CN115453619A - High-resolution seismic imaging method and system for generating countermeasure network based on conditions - Google Patents
High-resolution seismic imaging method and system for generating countermeasure network based on conditions Download PDFInfo
- Publication number
- CN115453619A CN115453619A CN202211113440.9A CN202211113440A CN115453619A CN 115453619 A CN115453619 A CN 115453619A CN 202211113440 A CN202211113440 A CN 202211113440A CN 115453619 A CN115453619 A CN 115453619A
- Authority
- CN
- China
- Prior art keywords
- seismic imaging
- noise
- imaging data
- generator
- free
- 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.)
- Pending
Links
- 238000003384 imaging method Methods 0.000 title claims abstract description 308
- 238000012549 training Methods 0.000 claims description 103
- 238000000034 method Methods 0.000 claims description 23
- 238000009826 distribution Methods 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 8
- 238000005520 cutting process Methods 0.000 claims description 7
- 238000012952 Resampling Methods 0.000 claims description 5
- 230000017105 transposition Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 230000006870 function Effects 0.000 description 55
- 238000013508 migration Methods 0.000 description 7
- 230000005012 migration Effects 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 5
- 238000010276 construction Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000002310 reflectometry Methods 0.000 description 3
- 150000003839 salts Chemical class 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 238000013256 Gubra-Amylin NASH model Methods 0.000 description 1
- 241001647769 Mirza Species 0.000 description 1
- 241001323321 Pluto Species 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013434 data augmentation Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000005755 formation reaction Methods 0.000 description 1
- 230000009191 jumping Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 230000000007 visual effect Effects 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
- G01V1/282—Application of seismic models, synthetic seismograms
-
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Remote Sensing (AREA)
- General Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Environmental & Geological Engineering (AREA)
- Geology (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Geophysics (AREA)
- Theoretical Computer Science (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Processing (AREA)
Abstract
The invention provides a high-resolution seismic imaging method and a high-resolution seismic imaging system for generating a countermeasure network based on conditions, and belongs to the technical field of high-resolution seismic imaging.
Description
Technical Field
The invention relates to the technical field of high-resolution seismic imaging, in particular to a high-resolution seismic imaging method and system based on a condition generation countermeasure network.
Background
Seismic imaging is an important geophysical method for understanding geological structures of underground depth space, resource storage and the like, and compared with other heavy magnetoelectric exploration methods, a seismic method and a seismic graph are methods with the highest resolution capability on stratum characteristics. However, in most cases, seismic imaging is based on the law of motion of the primary reflected wavefield, and contains various types of noise (such as train, wire and other unwanted waveform interference information) which pollute reflection data, the noise exists on a seismic imaging section, the quality of the seismic imaging is seriously reduced, the resolution is sacrificed, and in serious cases, a seismic imaging interpreter cannot identify stratigraphic units from the noise, and high-resolution seismic identification is limited.
Disclosure of Invention
The invention aims to provide a high-resolution seismic imaging method and a high-resolution seismic imaging system for generating a countermeasure network based on conditions.
In order to achieve the purpose, the invention provides the following scheme:
a high resolution seismic imaging method for generating a countermeasure network based upon conditions, the high resolution seismic imaging method comprising the steps of:
constructing a training data set comprising a plurality of pieces of training data; the training data comprises real noise-free seismic imaging data and noise-containing seismic imaging data corresponding to the real noise-free seismic imaging data;
generating a countermeasure network based on the condition, and constructing a seismic imaging model; the seismic imaging model comprises a generator and a discriminator; the generator is used for predicting noise-free seismic imaging data corresponding to the input noise-containing seismic imaging data to obtain predicted noise-free seismic imaging data; the discriminator is used for judging the probability that the noise-free seismic imaging data input into the discriminator comes from the training data set; the noise-free seismic imaging data input into the discriminator comprises real noise-free seismic imaging data in a training dataset and the predicted noise-free seismic imaging data;
performing iterative training on the seismic imaging model by using the training data set, and updating the parameters of the discriminator and the parameters of the generator until the discriminator judges that the probability of the predicted noise-free seismic imaging data from the training data set is higher than a preset threshold value, so as to obtain a trained generator;
and acquiring noise-containing seismic imaging data to be converted, and inputting the noise-containing seismic imaging data to be converted into the generator after training to obtain predicted noise-free seismic imaging data corresponding to the noise-containing seismic imaging data to be converted.
