A Knowledge-Guided Method for Separating P-waves and S-waves in Complex Geological Structures

Abstract

This invention discloses a knowledge-guided method for separating P-waves and S-waves in complex geological structures, applied to the field of seismic data processing. It addresses the problem that existing technologies struggle to effectively separate P-waves and S-waves in complex geological structures. First, this invention utilizes a recurrent neural network for elastic wave full-waveform inversion to obtain knowledge representations of P-waves and S-waves in complex structures. Second, using these knowledge representations, a dual-branch autoencoder network is constructed. One branch separates P-waves under the guidance of the P-wave knowledge representation, while the other branch separates S-waves under the guidance of the S-wave knowledge representation. Finally, an autoencoder network architecture is constructed that comprehensively constrains wavefield reconstruction error, P-wave knowledge-guided error, and S-wave knowledge-guided error. Theoretical analysis and numerical calculations demonstrate that the proposed method can effectively separate P-waves and S-waves in complex geological structures, reducing dependence on data samples and exhibiting a certain degree of robustness.

Description

Technical Field

[0001] This invention belongs to the field of earthquake data processing, and specifically relates to an earthquake P-wave and S-wave separation technology. Background Technology

[0002] P-wave and S-wave separation is a crucial step in multi-component seismic data processing, and its effectiveness directly impacts the overall quality of the processed data. In elastic media, seismic records contain multiple wavefield components, primarily P-waves (P waves) and S-waves (S waves). In complex geological structures, P-waves and S-waves undergo intricate transformations and mixing, affecting the quality of subsurface geological imaging. Furthermore, due to the complexity of the P-wave and S-wave fields, accurate wavefield identification and separation based solely on salient signal characteristics is challenging. Therefore, achieving effective P-wave and S-wave separation is essential for multi-component seismic data processing.

[0003] Currently, methods for achieving P-wave and S-wave separation include transform domain-based methods, wavelength extension methods, and artificial intelligence methods. Transform domain-based methods are effective for P-wave and S-wave separation under simple geological structures, but under complex geological structures, the characteristic energies of P-waves and S-waves overlap in the transform domain, resulting in impure wavefields. Wavefield extension methods extend the seismic record downwards from the recording plane to a reference depth, achieving P-wave and S-wave separation through polarization vectors. Subsequently, the separated P-wave and S-wave records are extracted at the reference depth and extended upwards to achieve P-wave and S-wave separation. This method achieves good P-wave and S-wave separation when the model parameters are accurate, but it is overly dependent on accurate model parameters. Artificial intelligence methods do not require accurate parameters, simplifying the wavefield separation process, but they are typically data-driven, making them less interpretable and relying on large amounts of high-quality data labels.

[0004] In summary, the existing technology has the following problems:

[0005] 1. Current transform domain-based methods are effective in simple geological structures, but may lead to impurities in P-waves and S-waves in complex geological conditions;

[0006] 2. Although the wavelength extension method works well when the model parameters are accurate, it is necessary to find ways to reduce its dependence on accurate parameters due to the uncertainty of parameters in actual geological conditions.

[0007] 3. Although intelligent methods do not require accurate model parameters, their interpretability is relatively poor, and they rely on a large number of high-quality data samples. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a knowledge-guided method for separating P-waves and S-waves in complex geological structures. This method extracts and represents the knowledge contained within the data itself, and uses this knowledge representation to guide an autoencoder network to achieve the separation of P-waves and S-waves. This reduces the dependence of traditional artificial intelligence methods on data samples, and can effectively separate P-waves and S-waves in complex geological structures, exhibiting a certain degree of robustness and the ability to analyze specific data.

[0009] The technical solution adopted in this invention is as follows: a knowledge-guided method for separating P-waves and S-waves in complex geological structures. First, knowledge extraction is achieved by using RNN to perform full waveform inversion of elastic waves, thereby obtaining knowledge representations of P-waves and S-waves in complex structures. Second, using the knowledge representation results of P-waves and S-waves in complex structures, a dual-branch autoencoder network is constructed. One branch achieves P-wave separation under the guidance of P-wave knowledge representation, and the other branch achieves S-wave separation under the guidance of S-wave knowledge representation.

