A data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile

Abstract

The application discloses a kind of vertical seismic profile up-and-down wave intelligent separation methods based on data driving, comprising the following steps: S1: collection well logging data, including velocity information and density information;S2: establish formation model and define seismic observation system, then based on vertical seismic profile wave field propagation equation numerical simulation, obtain simulated up-and-down wave data, and combined to obtain simulated total wave field data;S3: establish the neural network of multi-task learning, according to the data obtained in step S2 the neural network is trained, and obtains regression model;S4: carry out feasibility test, when the result is not up to standard, adjust training sample and optimize network parameter, return step S3 and retrain;When the result is up to standard, the regression model obtained in step S3 is used as the final separation model, and the target vertical seismic profile data is separated into up-and-down wave using it.The application can more intelligently, efficiently and accurately separate the up-and-down wave of vertical seismic profile data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wave field separation technology, and particularly relates to an intelligent separation method for upgoing and downgoing waves on a vertical seismic profile based on data driving. BACKGROUND

[0002] A vertical seismic profile (VSP) is a method for observing a seismic wave field by using a borehole geophone or a downhole source, and has the advantages of less interference and high resolution compared with conventional surface seismic technology. However, the wave field data received by the vertical seismic profile technology is a composite signal, which not only has downgoing waves (ground reverse waves) propagating downward, but also has upgoing waves (stratum reflected waves) propagating upward. The upgoing and downgoing waves interfere with each other and alias, so that the entire wave field signal becomes complex and disordered. Therefore, the original wave field needs to be accurately separated. The separation of upgoing and downgoing waves is the most critical and basic link of the vertical seismic profile data processing, and the separation quality determines the imaging accuracy and the rationality of further interpretation.

[0003] Most conventional separation methods for upgoing and downgoing waves of a vertical seismic profile are based on the difference in wave field apparent velocity and polarization direction filtering, and can be divided into separation methods based on time domain filtering or transform domain filtering. The separation method based on time domain filtering needs to perform first arrival picking and static time shift, otherwise the wave field energy will be dispersed, and high-precision separation requires time-varying filtering windows, which greatly increases the calculation amount. The separation method based on transform domain filtering can quickly realize separation, but it is difficult to overcome the "false axis" caused by the Fourier transform mechanism, so tedious data preprocessing or subsequent processing is inevitable. At the same time, due to the difference in technical adaptability of different data of the existing separation methods, geological prior information is often needed, and subjective manual parameter adjustment is involved. This not only makes the technical scheme process cumbersome, but also affects the accuracy of the separation result due to subjective parameter adjustment, and it is difficult to realize automatic, efficient and high-precision wave field separation. SUMMARY

[0004] In view of the above problems, the present application aims to provide an intelligent separation method for upgoing and downgoing waves of a vertical seismic profile based on data driving, which realizes intelligent separation of wave field in a data-driven manner by modeling data with real logging data and training a multi-task learning neural network to obtain a regression model.

[0005] The technical scheme of the present application is as follows:

[0006] An intelligent separation method for upgoing and downgoing waves of a vertical seismic profile based on data driving, comprising the following steps:

[0007] S1: Collect logging data, wherein the logging data includes velocity information and density information;

[0008] S2: establishing a formation model and defining a seismic observation system according to the logging data, then performing numerical simulation based on a vertical seismic profile wave field propagation equation to obtain simulated upgoing and downgoing wave data, and combining the simulated upgoing and downgoing wave data to obtain simulated total wave field data;

[0009] S3: establishing a multi-task learning neural network, taking the simulated total wave field data as data input and the simulated upgoing and downgoing wave data as data output, training the multi-task learning neural network to obtain a regression model;

[0010] S4: performing a feasibility test on the regression model,

[0011] when the result of the feasibility test is not up to standard, adjusting the training sample of the multi-task learning neural network and optimizing network parameters, and returning to step S3 for retraining;

[0012] when the result of the feasibility test is up to standard, taking the regression model obtained in step S3 as a final separation model, and using the final separation model to perform upgoing and downgoing wave separation on target vertical seismic profile data.