Optionally, the generator is a U-net network; the generator comprises an encoder and a decoder; the encoder comprises a plurality of convolutional layers; the decoder includes the same number of transposed convolutional layers as those in the encoder; and jump connection is adopted between each convolution layer of the encoder and each transposition convolution layer of the decoder.
Optionally, in the encoder, a leakage ReLU and normalization processing are adopted between each convolution layer except for the first convolution layer; in the decoder, each transposed convolutional layer except the last layer of transposed convolutional layer adopts ReLU and normalization processing; the final layer of the transposition convolution layer of the decoder adopts a Tanh nonlinear activation function.
Optionally, the iteratively training the seismic imaging model by using the training data set, and updating the parameters of the discriminator and the parameters of the generator until the discriminator judges that the probability that the predicted noise-free seismic imaging data comes from the training data set is higher than a preset threshold specifically includes:
inputting a plurality of pieces of noise-containing seismic imaging data in a training data set into the generator respectively as conditions to obtain a plurality of pieces of predicted noise-free seismic imaging data;
respectively scoring the plurality of pieces of predicted noise-free seismic imaging data, the real noise-free seismic imaging data corresponding to the plurality of pieces of noise-containing seismic imaging data and the random noise-free seismic imaging data by using the discriminator, and updating parameters of the discriminator according to a loss function of the discriminator so that the discriminator determines that the probability that the real noise-free seismic imaging data comes from the training data set is higher than a threshold value and the probability that the predicted noise-free seismic imaging data comes from the training data set is lower than a preset threshold value, thereby obtaining the optimized discriminator;
scoring the plurality of pieces of predicted noise-free seismic imaging data by using the optimized discriminator, updating parameters of the generator according to a loss function of the generator, and enabling the generator to enable the predicted noise-free seismic imaging data obtained according to the noise-containing seismic imaging data input into the generator to be close to the real noise-free seismic imaging data corresponding to the noise-containing seismic imaging data input into the generator to obtain an optimized generator;
iteratively updating the parameters of the discriminator and the parameters of the generator until the objective function of the seismic imaging model converges; the objective function is that the loss function of the discriminator is the largest and the loss function of the generator is the smallest.
Optionally, the objective function of the seismic imaging noise pressure modeling is as follows:
wherein G is * Is an objective function, theta, of the seismic imaging noise pressure modeling G As a parameter of the generator, θ D As a parameter of the discriminator, L D (θ G ,θ D ) As a loss function of said arbiter, L G (θ G ) For the loss function of the generator, μ is a hyperparameter for balancing the arbiter and the generator; the objective function convergence is achieved with the conditions that the penalty function of the arbiter is maximum and the penalty function of the generator is minimum.
Optionally, the penalty function of the discriminator is as follows:
wherein L is D (θ G ,θ D ) Is a loss function of the discriminator,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is the latent variable for generating the noise-free seismic imaging image,
noise-free seismic imaging data, P, generated for the generator g Distribution, P, of noise-free seismic imaging data generated for the generator r To train the distribution of noise-free seismic imaging data in the dataset,
for the distribution of random noise-free seismic imaging data,
for random noise-free seismic imaging data,
randomly chosen from the data generated by the generator and the data of the training data set,
ε is a uniform random number between 0 and 1, θ G As a parameter of the generator, θ D Lambda is a penalty factor for the parameters of the discriminator,
is a gradient penalty term.
Optionally, the loss function of the generator is as follows:
wherein L is G (θ G ) Is a loss function of the generator, L GAN Is a score of the discriminator, L 1 An L1-norm term for the noise-free seismic imaging data generated by the generator and corresponding noise-free seismic imaging data in the training dataset,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is the latent variable for generating the noise-free seismic imaging image, θ G As a parameter of the generator, θ D Is a parameter of the discriminator.