[0010] The beneficial effects of this invention are as follows: Traditional P-wave and S-wave separation methods struggle to separate P-waves and S-waves in complex geological structures, or rely excessively on accurate model parameters. Artificial intelligence methods, on the other hand, rely excessively on high-quality data samples. This invention fully utilizes the information inherent in the observation records themselves, employing RNNs to extract and represent knowledge from the records. While reconstructing the observation records through an autoencoder network, the knowledge representation guides the network to achieve P-wave and S-wave separation. This invention reduces reliance on data samples, has been validated in complex geological models, and provides a new approach to P-wave and S-wave separation in complex geological structures. Attached Figure Description

[0011] Figure 1 It is a self-encoder structure.

[0012] Figure 2 This is a flowchart of the solution of the present invention.

[0013] Figure 3 It is an RNN unit structure.

[0014] Figure 4 To implement knowledge representation for RNNs.

[0015] Figure 5 To implement the full waveform inversion process for RNN.

[0016] Figure 6 A knowledge-guided self-encoding network structure.

[0017] Figure 7 For the complex Model-A model;

[0018] Wherein, (a) is the longitudinal wave velocity model and (b) is the transverse wave velocity model.

[0019] Figure 8 The velocity model is obtained by RNN inversion;

[0020] Wherein, (a) is the longitudinal wave velocity model and (b) is the transverse wave velocity model.

[0021] Figure 9 The results of the longitudinal wave knowledge representation and the residuals with the real data;

[0022] (a) represents the result of P-wave knowledge representation, (b) represents the actual P-wave record, and (c) represents the residual between the knowledge representation result and the actual wave field.

[0023] Figure 10 The results of the knowledge representation of transverse waves and the residuals with the actual data;

[0024] (a) represents the result of S-wave knowledge representation, (b) represents the actual S-wave record, and (c) represents the residual between the knowledge representation result and the actual wave field.

[0025] Figure 11 The results of longitudinal wave separation and residuals;

[0026] Wherein, (a) is the predicted P-wave result, (b) is the actual P-wave, and (c) is the residual between the predicted P-wave and the actual P-wave.

[0027] Figure 12 The results of shear wave separation and residuals;

[0028] Wherein, (a) is the predicted S-wave result, (b) is the actual S-wave, and (c) is the residual between the predicted S-wave and the actual S-wave.

[0029] Figure 13 To reconstruct the mixed wave field results;

[0030] Wherein, (a) is the predicted mixed wave field result, (b) is the actual mixed wave field, and (c) is the residual between the predicted mixed wave field and the actual mixed wave field.

[0031] Figure 14 For Model-B complex models;

[0032] Wherein, (a) is the longitudinal wave velocity model and (b) is the transverse wave velocity model.

[0033] Figure 15 Knowledge representation of longitudinal and transverse waves with salt-and-pepper noise;

[0034] Among them, (a) is the knowledge representation of P-wave with salt-and-pepper noise, (b) is the knowledge representation of S-wave with salt-and-pepper noise, and (c) is the knowledge representation of mixed wave field with salt-and-pepper noise.

[0035] Figure 16The result is for longitudinal wave separation;

[0036] Wherein, (a) is the predicted P-wave result, (b) is the actual P-wave, and (c) is the residual between the predicted P-wave and the actual P-wave.

[0037] Figure 17 The result is for shear wave separation;

[0038] Wherein, (a) is the predicted S-wave result, (b) is the actual S-wave, and (c) is the residual between the predicted S-wave and the actual S-wave.

[0039] Figure 18 This is the U-Net network structure. Detailed Implementation

[0040] To facilitate understanding of the technical content of this invention by those skilled in the art, the following technical terms will be explained first:

[0041] 1. Self-encoder

[0042] Early research on autoencoders addressed the "encoder problem," or dimensionality reduction challenge, in neural network representation learning. In 1985, Ackley, Hinton, and Sejnowski first attempted autoencoder algorithms on Boltzmann machines. In 1986, autoencoders became one of the implementations of the backpropagation algorithm, known as "self-supervised backpropagation." In 1987, LeCun constructed an autoencoder neural network using a multilayer perceptron (MLP) and applied it to data denoising. Around the same time, Borald and Kamp's MLP autoencoder research also garnered significant attention. In 1994, Hinton and Zemel proposed the "minimum description length principle," constructing the first generative model based on an autoencoder. These studies laid the foundation for the development of autoencoders and expanded their application areas in neural networks. Figure 1 As shown, an autoencoder consists of two main components: an encoder and a decoder. The encoder is responsible for encoding the input, while the decoder uses the encoding to reconstruct the input. Although the simplest autoencoder can achieve this function by simply repeating the original signal, autoencoders are typically designed to approximately reconstruct the input signal, containing only the most relevant parts of the original signal.

