disclosure can further utilize stratigraphic continuity, phase axis extension patterns, and tectonic boundary information to guide the network in reconstructing missing regions, thereby effectively reducing structural distortion and improving reconstruction accuracy and geological rationality in complex missing scenarios.

[0094] In some embodiments, the joint loss can be expressed as:

[0095] in, For the losses in rebuilding the missing areas, To preserve spectrum loss, For structural continuity loss, These are the weighting coefficients.

[0096] In some embodiments, the weight coefficients of the joint loss function can be dynamically adjusted. As an example, during the training of the target model, the electronic device can adjust the loss for reconstructing missing regions. Weighting coefficients By gradually reducing the frequency, electronic devices can maintain spectrum preservation loss. Weighting coefficients Without changing, electronic devices can adjust for structural continuity losses. Weighting coefficients Gradually increase.

[0097] As an example, the weighting coefficients and The dynamic adjustment can be achieved using the following formula:

[0098] in, For the current training round, For the total number of training rounds, As the initial reconstruction loss weight, For the spectral loss weight, The initial structural loss weights, The final structural loss weights.

[0099] As an example, weighting coefficients , and Dynamic adjustments can be made in the following way: at the start of training, The value is greater than , The value is greater than For example, the weighting coefficients can be set to... , , ,Right now This allows the target model to prioritize learning amplitude recovery during the initial training phase. As the number of training epochs increases, The value of gradually decreases. As the values ​​of gradually increase, the network focuses more on structural continuity in the later stages of training. For example, at the end of training, the weight coefficients can be adjusted to... , , .

[0100] Through the above dynamic adjustments, stable convergence can be ensured in the early stage of training, and the structural recovery ability can be enhanced in the later stage of training, thereby improving the reconstruction effect of complex geological structures.

[0101] It should be noted that the joint loss function constructed in this embodiment is not a simple error weighting, but a triple constraint mechanism designed for the characteristics of seismic data: the missing area reconstruction loss focuses on the recovery of missing areas, the spectrum preservation loss ensures that high-frequency seismic information is not lost, and the structural continuity loss maintains the continuity of the phase axis and geological structure.

[0102] In some embodiments, after determining the joint loss, the electronic device may employ an optimizer to optimize the parameters of the target model. As an example, the optimizer could be the AdamW optimizer, used to update the network parameters of the target model based on the joint loss. During the optimization process, the electronic device may employ a cosine annealing learning rate strategy to dynamically adjust the learning rate, which can be expressed as:

[0103] Where T represents the total number of training rounds, and t represents the current training round. The initial learning rate, This is the minimum learning rate.

[0104] In some embodiments, to improve the adaptability of the target model to complex missing scenarios, the electronic device can randomly generate new missing patterns before the start of each training cycle to dynamically update the missing pattern encoding matrix of the training samples. As an example, the new missing patterns may include at least one of random missing and regular missing patterns.

[0105] In this way, the embodiments of this disclosure can avoid the target model memorizing fixed missing patterns by randomly generating missing patterns before the start of each training cycle, thereby effectively improving the generalization ability and engineering applicability of the target model in various scenarios such as random missing patterns, regular missing patterns, and large-area missing patterns.

[0106] In some embodiments, after training is complete, the electronic device can input the missing seismic data to be processed into the trained target model, and the target model will output reconstructed complete seismic data. The reconstructed complete seismic data can be used for subsequent processing steps such as seismic imaging, velocity modeling, reservoir prediction, and full waveform inversion to improve the data quality and accuracy of subsequent processing steps.

[0107] For ease of description, the following text will refer to Figure 6 and Figure 7 The embodiments of this disclosure will be described.

[0108] like Figure 6 As shown, Figure 6 This is a schematic diagram of the target model structure provided in this embodiment. The target model adopts an encoder-decoder architecture, with the encoder (downsampling path) on the left and the decoder (upsampling path) on the right.

[0109] In the encoder, input data (e.g., seismic data to be reconstructed) is used. The input data consists of two features (M and S), which are first feature representations and second feature representations, respectively. These features are then processed through multiple downsampling modules. Each downsampling module includes two 3×3 convolutional layers (each followed by an activation function) and a 2×2 max-pooling layer (with a stride of 2). Through this structure, the spatial size of the input data can be gradually halved (i.e., from H...). W changes sequentially to H / 2 W / 2, H / 4 W / 4, H / 8 (W / 8, etc.), while the number of channels increases layer by layer to extract multi-scale deep semantic features. In this process, the encoded features output by the encoder at each downsampling stage are passed to the decoding module of the corresponding layer of the decoder through skip connections.

