MPST network model-based double-shot aliasing noise removal method and device

Abstract

The invention discloses a double-shot aliasing noise removal method and device based on an MPST network model, and the method comprises the steps: obtaining a first seismic signal with noise and a second seismic signal without noise, and enabling the first seismic signal and the second seismic signal to serve as an original data set; dividing the original data set into a training set, a verification set and a test set according to a predetermined proportion; pre-training the training set, and training an initial MPST model based on the pre-trained training set; when the training frequency reaches a preset frequency, verifying the initial MPST model based on a verification set; when the verification result is in the preset stable range, the parameters of the MPST network model are updated to the network model parameters at the moment, noise removal is carried out based on the updated MPST network model, aliasing noise can be effectively removed through the MPST model under the condition that seismic data are obviously insufficient, and the noise removal efficiency and effect are improved.

Description

Technical Field

[0001] The present application relates to the technical field of double-shot aliasing noise removal, and more specifically, to a double-shot aliasing noise removal method and device based on an MPST network model. Background Art

[0002] Currently, oil reservoirs cannot meet social demand, and undiscovered reserves remain a critical task. Therefore, in recent years, my country has intensified its exploration efforts to secure oil and gas resources. As exploration areas expand, the traditional single-shot acquisition method, which is inefficient, time-consuming, and costly, is gradually being replaced by multi-shot acquisition. This approach not only saves costs, improves efficiency, and ensures acquisition quality. However, since the aliased data from multi-shot acquisition cannot be directly used in subsequent processing and interpretation, alias separation is required to ensure data quality for subsequent processing. Therefore, as aliased acquisition becomes the mainstream, alias separation technology has become increasingly important. Therefore, developing a method for efficiently separating aliased data is of great practical significance.

[0003] Here, one shot from a double-shot aliasing is treated as the signal, while the other shot is treated as aliasing noise. Because aliasing data separation requires the complete separation of similar waves, the aliasing noise must be completely removed while preserving the energy of the single-shot signal. Traditional methods for removing aliasing noise utilize filtering, but due to the similarity of the two-shot signals and their similar energy intensities, filtering methods can easily lose significant signal energy. Furthermore, the large volume of seismic data places high demands on computing power.

[0004] Therefore, a better solution is urgently needed. Summary of the Invention

[0005] The present invention provides a method and device for removing double-shot aliasing noise based on an MPST network model, which is used to solve the technical problems of poor double-shot aliasing noise removal effect and low efficiency in the prior art. The method comprises:

[0006] Acquire a first seismic signal with noise and a second seismic signal without noise, and use the first seismic signal and the second seismic signal as original data sets;

[0007] Dividing the original data set into a training set, a validation set, and a test set according to a predetermined ratio;

[0008] Pre-training the training set, and training an initial MPST model based on the pre-trained training set;

[0009] When the number of training times reaches a predetermined number, verifying the initial MPST model based on the verification set;

[0010] When the verification result is within the preset stable range, the parameters of the MPST network model are updated to the network model parameters at this time, and noise removal is performed based on the updated MPST network model.

[0011] In some specific embodiments, the training set is pre-trained, specifically:

[0012] Special values and unexpected values in the training set are removed, and the training set after the special values and unexpected values are removed is cut into a set of signal slices of a predetermined size.

[0013] In some specific embodiments, the initial MPST model is trained based on the pre-trained training set, specifically:

[0014] Three parallel branches are pre-set, and the input original seismic signal data is downsampled using a multi-scale segmentation method to obtain a first segmentation result, a second segmentation result, and a third segmentation result. The first segmentation result, the second segmentation result, and the third segmentation result correspond to the segmentation results obtained under each of the three parallel branches, respectively. The size of the feature map output by the first segmentation result is half the size of the input original data, the size of the feature map output by the second segmentation result is one-quarter the size of the input original data, and the size of the feature map output by the third segmentation result is one-eighth the size of the input original data;

[0015] In each branch, self-attention features are extracted on the first segmentation result, the second segmentation result, and the third segmentation result through a multi-head self-attention mechanism to obtain a first downsampling result, a second downsampling result, and a third downsampling result;

[0016] Inputting the first down-sampling result, the second down-sampling result, and the third down-sampling result into an up-sampling layer and performing up-sampling to obtain an up-sampling result;

[0017] Performing feature fusion on the upsampling results to obtain a fused signal feature map;

[0018] The fused signal feature map is input into the convolution layer to obtain a denoised seismic signal.

[0019] In some specific embodiments, the downsampling of the pre-trained training set is performed to obtain a downsampling result by a downsampling formula, and the downsampling formula is specifically:

[0020]

[0021] Among them, H r is the size of the downsampling result after downsampling, H o is the size of the input signal during downsampling, K is the convolution kernel size, and B ris the number of channels of the downsampling result, S is the step size of the convolution kernel movement during the downsampling process, and P is the number of circles that need to fill the edge value of the input signal.