[0013] Preferably, in step S2, the vertical seismic profile wave field propagation equation is:

[0014]

[0015]

[0016] wherein U n and U n+1 are upgoing wave data of the nth layer and the nth+1 layer, respectively; D n and D n+1 are downgoing wave data of the nth layer and the nth+1 layer, respectively; R n is a reflection coefficient of the nth layer of underground medium; P n is a forward propagation action of the nth layer of underground medium; P n -1 is a backward propagation action of the nth layer of underground medium.

[0017] Preferably, when the absorption and attenuation effect of the formation is considered, the forward propagation action P n of the nth layer of underground medium is calculated by the following formula:

[0018]

[0019] wherein e is a natural base; f is frequency; h n is layer thickness; Q n is quality factor; vn V is the velocity; i is the imaginary number.

[0020] As a preference, the quality factor Q n calculated by the following formula:

[0021]

[0022] In the formula: Q v,n is the velocity-dependent quality factor; Q ρ,n is the density-dependent quality factor;

[0023] Q v,n and Q ρ,n are calculated by the following formula:

[0024]

[0025] In the formula: Q ep,n is the quality factor related to the elastic parameter in the nth layer of the underground medium; Q1 is the quality factor of the first layer of the underground medium; ep n is the elastic parameter of the nth layer; ep0 is the elastic parameter at the free surface; Q0 is the quality factor at the free surface; ep1 is the elastic parameter of the first layer.

[0026] As a preference, in step S2, when numerically simulating based on the wave field propagation equation of the vertical seismic section, the following sub-steps are specifically included:

[0027] S21: defining a layer matrix M n :

[0028]

[0029] S22: calculating the upgoing wave data and the downgoing wave data in the whole stratum propagation process according to the layer matrix M n :

[0030]

[0031] In step S2, when combining the upgoing wave data and the downgoing wave data to obtain the total wave field data, the total wave field data is calculated by the following formula:

[0032] W = U + D (8)

[0033] In the formula: W is the total wave field data; U is the upgoing wave data; D is the downgoing wave data.

[0034] As preferred, in step S3, the multi-task learning neural network comprises eight convolutional layers, wherein the first three convolutional layers are used to build the encoding layer of the neural network, and the receptive field of the neural network is enlarged by using the max pooling layer in the encoding process; the fourth convolutional layer is the intermediate layer of the neural network; the last four convolutional layers are used to build the decoding process of the neural network, and the inverse process of the pooling is simulated by using the up-sampling layer which is symmetrical to the pooling layer in the decoding process; the first six convolutional layers are activated by using the activation function one, and the seventh convolutional layer and the eighth convolutional layer are multi-task shunted and data exported by using the activation function two.

[0035] As preferred, each convolutional layer is a 2D convolution operation with a 3*3 kernel, and is matched with batch normalization BN; the max pooling layer adopts a 2*2 kernel.

[0036] As preferred, the activation function one adopts the ReLU function, and the activation function two adopts the Tanh function.

[0037] As preferred, in step S1, when collecting the logging data, logging data from different geological backgrounds and different reservoir types are collected.

[0038] The beneficial effects of the present application are:

[0039] The present application can more intelligently, efficiently and accurately separate the up-going and down-going waves of the vertical seismic profile data by using logging data for data modeling and combining the multi-task data driving of the neural network. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0041] Figure 1 The flowchart of the present application based on data-driven intelligent separation method of up-going and down-going waves of vertical seismic profile;

[0042] Figure 2 The structure diagram of the multi-task learning neural network of one specific embodiment;

[0043] Figure 3 The data modeling diagram of one specific embodiment of the Ordos Basin well; wherein, Figure 3 (a) is the well curve needed for data modeling, Figure 3 (b) is the formation model built based on the well curve and the defined observation system;

[0044] Figure 4 Fig. 1 shows a schematic diagram of modeling results of data of a well in Ordos Basin for an embodiment; wherein, Figure 4 (a) is simulated upgoing wave data, Figure 4 (b) is simulated downgoing wave data, Figure 4 (c) is total wavefield data obtained by stacking and combining;