Optionally, before the iteratively training the seismic imaging model with the training data set, updating the parameters of the discriminator and the parameters of the generator until the discriminator judges that the probability that the predicted noise-free seismic imaging data is from the training data set is higher than a preset threshold, the high-resolution seismic imaging method further includes:
and performing image resampling on the original noise-free seismic imaging data and the original noise-containing seismic imaging data according to the preset sizes of the generator and the discriminator aiming at any group of original noise-free seismic imaging data and the corresponding original noise-containing seismic imaging data to obtain the cutting noise-free seismic imaging data and the corresponding cutting noise-containing seismic imaging data which accord with the preset sizes.
Optionally, after the image resampling is performed on the original noiseless seismic imaging data and the original noisy seismic imaging data according to the preset sizes of the generator and the discriminator to obtain the clipped noiseless seismic imaging data and the corresponding clipped noisy seismic imaging data which conform to the preset sizes, the high-resolution seismic imaging method further includes:
performing data expansion processing on any group of the cut noise-free seismic imaging data and the corresponding cut noise-containing seismic imaging data to obtain new training data; the data expansion process includes any one of random inversion, random rotation, and random noise.
Corresponding to the aforementioned high-resolution seismic imaging method, the invention also provides a high-resolution seismic imaging system for generating a confrontational network based on conditions, which when executed by a computer performs the high-resolution seismic imaging method for generating a confrontational network based on conditions as described above.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention provides a high-resolution seismic imaging method for generating an anti-network based on conditions, which is characterized in that the anti-network is generated based on the conditions, a seismic imaging noise suppression model is constructed, on the basis of randomly generating noise-free seismic imaging data by the conventional generation anti-network, the noise-containing seismic imaging data is used as the conditions, the noise-free seismic imaging data output by a generator can keep the structural characteristics of the noise-containing seismic imaging data, the effect of suppressing noise in the noise-containing seismic imaging data is achieved, the parameters of the generator and a discriminator are iteratively updated through a training data set, effective signals are identified and noise signals are suppressed through a mutual game mode of the generator and the discriminator until the discriminator cannot judge whether the input noise-free seismic imaging data is the noise-free seismic imaging data generated by the generator, the noise-free seismic imaging data obtained after the noise-containing seismic imaging data is denoised by the generator can be ensured to be more fit with the real noise-free seismic imaging data, the noise removal of the seismic imaging data is realized, and the high-resolution seismic imaging data has great application value and practical significance in promoting intelligent seismic data processing and high-efficiency exploration.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
Fig. 1 is a flowchart of a high-resolution seismic imaging method for generating a countermeasure network based on conditions according to embodiment 1 of the present invention;
FIG. 2 is a schematic illustration of noisy seismic imaging data and noise-free seismic imaging data in a high resolution seismic imaging method as provided in example 1 of the invention;
fig. 3 is a schematic structural diagram of a generator in the high-resolution seismic imaging method according to embodiment 1 of the present invention;
fig. 4 is a schematic structural diagram of a discriminator in the high-resolution seismic imaging method according to embodiment 1 of the present invention;
FIG. 5 is a schematic diagram of the input and output of a generator and an arbiter in the high resolution seismic imaging method according to embodiment 1 of the present invention; wherein D represents a discriminator and G represents a generator;
fig. 6 is a schematic diagram of an intermediate result in a training process in the high-resolution seismic imaging method according to embodiment 1 of the present invention;
fig. 7 is a schematic diagram of a generator in the high-resolution seismic imaging method according to embodiment 1 of the present invention generating noise-free seismic imaging data from noisy seismic imaging data; wherein a) represents noisy seismic imaging data, b) represents noise-free seismic imaging data;
fig. 8 is a schematic structural diagram of a high-resolution seismic imaging system for generating a countermeasure network based on conditions according to embodiment 2 of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the 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, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Currently, in the field of seismic data processing, deep Neural Networks (DNNs) (Lecun et al, 2015) have received much attention as a promising framework of methods that provide the most advanced capabilities for image classification, speech recognition, target detection, and many other fields related to processing unstructured data. The method has strong application capability in the field of geophysical. The DNN consists of multiple layers for learning the characteristics of given data.
