Model training method, VSP wave field separation method and device and medium

Abstract

The invention provides a model training method, a VSP wave field separation method, a VSP wave field separation device and a medium, and relates to the field of oil-gas exploration. The method comprises the following steps: obtaining VSP sample data, constructing a first depth sparse self-encoding network, obtaining gradient features presented by different wave fields represented by the VSP sample data, carrying out wave field separation on the VSP sample data according to the gradient features to obtain different wave fields corresponding to a second depth sparse self-encoding network and the VSP sample data, and obtaining reconstructed VSP data according to the different wave fields. And calculating reconstruction loss, gradient loss and sparse penalty term loss, and further inputting the loss into a second deep sparse self-encoding network for iterative updating to obtain a trained deep sparse self-encoding network. According to the method, gradient features presented by different wave fields are utilized, a deep sparse self-encoding network is taken as a basis, the deep learning network which is suitable for VSP wave field separation and is high in practicability is obtained by adopting unsupervised model training, and high-precision VSP wave field separation is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of oil and gas exploration, and in particular to a model training method, a VSP wave field separation method, a device and a medium. BACKGROUND

[0002] With the continuous development of seismic exploration technology, the vertical seismic profile (VSP) method is produced. The vertical seismic profile method is to observe the seismic wave field in the well, place the geophone in the well at different depths to record the seismic signals generated by the ground source, obtain the vertical seismic profile, and perform VSP data, which is mainly used for detecting the geological structure and stratum characteristics near the well. Since the VSP data contains various wave field information, and specific wave field information usually has specific geological significance, high-precision wave field separation is beneficial to better understanding of the attribute information of the underground medium, which is of great significance for oil and gas reservoir prediction and oil and gas exploration.

[0003] In recent years, deep learning (DL) has developed rapidly, and the industry has tried to separate VSP wave fields based on deep learning. For example, Cao et al. based on the idea of generative adversarial network, used asymmetric convolutional blocks to separate the uplink and downlink waves of VSP data; Wei et al. proposed a wave field separation method for applying conditional generative adversarial network to directly reconstruct the wave field of the P-wave from the full elastic wave field VSP data. However, the implementation of these methods requires a large amount of labeled data, and only good results are obtained on model data. When applied to the wave field separation of actual VSP data, it is difficult to obtain good results. In order to avoid the difficulty of separating the wave field of actual VSP data by the deep learning model trained by the synthetic data label, Tao et al. proposed a method based on convolutional auto encoder (CAE), which artificially selected reasonable samples through frequency wave number domain (F-k) filtering to construct a training set. However, the processing procedure of this method is complex, and the training and adaptability of the model performance are constrained by the degree of artificial selection of reasonable samples.

[0004] Therefore, there is an urgent need to provide a practical VSP wave field separation scheme. SUMMARY

[0005] The embodiments of the present application provide a model training method, a VSP wave field separation method, a device and a medium, to solve the problem that it is difficult to obtain good results when separating the VSP wave field by using related technologies.

[0006] In a first aspect, the embodiments of the present application provide a model training method, comprising:

[0007] obtaining VSP sample data;

[0008] construct a first deep sparse auto-encoder network based on the VSP sample data;

[0009] obtain gradient features presented by different wave fields in a VSP wave field represented by the VSP sample data;

[0010] perform wave field separation on the VSP sample data based on the first deep sparse auto-encoder network according to the gradient features presented by different wave fields, to obtain a second deep sparse auto-encoder network and different wave fields corresponding to the VSP sample data;

[0011] obtain reconstructed VSP data according to the different wave fields corresponding to the VSP sample data, and calculate a reconstruction loss, a gradient loss and a sparse penalty term loss of the reconstructed VSP data relative to the VSP sample data;

[0012] input the reconstruction loss, the gradient loss and the sparse penalty term loss into the second deep sparse auto-encoder network, and iteratively update the second deep sparse auto-encoder network to obtain a deep sparse auto-encoder network.

[0013] In a possible implementation, the first deep sparse auto-encoder network includes sparse auto-encoders and decoders corresponding to different wave fields, and the wave field separation on the VSP sample data based on the first deep sparse auto-encoder network according to the gradient features presented by different wave fields to obtain the second deep sparse auto-encoder network and the different wave fields corresponding to the VSP sample data includes: for a sparse auto-encoder and a decoder of a target wave field in the different wave fields, inputting the VSP sample data into the sparse auto-encoder to obtain sample low-dimensional features, inputting the sample low-dimensional features into the decoder, and decoding the sample low-dimensional features under the gradient features presented by the target wave field to obtain the target wave field corresponding to the second deep sparse auto-encoder network and the VSP sample data.

[0014] In a possible implementation, the gradient features presented by different wave fields in a VSP wave field represented by the VSP sample data are obtained by: for each wave field in the VSP wave field represented by the VSP sample data, obtaining a phase axis extension angle of the wave field in an extension direction of a seismic phase axis; and obtaining the gradient features presented by the wave field according to the phase axis extension angle.

[0015] In a possible implementation, the gradient features presented by the wave field are obtained according to the phase axis extension angle by: determining a convolution kernel based on the phase axis extension angle; and performing a convolution operation on the phase axis extension angle and the convolution kernel to obtain the gradient features presented by the wave field.

[0016] In a possible implementation, the reconstruction loss, the gradient loss and the sparse penalty term loss are input into the second deep sparse auto-encoding network, the second deep sparse auto-encoding network is iteratively updated, and a deep sparse auto-encoding network is obtained, including: determining a total loss according to the reconstruction loss, the gradient loss and the sparse penalty term loss; iteratively updating the second deep sparse auto-encoding network according to the total loss; wherein the total loss satisfies the following formula:

[0017] Loss = Loss recon + λ1Loss grad + λ2Loss spar

[0018] In the above formula, Loss represents the total loss, Loss recon represents the reconstruction loss, Loss grad represents the gradient loss, Loss spar represents the sparse penalty term loss, and λ1 and λ2 are weights of the corresponding terms.

[0019] In a possible implementation, before the VSP sample data is separated based on the gradient features presented by different wave fields according to the first deep sparse auto-encoding network, the method further includes: preprocessing the VSP sample data to obtain preprocessed data, and the preprocessing includes cubic polynomial interpolation and resampling.

[0020] In a possible implementation, the model training method further includes: adding noise to the VSP sample data to obtain noise-added data, and the noise includes Gaussian noise and / or uniform noise.

[0021] In a possible implementation, the preprocessed data, the noise-added data and the gradient features presented by different wave fields are input into the second deep sparse auto-encoding network, the second deep sparse auto-encoding network is iteratively updated, and a deep sparse auto-encoding network is obtained.

