Mixed-gas separation model training method and related methods and apparatuses

Abstract

The application discloses a mixed-sampling separation model training method and related methods and devices. The method comprises: performing trace set extraction on collected multiple groups of aliasing data, corresponding main gun data and excitation time intervals, taking the obtained seismic data corresponding to the geophone trace set as a sample set; using the aliasing data in the sample set and the corresponding excitation time interval as input data, and the corresponding main gun data as label data, training a mixed-sampling separation model, wherein the mixed-sampling separation model comprises a deep learning network and a geophysical guide network; and when a preset training end condition is met, obtaining a trained mixed-sampling separation model. The method can effectively suppress adjacent gun interference and effectively protect the main gun data, can greatly improve the separation efficiency and precision, improve the signal-to-noise ratio of seismic data, and provide high-quality basic data for subsequent quantitative interpretation of seismic data and high-precision reservoir prediction.

Description

Technical Field

[0001] This invention relates to the field of seismic data processing in oil and gas geophysical exploration, and particularly to a method for training a mixed production separation model and related methods and apparatus. Background Technology

[0002] In oil and gas geophysical exploration, seismic excitation equipment is used to generate earthquakes, which are then used for exploration. Separation of seismic data from the main source earthquake is the process of separating the seismic data generated by the main source from the cascaded data of seismic data from adjacent sources. Here, the main source refers to the selected seismic source, and adjacent sources refer to other seismic sources. Main source data refers to the data from the main source earthquake. Cascaded data is the data from the cascaded seismic data of the main source earthquake and adjacent sources.

[0003] In geological exploration, a selected location is typically used as a shot point to conduct artificial seismic tests. The resulting data is then analyzed to study the geological structure. Before triggering an artificial seismic test at the shot point, geophones are placed at multiple locations on the surface. These geophones sample and record seismic data at fixed time intervals, and the data recorded by each geophone is called a seismic trace. The geophones are arranged in a straight line with the shot point, ensuring that the recorded seismic traces comprehensively reflect the earthquake's propagation process from the shot point. The fixed spacing between the geophones ensures that the seismic traces represent earthquake propagation at fixed intervals, exhibiting strong regularity and facilitating subsequent data analysis. The spacing between the geophones is called the trace spacing. After triggering an artificial seismic test at the shot point, the seismic traces recorded by each geophone are arranged according to their geographical location to form the shot gather record. The horizontal axis of the shot gather record represents the seismic trace number, and the vertical axis represents time.

[0004] Traditional seismic acquisition techniques set sufficiently long excitation times for adjacent shots relative to the main shot, ensuring minimal overlap between the main shot and adjacent shot seismic data, thus eliminating the need for seismic separation. Multi-source seismic acquisition techniques simultaneously or with delays excite multiple sets of sources, shortening acquisition time, significantly improving efficiency, and reducing costs. Seismic separation is a crucial processing technique in multi-source overlapping acquisition. Currently, conventional adjacent shot interference suppression techniques can be broadly categorized into two types: filtering-based suppression methods and inversion-based suppression methods. Filtering-based methods approximate adjacent shot interference on non-shot gathers (such as common receiver gathers) as random noise, using methods like frequency-wavenumber domain dip filtering and space-time domain median filtering to suppress the interference. These methods offer advantages such as high computational speed and ease of implementation. Sparse inversion-based methods treat the mixed acquisition data as observation data, the effective seismic records to be recovered as model variables, and adjacent shot interference suppression as an inverse problem, solving it using a sparse-constrained training inversion method. Compared with filtering methods, effective seismic records separated by inversion methods have higher signal-to-noise ratio and fidelity. Summary of the Invention

[0005] The inventors discovered that existing mixed-data separation methods mainly fall into two categories: filtering and inversion. Filtering-based methods are less effective at separating low signal-to-noise ratio data. Sparse inversion-based methods have low computational efficiency, making them difficult to apply in practice. Conventional adjacent-gun interference suppression techniques are essentially model-driven algorithms, requiring the design of a differentiation algorithm based on the differences between main gun data and adjacent-gun interference. They assume the discontinuity of adjacent-gun interference in the non-gun domain to distinguish between the main gun signal and interference noise. However, actual data is often affected by various noises, leading to a deterioration in the continuity of the effective main gun signal in the non-gun domain. Therefore, relying solely on a single, artificial assumption to suppress adjacent-gun interference cannot achieve ideal results. To at least partially address the technical problems of existing technologies, the inventors have developed this invention, providing a mixed-data separation model training method and related methods and apparatus through specific implementation methods.

[0006] In a first aspect, embodiments of the present invention provide a method for training a mixed sampling separation model, comprising:

[0007] Gatherings were extracted from the multiple sets of aliased data, the corresponding main gun data, and the excitation time interval. The receiver gathers corresponding to the acquired seismic data were used as the sample set.