[0045] Figure 5 Fig. 2 shows a schematic diagram of performance results of separating verification set data by using a regression model for an embodiment; wherein, Figure 5 (a) is verification set data, Figure 5 (b) is reference upgoing wave, Figure 5 (c) is reference downgoing wave, Figure 5 (d) is upgoing wave obtained by separating according to the present application, Figure 5 (e) is downgoing wave obtained by separating according to the present application;

[0046] Figure 6 Fig. 3 shows a schematic diagram of frequency wave number spectrum analysis results of separating results of verification set data as shown in Fig. 2; wherein, Figure 5 (a) is F-K spectrum of reference upgoing wave data, Figure 6 (b) is F-K spectrum of upgoing wave data separated according to the present application, Figure 6 (c) is F-K spectrum of reference downgoing wave data, Figure 6 (d) is F-K spectrum of downgoing wave data separated according to the present application; Figure 6

[0047] Fig. 4 shows a schematic diagram of separating results of vertical seismic profile data of a practical work area in Sichuan Basin for an embodiment; wherein, Figure 7 (a) is vertical seismic profile data of a practical work area in Sichuan Basin, Figure 7 (b) is upgoing wave result obtained by separating according to the present application, Figure 7 (c) is downgoing wave result obtained by separating according to the present application. Figure 7 DETAILED DESCRIPTION The present application will be further described below in conjunction with the drawings and embodiments. It should be noted that the embodiments in the present application and the technical features in the embodiments can be combined with each other without conflict. It should be noted that all the technical and scientific terms used in the present application have the same meaning as that generally understood by the ordinary skilled in the art to which the present application belongs, unless otherwise specified. The similar words such as "comprise" or "contain" and the like used in the present application mean that the elements or objects before the words cover the elements or objects listed after the words and their equivalents, and do not exclude other elements or objects.

[0048] As

[0049] Figure 1 ​As shown, the present application provides a data-driven intelligent separation method for upgoing and downgoing waves on a vertical seismic profile, comprising the following steps:

[0050] S1: collecting logging data, wherein the logging data comprises velocity information and density information;

[0051] In one specific embodiment, when collecting logging data, logging data from different geological backgrounds and different reservoir types is collected. In this embodiment, by collecting logging data from different geological backgrounds and different reservoir types, the regression model obtained by subsequent training can be applied to the separation of upgoing and downgoing waves of vertical seismic profile data in various different work areas.

[0052] It should be noted that when using the present application, logging data from the same or similar work area as the target vertical seismic profile work area can also be collected according to the geological background and reservoir type of the target vertical seismic profile data. In this way, the regression model obtained by training can be more accurate when separating the target vertical seismic profile data.

[0053] In addition, when collecting logging data, the logging data can be preprocessed as needed to delete invalid values in the collected data. The method for identifying invalid values is a prior art, which will not be described here.

[0054] S2: establishing a formation model and defining a seismic observation system according to the logging data, then performing numerical simulation based on the vertical seismic profile wave field propagation equation to obtain simulated upgoing and downgoing wave data, and combining the simulated upgoing and downgoing wave data to obtain simulated total wave field data;

[0055] In one specific embodiment, the vertical seismic profile wave field propagation equation is:

[0056]

[0057]

[0058] wherein U n , U n+1 are the upgoing wave data of the nth layer and the nth+1 layer respectively; D n , D n+1 are the downgoing wave data of the nth layer and the nth+1 layer respectively; R n is the reflection coefficient of the nth layer of underground medium; P n is the forward propagation effect of the nth layer of underground medium; P n -1 is the backward propagation effect of the nth layer of underground medium.

[0059] Optionally, when considering the absorption and attenuation effect of the formation, the forward propagation effect Pn The calculation is made by the following formula:

[0060]

[0061] In the formula, e is the natural base number; f is the frequency; h n is the layer thickness; Q n is the quality factor; v n is the velocity; and i is the imaginary number.

[0062] In the above embodiment, the result obtained by the present application is more in line with the actual geological conditions and more accurate by considering the absorption attenuation effect of the stratum.

