阅读说明：本技术 一种基于深度学习照明分析的局部偏移成像方法 (Local offset imaging method based on deep learning illumination analysis ) 是由 荣超 贾晓峰 于 2020-12-21 设计创作，主要内容包括：本发明涉及一种基于深度学习照明分析的局部偏移成像方法,包括如下步骤：步骤1)构建二维地质速度模型,利用传统照明分析方法得到所述二维地质速度模型对应单炮照明结果；步骤2)基于地质速度模型与对应单炮照明结果构建训练数据集,输入构建的Unet神经网络进行训练；步骤3)利用训练好的神经网络预测单炮照明结果,从单炮叠加照明图看出地质速度模型中照明能量的具体分布情况,手动选择像素值小于预定阈值的区域,将该区域定义为弱照明区域；步骤4)依据单炮在弱照明区域照明强度,按照贡献大小筛选出炮集；步骤5)将筛选出来的炮集与全部炮集依据正传波场与检波器反传波场进行偏移成像；最后截取弱照明区域成像结果。(The invention relates to a local offset imaging method based on deep learning illumination analysis, which comprises the following steps: step 1), constructing a two-dimensional geological speed model, and obtaining a single-shot illumination result corresponding to the two-dimensional geological speed model by using a traditional illumination analysis method; step 2) constructing a training data set based on the geological speed model and the corresponding single-shot illumination result, and inputting the constructed Unet neural network for training; step 3) predicting a single shot illumination result by using the trained neural network, finding out the specific distribution condition of illumination energy in the geological speed model from the single shot superposition illumination map, manually selecting a region with a pixel value smaller than a preset threshold value, and defining the region as a weak illumination region; step 4) screening shot sets according to the contribution size and the illumination intensity of the single shot in the weak illumination area; step 5) carrying out migration imaging on the screened shot sets and all shot sets according to a forward transmission wave field and a backward transmission wave field of the detector; and finally intercepting the imaging result of the weak illumination area.)

1. A local offset imaging method based on deep learning illumination analysis is characterized by comprising the following steps:

step 1), constructing a two-dimensional geological speed model, and obtaining a single-shot illumination result corresponding to the two-dimensional geological speed model by using a traditional illumination analysis method;

step 2) constructing a training data set based on the geological speed model and the corresponding single-shot illumination result, and inputting the constructed Unet neural network for training;

step 3) predicting a single shot illumination result by using the trained neural network, finding out the specific distribution condition of illumination energy in the geological speed model from the single shot superposition illumination map, manually selecting a region with a pixel value smaller than a preset threshold value, and defining the region as a weak illumination region;

step 4) screening shot sets according to the contribution size and the illumination intensity of the single shot in the weak illumination area;

step 5) carrying out migration imaging on the screened shot sets and all shot sets according to a forward transmission wave field and a backward transmission wave field of the detector; and finally intercepting the imaging result of the weak illumination area.

2. The local offset imaging method based on deep learning illumination analysis according to claim 1, wherein the step (1) constructs a two-dimensional geological velocity model, the illumination analysis is to write matlab program to randomly generate some underground velocity structure models for underground complex geological structures, and inclined stratums, folds, faults and high-speed abnormal body structures are added into the models to ensure the richness of the models.

3. The local offset imaging method based on deep learning illumination analysis according to claim 1, wherein the step (1) obtains the illumination result of the model corresponding to the single shot by using a traditional illumination analysis method, and the local offset imaging method based on the two-dimensional constant-density time domain acoustic wave equation is as follows:

where v (x, z) is the seismic velocity in space (x, z), u (x, z, t) is the wavefield in space (x, z) at time t, and s (t) is the seismic source at time t;

defining the single-shot source illumination intensity of one point in the space as follows:

whereinThe source illumination intensity of the ith shot;

the illumination intensity of the N-shot seismic sources is the sum of the illumination intensities of all the single-shot seismic sources:

4. the local migration imaging method based on deep learning illumination analysis according to claim 1, wherein the step (2) is to construct a training data set based on a geological velocity model and a corresponding single shot illumination result, add the location information of shot points into the geological velocity model and store the location information in an RGB picture, and obtain an input data set of a neural network, and the method comprises the following steps: adding a point Gaussian function at the position corresponding to the shot point to simulate the situation of the shot point; for the location of the shot point (x)₀,y₀) Then, adding a two-dimensional gaussian function to the model is:

where (x, y) is the model midpoint coordinate, and c1 and c2 define the variance in the x and y directions.