[0022] In some specific embodiments, the multi-head self-attention mechanism includes W-MSA window attention and SW-MSA window attention, three parallel branches are pre-set, and the downsampling result is multi-scale segmented by the multi-head self-attention mechanism to obtain a first segmentation result, a second segmentation result, and a third segmentation result, which are specifically determined by the following formula:

[0023]

[0024] in, Represents the output signal characteristic diagram of W-MSA, Z l represents the output signal feature map of MLP, It represents the output signal characteristic graph of SW-MSA, l+1 represents the number of network layers running in the next layer, and l-1 represents the number of network layers running in the previous layer;

[0025] Q i =QW i Q ,K i =KW i K ,V i =VW i V ;

[0026] Where i is the number of heads in the multi-head self-attention mechanism, Q represents the query in the attention mechanism, K represents the key in the attention mechanism, and V represents the value in the attention mechanism. i Q 、W i K 、W i V is the parameter matrix corresponding to the calculation of Q, K, and V;

[0027] MultiHead(Q,K,V)=concact(F 1, F 2, F3...F i );

[0028] Among them, MulitiHead represents the multi-head attention mechanism, and concact represents the weight matrix for jointly calculating self-attention.

[0029] In some specific embodiments, the first downsampling result, the second downsampling result, and the third downsampling result are input into an upsampling layer and upsampled to obtain an upsampling result. The upsampling layer includes a first upsampling module, a second upsampling module, a third upsampling module, a fourth upsampling module, a fifth upsampling module, and a sixth upsampling module, specifically:

[0030] Upsampling the third downsampling result by the third upsampling module to obtain a third upsampling result, wherein the segmentation scale of the feature map in the third upsampling result is 4;

[0031] Upsampling the third upsampling result by the fifth upsampling module to obtain a fifth upsampling result, wherein the segmentation scale of the feature map in the fifth upsampling result is 2;

[0032] Upsampling the fifth upsampling result by the sixth upsampling module to obtain a sixth upsampling result, wherein the segmentation scale of the feature map in the sixth upsampling result is 1;

[0033] The second down-sampling result and the third up-sampling result are used as inputs of a second up-sampling module, and up-sampled by the second up-sampling module to obtain a second up-sampling result, wherein the segmentation scale of the feature map in the second up-sampling result is 2;

[0034] Using the second upsampling result as an input of a fourth upsampling module, performing upsampling by the fourth upsampling module to obtain a fourth upsampling result, wherein the segmentation scale of the feature map in the fourth upsampling result is 1;

[0035] The first downsampling result, the fifth upsampling result, and the second upsampling result are used as inputs of a first upsampling module, and upsampling is performed by the first upsampling module to obtain a first upsampling result, wherein the segmentation scale of the feature map in the first upsampling result is 1;

[0036] The first up-sampling result, the fourth up-sampling result, and the sixth up-sampling result are used as the up-sampling results.

[0037] In some specific embodiments, the upsampling module in the upsampling layer is implemented by an upsampling formula, which is specifically:

[0038]

[0039] Among them, H ois the signal size after deconvolution, H is the input signal size of the upsampling layer, S is the convolution kernel moving step, P is the edge padding value of the input signal of the upsampling layer, K is the deconvolution kernel size, fixed to 3×3, B0 and B are the number of channels before and after deconvolution, respectively.

[0040] Accordingly, the present invention also proposes a double-shot aliasing noise removal device based on the MPST network model, the device comprising:

[0041] an acquisition module, configured to acquire a first seismic signal with noise and a second seismic signal without noise, and use the first seismic signal and the second seismic signal as original data sets;

[0042] A partitioning module is used to divide the original data set into a training set, a validation set and a test set according to a predetermined ratio;

[0043] A training module, configured to pre-train the training set and train an initial MPST model based on the pre-trained training set;

[0044] A verification module, configured to verify the initial MPST model based on a verification set when the number of training times reaches a predetermined number;

[0045] The noise removal module is used to update the parameters of the MPST network model to the network model parameters at this time when the verification result is within a preset stable range, and to perform noise removal based on the updated MPST network model.

[0046] One embodiment of the present invention also provides a computing device, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the double-shot aliasing noise removal method based on the MPST network model as described above are implemented.

[0047] One embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the double-shot aliasing noise removal method based on the MPST network model as described above.