[0008] Using the aliased data and corresponding excitation time intervals in the sample set as input data, and the corresponding main gun data as label data, a mixed-mining separation model is trained. The mixed-mining separation model includes a deep learning network and a geophysical guidance network.

[0009] When the preset training termination conditions are met, the trained mixed sampling separation model is obtained.

[0010] In some optional embodiments, the step of extracting gathers from the acquired multiple sets of aliased data, the corresponding main gun data, and the excitation time interval, and using the receiver gathers corresponding to the acquired seismic data as a sample set, includes:

[0011] Seismic data was acquired using at least one of the following methods:

[0012] Select the seismic data of two shots that are not overlapped, and superimpose the seismic data of the two shots at random time intervals to obtain the seismic overlapped data of the main shot and the adjacent shot. The random time interval is the excitation time interval corresponding to the overlapped data.

[0013] Numerical simulation techniques are used to simulate conventionally acquired seismic data to obtain main gun data, and to simulate aliased acquired seismic data to obtain aliased seismic data and corresponding excitation time intervals.

[0014] Multiple sets of aliased data were collected, and the main gun data and the corresponding firing time interval were separated using aliasing separation technology.

[0015] The acquired seismic data is transformed into receiver gathers, and these receiver gathers are used as a sample set.

[0016] In some optional embodiments, before training the mixed-data separation model using the aliased data and corresponding firing time intervals from the sample set as input data and the corresponding main gun data as label data, the following steps are included:

[0017] Construct an aliasing matrix based on the excitation time interval of each sample in the sample set;

[0018] Initialize the parameters of the mixed sampling and separation model This includes deep learning network parameters Θ and geophysical guidance network tradeoff parameters δ and η;

[0019] The aliasing matrix is ​​used to determine the initial value x0 of the main gun calculation data for the mixed-sampling separation model;

[0020] The initial training iterations are 0.

[0021] In some optional embodiments, constructing the aliasing matrix based on the excitation time interval of each sample in the sample set includes:

[0022] Convert the excitation time interval of each sample in the sample set into an integer;

[0023] Based on the integerized excitation time interval of each sample in the sample set, determine the delay matrix of each sample corresponding to different excitation time intervals;

[0024] Summing the delay matrix yields the aliasing matrix.

[0025] In some optional embodiments, determining the delay matrix for each sample corresponding to different excitation time intervals based on the integerized excitation time interval of each sample in the sample set includes:

[0026] Set an N t Matrix E m m is the excitation time interval, N t It is a positive integer;

[0027] matrix E m The row ordinal number minus the column ordinal number equals the integerized excitation time interval m of each sample in the sample set. The elements are assigned a value of 1, and the other elements are assigned a value of 0, to obtain the delay matrix corresponding to the sample with an excitation time interval of m.

[0028] Repeat the above process to obtain the delay matrix corresponding to samples with different excitation time intervals.

[0029] In some optional embodiments, summing the delay matrix to obtain the aliasing matrix includes:

[0030] The delay matrices of each sample at the same excitation time interval are summed, and the sums of the delay matrices corresponding to different excitation time intervals are added together to obtain the aliasing matrix.

[0031] In some optional embodiments, determining the initial value x0 of the main gun calculation data for the mixed-sampling separation model using the aliasing matrix includes:

[0032] According to x0=Γ T d, determine the initial value x0 of the main gun calculation data, where d is the aliased data in the sample set, Γ T It is the transpose of the aliased matrix.

[0033] In some optional embodiments, the step of using the aliased data and corresponding firing time intervals from the sample set as input data, and the corresponding main gun data as label data, to train the aliasing separation model includes:

[0034] Input δ and η into the geophysical guidance network get

[0035] The geophysical guidance network is used to weigh the parameters δ, η, and the main gun calculation data x as input parameters to a deep learning network, and based on... We obtain new δ, η, and main gun calculation data x;

[0036] Calculate the loss function We obtain a new Θ, where d i For a sample of aliased data, x i For d i Corresponding main gun data, m i For d i The corresponding activation time interval, where N is the total number of samples in the sample set, and i is the index of the sample in the sample set, Net(d i m i ;Θ) is the mixed sampling separation model based on Θ, d i and m i The corresponding main gun calculation data obtained;

[0037] The mixed sampling separation model is updated with new δ, η, and Θ as new parameters, and the number of training iterations is also updated.

[0038] In some optional embodiments, after updating the number of training iterations, the method further includes:

[0039] If the maximum number of training iterations is not met, the parameters of the current network's hybrid sampling and separation model are adjusted using the backpropagation algorithm.

