Vertical seismic profile wave field separation method, device, equipment and medium

Abstract

The invention provides a vertical seismic profile wave field separation method and device, equipment and a medium, and belongs to the technical field of artificial intelligence, and the method comprises the steps: obtaining a plurality of full-wave data samples in a time-frequency domain in a model training stage, taking difference minimization between upgoing wave data corresponding to full-wave data samples determined by the deep neural network model and the multi-frame Wiener filter and the full-wave data samples as a training target, and performing iterative training on the deep neural network model by using a plurality of full-wave data samples to obtain a target model; in the model reasoning stage, converting the obtained target full-wave data in the time domain into time-frequency domain data from time domain data, and inputting the time-frequency domain data into the target model to obtain target upgoing wave data in the time-frequency domain; and determining target downlink wave data according to a difference value between the target full wave data and the target uplink wave data in the time domain.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and medium for separating vertical seismic profile wave fields. Background Art

[0002] In order to solve the problems of insufficient labeled data and difficulty in manually annotating data in vertical seismic profiling (VSP) wavefield separation scenarios, unsupervised deep learning methods that do not require manual data annotation have gradually attracted attention.

[0003] However, the current VSP wavefield separation method based on unsupervised deep learning requires the use of prior information to achieve wavefield separation, but it is often difficult to accurately obtain prior information in actual scenarios, resulting in poor wavefield separation effects. Summary of the invention

[0004] In view of the above problems, the embodiments of the present application provide a vertical seismic profile wavefield separation method, device, equipment and medium to overcome the above problems or at least partially solve the above problems.

[0005] A first aspect of an embodiment of the present application provides a vertical seismic profile wavefield separation method, the method comprising:

[0006] In the model training stage, a plurality of full-wave data samples in the time-frequency domain are obtained, each of the full-wave data samples comprising: full-wave data of a plurality of adjacent channels in the time-frequency domain, and a plurality of detectors at adjacent positions corresponding to the plurality of adjacent channels;

[0007] In the model training stage, the training goal is to minimize the difference between the upgoing wave data of multiple adjacent channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain, and the deep neural network model is iteratively trained using the multiple full-wave data samples in the time-frequency domain to obtain a target model, wherein the deep neural network model is used to determine the upgoing wave data of the target channel in the multiple adjacent channels associated with it in the time-frequency domain based on the full-wave data samples, and the multi-frame Wiener filter is used to determine the upgoing wave data of other channels in the multiple adjacent channels associated with it in the time-frequency domain based on the upgoing wave data determined by the deep neural network model;

[0008] In the model inference stage, the vertical seismic profile wave field array fed back by the detector is obtained to obtain the target full-wave data in the time domain. The target full-wave data in the time domain is converted from time domain data to time-frequency domain data and then input into the target model to obtain the target upgoing wave data in the time-frequency domain. The target upgoing wave data in the time-frequency domain is converted from time-frequency domain data to time domain data, and the target full-wave data in the time domain is subtracted from the target upgoing wave data in the time domain to obtain the target downgoing wave data in the time domain.

[0009] As a possible implementation manner, the full-wave data samples are obtained by converting the full-wave data of the adjacent multiple channels in the time domain from time domain data to time-frequency domain data; before converting the full-wave data of the adjacent multiple channels in the time domain from time domain data to time-frequency domain data, the method further includes:

[0010] Determine a dynamic time adjustment distance between first full-wave data and second full-wave data, wherein the first full-wave data and the second full-wave data are: two full-wave data having a time offset in full-wave data of adjacent multiple channels of the full-wave data sample in the time domain;

[0011] Determine an alignment path according to the dynamic time adjustment distance, and use the alignment path to perform a leveling and alignment process on the first full-wave data and the second full-wave data;

[0012] The dynamic time adjustment distance is determined based on the following formula:

[0013] γ(i,j)=d(q i ,c j )+min{γ(i-1,j-1),γ(i-1,j),γ(i,j-1)}

[0014] Wherein, γ(i,j) represents the cumulative distance from the first i points of the first full-wave data to the first j points of the second full-wave data. When i is equal to the length of the first full-wave data and j is equal to the length of the second full-wave data, γ(i,j) is the dynamic time adjustment distance; d(q i ,c j ) represents the distance between the i-th point of the first full-wave data and the j-th point of the second full-wave data.

[0015] As a possible implementation, the difference between the upgoing wave data of the adjacent multiple channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain is measured by a target loss function, and the target loss function is expressed as follows:

[0016]

[0017] Among them, L WC Indicates the waveform consistency loss, Y p1 (t,f) represents the real full-wave data sample of the p1th channel at time t and frequency f. represents the upgoing wave data of the p1th channel at time t and frequency f determined based on the deep neural network model and the multi-frame Wiener filter, C represents the number of the multiple channels, Re(·) and Im(·) are used to extract the real part and the imaginary part, respectively, |·| is used to calculate the amplitude, and ‖·‖1 is used to calculate the L1 norm.

[0018] As a possible implementation manner, the upgoing wave data of the target channel in the time-frequency domain is obtained by executing the following steps by the deep neural network model:

[0019] For the full-wave data samples input into the deep neural network model, the two-dimensional convolution layer Conv 2D and the global layer normalization gLN are used in sequence to perform data transformation;

[0020] The data obtained after the data transformation is processed by using two first deep neural network units and a second deep neural network unit in sequence;

[0021] The processed data is transformed and restored using a two-dimensional deconvolution layer Deconv 2D to obtain the upgoing wave data of the target channel in the time-frequency domain.

[0022] As a possible implementation, the first deep neural network unit includes: an expansion layer, a layer normalization, a bidirectional long short-term memory neural network, a one-dimensional deconvolution layer, and a residual layer connected in sequence;

[0023] The second deep neural network unit includes: an attention mechanism module, a two-dimensional convolutional layer, a parameterized rectified linear unit, layer normalization along the channel and frequency dimensions, and a residual layer connected in sequence.

[0024] As a possible implementation manner, the upgoing wave data of other channels in the adjacent multiple channels in the time-frequency domain are determined by the following equations:

[0025]

[0026] in, represents the upgoing wave data from the first channel of the plurality of adjacent channels at time t and frequency f The obtained stacked tensor, the first channel is the target channel, I and J represent the hyperparameters of the multi-frame Wiener filter; represents the upgoing wave data of the p2th channel in the plurality of adjacent channels at time t and frequency f, represents the filter corresponding to the p2th channel, the p2th channel is the other channels among the plurality of adjacent channels except the first channel, and C represents the number of the plurality of adjacent channels.