[0048] By applying the above technical solution, a double-shot aliasing noise removal method based on the MPST network model is proposed, the method comprising: obtaining a first seismic signal with noise and a second seismic signal without noise, and using the first seismic signal and the second seismic signal as original data sets; dividing the original data sets into a training set, a validation set and a test set according to a predetermined ratio; pre-training the training set, and training the initial MPST model based on the pre-trained training set; when the number of training times reaches a predetermined number, verifying the initial MPST model based on the validation set; when the verification result is within a preset stable range, updating the parameters of the MPST network model to the network model parameters at this time, and performing noise removal based on the updated MPST network model. When the seismic data is obviously insufficient, the MPST model can effectively remove the aliasing noise and improve the efficiency and effect of noise removal. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0050] Figure 1 This is a flow chart of a method for removing double-shot aliasing noise based on an MPST network model provided in an embodiment of the present application;

[0051] Figure 2 Schematic diagram of the W-MSA segmentation process provided in an embodiment of the present application;

[0052] Figure 3 Schematic diagram of the SW-MSA process provided in an embodiment of the present application;

[0053] Figure 4 is a schematic diagram of experimental results of horizontal layered model simulation data provided in an embodiment of the present application;

[0054] Figure 5 This is a residual comparison diagram of the horizontal layered model simulation data provided in the embodiment of the present application;

[0055] Figure 6 3. It is a schematic diagram of the experimental results of the three-dimensional overthrust simulation data provided in the embodiment of the present application;

[0056] Figure 7 This is a residual comparison diagram of the simulation data of the three-dimensional overthrust model provided in the embodiment of the present application;

[0057] Figure 8 This is a schematic diagram of the structure of the upsampling process provided in an embodiment of the present application;

[0058] Figure 9 1 is a schematic structural diagram of a double-shot aliasing noise removal device based on an MPST network model provided in an embodiment of the present application;

[0059] Figure 10 This is a structural block diagram of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] The following description sets forth many specific details to facilitate a thorough understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0061] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms "a," "an," and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0062] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0063] like Figure 1 As shown, the present application proposes a double-shot aliasing noise removal method based on the MPST network model, the method comprising the following steps:

[0064] Step S101 : Acquire a first seismic signal with noise and a second seismic signal without noise, and use the first seismic signal and the second seismic signal as original data sets.

[0065] Step S102: Divide the original data set into a training set, a validation set, and a test set according to a predetermined ratio.

[0066] In this embodiment, seismic signals with noise and seismic signals without noise are first obtained as training data sets for the model. To facilitate subsequent verification, the data sets are divided into training set, validation set and test set in a ratio of 7:1:1. It should be noted that the above division ratio is only used as a preferred scheme, and not a limitation on the division ratio. Therefore, it can be understood that those skilled in the art can adjust the division ratio according to actual needs.

[0067] Step S103: pre-train the training set, and train the initial MPST model based on the pre-trained training set.

[0068] In a possible implementation, pre-training is performed on the training set, specifically:

[0069] Special values and unexpected values in the training set are removed, and the training set after the special values and unexpected values are removed is cut into a set of signal slices of a predetermined size.

[0070] In this embodiment, the seismic signals with noise and the seismic signals without noise in the training set are preprocessed and clipped. The preprocessing can exclude special values and unexpected values in the seismic signals, and the clipping is to clip the seismic signals into data of a fixed size so that they can be input into the MPST model for training. The noisy data is input as a sample, and the noise-free data is input as a label. For example, when using the network model to process seismic signals with aliasing noise, the noisy signal graph and the noise-free signal graph need to be clipped into a set of signal slices of 255×255 size (the size can be adjusted as needed), and then this set of signal slices is input into the network model for training.

[0071] In one possible implementation, the initial MPST model is trained based on the pre-trained training set, specifically:

[0072] Three parallel branches are pre-set, and the input original seismic signal data is downsampled using a multi-scale segmentation method to obtain a first segmentation result, a second segmentation result, and a third segmentation result. The first segmentation result, the second segmentation result, and the third segmentation result correspond to the segmentation results obtained under each of the three parallel branches, respectively. The size of the feature map output by the first segmentation result is half the size of the input original data, the size of the feature map output by the second segmentation result is one-quarter the size of the input original data, and the size of the feature map output by the third segmentation result is one-eighth the size of the input original data;

[0073] In each branch, self-attention features are extracted on the first segmentation result, the second segmentation result, and the third segmentation result through a multi-head self-attention mechanism to obtain a first downsampling result, a second downsampling result, and a third downsampling result;

[0074] Inputting the first down-sampling result, the second down-sampling result, and the third down-sampling result into an up-sampling layer and performing up-sampling to obtain an up-sampling result;

[0075] Performing feature fusion on the upsampling results to obtain a fused signal feature map;

[0076] The fused signal feature map is input into the convolution layer to obtain a denoised seismic signal.

[0077] In this embodiment, the segmentation scales of the three parallel branches are different. Specifically, the current model is divided into three scales, wherein the segmentation scales of the original seismic signal are 8, 4, and 2 respectively.

[0078] In this embodiment, the first segmentation result, the second segmentation result and the third segmentation result generated by each branch are input to the upsampling layer for upsampling, and the upsampling layer finally outputs the upsampling result. After the upsampling result is feature fused, a fused signal feature map is obtained.