[0022] In a second aspect, the present application provides a VSP wave field separation method, including:

[0023] Obtaining VSP data of a target well;

[0024] Obtaining gradient features presented by different wave fields in a VSP wave field represented by the VSP data, and preprocessing the VSP data to obtain preprocessed VSP data, and the preprocessing includes cubic polynomial interpolation and resampling;

[0025] Inputting the preprocessed VSP data into a deep sparse auto-encoding network for wave field separation, separating the preprocessed VSP data based on the gradient features presented by different wave fields to obtain different wave fields corresponding to the target well, and the deep sparse auto-encoding network is obtained by the model training method of any one of the first aspect.

[0026] In a third aspect, the present application provides a model training device, comprising:

[0027] a first obtaining module, configured to obtain VSP sample data;

[0028] a constructing module, configured to construct a first deep sparse auto-encoding network based on the VSP sample data;

[0029] a second obtaining module, configured to obtain gradient features presented by different wave fields in a VSP wave field represented by the VSP sample data;

[0030] a separating module, configured to perform wave field separation on the VSP sample data based on the first deep sparse auto-encoding network according to the gradient features presented by the different wave fields, to obtain a second deep sparse auto-encoding network and different wave fields corresponding to the VSP sample data;

[0031] a processing module, configured to obtain reconstructed VSP data according to the different wave fields corresponding to the VSP sample data, and to calculate a reconstruction loss, a gradient loss and a sparse penalty term loss of the reconstructed VSP data relative to the VSP sample data;

[0032] an adjusting module, configured to input the reconstruction loss, the gradient loss and the sparse penalty term loss into the second deep sparse auto-encoding network, to iteratively update the second deep sparse auto-encoding network, and to obtain a deep sparse auto-encoding network.

[0033] In a fourth aspect, the present application provides a VSP wave field separation device, comprising:

[0034] an obtaining module, configured to obtain VSP data of a target well;

[0035] a preprocessing module, configured to obtain gradient features presented by different wave fields in a VSP wave field represented by the VSP data, and to preprocess the VSP data to obtain preprocessed VSP data, the preprocessing including cubic polynomial interpolation and resampling;

[0036] a separating module, configured to input the preprocessed VSP data into a deep sparse auto-encoding network for wave field separation, to perform wave field separation on the preprocessed VSP data based on the gradient features presented by the different wave fields, and to obtain different wave fields corresponding to the target well, the deep sparse auto-encoding network being obtained through the model training method of any one of the first aspect.

[0037] In a fifth aspect, the present application provides an electronic device, comprising a processor and a memory in communication with the processor;

[0038] the memory, configured to store computer execution instructions;

[0039] A processor is configured to execute computer-executable instructions stored in a memory to implement the method of the first aspect and / or the second aspect.

[0040] In a sixth aspect, the present application provides a computer-readable storage medium, which stores computer-executable instructions, and the computer-executable instructions are configured to implement the method of the first aspect and / or the second aspect when executed.

[0041] In a seventh aspect, the present application provides a computer program product, which comprises a computer program, and the computer program is configured to implement the method of the first aspect and / or the second aspect when executed.

[0042] The model training method, the VSP wave field separation method, the device and the medium provided by the embodiments of the present application are as follows: VSP sample data is obtained, a first deep sparse auto-encoding network is constructed based on the VSP sample data, and gradient features presented by different wave fields in the VSP wave field represented by the VSP sample data are obtained. According to the gradient features presented by different wave fields, the VSP sample data is separated based on the first deep sparse auto-encoding network to obtain a second deep sparse auto-encoding network and different wave fields corresponding to the VSP sample data. Then, according to the different wave fields corresponding to the VSP sample data, reconstructed VSP data is obtained, and a reconstruction loss, a gradient loss and a sparse penalty term loss of the reconstructed VSP data relative to the VSP sample data are calculated. Further, the reconstruction loss, the gradient loss and the sparse penalty term loss are input into the second deep sparse auto-encoding network, and the second deep sparse auto-encoding network is iteratively updated to obtain a trained deep sparse auto-encoding network. In this process, by taking the gradient features presented by different wave fields in the VSP wave field as a constraint and taking the deep sparse auto-encoding network as a basis, a deep learning network suitable for VSP wave field separation and having strong practicability is trained in an unsupervised mode without manual labeling, so as to improve the wave field separation effect when the deep sparse auto-encoding network is applied to actual VSP data, and to realize high-precision separation of the VSP wave field, thereby laying a solid foundation for imaging and corridor stacking of the VSP data. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings, which are incorporated herein and form a part of the specification, illustrate embodiments consistent with the present application and, together with the description, further serve to explain the principles of the application.

[0044] Figure 1 An application scenario diagram of the model training method provided by the exemplary embodiments of the present application is shown in FIG. 1.

[0045] Figure 2 A flowchart of the model training method provided by the exemplary embodiments of the present application is shown in FIG. 2.

[0046] Figure 3A schematic diagram of the difference of the extension direction slope of seismic events in VSP upgoing and downgoing wave field is provided for the exemplary embodiments of the present application.

[0047] Figure 4 A schematic diagram of the structure of deep sparse auto-encoding network is provided for the exemplary embodiments of the present application.

[0048] Figure 5 A schematic diagram of the flow of VSP wave field separation method is provided for the exemplary embodiments of the present application.

[0049] Figure 6 A schematic diagram of VSP data of target well is provided for the exemplary embodiments of the present application.

[0050] Figure 7 A schematic diagram of upgoing wave field of target well is provided for the exemplary embodiments of the present application.

[0051] Figure 8 A schematic diagram of downgoing wave field of target well is provided for the exemplary embodiments of the present application.

[0052] Figure 9 Another schematic diagram of the flow of VSP wave field separation method is provided for the exemplary embodiments of the present application.

[0053] Figure 10 A schematic diagram of the structure of model training device is provided for the exemplary embodiments of the present application.

[0054] Figure 11 A schematic diagram of the structure of VSP wave field separation device is provided for the exemplary embodiments of the present application.

[0055] Figure 12 A schematic diagram of the structure of electronic device is provided for the exemplary embodiments of the present application.

[0056] The specific embodiments of the present application have been shown by way of example in the above figures, and will be described in more detail hereafter. These figures and this written description are not intended to limit the scope of the inventive concept in any way, but rather to illustrate the inventive concept by reference to specific embodiments. DETAILED DESCRIPTION

[0057] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The same reference numbers in different drawings represent the same or similar elements. The following detailed description is not intended to limit the scope of the present application, but rather to explain the embodiments of the present application as set out in the claims. Other embodiments will be apparent to those of ordinary skill in the art from this disclosure.

[0058] The terms “first,” “second,” etc., used in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, products, or apparatus.