[0027] As a possible implementation manner, converting the target full-wave data in the time domain from time domain data to time-frequency domain data and then inputting it into the target model comprises:

[0028] By short-time Fourier transform, the target full-wave data in the time domain is converted from time domain data to time-frequency domain data to obtain the target full-wave data in the time-frequency domain;

[0029] After stacking the real part and the imaginary part of the target full-wave data in the time-frequency domain, the stacked parts are input into the target model;

[0030] Converting the target upgoing wave data in the time-frequency domain from time-frequency domain data to time domain data comprises:

[0031] The target upgoing wave data in the time-frequency domain is converted from time-frequency domain data to time-domain data by inverse short-time Fourier transform, so as to obtain the target upgoing wave data in the time domain.

[0032] A second aspect of the embodiments of the present application provides a vertical seismic profile wave field separation device, the device comprising:

[0033] The first training module is used to obtain a plurality of full-wave data samples in the time-frequency domain during the model training phase, each of the full-wave data samples comprising: full-wave data of a plurality of adjacent channels in the time-frequency domain, and a plurality of detectors at adjacent positions corresponding to the plurality of adjacent channels;

[0034] A second training module is used for iteratively training the deep neural network model using the multiple full-wave data samples in the time-frequency domain to obtain a target model, with the goal of minimizing the difference between the upgoing wave data of multiple adjacent channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and their corresponding full-wave data samples in the time-frequency domain during the model training phase, wherein the deep neural network model is used to determine the upgoing wave data of a target channel in the multiple adjacent channels associated therewith in the time-frequency domain based on the full-wave data samples, and the multi-frame Wiener filter is used to determine the upgoing wave data of other channels in the multiple adjacent channels associated therewith in the time-frequency domain based on the upgoing wave data determined by the deep neural network model;

[0035] A model inference module is used to obtain the vertical seismic profile wave field array fed back by the detector to obtain target full-wave data in the time domain, convert the target full-wave data in the time domain from time domain data to time-frequency domain data and then input it into the target model to obtain target upgoing wave data in the time-frequency domain, convert the target upgoing wave data in the time-frequency domain from time-frequency domain data to time domain data, and use the target full-wave data in the time domain to subtract the target upgoing wave data in the time domain to obtain target downgoing wave data in the time domain.

[0036] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the vertical seismic profile wave field separation method described in the first aspect are implemented.

[0037] According to a fourth aspect of the embodiments of the present application, a computer-readable storage medium is provided, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the vertical seismic profile wave field separation method described in the first aspect are implemented.

[0038] The embodiments of the present application include the following advantages: by virtue of the advantage that the deep neural network model can extract rich features of waveform data, in the model training stage, based on the training target and the acquired full-wave data samples, the deep neural network model is iteratively trained in an unsupervised learning manner to obtain the target model, and in the model inference stage, the target model is used to determine the upgoing wave data in the time-frequency domain, and then the full-wave data is used to subtract the upgoing wave data in the time domain to obtain the downgoing wave data, thereby realizing the separation of the upgoing and downgoing waves, thereby avoiding the introduction of manually labeled data or prior information in the wave field separation process, thereby improving the wave field separation effect; and in the model training stage, a multi-frame Wiener filter is introduced to determine the upgoing wave data of other channels in the time-frequency domain, and then the deep neural network model is optimized by minimizing the difference between the multi-channel full-wave data and the multi-channel upgoing wave data, thereby ensuring that when the model is trained for a multi-channel array, the travel time difference between the upgoing wave and the downgoing wave can be maximized, so as to further improve the wave field separation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0040] Figure 1is an implementation flow chart of a vertical seismic profile wave field separation method in an embodiment of the present application;

[0041] Figure 2 It is a schematic diagram of a before-and-after comparison effect of flattening a VSP wave field using a DTW algorithm in an embodiment of the present application;

[0042] Figure 3 is a schematic diagram of a model optimization process in an embodiment of the present application;

[0043] Figure 4 is a schematic diagram of a deep neural network model structure in an embodiment of the present application;

[0044] Figure 5 is a schematic diagram of a VSP principle in an embodiment of the present application;

[0045] Figure 6 It is a schematic diagram of the separation effect of uplink and downlink waves of a VSP in an embodiment of the present application;

[0046] Figure 7 It is a schematic diagram of another VSP uplink and downlink wave separation effect in an embodiment of the present application;

[0047] Figure 8 It is a structural schematic diagram of a vertical seismic profile wave field separation device in an embodiment of the present application;

[0048] Fig. 9 It is a schematic diagram of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0049] To facilitate understanding of the technical solutions provided by the present application, the main technical concepts involved in the embodiments of the present application are briefly described below.

[0050] Dynamic Time Warping (DTW): is an algorithm based on dynamic programming strategy to align two time series nonlinearly in the time domain in order to correctly calculate the similarity between the two.

[0051] Deep Neural Network (DNN) model: is a computational model that simulates the structure and function of the human brain neural network. Its basic unit is the neuron. Each neuron receives input from other neurons and changes the impact of the input on the neuron by adjusting the weight.

[0052] Multi-frame Wiener filter (MFWF): A linear filter that uses the least squares as the optimal criterion for multi-frame data.

[0053] Multi-channel loss (MC loss): is a function used to estimate the degree of inconsistency between the predicted value and the true value of the multi-channel neural network model.

[0054] Two-Dimensional Convolutional layer (Conv 2D): It is a key component in convolutional neural networks. It is mainly used to process data with a grid structure and detect local features in the input data by learning a set of filters.

[0055] Global Layer Normalization (gLN): is a normalization technique applied in neural networks that calculates the mean and variance of all samples and channels in the entire batch, thereby achieving a more stable training process and better model performance.

[0056] Two-Dimensional Deconvolution (Deconv 2D): It is the inverse process of the convolution operation.

[0057] Bidirectional Long Short Term Memory neural network (BiLSTM): is a sequence model that combines two Long Short Term Memory neural network (LSTM) units in opposite directions. It can capture contextual information in the sequence from both the forward and backward directions, thereby enhancing the model's ability to understand sequence data.

[0058] Parametric Rectified Linear Unit (PReLU): is an activation function that introduces parameters based on the Rectified Linear Unit (ReLU) so that the function is not completely zero on the negative semi-axis, which helps to alleviate the gradient vanishing problem and allows the network to learn non-zero slopes, improving the flexibility and expressiveness of the model.

[0059] Layer Normalization along the channel and frequency dimensions (cfLN): is a normalization technique specifically designed to process data with channel and frequency dimensions, commonly found in neural networks that process audio signals or spectrograms.