[0079] In a possible implementation, downsampling the pre-trained training set to obtain a downsampling result is performed using a downsampling formula, and the downsampling formula is specifically:

[0080]

[0081] Among them, H r is the size of the downsampling result after downsampling, H o is the size of the input signal during downsampling, K is the convolution kernel size, and B r is the number of channels of the downsampling result, S is the step size of the convolution kernel movement during the downsampling process, and P is the number of circles that need to fill the edge value of the input signal.

[0082] In this embodiment, after the seismic signal is input into the MPST model, its features need to be extracted and it needs to enter the coding layer according to the coding method. In order to ensure the computing power of the seismic signal, it is usually divided into a series of slices of equal size. However, in order to increase the computing power, obtain slices of different sizes at the same time, and ensure the attention of different perspectives, three branches are added for segmentation. In order to ensure the normal computing power of most computers, the three branches are fixed to 8, 4, and 2 in parallel in the present invention, and the results are finally combined by averaging. The signal needs to be downsampled before segmentation. The downsampling formula is expressed as follows:

[0083]

[0084] Where H r is the size of the output signal after downsampling and embedding, H o is the original input signal size, K is the convolution kernel size, Br is the number of channels in the output signal after downsampling and embedding. Due to the need for downsampling and embedding, S in the formula is the step size of the convolution kernel during the downsampling process, which is fixed to 1 in actual calculations. P is the number of cycles required to fill the edge values of the input signal, which is fixed to 0 here.

[0085] In one possible implementation, the multi-head self-attention mechanism includes W-MSA window attention and SW-MSA window attention, pre-sets three parallel branches, and performs multi-scale segmentation on the downsampling result through the multi-head self-attention mechanism to obtain a first segmentation result, a second segmentation result, and a third segmentation result, which are specifically determined by the following formula:

[0086]

[0087] in, Represents the output signal characteristic diagram of W-MSA, Z l represents the output signal feature map of MLP, It represents the output signal characteristic graph of SW-MSA, l represents the number of network layers of the model operation, l+1 represents the number of network layers of the next layer operation, and l-1 represents the number of network layers of the previous layer;

[0088] Q i =QW i Q ,K i =KW i K ,V i =VW i V ;

[0089] Where i is the number of heads in the multi-head self-attention mechanism, Q represents the query in the attention mechanism, K represents the key in the attention mechanism, and V represents the value in the attention mechanism. i Q 、W i K 、W i V is the parameter matrix corresponding to the calculation of Q, K, and V;

[0090] MultiHead(Q,K,V)=concact(F 1, F 2, F3...F i );

[0091] Among them, MulitiHead represents the multi-head attention mechanism, and concact represents the weight matrix for jointly calculating self-attention.

[0092] In this embodiment, Figure 2 and Figure 3 As shown in Figure 2, they are schematic diagrams of the segmentation of W-MSA window attention and SW-MSA window attention.

[0093] In this example, the first row of formulas describes the overall process of the network structure. The second row of formulas describes the core formula of the entire network structure, the attention mechanism. The third row of formulas describes the multi-head attention formula, which is also a fundamental core formula in self-attention. Specifically, the W-MSA in the first row of formulas contains the third row of formulas, which in turn contains the second row of formulas.

[0094] In this embodiment, MLP is a substructure in the network model, which is used to perform fusion processing on the feature maps after self-attention, and the model is a multi-layer perceptron.

[0095] In one possible implementation, the first downsampling result, the second downsampling result, and the third downsampling result are input into an upsampling layer and upsampled to obtain an upsampling result. The upsampling layer includes a first upsampling module, a second upsampling module, a third upsampling module, a fourth upsampling module, a fifth upsampling module, and a sixth upsampling module, specifically:

[0096] Upsampling the third downsampling result by the third upsampling module to obtain a third upsampling result, wherein the segmentation scale of the feature map in the third upsampling result is 4;

[0097] Upsampling the third upsampling result by the fifth upsampling module to obtain a fifth upsampling result, wherein the segmentation scale of the feature map in the fifth upsampling result is 2;

[0098] Upsampling the fifth upsampling result by the sixth upsampling module to obtain a sixth upsampling result, wherein the segmentation scale of the feature map in the sixth upsampling result is 1;

[0099] The second down-sampling result and the third up-sampling result are used as inputs of a second up-sampling module, and up-sampled by the second up-sampling module to obtain a second up-sampling result, wherein the segmentation scale of the feature map in the second up-sampling result is 2;

[0100] Using the second upsampling result as an input of a fourth upsampling module, performing upsampling by the fourth upsampling module to obtain a fourth upsampling result, wherein the segmentation scale of the feature map in the fourth upsampling result is 1;

[0101] The first downsampling result, the fifth upsampling result, and the second upsampling result are used as inputs of a first upsampling module, and upsampling is performed by the first upsampling module to obtain a first upsampling result, wherein the segmentation scale of the feature map in the first upsampling result is 1;

[0102] The first up-sampling result, the fourth up-sampling result, and the sixth up-sampling result are used as the up-sampling results.

