A method for mixed data shot-hydrophone domain multi-stage joint iterative separation

Abstract

The application discloses a kind of offshore towed cable double sides simultaneous excitation aliasing data shot domain-receiver point domain multi-level joint iterative separation method, belong to oil and gas reservoir seismic exploration efficient acquisition and processing technical field, including the following steps, S1. obtaining the aliasing data to be separated, pre-processing is carried out to seismic data;S2. obtaining shot domain-receiver point domain neural network seismic training data set {B, P};S3. constructing shot domain-receiver point domain neural network model;S4. using training data set, multiple iteration training is carried out to the shot domain-receiver point domain neural network model that is pre-built;S5. using the trained shot domain-receiver point domain neural network model to the aliasing data to be separated carries out intelligent separation, obtains the separated seismic data.The application trains multiple sets of multi-data domain network models using iterative strategy, captures the mapping relationship between different aliasing degrees, different data domain characteristics and label data from a set of training data, thereby significantly improving the network model training process, while ensuring that the algorithm is efficiently implemented, ensuring the quality of the processed data.

Description

Technical Field

[0001] This invention relates to the field of efficient acquisition technology for seismic exploration of oil and gas reservoirs, and in particular to the field of a multi-level joint iterative separation method for shot-detector point domain of simultaneously excited aliased data from both sides of a towed cable at sea. Background Technology

[0002] In aliased data simultaneously excited by both sides of a towed cable at sea, by using a time-delay coding strategy during acquisition, aliasing noise becomes incoherent noise in the common receiver gather, while the effective signal remains coherent. However, the presence of locally continuous coherent aliasing noise can severely affect the separation results of the aliased data. To suppress aliasing noise in the data, a multi-level joint iterative separation algorithm for gun-receiver domain aliased data simultaneously excited by both sides of a towed cable at sea was developed. Denoising-based algorithms mainly treat aliasing noise as incoherent random noise and directly suppress the noise in the aliased record. Inversion-based algorithms, on the other hand, use time-delay coding and sparse transformation to iteratively estimate aliasing noise. Compared to direct denoising, the inversion-based algorithm has better separation performance, but its computational cost is significantly higher. Previous separation methods are often limited to single data domain separation, making it difficult to address the problem that denoising operators cannot completely suppress aliasing noise when locally continuous coherent areas exist. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. To achieve the above-mentioned objective, in a first aspect, this invention proposes a multi-level joint iterative separation method for shot-receiver point domain of aliased data, which includes the following steps:

[0004] S1. Obtain the aliased data to be separated and preprocess the seismic data;

[0005] S2. Obtain the gun-detector point domain neural network seismic training dataset {B, P};

[0006] S3. Constructing a gun-detector neural network model: The gun-detector neural network model includes a multi-level network model with two data domains.

[0007] S4. Using the training dataset, perform multiple iterative training on the pre-built gun-detector point-domain neural network model;

[0008] S5. The trained shot-receiver point-domain neural network model is used to intelligently separate the aliased data to be separated, and the separated seismic data is obtained.

[0009] Secondly, this invention proposes a multi-level joint iterative separation device for gun-receiver point domain of aliased data, the device comprising:

[0010] The data acquisition module is used to acquire the aliased data to be separated and to preprocess the seismic data;

[0011] The training dataset module is used to obtain the shot-receiver point domain neural network seismic training dataset {B, P}.

[0012] The model building module is used to build a gun-receiver point domain neural network model: the gun-receiver point domain neural network model includes a multi-level network model of two data domains.

[0013] The model training module is used to perform multiple iterations of training on a pre-built gun-detector point-domain neural network model using the training dataset.

[0014] The data separation module is used to intelligently separate the aliased data to be separated using a trained shot-receiver point-domain neural network model, so as to obtain the separated seismic data.

[0015] Thirdly, the present invention proposes an electronic device comprising: a processor and a memory; the memory storing computer execution instructions; the processor executing the computer execution instructions stored in the memory, such that the at least one processor executes the gun-detector point domain multi-level joint iterative separation method for the aliased data as described in the first aspect above and various possible aspects of the first aspect.