[0060] Vertical Seismic Profiling (VSP) is a seismic exploration technology that receives seismic waves excited by ground sources by arranging geophones in the well and obtains vertical seismic profiles.

[0061] VSP is mainly used in oil and gas exploration, geothermal energy exploration, geological structure research, earthquake monitoring and research, etc. It arranges detectors in the well to receive seismic waves excited by the ground source and obtain vertical seismic profiles. VSP provides the most direct correspondence between underground formation results and measurement parameters. Compared with conventional ground seismic exploration technology, it has the characteristics of small interference, rich wave field information and good imaging effect. Take oil as an example. As one of the scarce resources, oil reserves are gradually decreasing, there is a lack of new oil exploration areas, and the oil production in old areas is too low. This puts higher requirements on oil exploration work. VSP, as a more refined exploration method, can better meet the high requirements of oil exploration work and has been widely used.

[0062] It should be noted that the main VSP wave field can be divided into upgoing waves and downgoing waves according to the direction of wave propagation to the receiving point, corresponding to the wave fields below and above the receiving point, respectively. Due to the weakening of wave field reflection energy, upgoing waves are usually weaker than downgoing waves and more difficult to identify. In VSP data interpretation, upgoing wave data is mainly used. In addition to the problems of random and coherent interference and low data signal-to-noise ratio, the original VSP data also has the problem of overlapping upgoing and downgoing waves, which makes it difficult to interpret the original VSP data directly. Therefore, it is necessary to use the VSP wave field separation method to separate the upgoing and downgoing waves, which is also conducive to improving the effect of subsequent VSP imaging.

[0063] Among the traditional VSP wavefield separation methods, median filtering, frequency wavenumber filtering (FK filtering) and Radon transform are widely used. Among them, median filtering selects the middle value of each window, which can enhance the stability along one direction and retain the ability to reject anomalies, but it highly requires the consistency of the depth waveform and requires manual selection of parameters; FK filtering and Radon transform work in the transform domain, highlighting the apparent velocity difference between the uplink and downlink wavefields, making filtering easier, but they have the defects of truncation artifacts and spatial aliasing, and insufficient and uneven depth sampling will also cause distortion in the spatial transformation process.

[0064] With the development of artificial intelligence technology, deep learning-based methods have achieved good results in various fields. Many deep learning studies have also been carried out on the problem of wave field separation and have been proven to be effective. Most deep learning methods optimize neural networks by minimizing the difference between actual labels and predicted labels. However, due to the lack of label data and the difficulty of manually annotating data in the VSP wave field separation scenario, synthetic data is often used to obtain the true label value. However, this method faces the problem of distribution differences between synthetic data and actual data, resulting in a model that has a good fitting effect on synthetic data but has a poor effect on real data.

[0065] In order to solve the problems of insufficient labeled data and difficulty in manually annotating data in VSP wavefield separation scenarios, and to avoid the influence of synthetic data on the wavefield separation effect, unsupervised deep learning methods that do not require manual data annotation have gradually attracted attention. It is understandable that unsupervised deep learning methods extract the inherent laws of data by automatically discovering complex structures and features in the data, thus avoiding the use of manually annotated data.

[0066] However, the current VSP wavefield separation method based on unsupervised deep learning requires the use of prior information to achieve wavefield separation, but it is often difficult to accurately obtain prior information in actual scenarios, resulting in poor wavefield separation effects.

[0067] Taking the unsupervised VSP wavefield separation method based on dual convolutional autoencoders (DualCAE) as an example, this method uses the directional continuity and zero mean characteristics of the uplink and downlink wavefields to design gradient loss and cosine similarity loss for model constraints, so that a good wavefield separation effect can be achieved without any manual annotation data. However, this method requires the slopes of the uplink and downlink waves to be determined in advance to provide prior information for the subsequent gradient loss; however, in actual situations, the slopes of the uplink and downlink waves are often difficult to obtain accurately due to noise interference, which leads to poor wavefield separation effects of this method.

[0068] In view of the problems existing in the above-mentioned related technologies, the embodiments of the present application propose a method, device, equipment and medium for wavefield separation of vertical seismic profiles. In the model training stage, the upgoing wave data of the target channel in the time-frequency domain is estimated by a deep neural network model, and then the reversibility of the wave is used to estimate the upgoing wave data of other channels in the time-frequency domain using MFWF according to the model estimation results, and the deep neural network model is optimized by minimizing the difference between multi-channel full-wave data and multi-channel upgoing wave data (i.e., the estimation results of the deep neural network model and MFWF) to obtain the target model, thereby ensuring that when the model is trained for a multi-channel array, the travel time difference between the upgoing wave and the downgoing wave can be maximized, thereby ensuring the subsequent wavefield separation effect; in the model reasoning stage, the target model is used to determine the upgoing wave data in the time-frequency domain, and then the full-wave data is used to subtract the upgoing wave data in the time domain to obtain the downgoing wave data, thereby realizing the separation of the upgoing and downgoing waves, thereby avoiding the introduction of manually labeled data or prior information in the wavefield separation process, thereby achieving effective improvement of the wavefield separation effect.

[0069] In combination with the accompanying drawings, a vertical seismic profile wave field separation method, device, equipment and medium provided in the embodiments of the present application are described in detail through some embodiments and their application scenarios.

[0070] First, refer to Figure 1 FIG. 1 is a flowchart of a vertical seismic profile wave field separation method provided in an embodiment of the present application, wherein the method comprises the following steps:

[0071] Step S11: In the model training stage, a plurality of full-wave data samples are obtained in the time-frequency domain, each of the full-wave data samples comprising: full-wave data of a plurality of adjacent channels in the time-frequency domain, and the plurality of adjacent channels correspond to a plurality of adjacent detectors.

[0072] The adjacent multiple channels correspond one-to-one to the adjacent multiple detectors, and the multiple detectors are pre-arranged at different depths in the wellbore to receive effective waves, so as to observe and study the vertical changes of the geological profile in the well.

[0073] In the specific implementation, for the model training stage, the VSP wave field arrays fed back by multiple geophones whose position distance is within the set range (i.e., adjacent to each other) are obtained to obtain the full-wave data of multiple adjacent channels in the time domain. It can be understood that by obtaining the VSP wave field arrays fed back by multiple geophones whose positions are adjacent to each other, the correlation between the multiple VSP wave field arrays (i.e., the full-wave data of the multiple channels in the time domain) can be guaranteed, thereby ensuring the accuracy of the upgoing wave data subsequently determined for the multiple channels (for example, determined using MFWF).