[0103] In this embodiment, the multi-scale aggregation Swin-Transformer model proposed in the present invention improves the model structure on the basis of the original model, innovatively proposes features at different levels, and combines these features in an aggregated connection manner. Unlike the original model, the newly proposed network model pays more attention to the connection between features and improves the connection method between features. It is not just a parallel operation of different branches, but uses the features extracted at different scales to aggregate and connect layer by layer.

[0104] When the MPST model calculates the seismic signal features at each scale, the branch with a larger segmentation scale will gradually increase the signal feature size in the branch during the decoding process. At this time, the size of the restored signal feature is the same as the size of the seismic signal feature map calculated by other branches with smaller segmentation scales.

[0105] Based on this, the seismic signal feature map connecting these two branches serves as the input to the next layer. The current model is divided into three scales, with the original seismic signal segmented at scales of 8, 4, and 2, respectively. During the upsampling process after segmentation, the signal feature map with a segmentation scale of 8 can be channel-fused with the signal feature map with a segmentation scale of 4 after a single deconvolution, and serve as the input to the next upsampling layer. Similarly, the signal feature map with a segmentation scale of 4 can be channel-fused with the signal feature map with a segmentation scale of 2 after a single deconvolution, and serve as the input to the subsequent layer. While focusing on both fine and global features, it can better integrate features from the previous scale into the next scale, avoiding signal loss.

[0106] In this embodiment, the specific upsampling process is as follows: Figure 8 As shown, the first down-sampling result, the second down-sampling result and the third down-sampling result are input into the up-sampling layer and up-sampled to obtain the up-sampling result. Since the sizes of the three down-sampling results are different, that is, the sizes of the first down-sampling result, the second down-sampling result and the third down-sampling result are respectively one-half, one-quarter and one-eighth of the original data size, the up-sampling layer also has three branches, which respectively restore the sizes of the three down-sampling results.

[0107] When upsampling the third downsampling result, the upsampling branch has three upsampling modules. After calculation by the third upsampling module, the resulting feature map size is one-quarter the original data size. After calculation by the fifth upsampling module, the resulting feature map size is one-half the original data size. After calculation by the sixth upsampling module, the resulting feature map size is the original data size.

[0108] When upsampling the second downsampling result, this upsampling branch has two upsampling modules. Since the size of the feature map obtained after the third downsampling result passes through the third upsampling module is the same as the size of the current second downsampling result, this feature map can be fused with the second downsampling result through channel stacking and input into the second upsampling module of this branch. After calculation by the second upsampling module, the resulting feature map size is half of the original data size. After calculation by the fourth upsampling module, the resulting feature map size is the original data size.

[0109] When upsampling the first downsampling result, the upsampling branch has a common upsampling module. Because the size of the feature map obtained after the third downsampling result passes through the second upsampling module and the size of the feature map obtained after the second downsampling result passes through one upsampling module are the same as the size of the current first downsampling result, the feature maps of these three branches can be fused through channel stacking and input into the first upsampling module of this branch. After calculation by the first upsampling module, the resulting feature map size is the original data size.

[0110] After completing the above three branches, that is, the results of the sixth upsampling module, the fourth upsampling module, and the first upsampling module are three feature maps of the same size as the original data. The three feature maps are fused through channel stacking and convolution operations to output the final denoising result.

[0111] In one possible implementation, the first downsampling result, the second downsampling result, and the third downsampling result are input into an upsampling layer and upsampled to obtain an upsampling result, which is specifically achieved by an upsampling formula, and the upsampling formula is specifically:

[0112]

[0113] Among them, H o is the signal size after deconvolution, H is the input signal size of the upsampling layer, S is the convolution kernel moving step, P is the edge padding value of the input signal of the upsampling layer, K is the deconvolution kernel size, fixed to 3×3, B0 and B are the number of channels before and after deconvolution, respectively.

[0114] In this embodiment, after downsampling, the image is reduced, and deconvolution is used to restore the feature map to its original size. The deconvolution formula is as follows:

[0115]

[0116] Where H ois the size of the signal after deconvolution, i.e., the size of the original input signal. H is the size of the input signal. S is the convolution kernel movement step, which is fixed at 1. P is the input signal edge padding value, which is fixed at 0. K is the deconvolution kernel size, which is fixed at 3×3. B0 and B are the number of channels before and after deconvolution, respectively.

[0117] After feature extraction is completed in each scale branch, a signal feature map with the same scale is ultimately obtained. At this point, each branch generates an output. To fuse the results at each scale, a multi-scale signal feature fusion output layer is added at the end of the model. The multi-scale signal feature fusion output layer performs channel fusion on the feature signal maps at different scales. The signal feature maps with different channels are then input into the convolutional layer. Convolution is then used to map the relationship between the seismic signal feature map and the true signal, ultimately yielding a denoised seismic signal.

[0118] Step S104: When the number of training times reaches a predetermined number, the initial MPST model is verified based on a verification set.