Convolutional Neural Networks (CNNs) are the most popular DNN architectures for image classification (Krizhevsky et al, 2017), semantic segmentation (Ronneberger et al, 2015), and data reconstruction (Zhang et al, 2017). CNNs use convolution, combining, and shared weights on multiple layers of DNNs instead of multi-layer perceptrons (MLPs), which greatly reduces the parameters that need to be trained. CNNs, which consist of convolutional layers only, are called complete convolutional networks (FCNs) (Long et al, 2015), which have proven to be an end-to-end pixel-level semantic segmentation solution. In the field of exploratory seismology, CNNs have been successfully applied to fault detection (Xiong et al, 2018), first arrival wave picking (Yuan et al, 2018), full waveform inversion (Lewis and Vigh, 2017), salt dome structure identification classification (Shi et al, 2018) and parametric model construction (Zhang et al, 2022).
In many types of network architectures, generative countermeasure networks (GANs) (Goodfello et al, 2014) are essentially a generative model for learning data sets and generating artificial objects similar to those in real life. This is a novel approach and the idea of GAN is simple. The GAN consists of two DNNs, a generator G and a discriminator D. G creates a dummy object and D tells the difference between the generated object and the real object. G and D are just as two competing players, and when they both play in the best mode, the balance will be achieved if their opponents are the best. GAN is a research hotspot in recent years, and several GANs with different structures have been proposed by academia (Radford et al, 2015, mao et al, 2016 li et al, 2018) to generate data for specific regions. The condition GAN is an extension of GAN (Mirza and osidero, 2014). The only difference between the conditions GAN and GAN is that the former additionally provides a given data as input to the generator, not just random noise. Conditional GAN can be used for image super resolution (Ledig et al, 2017), image deblurring (Kupyn et al, 2017), and image-to-image conversion (Isola et al, 2017).
In high resolution seismic imaging tasks, the reflected signal needs to be undistorted and where the noise needs to be trained to be removed. CGAN was trained on deblurred GAN (Deblu-GAN), which was developed from pix2pix GAN, which implements an end-to-end dynamic deblurring method. The advantage of the Deblu-GAN is that the visual performance is improved while the image structure similarity is maintained, and the task target is met.
The invention aims to provide a high-resolution seismic imaging method and system for generating a confrontation network based on conditions.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
Example 1:
the embodiment provides a high-resolution seismic imaging method for generating a countermeasure network based on conditions, as shown in a flow chart shown in fig. 1, and the high-resolution seismic imaging method comprises the following steps:
s1, constructing a training data set comprising a plurality of pieces of training data; the training data includes true noise-free seismic imaging data and noise-containing seismic imaging data corresponding to the true noise-free seismic imaging data. In this embodiment, the real noise-free seismic imaging data is a reflectivity image obtained through an accurate velocity model, the noise-containing seismic imaging data is migration imaging data obtained through an RTM migration technique, and the difference between the two data is shown in fig. 2, where the left side is migration imaging with noise, and the right side is a noise-free reflectivity image.
S2, generating a countermeasure network based on conditions, and constructing a seismic imaging model; the seismic imaging model comprises a generator and a discriminator; the generator is used for predicting noise-free seismic imaging data corresponding to the input noise-containing seismic imaging data to obtain predicted noise-free seismic imaging data; the discriminator is used for judging the probability that the noise-free seismic imaging data input into the discriminator comes from the training data set; the noise-free seismic imaging data input into the discriminator includes true noise-free seismic imaging data and predicted noise-free seismic imaging data in the training dataset.
S3, iteratively training the seismic imaging model to obtain a trained seismic imaging model; in this embodiment, the training data set is used to perform iterative training on the seismic imaging model, and the parameters of the discriminator and the parameters of the generator are updated until the discriminator determines that the probability of predicting the noise-free seismic imaging data from the training data set is higher than the preset threshold, so as to obtain the generator after training.
S4, applying the trained seismic imaging model; in this embodiment, noise-containing seismic imaging data to be converted is acquired, and the noise-containing seismic imaging data to be converted is input to a generator that has been trained, so as to obtain predicted noise-free seismic imaging data corresponding to the noise-containing seismic imaging data to be converted.
Specifically, in the embodiment, the generator of the seismic imaging model is a U-net network, and the structure of the U-net network is shown in fig. 3, wherein the U-net network comprises an encoder and a decoder; wherein, the encoder comprises a plurality of convolution layers, and the decoder comprises the same number of transposed convolution layers as the convolution layers in the encoder; and jump connection is adopted between each convolution layer of the encoder and each transposition convolution layer of the decoder. In the encoder, leaky ReLU and normalization processing are adopted among convolution layers except the first layer of convolution layer; in the decoder, the transposed convolution layers except the last layer adopt Tanh nonlinear activation function, the other transposed convolution layers adopt ReLU and normalization processing, and the discriminator adopts a Patch-GAN model, and the structure is shown in FIG. 4.