[0110] In the decoder, the feature map sequentially passes through multiple upsampling modules. Each upsampling module includes a 2×2 transposed convolution (with a stride of 2) to restore the spatial resolution of the feature map, and two 3×3 convolutional layers (each followed by an activation function). During the upsampling process, the decoder receives the encoded features of the corresponding layer from the encoder through skip connections and concatenates these encoded features with the upsampled feature map, thereby achieving the fusion of shallow detail features and deep semantic features. Finally, the decoder outputs a feature map with the same spatial size as the input data, and maps the number of channels to the target number of channels through a 1×1 convolutional layer to obtain the reconstructed seismic data D'.

[0111] It should be noted that, Figure 6 The target model structure shown is only an example. The number of downsampling times in the encoder, the number of upsampling times in the decoder, and the number and size of each layer of convolutional kernels can be adjusted according to the actual application scenario. This disclosure does not limit this.

[0112] like Figure 7 As shown, Figure 7 This is a schematic diagram of the overall process of a model training method for seismic data reconstruction provided in this embodiment. The flowchart illustrates the entire process from data acquisition to network training, including key steps such as data acquisition, missing sample construction, missing sample-aware encoding, structural continuity feature extraction, seismic data reconstruction network, structure-guided reconstruction, joint loss function construction, and network training.

[0113] In the data acquisition phase (S1), the electronic device acquires complete seismic data as training labels. In the missing sample construction phase (S2), the electronic device performs missing data processing on the complete seismic data according to a preset missing data strategy, constructing seismic data with missing regions (i.e., seismic data to be reconstructed). The missing data D is then processed to generate a corresponding missing pattern encoding matrix M (i.e., the first feature representation M), where matrix element 1 represents valid data and 0 represents missing data. In the missing data-aware encoding stage (S3), the electronic device fuses the missing data D with the missing pattern encoding matrix M along the channel dimension, enabling the network to identify the location of the missing region.

[0114] In the structural continuity feature extraction stage (S4), the electronic device extracts the structural features of the seismic data to be reconstructed through multi-scale convolution operations. Specifically, 3×3 convolution kernels, 5×5 convolution kernels, and 7×7 convolution kernels are used to extract geological structural features at different scales, and the multi-scale features are fused to generate a structural feature map S (i.e., the second feature representation S).

[0115] In the seismic data reconstruction network stage (S5), the electronic device inputs the missing data D, the missing pattern coding matrix M, and the structural feature map S into the encoder-decoder network. In the structure-guided reconstruction stage (S6), the electronic device uses the structural feature map S to guide the deep features F output by the encoder, generating guided enhancement features. This feature is then input into the decoder for upsampling and recovery to obtain the reconstructed seismic data. .

[0116] In the joint loss function construction phase (S7), the electronic device constructs the joint loss function L. In the network training phase (S8), the electronic device calculates the gradient based on the joint loss L, optimizes the network parameters (i.e., the target model) through the backpropagation algorithm, and continuously optimizes the network parameters through multiple rounds of iterative training until the model converges.

[0117] After training is complete, the data reconstruction and output phase (S9) begins. Electronic equipment inputs the missing seismic data to be processed into the trained target model, which then outputs the reconstructed complete seismic data. This reconstructed complete seismic data can be used in subsequent processing steps such as seismic imaging, velocity modeling, reservoir prediction, and full waveform inversion to improve the data quality and accuracy of these processes.

[0118] In this way, the network training mechanism constructed in this embodiment of the present disclosure, which integrates missing pattern perception, structural continuity constraints, and dynamic missing enhancement training, enables the target model to simultaneously learn the recovery rules of missing regions, seismic spectrum characteristics, and stratigraphic structural continuity characteristics, thereby improving the model's reconstruction accuracy and generalization ability in scenarios of random missing, rule-based missing, and large-area missing.

[0119] This disclosure also provides corresponding apparatus for implementing the above methods or steps. Figure 8 This is a schematic diagram of the structure of a device 800 for model training for seismic data reconstruction provided in an embodiment of this disclosure.

[0120] In this embodiment, the device 800 can be disposed within an electronic device and is understood as a part of the functional modules of the aforementioned electronic device. Specifically, the electronic device can be a server or a terminal, wherein the terminal can specifically include mobile phones, computers, or tablet computers, etc., without limitation.

