CN111983681A - Seismic wave impedance inversion method based on countermeasure learning - Google Patents
Seismic wave impedance inversion method based on countermeasure learning Download PDFInfo
- Publication number
- CN111983681A CN111983681A CN202010895121.2A CN202010895121A CN111983681A CN 111983681 A CN111983681 A CN 111983681A CN 202010895121 A CN202010895121 A CN 202010895121A CN 111983681 A CN111983681 A CN 111983681A
- Authority
- CN
- China
- Prior art keywords
- seismic
- data
- network
- training
- impedance
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 76
- 238000012549 training Methods 0.000 claims description 54
- 230000006870 function Effects 0.000 claims description 24
- 238000010606 normalization Methods 0.000 claims description 4
- 230000003042 antagnostic effect Effects 0.000 claims 4
- 238000012545 processing Methods 0.000 abstract description 4
- 230000008569 process Effects 0.000 description 14
- 238000013135 deep learning Methods 0.000 description 7
- 238000004088 simulation Methods 0.000 description 7
- 238000007796 conventional method Methods 0.000 description 6
- 239000011159 matrix material Substances 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000002194 synthesizing effect Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 230000004304 visual acuity Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000007620 mathematical function Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000002948 stochastic simulation Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
- G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
- G01V1/30—Analysis
- G01V1/306—Analysis for determining physical properties of the subsurface, e.g. impedance, porosity or attenuation profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V2210/00—Details of seismic processing or analysis
- G01V2210/60—Analysis
- G01V2210/62—Physical property of subsurface
- G01V2210/622—Velocity, density or impedance
- G01V2210/6226—Impedance
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Remote Sensing (AREA)
- Acoustics & Sound (AREA)
- Environmental & Geological Engineering (AREA)
- Geology (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Geophysics (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention discloses a seismic wave impedance inversion method based on countermeasure learning, which is applied to the field of seismic data processing and aims at the problem of lack of logging data in the traditional high-resolution seismic wave impedance inversion method.
Description
Technical Field
The invention belongs to the field of seismic data processing, and particularly relates to a seismic impedance inversion technology.
Background
The seismic wave impedance inversion obtains the underground wave impedance information by artificially exciting the reflected seismic waves, and is an important technical means for predicting oil and gas reservoirs. The method is limited by the frequency bandwidth of seismic waves, and high-resolution seismic wave impedance information is difficult to obtain by a conventional seismic inversion method, so that the conventional high-resolution seismic inversion method is still a difficult problem of seismic wave impedance inversion at present. At present, the most common high-resolution seismic wave impedance inversion technical means is a geostatistical inversion method. The technology utilizes the longitudinal high resolution of the logging data and the transverse distribution characteristics of the seismic data to realize the high resolution well interpolation under the control of the seismic data. The specific implementation process comprises the steps of constructing a variation function and statistical distribution through known logging data, taking the statistical distribution as prior information, taking seismic data as a likelihood function, and obtaining posterior information under a Bayesian framework to realize the impedance curve generation of a non-well region. From the implementation process, it can be found that the prior information of the conventional geostatistical inversion method based on the variation function depends on well data, and if the well logging data samples are few or the heterogeneity of the underground reservoir is severe, the obtained prior information is insufficient and incomplete, which inevitably causes the inversion result to have deviation. Therefore, to obtain inversion results with high accuracy and precision, a large amount of well data is required, and in practice, the well logging data is always very limited.
Geostatistical inversion is the main method for obtaining high-resolution wave impedance inversion results at present, but the existing geostatistical inversion method still has some problems. First, a conventional two-point geostatistical inversion method relies on known well logging data to construct a variation function, that is, prior information obtained by geostatistical is closely related to the well logging data. This presents two serious problems: firstly, when the logging data is less, the obtained prior information is incomplete; secondly, when the geological structure of the underground becomes very complex, the geological information description based on the variation function is not sufficient. The later proposed multi-point geostatistical inversion method overcomes the defect that two-point geostatistical cannot well describe the complex underground geological structure to a certain extent, but introduces a new problem. First, the stochastic simulation process of multi-point geostatistics is very time consuming [9 ]. Secondly, the requirement of multi-point geostatistics on training pictures is high; it requires that the geologic features reflected by the training images must be consistent with the reservoir features to be simulated. In practice, it is often difficult to obtain such a high quality training picture.
Disclosure of Invention
In order to solve the problem of lack of logging data in the traditional high-resolution seismic wave impedance inversion method, the invention realizes the high-resolution seismic wave impedance inversion method based on a generated countermeasure Network (GAN).