Specifically to this embodiment, step S3 may specifically include the following steps:
s31, generating a plurality of pieces of predicted noise-free seismic imaging data by using a generator; in this embodiment, a plurality of pieces of noise-containing seismic imaging data in the training dataset are input to the generator as conditions, respectively, to obtain a plurality of pieces of predicted noise-free seismic imaging data.
S32, updating parameters of the discriminator according to the loss function of the discriminator; specifically, a discriminator is used for respectively scoring a plurality of pieces of predicted noiseless seismic imaging data, real noiseless seismic imaging data corresponding to a plurality of pieces of noisy seismic imaging data and random noiseless seismic imaging data, and parameters of the discriminator are updated according to a loss function of the discriminator, so that the probability that the discriminator judges that the real noiseless seismic imaging data come from a training data set is higher than a preset threshold value and the probability that the predicted noiseless seismic imaging data come from the training data set is lower than the preset threshold value, namely the discriminator can discriminate whether the noiseless seismic imaging data input into the discriminator are predicted noiseless seismic imaging data generated by a generator, and the optimized discriminator is obtained.
S33, updating the parameters of the generator according to the parameters of the discriminator and the loss function of the generator; specifically, the optimized discriminator is used for scoring a plurality of pieces of predicted noise-free seismic imaging data, parameters of the generator are updated according to a loss function of the generator, and the predicted noise-free seismic imaging data obtained by the generator according to the noise-containing seismic imaging data input into the generator is close to the real noise-free seismic imaging data corresponding to the noise-containing seismic imaging data input into the generator, so that the optimized generator is obtained.
S34, judging whether the objective function of the seismic imaging model is converged; if not, jumping to the step S32, and repeatedly executing the steps S32-S33 to iteratively update the parameters of the discriminator and the parameters of the generator; and if so, stopping iterative updating. Specifically, the objective function of the seismic imaging model is that the loss function of the discriminator is the largest and the loss function of the generator is the smallest; when the objective function of the seismic imaging model is converged, the discriminator judges that the probability of the predicted noise-free seismic imaging data from the training data set is higher than a preset threshold value, namely, whether the noise-free seismic imaging data input into the discriminator is the predicted noise-free seismic imaging data predicted by the generator cannot be judged, and the purpose that the discriminator is successfully deceived by the generator is achieved.
The objective function of the seismic imaging noise pressure modeling is shown as follows:
wherein, G * Modeling the objective function, theta, for seismic imaging noise G To be a parameter of the generator, θ D As a parameter of the discriminator, L D (θ G ,θ D ) As a loss function of the discriminator, L G (θ G ) To generate the loss function of the generator, μ is the hyperparameter used to balance the arbiter and the generator; the arbiter expects the objective function to be the largest and the generator expects the objective function to be the smallest.
Loss function L of discriminator D (θ G ,θ D ) As shown in the following formula:
wherein, the first and the second end of the pipe are connected with each other,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is the latent variable for generating the noise-free seismic imaging image,
noise-free seismic imaging data, P, generated for the generator g Distribution of noise-free seismic imaging data generated for a generator, P r To train the distribution of noise-free seismic imaging data in the dataset,
for the distribution of random noise-free seismic imaging data,
for random noise-free seismic imaging data,
randomly chosen from the data generated by the generator and the data of the training data set,
ε is a uniform random number between 0 and 1, θ G To the parameters of the generator, θ D Lambda is the penalty factor for the parameters of the discriminator,
is a gradient penalty term.
Loss function L of the generator G (θ G ) As shown in the following formula:
wherein L is GAN As a score term for the discriminator, L 1 An L1-norm term for the noise-free seismic imaging data generated by the generator and corresponding noise-free seismic imaging data in the training dataset,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is the latent variable for generating the noise-free seismic imaging image, θ G To be a parameter of the generator, θ D Are parameters of the discriminator.