[0063] Optionally, the quality factor Q n is calculated by the following formula:

[0064]

[0065] In the formula, Q v,n is the velocity-dependent quality factor; Q ρ,n is the density-dependent quality factor;

[0066] Q v,n and Q ρ,n are calculated by the following formula:

[0067]

[0068] In the formula, Q ep,n is the quality factor related to the elastic parameters (i.e. velocity v and density p) in the nth layer of the underground medium; Q1 is the quality factor of the first layer of the underground medium; ep n is the elastic parameter of the nth layer; ep0 is the elastic parameter at the free surface; Q0 is the quality factor at the free surface; and ep1 is the elastic parameter of the first layer.

[0069] In one specific embodiment, when the numerical simulation is based on the wave field propagation equation of the vertical seismic profile, the following sub-steps are specifically included:

[0070] S21: defining a layer matrix M n that includes reflection and attenuation physical effects:

[0071]

[0072] S22: calculating the upgoing wave data and downgoing wave data in the whole stratum propagation process according to the layer matrix M n :

[0073]

[0074] In step S2, the upgoing wave data and the downgoing wave data are combined to obtain total wave field data, and the total wave field data is calculated by the following formula:

[0075] W = U + D (8)

[0076] In the formula, W is total wave field data, U is upgoing wave data, and D is downgoing wave data.

[0077] S3: A multi-task learning neural network is established, the simulated total wave field data is used as data input, the simulated upgoing wave data and downgoing wave data are used as data output, the multi-task learning neural network is trained, and a regression model is obtained.

[0078] In a specific embodiment, as shown in the figure, Figure 2 the multi-task learning neural network includes eight convolutional layers, the first three convolutional layers are used to build an encoding layer of the neural network, and a maximum pooling layer is used in the encoding process to expand the receptive field of the neural network; the fourth convolutional layer is an intermediate layer of the neural network; the last four convolutional layers are used to build a decoding process of the neural network, and an up-sampling layer symmetric to the pooling layer is used in the decoding process to simulate the inverse process of the pooling; the first six convolutional layers are activated using an activation function one, the seventh convolutional layer and the eighth convolutional layer are multi-task shunted, and data is exported using an activation function two.

[0079] Optionally, each convolutional layer is a 2D convolution operation with a 3*3 kernel, and is matched with batch normalization BN; and the maximum pooling layer adopts a 2*2 kernel. In this embodiment, the batch normalization BN can accelerate the convergence speed of the neural network, the maximum pooling layer can expand the receptive field of the neural network, reduce the information redundancy introduced by the convolution operation, and improve the performance of the neural network.

[0080] Optionally, the activation function one adopts a ReLU function, and the activation function two adopts a Tanh function. In this embodiment, the expressions of the two activation functions are respectively:

[0081] ReLU(x) = max(0, x) (9)

[0082]

[0083] It should be noted that the selection of the activation function is a strategic problem in deep learning. In this embodiment, the ReLU function capable of enhancing the nonlinear expression ability of the network is used as the activation function one, and the Tanh function with a value range more suitable for seismic data export is used as the activation function two. The two activation functions jointly act on the activation expression of the neural network, and can achieve higher quality upgoing wave and downgoing wave output effect. It should be noted that when the present application is used, other activation functions in the prior art can also be selected according to the accuracy and other requirements.

[0084] In the present application, the neural network adopted is a multi-task learning neural network, which can avoid separate training and thus realize more efficient and intelligent separation of upgoing and downgoing waves. It should be noted that in addition to the multi-task learning neural network adopted in the above-mentioned embodiments, other multi-task learning networks in the prior art can also be used for training, and the regression model obtained by training can also realize intelligent separation of upgoing and downgoing waves of the target vertical seismic profile data; only the separation accuracy of the existing other multi-task learning networks does not reach the separation accuracy of the multi-task learning neural network described in the present application, and when used, the separation accuracy requirement can be selected, and as for intelligent separation, the other multi-task learning networks in the prior art can also be applicable to the present application.

[0085] When the multi-task learning neural network is trained by using the data obtained in step S2, the intelligent separation of upgoing and downgoing waves can be regarded as a deep learning multi-task mapping:

[0086]

[0087] In the formula: X is the input of the neural network; F i is the i-th mapping relationship to be solved; Y i is the i-th output task component; and m is the number of tasks.