The technical scheme adopted by the invention is as follows: a seismic wave impedance inversion method based on countermeasure learning, the method being based on a countermeasure network comprising: a generator network and a discriminator network;
in the stage of training the countermeasure network, the input of the generator network is the noise data with the same dimension as the training data, and the output of the generator network is the impedance data; the input of the discriminator network comprises real seismic data and synthetic seismic data, and the synthetic seismic data is synthesized with the seismic wavelets according to the impedance data output by the generator network; alternately training the generator network and the discriminator network according to the output of the discriminator network;
and performing wave impedance inversion by adopting the trained confrontation network.
In the stage of training against the network, the input training data of the generator network is the acquired seismic data, which is taken as condition data by splicing behind the noise data in each trace along the length of the single trace data.
The noisy data is sampled from a standard normal distribution and has the same dimensions as the training data.
In the stage of the training of the confrontation network, the synthetic seismic data input by the discriminator network specifically comprises the following steps: and after carrying out inverse normalization on the impedance data output by the generator network, carrying out convolution on the impedance data and the seismic wavelets to obtain synthetic seismic data.
When the trained confrontation network is used for wave impedance inversion, the input of the generator network is noise data sampled from standard normal distribution and normalized seismic data with the same dimensionality as the data.
The loss function of the generator network when generating the pre-training of the countermeasure network is:
wherein, SeismicGRepresenting a Seismic recording synthesized using the generated impedance, SeismicRRepresenting a real seismic recording.
The loss function of the generator network in the training of the countermeasure network is generated as follows:
LossG=cross_entropy(SeismicG,1)
+norm(SeismicG-SeismicR)
+norm(Gip-initip)
wherein, SeismicGRepresenting a Seismic recording synthesized using the generated impedance, SeismicRRepresenting real seismic records, GipRepresenting the impedance, init, generated by the generatoripRepresenting the initial impedance model.
The invention has the beneficial effects that: according to the method, the generation of the countermeasure network and the seismic wave impedance inversion are combined, well data distribution information is obtained through pre-training of the neural network, the training generation of the countermeasure network and the wave impedance inversion process are combined, a high-resolution wave impedance inversion result can be obtained without logging data participation in the inversion process, and the serious dependence of the traditional high-resolution seismic wave impedance inversion on logging data is solved; the method of the invention has the following advantages:
(1) introducing a condition generation countermeasure network, which can not only describe more complex underground geological information, but also introduce seismic data as a condition-guided inversion result;
(2) seismic data are introduced as training data, and the serious dependence of the traditional inversion method on logging data is broken through.
Drawings
FIG. 1 is a flow chart of a wave impedance inversion technique;
FIG. 2 is a two-point geostatistical technique flow diagram;
FIG. 3 is a flow diagram of a multi-point geostatistical technique;
FIG. 4 is a schematic flow chart of generating a countermeasure network;
FIG. 5 is a seismic wave impedance inversion method based on a generative countermeasure network;
FIG. 6 is a high resolution seismic wave impedance inversion network structure based on a generative countermeasure network;
FIG. 7 is a flow chart of a conditional generation countermeasure network;
FIG. 8 is the result of a single pass inversion after physical steering is added;
FIG. 9 is the result of a single pass inversion after the initial model constraints are added;
FIG. 10 is a model wave impedance and seismic section;
wherein, fig. 10(a) is a real impedance model, and fig. 10(b) is a synthetic record;
FIG. 11 is a graph comparing inversion results;
wherein, fig. 11(a) is the inversion result of the conventional method, and fig. 11(b) is the inversion result of the present invention;
FIG. 12 is a single pass comparison of the results of model inversion;
FIG. 13 is a single-pass comparison of the results of an actual data inversion;
wherein, fig. 13(a) the inversion result of the actual data using the method of the present invention, and fig. 13(b) the inversion result of the actual data using the conventional method;
FIG. 14 is a cross-sectional comparison of the results of an actual data inversion;
wherein, FIG. 14(a) is a section of the impedance inversion result of the method of the invention of the well A, and FIG. 14(b) is a section of the conventional inversion result of the well A;
FIG. 15 is a slice comparison of the results of an actual data inversion;
fig. 15(a) is a seismic data slice, fig. 15(b) is an inversion result slice of the method of the present invention, and fig. 15(c) is a conventional inversion result slice.