[0043] The most basic form of an autoencoder is a feedforward, non-recurrent neural network, similar to a single-layer perceptron in a multilayer perceptron. It contains one or more hidden layers connecting the input and output. The output layer has the same number of nodes as the input layer, and its goal is to minimize the difference between the input and output, thereby reconstructing the input. It's important to note that autoencoders differ from supervised learning; their goal is to reconstruct the input rather than predict a target value given the input, thus classifying them as unsupervised learning.

[0044] 2. Full waveform inversion

[0045] In traditional seismic exploration, exploration teams generate seismic waves using active source methods (such as blasting or air gun sources), and then receive the reflected and refracted signals of these waves using a set of geophones. Traditional seismic migration methods primarily involve a series of processing steps, including normal time difference correction, removal of multiples and background noise, and migration correction, to ultimately obtain the layered velocity structure of the subsurface medium. This method, by processing the original artillery record map, obtains a general image of the subsurface medium distribution, thereby determining the location and depth of the reservoir and improving the exploration success rate.

[0046] With the increasing complexity of mineral resource exploration, the demand for subsurface structure information is becoming increasingly urgent. Full waveform inversion, a technique based on numerical methods to simulate wavefield propagation in different velocity models, continuously adjusts the velocity model by comparing the errors between synthetic and observed seismic data, performing wavefield simulations until the residual energy error between the synthetic and observed data is within 5%. Compared to traditional seismic migration, full waveform inversion more directly generates synthetic seismic maps without relying on manual processing steps, thus providing more accurate subsurface structure information. This method demonstrates superior performance in situations with significant geological structural variations and high complexity.

[0047] The proposed P-wave and S-wave separation method consists of two stages. The first stage is knowledge extraction and representation, and the second stage is knowledge-guided autoencoder network to achieve P-wave and S-wave separation, such as... Figure 2 As shown, the solution of the present invention includes the following implementation process:

[0048] 1. Representation of knowledge about longitudinal and transverse waves

[0049] The wave field received by the detector at each instant satisfies the wave equation. In an isotropic medium, the longitudinal wave is a divergence-free field, and the transverse wave is an irrotational field. Based on Helmholtz theory, Jianlei et al. proposed the equivalent first-order velocity-stress equation for elastic waves with separated longitudinal and transverse waves:

[0050]

[0051] Where v x and vz These are the velocity components of the particle along the x and z directions, respectively; v xp and v xs These are the P-wave velocity components and S-wave velocity components in the x-direction, respectively; v zp and v zs These are the P-wave velocity components and S-wave velocity components in the z-direction, respectively; τ xx and τ zz These are the normal stresses along the x and z directions, respectively; τ xz It is shear stress; λ and μ are Lamé coefficients; ρ is density; V p and V s These are the longitudinal wave velocity and the transverse wave velocity, respectively; v p =v xp +v zp For an irrotational pure P-wave field, v s =v xs +v zs For a divergence-free pure S-wave field; s is the excitation source acting on the stress, usually a Ricker wavelet, whose expression is as follows:

[0052]

[0053] Among them, f m The dominant frequency of the seismic source is represented. Equation (1) is numerically simulated using a finite difference method, with the initial condition in the wave equation being 0. The iterative process of the finite difference simulation depends on the current input and the state of the wave equation at the previous time. The physical process of wave equation propagation can be simulated using RNN (Recursive Neural Network) units. At a given time t, such as... Figure 3 As shown, the RNN unit stores parameters such as the current wave field, P-wave and S-wave velocity models, density model, normal stress, and shear stress.

[0054]

[0055] Where T represents the transpose operation of the matrix, given an accurate P-wave and S-wave velocity model V p V s And when using the density model ρ, the RNN element corresponds to the current source s in the sequence. t and the RNN unit state h from the previous time step t-1 Perform the operation and output the current time-current mixed wave field record U. t P-wave field record U pt S-wave field record U st And an updated RNN unit state h t The entire wave equation update process can be described as follows:

[0056]

[0057] A describes the elastic wave equation with respect to the RNN element storage variable h in the absence of a seismic source. t-1 Update; P (i) and P (o) This indicates a linear operation that samples a specific row or column of a variable. For example... Figure 4 As shown, the forward modeling of the elastic wave equation (1) is achieved using an RNN. Figure 4 As shown, wavefield snapshots are obtained at each moment, including P-waves, S-waves, and mixed wavefields. By setting the position of the detector, the seismic record corresponding to the current moment can be received, realizing the extraction and representation of knowledge about P-waves and S-waves.