[0121] like Figure 8 As shown, the device 800 includes: an acquisition module 810 for acquiring raw seismic data, which includes the number of time sampling points and the number of seismic traces; a determination module 820 for determining, based on the raw seismic data, the seismic data to be reconstructed and a first feature representation of the seismic data to be reconstructed, the first feature representation being used to characterize the missing regions and distribution characteristics of the seismic data to be reconstructed; an execution module 830 for performing multi-scale structural feature extraction operations on the seismic data to be reconstructed to determine a second feature representation of the seismic data to be reconstructed, the second feature representation being used to characterize the continuity of the structural features of the seismic data to be reconstructed; a processing module 840 for processing the seismic data to be reconstructed, the first feature representation, and the second feature representation using a target model to determine the reconstructed seismic data; and a training module 850 for training the target model based on the differences between the reconstructed seismic data and the raw seismic data.

[0122] In some embodiments, the execution module 830 may also be used to: perform convolution operations on the seismic data to be reconstructed using a set of convolution kernels of different scales to extract geological structural features at different scales; and fuse the geological structural features at different scales to determine a second feature representation of the seismic data to be reconstructed.

[0123] In some embodiments, the target model includes an encoder and a decoder, and the processing module 840 is further configured to: determine a structure-guided weight matrix based on the second feature representation, the structure-guided weight matrix being used to characterize the degree of structural continuity at each location in the seismic data to be reconstructed; perform enhancement processing on the third feature representation based on the structure-guided weight matrix to determine structure-guided enhancement features; and process the structure-guided enhancement features using the decoder to determine the reconstructed seismic data.

[0124] In some embodiments, the processing module 840 is further configured to: acquire multiple different levels of coding features generated by the encoder; perform a fusion operation on the structure-guided enhancement features and the multiple different levels of coding features to determine fused features; perform a convolution operation on the fused features to determine final reconstructed features; and process the final reconstructed features using the decoder.

[0125] In some embodiments, the training module 850 is further configured to: determine a joint loss based on the difference between the reconstructed seismic data and the original seismic data; and adjust the parameters of the target model based on the joint loss to train the target model.

[0126] In some embodiments, the joint loss includes at least one of the following: missing region reconstruction loss, used to characterize the reconstruction error between the reconstructed seismic data and the original seismic data in the missing region; spectral preservation loss, used to characterize the spectral difference between the reconstructed seismic data and the original seismic data; and structural continuity loss, used to characterize the structural continuity difference between the reconstructed seismic data and the original seismic data.

[0127] In some embodiments, the apparatus 800 may further include: an adjustment module, configured to gradually decrease the weight coefficient of the missing region reconstruction loss during the training of the target model; keep the weight coefficient of the spectrum preservation loss unchanged during the training of the target model; and gradually increase the weight coefficient of the structural continuity loss during the training of the target model.

[0128] It should be noted that, Figure 8 The apparatus 800 shown can perform the various steps in the above method embodiments and achieve the various processes and effects in the above method embodiments, which will not be elaborated here.

[0129] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. In this embodiment, Figure 9 The electronic devices shown can be servers or terminals, and terminals specifically include mobile phones, computers, or tablets, etc., without limitation.

[0130] like Figure 9As shown, the electronic device may include a processor 910 and a memory 920 storing computer program instructions.

[0131] Specifically, the processor 910 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this disclosure.

[0132] Memory 920 may include a large-capacity storage for information or instructions. For example, and not limitingly, memory 920 may include a hard disk drive (HDD), a floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 920 may include removable or non-removable (or fixed) media. Where appropriate, memory 920 may be internal or external to the integrated gateway device. In a particular embodiment, memory 920 is a non-volatile solid-state memory. In a particular embodiment, memory 920 includes read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (Electrically Programmable ROM, EPROM), an electrically erasable programmable PROM (EEPROM), an electrically alterable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0133] The processor 910 reads and executes computer program instructions stored in the memory 920 to perform the methods or steps provided in the embodiments of this disclosure.

[0134] In one example, the electronic device may also include a transceiver 930 and a bus 940. Wherein, as... Figure 9 As shown, the processor 910, memory 920 and transceiver 930 are connected via bus 940 and communicate with each other.

[0135] Bus 940 may include hardware, software, or both. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industrial Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 940 may include one or more buses.

[0136] This disclosure also provides a computer-readable storage medium that can store a computer program that, when executed by a processor, causes the processor to implement the methods provided in this disclosure.