Detailed Description
To facilitate understanding of the technical contents of the present invention by those skilled in the art, the description will first be made of the prior art related to the technical contents of the present invention:
the related prior art is as follows:
1. seismic wave impedance inversion method
The seismic wave impedance inversion is a seismic special processing interpretation technology for inverting the stratum wave impedance by using seismic data, and is a seismic inversion technology which effectively combines the post-stack seismic data, logging data and geological interpretation. The method not only enables geological researchers to better know geological characteristics, but also has important instructive significance for oil and gas exploration. The method has the core idea that a wave impedance profile with high readability is finally obtained by iteratively updating a given initial wave impedance model by utilizing the characteristics of high longitudinal resolution of logging data and high transverse resolution of seismic data and combining geological interpretation data. The seismic wave impedance inversion method starts in the last 70 th century, and the post-stack one-dimensional wave impedance inversion based on the convolution model is mainly researched at the time. The generalized linear inversion method of seismic data appeared in 1983 enables the wave impedance inversion technology to enter a new development stage. In the 90 s, the Li soldier also provided a method combining recursive inversion and bandwidth constraint inversion.
The principle of seismic wave impedance inversion is as follows, and according to the Robinson convolution model, the post-stack seismic traces can be represented by the convolution of wavelets and reflection coefficients:
s(t)=w(t)*r(t)+n(t) (1)
wherein s (t) represents seismic trace data, w (t) represents seismic wavelets, r (t) represents formation reflection coefficients, and n (t) represents random noise. The model can also be expressed as:
wherein, Wi-j+1Represents a wavelet, rjDenotes the zero-bias reflection coefficient, niRepresenting random noise and N representing the number of points of the sequence of reflection coefficients.
The above formula is expressed in matrix form:
s=Wr (3)
w represents a seismic wavelet matrix, and r represents a reflection coefficient sequence;
according to Russell's approximation formula, when the difference between the formation wave impedance coefficients is much smaller than the absolute value of the wave impedance, there is the following relationship:
wherein Z isiIs the wave impedance coefficient of the ith layer. The impedance can be converted into the reflection coefficient by the formula, and conversely, the impedance can be converted into the reflection coefficient. The above formula is expressed in matrix form:
r=Dl (5)
the above-mentioned binding formula (3) gives:
s=WDl (7)
then, let G be WD, the forward model of the post-stack wave impedance can be expressed as:
s=Gl (8)
wherein, the vector s represents the post-stack seismic record, D represents a difference matrix, G represents a forward matrix, and l represents a stratum wave impedance sequence to be inverted. The flow of the wave impedance inversion is shown in fig. 1.
2. Geostatistical inversion method
Since this century, nonlinear inversion methods have been used in wave impedance inversion, making this technology a tremendous advance. Among them, the rise of geostatistics has led geostatistical inversion methods to be widely paid attention by overseas and overseas scholars, and has achieved very important development results.
Geostatistical inversion describes a prior density function of a solution space by using geostatistical information, and mainly comprises two parts of random modeling and simulation result optimization. The method has the characteristics that the advantages of stochastic modeling and seismic inversion are combined, and the inversion result breaks through the limitation of seismic frequency bandwidth, so that a high-resolution stratum wave impedance inversion result can be obtained. Moreover, the wave impedance result obtained by the geostatistical inversion is well matched with the logging data. Traditional geostatistical attempts to describe geological information using a spatial relationship between two points. Firstly, calculating a variation value between every two known well points in an area to be inverted, and constructing a variation function according to the variation value, wherein the expression is as follows:
r(x,d)=0.5*E[Z(x)-Z(x+d)]2=r(d) (10)
that is, at a certain position x, the difference between the parameters of distance d is obtained, and then the half-variance is taken, and the variation value is denoted as r (x, d). Secondly, the mean and variance of the points to be simulated are estimated by using a kriging method. By Z*(x0) Is noted as a certain position x0Analog value of (A;)iRepresenting the weight of the kriging interpolation, the calculation method of the estimated value of the point to be simulated is as follows:
and finally, constructing Gaussian distribution by using the mean value and the variance, and sampling to obtain an analog value of the point. Finally, the simulation result is convolved with the wavelet, and the simulation result is optimized by synthesizing errors between the record and the actual seismic data. The complete process is shown in figure 2:
with the progress of research, people gradually find that geological information of underground complex structures cannot be described only by the spatial relationship of two points. In order to overcome the defects, people also provide a multipoint geostatistical inversion method. Multi-point geostatistics describe subsurface geologic structure information using a combination pattern of multiple points in space, and are therefore more suitable for geologic information simulation with complex structures. First, a training image containing prior geological knowledge of a work area to be studied is obtained. Second, the training image is scanned with the template to generate a number of data events. Then, searching a prototype closest to the point to be simulated from the data event library, and constructing a conditional probability density function according to the prototype for random simulation. Finally, the simulation result is convolved with the wavelet, and the simulation result is optimized by synthesizing errors between the record and the actual seismic data. The complete flow of the geostatistical inversion is shown in figure 3.