[0059] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0060] Currently, numerous experts and scholars have conducted research on wavefield separation techniques. Existing research methods mainly utilize wavefield characteristics such as apparent velocity and polarization in seismic wave kinematics and dynamics. These methods include median filtering, Radon transform, FK filtering, variable apparent velocity wavefield separation, polarization filtering, wavelet analysis transform, and wave equation methods. Among these, the median filtering method sorts the data within each analysis window and selects the value at the middle position as the output to achieve wavefield stability analysis in a certain direction. However, median filtering requires good consistency between wavefields at different depths. Radon transform and FK filtering both require transforming the time (T)-depth (D) domain data to other domains and then using the apparent velocity differences between different wavefields to construct suitable filters to achieve separation between different wavefields. However, these methods require uniform sampling rate of the data to be processed and significant wavefield characteristics. Wave equation-based methods, based on elastic wave methods, use the vertical (horizontal) component of the wave field to correct the horizontal (vertical) component, thereby eliminating the shear (longitudinal) wave component in the vertical (horizontal) component and realizing the analysis of longitudinal and shear waves. However, the drawback of this type of method is that it requires the propagation speed of longitudinal and shear waves in the strata and has high requirements for the signal-to-noise ratio of the data. Polarization filtering starts from the dynamic characteristics of the wave and uses the difference in polarization of longitudinal and shear waves to separate longitudinal and shear waves. However, this type of method requires good matching of signal characteristics and still requires knowledge of the longitudinal and shear wave velocities in each stratum.

[0061] In summary, VSP wavefield separation methods based on physical mechanisms are limited by the assumptions underlying the methods used, making it difficult to meet the VSP wavefield separation requirements under different conditions.

[0062] Furthermore, in deep learning-based techniques for VSP wavefield separation, the label data used to train the deep learning model is mostly synthetic data. The difference in data distribution characteristics between synthetic data and actual VSP data limits the practical application of deep learning networks (models) trained on labeled data, making it difficult for them to achieve good results when applied to real VSP data. To avoid the challenge of separating the wavefield of actual VSP data using deep learning models trained on synthetic data labels, Tao et al. proposed a method based on a convolutional autoencoder (CAE). This method uses frequency-wavenumber domain (Fk) filtering to manually select appropriate samples to construct the training set. However, this method has a complex processing flow, and the training and adaptability of the model are constrained by the appropriateness of the manual selection.

[0063] Based on the above, this application provides a VSP wavefield separation scheme, which uses a deep sparse autoencoder network constrained by gradient features for VSP wavefield separation. It is highly practical and can separate VSP wavefields with high accuracy.

[0064] Figure 1 A schematic diagram illustrating an application scenario of the model training method provided as an exemplary embodiment of this application. For example... Figure 1 As shown, the application scenario includes a first client 11, a server 12, and a second client 13, wherein there can be at least one first client 11 and at least one second client 13. In practical applications, for example, after researchers or other relevant personnel obtain VSP data, they process the VSP data through the first client 11 to obtain VSP sample data and send it to the server 12 for storage, or they generate VSP sample data through the first client 11 and send it to the server 12 for storage; when the server 12 detects a model training instruction issued by the user through the second client 13, it executes the model training method provided in this application based on the stored VSP sample data to obtain a trained deep sparse autoencoder network, and sends the result of the deep sparse autoencoder network training to the second client 13, thereby allowing relevant personnel to know that the deep sparse autoencoder network has been trained.

[0065] It should be noted that server 12 can also be replaced by a server cluster or other computing devices with a certain computing power. Both the first client 11 and the second client 13 can be mobile phones, computers, laptops, or personal digital assistants (PDAs).

[0066] The following is combined Figure 1 Application scenarios, refer to Figure 2 This document describes a model training method according to exemplary embodiments of this application. It should be noted that the above application scenarios are shown only to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited to those described herein. Figure 1 The limitations of the application scenarios shown.

[0067] Figure 2 This is a schematic flowchart of a model training method provided for an exemplary embodiment of this application. Figure 2 As shown, the model training method in this application embodiment includes the following steps:

[0068] S201. Obtain VSP sample data.

[0069] For example, such as Figure 1 As shown, the first client 11 obtains VSP data to be separated from relevant geological exploration departments or relevant geological information platforms, processes the VSP data to obtain VSP sample data, and sends it to the server 12; correspondingly, the server 12 receives the VSP sample data sent by the first client 11.

[0070] Optionally, the processing of VSP data may include noise reduction and other interference filtering processes, and / or data enhancement processes such as interpolation, sampling, and data fine-tuning, etc.

[0071] S202. Construct the first deep sparse autoencoder network based on VSP sample data.

[0072] For example, a first deep sparse autoencoder network is constructed based on VSP sample data according to preset rules. It can be understood that the process of constructing the first deep sparse autoencoder network is the same as the process of constructing the initial deep sparse autoencoder network in related technologies, and will not be elaborated here.

[0073] S203. Obtain the gradient characteristics of different wavefields in the VSP wavefield represented by the VSP sample data.

[0074] For example, the VSP wavefield represented by VSP sample data can include both ascending and descending wavefields. Since the seismic phase axes in the ascending and descending wavefields have different directions of extension in the VSP data, in the time (T)-depth (D) domain (TD domain), the ascending and descending waves can be distinguished by the polarity of the slope of the seismic phase axis; that is, the continuous direction of the ascending wave exhibits a positive slope, and the continuous direction of the descending wave exhibits a negative slope. For example, Figure 3 A schematic diagram illustrating the difference in the slope of the extension direction of the seismic phase axis in the uplink and downlink wavefields of the VSP, provided as an exemplary embodiment of this application. (See diagram for reference.)Figure 3 As shown, the solid white arrows represent downward waves, and the dashed white arrows represent upward waves. Although there are some differences in direction, the seismic waves formed at the reflection interface have good continuity. Relevant personnel can use the corresponding VSP data processing software to interactively track the phase axes of the upward and downward wave fields that reflect high energy, in order to obtain the extension curves of the phase axes of the high-energy reflected seismic waves in the upward and downward wave fields. Then, the gradient characteristics presented by the extension curves can be determined by calculation.

[0075] S204. Based on the gradient characteristics of different wave fields, the VSP sample data is separated by the first deep sparse autoencoder network to obtain the different wave fields corresponding to the second deep sparse autoencoder network and the VSP sample data.

[0076] Correspondingly, the first deep sparse autoencoder network can be understood as a deep learning network containing a sparse autoencoder. The sparse autoencoder is an unsupervised machine learning algorithm that continuously adjusts its parameters by calculating the error between the autoencoder's output and the original input.

[0077] This step also inputs the gradient features of different wavefields as constraints into the first deep sparse autoencoder network. The first deep sparse autoencoder network performs wavefield separation on the VSP sample data based on the gradient features of different wavefields, thereby obtaining the second deep sparse autoencoder network after wavefield separation and the different wavefields corresponding to the VSP sample data.