[0016] Fourthly, the present invention proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the gun-detector point domain multi-level joint iterative separation method for the aliased data as described in the first aspect and various possible aspects of the first aspect.

[0017] The beneficial effects of this invention are: by using an iterative strategy to train multiple sets of multi-data-domain network models, the mapping relationship between different degrees of aliasing, different data-domain features and label data can be captured from a set of training data, thereby significantly improving the network model training process and ensuring the quality of the processed data while ensuring the efficient implementation of the algorithm. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the principle of a multi-level joint iterative separation method provided by the present invention;

[0019] Figure 2 This is a schematic diagram of a neural network model structure for the gun domain-detector domain provided by the present invention;

[0020] Figure 3 A flowchart of a multi-level joint iterative separation method for the gun domain and receiver point domain provided by the present invention;

[0021] Figure 4A comparison chart of the training effects of two different separation methods provided in an embodiment of the present invention;

[0022] Figure 5 This is a conventional single-data-domain neural network iterative prediction result provided in an embodiment of the present invention;

[0023] Figure 6 The results of multi-level joint iterative separation and prediction of the gun domain and receiver domain provided in the embodiments of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] like Figure 1 As shown, this invention proposes a multi-level joint iterative separation method for shot domain-receiver point domain of aliased data. The method includes the following steps: S1. Obtain the aliased data to be separated and preprocess the seismic data;

[0026] Specifically, for aliased seismic data, it was noted that the aliased data on the shot gather record is characterized by left-side data aliasing right-side data, and the inclination angle of the in-phase axis differs, which can be learned by neural networks. This is more suitable for the multi-level joint iterative data separation method of shot domain-sensor domain. Therefore, the simulation is carried out by simultaneous bilateral excitation of the marine towed cable, which is aimed at the aliased data of the marine towed cable bilateral excitation.

[0027] S2. Obtain the gun-detector point domain neural network seismic training dataset {B, P};

[0028] The artillery training set {B} is obtained based on the towed cable data collection dataset. CSG P CSG Data collected using offshore towlines can be used to simulate aliased data, and data excited from one side can be flipped to simulate data excited from the other side of the towline.

[0029] Specifically, step S2 includes: constructing gun-domain training data, using the right-side data as the primary source and the left-side data as the secondary source, and the input for neural network training is as follows:

[0030]

[0031] in This represents the aliasing record in the forward modeling of the gun domain, where τ represents the attenuation factor used to attenuate the intensity of the secondary source data in the gun domain training set. The gun-domain excitation delay operator is represented by Γ.i P(x,t)=P(x,t+Δt i,j ), Δt i,j This represents the delay time of the j-th trace from the i-th source. For the shot domain data, the delay time of each trace is the same. This represents the single-source record of the j-th earthquake source.

[0032] Similarly, obtain the detection point domain training set {B} CRG P CRG Using the same dataset as the gun domain, construct the receiver point domain training data, modify the delay time encoding to create receiver point domain data, and the input for neural network training is as follows:

[0033]

[0034] in This represents the forward-modeled receiver point domain record. This represents the point-domain excitation delay operator, whose mode of action is Γ. i P(x,t)=P(x,t+Δt i,j ), Δt i,j This represents the delay time of the j-th trace from the i-th source. The delay time of each trace varies depending on the excitation. This represents the single-source record of the j-th earthquake source.

[0035] S3. Constructing a gun-detector neural network model: The gun-detector neural network model includes a multi-level network model with two data domains.

[0036] Specifically, step S3 includes: the network consists of a multi-level U-shaped encoder-decoder, the encoder and decoder of the network are based on an attention mechanism, including an input module, a downsampling stage, an upsampling stage, and an output module; the input module is used to send the training sample data input to the network to the downsampling stage; the downsampling stage is used to extract features from the input training aliased data; the upsampling stage is used to integrate and concatenate the features extracted by the downsampling stage, and output the final result through the output module.

[0037] like Figure 2 The diagram shown is a specific network model structure diagram of the present invention. The network consists of a multi-level U-shaped encoder-decoder. The first two levels of upsampling and downsampling are composed of StVit modules, which increase the speed of feature extraction while calculating the attention mechanism. The forward and inverse discrete Haar wavelet transforms are used for downsampling and upsampling, and the deepest layer is composed of a Swin-Transformer structure, which extracts features of three structural regions: horizontal, local, and vertical, according to different window division methods.

[0038] The first two stages of upsampling and downsampling consist of ST-ViT modules. The attention mechanism in the ST-ViT structure is calculated as follows:

[0039]

[0040] In the formula, Z represents the token of size obtained by dividing the token. express transpose, 'i' represents the scaling factor, 'softmax' is the normalized exponential function, and 'i' represents the number of iterations in the network. Redundancy is reduced by iteratively updating the token structure, increasing learnable parameters while decreasing computation time. Simultaneously, the network uses forward and inverse Haar wavelet transforms for downsampling and upsampling, respectively, as shown in the following formula:

[0041]

[0042] Here, D represents Discrete Haar wavelet transform, and x represents the seismic data input to the network. 1,A x 1,H x 1,V x 1,D Corresponding to the approximate matrix, vertical details, horizontal details, and diagonal details, the specific calculation methods are as follows:

[0043]

[0044]

[0045] The Discrete Hal wavelet transform decomposes the feature maps in the network from one channel to four channels, reducing the image size to half its original dimension. This transformation replaces the downsampling operation of the convolutional layer by effectively compressing the data dimension while preserving and extracting key features. Similarly, the inverse wavelet transform is the reverse operation of the above steps and is used for upsampling of the network.

[0046] The deepest layer consists of a DT (Digital Transmission) structure, which extracts features from three structural regions—horizontal, local, and vertical—based on different window partitioning methods. The DT structure computes the self-attention mechanism by dividing the window regions into three different types:

[0047] z:[z h ,z v ,z l ]∈R N×D ,

[0048] z h ∈R 1×W×C ,

[0049] zv ∈R H×1×C ,

[0050] z l ∈R L×L×C

[0051] Here, H, W, and C represent the resolution (height and width) and the number of channels for the feature map input to the attention layer, L represents the resolution of the local window, which is a hyperparameter, and Z represents the size of the token divided into N×D segments;

[0052] In the DT structure, the attention mechanism is calculated as follows:

[0053] [Q,K,V]=Linear(z)

[0054]

[0055] Atten(Q,K,V)=AV

[0056] In the formula, K T This represents the transpose of K. is the scaling factor, softmax represents the normalized exponential function, and Q, K, and V represent the Query, Key, and Value values ​​in the attention mechanism. This structure serves as the deepest encoder-decoder layer to extract data features and reconstruct the output image.

[0057] S4. Using the training dataset, perform multiple iterative training on the pre-built gun-detector point-domain neural network model;

[0058] Specifically, step S4 includes: dividing the earthquake dataset {B, P} into 96*96 pixel blocks, which becomes {Y, X} as the input X and the expected output label Y of the shot-receiver point domain neural network. The closer the predicted result is to the true value Y, the smaller the loss. The neural network updates the network parameters θ by minimizing the loss function. By minimizing the loss function, the prediction result closest to the true value is achieved.

[0059] The loss functions for the gun domain and the receiver point domain are as follows:

[0060]

[0061] Θ represents the network parameters. This represents the predicted values ​​of the input training dataset X using network parameters Θ. AGC stands for gain control. Gain control is applied to the gun domain records during loss calculation to amplify the error of the residual weak signal and increase the features in the training set. This represents the square of the L2 norm.

[0062] Specifically, using the training dataset, the pre-built neural network model is trained through multiple iterations: a gun-receiver point domain neural network model is constructed, and the initial values ​​of the model parameters are determined according to L... crg L csg The parameters of the first-level model are obtained through separate training. and Using the trained network model parameters and iterative formulas, the aliased data is processed to obtain an updated dataset:

[0063]

[0064] Where Γ i -1 The time delay inverse operator of source i operates in the manner of Γ. i P(x,t)=P(x,t-Δt i,j ), Δt i,j This represents the delay time of the j-th trace from the i-th source. In the first iteration, i = 1. In the original training data, this corresponds to... In the next iteration, the input data becomes according to the iteration formula. The network expects the output (label) to remain clean data P, according to the input X. 1 The incoherent aliasing noise in the desired output label Y is estimated by the network and subtracted from the mixed record; after several iterations, It gradually converges to a clean seismic record M, which is the closest to P;

[0065] Where Γ n Let represent the aliasing factor of the nth earthquake source; applying iterative updates to network training yields the final loss function as follows:

[0066]

[0067] right Using the same iterative training method, θ i-1 Let θ represent the network parameters obtained during the (i-1)th training iteration. i This represents the network parameters obtained during the i-th training iteration. The value of i during iterative training is fixed depending on the situation.

[0068] S5. The trained shot-receiver point-domain neural network model is used to intelligently separate the aliased data to be separated, and the separated seismic data is obtained.

[0069] Specifically, step S5 includes: training a total of 2*i sets of network models; for the aliased data D to be separated, first extracting the gun domain records and separating them according to the iterative formula:

[0070]

[0071] Where M i This represents the separation result of the j-th iteration. When j=1, After iteration saturation, the iterative data domain is transformed to the detector point domain, and separation is performed according to the iterative formula:

[0072]

[0073] At this point, the first iteration of the gun-receiver domain has been completed, and the network parameters have been updated to θ. 2 Repeat the above steps to complete the iterative separation of the i-th level network.

[0074] Specifically, in the embodiments of this application, a lightweight network example is used for illustration. For fair comparison, the number of iterations is set to 3 during the iterative training process, and each iteration is set to train for 200 epochs, that is, a total of three sets of network parameters are to be trained.

[0075] This embodiment uses synthetic seismic records and a network model to observe the noise separation effect of an iteratively trained neural network. Furthermore, for a fair comparison, we substitute the parameters of a single-level neural network into the iterative formula. Conduct a comparative evaluation experiment.

[0076] Figure 4 The graph compares the separation effects of two different training strategies: traditional single-data-domain neural network iteration and the multi-level joint iterative separation of the gun-detector domain in this invention. Both strategies iterate 15 times (saturated). The green line represents the separation effect of single-level training. It can be seen that when the signal-to-noise ratio increases to a certain extent, the network can no longer identify the residual noise in the data, and the signal-to-noise ratio reaches stability during iteration. The red line represents the separation method in this paper, which separates the data by iterative encoding of the common gun point CSG gather and the common detector point CRG gather. It can be seen that the iterative signal-to-noise ratio has three segments, which effectively solves the problem that a single network cannot identify residual noise.

[0077] Meanwhile, to visually demonstrate the effectiveness of this method, the detection results of the algorithm were visualized and analyzed, such as... Figure 5 The image shows the iterative prediction results of a traditional single-data-domain neural network. Figure 5 In the diagram (a), the result of 15 iterations of a traditional single-level neural network is shown, with a final signal-to-noise ratio of 22.12 dB. Figure 5 (b) in the figure represents the residual corresponding to the processing result.

[0078] Figure 6To utilize the multi-level joint iterative separation and prediction results of the gun domain and receiver point domain in this invention, Figure 6 (a) shows the result of joint iterative processing of the gun domain and receiver point domain using the iterative training framework, with a signal-to-noise ratio of 27.96 dB. Figure 5 In the diagram, (b) represents the corresponding residual. From... Figure 5 and Figure 6 As can be seen, deep convolutional neural networks based on iterative training frameworks have better dealiasing performance.

[0079] In summary, this invention employs a multi-data-domain joint iterative method that can be integrated into the separation of shot domain records. Currently, high-efficiency multi-source acquisition technology is a hot topic in seismic data acquisition. High-quality separation and high-precision de-aliasing of aliased data can, to some extent, promote the implementation of aliased acquisition and reduce acquisition costs. It primarily addresses the problem of local continuous coherence of aliasing noise in a single data domain during the separation of aliased data from bilateral excitation at sea.

[0080] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A multi-level joint iterative separation method for shot-receiver point domain of aliased data, characterized in that: Includes the following steps: S1. Obtain the aliased data to be separated, preprocess the seismic data, and note that the aliased data on the shot gather record is characterized by left-side data aliasing right-side data, and there are differences in the dip angle of the phase axis, which can be used for neural network learning. S2. Obtain the seismic training dataset for the shot-detector point domain neural network; S3. Constructing a gun-receiver point-domain neural network model, specifically including: the gun-receiver point-domain neural network model consists of a multi-level U-shaped encoder-decoder, the encoder and decoder of the network are based on an attention mechanism, including an input module, a downsampling stage, an upsampling stage, and an output module; the input module is used to send the training sample data input to the network to the downsampling stage; the downsampling stage is used to extract features from the input training aliased data; the upsampling stage is used to integrate and concatenate the features extracted by the downsampling stage, and output the final result through the output module; S4. Using the training dataset, perform multiple iterations to train the pre-built gun-detector point-domain neural network model; the first two stages of upsampling and downsampling consist of ST-ViT modules, and the calculation method of the attention mechanism in this structure is as follows: In the formula, Z represents the segmented word units. express transpose, Softmax is a scaling factor, and it is represented by a normalized exponential function. This represents the number of iterations calculated in the network. The network uses forward and inverse Haar wavelet transforms for downsampling and upsampling, respectively, and the formula is as follows: For discrete Haar wavelet transform, To input earthquake data into the network, The calculation methods for the approximate matrix, vertical details, horizontal details, and diagonal details are as follows: The deepest layer consists of a DT structure, which acts as the deepest encoder-decoder to extract data features and reconstruct the output image. Depending on the window division method, it extracts three structural regions: horizontal, local, and vertical. S5. The trained shot-receiver point-domain neural network model is used to intelligently separate the aliased data to be separated, and the separated seismic data is obtained.

2. The multi-level joint iterative separation method according to claim 1, characterized in that: Step S2 includes step S2.1, constructing the gun domain training data, using the right-side data as the primary source and the left-side data as the secondary source. The input for neural network training is as follows: in This represents the aliasing record of the gun domain in the forward modeling, i=1,2, This represents the attenuation factor, used to attenuate the intensity of the secondary source data in the gun domain training set. This represents the delay time operator for the artillery domain. For artillery domain data, the delay time is the same for each trajectory. This indicates a single-source CSG record.

3. The multi-level joint iterative separation method according to claim 2, characterized in that: Step S2 includes step S2.2, which uses the same dataset as the gun domain to construct receiver point domain training data, modifies the delay time encoding to create receiver point domain data, and the input for neural network training is as follows: in This represents the forward-modeled receiver point domain record. This represents the detector point domain delay time operator. For detector point domain data, the delay time for each trace varies depending on each excitation. This indicates a single-source CRG record.

4. The multi-level joint iterative separation method according to claim 3, characterized in that: Step S4 specifically includes dividing the pre-configured shot-receiver neural network seismic dataset into 96*96 pixel blocks, which serve as the input X and desired output label Y of the shot-receiver neural network. The loss functions for the receiver domain and shot domain are as follows: Represents network parameters, Indicates the use of network parameters The predicted value for the input training dataset X. This indicates gain control; the gun domain recorder uses gain control to amplify the error of the residual weak signal when calculating the loss. express The square of the norm; the closer the predicted result is to the true value Y, the smaller the loss.

5. A multi-level joint iterative separation device for shot-receiver point domain of aliased data, characterized in that, The device is used to perform the multi-level joint iterative separation method for gun domain-detector point domain of aliased data as described in claim 1.

6. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing the method as described in any one of claims 1 to 4.

Images