3. Generating a countermeasure network
Deep learning is a branch of machine learning, and is an algorithm for performing characterization learning on large-scale data by taking an artificial neural network as a framework. Beginning in the 60's of the 20 th century, Ivakhnenko et al proposed a multi-layered depth sensor for feed forward, an original model of convolutional neural networks was proposed in 1984, a deep learning concept was proposed in 2006 G.Hinton, and a deep confidence network and a layer-by-layer greedy algorithm proposed by its team opened the door to deep learning. But the first application to deep learning was in 1989, y. In 1998, y.lecun proposed a Convolutional Neural Network (CNN), a common model for deep learning, which pushed deep learning to a new height. In 2014, Goodfellow et al proposed generation of confrontation Network (GAN), and have received extensive attention from researchers and industries at home and abroad, and have become a research hotspot in the field of artificial intelligence.
Generating a countermeasure Network (GAN) based on the idea of zero sum game can obtain the distribution of data through unsupervised learning and Generate more realistic data. It is mainly composed of two parts: generator networks (Generator, G) and Discriminator networks (Discriminator, D). During the training process, the generator network continuously learns the data distribution of the real samples and continuously generates "false" samples that are close to the real data distribution. The discriminator network discriminates the input real sample and the false sample generated by the generator respectively; if the input is a real sample, the judgment result is 1; if the input is a "false" sample generated by the generator, the decision is 0. With the continuous training, the two will reach a balance finally, namely the generator will generate the sample data which is false or spurious, and the judger will always output the judgment result to the sample data generated by the generator to be about 0.5. As shown in fig. 4, a schematic flow chart of generating a countermeasure Network (GAN) is shown.
The invention is further explained below with reference to the drawings:
aiming at the defects of geostatistical inversion in the prior art, the invention introduces a generation assisted confrontation Network (GAN) in the field of deep learning. The method aims to obtain a high-resolution wave impedance inversion result by using a generation countermeasure network technology in seismic wave impedance inversion, improve the prediction precision of an oil and gas reservoir and reduce the dependence of the high-resolution inversion on logging data. The main contents of the technical scheme of the invention are as follows:
(1) a more complete picture of the complex underground geological structure is obtained by generating the learning and representation capability of the confrontation network.
(2) Seismic data is introduced into the training data to reduce the heavy dependence on well log data.
As shown in fig. 5, the implementation process of the present invention includes the following steps:
and S1, acquiring training data. It is known that training a neural network with excellent performance requires a large amount of training data, and the logging data in actual work is often very limited, which makes the idea of network training using logging data difficult to realize. In order to solve the problem, seismic data which are easy to acquire in actual work are used as training data, the problem of insufficient training data is solved, and meanwhile the dependence degree of inversion results on logging data is greatly reduced. Meanwhile, in order to make the network easier to train, we normalize the seismic data, and the expression is as follows:
wherein d isstdRepresenting normalized seismic data, d representing raw seismic data, μ representing the mean of the seismic data, and σ representing the standard deviation of the seismic data.
And S2, acquiring input data during generator network training. In the original generation countermeasure network, the input data of the generator is a noise vector sampled from the gaussian distribution, which results in poor network interactivity. To address this issue, the present invention decides to employ a Conditional Generation Adaptive Network (CGAN) framework. The input data to the network also incorporates condition data with additional information compared to the original generated countermeasure network. To obtain qualified input data, noise data sampled from a standard normal distribution of the same dimensions as the training data in step S1 is first generated. Then, the seismic data in step S1 is used as conditional data, and each trace is spliced behind the corresponding noise data in the direction of the single trace data length.
And S3, acquiring input data of the discriminator network. As can be seen from the principle of generating a countermeasure network, the input data of the discriminator network are: real seismic data and seismic data synthesized from the impedance data generated by the generator and the seismic wavelets. At this time, since the generation result of the generator is a normalized impedance result between 0 and 1, it is necessary to denormalize the normalized impedance result before synthesizing with the wavelet. At this time, the synthetic seismic data as "false" samples is consistent with the value range of the original seismic data, and therefore the real sample data input to the discriminator should be the original seismic data without being subjected to the normalization processing.
And S4, training a network. Firstly, the training data obtained in the step S2 is input into the generator network by taking patch every time to obtain the normalized impedance generated by the generator, the impedance result is convoluted with the seismic wavelet after being denormalized, and the convolution result is false sample data. And respectively inputting the true sample data and the false sample data into a discriminator network to obtain a discrimination result so as to carry out alternate training on the generator and the discriminator.
And S5, inversion. When the network is trained, the wave impedance inversion can be performed. The input data at this time is noise data sampled from a standard normal distribution and normalized seismic data of the same dimension as the data, and it should be noted that the two sets of data should be spliced in the manner of step S2 at this time. The result obtained by the input data is the normalized wave impedance inversion result, and the inversion result is subjected to inverse normalization to obtain the final wave impedance inversion result.