[0078] Considering that different geological target layers have different requirements for VSP data, resulting in differences in the depth interval of detectors in the same well and the data sampling interval in different wells, this application embodiment addresses this problem by performing cubic polynomial interpolation and resampling on the VSP sample data. This allows the processed VSP data to have the same detector depth interval and time sampling interval, making it easier for the aforementioned depth dilution autoencoder network to adapt to wavefield separation of VSP data from different wells.

[0079] Therefore, in some embodiments, before performing wavefield separation on the VSP sample data based on the gradient characteristics of different wavefields, the method further includes: preprocessing the VSP sample data to obtain preprocessed data, wherein the preprocessing includes cubic polynomial interpolation and resampling.

[0080] Cubic polynomial interpolation is a numerical analysis method used to construct a cubic polynomial function from known data points to interpolate between these points. Through cubic polynomial interpolation, a curve can be accurately fitted at given data points, thereby inferring the numerical values ​​between the data points. This interpolation method is commonly used for data smoothing and curve fitting. Resampling, in statistics and data analysis, refers to resampling existing data to change the sampling frequency or time interval, or to change the number of data points, thereby adapting to specific analytical needs. Resampling can include both upsampling (adding data points) and downsampling (reducing data points), aiming to adjust the granularity of the data for better analysis and processing.

[0081] S205. Based on the different wavefields corresponding to the VSP sample data, obtain the reconstructed VSP data, and calculate the reconstruction loss, gradient loss, and sparsity penalty term loss of the reconstructed VSP data relative to the VSP sample data.

[0082] Optionally, obtaining reconstructed VSP data based on different wavefields corresponding to the VSP sample data may include: accumulating different wavefields corresponding to the VSP sample data to obtain reconstructed VSP data.

[0083] For example, taking the VSP wavefield represented by VSP sample data as an example, which includes the up-going wavefield and the down-going wavefield, the up-going wavefield corresponding to the VSP sample data is added to the down-going wavefield corresponding to the VSP sample data to obtain the reconstructed VSP data.

[0084] It should be noted that the "adding the uplink wave field corresponding to the VSP sample data to the downlink wave field corresponding to the VSP sample data" described here is only one specific implementation of reconstructing VSP data. In practical applications, it can be set according to actual needs. In other words, reconstructing VSP data can also be done through other operations besides addition.

[0085] Furthermore, considering that the problem to be solved in this application can be described as a constrained optimization problem as follows:

[0086]

[0087] Where x represents VSP sample data, U represents the uplink field output by the uplink field decoder, D represents the downlink field output by the downlink field decoder, and ||.|| measures the difference between x and U and D. To optimize ψ u and ψ d The regularization operator for the component, ψ u For the components of the ascending wave field, ψ d R represents the components of the downlink wave field. T×N For a deep neural network function, θu θ represents the parameters of the up-wave field encoder and decoder. d The parameters represent the downlink wave field encoder and decoder, where T is the number of sampling points in the time domain and N is the number of detectors in the well.

[0088] Based on the above theory, the final wave field is shown below:

[0089]

[0090] in, These are the parameters of the uplink field decoder and encoder after iteration. These are the parameters of the downlink wavefield decoder and encoder after iteration. For the upward wave field after iteration, This is the downlink wave field after iteration.

[0091] Based on the constrained optimization problem described above and the network structure of the deep sparse autoencoder network, two regularization terms can be introduced as constraints for solving the above equations when constructing the loss function. These two regularization terms are the gradient loss and the sparsity penalty loss. These two regularization terms can fully leverage the strengths of machine learning algorithms and utilize the characteristics of VSP data, enabling the deep sparse autoencoder network to better learn the wavefield features contained in the VSP sample data, thereby enhancing the performance and applicability of the deep sparse autoencoder network and achieving high-precision VSP wavefield separation. Therefore, the loss function in this embodiment can include: reconstruction loss, gradient loss, and sparsity penalty loss.

[0092] Based on VSP sample data and reconstructed VSP data, the reconstruction loss, gradient loss, and sparsity penalty term loss of the reconstructed VSP data relative to the VSP sample data are calculated according to preset rules. Among them, the reconstruction loss is used to measure the difference between the original VSP data and the reconstructed VSP data; the gradient loss is used to enhance the continuity of the wavefield and the separation of the uplink and downlink wavefields; the sparsity penalty term loss is used to limit the average activation of neurons in the hidden layer of the deep sparse autoencoder network, forcing only a portion of neurons to be activated, thereby increasing the robustness of the deep sparse autoencoder network.

[0093] S206. Input the reconstruction loss, gradient loss, and sparse penalty term loss into the second deep sparse autoencoder network, and iteratively update the second deep sparse autoencoder network to obtain the deep sparse autoencoder network.

[0094] In this step, any one of the reconstruction loss, gradient loss, and sparse penalty term loss can be sequentially input into the model parameters of the second deep sparse autoencoder network. For example, the reconstruction loss can be input into the second deep sparse autoencoder network first to update its model parameters. Then, steps S203, S204, S205, and S206 can be executed again for the same input to achieve iterative updates to the second deep sparse autoencoder network. Alternatively, logical operations can be performed on the reconstruction loss, gradient loss, and sparse penalty term loss to obtain a total loss value. The model parameters of the second deep sparse autoencoder network can be updated based on this total loss value, and the second deep sparse autoencoder network can be iteratively updated to achieve the effect of model training, thereby obtaining a trained deep sparse autoencoder network.

[0095] It is understandable that the model training process ends similarly to related technologies. It can be considered that the deep sparse autoencoder network is trained after the set number of training iterations is reached, or the trained deep sparse autoencoder network is obtained when the loss value is less than a set threshold.

[0096] The model training method provided in this application uses the gradient features of different wavefields in the VSP wavefield as constraints, and is based on a deep sparse autoencoder network. Combining the advantages of deep neural networks, it adopts an unsupervised training mode to obtain a deep learning network that is suitable for VSP wavefield separation and has strong practicality. It does not require manual labeling, further enhancing the performance and applicability of the deep learning network, thereby improving the wavefield separation effect when applied to actual VSP data, achieving high-precision VSP wavefield separation, and laying a solid foundation for VSP data imaging and corridor overlay.

[0097] In some embodiments, the first deep sparse autoencoder network includes sparse autoencoders and decoders corresponding to different wavefields. Correspondingly, based on the gradient features exhibited by different wavefields, wavefield separation is performed on the VSP sample data using the first deep sparse autoencoder network to obtain the second deep sparse autoencoder network and the different wavefields corresponding to the VSP sample data. This can include: for the sparse autoencoder and decoder of the target wavefield in different wavefields, inputting the VSP sample data into the sparse autoencoder to obtain low-dimensional features of the samples, inputting the low-dimensional features of the samples into the decoder, and decoding the low-dimensional features of the samples under the constraint of the gradient features exhibited by the target wavefield to obtain the target wavefield corresponding to the second deep sparse autoencoder network and the VSP sample data.

