Intelligent surface wave suppression methods, systems, storage media, and electronic devices for seismic data

Abstract

This invention belongs to the technical field of oil and gas geophysical exploration, and relates to a method, system, storage medium, and electronic device for intelligent surface wave suppression of seismic data. The method includes constructing seismic shot gather data containing surface waves and obtaining clean learning labels; building a QAU-Net network architecture based on the U-Net network architecture and introducing an attention mechanism; training a network model for suppressing surface waves; and evaluating the training effect of the network model. This invention improves the accuracy of seismic surface wave suppression.

Description

Technical Field

[0001] This invention belongs to the technical field of oil and gas geophysical exploration, and relates to a method, system, storage medium and electronic device for intelligent surface wave suppression of seismic data. Background Technology

[0002] Seismic surface waves, primarily Rayleigh and Love waves, are common and powerful wave types in seismic exploration. Although they carry rich shallow geological information, their low-frequency and high-amplitude characteristics often mask more critical body wave signals (such as P-waves and S-waves), posing a significant challenge to the quality of seismic data and subsequent geological interpretation. Therefore, effectively suppressing or eliminating surface waves while preserving valuable geological information has become a research hotspot in the field of seismic exploration.

[0003] Traditional methods for suppressing seismic surface waves primarily focus on filtering techniques in the frequency, wavenumber, and spatial domains. These methods utilize the differences between seismic surface waves and body waves in different domains to achieve suppression. Frequency domain filtering leverages the low-frequency characteristics of seismic surface waves, using appropriately designed bandpass filters to attenuate low-frequency components and suppress them. Wavenumber domain filtering utilizes the differences between seismic surface waves and body waves in the wavenumber domain, separating them using wavenumber filters. The advantage of wavenumber domain filtering is its ability to more precisely control the degree of surface wave suppression. Similar methods include τ-p transform, wavelet transform, local time-frequency decomposition, curvelet transform, and sparse decomposition. However, in practical applications, these methods are limited by the requirements of array deployment.

[0004] Furthermore, the approximately linear characteristics of seismic surface waves can be utilized to eliminate their influence, such as plane wave decomposition, singular value decomposition, interferometry removal, and prediction error filters (PEF). Although these methods can reduce the interference of surface waves on the effective signal to some extent, they inevitably damage the reflection information of the effective signal. Surface wave prediction and suppression algorithms based on dispersion characteristics fully utilize the dispersion characteristics of surface waves. By analyzing the data to obtain dispersion curves, a surface wave model is constructed, and finally, surface waves are removed from the data through reduction strategies. Commonly used methods include high-precision linear Radon method, fk transform method, wavelet transform method, τ-p transform method, and phase shift method. However, these methods have low accuracy in the low-frequency band, which can easily affect the accuracy of seismic surface wave calculations.

[0005] Based on the above problems, although traditional filtering methods can weaken the influence of surface waves to a certain extent, they also have the disadvantage of damaging the effective signal reflection information. Dispersion-based surface wave prediction and suppression algorithms rely on the accuracy of dispersion spectrum calculation and acquisition. Therefore, it is urgent to develop a high-performance intelligent surface wave suppression method and system to improve the accuracy of seismic surface wave suppression. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method, system, storage medium, and electronic device for intelligent surface wave suppression of seismic data.

[0007] In a first aspect, the present invention provides a method for intelligent suppression of surface waves in seismic data, employing the following technical solution:

[0008] A method for intelligent surface wave suppression of seismic data is disclosed. The method includes: constructing seismic shot gather data containing surface waves and obtaining clean learning labels; based on the U-Net network architecture, introducing an attention mechanism and building a QAU-Net network architecture; training a network model for suppressing surface waves; and evaluating the training effect of the network model.

[0009] Preferably, the construction of seismic shot gather data containing surface waves includes: establishing a model including a low-velocity caprock and a high-velocity subsurface in half-space for forward modeling of pure surface waves; the aim is to acquire pure surface waves separately.