Fig. 6 is a diagram showing a high-resolution seismic wave impedance inversion network structure based on a generation countermeasure network, wherein D in fig. 6 represents a discriminator network in the generation countermeasure network, and G represents a generator network therein. The sesimic represents a real Seismic record, and the Input represents the Input data obtained in step S2. Output in fig. 6 is normalized impedance data generated by the generator network, and a reflection coefficient calculated after the impedance is inversely normalized is below output. The left side of the reflection coefficient in turn represents the seismic wavelet and the synthetic seismic data synthetic seismic.
During the network training process, the network input of the generator is a four-dimensional tensor, and the dimension of the tensor is equal to [ Batch × NP × NT × channel ], wherein the Batch represents the number of samples of the input; NP indicates the height of the input sample, which is equal to the number of data points per pass; NT represents the width of the input sample, i.e., refers to the number of seismic traces per input; the channel indicates the number of channels of the input sample, here equal to 1. Each sample is a patch, and is randomly selected from the training data of step S3. And, the dimension of the output and the meaning represented by the dimension are the same as the input, three full connection layers are arranged between the input and the output, and the activation functions of the three full connection layers are Sigmoid functions, and the expression is as follows:
sigmoid is a mathematical function and x is the variable of this function. This variable corresponds to the invention's output value of the last fully connected layer of the network that has not been subjected to the activation function.
We solve the problem of training sample deficiency by using seismic data as a training dataset. In order to accelerate the training of the network and make full use of the existing logging data, the impedance data obtained from the logging data and the corresponding seismic data are first combined into "tag-training sample" data pairs, and these data pairs are used to pre-train the generator network in a supervised manner. The loss function during the pre-training process is calculated as follows:
wherein the first term [ Giplog(Realip)+(1-Gip)log(1-Realip)]Representing the cross entropy of the impedance generated by the generator with the true impedance, the second termShown is the L2 loss function of the true impedance versus the generator generated impedance. GipRepresenting the impedance generated by the generator, RealipRepresenting the true impedance value.
Then, after the generator network is pre-trained, the entire generated countermeasure network is trained by alternating training generator networks and discriminator networks. First, we first introduce the loss function originally generated against the network:
wherein G, D denotes the generator network and the discriminator network, x, P, respectivelydata(x) Respectively representing the true sample data and the true distribution to which it is subjected, z, Pz(z) represents the input noise data and the distribution learned by the generator, respectively.
The core idea is as follows: first, the discriminator network D is made to learn to discriminate between false samples and true samples. After D has a certain resolving power, the generator network G tries to fool D with generated false samples and let D believe that these false samples are authentic. With the enhancement of the discrimination capability of the D on the true samples and the false samples (namely, more and more learned sample data), the G is more and more difficult to cheat the D, so that the G also continuously improves the capability of generating the false samples. The discrimination capability of D and the capability of G counterfeiting are greatly improved in the game by circulating for many times. In the training process of generating the countermeasure network, a discriminator network D is trained by maximizing the probability of correctly distributing labels to training samples and G generation samples respectively; the generator network G is also trained by constantly minimizing log (1-D (G (z))).
D has a certain resolving power and is understood to be: by feeding the true and false samples to the discriminators, respectively, the network of discriminators can correctly distinguish which ones of these samples (where a portion takes on a value greater than or equal to 1 but less than the total number of all samples) are true and which ones are false. However, the resolution is not strong enough, so that it is not possible to correctly distinguish which samples are true and which are false.
Here, in order to avoid the problem of poor Network interactivity caused by only random noise input by the originally generated countermeasure Network, the present invention determines to enhance the Network interactivity by using a Conditional generation countermeasure Network (CGAN) framework. The conditional generation countermeasure network allows us to add additional information, which may be category information or any other information, to target the results generated in the originally generated countermeasure network. The loss function against the network is generated conditional on:
wherein G, D distinguishes between the representation generator network and the discriminator network, x, Pdata(x) Respectively representing the true sample data and the true distribution to which it is subjected, z, Pz(z) represents the input noise data and the distribution learned by the generator, respectively, and y is the condition data. As can be seen from the formula, the generator produces results with or constrained by y information. Therefore, in order to keep the inversion result obtained by the method consistent with the seismic data, the seismic data serving as conditional data are spliced behind the noise vector, so that the inversion result is restrained. As shown in fig. 7, a flow framework for generating a countermeasure network for a condition.