[0088] In the positive process of solving F, the single-component task i performed by the neural network Net can be expressed as:

[0089]

[0090] For the upgoing and downgoing wave data features extracted by the neural network training, they can be approximately expressed as:

[0091] y = BN [f (x * w + b)] (13)

[0092] In the formula: y is the output feature; BN is the batch normalization operation; f is the activation function; x is the input data; w is the network weight; and b is the bias unit.

[0093] S4: performing a feasibility test on the regression model,

[0094] When the result of the feasibility test does not meet the standard, the training samples of the multi-task learning neural network are adjusted, the network parameters are optimized, and step S3 is returned to retrain;

[0095] When the result of the feasibility test meets the standard, the regression model obtained in step S3 is used as the final separation model, and the final separation model is used to separate the upgoing and downgoing waves of the target vertical seismic profile data.

[0096] It should be noted that whether the result of the feasibility test meets the standard is determined according to a threshold set by a person. In a specific embodiment, the feasibility of the regression model is determined by structural similarity (SSIM) and F-K spectrum, and optionally, when the SSIM value between the prediction result and the reference result is greater than or equal to 0.9, it is considered that the separation of the (model data) is effective; the F-K spectrum can further evaluate the amplitude preservation of the separation result, and the more similar the F-K spectrum of the prediction result is to the reference spectrum, the better the separation effect is.

[0097] In a specific embodiment, the data-driven up-and-down wave intelligent separation method based on vertical seismic profile is used to separate the up-and-down wave of the vertical seismic profile data of a practical work area in Sichuan Basin, which specifically includes the following steps:

[0098] Logging data is collected. In this embodiment, the collected logging data includes 10 logging data from Ordos Basin, Qiongzhou Southeast Basin and the like. The collected logging data is sorted, the P-wave velocity and density curves are selected, the units are unified to the meter system, the invalid values in the data are removed, the quality factor is converted according to the formula (4)-(5), and the logging data is divided into several strata according to the thickness of 1 m, wherein the velocity, density and other parameters of each stratum are the average values of the logging parameters of the layer.

[0099] In this embodiment, the data modeling of the target layer section of 500-1000 m of a well in Ordos Basin is as shown in Figure 3 , wherein Figure 3 (a) is the P-wave velocity, density and converted quality factor curve of the well section, Figure 3 (b) is a stratum model established according to the information shown in Figure 3 (a), wherein the background of the stratum model is filled with P-wave velocity.

[0100] Based on the stratum model shown in Figure 3 (b), a seismic observation system is defined, the wellhead and the source geographical point overlap, which is used to simulate ideal zero-offset vertical seismic profile data, the receivers are placed equidistantly in the well, the interval is set to 10 m, the receiver sampling rate is 2 milliseconds, and the sampling time is 0.3 seconds to 1.5 seconds.

[0101] Based on the stratum model shown in Figure 3 (b) and the defined observation system, numerical simulation is performed by using the vertical seismic profile wave field propagation equation shown in the formula (1)-(2), and the simulation result is as shown in Figure 4 , wherein Figure 4 the longitudinal and transverse coordinates in Figure 3 (b) are opposite to the longitudinal and transverse coordinates of the stratum model shown in Figure 3 (b), the transverse coordinate of the simulation data is the depth of the receiver, and the longitudinal coordinate is the wave field propagation time; wherein,Figure 4 (a) shows the obtained simulated upflow wave data. Figure 4 (b) The obtained simulated downwave data. Figure 4 (c) is Figure 4 (a) Simulated upwave data and Figure 4 (b) The simulated total wavefield data obtained by superimposing and combining the simulated downflow wave data can be regarded as the ideal vertical seismic profile wavefield signal of a target section of a well in the Ordos Basin.

[0102] Using the above method, data modeling was carried out on the remaining 9 wells to simulate different types of uplink and downlink wave data. The data obtained from these 10 wells were used to construct an independent and complete training set, giving it the advantages of being physically driven and geologically significant.