In order to make the size of the noisy seismic imaging data required to be input into the seismic imaging model consistent, before step S1, the high resolution seismic imaging method further comprises:
and performing image resampling on the original noise-free seismic imaging data and the original noise-containing seismic imaging data according to the preset sizes of the generator and the discriminator aiming at any group of original noise-free seismic imaging data and the corresponding original noise-containing seismic imaging data to obtain the cutting noise-free seismic imaging data and the corresponding cutting noise-containing seismic imaging data which accord with the preset sizes.
Further, in order to improve the data volume participating in training the seismic imaging model, the high-resolution seismic imaging method further includes:
carrying out data expansion processing on any one group of the cut noise-free seismic imaging data and the corresponding cut noise-containing seismic imaging data to obtain new training data; the data expansion process includes any one of random inversion, random rotation, and random noise.
The high-resolution seismic imaging method based on the condition generation countermeasure network provided by the invention is described below by combining a specific example, and comprises the following steps:
a1, performing pre-stack migration imaging on seismic shot gather data of a target work area to obtain migration imaging data, converting the migration imaging data into an image format to obtain a training data set comprising a plurality of pieces of noise-containing seismic imaging data (migration imaging data) and corresponding noise-free seismic imaging data (reflectivity data), and using the training data set as data input by a training seismic imaging model.
And A2, randomly cutting out 32 images with the size of 256x256 from the training data set, and performing data augmentation expansion before inputting the condition GAN. For example, in the embodiment, since the generator uses a U-net network and the discriminator uses a markov discriminator (Patch-GAN), 32 or 64 images with a size of 256 × 256 need to be randomly cropped from the training data set. And carrying out random rotation, random inversion or random noise on the obtained image data.
And A3, generating a confrontation network based on the condition, and constructing a seismic imaging model. The conditional generation countermeasure network consists of two DNNs (producers and discriminators) that game each other. In one aspect, a generator captures conditional data distribution and generates spurious data to spoof the arbiter upon inputting additional conditional data into the data space. On the other hand, the discriminator D estimates the conditional probability that the sample input into the discriminator comes from the training data set to discriminate between true and false. FIG. 5 illustrates the structure of a conditionally generated countermeasure network, with the generator and the discriminator having respective independent loss functions, and the generator reconstructing a resulting, i.e., generated, noise-free seismic image G (z | y) from a noise profile z and an applied condition (noisy seismic image) y; the discriminator determines the probability D (x | y) that the noise-free seismic image is from the training data set based on the input noise-containing seismic image y and the noise-free seismic image x corresponding to y, note that the noise-free seismic image x corresponding to y input to the discriminator here may be the true noise-free seismic image (i.e., the noise-free reflectance image) from the training data set or may be the predicted noise-free seismic image predicted by the generator.
A31, constructing a generator network structure by using a U-net network, and defining a loss function; u-net is named after its U-network architecture, which is, in theory, a coding-decoding fully-connected network (FCN) model consisting of a compressed down-sampling path and an extended up-sampling path symmetrical thereto. In the down-sampling stage, the image size gradually decreases as the number of channels increases. The process is reversed when the data flow enters the upsampling stage through a bottleneck layer in the middle of the network. One advantage of the FCN model is that the input size of the image is not limited. Thus, images of any size can be inferred without modifying the architecture. An efficient way is to communicate low-level information over a network. The U-net network employs a hopping connection between the ith layer and the nth-i layer, where n is the depth of the U-net network, which has proven to be effective over small data sets while avoiding the over-fitting problem.
The architecture of the U-net used in the generator in this example includes an encoder and a decoder. There are 16 layers in the encoder-decoder architecture, and 54409603 weight parameters need to be trained. In the encoder section, 8 convolutional layers are included, and the remaining layers, except the first convolutional layer, employ a leaky ReLU with a slope of 0.2 and normalization processing. In the decoder section, 8 transposed convolutional layers are included, the first 7 layers employ ReLU and normalization, the last layer employs Tanh as the nonlinear activation function, and 50% probability Dropout is applied in the middle layers of the decoder section to enhance generalization performance and alleviate the overfitting problem.