Meanwhile, as can be seen from the foregoing, in order to solve the problems of insufficient training data and dependence of inversion results on logging data, we do not make the discriminator network directly learn to discriminate true and false impedance data, but indirectly discriminate impedance by discriminating between true and false seismic data, so that abundant seismic data are ingeniously used to replace limited impedance data as training data, and the above problems are effectively solved. In addition, in order to enable the network to be optimized more correctly in the training process, a physical guiding mechanism is introduced, namely an L2 norm of the residual error of the synthetic seismic record and the real seismic record is added into the loss function. Thus, we modify the loss function of the generator network as follows:
wherein, SeismicGRepresenting a Seismic recording synthesized using the generated impedance, SeismicRRepresenting a real seismic recording. However, the test results show that although we can invert the impedance variation, the correct range of the value domain cannot be obtained, as shown in fig. 8:
in order to solve the problem, the invention adds the L2 norm of the initial model and the generated impedance residual as a regular term to the final loss function to constrain the value range of the inversion result, and the final loss function is modified as follows:
wherein, SeismicGRepresenting a Seismic recording synthesized using the generated impedance, SeismicRRepresenting real seismic records, GipRepresenting the impedance, init, generated by the generatoripRepresenting the initial impedance model. As shown in fig. 9, the result of the single pass inversion after the initial model constraint is added. From the data, the inversion result has good goodness of fit with actual data.
In order to verify the effect of the invention, a Marmousi model is used as model data to generate post-stack seismic records to verify the effectiveness of the invention. The known Marmousi impedance model is shown in fig. 10(a), and the resultant data obtained by convolving the Marmousi model with a 30Hz Ricker wavelet is shown in fig. 10 (b). It has 377 channels, each channel having 350 samples, with a sampling interval of 4 ms.
Fig. 11 is an obtained inversion result of a conventional inversion and a high-resolution wave impedance inversion method based on a generation countermeasure network. Comparing the results, it can be seen that the resolution of the conventional inversion is far less high than that of the high resolution inversion method based on the generation of the countermeasure network.
The high-resolution wave impedance inversion method based on the generated countermeasure network does not use logging data required by the conventional method to participate in the inversion process, and the accuracy and the resolution of the inversion result are obviously higher than those of the conventional inversion method. And as shown in fig. 12, the inversion result obtained by the method provided by the invention has high coincidence degree with the well on a single channel.
To further illustrate the practical application capability of the method of the present invention, we apply the method of the present invention to a work area in northeast China. Firstly, interpolation is carried out on logging data by a geostatistical method, then transverse and longitudinal Butterworth filtering is carried out on the interpolation result to obtain an initial model, and the filtering coefficient is 0.8. Then, training data is obtained and the network is trained according to the process of the invention.
To illustrate the problem, we compared the inversion results of the conventional method and the present invention. Fig. 13 is a single-pass comparison of the wave impedance inversion results, and it can be found that the inversion result (dotted line) of the method of the present invention is closer to the real result (solid line), and the result accuracy is better than that of the conventional method.
As shown in fig. 14 and 15, from the comparison between the inversion result of the well-crossing profile and the slice inversion result, it can be found that the inversion result obtained by the method of the present invention has high goodness of fit with the logging data; and the data can be found to be more in line with the actual geological rule by comparing with the slice of the seismic data.
It can be seen from the comparison between the inversion result obtained by the conventional inversion method of the actual work area and the inversion result obtained by the high-resolution wave impedance inversion method based on the generated countermeasure network, and the resolution of the inversion result of the method provided by the invention is obviously higher than that of the inversion result obtained by the conventional method. And the whole has better transverse continuity. Compared with the seismic slice, the inversion result has higher goodness of fit with the seismic data, better accords with the actual geological rule, and can provide better guidance for further oil and gas exploration.
It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Various modifications and alterations to this invention will become apparent to those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.
Claims (7)
1. A seismic wave impedance inversion method based on countermeasure learning is characterized in that a countermeasure network based on the method comprises the following steps: a generator network and a discriminator network;
in the stage of training the countermeasure network, the input of the generator network is the noise data with the same dimension as the training data, and the output of the generator network is the impedance data; the input of the discriminator network comprises real seismic data and synthetic seismic data, and the synthetic seismic data is synthesized with the seismic wavelets according to the impedance data output by the generator network; alternately training the generator network and the discriminator network according to the output of the discriminator network;
and performing wave impedance inversion by adopting the trained confrontation network.
2. The seismic wave impedance inversion method based on the countermeasure learning of claim 1, wherein in the stage of the countermeasure network training, the training data inputted from the generator network is the acquired seismic data, and the seismic data is taken as condition data by splicing behind the noise data in each trace along the length of the single trace data.