[0098] For example, Figure 4 A schematic diagram of the structure of a deep sparse autoencoder network provided as an exemplary embodiment of this application. For example... Figure 4As shown, the deep sparse autoencoder network includes an uplink wave field encoder, an uplink wave field decoder, a downlink wave field encoder, and a downlink wave field decoder. The uplink wave field encoder and decoder are used to predict the uplink wave field, and the downlink wave field encoder and decoder are used to predict the downlink wave field. All four encoders and decoders are composed of deep convolutional networks, and their network structures can have the same depth. By utilizing convolution, max pooling, and upsampling operations in the deep convolutional network, the prediction of the corresponding wave field is achieved. The goal is to obtain reconstructed VSP data of the uplink and downlink wave fields separated by this network structure, which is approximately the same size as the input data.

[0099] Optionally, since the value range of VSP data includes negative values, the uplink and downlink wave field encoders can use the tanh function as the activation function.

[0100] based on Figure 4 The structure shown involves inputting VSP sample data into an up-wave field encoder, which maps the VSP sample data into a low-dimensional space to extract key features, i.e., low-dimensional features of the samples. These low-dimensional features are then input into an up-wave field decoder, which decodes the low-dimensional features under the gradient feature constraints of the target wave field to obtain the up-wave field corresponding to the VSP sample data. Simultaneously, the VSP sample data is input into a down-wave field encoder, which maps the VSP sample data into a low-dimensional space to extract key features, i.e., low-dimensional features of the samples. These low-dimensional features are then input into a down-wave field decoder, which decodes the low-dimensional features under the gradient feature constraints of the target wave field to obtain the down-wave field corresponding to the VSP sample data.

[0101] Furthermore, the uplink wave field corresponding to the VSP sample data is added to the downlink wave field corresponding to the VSP sample data to obtain the reconstructed VSP data.

[0102] In some embodiments, obtaining the gradient characteristics of different wavefields in the VSP wavefield represented by the VSP sample data as described above may include: for each wavefield in the VSP wavefield represented by the VSP sample data, obtaining the phase axis extension angle of the wavefield in the direction of the seismic phase axis extension; and obtaining the gradient characteristics of the wavefield based on the phase axis extension angle.

[0103] For example, still refer to Figure 3The difference in the slope of the extension direction of the seismic phase axis in the VSP uplink and downlink wave fields is shown. Interpreters calculate the extension angle of the phase axis of the uplink and downlink waves respectively based on the significant extension direction of the seismic phase axis in the uplink and downlink wave fields. For example, phase axis tracking is performed on the uplink and downlink wave fields that reflect strong energy to obtain the extension curve of the seismic phase axis of the strong energy reflection in the uplink and downlink wave fields. Then, the angle of the extension curve with respect to the horizontal axis is determined by calculation.

[0104] In some embodiments, the in-phase axis extension angles of the uplink and downlink wave fields at unpicked positions can be obtained by interpolation, thereby obtaining the in-phase axis extension angles of the uplink and downlink wave fields.

[0105] Furthermore, based on the extension angle of the phase axis of the up-going wave field, the gradient characteristics of the up-going wave field are obtained according to a preset rule, i.e., the gradient operator of the up-going wave field; based on the extension angle of the phase axis of the down-going wave field, the gradient characteristics of the down-going wave field are obtained according to a preset rule, i.e., the gradient operator of the down-going wave field.

[0106] Based on the above embodiments, in some embodiments, obtaining the gradient characteristics of the wave field based on the phase axis extension angle may include: determining the convolution kernel based on the phase axis extension angle; performing a convolution operation on the phase axis extension angle and the convolution kernel to obtain the gradient characteristics of the wave field.

[0107] The convolution kernels determined by different in-phase axis extension angles are different. For example, the Soble operator can be used to determine the convolution kernel based on the in-phase axis extension angle of the uplink and downlink wave fields, thereby constructing gradient features along any in-phase axis extension angle.

[0108] By default, the in-phase axis extension angle of the up-going wave field is denoted as ang_u, and the gradient characteristic of the up-going wave field is denoted as grad_u; the in-phase axis extension angle of the down-going wave field is denoted as ang_d, and the gradient characteristic of the down-going wave field is denoted as grad_d. For example, grad_u and grad_d can be calculated using the following formulas:

[0109]

[0110]

[0111] in, and Here is an example of a convolution kernel.

[0112] The above embodiments provide how to obtain the gradient characteristics of the wave field.

[0113] In some embodiments, the reconstruction loss, gradient loss, and sparse penalty term loss are input into a second deep sparse autoencoder network, and the second deep sparse autoencoder network is iteratively updated to obtain a deep sparse autoencoder network. This includes: determining the total loss based on the reconstruction loss, gradient loss, and sparse penalty term loss; and iteratively updating the second deep sparse autoencoder network based on the total loss. The total loss satisfies the following formula:

[0114] Loss = Loss recon +λ1Loss grad +λ2Loss spar

[0115] In the above formula, Loss represents the total loss. recon Loss indicates reconstruction losses. grad Represents gradient loss, Loss spar Let λ1 and λ2 represent the sparse penalty term loss, and let λ1 and λ2 be the weights of the corresponding terms (i.e., the weights of the regularization term).

[0116] For example, the Euclidean distance metric Loss between the VSP sample data and the reconstructed VSP data is used. recon Correspondingly, Loss recon Satisfy the following formula:

[0117]

[0118] Loss grad Satisfying the formula:

[0119]

[0120] Where ⊙ represents the Hadamard product, which is the result of multiplying corresponding elements of two matrices, and * represents the convolution operator.

[0121] Additionally, Loss spar This can be achieved by adding a sparsity constraint, specifically satisfying the following formula:

[0122]

[0123] Where s is the number of neurons in the hidden layer, α is the output of the hidden layer, and ρ is the desired neuron activation level. It is the calculated average activation level. The KL divergence can be calculated using the following formula:

[0124]

[0125] Correspondingly, during model training, the weight parameters λ1 and λ2 of the regularization term need to be reasonably selected. If they are too large or too small, it will affect the separation effect of the wave field. Then, based on the loss (i.e., the total loss), the model parameters of the deep sparse autoencoder network are adjusted, and iterative training is performed.

[0126] In some embodiments, the model training method further includes adding noise to the VSP sample data to obtain noisy data, such as Gaussian noise and / or uniform noise. In one implementation, the noise includes Gaussian noise; in another implementation, the noise includes uniform noise; and in yet another implementation, the noise includes both Gaussian noise and uniform noise.