5. The local offset imaging method based on deep learning illumination analysis according to claim 1, wherein in the step (2), a training data set is constructed based on a geological speed model and a corresponding single shot illumination result, the geological model single shot illumination result is converted into an RGB picture to be stored, and a label data set of a neural network is obtained; and based on the traditional Unet network framework, adjusting network parameters to obtain the network structure of the project, and loading the input training set and the label data set into the neural network for training.

6. The local offset imaging method based on deep learning illumination analysis according to claim 1, wherein the step (3) superimposes the single-shot prediction results of the neural network to obtain a multi-shot energy illumination map, the energy intensity of each region is seen from the illumination map, the superimposed illumination map is read by means of a matlab program, and then the size of the weak illumination region is manually selected.

7. The local offset imaging method based on deep learning illumination analysis according to claim 1, wherein the step (4) screens shot gathers according to contribution size according to illumination intensity of single shots in weak illumination areas; and averaging the weak illumination areas according to the superimposed illumination map to obtain average illumination, averaging the illumination results of each single shot in the weak illumination areas, and if the average value is greater than the average illumination, determining the shot as a screened shot set.

8. The local migration imaging method based on deep learning illumination analysis as claimed in claim 1, wherein the step (5) is to perform migration imaging on the screened shot gathers and all shot gathers according to forward wave field and backward wave field of the detector, finally intercept the imaging result of the weak illumination area, record only the forward wave field of the upper boundary of the weak illumination area, and then apply the backward wave field of the detector to perform imaging at each point in space by using cross-correlation imaging conditions; wherein the cross-correlation imaging condition is formulated as,

wherein r (x, z, t) is a backward wave field of the detector at the time t at the space point (x, z), and s (x, z, t) is a forward wave field of the seismic source at the time t at the space point (x, z); image (x, z) is the imaging result at space (x, z).

Technical Field

The invention realizes the one-way illumination of the underground medium by using a deep learning method, particularly screens out the cannons contributing to the target area according to the energy distribution of the single cannon, and realizes the offset imaging of the target area by using fewer cannons.

Background

Geophysical exploration faces the problem of increasingly complex exploration targets as the degree of exploration and production increases. Seismic exploration is the core technology of oil and gas exploration, and at present, with the increase of exploration depth, the more complex exploration area structure and the development of offshore exploration from shallow sea to deep sea, the seismic exploration technology and the processing method must be improved. Migration imaging is a key step in seismic data processing and has been the core of seismic processing algorithm research. From the initial manual shift to the present reverse time shift algorithm, a number of different shift algorithms are continually being proposed and improved. The steps of the migration algorithm from migration are mainly divided into prestack and poststack migration, and the migration can be divided into time migration and depth migration according to the time or depth as the longitudinal coordinate in the obtained migration profile. Migration methods commonly used in the industry, based on the basic principles of the algorithm, can be classified into Kirchhoff integration and wave equation migration, both of which can be used for prestack and poststack migration as well as time migration and depth migration. The Kirchoff integration method is based on the principle of geometric optics to solve the high-frequency progressive solution of the wave equation. The method utilizes high-frequency approximation of a ray theory, and solves the analytic solution of the wave equation by solving the Green function of seismic waves propagated underground after the shot source is excited. The wave equation migration algorithm mainly comprises a one-way migration method and a reverse time migration method, wherein the two methods are used for simulating the propagation process of seismic waves in the underground medium by directly solving a wave equation. The wave equation method can be further classified into a finite difference method, a fourier transform method, and a mixed domain method according to the difference of the solving method. Yoon and the like realize three-dimensional sound wave reverse time migration; zhang et al propose a strategy for achieving true amplitude reverse time migration; soubaras and the like provide reverse time migration imaging by a two-step display matching method, so that the time step of wave field extrapolation can be increased; zhang et al and Huang et al achieve stable high quality reverse time migration imaging of TTI media; zhang et al derive the time domain viscous acoustic wave equation and achieve reverse time migration imaging with Q-compensation. Many scholars in China also carry out a great deal of research on reverse time migration, and research methods comprise a finite element method, a finite difference method, a spectral element method, a mixing method and the like. The doherty excitation method applies a reverse time migration method to anisotropic media and elastic wave multi-component data to obtain a good application effect; liuhong Wei and the like analyze the denoising technology and the storage problem of pre-stack reverse time migration, and realize the high-order finite difference method reverse time migration GPU acceleration algorithm.