[0010] A subsurface structural model was constructed to perform forward modeling of subsurface effective waves, generating subsurface effective waves free from surface wave interference.

[0011] The forward modeling results of pure surface waves and effective subsurface waves are superimposed to obtain seismic shot gather data containing surface waves.

[0012] Preferably, obtaining clean learning labels includes: constructing a subsurface structural model to perform forward modeling of subsurface effective waves, generating clean data free from surface wave interference, and obtaining clean learning labels.

[0013] Preferably, the QAU-Net network architecture is built by introducing an attention mechanism based on the U-Net network architecture, which includes: introducing an attention mechanism between neurons in skip connections, so that the QAU-Net network ignores unimportant feature vectors and focuses on calculating useful vectors.

[0014] Preferably, the training of the network model for suppressing surface waves includes: using the seismic shot gather data containing surface waves as input to the network model, and using the learned labels as the expected output of the network model;

[0015] By using appropriate optimization algorithms to train the network model and continuously iterating and updating the network parameters, the network output gradually approaches the expected output of the network model.

[0016] Preferably, the training network model for suppressing surface waves includes a network output that can be represented by the following formula:

[0017] s out (t)=f θ (ε,s(t))

[0018] Where f represents the network model, ε represents the editable parameters in the network, θ represents all training weight parameters, and s(t) represents data containing surface waves.

[0019] Preferably, the changes in the loss function and the performance of the network model on the validation set are monitored in real time to evaluate the training effect of the network model.

[0020] Secondly, this invention provides an intelligent surface wave suppression system for seismic data, employing the following technical solution:

[0021] A smart surface wave suppression system for seismic data, comprising a data preparation and label creation module, a network model building module, a network model training module, and a network model optimization module;

[0022] The data preparation and labeling module is used to construct seismic shot gather data containing surface waves and obtain clean learning labels.

[0023] The network model building module is used to introduce an attention mechanism and build a QAU-Net network;

[0024] The network model training module is used to train the surface wave suppression model;

[0025] The network model optimization module is used to evaluate the training effect of the network model.

[0026] Thirdly, the present invention provides a computer storage medium, which adopts the following technical solution:

[0027] A computer storage medium storing one or more programs, which, when executed, can implement the above-described intelligent surface wave suppression method for seismic data.

[0028] Fourthly, the present invention provides a device that adopts the following technical solution:

[0029] A device includes a processor, a communication interface, a memory, and a communication bus; the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory stores at least one program in the computer storage medium that can be loaded and executed by the processor.

[0030] In summary, the present invention has the following beneficial technical effects:

[0031] 1. The quadratic neurons in the intelligent surface wave suppression method for seismic data have higher nonlinear characteristics than traditional linear neurons, which improves the fitting ability of the neural network and enables convergence within a shorter training cycle. They have a strong ability to suppress surface waves and recover weak signals.

[0032] 2. The method of separately generating surface waves and effective waves in the intelligent surface wave suppression method for seismic data makes the labels learned by the network more accurate. The random combination of surface waves and effective waves can enrich the training set, which is more beneficial to the network's learning. Finally, through testing with validation data, it has better suppression performance for surface waves.

[0033] 3. The intelligent surface wave suppression method for seismic data ensures that the model can accurately distinguish between surface waves and effective waves during the learning process, achieving high-precision intelligent surface wave suppression, improving the accuracy of labels, and providing greater potential and space for model learning. Attached Figure Description

[0034] Figure 1 The flowchart of the intelligent surface wave suppression method for seismic data has been made public.

[0035] Figure 2 It discloses a schematic diagram of the structure of the forward model on the training set.

[0036] Figure 3 It discloses structural diagrams of traditional linear neuron architecture and secondary neuron architecture.

[0037] Figure 4 The attention mechanism and the structural diagram of the QAU-Net network are disclosed.

[0038] Figure 5 It is a schematic diagram of the training set creation and model training process for surface wave intelligent suppression.

[0039] Figure 6 It is a diagram showing the dataset used to train the network model.