[0127] Gaussian noise refers to a type of noise whose probability density function follows a Gaussian distribution (i.e., a normal distribution). Common Gaussian noises include fluctuation noise, cosmic noise, thermal noise, and shot noise. Correspondingly, adding Gaussian noise to VSP sample data can be done in the following steps: 1) Generate random numbers that follow a Gaussian distribution using a random number generator; 2) Adjust the mean and variance of the generated Gaussian noise according to specific needs. Typically, the mean can be set to 0, and the variance can be adjusted as needed; 3) Add the generated Gaussian noise to the VSP sample data, which can be achieved by simply adding the generated Gaussian noise value to the VSP sample data. Uniform noise is a type of random noise that is uniformly distributed within a specific range. In signal processing and communication systems, uniform noise is often used to simulate random interference or noise in the system. Correspondingly, adding uniform noise to the VSP sample data can be done in the following steps: 1) Determine the amplitude range of the uniform noise, i.e., the minimum and maximum values ​​of the noise, which determines the amplitude of the noise signal; 2) Use a random number generator to generate uniformly distributed random numbers, ensuring that the generated random numbers are within the amplitude range determined in the previous step; 3) Add the generated random numbers to the VSP sample data, which can be achieved through a simple addition operation. The added noise signal can be the noise of a single sample point or the noise of the entire signal sequence; 4) As needed, the intensity or amplitude range of the noise can be adjusted to achieve the desired noise level.

[0128] This application embodiment adds different levels of noise to the VSP sample data to simulate random interference during signal transmission or data acquisition, thereby enhancing the robustness and performance of deep learning networks.

[0129] Furthermore, the deep sparse autoencoder network in this embodiment can also draw on the idea of ​​a denoising autoencoder, where the statistical distribution parameters of the added noise can be obtained by statistically analyzing the noise data using conventional denoising methods. In some embodiments, the preprocessed data, the noise-added data, and the gradient features of different wavefields are input into a second deep sparse autoencoder network, and the second deep sparse autoencoder network is iteratively updated to obtain the deep sparse autoencoder network.

[0130] For example, starting from the initial model, a model training method is provided based on the above embodiments. The obtained preprocessed data, noisy data, and gradient features of different wavefields are used to continuously update the model, thereby obtaining a trained deep sparse autoencoder network, which enables the model to better learn the different wavefield features contained in the VSP data.

[0131] The following section explains the application of deep sparse autoencoder networks in VSP wavefield separation through specific implementation methods.

[0132] Figure 5 This is a schematic flowchart of a VSP wavefield separation method provided as an exemplary embodiment of this application. Figure 5 As shown, the VSP wavefield separation method includes the following steps:

[0133] S501. Obtain VSP data for the target well.

[0134] In this step, still refer to Figure 1 The first client 11 obtains VSP data for wavefield separation from relevant geological exploration departments or geological information platforms, processes the VSP data to obtain VSP data for the target well, and sends it to the server 12; correspondingly, the server 12 receives the VSP data for the target well sent by the first client 11. For example, Figure 6 A schematic diagram of VSP data for a target well provided as an exemplary embodiment of this application.

[0135] S502. Obtain the gradient characteristics of different wavefields in the VSP wavefield represented by the VSP data, and preprocess the VSP data to obtain preprocessed VSP data. The preprocessing includes cubic polynomial interpolation and resampling.

[0136] In this step, the gradient characteristics of different wavefields in the VSP wavefield represented by the VSP data are obtained, and the operations of cubic polynomial interpolation and resampling of the VSP data are the same as those for the VSP sample data described above, and will not be repeated here.

[0137] S503. Input the preprocessed VSP data into a deep sparse autoencoder network for wave field separation. Based on the gradient characteristics of different wave fields, perform wave field separation on the preprocessed data to obtain different wave fields corresponding to the target well.

[0138] The deep sparse autoencoder network is obtained through the model training method described above. For example, the preprocessed VSP data is input into the trained deep sparse autoencoder network for wavefield separation. Based on the gradient features of the uplink and downlink wavefields, wavefield separation is performed on the preprocessed VSP data to obtain the uplink and downlink wavefields corresponding to the target well. For instance, VSP data collected from a well in the Tarim Basin (see...) Figure 6 The VSP wavefield separation method provided in the embodiments of this application is used for, for example, Figure 6 Wavefield separation was performed on the VSP data of the target well shown, and the uplink and downlink wavefields were obtained as follows: Figure 7 and Figure 8 As shown.

[0139] Figure 9 Another schematic diagram of the VSP wavefield separation method provided for an exemplary embodiment of this application. (See attached diagram.) Figure 9 As shown, the VSP wavefield separation method includes the following steps:

[0140] S901. Obtain VSP sample data.

[0141] S902. Preprocess the VSP sample data to obtain preprocessed VSP sample data. The preprocessing includes cubic polynomial interpolation and resampling.

[0142] S903. Obtain the gradient characteristics of different wavefields in the VSP wavefield represented by the VSP sample data.

[0143] S904. Add noise to the VSP sample data to obtain noisy data.

[0144] For example, in one implementation, the noise includes Gaussian noise; in another implementation, the noise includes uniform noise; and in yet another implementation, the noise includes both Gaussian noise and uniform noise.

[0145] S905. Construct a deep sparse autoencoder network based on gradient feature constraints using VSP sample data.

[0146] S906. Establish the total loss function.

[0147] The total loss function satisfies the following formula:

[0148] Loss = Loss recon +λ1Loss grad +λ2Lossspar

[0149] In the above formula, Loss represents the total loss. recon Loss indicates reconstruction losses. grad Represents gradient loss, Loss spar Let λ1 and λ2 represent the sparse penalty term loss, and let λ1 and λ2 be the weights of the corresponding terms (i.e., the weights of the regularization term).

[0150] S907. Set the hyperparameters of the deep sparse autoencoder network.

[0151] For example, hyperparameters include the learning rate, the weight of the regularization term, the number of neurons in the hidden layer, and the number of iterations.

[0152] S908. Iteratively train the deep sparse autoencoder network to obtain the trained deep sparse autoencoder network.

[0153] Based on the preprocessed VSP sample data, the noise-added data, the gradient features of different wave fields, and the total loss function, the deep sparse autoencoder network is iteratively trained to obtain the trained deep sparse autoencoder network.

[0154] S909. Input the preprocessed VSP data of the target well into the trained deep sparse autoencoder network to obtain different wavefields corresponding to the target well.

[0155] The preprocessing of the VSP data for the target well is the same as in S902. Correspondingly, by inputting the preprocessed VSP data of the target well into a trained deep sparse autoencoder network, the uplink and downlink wavefields corresponding to the target well are obtained.