[0074] Then, the full-wave data of the adjacent multiple channels in the time domain are converted from time domain data to time-frequency domain data, so as to obtain the full-wave data of the adjacent multiple channels in the time-frequency domain (i.e., the full-wave data samples). The embodiment of the present application does not specifically limit the method used to convert the time domain data into the time-frequency domain data.

[0075] Optionally, considering that the wavefield array of the VSP is formed by multiple detectors distributed at different depths in the wellbore, the wavefield array fed back by each detector may be offset in time with different depths. Therefore, after obtaining the full-wave data of the multiple adjacent channels in the time domain, each full-wave data can be flattened and aligned based on a sequence alignment method to eliminate the time offset of each full-wave data and align them in time. The embodiment of the present application does not impose any specific limitation on the sequence alignment method.

[0076] Step S12: In the model training stage, the training goal is to minimize the difference between the upgoing wave data of multiple adjacent channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain, and the deep neural network model is iteratively trained using the multiple full-wave data samples in the time-frequency domain to obtain a target model, wherein the deep neural network model is used to determine the upgoing wave data of the target channel in the time-frequency domain associated with the multiple adjacent channels based on the full-wave data samples, and the multi-frame Wiener filter is used to determine the upgoing wave data of other channels in the time-frequency domain associated with the multiple adjacent channels based on the upgoing wave data determined by the deep neural network model.

[0077] In the specific implementation, for the model training stage, the difference between the full-wave data sample (also called the real full-wave data sample, which contains the real full-wave data of multiple channels) and the up-wave data of the target channel determined by the model (i.e., the model predicts and estimates) and the up-wave data of other channels determined by the multi-frame Wiener filter is minimized as the training target to adjust the parameters of the deep neural network model (which is a multi-channel single-output network model). Therefore, by introducing the multi-frame Wiener filter to constrain the deep neural network model in the model training stage, it is possible to avoid the situation where the up-wave generated by the model completely overlaps with the original full wave.

[0078] For example, a loss function can be designed to characterize the difference between the complete waveform spectrum (i.e., the full-wave spectrum) and the uplink wave spectrum determined by the model and the multi-frame Wiener filter, and then the loss function is used to iteratively train the model, and the converged deep neural network model is used as the target model.

[0079] Step S13: In the model inference stage, the vertical seismic profile wave field array fed back by the detector is obtained to obtain the target full-wave data in the time domain, the target full-wave data in the time domain is converted from time domain data to time-frequency domain data and then input into the target model to obtain the target upgoing wave data in the time-frequency domain, the target upgoing wave data in the time-frequency domain is converted from time-frequency domain data to time domain data, and the target full-wave data in the time domain is subtracted from the target upgoing wave data in the time domain to obtain the target downgoing wave data in the time domain.

[0080] In the specific implementation, for the model reasoning stage, the target full-wave data in the time domain is converted from time domain data to time-frequency domain data, so that the target model can analyze and infer the frequency characteristics of the input data, thereby improving the accuracy of the model output results, so that the target upgoing wave data determined by the target model can be closer to the actual upgoing wave data.

[0081] Then, after the target upgoing wave data in the time-frequency domain is converted from the time-frequency domain data to the time domain data, the target downgoing wave data in the time domain can be obtained by subtracting the upgoing wave data in the time domain from the target full-wave data in the time domain (i.e., the vertical seismic profile wave field array fed back by the detector). The embodiment of the present application does not specifically limit the method used to convert the time-frequency domain data into the time domain data.

[0082] By adopting the technical solution of the embodiment of the present application, with the advantage that the deep neural network model can extract rich features of waveform data, the deep neural network model is iteratively trained in an unsupervised learning manner based on the training target and the acquired full-wave data samples in the model training stage to obtain the target model, and the target model is used to determine the upgoing wave data in the time-frequency domain in the model inference stage, and then the upgoing wave data is subtracted from the full-wave data in the time domain to obtain the downgoing wave data, thereby realizing the separation of the upgoing and downgoing waves, thereby avoiding the introduction of manually labeled data or prior information in the wave field separation process, thereby improving the wave field separation effect; and in the model training stage, a multi-frame Wiener filter is introduced to determine the upgoing wave data of other channels in the time-frequency domain, and then the deep neural network model is optimized by minimizing the difference between the multi-channel full-wave data and the multi-channel upgoing wave data, thereby ensuring that when the model is trained for a multi-channel array, the travel time difference between the upgoing wave and the downgoing wave can be maximized to further improve the wave field separation effect.

[0083] As a possible implementation manner, the full-wave data samples are obtained by converting the full-wave data of the adjacent multiple channels in the time domain from time domain data to time-frequency domain data; before converting the full-wave data of the adjacent multiple channels in the time domain from time domain data to time-frequency domain data, the method further includes:

[0084] Determine a dynamic time adjustment distance between first full-wave data and second full-wave data, wherein the first full-wave data and the second full-wave data are: two full-wave data having a time offset in the full-wave data of the adjacent multiple channels in the time domain;

[0085] Determine an alignment path according to the dynamic time adjustment distance, and use the alignment path to perform a leveling and alignment process on the first full-wave data and the second full-wave data;

[0086] The dynamic time adjustment distance is determined based on the following formula:

[0087] γ(i,j)=d(q i ,c j )+min{γ(i-1,j-1),γ(i-1,j),γ(i,j-1)}

[0088] Wherein, γ(i,j) represents the cumulative distance from the first i points of the first full-wave data to the first j points of the second full-wave data. When i is equal to the length of the first full-wave data and j is equal to the length of the second full-wave data, γ(i,j) is the dynamic time adjustment distance; d(q i ,c j ) represents the distance between the i-th point of the first full-wave data and the j-th point of the second full-wave data.

[0089] In this embodiment, considering that the wavefield array of VSP is formed by multiple detectors distributed at different depths in the wellbore, the wavefield array fed back by each detector is offset in time with different depths. Therefore, the present application applies the DTW algorithm to wavefield separation, and pre-flattens and aligns each full-wave data based on the DTW algorithm to eliminate the temporal offset of each data, thereby improving the robustness of the wavefield separation method.

[0090] Specifically, DTW is an algorithm for measuring the similarity between two sequences. The core idea is to find a nonlinear correspondence between the two sequences so that the distance between them is minimized. Assume that there are two sequences Q1 and C1 (i.e., the first full-wave data and the second full-wave data) of length m and n respectively, and construct an n×m matrix. The position (i, j) of the matrix is ​​used to store d(q i ,c j ), and let W = w1, w2, ..., w k ,...,w K is the alignment path, w k Corresponding to the kth position in the matrix, the alignment path can be determined according to the following formula:

[0091]

[0092] Among them, K∈[max(m,n),m+n-1), DTW(Q1,C1) represents the dynamic time adjustment distance (i.e., DTW distance) between sequence Q1 and sequence C1, and when i is equal to the length m of the first full-wave data (i.e., sequence Q1) and j is equal to the length n of the second full-wave data (i.e., sequence C1), γ(i,j) is the DTW distance, that is, γ(m,n)=DTW(Q1,C1); therefore, DTW distance calculation can be regarded as a dynamic programming problem, and by solving γ(m,n), the corresponding alignment path can be obtained by inferring from the result.