[0040] Secondly, embodiments of the present invention provide a mixed mining separation method, characterized in that it is implemented using a mixed mining separation model obtained by the aforementioned mixed mining separation model training method, including:

[0041] The seismic aliasing data to be separated is subjected to gather extraction to obtain the receiver gather corresponding to the seismic aliasing data to be separated.

[0042] The data of the receiver gathers corresponding to the seismic aliasing data to be separated are input into the seismic aliasing model, and the main gun data is obtained based on the output separated data.

[0043] In some optional embodiments, obtaining the main gun data based on the separated output data includes:

[0044] The separated output data is transformed back into gun data using the method of gather extraction, and the gun data is the main gun data.

[0045] Thirdly, embodiments of the present invention provide a mixed sampling separation model generation device, characterized in that it includes:

[0046] Sample set generation module: used to transform multiple sets of aliased data and their corresponding main gun data and corresponding excitation time intervals into detector gathers, and use the detector gathers as sample sets; extract a preset number of samples from the sample set to generate a sample set;

[0047] The mixed mining separation model training module is used to input the sample set into the mixed mining separation model composed of a deep learning network and a geophysical guidance network for training until the preset training termination conditions are met, and the trained mixed mining separation model is obtained.

[0048] Fourthly, embodiments of the present invention provide a mixed sampling and separation device, characterized in that it comprises:

[0049] Hybrid data transformation module: used to extract gathers from the seismic aliasing data to be separated, and obtain the receiver gathers corresponding to the seismic aliasing data to be separated;

[0050] Mixed data separation module: used to input the data of the detector point gather into the mixed data separation model obtained by the mixed data separation model training method of any one of claims 1-9, obtain the separated data, and then transform it back into shot gather according to the gather extraction method.

[0051] Based on the same inventive concept, embodiments of the present invention also provide a computer storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned mixed-sampling separation model training method or mixed-sampling separation method.

[0052] Based on the same inventive concept, embodiments of the present invention also provide a terminal device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned mixed mining separation model training method or mixed mining separation method.

[0053] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0054] Embodiment 1 of the present invention provides a method for training a mixed mining separation model. By using seismic data obtained through various methods, a mixed mining separation model including a deep learning network and a geophysical guidance network is trained, resulting in a mixed mining separation model that can effectively suppress interference from adjacent shots and effectively protect the main shot data, while significantly improving computational efficiency.

[0055] The second embodiment of this invention provides a mixed-source separation method, which uses a trained mixed-source separation model including a deep learning network and a geophysical guidance network to perform mixed-source separation. This method can effectively suppress interference from adjacent shots and effectively protect the main shot data, significantly improve separation efficiency and accuracy, increase the signal-to-noise ratio of seismic data, and provide high-quality basic data for subsequent quantitative interpretation of seismic data and high-precision reservoir prediction.

[0056] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0057] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of a mixed sampling separation model training method in an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the deep learning network structure in an embodiment of the present invention;

[0061] Figure 3 This is a diagram showing the actual marine mixed mining record in an embodiment of the present invention;

[0062] Figure 4 This is a main gun recording diagram separated using an inversion method in an embodiment of the present invention;

[0063] Figure 5 This is an example of an adjacent-shot interference map separated using an inversion method in this embodiment of the invention.

[0064] Figure 6 This is a main gun recording diagram separated using the mixed-sampling separation model in an embodiment of the present invention;

[0065] Figure 7 The adjacent shot interference is separated using the mixed-sampling separation model in this embodiment of the invention;

[0066] Figure 8 This is a residual image of seismic data of an oilfield after co-production separation using the co-production separation model and co-production separation using the sparse inversion method, as shown in this embodiment of the invention.

[0067] Figure 9 This is a flowchart of a mixed sampling and separation method according to an embodiment of the present invention;

[0068] Figure 10 This is a simulated mixed sampling data diagram in an embodiment of the present invention;

[0069] Figure 11 This is a diagram of the main gun data without overlap in an embodiment of the present invention;

[0070] Figure 12This is a diagram of the main gun data after the mixed sampling and separation method in this embodiment of the invention;

[0071] Figure 13 This is a residual plot showing the processing result and the actual value in an embodiment of the present invention.

[0072] Figure 14 This is a block diagram of a mixed sampling separation model training device in an embodiment of the present invention;

[0073] Figure 15 This is a block diagram of a mixed sampling and separation device according to an embodiment of the present invention. Detailed Implementation

[0074] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0075] To address the problems existing in the prior art, embodiments of the present invention provide a method for training a mixed sampling separation model, as well as related methods and apparatus.

[0076] Example 1

[0077] Embodiment 1 of the present invention provides a training method for a mixed sampling separation model, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0078] Step S101: Extract gathers from the multiple sets of aliased data, the corresponding main gun data, and the excitation time interval, and use the receiver gathers corresponding to the acquired seismic data as the sample set.