The seismic illumination analysis utilizes a seismic forward modeling method facing geological targets, an observation system and a subsurface geological structure are known, and the propagation distribution situation of seismic wave energy in a subsurface complex structure is known and researched. In data acquisition, seismic lighting can assist the design of an observation system (Hoffmann, 2001), in data processing, the result of lighting energy analysis is used for assisting data regularization, processing missing data and irregular shot-geophone point distribution in actual data, and compensating a weakened lighting energy shadow part, so that the application value of the system is immeasurable. The illumination operator gradually develops from a one-way wave forward operator and a wave equation finite difference operator based on a Gaussian ray beam. The illumination technology utilizes ray tracing at the earliest time to obtain the covering times of each underground reflection interface. Beykin (1985) proposes a method for evaluating imaging resolution by using the number of times of coverage based on generalized Laden transform and Born approximation; lecote and Gelius (1998) and Gelius and lecote (2000) propose methods for calculating the number of times of coverage of the ground scattered waves using a ray tracing method; the illumination analysis based on ray tracing provides directional illumination information, the calculation speed is high, but only the kinematic characteristics of seismic waves are reflected, and the calculation result is reliable when the medium is uniform; for complex structures (such as salt domes, paste bodies and the like), due to high-frequency ray approximation, the method generates shadow areas and defocusing areas in illumination areas, and calculation errors are generated. To avoid the disadvantages of this algorithm, the illumination analysis introduces wave equations with certain effects. The wave theory is applied to the study of the illumination condition under complex geology (Wu and Chen, 2001; Wu and Chen, 2002; Xie and Wu, 2002; Xie et al, 2005; Wu and Chen, 2006; Xie et al, 2006; Xie and Yang, 2008), Wu et al propose the concept of optimizing and designing the parameters of the acquisition system by means of illumination analysis; liwanwan (2008) performs illumination analysis by using a two-pass wave operator; DongLiang et al (2006) utilize seismic wave illumination analysis to obtain energy distributions in the CRP sense for seismic data acquisition design and optimization of observation systems. The illumination analysis method based on the ray has fast calculation speed, but has poor illumination result for a more complex geological model; the illumination analysis method based on the wave equation has accurate illumination results, but needs a large amount of calculation time. In recent years, with the performance of the machine being greatly improved, the deep learning method can well solve some nonlinear mapping problems, a large batch of data sets are used for training, a good network model can be obtained, and the calculation rate is greatly improved.

Disclosure of Invention

In order to solve the technical problems, the invention provides a local imaging algorithm based on deep learning illumination analysis, which utilizes illumination analysis to research the energy distribution condition of seismic waves generated by seismic source excitation and propagated in an underground medium, can optimize an observation system, enables the energy distribution in the medium to be optimal, and can perform energy compensation at the same time. The method based on deep learning can quickly obtain the energy distribution condition of the underground geologic body, then a shot set which contributes greatly to a target area is screened out through an energy angle, and partial shot sets are utilized to further image and compensate the area, so that the calculation time can be greatly shortened, and the working efficiency is improved.