[0093] After the alignment path is obtained, the first full-wave data and the second full-wave data are stretched and compressed along the inter-axis according to the alignment path to achieve a flattening alignment process. The schematic diagram of the before-and-after comparison effect of flattening the VSP wave field using the DTW algorithm is as follows: Figure 2 Then, the above steps are repeated for each data with time offset in each full-wave data, so as to eliminate the time offset of each full-wave data.

[0094] As a possible implementation method, refer to Figure 3 The schematic diagram of the model optimization process shown in FIG. 1 shows that the difference between the upgoing wave data of multiple adjacent channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain is measured by the target loss function, and the target loss function is expressed as follows:

[0095]

[0096] Among them, L WC Indicates the waveform consistency loss, Y p1 (t,f) represents the real full-wave data sample of the p1th channel at time t and frequency f. represents the upgoing wave data of the p1th channel at time t and frequency f determined based on the deep neural network model and the multi-frame Wiener filter (that is, the upgoing wave estimation result of the target channel is determined by the deep neural network model, and the upgoing wave estimation results of other channels are determined by the multi-frame Wiener filter according to the estimation results of the aforementioned model), C represents the number of the multiple adjacent channels, Re(·) and Im(·) are used to extract the real part and the imaginary part, respectively, |·| is used to calculate the amplitude, and ‖·‖1 is used to calculate the L1 norm.

[0097] In this implementation, a multi-channel loss function is designed to optimize the deep neural network model. By comprehensively considering the upgoing waves estimated by the model and the multi-frame Wiener filter for all channels, the model is fine-tuned with the goal of minimizing the difference between the complete waveform spectrum of all channels and the estimated upgoing wave spectrum. This ensures that when the model is trained for a multi-channel array, the travel time difference between the upgoing wave and the downgoing wave can be maximized, thereby improving the wave field separation effect.

[0098] As a possible implementation manner, the upgoing wave data of the target channel in the time-frequency domain is obtained by executing the following steps by the deep neural network model:

[0099] For the full-wave data samples input into the deep neural network model, the two-dimensional convolution layer Conv 2D and the global layer normalization gLN are used in sequence to perform data transformation;

[0100] The data obtained after the data transformation is processed by using two first deep neural network units and a second deep neural network unit in sequence;

[0101] The processed data is transformed and restored using a two-dimensional deconvolution layer Deconv 2D to obtain the upgoing wave data of the target channel in the time-frequency domain.

[0102] In this embodiment, reference Figure 4 The schematic diagram of the deep neural network model structure shown is that in order to improve the estimation accuracy of the model for upgoing waves, the present application designs a deep neural network model suitable for upgoing wave estimation, and iteratively trains it in an unsupervised learning manner to obtain the target model.

[0103] Specifically, the deep neural network model includes: a two-dimensional convolution layer, a global layer normalization, two first deep neural network units, a second deep neural network unit, and a two-dimensional deconvolution layer connected in sequence. After the full-wave data sample is input into the deep neural network model, it will first be transformed by the two-dimensional convolution layer and the global layer normalization, and then processed by the two first deep neural network units and the second deep neural network unit in sequence, and finally after the two-dimensional deconvolution layer restores the data transformation, the model outputs the predicted uplink wave data.

[0104] Optionally, the first deep neural network unit comprises: an expansion layer, a layer normalization, a bidirectional long short-term memory neural network, a one-dimensional deconvolution layer and a residual layer connected in sequence;

[0105] The second deep neural network unit includes: an attention mechanism module, a two-dimensional convolutional layer, a parameterized rectified linear unit, layer normalization along the channel and frequency dimensions, and a residual layer connected in sequence.

[0106] As a possible implementation method, in order to reduce the complexity of data conversion and model processing, the full-wave data in the time domain (such as the target full-wave data in the time domain) is first converted from time domain data to time-frequency domain data through short-time Fourier Transform (STFT) to obtain full-wave data in the time-frequency domain (such as the target full-wave data in the time-frequency domain).

[0107] Exemplarily, the full-wave data of multiple adjacent channels in the time-frequency domain can be expressed as follows:

[0108] Y p (t,f)=U p (t,f)+D p (t,f),p∈{1,…,C}Equation (1)

[0109] Wherein, C represents the number of the adjacent multiple channels, Y p (t,f) represents the full-wave data of the pth channel in the time-frequency domain (i.e., the STFT vector of the full wave at time t and frequency f), U p (t,f) represents the upgoing wave data of the pth channel in the time-frequency domain (i.e., the STFT vector of the upgoing wave at time t and frequency f), D p (t,f) represents the downlink wave data of the pth channel in the time-frequency domain (i.e., the STFT vector of the downlink wave at time t and frequency f), Y p (t,f),U p (t,f) and Represents a complex number domain.

[0110] In order to facilitate the model to process the data of each channel, the real part and the imaginary part of the full-wave data of each channel in the time-frequency domain are stacked (for example, the real part and the imaginary part of the target full-wave data in the time-frequency domain are stacked), and then the data obtained after stacking is The target model is then inputted, and the model generates an estimated upgoing wave for each channel (such as the target upgoing wave data in the time-frequency domain). Stacking means packing each object involved in the stacking into a higher-dimensional tensor. For example, by stacking two one-dimensional data, a two-dimensional data (i.e., a higher-dimensional tensor) can be obtained.

[0111] As a possible implementation method, in order to reduce the complexity of data conversion, the upgoing wave data in the time-frequency domain (such as the target upgoing wave data in the time-frequency domain) is converted from time-frequency domain data to time domain data through inverse short-time Fourier transform to obtain the upgoing wave data in the time domain (such as the target upgoing wave data in the time domain).

[0112] As a possible implementation manner, the upgoing wave data of other channels in the adjacent multiple channels in the time-frequency domain are determined by the following equations:

[0113]

[0114] in, represents the upgoing wave data from the first channel of the plurality of adjacent channels at time t and frequency f The obtained stacked tensor, the first channel is the target channel, I and J represent the hyperparameters of the multi-frame Wiener filter; represents the upgoing wave data of the p2th channel in the plurality of adjacent channels at time t and frequency f, represents the filter corresponding to the p2th channel, the p2th channel is the other channels among the plurality of adjacent channels except the first channel, and C represents the number of the plurality of adjacent channels.