[0079] In some alternative embodiments, seismic data is acquired using at least one of the following methods:

[0080] Seismic data from two shots that are not overlapped are selected and superimposed at random time intervals to obtain overlapping seismic data of the main shot and the adjacent shot. The random time interval is the excitation time interval corresponding to the overlapping data. For example, using conventional seismic acquisition techniques to obtain seismic data from the main shot and the adjacent shot that are not overlapped, the excitation time interval for the adjacent shot relative to the main shot is set to be sufficiently long, i.e., the excitation time interval is long enough, so that the seismic data from the main shot and the adjacent shot hardly overlap, eliminating the need for seismic separation. Then, the seismic data from the main shot and the adjacent shot that are not overlapped are superimposed at random time intervals, the random time interval being the excitation time interval corresponding to the overlapping data. The main shot data has already been acquired using conventional seismic acquisition techniques.

[0081] Numerical simulation technology is used to simulate conventionally acquired seismic data to obtain main gun data, and simulate aliased acquired seismic data to obtain aliased seismic data and corresponding excitation time intervals. Multiple sets of aliased data are then separated using aliasing separation technology to separate the main gun data and corresponding excitation time intervals.

[0082] The acquired seismic data is transformed into receiver gathers, and these receiver gathers are used as a sample set.

[0083] Step S102: Construct an aliasing matrix based on the excitation time interval of each sample in the sample set.

[0084] Convert the excitation time interval of each sample in the sample set into an integer; for example, the excitation time interval of each sample can be rounded up, rounded to the nearest integer, or the decimal part can be directly deleted, etc.

[0085] Based on the integerized excitation time interval of each sample in the sample set, determine the delay matrix of each sample corresponding to different excitation time intervals;

[0086] In some alternative embodiments, an N is set. t Matrix E m m is the excitation time interval, N t It is a positive integer;

[0087] matrix E m The row ordinal number minus the column ordinal number equals the integerized excitation time interval m of each sample in the sample set. The elements are assigned a value of 1, and the other elements are assigned a value of 0. This yields the delay matrix corresponding to the sample with an excitation time interval of m.

[0088] Repeat the above process to obtain the delay matrix corresponding to samples with different excitation time intervals.

[0089] The delay matrix can be calculated as: E m Let the delay matrix be defined as:

[0090]

[0091] Where Nt is the number of columns in the main gun's calculation data x. For example, when m = 0, E0 = I, where I is the identity matrix; when m = 1, we have

[0092]

[0093] Summing the delay matrix yields the aliasing matrix.

[0094] In some optional embodiments, the delay matrices of each sample for the same excitation time interval are summed, and the sums of the delay matrices corresponding to different excitation time intervals are added together to obtain an aliasing matrix.

[0095] Step S103: Initialize the parameters of the mixed sampling and separation model.

[0096] Initialize the parameters of the mixed sampling and separation model This includes the deep learning network parameters Θ and the geophysical guidance network tradeoff parameters δ and η. For example, δ and η are initialized to 0.1 and 0.9, respectively.

[0097] Step S104: Determine the initial values ​​of the main gun calculation data.

[0098] The aliasing matrix is ​​used to determine the initial value x0 of the main gun calculation data for the mixed-sampling separation model.

[0099] Set the aliasing matrix Γ such that it satisfies:

[0100] d=Γx

[0101] Where d represents aliased data; x represents the main gun calculation data.

[0102] According to the formula d=Γx, we can obtain x=Γ. -1 d, where Γ -1 Let Γ be the inverse matrix.

[0103] However, due to the underdeterminism of matrix Γ, its inverse matrix does not exist, therefore Γ is used. T Replacement Γ -1 Calculate the initial guess for x:

[0104] That is, x0 = Γ T d;

[0105] Therefore, in some alternative embodiments, according to x0=Γ T d, calculate the initial value x0 of the main gun calculation data, where d is the aliased data in the sample set, Γ T It is the transpose of the aliased matrix.

[0106] The initial training iterations are set to 0. Since the initial values ​​for the main gun's calculations are obtained without prior training, the initial training iterations are set to 0 after determining the initial values ​​and before the first training session. Recording the number of training iterations limits the maximum number of training iterations, suppressing overfitting in the model. Reaching the maximum number of training iterations can also be used as a preset training termination condition.

[0107] Step S105: Using the aliased data and corresponding excitation time intervals in the sample set as input data, and the corresponding main gun data as label data, train the aliasing separation model. The aliasing separation model includes a deep learning network and a geophysical guidance network.

[0108] Reference Figure 2As shown, the deep learning network adopts a U-shaped network structure, which consists of a feature encoding layer, a downsampling layer, a feature decoding layer, an upsampling layer, and an output layer.