[0156] In summary, this application has at least the following advantages:

[0157] I. By using the gradient characteristics of different wavefields in the VSP wavefield as constraints, and based on a sparse autoencoder network, combined with the advantages of deep neural networks, a deep learning network suitable for VSP wavefield separation is constructed, thereby achieving high-precision VSP wavefield separation. In addition, since this deep learning network adopts an unsupervised mode, no manual labeling is required, which greatly improves the universality of the solution and further enhances the performance and applicability of the deep learning network. This improves the wavefield separation effect when applied to actual VSP data, achieving high-precision VSP wavefield separation and laying a solid foundation for VSP data imaging and corridor overlay.

[0158] Second, by adding two regularization terms to the total loss function, the advantages of machine learning algorithms and the characteristics of VSP data are fully utilized during model training. This allows the deep sparse autoencoder network to better learn the wavefield features contained in the VSP sample data, thereby enhancing the performance and applicability of the deep sparse autoencoder network and achieving high-precision VSP wavefield separation.

[0159] Third, by adding noise to the VSP sample data, random interference during signal transmission or data acquisition can be simulated, thereby enhancing the robustness and performance of deep learning networks.

[0160] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0161] Figure 10 A schematic diagram of a model training apparatus provided for an exemplary embodiment of this application. Figure 10 As shown, the model training device 100 includes a first acquisition module 101, a construction module 102, a second acquisition module 103, a separation module 104, a processing module 105, and an adjustment module 106, wherein:

[0162] The first acquisition module 101 is used to acquire VSP sample data;

[0163] Module 102 is used to construct a first deep sparse autoencoder network based on VSP sample data;

[0164] The second acquisition module 103 is used to acquire the gradient features of different wavefields in the VSP wavefield represented by the VSP sample data.

[0165] The separation module 104 is used to perform wavefield separation on VSP sample data based on the gradient characteristics presented by different wavefields, and obtain different wavefields corresponding to the second deep sparse autoencoder network and VSP sample data.

[0166] The processing module 105 is used to obtain reconstructed VSP data based on different wavefields corresponding to VSP sample data, and to calculate the reconstruction loss, gradient loss and sparsity penalty term loss of the reconstructed VSP data relative to the VSP sample data.

[0167] The adjustment module 106 is used to input the reconstruction loss, gradient loss and sparse penalty term loss into the second deep sparse autoencoder network, and iteratively update the second deep sparse autoencoder network to obtain the deep sparse autoencoder network.

[0168] In one possible implementation, the first deep sparse autoencoder network includes sparse autoencoders and decoders corresponding to different wavefields. The separation module 104 can be specifically used to: input VSP sample data into the sparse autoencoder to obtain low-dimensional features of the sample for the sparse autoencoder and decoder of the target wavefield in different wavefields, input the low-dimensional features of the sample into the decoder, and decode the low-dimensional features of the sample under the constraint of the gradient features presented by the target wavefield to obtain the second deep sparse autoencoder network and the target wavefield corresponding to the VSP sample data.

[0169] In one possible implementation, the second acquisition module 103 may be specifically used to: acquire the phase axis extension angle of each wavefield in the VSP wavefield represented by the VSP sample data in the direction of the seismic phase axis extension; and obtain the gradient characteristics of the wavefield based on the phase axis extension angle.

[0170] In one possible implementation, the second acquisition module 103 can also be used to: determine the convolution kernel based on the phase axis extension angle; and perform a convolution operation on the phase axis extension angle and the convolution kernel to obtain the gradient characteristics of the wave field.

[0171] In one possible implementation, the adjustment module 106 may be specifically used to: determine the total loss based on the reconstruction loss, gradient loss, and sparse penalty term loss; and iteratively update the second deep sparse autoencoder network based on the total loss; wherein the total loss satisfies the following formula:

[0172] Loss = Loss recon +λ1Loss grad +λ2Loss spar

[0173] In the above formula, Loss represents the total loss. recon Loss indicates reconstruction losses. grad Represents gradient loss, Loss spar Let λ1 and λ2 represent the sparse penalty term loss, and let λ1 and λ2 be the weights of the corresponding terms.

[0174] In one possible implementation, the second acquisition module 103 can also be used to: preprocess the VSP sample data before performing wavefield separation on the VSP sample data based on the gradient characteristics presented by different wavefields, to obtain preprocessed data, wherein the preprocessing includes cubic polynomial interpolation and resampling.

[0175] In one possible implementation, the second acquisition module 103 can also be used to: add noise to the VSP sample data to obtain noisy data, wherein the noise includes Gaussian noise and / or uniform noise.

[0176] The model training device provided in this application embodiment can execute the technical solution shown in the above model training method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0177] Figure 11 A schematic diagram of the structure of a VSP wavefield separation device provided as an exemplary embodiment of this application. Figure 11 As shown, the VSP wavefield separation device 110 includes an acquisition module 111, a preprocessing module 112, and a separation module 113, wherein:

[0178] Module 111 is used to acquire VSP data of the target well;

[0179] The preprocessing module 112 is used to obtain the gradient characteristics of different wavefields in the VSP wavefield represented by the VSP data, and to preprocess the VSP data to obtain preprocessed VSP data. The preprocessing includes cubic polynomial interpolation and resampling.

[0180] The separation module 113 is used to input the preprocessed VSP data into a deep sparse autoencoder network for wave field separation. Based on the gradient characteristics of different wave fields, the preprocessed VSP data is separated into different wave fields corresponding to the target well.

[0181] The deep sparse autoencoder network is trained using the model training method described above.

[0182] The apparatus provided in this embodiment can be used to execute the technical solution shown in the above-described VSP wave field separation method embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0183] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into a chip within the device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0184] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-On-a-Chip (SOC).

[0185] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Video Discs, DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0186] Figure 12 A schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application. For example... Figure 12 As shown, the electronic device 120 of this embodiment includes:

[0187] At least one processor 121; and a memory 122 communicatively connected to said at least one processor;

[0188] The memory 122 stores instructions that can be executed by the at least one processor 121 to cause the electronic device to perform the method as described in any of the above embodiments.

[0189] Alternatively, the memory 122 can be either standalone or integrated with the processor 121.

[0190] The memory 122 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0191] The processor 121 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. Specifically, when implementing the model training method or VSP wave field separation method described in the foregoing method embodiments, the electronic device may be, for example, an electronic device with processing capabilities such as a server.

[0192] Optionally, the electronic device may also include a communication interface 123. In specific implementations, if the communication interface 123, memory 122, and processor 121 are implemented independently, they can be interconnected via a bus to complete communication. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.

[0193] Optionally, in a specific implementation, if the communication interface 123, memory 122 and processor 121 are integrated on a single chip, then the communication interface 123, memory 122 and processor 121 can communicate through an internal interface.