[0115] In this embodiment, the present application considers that filtering the spectrum of the uplink wave data U1(t, f) of the target channel can obtain the uplink wave data U1(t, f) of the pth channel (ie, other channels). p (t,f) spectrum, that is, according to the reversibility of wave propagation, U p The spectrum of (t,f) can be generated by forward or backward propagation of the spectrum of U1(t,f).

[0116] Therefore, this application uses the multi-frame Wiener filter (MFWF) and the model to estimate the upgoing wave data of the target channel. Multiply them together to predict the uplink wave data of other channels. Based on this, the above equation (1) can be restated as the following equation (3) for the 2nd to Cth channels (i.e. other channels):

[0117]

[0118] Wherein, C represents the number of the adjacent multiple channels, Y p (t,f) represents the full-wave data of the pth channel (here it means other channels except the first channel) in the time-frequency domain. represents the upgoing wave data from the first channel (i.e., the target channel) of the plurality of adjacent channels at time t and frequency f The obtained stacked tensors, I and J represent the hyperparameters of the multi-frame Wiener filter, represents the (I+J+1)-dimensional complex field; is a time-invariant linear filter, (·) H represents the conjugate transpose, represents the complex domain; Dp (t,f) represents the downlink wave data of the pth channel in the time-frequency domain.

[0119] based on The stacked time-frequency spectrum can be estimated by solving the minimization problem corresponding to the following equation (4) to estimate the filter at each frequency

[0120]

[0121] It can be understood that equation (4) is a linear regression problem, and its closed-form solution (or analytical solution) is shown in the following equation (5):

[0122]

[0123] Finally, the parameter p2 is used to replace the parameter p in the above equation (4), and the filter determined based on the above equation (4) can be obtained. The above equation (2) is used to obtain the upgoing wave data estimated based on the multi-frame Wiener filter:

[0124] Based on the above examples and embodiments, it can be seen that VSP is an exploration technology that receives effective waves by arranging detectors at different depths in the wellbore. The schematic diagram of the VSP principle is as follows: Figure 5 shown.

[0125] Since VSP observes and studies the vertical changes of geological profiles in the well, compared with ground seismic exploration technology, the kinematic and dynamic characteristics of seismic waves in VSP are more obvious; and VSP has a higher signal-to-noise ratio, because VSP data only passes through the surface once, which makes the wavefront diffusion, stratum absorption, low-velocity zone and other waves have little influence and transformation, and thus makes the wave distortion and interference less. In addition, the receiving point of the vertical seismic profile is on or near the interface, so the purer sub-wave waveform related to the interface can be directly recorded. In addition, ground seismic exploration technology cannot record the downlink wave field, while VSP can record uplink and downlink waves at the same time, which is convenient for studying the directional characteristics of seismic waves; and it is easier, more real and has higher fidelity to extract velocity parameters, amplitude information, lithology parameters, etc. from VSP data; therefore, VSP can be well applied to oil and gas exploration, geothermal energy exploration, geological structure research, earthquake monitoring and research and other fields.

[0126] However, since the detector and the source are not located on the same horizontal plane, the upgoing and downgoing waves containing different stratigraphic information are intertwined and superimposed on each other, making it difficult to directly use them for stratigraphic interpretation. Therefore, only by correctly identifying, separating and extracting the upgoing and downgoing waves can the useful information of VSP be fully utilized.

[0127] In view of the above problems, the present application uses a target model (i.e., a deep neural network model that converges after iterative training) to extract the rich features of the VSP waveform data, and estimates the upgoing wave data based on it, and then uses the full wave data to subtract the estimated upgoing wave data to obtain the estimated downgoing wave data, thereby achieving the separation of the upgoing and downgoing waves. The VSP wave field separation method provided in the embodiment of the present application can achieve a good VSP upgoing and downgoing wave separation effect on real data without introducing prior information. The schematic diagram of the VSP upgoing and downgoing wave separation effect is shown in FIG. Figure 6 and Figure 7 shown.

[0128] In addition, during the model training stage, the present application optimizes the deep neural network model parameters by maximizing the unsupervised learning method of the deep neural network model and the multi-frame Wiener filter for the similarity between the upgoing waves predicted by each channel and the original full wave, thereby maximizing the travel time difference between the upgoing waves and the downgoing waves, thereby improving the effect of wave field separation, and making it unnecessary to use synthetic data in the model training process, that is, unsupervised learning can be carried out directly on real data, thereby greatly avoiding the problem that the model is not suitable for real data due to the inconsistent distribution of synthetic data and real data.

[0129] For the method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present application are not limited by the order of the actions described, because according to the embodiments of the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present application.

[0130] Second, Figure 8 : is a structural schematic diagram of a vertical seismic profile wave field separation device according to an embodiment of the present application, the device comprising:

[0131] The first training module 810 is used to obtain a plurality of full-wave data samples in the time-frequency domain during the model training phase, each of the full-wave data samples comprising: full-wave data of a plurality of adjacent channels in the time-frequency domain, and a plurality of detectors at adjacent positions corresponding to the plurality of adjacent channels;

[0132] The second training module 820 is used to, in the model training stage, take minimizing the difference between the upgoing wave data of multiple adjacent channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain as the training goal, and use the multiple full-wave data samples in the time-frequency domain to iteratively train the deep neural network model to obtain a target model, wherein the deep neural network model is used to determine the upgoing wave data of the target channel in the multiple adjacent channels associated with it in the time-frequency domain based on the full-wave data samples, and the multi-frame Wiener filter is used to determine the upgoing wave data of other channels in the multiple adjacent channels associated with it in the time-frequency domain based on the upgoing wave data determined by the deep neural network model;

[0133] The model inference module 830 is used to obtain the vertical seismic profile wave field array fed back by the detector, obtain the target full-wave data in the time domain, convert the target full-wave data in the time domain from time domain data to time-frequency domain data, and then input it into the target model to obtain the target upgoing wave data in the time-frequency domain, convert the target upgoing wave data in the time-frequency domain from time-frequency domain data to time domain data, and use the target full-wave data in the time domain to subtract the target upgoing wave data in the time domain to obtain the target downgoing wave data in the time domain.