[0119] In this embodiment, data is input into the MPST model for processing. If the current number of iterations does not reach the preset value, the next round of training is continued. Once the required number of iterations is met, the validation set is used to evaluate network performance. If the validation results meet the requirements, training ends and the network model parameters are saved.

[0120] During the network model training process, the real-time status and training saturation of the network model can be viewed through the validation set. When the results of the validation set tend to be stable, the network model parameters are saved. At this time, the MPST model has strong aliasing noise suppression performance.

[0121] Finally, the data in the test set is input into the MPST network model that performs well on the validation set. At this time, the aliasing noise suppression effect demonstrated by the MPST network model on the test set is the most realistic MPST model's effect of suppressing aliasing noise in seismic signals.

[0122] By inputting the original noise-free signal and the signal with double-shot aliasing noise, the parameters in the model are continuously updated during the training process, and a mapping relationship between the original noise-free signal and the signal with double-shot aliasing noise is established. The aliasing noise that needs to be removed in the double-shot aliasing noise is identified on a large scale, while the denoised original signal is restored on a small scale.

[0123] Step S105: When the verification result is within a preset stable range, the parameters of the MPST network model are updated to the current network model parameters, and noise removal is performed based on the updated MPST network model.

[0124] In addition, this application also evaluates the noise removal effect of the MPST model, such as Figure 4 and Figure 5 Shown are the experimental results of the horizontal layered model simulation data and the residual comparison diagram of the horizontal layered model simulation data.

[0125] 1. Horizontal layered model experiment

[0126] First, by establishing a three-dimensional horizontal layered model and setting five layers of different speeds, the speeds from top to bottom are 3000m / s, 3200m / s, 3350m / s, 3750m / s, and 4000m / s, respectively, to verify the effect of Multiscale Polymerization Swin-Transformer (MPST) on the simulated aliasing noise generated in the horizontal layered model. The GAN network, Unet network, and Swin-Transformer network commonly used in the field of seismic denoising are used as comparison networks to evaluate the feasibility of MPST for aliasing noise suppression. The test results are as follows Figure 4 From the result graph, Unet has a better effect on signal removal in the aliasing area, but there are still some residual signals in the non-aliasing area that are not suppressed cleanly, and there is a small part of the signal missing in the lower right corner; Swin-Transformer also has the problem of a small part of the signal missing in the lower right corner, but the overall signal removal is not clean and there are residuals, and the structure of the first arrival wave is changed, and the original signal is changed; GAN and MPST, from the result graph, both completely suppress the aliasing noise, and the naked eye cannot see Figure 4 (e) and 4(f) show the difference. Therefore, we will calculate the residual to assist in the judgment.

[0127] Figure 5 (a)-(d) are the residual results of Unet, Swin-Transformer, GAN, and MPST. Figure 5 (a) shows that Unet has a good removal effect, and the aliasing noise is almost completely removed, but there are still some local residues, and a large amount of valid signals also remain, indicating that the valid signals are not well preserved. Figure 5 (b) It can be seen that the Swin-Transformer removal effect is poor, and a large amount of aliasing noise remains. At the same time, the effective signal is also removed, and the newly added noise changes the effective signal at the top. Figure 5 (c) It can be seen that the experimental results of the GAN network are better than those of the Unet and Swin-Transformer methods. The aliasing noise is effectively removed and the original signal is well preserved, but there are still some local effective signals remaining. Figure 5(d) is the residual result of the MPST method proposed in this chapter. It can be seen from the figure that the aliasing noise is completely removed and the effective signal is also completely retained, and the removal performance is the best.

[0128] 2. Overthrust model experiment

[0129] After completing the horizontal layered model experiments, further experiments were conducted on an overthrust model to further verify the feasibility of the MPST method. The experimental results show that while the Unet method attenuates some noise, it cannot effectively suppress aliasing noise. Figure 6 (d) is the experimental result of the Swin-Transformer method. It can be seen from the figure that this method can hardly remove the aliasing noise, and it will also change the structure of the original signal and produce a large amount of new noise. Figure 6 (e) shows the removal effect of the GAN method. It can be seen from the figure that this method is also unable to effectively remove aliasing noise on complex data, and the intensity of the effective signal energy is also weakened. Figure 6 (f) is the MPST method, which can not only effectively remove aliasing noise, but also keep the energy intensity consistent with the original effective signal, and the effective signal structure does not change. Compared with other comparative experimental methods, the MPST method has the best suppression effect, which proves the applicability and feasibility of the MPST method.

[0130] The residual results can more intuitively judge the ability of the four methods to remove noise and retain effective signals. Since the three-dimensional overthrust model is more inclined to the actual underground situation, it is necessary to conduct simulation experiments. And through this experiment, it was found that the results of the GAN network deteriorated when training complex simulated data, and it could not effectively suppress aliasing noise, such as Figure 7 (c) shown. Figure 7 (a) is the Unet residual result. From the figure, we can see that the noise is not completely suppressed, and the energy of the effective signal is also lost. Figure 7 (b) is the residual result of Swin-Transformer. From the figure, the effect is poor, the original signal is changed, and new noise is added. Figure 7 (d) is the residual result of the MPST network proposed in this paper. It can be seen from this figure that while the aliasing noise is removed, the effective signal is retained as much as possible. Although a small amount of effective signal is also removed, overall, this network has the best performance.