The technical scheme of the invention is as follows: a local migration imaging method based on deep learning illumination analysis utilizes deep learning to realize the distribution condition of energy in a wave field, takes synthetic data as an example, and comprises the following steps:

step 1), constructing a two-dimensional geological speed model, and obtaining a single-shot illumination result corresponding to the two-dimensional geological speed model by using a traditional illumination analysis method;

step 2) constructing a training data set based on the geological speed model and the corresponding single-shot illumination result, and inputting the constructed Unet neural network for training;

step 3) predicting a single shot illumination result by using the trained neural network, finding out the specific distribution condition of illumination energy in the geological speed model from the single shot superposition illumination map, manually selecting a region with a pixel value smaller than a preset threshold value, and defining the region as a weak illumination region;

step 4) screening shot sets according to the contribution size and the illumination intensity of the single shot in the weak illumination area;

step 5) carrying out migration imaging on the screened shot sets and all shot sets according to a forward transmission wave field and a backward transmission wave field of the detector; and finally intercepting the imaging result of the weak illumination area.

Further, the two-dimensional geological speed model is constructed in the step (1), lighting analysis is to write matlab program to randomly generate some underground speed structure models aiming at underground complex geological structures, and inclined stratums, folds, faults and high-speed abnormal body structures are added into the models to ensure the richness of the models.

Further, in the step (1), a single shot illumination result corresponding to the model is obtained by using a traditional illumination analysis method, and the single shot illumination result is obtained according to a two-dimensional constant-density time domain acoustic wave equation:

where v (x, z) is the seismic velocity in space (x, z), u (x, z, t) is the wavefield in space (x, z) at time t, and s (t) is the seismic source at time t;

defining the single-shot source illumination intensity of one point in the space as follows:

whereinThe source illumination intensity of the ith shot;

the illumination intensity of the N-shot seismic sources is the sum of the illumination intensities of all the single-shot seismic sources:

further, the step (2) is to construct a training data set based on the geological speed model and the corresponding single shot illumination result, add the position information of the shot points into the geological speed model and store the position information into an RGB picture, so as to obtain an input data set of the neural network: adding points at the corresponding shot positionsA Gaussian function simulating the condition of shot points; for the location of the shot point (x)₀,y₀) Then, adding a two-dimensional gaussian function to the model is:

where (x, y) is the model midpoint coordinate, and c1 and c2 define the variance in the x and y directions;

further, the training data set is constructed based on the geological speed model and the corresponding single shot illumination result, the geological model single shot illumination result is converted into an RGB picture to be stored, and a label data set of the neural network is obtained; and based on the traditional Unet network framework, adjusting network parameters to obtain the network structure of the project, and loading the input training set and the label data set into the neural network for training.

Further, in the step (3), the single shot prediction results of the neural network are superposed to obtain a multi-shot energy illumination map, the energy intensity of each area is seen from the illumination map, the superposed illumination map is read by means of a matlab program, and then the size of the weak illumination area is manually selected.

Further, the shot set is screened out according to the contribution size and the illumination intensity of the single shot in the weak illumination area in the step (4); and averaging the weak illumination areas according to the superimposed illumination map to obtain average illumination, averaging the illumination results of each single shot in the weak illumination areas, and if the average value is greater than the average illumination, determining the shot as a screened shot set.

Further, the shot gathers and all shot gathers which are screened out are subjected to offset imaging according to forward wave fields and backward wave fields of the detector in the step (5), finally, the imaging result of the weak illumination area is intercepted, only the forward wave field of the upper boundary of the weak illumination area is recorded, and then the backward wave field of the detector is imaged at each point of the space by applying the cross-correlation imaging condition; wherein the cross-correlation imaging condition is formulated as,

wherein r (x, z, t) is a backward wave field of the detector at the time t at the space point (x, z), and s (x, z, t) is a forward wave field of the seismic source at the time t at the space point (x, z); image (x, z) is an imaging result at the space (x, z);

and (5) performing offset imaging on the screened shot gathers and all shot gathers according to a forward wave field and a backward wave field of a detector, and finally intercepting the imaging result of the weak illumination area.

Has the advantages that:

1) the method realizes the single shot illumination of the underground speed model by using the deep learning convolutional neural network, and can quickly obtain the energy distribution condition of the geological model according to the given shot point position;

2) the shot set which has larger contribution to the weak illumination area is screened out according to the illumination result of the single shot; offset imaging of the weak illumination area is realized through the shot gathers, and the calculation efficiency is greatly improved; in the actual acquisition, a reference function can be provided for the acquisition design of the weak illumination area.