[0134] By adopting the technical solution of the embodiment of the present application, with the advantage that the deep neural network model can extract rich features of waveform data, the deep neural network model is iteratively trained in an unsupervised learning manner based on the training target and the acquired full-wave data samples in the model training stage to obtain the target model, and the target model is used to determine the upgoing wave data in the time-frequency domain in the model inference stage, and then the upgoing wave data is subtracted from the full-wave data in the time domain to obtain the downgoing wave data, thereby realizing the separation of the upgoing and downgoing waves, thereby avoiding the introduction of manually labeled data or prior information in the wave field separation process, thereby improving the wave field separation effect; and in the model training stage, a multi-frame Wiener filter is introduced to determine the upgoing wave data of other channels in the time-frequency domain, and then the deep neural network model is optimized by minimizing the difference between the multi-channel full-wave data and the multi-channel upgoing wave data, thereby ensuring that when the model is trained for a multi-channel array, the travel time difference between the upgoing wave and the downgoing wave can be maximized to further improve the wave field separation effect.

[0135] Optionally, the full-wave data samples are obtained by converting full-wave data of a plurality of adjacent channels in the time domain from time domain data to time-frequency domain data; the device further comprises:

[0136] A first alignment module is used to determine a dynamic time adjustment distance between first full-wave data and second full-wave data before converting the full-wave data of the adjacent multiple channels in the time domain from time domain data to time-frequency domain data, wherein the first full-wave data and the second full-wave data are: two full-wave data that are offset in time in the full-wave data of the adjacent multiple channels in the time domain;

[0137] A second alignment module, used for determining an alignment path according to the dynamic time adjustment distance, and using the alignment path to perform a leveling and alignment process on the first full-wave data and the second full-wave data;

[0138] The dynamic time adjustment distance is determined based on the following formula:

[0139] γ(i,j)=d(q i ,c j )+min{γ(i-1,j-1),γ(i-1,j),γ(i,j-1)}

[0140] Wherein, γ(i,j) represents the cumulative distance from the first i points of the first full-wave data to the first j points of the second full-wave data. When i is equal to the length of the first full-wave data and j is equal to the length of the second full-wave data, γ(i,j) is the dynamic time adjustment distance; d(q i ,c j ) represents the distance between the i-th point of the first full-wave data and the j-th point of the second full-wave data.

[0141] Optionally, the difference between the upgoing wave data of the adjacent multiple channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain is measured by a target loss function, and the target loss function is expressed as follows:

[0142]

[0143] Among them, L WC Indicates the waveform consistency loss, Y p1 (t,f) represents the real full-wave data sample of the p1th channel at time t and frequency f. represents the upgoing wave data of the p1th channel at time t and frequency f determined based on the deep neural network model and the multi-frame Wiener filter, C represents the number of the multiple adjacent channels, Re(·) and Im(·) are used to extract the real part and the imaginary part, respectively, |·| is used to calculate the amplitude, and ‖·‖1 is used to calculate the L1 norm.

[0144] Optionally, the second training module 820 is further configured to perform the following steps through the deep neural network model:

[0145] For the full-wave data samples input into the deep neural network model, the two-dimensional convolution layer Conv 2D and the global layer normalization gLN are used in sequence to perform data transformation;

[0146] The data obtained after the data transformation is processed by using two first deep neural network units and a second deep neural network unit in sequence;

[0147] The processed data is transformed and restored using a two-dimensional deconvolution layer Deconv 2D to obtain the upgoing wave data of the target channel in the time-frequency domain.

[0148] Optionally, the first deep neural network unit comprises: an expansion layer, a layer normalization, a bidirectional long short-term memory neural network, a one-dimensional deconvolution layer and a residual layer connected in sequence;

[0149] The second deep neural network unit includes: an attention mechanism module, a two-dimensional convolutional layer, a parameterized rectified linear unit, layer normalization along the channel and frequency dimensions, and a residual layer connected in sequence.

[0150] Optionally, the upgoing wave data of other channels in the adjacent multiple channels in the time-frequency domain are determined by the following equations:

[0151]

[0152] in, represents the upgoing wave data from the first channel of the plurality of adjacent channels at time t and frequency f The obtained stacked tensor, the first channel is the target channel, I and J represent the hyperparameters of the multi-frame Wiener filter; represents the upgoing wave data of the p2th channel in the plurality of adjacent channels at time t and frequency f, represents the filter corresponding to the p2th channel, the p2th channel is the other channels among the plurality of adjacent channels except the first channel, and C represents the number of the plurality of adjacent channels.

[0153] Optionally, the model inference module 830 is also used to convert the target full-wave data in the time domain from time domain data to time-frequency domain data through short-time Fourier transform, to obtain the target full-wave data in the time-frequency domain, and stack the real part and imaginary part of the target full-wave data in the time-frequency domain before inputting it into the target model.

[0154] Optionally, the model inference module 830 is further used to convert the target upgoing wave data in the time-frequency domain from time-frequency domain data to time domain data through inverse short-time Fourier transform to obtain the target upgoing wave data in the time domain.

[0155] It should be noted that the device embodiment is similar to the method embodiment, so the description is relatively simple, and the relevant parts can be referred to the method embodiment.

[0156] The present application also provides an electronic device, referring to Fig. 9 , Fig. 9 Schematic diagram of an electronic device proposed in an embodiment of the present application. Fig. 9 As shown, the electronic device 100 includes: a memory 110 and a processor 120. The memory 110 and the processor 120 are connected via a bus communication. A computer program is stored in the memory 110. The computer program can be run on the processor 120 to implement the steps in the vertical seismic profile wave field separation method disclosed in the embodiment of the present application.

[0157] The embodiment of the present application also provides a computer-readable storage medium on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the vertical seismic profile wavefield separation method disclosed in the embodiment of the present application is implemented.

[0158] The embodiment of the present application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the vertical seismic profile wavefield separation method disclosed in the embodiment of the present application.

[0159] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0160] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, systems, devices, storage media, and program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0162] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0163] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0164] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0165] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0166] The vertical seismic profile wave field separation method, device, equipment and medium provided by the present application are introduced in detail above. The principles and implementation methods of the present application are explained in this article using specific examples. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A vertical seismic profile wave field separation method, characterized in that: The method comprises: In the model training stage, a plurality of full-wave data samples in the time-frequency domain are obtained, each of the full-wave data samples comprising: full-wave data of a plurality of adjacent channels in the time-frequency domain, and a plurality of detectors at adjacent positions corresponding to the plurality of adjacent channels; In the model training stage, the training goal is to minimize the difference between the upgoing wave data of multiple adjacent channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain, and the deep neural network model is iteratively trained using the multiple full-wave data samples in the time-frequency domain to obtain a target model, wherein the deep neural network model is used to determine the upgoing wave data of the target channel in the multiple adjacent channels associated with it in the time-frequency domain based on the full-wave data samples, and the multi-frame Wiener filter is used to determine the upgoing wave data of other channels in the multiple adjacent channels associated with it in the time-frequency domain based on the upgoing wave data determined by the deep neural network model; In the model inference stage, the vertical seismic profile wave field array fed back by the detector is obtained to obtain the target full-wave data in the time domain. The target full-wave data in the time domain is converted from time domain data to time-frequency domain data and then input into the target model to obtain the target upgoing wave data in the time-frequency domain. The target upgoing wave data in the time-frequency domain is converted from time-frequency domain data to time domain data, and the target full-wave data in the time domain is subtracted from the target upgoing wave data in the time domain to obtain the target downgoing wave data in the time domain.