[0109] The feature coding layer consists of three sets of convolutional layers with 64 channels and a filter size of 3. Each set of convolutional layers contains the ReLU (Rectified Linear Unit) activation function.

[0110] The downsampling layer consists of a set of convolutional layers with 64 channels, a filter size of 3, and a stride of 2. Each set of convolutional layers contains the ReLU activation function.

[0111] The feature decoding layer consists of a set of convolutional layers with 64 channels, a filter size of 1, and containing the ReLU activation function; three sets of convolutional layers with 64 channels, a filter size of 3, and containing the ReLU activation function; and a set of convolutional layers with 64 channels, a filter size of 1, and containing no activation function.

[0112] The downsampling layer consists of a set of deconvolutional layers with 64 channels, a filter size of 3, and a stride of 2.

[0113] The output layer consists of a set of convolutional layers with 1 channel and 3 filter size.

[0114] In some specific embodiments, δ and η are input into the geophysical guidance network. get In the next training iteration, after updating δ and η, we obtain the new... Similarly, with each update of δ and η, we can obtain the new value corresponding to each update.

[0115] The geophysical guidance network is used to weigh the parameters δ, η, and the main gun calculation data x as input parameters to a deep learning network, and based on... The new δ, η, and main gun calculation data x are obtained; in the next training, the obtained new δ, η, and main gun calculation data x are input into the deep learning network, and the new δ, η, and main gun calculation data x are obtained from the new δ and η are used to calculate the new δ and η. The next training iteration yields δ, η, and main gun calculation data x. Similarly, each time δ, η, and main gun calculation data x are input into the deep learning network, and the network is trained based on the input δ and η. Each time, new δ, η, and main gun calculation data x will be obtained.

[0116] Calculate the loss function We obtain a new Θ, where d i For a sample of aliased data, x i For d i Corresponding main gun data, m iFor d i The corresponding activation time interval, where N is the total number of samples in the sample set, and i is the index of the sample in the sample set, Net(d i m i ;Θ) is the mixed sampling separation model based on Θ, d i and m i The corresponding main gun calculation data is obtained. The loss function represents the value of Θ when the sum of the differences between the main gun calculation data corresponding to each sample obtained from the mixed sampling separation model and the corresponding main gun data is minimized. The minimization process can be implemented using the Adam optimization algorithm. Similarly, Θ is updated according to the loss function after each training iteration.

[0117] The mixed sampling separation model is updated with new parameters δ, η, and Θ, and the training iterations are also updated. After the above training process, the mixed sampling separation model is updated, and the training iterations are updated by incrementing the training iterations by 1.

[0118] After completing step S105, a judgment is made. If the preset training termination condition has not been met, step S106 is executed; otherwise, step S107 is executed. For example, if reaching the maximum number of training iterations is set as the preset training termination condition, then if the maximum number of training iterations has not been reached, step S106 is executed; otherwise, step S107 is executed. If the maximum number of training iterations is not set, the default maximum number of training iterations is 50.

[0119] Step S106: Adjust the parameters of the current network's mixed sampling separation model using the backpropagation algorithm.

[0120] The parameters of the current network's mixed sampling and separation model are adjusted using the backpropagation algorithm. This includes deep learning network parameters Θ and geophysical guidance network tradeoff parameters δ and η.

[0121] After executing step S106, execute step S105.

[0122] Step S107: Terminate training to obtain the trained smuggling separation model. The most recently updated smuggling separation model after training is terminated is taken as the trained smuggling separation model.

[0123] For example, in an application case of Block A in an oil field, the data being processed was marine seismic data. Synchronous source excitation was used, resulting in severe aliasing, which affected further processing and interpretation of the seismic data. Figure 3 This is mixed seismic data from this exploration block. Figure 4 The results of using the sparse inversion method for adjacent-shot interference separation show that the mixed-sampled signals are effectively separated. Figure 5To separate adjacent shot interference using the sparse inversion method, although this method effectively separates adjacent shot interference, it requires a lot of computer resources, resulting in a small amount of data that can be processed each time and low processing efficiency. The straight-line-like trace indicated by the black arrow in the figure shows the separation effect. Figure 6 For data obtained after merging and separating using this invention, Figure 7 This is the adjacent gun interference separated by the present invention. Figure 8 To compare the data after suppressing interference from adjacent guns using two different methods, the present invention demonstrates better separation performance. This indicates that the method of the present invention effectively suppresses interference from adjacent guns while effectively protecting the main gun data and significantly improving computational efficiency.

[0124] In this embodiment, a mixed mining separation model, including a deep learning network and a geophysical guidance network, was trained using seismic data obtained through various methods. This resulted in a mixed mining separation model that can effectively suppress interference from adjacent guns and effectively protect the main gun data, while also significantly improving computational efficiency.