[0194] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0195] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed, they are used to implement the method steps as described in the above method embodiments. The specific implementation methods and technical effects are similar and will not be repeated here.

[0196] The aforementioned computer-readable storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0197] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in a model training device or a VSP wavefield separation device.

[0198] This application also provides a computer program product, including a computer program, which, when executed, implements the method steps as described in the above method embodiments. The specific implementation and technical effects are similar and will not be repeated here.

[0199] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0200] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A model training method, characterized in that, include: Obtain vertical seismic profile (VSP) sample data; A first deep sparse autoencoder network is constructed based on the VSP sample data; Obtain the gradient characteristics of different wavefields in the VSP wavefield represented by the VSP sample data; Based on the gradient characteristics of the different wave fields, wave field separation is performed on the VSP sample data using the first deep sparse autoencoder network to obtain the second deep sparse autoencoder network and the different wave fields corresponding to the VSP sample data. Based on the different wavefields corresponding to the VSP sample data, reconstructed VSP data is obtained, and the reconstruction loss, gradient loss, and sparsity penalty term loss of the reconstructed VSP data relative to the VSP sample data are calculated. The reconstruction loss, the gradient loss, and the sparse penalty term loss are input into the second deep sparse autoencoder network, and the second deep sparse autoencoder network is iteratively updated to obtain the deep sparse autoencoder network.

2. The model training method according to claim 1, characterized in that, The first deep sparse autoencoder network includes sparse autoencoders and decoders corresponding to different wavefields. The step of performing wavefield separation on the VSP sample data based on the gradient features exhibited by the different wavefields, thereby obtaining a second deep sparse autoencoder network and the different wavefields corresponding to the VSP sample data, includes: For sparse autoencoders and decoders of target wavefields in different wavefields, the VSP sample data is input into the sparse autoencoder to obtain low-dimensional features of the samples. The low-dimensional features of the samples are then input into the decoder. Under the constraint of the gradient features presented by the target wavefield, the low-dimensional features of the samples are decoded to obtain the second deep sparse autoencoder network and the target wavefield corresponding to the VSP sample data.

3. The model training method according to claim 1 or 2, characterized in that, The step of obtaining the gradient features of different wavefields in the VSP wavefield represented by the VSP sample data includes: For each wavefield in the VSP wavefield represented by the VSP sample data, obtain the in-phase axis extension angle of the wavefield in the extension direction of the seismic in-phase axis; The gradient characteristics of the wave field are obtained based on the extension angle of the phase axis.

4. The model training method according to claim 3, characterized in that, The step of obtaining the gradient characteristics of the wave field based on the extension angle of the phase axis includes: The convolution kernel is determined based on the extension angle of the phase axis. By performing a convolution operation on the extension angle of the phase axis and the convolution kernel, the gradient characteristics of the wave field are obtained.

5. The model training method according to claim 1 or 2, characterized in that, The step of inputting the reconstruction loss, the gradient loss, and the sparse penalty term loss into the second deep sparse autoencoder network, and iteratively updating the second deep sparse autoencoder network to obtain the deep sparse autoencoder network includes: The total loss is determined based on the reconstruction loss, the gradient loss, and the sparse penalty term loss; The second deep sparse autoencoder network is iteratively updated based on the total loss. The total loss satisfies the following formula: Loss=Loss recon +λ1Loss grad +λ2Loss spar In the above formula, Loss represents the total loss. recon The Loss represents the reconstruction loss. grad The gradient loss, Loss spar Let λ1 and λ2 represent the sparse penalty term loss, where λ1 and λ2 are the weights of the corresponding terms, respectively.

6. The model training method according to claim 1 or 2, characterized in that, Before performing wavefield separation on the VSP sample data based on the gradient features presented by the different wavefields, the method further includes: The VSP sample data is preprocessed to obtain preprocessed data. The preprocessing includes cubic polynomial interpolation and resampling.

7. The model training method according to claim 6, characterized in that, Also includes: Noise is added to the VSP sample data to obtain noisy data, wherein the noise includes Gaussian noise and / or uniform noise.

8. The model training method according to claim 7, characterized in that, Also includes: The preprocessed data, the noisy data, and the gradient features of different wave fields are input into the second deep sparse autoencoder network, and the second deep sparse autoencoder network is iteratively updated to obtain the deep sparse autoencoder network.

9. A VSP wavefield separation method, characterized in that, include: Acquire vertical seismic profile (VSP) data of the target well; The gradient features of different wavefields in the VSP wavefield represented by the VSP data are obtained, and the VSP data is preprocessed to obtain preprocessed VSP data. The preprocessing includes cubic polynomial interpolation and resampling. The preprocessed VSP data is input into a deep sparse autoencoder network for wave field separation. Based on the gradient features of the different wave fields, the preprocessed VSP data is separated into different wave fields corresponding to the target well. The deep sparse autoencoder network is obtained by the model training method as described in any one of claims 1 to 8.

10. A model training device, characterized in that, include: The first acquisition module is used to acquire vertical seismic profile (VSP) sample data. The construction module constructs a first deep sparse autoencoder network based on the VSP sample data; The second acquisition module is used to acquire the gradient features of different wavefields in the VSP wavefield represented by the VSP sample data; The separation module is used to perform wavefield separation on the VSP sample data based on the gradient characteristics presented by the different wavefields, and obtain the second deep sparse autoencoder network and the different wavefields corresponding to the VSP sample data. The processing module is used to obtain reconstructed VSP data based on different wavefields corresponding to the VSP sample data, and to calculate the reconstruction loss, gradient loss and sparsity penalty term loss of the reconstructed VSP data relative to the VSP sample data. The adjustment module is used to input the reconstruction loss, the gradient loss, and the sparse penalty term loss into the second deep sparse autoencoder network, and iteratively update the second deep sparse autoencoder network to obtain a trained deep sparse autoencoder network.

11. A VSP wavefield separation device, characterized in that, include: The acquisition module is used to acquire the vertical seismic profile (VSP) data of the target well. The preprocessing module is used to obtain the gradient features of different wavefields in the VSP wavefield represented by the VSP data, and to preprocess the VSP data to obtain preprocessed VSP data. The preprocessing includes cubic polynomial interpolation and resampling. The separation module is used to input the preprocessed VSP data into a deep sparse autoencoder network for wave field separation. Based on the gradient features of the different wave fields, the preprocessed VSP data is separated into different wave fields corresponding to the target well. The deep sparse autoencoder network is obtained by the model training method as described in any one of claims 1 to 8.

12. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory is used to store computer-executed instructions; The processor is configured to execute the computer execution instructions to implement the method as described in any one of claims 1 to 9.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the method as described in any one of claims 1 to 9.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed, it implements the method as described in any one of claims 1 to 9.

Images