2. The method according to claim 1, characterized in that: The full-wave data samples are obtained by converting the full-wave data of a plurality of adjacent channels in the time domain from time domain data to time-frequency domain data; Before converting the full-wave data of the adjacent multiple channels in the time domain from time domain data to time-frequency domain data, the method further includes: Determine a dynamic time adjustment distance between first full-wave data and second full-wave data, wherein the first full-wave data and the second full-wave data are: two full-wave data offset in time among the full-wave data of the adjacent multiple channels in the time domain; Determine an alignment path according to the dynamic time adjustment distance, and use the alignment path to perform a leveling and alignment process on the first full-wave data and the second full-wave data; The dynamic time adjustment distance is determined based on the following formula: γ(i,j)=d(q i ,c j )+min[γ(i-1,j-1),γ(i-1,j),γ(i,j-1)} Wherein, γ(i,j) represents the cumulative distance from the first i points of the first full-wave data to the first j points of the second full-wave data. When i is equal to the length of the first full-wave data and j is equal to the length of the second full-wave data, γ(i,j) is the dynamic time adjustment distance; d(q i ,c j ) represents the distance between the i-th point of the first full-wave data and the j-th point of the second full-wave data.

3. The method according to claim 1, characterized in that The difference between the upgoing wave data of the adjacent multiple channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and the corresponding full-wave data samples in the time-frequency domain is measured by the target loss function, and the target loss function is expressed as follows: Among them, L WC Indicates the waveform consistency loss, Y p1 (t,f) represents the real full-wave data sample of the p1th channel at time t and frequency f. represents the upgoing wave data of the p1th channel at time t and frequency f determined based on the deep neural network model and the multi-frame Wiener filter, C represents the number of the multiple adjacent channels, Re(·) and Im(·) are used to extract the real part and the imaginary part, respectively, |·| is used to calculate the amplitude, and ‖·‖1 is used to calculate the L1 norm.

4. The method according to claim 1, characterized in that The upgoing wave data of the target channel in the time-frequency domain is obtained by executing the following steps by the deep neural network model: For the full-wave data samples input into the deep neural network model, the two-dimensional convolution layer Conv 2D and the global layer normalization gLN are used in sequence to perform data transformation; The data obtained after the data transformation is processed by using two first deep neural network units and a second deep neural network unit in sequence; The processed data is transformed and restored using a two-dimensional deconvolution layer Deconv 2D to obtain the upgoing wave data of the target channel in the time-frequency domain.

5. The method according to claim 4, characterized in that The first deep neural network unit comprises: an expansion layer, a layer normalization, a bidirectional long short-term memory neural network, a one-dimensional deconvolution layer and a residual layer connected in sequence; The second deep neural network unit includes: an attention mechanism module, a two-dimensional convolutional layer, a parameterized rectified linear unit, layer normalization along the channel and frequency dimensions, and a residual layer connected in sequence.

6. The method according to any one of claims 1 to 5, characterized in that: The upgoing wave data of other channels in the adjacent multiple channels in the time-frequency domain are determined by the following equations: in, represents the upgoing wave data from the first channel of the plurality of adjacent channels at time t and frequency f The obtained stacked tensor, the first channel is the target channel, I and J represent the hyperparameters of the multi-frame Wiener filter; represents the upgoing wave data of the p2th channel in the plurality of adjacent channels at time t and frequency f, represents the filter corresponding to the p2th channel, the p2th channel is the other channels among the plurality of adjacent channels except the first channel, and C represents the number of the plurality of adjacent channels.

7. The method according to any one of claims 1 to 5, characterized in that: The target full-wave data in the time domain is converted from time domain data to time-frequency domain data and then input into the target model, comprising: By short-time Fourier transform, the target full-wave data in the time domain is converted from time domain data to time-frequency domain data to obtain the target full-wave data in the time-frequency domain; After stacking the real part and the imaginary part of the target full-wave data in the time-frequency domain, the stacked parts are input into the target model; Converting the target upgoing wave data in the time-frequency domain from time-frequency domain data to time domain data comprises: The target upgoing wave data in the time-frequency domain is converted from time-frequency domain data to time-domain data by inverse short-time Fourier transform, so as to obtain the target upgoing wave data in the time domain.

8. A vertical seismic profile wave field separation device, characterized in that: The device comprises: The first training module is used to obtain a plurality of full-wave data samples in the time-frequency domain during the model training phase, each of the full-wave data samples comprising: full-wave data of a plurality of adjacent channels in the time-frequency domain, and a plurality of detectors at adjacent positions corresponding to the plurality of adjacent channels; A second training module is used for iteratively training the deep neural network model using the multiple full-wave data samples in the time-frequency domain to obtain a target model, with the goal of minimizing the difference between the upgoing wave data of multiple adjacent channels in the time-frequency domain determined by the deep neural network model and the multi-frame Wiener filter and their corresponding full-wave data samples in the time-frequency domain during the model training phase, wherein the deep neural network model is used to determine the upgoing wave data of a target channel in the multiple adjacent channels associated therewith in the time-frequency domain based on the full-wave data samples, and the multi-frame Wiener filter is used to determine the upgoing wave data of other channels in the multiple adjacent channels associated therewith in the time-frequency domain based on the upgoing wave data determined by the deep neural network model; A model inference module is used to obtain the vertical seismic profile wave field array fed back by the detector to obtain target full-wave data in the time domain, convert the target full-wave data in the time domain from time domain data to time-frequency domain data and then input it into the target model to obtain target upgoing wave data in the time-frequency domain, convert the target upgoing wave data in the time-frequency domain from time-frequency domain data to time domain data, and use the target full-wave data in the time domain to subtract the target upgoing wave data in the time domain to obtain target downgoing wave data in the time domain.

9. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the vertical seismic profile wavefield separation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the vertical seismic profile wavefield separation method according to any one of claims 1 to 7 is implemented.

Images