[0125] Example 2

[0126] Embodiment 2 of the present invention provides a method for merging and separating samples, the process of which is as follows: Figure 9 As shown, it includes the following steps:

[0127] Step S201: Extract gathers from the seismic aliasing data to be separated to obtain the receiver gathers corresponding to the seismic aliasing data to be separated.

[0128] Step S202: Input the data of the receiver gather corresponding to the seismic aliasing data to be separated into the aliasing separation model, and obtain the main gun data based on the output separated data.

[0129] In some optional embodiments, the separated output data is transformed back into gun sets, i.e., main gun data, using a gather extraction method.

[0130] Mixed seismic data (such as...) Figure 10 As shown, the data is transformed into a receiver gather using conventional gather extraction methods. The receiver gather data is then input into a multi-information fusion-based mixed-sampling separation network model to obtain separated data. Finally, it is transformed back into a shot gather using conventional gather extraction methods (e.g., ...). Figure 11 (As shown). Figure 12 This is the main gun data that is not mixed. Figure 13 The residual between the processing result of this invention and the unmixed main gun data is shown by the black arrow in the figure, which indicates the straight line-like trace, showing the separation effect. The difference between the two is very small, proving that this embodiment can effectively suppress interference from adjacent guns and effectively protect the main gun data, and the computational efficiency is greatly improved.

[0131] For example, in an application case of Block A in an oil field, the data being processed was marine seismic data. Synchronous source excitation was used, resulting in severe aliasing, which affected further processing and interpretation of the seismic data. Figure 3 This is mixed seismic data from this exploration block. Figure 4 The results of using the sparse inversion method for adjacent-shot interference separation show that the mixed-sampled signals are effectively separated. Figure 5 To separate adjacent shot interference using the sparse inversion method, although this method is effective in separating adjacent shot interference, it requires a lot of computer resources, resulting in a small amount of data that can be processed each time and low processing efficiency. Figure 6 Data obtained after merging and separating using this invention Figure 7 This is the adjacent gun interference separated by the present invention. Figure 8 To compare the data after suppressing interference from adjacent guns using two different methods, the present invention demonstrates better separation performance. This indicates that the method of the present invention effectively suppresses interference from adjacent guns while effectively protecting the main gun data and significantly improving computational efficiency.

[0132] In this embodiment, by using a mixed-production separation model that includes a deep learning network and a geophysical guidance network, which has been trained, mixed-production separation can effectively suppress interference from adjacent shots and effectively protect the main shot data. This can significantly improve separation efficiency and accuracy, increase the signal-to-noise ratio of seismic data, and provide high-quality basic data for subsequent quantitative interpretation of seismic data and high-precision reservoir prediction.

[0133] Example 3

[0134] This invention provides a training device for a mixed sampling separation model, the structure of which is described in reference to... Figure 14 As shown, it includes:

[0135] Sample set generation module 101: is used to transform multiple sets of aliased data and their corresponding main gun data and corresponding excitation time intervals into detector gathers, and use the detector gathers as sample sets; and to extract a preset number of samples from the sample set to generate a sample set;

[0136] Mixed mining separation model training module 102: This module is used to input the sample set into the mixed mining separation model composed of a deep learning network and a geophysical guidance network for training until the preset training termination condition is met, thus obtaining the trained mixed mining separation model.

[0137] For example, in an application case of Block A in an oil field, the data being processed was marine seismic data. Synchronous source excitation was used, resulting in severe aliasing, which affected further processing and interpretation of the seismic data. Figure 3 This is mixed seismic data from this exploration block. Figure 4 The results of using the sparse inversion method for adjacent-shot interference separation show that the mixed-sampled signals are effectively separated. Figure 5To separate adjacent shot interference using the sparse inversion method, although this method is effective in separating adjacent shot interference, it requires a lot of computer resources, resulting in a small amount of data that can be processed each time and low processing efficiency. Figure 6 Data obtained after merging and separating using this invention Figure 7 This is the adjacent gun interference separated by the present invention. Figure 8 To compare the data after suppressing interference from adjacent guns using two different methods, the present invention demonstrates better separation performance. This indicates that the method of the present invention effectively suppresses interference from adjacent guns while effectively protecting the main gun data and significantly improving computational efficiency.

[0138] In this embodiment, a mixed mining separation model, including a deep learning network and a geophysical guidance network, was trained using seismic data obtained through various methods. This resulted in a mixed mining separation model that can effectively suppress interference from adjacent guns and effectively protect the main gun data, while also significantly improving computational efficiency.