[0103] The dataset obtained from the above simulation is divided into 3×38412 64×64 training blocks. Multi-task deep learning will be performed on the network in the form of training blocks. In this embodiment, the multi-task learning neural network used is as follows: Figure 2 As shown, the multi-task learning neural network consists of a total of eight basic convolutional layers. The first three basic convolutional layers are used to construct the network's encoding layer, the fourth layer is the network's intermediate layer, the last four basic convolutional layers are used to construct the network's decoding process, and multi-task splitting is performed in the last two convolutional layers.

[0104] In this embodiment, 100 iterative training cycles were set to obtain the regression model. During this process, the mean squared error loss function was changed from 3.3e -2 Reduced to 1.5e -6 This indicates that the training was effective.

[0105] A validation set was simulated and modified by adding operations such as two-dimensional Gaussian deformation to differentiate it from the training set data. This was used to test the separation performance of the regression model described in this invention. The results are as follows: Figure 5 As shown; where, Figure 5 (a) is the validation set data, which is Figure 5 (b) shows the reference upward wave and Figure 5 (c) shows the total wavefield data obtained by superimposing the reference downwave. Figure 5 (b) is the reference upward wave. Figure 5 (c) is the reference downward wave. Figure 5 (d) is the upward wave obtained by the present invention. Figure 5 (e) is the downlink wave obtained by the present invention.

[0106] contrast Figure 5 (b) and Figure 5 (d) and comparison Figure 5 (c) and Figure 5(e)It can be seen that there is little difference between the separation results of the application and the reference results. The structural similarity (SSIM) analysis is respectively performed on different wave field results, and the SSIM value between the separated upgoing wave of the application and the reference upgoing wave is 0.94, and the SSIM value of the downgoing wave is 0.97, which indicates that the similarity between the ideal results and the actual separation results is good, and there is no obvious image distortion.

[0107] Further frequency-wavenumber (F-K) spectrum analysis is carried out on the separation results, and the results are as shown in Figure 6 wherein, Figure 6 (a) and Figure 6 (b) are F-K spectra of the reference upgoing wave data and the separated upgoing wave data of the application respectively, Figure 6 (c) and Figure 6 (d) are F-K spectra of the reference downgoing wave data and the separated downgoing wave data of the application respectively. It can be seen from Figure 6 that there is no obvious energy loss in the F-K spectrum, indicating that the amplitude preservation of the separation method of the application is good.

[0108] According to the separation performance test of the above verification set, it can be known that the regression model used in the embodiment has good separation performance, and passes the technical feasibility verification, and is used as the final separation model of the application.

[0109] The final separation model obtained by the above steps is used for upgoing and downgoing wave separation test on the vertical seismic profile data of a practical work area in Sichuan Basin, and the results are as shown in Figure 7 wherein Figure 7 (a) is the vertical seismic profile data of a practical work area in Sichuan Basin, Figure 7 (b) is the upgoing wave result separated by the application, Figure 7 (c) is the downgoing wave result separated by the application. It can be seen from Figure 7 that compared with the model data, the real vertical seismic profile data has more complex and more disordered wave field characteristics, and the upgoing wave energy is severely suppressed by the downgoing wave, which increases the difficulty of accurate separation of the two wave fields; however, the application can still well separate the upgoing wave and the downgoing wave, and the application avoids the cumbersome data processing flow and does not need subjective threshold selection, thereby providing a new way for efficient and intelligent separation of upgoing and downgoing wave fields of vertical seismic profile data.

[0110] In summary, the application can more efficiently, intelligently and accurately separate the upgoing and downgoing waves of the vertical seismic profile data. Compared with the prior art, the application has significant progress.

[0111] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in any form. Although the present application has been described above with the preferred embodiments, it is not intended to limit the present application, and any person skilled in the art can make some changes or modifications to the above disclosed technical contents without departing from the technical solution of the present application, and any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application still belong to the scope of the technical solution of the present application.