[0058] 2. Full waveform inversion via RNN

[0059] The basic idea of ​​full waveform inversion is to compare the differences between the actual observed seismic record U and the seismic data calculated by numerical simulation. To invert the physical properties of the subsurface medium, a more accurate velocity model V is obtained from the observed mixed wavefield observation record U. p and V s The density model ρ is a key step in representing P-wave and S-wave knowledge. In the RNN network, the velocity and density parameters in the initial unit state h0 of the RNN are set as trainable parameters and initialized as follows:

[0060]

[0061] By using an RNN to perform forward modeling of the elastic wave equation, we can obtain the equation for the recording duration [0, t]. max Within the range, the RNN cell storage state h at each time step t ,t∈[0,t max A set of RNN state records H and the generated observation records are composed of […].

[0062]

[0063] t max This represents the recording duration, and its value can be set according to the actual application. In the following two models in this embodiment, the values ​​are 1280 and 3800, respectively.

[0064] The problem is transformed into updating the state h of the RNN cells via an RNN. t The velocity model V in s V p And the density model parameter ρ, which makes the predicted mixed observation record To approximate the original observation record U as closely as possible, it can be described as follows:

[0065]

[0066] Here, ||·|| represents the loss function used by the RNN to achieve full waveform inversion. Through multiple forward simulations, objective function calculations, and model parameter adjustments, the P-wave and S-wave velocity models and density models are gradually optimized. The RNN process for achieving full waveform inversion is as follows: Figure 5 As shown. The entire optimization process satisfies the physical process of elastic wave propagation, that is, the constraint condition of the full waveform inversion optimization problem is equation (1). In other words, full waveform inversion can be correctly understood as a special case of knowledge-guided RNN training. The RNN network minimizes the generated observation records. The difference between V and U yields a more accurate velocity model. P and V s And density model ρ.

[0067] 3. Knowledge-guided autoencoder networks

[0068] First, RNNs are used to perform full waveform inversion of elastic waves to extract knowledge and obtain knowledge representations of complex P-wave and S-wave structures; second, such as Figure 6 As shown, using the knowledge representation results of complexly constructed longitudinal and transverse waves, a two-branch autoencoder network is constructed, namely network P and network S, whose expressions are as follows:

[0069]

[0070] in, and These are the predicted wavefield records for the P-wave and S-wave, respectively. and Let θ represent the nonlinear mapping functions implemented by neural networks P and S, respectively. P and θ S Let P and S represent the trainable parameters of neural networks P and S, respectively. As shown in Figure 18, each branch uses a U-Net network structure, and network P outputs the predicted P-wave field record. Prediction results of S-wave field records output by network S Among them, network P represents the longitudinal wave knowledge representation U. P Under the guidance of [unclear], longitudinal wave separation is achieved, and network S represents transverse wave knowledge U. s The separation of transverse waves is achieved under the guidance of [the relevant technology / mechanism].

[0071] Finally, the wavefield reconstruction error L is constructed. rec Longitudinal wave knowledge guides error L p And the knowledge of transverse waves guides the error L sAn autoencoder architecture with comprehensive constraints. The specific expression for the overall loss function formed by the comprehensive constraints is as follows:

[0072]

[0073] Where λ and μ are empirical parameters, both of which are taken as 0.5 in this embodiment, and ||·||1 represents the 1-norm. U p =[U pt ] T ,t∈[0,t max ], U s =[U st ] T ,t∈[0,t max ].

[0074] The training stops after 4000 iterations, or when the total loss function stops decreasing and begins to oscillate.

[0075] Model-A used for velocity inversion is a partial Marmousi-ii model. Model-A includes velocity models for both P-waves and S-waves, such as... Figure 7 As shown. Model-A has a grid size of 100×100 and a spatial sampling interval of 10 meters. Forward modeling of the elastic wave equations is performed using an RNN network, employing a Ricker wavelet with a dominant frequency of 40Hz as the source points, spaced horizontally at 10-meter intervals (a total of 10 source points). The seismic detector array is located at a horizontal distance of 0.5 kilometers. The forward temporal simulation includes 1280 time sampling points with a time sampling interval of 0.5 milliseconds, covering a time range from 0 to 0.64 seconds.