[0139] Example 4

[0140] This invention provides a mixed sampling and separation device, the structure of which is as follows: Figure 15 As shown, it includes:

[0141] Hybrid data transformation module 201: used to extract gathers from the seismic aliasing data to be separated, and obtain the receiver gathers corresponding to the seismic aliasing data to be separated;

[0142] Mixed data separation module 202: is used to input the data of the detector point gather into the mixed data separation model obtained by the aforementioned mixed data separation model training method, obtain the separated data, and then transform it back into shot gather according to the gather extraction method.

[0143] For example, in an application case of Block A in an oil field, the data being processed was marine seismic data. Synchronous source excitation was used, resulting in severe aliasing, which affected further processing and interpretation of the seismic data. Figure 3 This is mixed seismic data from this exploration block. Figure 4 The results of using the sparse inversion method for adjacent-shot interference separation show that the mixed-sampled signals are effectively separated. Figure 5 To separate adjacent shot interference using the sparse inversion method, although this method is effective in separating adjacent shot interference, it requires a lot of computer resources, resulting in a small amount of data that can be processed each time and low processing efficiency. Figure 6 Data obtained after merging and separating using this invention Figure 7 This is the adjacent gun interference separated by the present invention. Figure 8 To compare the data after suppressing interference from adjacent guns using two different methods, the present invention demonstrates better separation performance. This indicates that the method of the present invention effectively suppresses interference from adjacent guns while effectively protecting the main gun data and significantly improving computational efficiency.

[0144] In this embodiment, by using a mixed-production separation model that includes a deep learning network and a geophysical guidance network, which has been trained, mixed-production separation can effectively suppress interference from adjacent shots and effectively protect the main shot data. This can significantly improve separation efficiency and accuracy, increase the signal-to-noise ratio of seismic data, and provide high-quality basic data for subsequent quantitative interpretation of seismic data and high-precision reservoir prediction.

[0145] Based on the same inventive concept, embodiments of the present invention also provide a computer storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned mixed-sampling separation model training method or mixed-sampling separation method.

[0146] Based on the same inventive concept, embodiments of the present invention also provide a terminal device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned mixed mining separation model training method or mixed mining separation method.

[0147] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0148] In the detailed description above, various features are combined together in a single embodiment to simplify this disclosure. This approach to disclosure should not be construed as reflecting an intention that embodiments of the claimed subject matter require more features than are explicitly stated in each claim. Rather, as reflected in the appended claims, the invention is presented with fewer features than all of the features in a single disclosed embodiment. Therefore, the appended claims are hereby explicitly incorporated into the detailed description, with each claim representing a separate preferred embodiment of the invention.

Claims

1. A training method for a mixed sampling separation model, characterized in that, include: Gatherings were extracted from the multiple sets of aliased data, the corresponding main gun data, and the excitation time interval. The receiver gathers corresponding to the acquired seismic data were used as the sample set. Using the aliased data and corresponding excitation time intervals in the sample set as input data, and the corresponding main gun data as label data, a mixed-mining separation model is trained. The mixed-mining separation model includes a deep learning network and a geophysical guidance network. When the preset training termination conditions are met, the trained mixed sampling separation model is obtained. The aliasing data includes an aliasing matrix constructed based on the excitation time interval of each sample in the sample set. The process of constructing the aliasing matrix includes: Convert the excitation time interval of each sample in the sample set into an integer; Based on the integerized excitation time interval of each sample in the sample set, a delay matrix is ​​determined for each sample corresponding to different excitation time intervals. In the delay matrix, the element whose row number minus column number equals the integerized excitation time interval of each sample in the sample set is assigned a value of 1, and other elements are assigned a value of 0. The delay matrices of each sample at the same excitation time interval are summed, and the sums of the delay matrices corresponding to different excitation time intervals are added together to obtain the aliasing matrix.

2. The method as described in claim 1, characterized in that, The process involves extracting gathers from multiple sets of aliased data, corresponding main gun data, and excitation time intervals. The resulting receiver gathers corresponding to the acquired seismic data are used as a sample set, including: Seismic data was acquired using at least one of the following methods: Select the seismic data of two shots that are not overlapped, and superimpose the seismic data of the two shots at random time intervals to obtain the seismic overlapped data of the main shot and the adjacent shot. The random time interval is the excitation time interval corresponding to the overlapped data. Numerical simulation techniques are used to simulate conventionally acquired seismic data to obtain main gun data, and to simulate aliased acquired seismic data to obtain aliased seismic data and corresponding excitation time intervals. Multiple sets of aliased data were collected, and the main gun data and the corresponding firing time interval were separated using aliasing separation technology. The acquired seismic data is transformed into receiver gathers, and these receiver gathers are used as a sample set.