Claims

1. A data-driven based intelligent separation of up-going and down-going waves on vertical seismic profile, characterized in that, Comprising the following steps: S1: collecting logging data, the logging data comprising velocity information and density information; S2: establishing a formation model and defining a seismic observation system according to the logging data, then performing numerical simulation based on a vertical seismic profile wave field propagation equation to obtain simulated upgoing and downgoing wave data, and combining the simulated upgoing and downgoing wave data to obtain simulated total wave field data; When performing numerical simulation based on the vertical seismic profile wave field propagation equation, the following sub-steps are included: S21 : Defining a layer matrix M including reflection and attenuation physical effects n : （6） where: R n is the reflection coefficient of the nth layer of subsurface media; P n is the forward propagating action of the nth layer of subsurface media; P n -1 is the backward propagating action of the nth layer of subsurface media; S22: According to the layer matrix M n Calculate upgoing and downgoing wave data during the propagation through the entire formation: （7） In the formula, D1, D n+1 respectively, are down wave data of the 1st layer and the n+1th layer; U1, U n+1 respectively, are up wave data of the 1st layer and the n+1th layer; When the upgoing and downgoing wave data are combined to obtain total wave field data, the total wave field data is calculated by the following formula: （8） In the formula, W is the total wave field data; U is the upgoing wave data; D is the downgoing wave data; S3: establishing a multi-task learning neural network, taking the simulated total wave field data as data input and the simulated upgoing and downgoing wave data as data output, training the multi-task learning neural network to obtain a regression model; S4: performing a feasibility test on the regression model, When the result of the feasibility test is not up to standard, adjusting the training samples of the multi-task learning neural network and optimizing the network parameters, and returning to step S3 for retraining; When the result of the feasibility test is up to standard, taking the regression model obtained in step S3 as the final separation model, and using the final separation model to perform upgoing and downgoing wave separation on target vertical seismic profile data.

2. The data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile according to claim 1, characterized in that, In step S2, the vertical seismic profile wave field propagation equation is: （1） （2） wherein: U n is the upgoing wave data for the nth layer; D n is the downgoing wave data for the nth layer.

3. The data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile of claim 2, wherein, When considering the absorption and attenuation effect of the strata, the forward propagation effect P of the nth subsurface medium n The calculation is performed using the following formula: （3） where: e is the natural base; f is the frequency; h n is the layer thickness; Q n is the quality factor; v n is the velocity; i is the imaginary number.

4. The data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile of claim 3, wherein, the quality factor Q n calculated by the formula: （4） where: Q v,n Q is the speed dependent quality factor; Q ρ,n Q is the density dependent quality factor; Q v,n and Q ρ,n are each calculated by the following formula: （5） where: Q ep,n is the quality factor associated with the elastic parameter in the nth layer of subsurface medium; Q1 is the quality factor of the first layer of subsurface medium; ep n is the elastic parameter of the nth layer; ep0 is the elastic parameter at the free surface; Q0 is the quality factor at the free surface; ep1 is the elastic parameter of the first layer.

5. The data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile according to claim 1, characterized in that, In step S3, the multi-task learning neural network comprises eight convolutional layers, of which the first three convolutional layers are used to build the encoding layer of the neural network, and a max-pooling layer is used in the encoding process to expand the receptive field of the neural network; the fourth convolutional layer is the intermediate layer of the neural network; the last four convolutional layers are used to build the decoding process of the neural network, and an up-sampling layer symmetric to the pooling layer is used in the decoding process to simulate the inverse process of the pooling; The first six convolutional layers are activated using activation function one, the seventh convolutional layer and the eighth convolutional layer are multi-task shunted, and data is exported using activation function two.

6. The data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile according to claim 5, characterized in that, Each convolutional layer is a 2D convolution operation with a 3*3 kernel, and is matched with batch normalization BN; the max-pooling layer uses a 2*2 kernel.

7. The data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile according to claim 5 or 6, characterized in that, The activation function one uses a ReLU function, and the activation function two uses a Tanh function.

8. The data-driven based up-going and down-going wave intelligent separation method on vertical seismic profile according to any one of claims 1-7, characterized in that, In step S1, when collecting logging data, logging data from different geological backgrounds and different reservoir types is collected.

Images