[0040] Figure 7 It is a diagram showing the test results of the publicly available dataset.

[0041] Figure 8 The structural diagram of the fk spectrum has been disclosed.

[0042] Figure 9 It discloses a framework diagram of a smart surface wave suppression method for seismic data.

[0043] Figure 10 It discloses a schematic diagram of an electronic device structure. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] The following is in conjunction with the appendix Figure 1-10 A preferred description of the present invention is provided below.

[0046] Currently, in the field of oil and gas exploration, the use of deep learning to process, interpret, and invert seismic data has become a research direction for many scholars. Deep learning has made great progress in picking first arrival waves, reservoir prediction, fault identification, and impedance inversion, but the intelligent suppression technology of seismic surface waves is still in its initial stage.

[0047] See appendix Figure 1 As shown, this invention proposes an intelligent surface wave suppression method for seismic data. This method mainly studies the application of machine learning methods in seismic exploration data processing, and realizes efficient and intelligent surface wave suppression of seismic data based on the QAU-Net (uadratic Convolution Kernel and Attention Mechanism U-Net) network.

[0048] The intelligent surface wave suppression method for seismic data includes the following steps:

[0049] S1. Data Preparation and Label Creation

[0050] Seismic surface wave noise is characterized by low velocity, dispersion, and high amplitude. Traditional methods involve constructing a low-velocity cap layer above the forward model to model surface waves. The thickness of this cap layer is typically controlled within one wavelength to ensure the dispersion characteristics of the surface waves can be captured. However, traditional methods have limitations in suppressing surface wave interference to obtain clean labels, thus limiting the learning potential of the forward model and making it difficult for the model to overcome the performance bottlenecks of traditional methods.

[0051] For example, the basic theory of deep learning is the Logistic Regression model, which generally assumes that for input data X, a predicted result Y can be obtained. The mathematical model can be expressed as:

[0052] Y = w T X+b (1)

[0053] In this neural network, the input data X is a matrix vertically stacked from the input features x1, x2, x3, etc.; w represents the weight coefficients, which are the coefficients of the linear relationship between the input features and the target variable; b represents the bias term, which is a constant term; X, w, and b represent the numerical matrix in the neural network; T represents the transpose operation; w T This represents the transpose of the weight matrix w;

[0054] Therefore, the deep learning process can be viewed as a matrix operation, and the final prediction result Y is obtained through a mathematical model.

[0055] To train the parameters w and b of the logistic regression model, the label Y needs to be input. * Define the loss function L(Y, Y) * The performance of the algorithm is measured by the following formula:

[0056]

[0057] When the loss function L(Y, Y) * Only when the values ​​are as small as possible can we find suitable parameters w and b.

[0058] In deep learning, neural networks process input data X and Y. * Feature learning is performed so that when deep learning encounters the same feature again, it can recognize it. However, it can only recognize features that exist in the input data, so it is necessary to create a targeted sample label dataset.

[0059] Currently, intelligent surface wave suppression methods are limited not only by the network's performance but also by low label generation quality. When generating labels, traditional filtering methods are used to create surface wave-free learning samples, but the learned forward model still suffers from the shortcomings of traditional filtering methods.

[0060] To address the limitations, a separate forward modeling method for seismic surface waves and subsurface structures can be used to create accurate label pairs with and without surface waves. This involves the following steps:

[0061] S11, Forward Modeling of Surface Waves and Effective Waves

[0062] 1. Pure surface wave forward modeling: A model including a low-velocity caprock and a high-velocity subsurface in half-space is established for pure surface wave forward modeling. The aim is to obtain the characteristic information of surface waves separately. The thickness of the low-velocity caprock is generally controlled within one wavelength to ensure the acquisition of the dispersion characteristics of surface waves.

[0063] 2. Subsurface effective wave forward modeling: Construct a subsurface structural model to realize the forward modeling simulation of subsurface effective waves (such as P-waves and S-waves). This step will generate clean data without surface wave interference, which will serve as the basis for subsequent production of surface wave-free labels.