[0076] like Figure 8 As shown, an approximate estimate of the real velocity model was obtained using an RNN network. Although the information on both sides of the model was inaccurate, the detailed information in the middle part of the model was recovered.

[0077] like Figure 9 , Figure 10 As shown, there is a structural correlation between the results of RNN knowledge representation of longitudinal and transverse waves and the actual longitudinal and transverse wave records. At the same time, the knowledge representation satisfies the wave propagation process and incorporates the physical mechanism of wave propagation.

[0078] The results of the knowledge representation of P and S waves are used to guide an autoencoder network to achieve P and S wave separation results, such as... Figure 11 and Figure 12 As shown, the loss of effective energy is effectively reduced in the separation results. This verifies the effectiveness of the proposed method for separating P-waves and S-waves under complex structural models.

[0079] The result of the autoencoder network reconstruction is as follows: Figure 13As shown, the residual energy is very small, verifying the fidelity of signal separation.

[0080] Model-B is a larger, more complex version of the Marmousi-ii model, such as... Figure 14 As shown. This model is used for robustness testing, aiming to evaluate the proposed method in more complex knowledge representations and such... Figure 15 The performance is shown under a noisy knowledge representation. An RNN was applied to the forward modeling of the elastic wave equations with a grid size of 480 × 1600 and a spatial sampling interval of 5 meters. The epicenter was located at a depth of 5.0 km below the surface, using a 40 Hz Ricker wavelet. The seismic detector array was located at a horizontal distance of 5.8 km. The simulation included 3800 time samples at a time sampling interval of 0.5 ms, covering a time range from 0.3 to 1.9 seconds.

[0081] Figure 7 , 8 In 14, Depth represents depth, Distance represents distance, and Velocity represents speed.

[0082] The representation results with knowledge of salt-and-pepper noise are as follows: Figure 15 As shown, due to the influence of noise, the reflected signals and weaker signals in the recording have been covered by noise, which poses a certain challenge to separation. Figure 16 and Figure 17 The results are shown below, with the separation guided by noisy knowledge. Although there is some loss of weak signals, the separation results are generally effective and close to the actual results, which proves that the method has a certain degree of robustness.

[0083] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A knowledge-guided method for separating P-waves and S-waves in complex geological structures, characterized in that, First, knowledge extraction is achieved by using RNN to perform full waveform inversion of elastic waves, obtaining knowledge representations of complex-structured P-waves and S-waves. Second, using the knowledge representations of complex-structured P-waves and S-waves, a dual-branch autoencoder network is constructed, in which one branch achieves P-wave separation under the guidance of P-wave knowledge representation, and the other branch achieves S-wave separation under the guidance of S-wave knowledge representation. The process of using RNN to perform full waveform inversion of elastic waves to extract knowledge is as follows: Record duration Within the range, that is , Represents the time step, where each RNN cell stores its state. A set of RNN state records and the generated observation records : ； in, The superscript T indicates transpose. and yes The velocity components of the particle along the x and z directions at time points; and These are the P-wave velocity component and the S-wave velocity component in the x-direction, respectively. and These are the P-wave velocity component and the S-wave velocity component in the z-direction, respectively. and These are the normal stresses along the x and z directions, respectively; It is shear stress; and It is the Lamé coefficient; It is density; and These are the longitudinal wave velocity and the transverse wave velocity, respectively. By updating the RNN unit state In , and This makes the predicted mixed observation records To approximate the original observation records as closely as possible. The description is as follows: ； According to the latest update , and The RNN unit is for the current earthquake source. and the RNN unit state at the previous time step Perform the operation and output the mixed wave field record at the current time. P-wave field record S-wave field record And an updated RNN unit state .

2. The knowledge-guided method for separating P-waves and S-waves in complex geological structures according to claim 1, characterized in that, The obtained knowledge representations of complex P-wave and S-wave structures include P-wave knowledge representations. Transverse wave knowledge representation , , .

3. The knowledge-guided method for separating P-waves and S-waves in complex geological structures according to claim 2, characterized in that, Both branches of the autoencoder in the dual-branch autoencoder network adopt the U-Net network structure.

4. The knowledge-guided method for separating P-waves and S-waves in complex geological structures according to claim 3, characterized in that, The loss function used to train the dual-branch autoencoder network is: ； in, and These are empirical parameters. The 1-norm is represented by and These are the predicted P-wave field records and the predicted S-wave field records, respectively.

Images