[0131] The embodiment of the present application also proposes a double-shot aliasing noise removal device based on the MPST network model, such as Figure 9 As shown, the device includes:

[0132] An acquisition module 10 is configured to acquire a first seismic signal with noise and a second seismic signal without noise, and use the first seismic signal and the second seismic signal as original data sets;

[0133] A partitioning module 20 is configured to partition the original data set into a training set, a validation set, and a test set according to a predetermined ratio;

[0134] A training module 30 is configured to pre-train the training set and train an initial MPST model based on the pre-trained training set;

[0135] A verification module 40 is configured to verify the initial MPST model based on a verification set when the number of training times reaches a predetermined number;

[0136] The noise removal module 50 is used to update the parameters of the MPST network model to the current network model parameters when the verification result is within a preset stable range, and perform noise removal based on the updated MPST network model.

[0137] Figure 10 The block diagram of a computing device 400 according to one embodiment of the present disclosure is shown. Components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and a database 450 is used to store data.

[0138] The computing device 400 also includes an access device 440 that enables the computing device 400 to communicate via one or more networks 460. Examples of such networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 440 may include one or more of any type of network interface (e.g., a network interface card (NIC)) whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, or a near field communication (NFC) interface.

[0139] In one embodiment of the present specification, the above components of the computing device 400 and Figure 10 Other components not shown in the figure may also be connected to each other, for example, via a bus. Figure 10 The computing device structure block diagram shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art may add or replace other components as needed.

[0140] Computing device 400 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook computer, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). Computing device 400 may also be a mobile or stationary server.

[0141] The processor 420 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-described method for removing two-shot aliasing noise based on the MPST network model. The above is a schematic diagram of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the above-described method for removing two-shot aliasing noise based on the MPST network model are based on the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the above-described method for removing two-shot aliasing noise based on the MPST network model.

[0142] An embodiment of the present specification further provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned method for removing double-shot aliasing noise based on the MPST network model.

[0143] The above is an illustrative embodiment of a computer-readable storage medium. It should be noted that the technical solution of this storage medium is based on the same concept as the technical solution of the MPST network model-based dual-shot aliasing noise removal method. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the MPST network model-based dual-shot aliasing noise removal method.

[0144] An embodiment of the present specification further provides a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the above-mentioned method for removing double-shot aliasing noise based on the MPST network model.

[0145] The above is an illustrative embodiment of a computer program according to this embodiment. It should be noted that the technical solution of this computer program is based on the same concept as the technical solution of the MPST network model-based dual-shot aliasing noise removal method. For details not described in detail in the computer program's technical solution, please refer to the description of the technical solution of the MPST network model-based dual-shot aliasing noise removal method.

[0146] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0147] The computer instructions include computer program code, which may be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0148] It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0149] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0150] The preferred embodiments disclosed above are intended only to help illustrate this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made based on the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that those skilled in the art can better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A double-shot aliasing noise removal method based on the MPST network model, characterized in that: The method comprises: Acquire a first seismic signal with noise and a second seismic signal without noise, and use the first seismic signal and the second seismic signal as original data sets; Dividing the original data set into a training set, a validation set, and a test set according to a predetermined ratio; Pre-training the training set, and training an initial MPST model based on the pre-trained training set; When the number of training times reaches a predetermined number, verifying the initial MPST model based on the verification set; When the verification result is within the preset stable range, the parameters of the MPST network model are updated to the network model parameters at this time, and noise removal is performed based on the updated MPST network model.

2. The method according to claim 1, characterized in that Pre-training is performed on the training set, specifically: Special values and unexpected values in the training set are removed, and the training set after the special values and unexpected values are removed is cut into a set of signal slices of a predetermined size.

3. The method according to claim 1, characterized in that The initial MPST model is trained based on the pre-trained training set, specifically: Three parallel branches are pre-set, and the input original seismic signal data is downsampled using a multi-scale segmentation method to obtain a first segmentation result, a second segmentation result, and a third segmentation result. The first segmentation result, the second segmentation result, and the third segmentation result correspond to the segmentation results obtained under each of the three parallel branches, respectively. The size of the feature map output by the first segmentation result is half the size of the input original data, the size of the feature map output by the second segmentation result is one-quarter the size of the input original data, and the size of the feature map output by the third segmentation result is one-eighth the size of the input original data; In each branch, self-attention features are extracted on the first segmentation result, the second segmentation result, and the third segmentation result through a multi-head self-attention mechanism to obtain a first downsampling result, a second downsampling result, and a third downsampling result; Inputting the first down-sampling result, the second down-sampling result, and the third down-sampling result into an up-sampling layer and performing up-sampling to obtain an up-sampling result; Performing feature fusion on the upsampling results to obtain a fused signal feature map; The fused signal feature map is input into the convolution layer to obtain a denoised seismic signal.