Drawings

FIG. 1 is a flow chart of a method of local offset imaging based on deep learning illumination analysis according to the present invention;

FIG. 2 a randomly generated velocity model;

fig. 3 training data set: (a) a neural network input; (b) outputting by a neural network;

FIG. 4 SEG salt dome model;

fig. 5 SEG salt dome model test results: (a) a neural network input; (b) single shot lighting prediction results;

fig. 6 SEG salt dome model processing results: (a) superposing illumination results on all the guns; (b) manually selecting a weak illumination area position;

FIG. 7 shot gather screening scenario: all the shots are black on the lower side of the figure, the selected shot gathers are black on the upper side of the figure, and the shots are unselected in white;

FIG. 8 reverse time migration results: (a) partial shot set migration results; (b) all shot-set migration results.

Detailed Description

The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments, and all other embodiments obtained by a person skilled in the art based on the embodiments of the present invention belong to the protection scope of the present invention without creative efforts.

As shown in fig. 1, according to an embodiment of the present invention, a local imaging algorithm based on deep learning illumination analysis is provided, which uses deep learning to predict the distribution of energy in a wave field, taking synthetic data as an example, and the process includes the following steps:

step 1), constructing a two-dimensional geological speed model, and obtaining a single-shot illumination result corresponding to the two-dimensional geological speed model by using a traditional illumination analysis method;

step 2) constructing a training data set based on the geological speed model and the corresponding single-shot illumination result, and inputting the constructed Unet neural network for training;

step 3) predicting a single shot illumination result by using the trained neural network, finding out the specific distribution condition of illumination energy in the geological speed model from the single shot superposition illumination map, manually selecting a region with a pixel value smaller than a preset threshold value, and defining the region as a weak illumination region;

step 4) screening shot sets according to the contribution size and the illumination intensity of the single shot in the weak illumination area;

step 5) carrying out migration imaging on the screened shot sets and all shot sets according to a forward transmission wave field and a backward transmission wave field of the detector; and finally intercepting the imaging result of the weak illumination area.

Further, the two-dimensional geological speed model is constructed in the step (1), lighting analysis is to write matlab program to randomly generate some underground speed structure models aiming at underground complex geological structures, and inclined stratums, folds, faults and high-speed abnormal body structures are added into the models to ensure the richness of the models.

Further, in the step (1), a single shot illumination result corresponding to the model is obtained by using a traditional illumination analysis method, and the single shot illumination result is obtained according to a two-dimensional constant-density time domain acoustic wave equation:

where v (x, z) is the seismic velocity in space (x, z), u (x, z, t) is the wavefield in space (x, z) at time t, and s (t) is the seismic source at time t;

defining the single-shot source illumination intensity of one point in the space as follows:

whereinThe source illumination intensity of the ith shot;

the illumination intensity of the N-shot seismic sources is the sum of the illumination intensities of all the single-shot seismic sources:

further, the step (2) is to construct a training data set based on the geological speed model and the corresponding single shot illumination result, add the position information of the shot points into the geological speed model and store the position information into an RGB picture, so as to obtain an input data set of the neural network: adding a point Gaussian function at the position corresponding to the shot point to simulate the situation of the shot point; for the location of the shot point (x)₀,y₀) Then, adding a two-dimensional gaussian function to the model is:

where (x, y) is the model midpoint coordinate, and c1 and c2 define the variance in the x and y directions;

further, the training data set is constructed based on the geological speed model and the corresponding single shot illumination result, the geological model single shot illumination result is converted into an RGB picture to be stored, and a label data set of the neural network is obtained; and based on the traditional Unet network framework, adjusting network parameters to obtain the network structure of the project, and loading the input training set and the label data set into the neural network for training.

Further, in the step (3), the single shot prediction results of the neural network are superposed to obtain a multi-shot energy illumination map, the energy intensity of each area is seen from the illumination map, the superposed illumination map is read by means of a matlab program, and then the size of the weak illumination area is manually selected.