[0064] For example, see Appendix Figure 2 (a) shows the low-velocity caprock model for forward modeling surface waves (six of which are examples) which can perform forward modeling of pure surface waves; (b) shows the subsurface structure model (six of which are examples) used for forward modeling of subsurface structure response.

[0065] S12, Data Overlay and Label Creation

[0066] 1. The forward modeling results of pure surface waves and effective subsurface waves in step S11 are superimposed to obtain complete seismic shot gather data containing surface waves;

[0067] 2. Clean data obtained from underground effective wave forward modeling are used as clean learning labels.

[0068] S2, Network Model Construction

[0069] Based on a quadratic neuron architecture with highly nonlinear characteristics, an attention mechanism is introduced to construct a network called QAU-Net (Quadratic Convolution Kernel and Attention Mechanism U-Net) with high fitting and fast convergence properties, including the following steps:

[0070] S21, QAU-Net network architecture design

[0071] 1. See appendix Figure 3 The traditional linear neuron architecture and the quadratic neuron architecture applied in this invention are shown respectively. The network of this invention adopts a U-shaped architecture, and the neurons in the encoder and decoder all use a quadratic neuron architecture, which increases the nonlinear characteristics of the network.

[0072] Appendix Figure 3 In the diagram, (a) represents a traditional linear neuron, and (b) represents a quadratic neuron; w represents the weight parameter, which connects the input node (the node in the previous layer) and the current node.

[0073] Xinput represents the input node of a neuron, which is the location where the neuron receives external information or the output of the neuron in the previous layer; Xout represents the output node of a neuron, which is the location where the neuron passes the processed information to the neuron in the next layer or outputs it to the outside.

[0074] This represents the addition operation; σ represents the convolution operation; b represents the bias term, a constant of the neuron; "⊙" represents the Hadamard product, which is the element-wise multiplication of two matrices or vectors of the same shape; σ represents the activation function, which determines the nonlinear transformation of the neuron's output information.

[0075] The U-shaped network architecture described above exhibits superior performance in many image recognition tasks where the input and output share the same spatial resolution. This strong performance is primarily attributed to the linked encoder-decoder architecture.

[0076] 2. See appendix Figure 4 Based on the U-Net network architecture, an attention mechanism is introduced between neurons in skip connections, which allows the neural network to ignore unimportant feature vectors and focus on computing useful vectors, thereby improving the network's computational efficiency and performance.

[0077] Among them, the U-Net network has the high-level features of deep networks while maintaining the model space structure that matches the output with the shallow low-level features; the features are downsampled at multiple spatial resolutions in the encoder and recombined with the corresponding upsampled features through skip links in the decoder, which improves the network's ability to capture key information.

[0078] Appendix Figure 4 (a) shows the attention mechanism, and (b) shows the QAU-Net network architecture. While discarding useless features that interfere with the fitting results, it can also improve the computation speed. The neurons in the downsampling and upsampling are built using higher-performance secondary neurons to construct the QAU-Net network.

[0079] S22, Network Parameter Initialization

[0080] We initialized the weight parameters, bias weights, and other parameters in the QAU-Net network in preparation for network training.

[0081] S3, Network Model Training

[0082] By using the label pairs with and without surface wave mapping relationships created in step S1, a high-performance surface wave intelligent suppression model is trained to achieve high-precision intelligent suppression of surface waves.

[0083] S31. Data Input and Output

[0084] See appendix Figure 5 This demonstrates the process of creating a training set and training a model for surface wave intelligent suppression.

[0085] ① represents the production process of surface wave data pairs. Based on the forward modeling of the constant elastic wave equation, pure surface waves and effective waves reflected from underground structures are derived from the low-velocity near-surface model and the underground structure model, respectively. The effective waves from the underground structure forward model are used as the input of the network through superposition.

[0086] The complete seismic shot gather data containing surface waves in step S12 is used as input to the network model;

[0087] The learned labels in step S12 serve as the expected output of the network model, ensuring that the effective wave information of the model's input matches the information of the learned labels perfectly.

[0088] Where ② represents the model training process, the model is QAU-Net, compared with the conventional U-Net model, the addition of the attention mechanism module increases the network computation efficiency, and the introduction of secondary neurons greatly improves the performance of the model, resulting in a network model that efficiently suppresses surface waves.

[0089] S32, Training the Network Model

[0090] By using appropriate optimization algorithms (such as Adam and SGD) to train the network model and continuously updating the network parameters, the network output gradually approaches the expected output of the network model. Deep learning has shown great advantages in the field of image feature extraction. The network acquires effective wave features through an end-to-end learning process to achieve the purpose of suppression.

[0091] See appendix Figure 6 The results show that (a) is pure surface wave data from the forward modeling of the low-velocity caprock, (b) is the effective wave response from the forward modeling of the subsurface structure, and (c) is the superposition result of (a) and (b), which serves as noisy data containing surface waves. This data is used as input to the network, and the richness of the data is increased by randomly combining (a) and (b).

[0092] First, the network output can be represented by the following formula:

[0093] s out (t)=f θ (ε,s(t)) (3)

[0094] Where f represents the network model, ε represents the editable parameters in the network, θ represents all training weight parameters, and s(t) represents data containing surface waves.

[0095] The training process of the model can be viewed as learning the mapping relationship and obtaining the optimal weights θ by minimizing the loss function based on the constructed pairs of surface wave data with and without surface waves. * This causes the network output s out (t) Approximating the expected output s of the network model * (t) can be expressed by the following formula:

[0096]

[0097] The type of neurons often determines the upper limit of a network. Typically, a neuron in the forward propagation of a convolutional layer can be represented by the following formula:

[0098] X out =WX+b (5)

[0099] Among them, X out denoted by , where X is the image input, W is the kernel weight, and b is the bias weight.

[0100] To improve the approximation ability of DNNs, the only way is to increase the weight coefficients by improving the geometry of the DNN (width, length, and special connections, etc.). The introduction of quadratic neural networks changes the uniqueness of the boosting path, increases the order of neurons, and thus increases the nonlinearity of the network. The improved quadratic neuron expression is as follows:

[0101] X out =W a X+W b X+W c X+b (6)

[0102] Among them, W a W b W c These represent different weighting coefficients; b represents the bias weight.

[0103] S4, Network Model Optimization

[0104] During training, the changes in the loss function and the network's performance on the validation set are monitored in real time to evaluate the training effect of the network.

[0105] Example

[0106] This invention is applied to forward modeling data outside the training set for validation; see appendix. Figure 7 As shown, (a) represents input data containing surface waves; (b) represents clean data without surface waves; (c) represents the output of a conventional U-Net network, showing incomplete suppression; (d) is the output of the present invention, which is highly similar to the clean data (b). Subtracting the two output results from the clean data yields the verification results (e) and (f), showing that the present invention has the smallest difference from the clean data.

[0107] The output of this invention can also well reproduce the characteristics of the effective wave in the frequency-wavenumber domain. See Appendix. Figure 8As shown, (a) shows the frequency-wavenumber domain characteristics of the input surface wave data, where low-speed, high-energy surface wave information is evident. After suppression by the present invention, the surface wave information in (b) is almost completely eliminated, and the effective wave in the frequency-wavenumber domain is similar to that of clean surface wave-free data.

[0108] See appendix Figure 9 This invention proposes an intelligent surface wave suppression system for seismic data, which includes a data preparation and label creation module, a network model building module, a network model training module, and a network model optimization module.

[0109] The data preparation and labeling module is used to construct seismic shot gather data containing surface waves and obtain clean learning labels.

[0110] The network model building module is used to introduce an attention mechanism and build a QAU-Net network;

[0111] The network model training module is used to train the surface wave suppression model;

[0112] The network model optimization module is used to evaluate the training effect of the network model.

[0113] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0114] This invention provides a computer storage medium storing instructions executable by at least one processor, which are executed by the at least one processor to enable the at least one processor to execute a computer program for a seismic data surface wave intelligent suppression method as described above.

[0115] The storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, read-only memory (ROM), random access memory (RAM), or optical discs.

[0116] This invention provides a device, see attached document. Figure 10The computer device, used to execute the aforementioned encoding management platform, includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities; the memory includes a non-volatile storage medium and internal memory; the processor executes the computer program to implement the steps of the aforementioned encoding management method; the non-volatile storage medium stores the operating system, computer program, and database; the internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium.

[0117] It should be noted that the user information (including but not limited to user device information and personal information) and data (including but not limited to data used for analysis, data stored and data displayed) involved in this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0118] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for intelligent surface wave suppression of seismic data, characterized in that, The method includes: Construct seismic shot gather data containing surface waves and obtain clean learning labels; Based on the U-Net network architecture, an attention mechanism is introduced to build the QAU-Net network architecture; Training a network model to suppress surface waves; Evaluate the training performance of the network model.

2. The intelligent surface wave suppression method for seismic data according to claim 1, characterized in that: The construction of seismic shot gather data containing surface waves includes: A model including a low-velocity caprock and a high-velocity subsurface in a semi-space was established for forward modeling of pure surface waves; the aim was to obtain pure surface waves independently. A subsurface structural model was constructed to perform forward modeling of subsurface effective waves, generating subsurface effective waves free from surface wave interference. The forward modeling results of pure surface waves and effective subsurface waves are superimposed to obtain seismic shot gather data containing surface waves.

3. The intelligent surface wave suppression method for seismic data according to claim 1, characterized in that: The process of obtaining clean learning labels includes: A subsurface structural model is constructed to perform forward modeling of effective subsurface waves, generating clean data free from surface wave interference and obtaining clean learning labels.

4. The intelligent surface wave suppression method for seismic data according to claim 1, characterized in that: The QAU-Net network architecture, based on the U-Net network architecture and incorporating an attention mechanism, includes: The introduced attention mechanism between skip connections allows the QAU-Net network to ignore unimportant feature vectors and focus on computing useful vectors.

5. The intelligent surface wave suppression method for seismic data according to claim 2, characterized in that: The network model for training and suppressing surface waves includes: The seismic shot gather data containing surface waves is used as the input to the network model, and the learned labels are used as the expected output of the network model. By using appropriate optimization algorithms to train the network model and continuously iterating and updating the network parameters, the network output gradually approaches the expected output of the network model.

6. The intelligent surface wave suppression method for seismic data according to claim 1, characterized in that: The network model for training and suppressing surface waves includes: The network output of the network model can be expressed by the following formula: s out (t)=f θ (ε,s(t)) Where f represents the network model, ε represents the editable parameters in the network, θ represents all training weight parameters, and s(t) represents data containing surface waves.

7. The intelligent surface wave suppression method for seismic data according to claim 1, characterized in that: The evaluation of the training performance of the network model includes: The changes in the loss function and the performance of the network model on the validation set are monitored in real time to evaluate the training effect of the network model.

8. A smart surface wave suppression system for seismic data, characterized in that, The system includes a data preparation and label creation module, a network model building module, a network model training module, and a network model optimization module. The data preparation and labeling module is used to construct seismic shot gather data containing surface waves and obtain clean learning labels. The network model building module is used to introduce an attention mechanism and build a QAU-Net network; The network model training module is used to train the surface wave suppression model; The network model optimization module is used to evaluate the training effect of the network model.

9. A computer storage medium, characterized in that, The device stores one or more programs that, when executed, can implement the intelligent surface wave suppression method for seismic data as described in any one of claims 1-7.

10. A device, characterized in that: It includes a processor, a communication interface, a memory, and a communication bus; the memory stores at least one program that can be loaded by the processor and executed from the computer storage medium as described in claim 9.

Images