3. The seismic wave impedance inversion method based on the countermeasure learning of claim 1, wherein the noise data are sampled from a standard normal distribution and have the same dimension as training data.
4. The seismic wave impedance inversion method based on countermeasure learning as claimed in claim 1, wherein at the stage of the countermeasure network training, the synthetic seismic data input by the discriminator network specifically is: and after carrying out inverse normalization on the impedance data output by the generator network, carrying out convolution on the impedance data and the seismic wavelets to obtain synthetic seismic data.
5. The seismic wave impedance inversion method based on the countermeasure learning of claim 1, wherein when the trained countermeasure network is used for wave impedance inversion, the input of the generator network is noise data sampled from a standard normal distribution and normalized seismic data with the same dimensionality as the data.
6. The seismic wave impedance inversion method based on antagonistic learning as claimed in claim 1, wherein the loss function of the generator network when generating the antagonistic network pre-training is:
wherein, SeismicGRepresenting a Seismic recording synthesized using the generated impedance, SeismicRRepresenting a real seismic recording.
7. The seismic wave impedance inversion method based on the antagonistic learning of claim 1, wherein the loss function of the generator network in the generation of the antagonistic network training is as follows:
LossG=cross_entropy(SeismicG,1)+norm(SeismicG-SeismicR)+norm(Gip-initip)
wherein, SeismicGRepresenting a Seismic recording synthesized using the generated impedance, SeismicRRepresenting real seismic records, GipRepresenting the impedance, init, generated by the generatoripRepresenting the initial impedance model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010895121.2A CN111983681B (en) | 2020-08-31 | 2020-08-31 | Seismic wave impedance inversion method based on countermeasure learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010895121.2A CN111983681B (en) | 2020-08-31 | 2020-08-31 | Seismic wave impedance inversion method based on countermeasure learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111983681A true CN111983681A (en) | 2020-11-24 |
CN111983681B CN111983681B (en) | 2021-09-14 |
Family
ID=73440849
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010895121.2A Active CN111983681B (en) | 2020-08-31 | 2020-08-31 | Seismic wave impedance inversion method based on countermeasure learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111983681B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112949089A (en) * | 2021-04-01 | 2021-06-11 | 吉林大学 | Aquifer structure inversion identification method based on discrete convolution residual error network |
CN113156513A (en) * | 2021-04-14 | 2021-07-23 | 吉林大学 | Convolutional neural network seismic signal denoising method based on attention guidance |
CN113191321A (en) * | 2021-05-21 | 2021-07-30 | 电子科技大学 | Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network |
CN113296150A (en) * | 2021-07-02 | 2021-08-24 | 清华大学 | High-dimensional closed-loop network seismic inversion method under logging constraint |
CN113608258A (en) * | 2021-07-16 | 2021-11-05 | 长江大学 | Self-consistent deep learning method for constructing high-resolution wave impedance inversion label |
CN113722893A (en) * | 2021-08-12 | 2021-11-30 | 中国石油大学(北京) | Seismic record inversion method, device, equipment and storage medium |
CN114563821A (en) * | 2020-11-27 | 2022-05-31 | 中国石油天然气股份有限公司 | Intelligent seismic wave impedance inversion label data manufacturing method and device |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102393532A (en) * | 2011-09-06 | 2012-03-28 | 电子科技大学 | Seismic signal inversion method |
CN103513278A (en) * | 2012-06-19 | 2014-01-15 | 中国石油化工股份有限公司 | Method for reservoir prediction by utilizing thickness of seismic wave group |
US20160209532A1 (en) * | 2014-11-17 | 2016-07-21 | Board Of Regents, The University Of Texas System | Applied interpolation techniques |
CN106997060A (en) * | 2017-06-14 | 2017-08-01 | 中国石油大学(华东) | A kind of seismic multi-attribute fusion method based on Shearlet fastICA |
CN109143353A (en) * | 2018-08-20 | 2019-01-04 | 电子科技大学 | A kind of pre-stack seismic waveform classification generating confrontation network based on depth convolution |
-
2020
- 2020-08-31 CN CN202010895121.2A patent/CN111983681B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102393532A (en) * | 2011-09-06 | 2012-03-28 | 电子科技大学 | Seismic signal inversion method |
CN103513278A (en) * | 2012-06-19 | 2014-01-15 | 中国石油化工股份有限公司 | Method for reservoir prediction by utilizing thickness of seismic wave group |
US20160209532A1 (en) * | 2014-11-17 | 2016-07-21 | Board Of Regents, The University Of Texas System | Applied interpolation techniques |
CN106997060A (en) * | 2017-06-14 | 2017-08-01 | 中国石油大学(华东) | A kind of seismic multi-attribute fusion method based on Shearlet fastICA |