Further, the shot set is screened out according to the contribution size and the illumination intensity of the single shot in the weak illumination area in the step (4); and averaging the weak illumination areas according to the superimposed illumination map to obtain average illumination, averaging the illumination results of each single shot in the weak illumination areas, and if the average value is greater than the average illumination, determining the shot as a screened shot set.

Further, the shot gathers and all shot gathers which are screened out are subjected to offset imaging according to forward wave fields and backward wave fields of the detector in the step (5), finally, the imaging result of the weak illumination area is intercepted, only the forward wave field of the upper boundary of the weak illumination area is recorded, and then the backward wave field of the detector is imaged at each point of the space by applying the cross-correlation imaging condition; wherein the cross-correlation imaging condition is formulated as,

wherein r (x, z, t) is a backward wave field of the detector at the time t at the space point (x, z), and s (x, z, t) is a forward wave field of the seismic source at the time t at the space point (x, z); image (x, z) is an imaging result at the space (x, z);

and (5) performing offset imaging on the screened shot gathers and all shot gathers according to a forward wave field and a backward wave field of a detector, and finally intercepting the imaging result of the weak illumination area.

According to one embodiment of the invention, the following is: and (4) testing by using the SEG salt dome model, and performing rarefaction treatment on the original model because the original model is huge. The real model is shown in fig. 2, the speed ranges from 1.5km/s to 4.5km/s, the model size is 230 × 676, and the grid distance dx ═ dz ═ 10 m. The data acquisition system consisted of 66 cannons and 676 geophones, which were evenly distributed across the entire earth's surface.

1) The generation of a speed model is realized under matlab, the illumination of a single shot and the offset imaging are realized under C language, and the training of a neural network is realized under python;

2) the program is written to generate a series of random geological speed models (such as figure 2) containing inclined layers, folds, faults and high-speed abnormal bodies;

3) obtaining a corresponding single shot illumination result of the generated geological speed model by using a two-way acoustic equation finite difference forward modeling method;

4) adding the position information of the cannon into the geological speed model, and transferring the position information into an RGB (red, green and blue) picture, and performing the same treatment on the illumination result (as shown in figure 3): FIG. 3(a) is the input to the training set, and FIG. 3(b) is the output of the training set;

5) modifying parameters such as the layer number, an optimization operator and the like based on the Unet network model to generate a neural network of the project, and inputting the picture training set into the network for training;

6) taking the SEG salt dome model as a test model (as shown in figure 4), adding shot point information, and inputting the shot point information into a neural network to predict single shot illumination: FIG. 5(a) is an input to a neural network, velocity model plus shot information; FIG. 5(b) is the output of the neural network, single shot illumination results;

7) superposing all the predicted single-shot illumination results to obtain the superposed illumination condition of the whole shot (as shown in figure 6 (a)); screening out obvious weak illumination areas according to the superposition illumination result (as shown in fig. 6 (b));

8) the illumination energy of each cannon contributes to a weak illumination area, the average illumination degree is obtained by averaging the weak illumination area according to a superposition illumination image, the illumination result of each single cannon is averaged in the weak illumination area, if the average value is larger than the average illumination degree, the cannon is a screened cannon set (shown in figure 7), compared with all the cannon sets which are black at the lower part of the cannon, the black at the upper part of the cannon is a selected cannon set, and the white is not selected;

9) recording a forward wave field in the forward modeling of the boundary on the SEG model weak illumination area;

10) and (3) performing migration imaging on the screened shot gathers and all shot gathers according to the recorded forward wave field and backward wave field of the detector by applying a cross-correlation imaging condition, and finally intercepting the imaging result of the weak illumination area (see figure 8): FIG. 8(a) shows the migration results of some of the shot sets selected, and FIG. 8(b) shows the migration results of all of the shot sets. The run-time comparison for reverse time migration is shown in table 1.

TABLE 1 reverse time migration time run time comparison results

	Full shot offset imaging	Partial shot offset imaging
			HP Z8G 4 work 10 nucleus	26 hours	9 hours

Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, but various changes may be apparent to those skilled in the art, and it is intended that all inventive concepts utilizing the inventive concepts set forth herein be protected without departing from the spirit and scope of the present invention as defined and limited by the appended claims.