3. The method as described in claim 1, characterized in that, Before training the mixed-sampling separation model, the process includes using the aliased data and corresponding firing time intervals from the sample set as input data, and the corresponding main gun data as label data. Construct an aliasing matrix based on the excitation time interval of each sample in the sample set; Initialize the parameters of the mixed sampling and separation model , Including deep learning network parameters And the geophysical guidance network tradeoff parameters δ, η; The aliasing matrix is ​​used to determine the initial value x0 of the main gun calculation data for the mixed-sampling separation model; The initial training iterations are set to 0.

4. The method as described in claim 1, characterized in that, The step of determining the delay matrix for each sample corresponding to different excitation time intervals based on the integerized excitation time interval of each sample in the sample set includes: Set an N t 1-th order matrix m is the excitation time interval, N t It is a positive integer; matrix The row ordinal number minus the column ordinal number equals the integerized excitation time interval m of each sample in the sample set. The elements are assigned a value of 1, and the other elements are assigned a value of 0, to obtain the delay matrix corresponding to the sample with an excitation time interval of m. Repeat the above process to obtain the delay matrix corresponding to samples with different excitation time intervals.

5. The method as described in claim 3, characterized in that, The step of using the aliasing matrix to determine the initial value x0 of the main gun calculation data for the mixed-sampling separation model includes: according to Determine the initial value x0 for the main gun calculation data, where d represents the aliased data in the sample set. It is the transpose of the aliased matrix.

6. The method as described in claim 3, characterized in that, The step of using the aliased data and corresponding firing time intervals in the sample set as input data, and the corresponding main gun data as label data, to train the mixed-sampling separation model includes: Input δ and η into the geophysical guidance network get ; The geophysical guidance network tradeoff parameters and the main gun calculation data x are input parameters as follows: The deep learning network, and based on This yields new δ, η, and main gun calculation data x; Calculate the loss function , get new , where d i For a sample of aliased data, x i For d i Corresponding main gun data, m i For d i The corresponding activation time interval, where N is the total number of samples in the sample set, and i is the index of the sample in the sample set. For the mixed sampling separation model based on d i and m i The corresponding main gun calculation data obtained; With new δ, η and For the new parameters, update the mixed sampling separation model and update the number of training iterations.

7. The method as described in claim 6, characterized in that, After updating the training iterations, the following is also included: If the maximum number of training iterations is not met, the parameters of the current network's hybrid sampling and separation model are adjusted using the backpropagation algorithm. .

8. A method for merging and separating samples, characterized in that, The mixed mining separation model is implemented using the mixed mining separation model training method as described in any one of claims 1-7, including: The seismic aliasing data to be separated is subjected to gather extraction to obtain the receiver gather corresponding to the seismic aliasing data to be separated. The data of the receiver gathers corresponding to the seismic aliasing data to be separated are input into the seismic aliasing model, and the main gun data is obtained based on the output separated data.

9. The method as described in claim 8, characterized in that, The process of obtaining main gun data based on the separated output data includes: The separated output data is transformed back into gun data using the method of gather extraction, and the gun data is the main gun data.

10. A mixed sampling separation model generation device, characterized in that, include: Sample set generation module: used to transform multiple sets of aliased data and their corresponding main gun data and corresponding excitation time intervals into detector point gathers, and use the detector point gathers as sample sets; A predetermined number of samples are extracted from the sample set to generate a sample set; The aliasing data includes an aliasing matrix constructed based on the excitation time interval of each sample in the sample set. The process of constructing the aliasing matrix includes: Convert the excitation time interval of each sample in the sample set into an integer; Based on the integerized excitation time interval of each sample in the sample set, a delay matrix is ​​determined for each sample corresponding to different excitation time intervals. In the delay matrix, the element whose row number minus column number equals the integerized excitation time interval of each sample in the sample set is assigned a value of 1, and other elements are assigned a value of 0. Summing the delay matrices of each sample within the same excitation time interval, and adding the sums of the delay matrices corresponding to different excitation time intervals, yields the aliasing matrix; The mixed mining separation model training module is used to input the sample set into the mixed mining separation model composed of a deep learning network and a geophysical guidance network for training until the preset training termination conditions are met, and the trained mixed mining separation model is obtained.

11. A mixed sampling and separation device, characterized in that, include: Hybrid data transformation module: used to extract gathers from the seismic aliasing data to be separated, and obtain the receiver gathers corresponding to the seismic aliasing data to be separated; Mixed data separation module: used to input the data of the detector point gather into the mixed data separation model obtained by the mixed data separation model training method of any one of claims 1-7, obtain the separated data, and then transform it back into shot gather according to the gather extraction method.

12. A computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which, when executed by a processor, implement the mixed-sampling separation model training method of any one of claims 1-7 or the mixed-sampling separation method of any one of claims 8-9.

13. A terminal device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the mixed sampling separation model training method of any one of claims 1-7 or the mixed sampling separation method of any one of claims 8-9.

Images