4. The method according to claim 3, characterized in that The downsampling of the pre-trained training set is performed to obtain a downsampling result by a downsampling formula, and the downsampling formula is specifically: Among them, H r is the size of the downsampling result after downsampling, H o is the size of the input signal during downsampling, K is the convolution kernel size, and B r is the number of channels of the downsampling result, S is the step size of the convolution kernel movement during the downsampling process, and P is the number of circles that need to fill the edge value of the input signal.

5. The method according to claim 3, characterized in that The multi-head self-attention mechanism includes W-MSA window attention and SW-MSA window attention. Three parallel branches are pre-set, and the downsampling result is multi-scale segmented by the multi-head self-attention mechanism to obtain the first segmentation result, the second segmentation result and the third segmentation result, which are specifically determined by the following formula: in, Represents the output signal characteristic diagram of W-MSA, Z l Represents the output signal feature map of MLP, It represents the output signal characteristic graph of SW-MSA, l represents the number of network layers of the model operation, l+1 represents the number of network layers of the next layer operation, and l-1 represents the number of network layers of the previous layer; Q i =QW i Q ,K i =KW i K ,V i =VW i V ; Where i is the number of heads in the multi-head self-attention mechanism, Q represents the query in the attention mechanism, K represents the key in the attention mechanism, and V represents the value in the attention mechanism. i Q 、W i K 、W i V is the parameter matrix corresponding to the calculation of Q, K, and V; MultiHead(Q,K,V)=concact(F 1, F 2, F3...F i ); Among them, MulitiHead represents the multi-head attention mechanism, and concact represents the weight matrix for jointly calculating self-attention.

6. The method according to claim 3, characterized in that The first downsampling result, the second downsampling result, and the third downsampling result are input into an upsampling layer and upsampled to obtain an upsampling result. The upsampling layer includes a first upsampling module, a second upsampling module, a third upsampling module, a fourth upsampling module, a fifth upsampling module, and a sixth upsampling module, specifically: Upsampling the third downsampling result by the third upsampling module to obtain a third upsampling result, wherein the segmentation scale of the feature map in the third upsampling result is 4; Upsampling the third upsampling result by the fifth upsampling module to obtain a fifth upsampling result, wherein the segmentation scale of the feature map in the fifth upsampling result is 2; Upsampling the fifth upsampling result by the sixth upsampling module to obtain a sixth upsampling result, wherein the segmentation scale of the feature map in the sixth upsampling result is 1; The second down-sampling result and the third up-sampling result are used as inputs of a second up-sampling module, and up-sampled by the second up-sampling module to obtain a second up-sampling result, wherein the segmentation scale of the feature map in the second up-sampling result is 2; Using the second upsampling result as an input of a fourth upsampling module, performing upsampling by the fourth upsampling module to obtain a fourth upsampling result, wherein the segmentation scale of the feature map in the fourth upsampling result is 1; The first downsampling result, the fifth upsampling result, and the second upsampling result are used as inputs of a first upsampling module, and upsampling is performed by the first upsampling module to obtain a first upsampling result, wherein the segmentation scale of the feature map in the first upsampling result is 1; The first up-sampling result, the fourth up-sampling result, and the sixth up-sampling result are used as the up-sampling results.

7. The method according to claim 6, characterized in that The first downsampling result, the second downsampling result, and the third downsampling result are input into the upsampling layer and upsampled to obtain an upsampling result, which is specifically achieved through an upsampling formula. The upsampling formula is specifically: Among them, H o is the signal size after deconvolution, H is the input signal size of the upsampling layer, S is the convolution kernel moving step, P is the edge padding value of the input signal of the upsampling layer, K is the deconvolution kernel size, fixed to 3×3, B0 and B are the number of channels before and after deconvolution, respectively.

8. A double-shot aliasing noise removal device based on the MPST network model, characterized in that: The device comprises: an acquisition module, configured to acquire a first seismic signal with noise and a second seismic signal without noise, and use the first seismic signal and the second seismic signal as original data sets; A partitioning module is used to divide the original data set into a training set, a validation set and a test set according to a predetermined ratio; A training module, configured to pre-train the training set and train an initial MPST model based on the pre-trained training set; A verification module, configured to verify the initial MPST model based on a verification set when the number of training times reaches a predetermined number; The noise removal module is used to update the parameters of the MPST network model to the network model parameters at this time when the verification result is within a preset stable range, and to perform noise removal based on the updated MPST network model.

9. A computing device, characterized in that include: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the double-shot aliasing noise removal method based on the MPST network model according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed by a processor, implement the steps of the double-shot aliasing noise removal method based on the MPST network model according to any one of claims 1 to 7.

Images