[0137] The aforementioned storage medium may, for example, include a memory 920 containing computer program instructions, which can be executed by a processor 910 of an electronic device to perform the methods provided in the embodiments of this disclosure. Optionally, the storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), compact disc ROM (CD-ROM), magnetic tape, floppy disk, and optical data storage device.

[0138] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A model training method for seismic data reconstruction, characterized in that, include: Acquire raw seismic data, which includes the number of time sampling points and the number of seismic traces; Based on the original seismic data, the seismic data to be reconstructed and a first feature representation of the seismic data to be reconstructed are determined. The first feature representation is used to characterize the missing regions and distribution characteristics of the seismic data to be reconstructed. A multi-scale structural feature extraction operation is performed on the seismic data to be reconstructed to determine a second feature representation of the seismic data to be reconstructed, the second feature representation being used to characterize the continuity of the structural features of the seismic data to be reconstructed; Using the target model, the seismic data to be reconstructed, the first feature representation, and the second feature representation are processed to determine the reconstructed seismic data; The target model is trained based on the differences between the reconstructed seismic data and the original seismic data.

2. The method according to claim 1, characterized in that, The step of performing multi-scale structural feature extraction on the seismic data to be reconstructed to determine the second feature representation of the seismic data to be reconstructed includes: A set of convolution kernels of different scales are used to perform convolution operations on the seismic data to be reconstructed to extract geological structural features at different scales. The geological structural features at different scales are fused to determine the second feature representation of the seismic data to be reconstructed.

3. The method according to claim 1, characterized in that, The target model includes an encoder and a decoder. The process of using the target model to process the seismic data to be reconstructed, the first feature representation, and the second feature representation to determine the reconstructed seismic data includes: The encoder is used to process the seismic data to be reconstructed, the first feature representation, and the second feature representation to obtain a third feature representation; Based on the second feature representation, a structural guiding weight matrix is ​​determined, which is used to characterize the degree of structural continuity at each location in the seismic data to be reconstructed. Based on the structure-guided weight matrix, the third feature representation is enhanced to determine the structure-guided enhanced feature; The decoder is used to process the structure-guided enhancement features to determine the reconstructed seismic data.

4. The method according to claim 3, characterized in that, The decoder is used to process the structure-guided enhancement features, including: Obtain the encoded features at multiple different levels generated by the encoder; The structure-guided enhancement features are fused with the multiple coding features at different levels to determine the fused features; Perform a convolution operation on the fused features to determine the final reconstructed features; The decoder is used to process the final reconstructed features.

5. The method according to claim 1, characterized in that, The step of training the target model based on the difference between the reconstructed seismic data and the original seismic data includes: Based on the differences between the reconstructed seismic data and the original seismic data, the joint loss is determined; Based on the joint loss, the parameters of the target model are adjusted to train the target model.

6. The method according to claim 5, characterized in that, The joint loss includes at least one of the following: Missing area reconstruction loss is used to characterize the reconstruction error between the reconstructed seismic data and the original seismic data in the missing area. Spectrum preservation loss is used to characterize the spectral difference between the reconstructed seismic data and the original seismic data; Structural continuity loss is used to characterize the difference in structural continuity between the reconstructed seismic data and the original seismic data.

7. The method according to claim 6, characterized in that, The method further includes: During the training of the target model, the weight coefficients of the reconstruction loss for the missing region are gradually reduced; During the training of the target model, the weight coefficients of the spectrum preservation loss are kept constant; During the training of the target model, the weight coefficients of the structural continuity loss are gradually increased.

8. An apparatus for training a model for seismic data reconstruction, characterized in that, include: The acquisition module is used to acquire raw seismic data, which includes the number of time sampling points and the number of seismic traces; The determination module is used to determine, based on the original seismic data, the seismic data to be reconstructed and a first feature representation of the seismic data to be reconstructed, wherein the first feature representation is used to characterize the missing regions and distribution characteristics of the seismic data to be reconstructed; An execution module is used to perform multi-scale structural feature extraction operations on the seismic data to be reconstructed, so as to determine a second feature representation of the seismic data to be reconstructed, wherein the second feature representation is used to characterize the continuity of the structural features of the seismic data to be reconstructed; The processing module is used to process the seismic data to be reconstructed, the first feature representation, and the second feature representation using the target model, so as to determine the reconstructed seismic data; The training module is used to train the target model based on the differences between the reconstructed seismic data and the original seismic data.

9. An electronic device, characterized in that, The electronic device includes: a processor; a memory for storing executable instructions; wherein the processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, causes the processor to implement the method described in any one of claims 1-7.

Images