CN109143353A (en) * | 2018-08-20 | 2019-01-04 | 电子科技大学 | A kind of pre-stack seismic waveform classification generating confrontation network based on depth convolution |
Non-Patent Citations (1)
Title |
---|
代丽艳,等: "微地震数据去噪方法综述", 《吉林大学学报(地球科学版)》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114563821A (en) * | 2020-11-27 | 2022-05-31 | 中国石油天然气股份有限公司 | Intelligent seismic wave impedance inversion label data manufacturing method and device |
CN114563821B (en) * | 2020-11-27 | 2024-05-28 | 中国石油天然气股份有限公司 | Intelligent seismic wave impedance inversion tag data manufacturing method and device |
CN112949089A (en) * | 2021-04-01 | 2021-06-11 | 吉林大学 | Aquifer structure inversion identification method based on discrete convolution residual error network |
CN113156513A (en) * | 2021-04-14 | 2021-07-23 | 吉林大学 | Convolutional neural network seismic signal denoising method based on attention guidance |
CN113156513B (en) * | 2021-04-14 | 2024-01-30 | 吉林大学 | Convolutional neural network seismic signal denoising method based on attention guidance |
CN113191321A (en) * | 2021-05-21 | 2021-07-30 | 电子科技大学 | Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network |
CN113191321B (en) * | 2021-05-21 | 2022-04-22 | 电子科技大学 | Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network |
CN113296150A (en) * | 2021-07-02 | 2021-08-24 | 清华大学 | High-dimensional closed-loop network seismic inversion method under logging constraint |
CN113296150B (en) * | 2021-07-02 | 2022-06-10 | 清华大学 | High-dimensional closed-loop network seismic inversion method under logging constraint |
CN113608258B (en) * | 2021-07-16 | 2024-01-19 | 长江大学 | Self-consistent deep learning method for constructing high-resolution wave impedance inversion label |
CN113608258A (en) * | 2021-07-16 | 2021-11-05 | 长江大学 | Self-consistent deep learning method for constructing high-resolution wave impedance inversion label |
CN113722893B (en) * | 2021-08-12 | 2023-06-09 | 中国石油大学(北京) | Seismic record inversion method, device, equipment and storage medium |
CN113722893A (en) * | 2021-08-12 | 2021-11-30 | 中国石油大学(北京) | Seismic record inversion method, device, equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN111983681B (en) | 2021-09-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111983681B (en) | Seismic wave impedance inversion method based on countermeasure learning | |
Li et al. | Deep-learning inversion of seismic data | |
CN115170943A (en) | Improved visual transform seabed substrate sonar image classification method based on transfer learning | |
CN116047583A (en) | Adaptive wave impedance inversion method and system based on depth convolution neural network | |
Yoo et al. | Impedance inversion based on domain adaptation technique with reconstruction | |
CN113625336A (en) | Seismic wave impedance thin layer inversion method based on full convolution neural network | |
Zu et al. | End-to-end deblending of simultaneous source data using transformer | |
Ferreira et al. | Generating sketch-based synthetic seismic images with generative adversarial networks | |
CN115639605B (en) | Automatic identification method and device for high-resolution fault based on deep learning | |
Dixit et al. | Genetic-evolutionary adaptive moment estimation-based semisupervised deep sequential convolution network for seismic impedance inversion: Application and uncertainty analysis | |
US20230222773A1 (en) | Method for predicting geological features from borehole image logs | |
Koochak et al. | A variability aware GAN for improving spatial representativeness of discrete geobodies | |
CN115731119A (en) | Automatic denoising method based on non-matching geological image | |
Kumar et al. | Deep learning swell noise estimation | |
CN116068644A (en) | Method for improving resolution and noise reduction of seismic data by using generation countermeasure network | |
Zhao et al. | Facies conditional simulation based on VAE-GAN model and image quilting algorithm | |
CN113253350A (en) | Porosity inversion method based on joint dictionary | |
CN113221651A (en) | Seafloor sediment classification method using acoustic propagation data and unsupervised machine learning | |
Jo et al. | Conditioning stratigraphic, rule-Based models with generative adversarial networks: A deepwater lobe, deep learning example | |
Cheung et al. | Stochastic representation and conditioning of process-based geological model by deep generative and recognition networks | |
CN117933319B (en) | Modeling method and system for deep random mixed medium non-stationary model and electronic equipment | |
Gao et al. | A novel automated machine-learning model for lithofacies recognition | |
CN118051793B (en) | Formation data real-time analysis and prediction method and system based on TRD construction method | |
CN117434607A (en) | Intelligent earthquake inversion prediction method based on deep learning | |
Bazargani et al. | Machine learning-based residual moveout picking |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |