Shot domain-detection point domain multi-stage joint iterative separation method for aliasing data

Abstract

The invention discloses a shot domain-detection point domain multi-stage joint iterative separation method for simultaneous excitation of aliasing data on both sides of an offshore towing cable, and belongs to the technical field of efficient acquisition and processing of seismic exploration of oil and gas reservoirs, and the method comprises the following steps: S1, obtaining aliasing data to be separated, and preprocessing the seismic data; s2, acquiring a shot domain-detection point domain neural network seismic training data set {B, P}; s3, constructing a shot domain-detection point domain neural network model; s4, performing multi-iteration training on a pre-established shot domain-detection point domain neural network model by using the training data set; and S5, performing intelligent separation on to-be-separated aliasing data by using the trained shot domain-detection point domain neural network model to obtain separated seismic data. According to the method, multiple sets of multi-data-domain network models are trained by utilizing an iteration strategy, and mapping relationships between different aliasing degrees, different data domain features and label data are captured from one set of training data, so that the network model training process is remarkably improved, and the processed data quality is ensured while the efficient implementation of the algorithm is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of efficient acquisition in seismic exploration of oil and gas reservoirs, and particularly to the technical field of a multi-level joint iterative separation method in the shot domain - geophone domain for aliased data simultaneously excited on both sides of a marine streamer. Background Art

[0002] In the aliased data simultaneously excited on both sides of a marine streamer, by using a time-delay coding strategy during the acquisition process, the aliased noise becomes incoherent noise in the common receiver gather, while the effective signal is coherent. However, there are parts where the local aliased noise is continuously coherent, which will seriously affect the separation result of the aliased data. In order to suppress the aliased noise in the data, a multi-level joint iterative separation algorithm in the shot domain - geophone domain for the aliased data simultaneously excited on both sides of a marine streamer has been developed. The algorithms based on denoising mainly regard the aliased noise as incoherent random noise and directly suppress the noise in the aliased record. The algorithms based on inversion use time-delay coding and sparse transformation to iteratively estimate the aliased noise. Compared with direct denoising, the algorithms based on the inversion idea have better separation effects, but the computational amount is significantly larger than that of the denoising algorithms. The previous separation processing is often limited to the separation in a single data domain and is difficult to solve the problem that the denoising operator is difficult to completely suppress when there are local continuously coherent parts in the aliased noise. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art. To achieve the above invention purpose, in the first aspect, the present invention proposes a multi-level joint iterative separation method in the shot domain - geophone domain for aliased data, and the method includes the following steps:

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

[0005] S2. Obtain the neural network seismic training data set {B, P} in the shot domain - geophone domain;

[0006] S3. Construct a neural network model in the shot domain - geophone domain: The neural network model in the shot domain - geophone domain includes multi-level network models in two data domains.

[0007] S4. Use the training data set to perform multiple iterative trainings on the pre-built neural network model in the shot domain - geophone domain;

[0008] S5. Use the trained neural network model in the shot domain - geophone domain to intelligently separate the aliased data to be separated and obtain the separated seismic data.

[0009] In the second aspect, the present invention proposes a multi-level joint iterative separation device in the shot domain - geophone domain for aliased data, and the device includes:

[0010] A data acquisition module, configured to acquire aliased data to be separated and preprocess seismic data;

[0011] A training dataset module, configured to acquire a seismic training dataset {B, P} for a shot-receiver domain neural network;

[0012] A model construction module, configured to construct a shot-receiver domain neural network model: The shot-receiver domain neural network model includes a multi-level network model for two data domains.

[0013] A model training module, configured to perform multiple iterative trainings on a pre-built shot-receiver domain neural network model by using the training dataset;

[0014] A data separation module, configured to intelligently separate the aliased data to be separated by using the trained shot-receiver domain neural network model to obtain the separated seismic data.

[0015] In a third aspect, the present invention provides an electronic device, including: a processor and a memory; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the shot-receiver 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.

[0016] In a fourth aspect, the present invention provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the processor executes the computer-executable instructions, the shot-receiver 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 is implemented.

[0017] The beneficial effects of the present invention are: By using an iterative strategy to train multiple sets of multi-data domain network models, the mapping relationship between different aliasing degrees, different data domain features and label data is captured from a set of training data, thereby significantly improving the network model training process, ensuring the data quality after processing while ensuring the efficient implementation of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0019] Figure 2 It is a schematic diagram of the structure of a shot-receiver domain neural network model provided by the present invention;

[0020] Figure 3 It is a flowchart of a shot-receiver domain multi-level joint iterative separation method provided by the present invention;

[0021] Figure 4Comparison chart of training effects of two different separation methods provided by the embodiments of the present invention;

[0022] Figure 5 Iterative prediction result of the traditional single-data-domain neural network provided by the embodiments of the present invention;

[0023] Figure 6 Multi-level joint iterative separation prediction result of shot domain - geophone domain provided by the embodiments of the present invention. Detailed implementation manners

[0024] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

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

[0026] Specifically, for the aliased seismic data, it is noted that the form of the aliased data on the shot gather record is left unilateral data aliasing right unilateral data, and the dip of the event has differences for the learning of the neural network, which is more suitable for the multi-level joint iterative data separation method of shot domain - geophone domain. Therefore, simultaneous bilateral excitation of the marine streamer is used for simulation, aiming at the aliased data of bilateral excitation of the marine streamer.

[0027] S2. Obtain the shot domain - geophone domain neural network seismic training data set {B, P};

[0028] According to the streamer acquisition data set, obtain the shot domain training set {B CSG , P CSG}. The data collected by the marine streamer can be used to simulate the aliased data, and the data of unilateral excitation is flipped as the data of the other side of the simulated streamer line excitation.

[0029] Specifically, step S2 includes: constructing the shot domain training data, using the right unilateral data as the main seismic source and the left unilateral data as the secondary seismic source, and the input of the neural network training is as follows:

[0030]

[0031] Where represents the forward shot domain aliased record, τ represents the attenuation factor, which is used to attenuate the intensity of the secondary seismic source data in the shot domain training set, represents the shot domain excitation delay operator, and its action mode is Γi P(x, t) = P(x, t + Δt i,j ), Δt i,j represents the delay time of the j-th trace of the i-th source. For shot gather data, the delay time of each trace is the same. represents the single-source record of the j-th source.

[0032] Similarly, obtain the geophone gather training set {B CRG , P CRG}: Use the same data set as the shot gather, construct the geophone gather training data, change the delay time encoding to create the geophone gather data, and the input for neural network training is as follows:

[0033]

[0034] where represents the forward modeling geophone gather record, represents the geophone gather excitation delay operator, and its action mode is Γ i P(x, t) = P(x, t + Δt i,j ), Δt i,j represents the delay time of the j-th trace of the i-th source. The delay time of each trace is different for each excitation. represents the single-source record of the j-th source.

[0035] S3. Construct a shot gather - geophone gather neural network model: The shot gather - geophone gather neural network model includes a multi-level network model for two data domains.

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

[0037] As Figure 2 shown, it is the 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 up and downsampling are composed of StVit modules, which increase the speed of extracting features while calculating the attention mechanism, and the forward and inverse discrete Haar wavelet transforms are used as down and upsampling. The deepest layer is composed of a Swin-Transformer structure, which extracts the features of three structural regions of horizontal - local - vertical according to different window partitioning methods.

[0038] The first two levels of upsampling and downsampling are composed of ST-ViT modules. The calculation method of the attention mechanism in the ST-ViT structure is as follows:

[0039]

[0040] In the formula, Z represents the token with a size obtained by partitioning, represents the transpose of, is the scaling factor, softmax represents the normalized exponential function, i represents the number of calculation iterations in the network. By iteratively updating the Token structure, its redundancy is reduced, the number of learnable parameters is increased while the calculation time is reduced. At the same time, the network uses the forward and inverse Haar wavelet transforms as downsampling and upsampling, and its formula is as follows:

[0041]

[0042] Here, D represents the discrete Haar wavelet transform, x represents the seismic data input into the network, x 1,A , x 1,H , x 1,V , x 1,D , corresponding to the approximation matrix, vertical detail, horizontal detail, and diagonal detail. The specific calculation method is as follows:

[0043]

[0044]

[0045] Under the effect of the discrete Haar wavelet transform, the feature map in the network is decomposed from one channel into four channels, and the image size is reduced to half of its original dimension This transform replaces the downsampling operation of the convolutional layer by effectively compressing the data dimension while retaining 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 is composed of the DT structure, which extracts the features of the horizontal-local-vertical three structural regions respectively according to different window partitioning methods. The DT structure calculates the self-attention mechanism by partitioning three different window regions. The partitioning of the window regions is of three 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 of the feature map input to the attention layer, L represents the resolution of the local window, which is a hyperparameter, and Z represents the tokens of size N×D divided out;

[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 represents the transpose of K, is the scaling factor, softmax represents the normalized exponential function, Q, K, and V represent the three values of Query, Key, and Value in the attention mechanism, and this structure is used as the deepest encoder-decoder to extract data features and reconstruct the output image.

[0057] S4. Use the training dataset to perform multiple iterative training on the pre-built shot domain - geophone domain neural network model;

[0058] Specifically, step S4 includes: dividing the seismic dataset {B, P} into blocks of 96*96 pixels, which becomes {Y, X} as the input X and the expected output label Y of the shot domain - geophone domain neural network. The closer the prediction result is to the true value Y, the smaller the loss. The neural network gradually updates the network parameters θ by minimizing the loss function, and by minimizing the loss function, it achieves a prediction result closest to the true value;

[0059] The loss functions for the shot domain and the geophone domain are respectively:

[0060]

[0061] Θ represents the network parameters, represents the predicted value of the input training dataset X using the network parameters Θ, AGC represents the gain control, and the shot domain record performs gain control when calculating the loss to amplify the error of the remaining weak signals and increase the features in the training set, represents the square of the L2 norm.

[0062] Specifically, using a training data set, a pre-built neural network model is trained through multiple iterations: a shot domain - geophone domain neural network model is constructed, and the initial values of the model parameters are determined. According to L crg and L csg the model parameters of the first level are respectively trained to obtain and The trained network model parameters are used to process the aliased data using an iterative formula to obtain an updated data set:

[0063]

[0064] where Γ i -1 is the inverse operator of the time delay of the i-th seismic source, and its action mode is Γ i P(x,t) = P(x,t - Δt i,j ), Δt i,j represents the delay time of the j-th trace of the i-th seismic source. At the first iteration, i = 1, which corresponds to in the original training data. At the next iteration, the input data becomes according to the iterative formula. The network expected output (label) is still the clean data P. According to the input X 1 and the expected output label Y, the incoherent aliased noise therein is estimated by the network and subtracted from the mixed record. After multiple iterations, it gradually converges to the clean seismic record M, that is, when it is closest to P;

[0065] where Γ n represents the aliasing factor of the n-th seismic source; applying the iterative update to the network training, the following final loss function is obtained:

[0066]

[0067] For the same iterative training method is used. θ i-1 represents the network parameters obtained from the (i - 1)-th training, and θ i represents the network parameters obtained from the i-th training of the network. The value of i for iterative training is determined according to the situation.

[0068] S5. Use the trained shot domain - geophone domain neural network model to intelligently separate the aliased data to be separated, and obtain the separated seismic data.

[0069] Specifically, step S5 includes: training a total of 2 * i sets of network models. For the aliased data D to be separated, first extract the shot domain record and separate it according to the iterative formula:

[0070]

[0071] where M i represents the separation result of the j-th iteration. When j = 1, After the iteration saturates, the iterative data domain is transformed to the receiver domain, and separation is performed according to the iterative formula:

[0072]

[0073] At this time, the first-level iteration of the shot domain - receiver domain is completed, and the network parameters are updated to θ 2 , and the above steps are repeated to complete the iterative separation of the i-th level network.

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

[0075] This embodiment uses synthetic seismic records and a network model to observe the noise separation effect of a neural network based on iterative training. And for fair comparison, we substitute the single-level neural network parameters into the iterative formula to conduct a comparative evaluation experiment.

[0076] Figure 4 is a comparison chart of the separation effects of two different training strategies: traditional single-data-domain neural network iteration and the multi-level joint iterative separation of the shot domain - receiver domain in the present invention. Both sets of strategies are iterated 15 times (saturated). Among them, the green line is the separation effect of the single-level training result. 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 iteration reaches stability. While the red line is the separation method in this paper, which is iteratively encoded and separated according to the common-shot CSG gather - common-receiver CRG gather. It can be seen that the iterative signal-to-noise ratio shows three segments, effectively solving the problem that it is difficult for a single network to identify residual noise.

[0077] At the same time, in order to intuitively demonstrate the effectiveness of this method, a visual analysis of the detection results of the algorithm is carried out. As Figure 5 shown is the iterative prediction result of the traditional single-data-domain neural network. Figure 5 In (a) of Figure 5 is the processing result of the traditional single-level neural network after 15 iterations, and the finally recovered signal-to-noise ratio is 22.12 dB.

[0078] Figure 6To adopt the prediction result of multi-level joint iterative separation in the shot domain - geophone domain of the present invention, Figure 6 where (a) in Figure 6 is the joint iterative processing result of the shot domain - geophone domain using the iterative training framework, and its signal-to-noise ratio is 27.96 dB. Figure 5 where (b) in Figure 5 is the corresponding residual. It can be seen from Figure 5 and Figure 6 that the deep convolutional neural network based on the iterative training framework has a better anti-aliasing effect.

[0079] In summary, the present invention adopts a multi-data domain joint iterative method that can incorporate the separation of shot domain records. At present, the multi-source high-efficiency acquisition technology is a hot spot in seismic data acquisition. The high-quality separation and high-precision de-mixing of aliased data can, to a certain extent, promote the implementation of aliased acquisition and reduce the acquisition cost, mainly solving the problem of local continuous coherence of aliasing noise in a single data domain faced by the separation of aliased data in offshore bilateral excitation.

[0080] The above description shows and describes a preferred embodiment of the present invention. However, as mentioned above, it should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications, and environments, and can be changed within the scope of the inventive concept described herein through the above teachings or the techniques or knowledge in related fields. And any changes and modifications made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A multi - level joint iterative separation method for aliased data in shot - receiver domain, characterized in that: It includes the following steps: S1. Obtain the aliased data to be separated and preprocess the seismic data; S2. Obtain the seismic training dataset {B, P} of the shot domain - geophone domain neural network; S3. Construct the neural network model of the shot domain - geophone domain; S4. Use the training dataset to perform multiple iterative trainings on the pre - built neural network model of the shot domain - geophone domain; S5. Use the trained neural network model of the shot domain - geophone domain to intelligently separate the aliased data to be separated and obtain the separated seismic data.

2. The multi-stage combined iterative separation method according to claim 1, wherein: Step S2.1 specifically includes constructing the shot domain training data, using the right - hand single - side data as the main seismic source and the left - hand single - side data as the secondary seismic source. The inputs for neural network training are as follows: Among them represents the forward modeling shot - gather aliased record, τ represents the attenuation factor, which is used to attenuate the intensity of the secondary source data in the shot - gather training set, represents the shot - gather excitation delay operator, and its action mode is Γ i P(χ,t) = P(x,t + Δt i,j ), Δt i,j represents the delay time of the j - th trace of the i - th source. For shot - gather data, the delay time of each trace is the same, represents the single - source record of the j - th source.

3. The multi-stage combined iterative separation method according to claim 1, characterized in that: Step S2.2 specifically includes using the same dataset as the shot domain to construct the geophone domain training data, changing the delay time encoding to create the geophone domain data. The inputs for neural network training are as follows: Among them represents the recorded data of the geophone domain in forward modeling represents the excitation delay operator in the geophone domain, and its action mode is Γ i P(x,t) = P(χ,t + Δt i,j ), where Δt i,j represents the delay time of the j-th trace of the i-th seismic source. The delay time of each trace varies according to each excitation represents the single-source record of the j-th seismic source 4. The multi-stage combined iterative separation method according to claim 1, characterized in that: Step S3 specifically includes that the neural network model of the shot domain - geophone domain is composed of multiple - level U - shaped encoder - decoders. The encoder - decoder of the network is based on the attention mechanism and includes an input module, a down - sampling section, an up - sampling section, and an output module; the input module is used to send the training sample data input to the network to the down - sampling section; the down - sampling section is used to extract features from the input training aliased data; the up - sampling section is used to integrate and splice the extracted features from the down - sampling section and output the final result through the output module.

5. The multi-stage combined iterative separation method according to claim 4, wherein: The first two levels of up - and down - sampling are composed of ST - ViT modules. The calculation method of the attention mechanism in this structure is as follows: where Z represents the token of size obtained by partitioning, denotes the transpose of, is the scaling factor, softmax represents the normalized exponential function, i represents the number of calculation iterations in the network, and the network uses the forward and inverse Haar wavelet transforms for downsampling and upsampling, and its formula is expressed as follows: D is the discrete Haar wavelet transform, x is the seismic data input to the network, X 1,A , χ 1,H , X 1,V , X 1,D corresponding to the approximation matrix, vertical detail, horizontal detail, and diagonal detail, and the calculation method is as follows: The deepest layer is composed of a DT structure, which is used as the deepest - layer encoder - decoder to extract data features and reconstruct the output image, and extracts three structural regions of horizontal - local - vertical according to different window division methods.

6. The multi-stage combined iterative separation method according to claim 1, wherein: Step S4 specifically includes dividing the set - up seismic dataset {B, P} of the shot domain - geophone domain neural network into blocks of 96 * 96 pixels as the input X and the expected output label Y of the shot domain - geophone domain neural network. The loss functions of the shot domain and the geophone domain are respectively: Θ represents network parameters, represents the predicted value of the input training dataset X using the network parameters Θ. AGC represents gain control. Gain control is performed during the calculation of the loss for the shot domain recording to amplify the error of the residual weak signal, represents the square of the L2 norm; the closer the prediction result is to the true value Y, the smaller the loss.

7. The multi-stage combined iterative separation method according to claim 6, wherein: According to the loss functions \(L\) in the shot domain and the geophone domain crg , \(L\) csg respectively train to obtain the model parameters of the first-level shot domain and geophone domain and Use the trained network model parameters to process the aliased data using the iterative formula to obtain an updated dataset: where Γ i -1 is the time-delay inverse operator of the source i, Δt i,j represents the delay time of the jth trace of the ith source, Γ n represents the aliasing factor of the nth source; The first iteration, i = 1, The corresponding original training data is After the next iteration, the input data is According to the input X 1 and the expected output label Y, subtract the incoherent aliasing noise from the mixed record; After multiple iterations, it gradually converges to the clean seismic record M, that is, when it is closest to P; Apply iterative update to the network training to obtain the final loss function: For using the same iterative training method, θ i-1 represents the network parameters obtained from the (i - 1)-th training, and q i represents the network parameters obtained from the i-th training of the network.

8. The multi-stage combined iterative separation method according to claim 1, wherein: Step S5 specifically includes training a total of 2 * i sets of network models, selecting the iterative data domain for encoding, and according to the sequence of shot domain - geophone domain - shot domain - geophone domain, for the aliased data D to be separated, first extract the shot domain records and separate them according to the iterative formula: M j represents the separation result of the j-th iteration. When j = 1, After iteration saturation, the iterative data domain is transformed to the geophone domain, and separation is performed according to the iterative formula: At this point, the first-level iteration of the shot domain - geophone domain is completed, and the network parameters are updated to θ 2 , and repeat the above steps until the iterative separation of the i-th level network is completed.

9. An alias data shot domain - geophone domain multi - level joint iterative separation device, characterized in that: It includes the following modules: Data acquisition module, used to obtain the aliased data to be separated and preprocess the seismic data; Training dataset module, used to obtain the seismic training dataset {B, P} of the shot domain - geophone domain neural network; Model construction module, used to construct the neural network model of the shot domain - geophone domain: the neural network model of the shot domain - geophone domain includes a multi - level network model of two data domains; Model training module, used to perform multiple iterative trainings on the pre - built neural network model of the shot domain - geophone domain using the training dataset; Data separation module, used to intelligently separate the aliased data to be separated using the trained neural network model of the shot domain - geophone domain and obtain the separated seismic data.

10. A computing device, characterized in that, It includes: 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, and the one or more programs include instructions for performing any of the methods recited in claims 1 to 8.

Images