The loss function of the generator mainly comprises two parts: the discriminator score part and the L1 norm loss term, L1 norm measures the distance between the actual output and the standard answer in order to make the output closer to the input. The L2 norm may cause output ambiguity. The result shows that the performance of the L1 norm in the seismic image denoising is superior to that of the L2 norm. In this example, the loss function of the generator can be expressed as:
wherein L is GAN A score of a discriminator, L 1 For the L1-norm term of the noise-free seismic imaging data generated by the generator and the corresponding noise-free seismic imaging data in the training dataset,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is the latent variable for generating the noise-free seismic imaging image, θ G To the parameters of the generator, θ D Are parameters of the discriminator.
A32, using a Markov chain countermeasure network Patch-GAN as a discriminator and defining a loss function of the discriminator; patch GAN is also a FCN model, proposed as a texture loss model. It treats the image as a markov random field, assuming that the independence between pixels is greater than one bin diameter. The Patch GAN runs over the image in a convolution manner, penalizing a portion of the image each time in proportion to the Patch. The final output of the arbiter is the average of all responses of Patch.
In this example, the Patch GAN arbiter is a five-layer FCN model with parameters 2764737, with the exception of the last layer, where the leakage ReLU with normalized sum slope of 0.2 is applied, and the final score of the arbiter is the average of the arbiter outputs.
Regarding the loss function of the arbiter, wasserstein GAN (WGAN-GP) with a gradient penalty is used in this example as the loss function for the arbiter training. By using the Wassertein-1 distance, rather than the Jensen-Shannon distance used in conventional GANs, WGAN-GP has proven to be powerful and robust in various GAN structures. Furthermore, the Wassertein-1 distance is related to the performance of the GAN, which allows the performance of the GAN to be evaluated by a loss function, whereas the Jensen-Shannon distance does not enable this function. In addition, the WGAN-GP enforces the Lipschitz constraint using gradient penalties rather than simple weight function compilation, making the training process more stable. The penalty function of the arbiter can be expressed as:
wherein the content of the first and second substances,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is the latent variable for generating the noise-free seismic imaging image,
for noise-free seismic imaging data generated by the generator, P g Distribution of noise-free seismic imaging data generated by a generator, P r To train the distribution of noise-free seismic imaging data in the dataset,
for the distribution of random noise-free seismic imaging data,
for random noise-free seismic imaging data,
randomly chosen from the data generated by the generator and the data of the training data set,
ε is a uniform random number between 0 and 1, θ G To the parameters of the generator, θ D As parameters of the discriminator, λ is a penalty coefficient,
is a gradient penalty term.
A33, when the arbiter tries to increase the arbiter loss, the generator tries to minimize the generator loss. Then the final objective function of the seismic imaging model is determined as follows:
where μ is a hyper-parameter for balancing the discriminator D and the generator G. In the experiment, μ =100 was set as recommended by Deblur GAN.
A4, iteratively training parameters of the seismic imaging model; the method comprises the following steps: and training the discriminator for multiple times, and updating parameters of the discriminator. The generator is then trained and generator parameters are updated based on the updated discriminator parameters. These two steps iterate back and forth until the generator converges to a steady state.
In this example, the arbiter is trained several times, and the parameters of the arbiter are updated. Secondly, when training the generator, the generator parameters are updated according to the updated discriminator parameters. These two steps of the discriminant training and the generator training are iterated back and forth until the seismic imaging model converges to a steady state. In the process of training the discriminator and the generator, parameters of the discriminator and parameters of the generator are optimized by adopting an Adam optimization method. The Adam optimization method introduces three hyper-parameters alpha and beta 1 ，β 2 Where α is the learning rate, β 1 ，β 2 Is the coefficient that calculates the running average of the gradient and its square. In the training process of the application, the distribution of the hyperparameters is respectively alpha =0.0001 and beta 1 ＝0，β 2 ＝0.9。
The deep learning framework used in this example was pytoch, training was performed on a single Nvidia GTX 3080Ti graphics device, and the seismic imaging model was trained using 6 classical synthetic seismic synthetic data (obtained using a speed model forward modeling) including canadian undulating surface synthetic data, hess oil company synthetic data, french oil research institute Marmousi synthetic data, pluto synthetic data, eastern geophysical salt dome synthetic data, and mexico gulf Sigsbee synthetic data, with the characteristics of the data set shown in table 1. The training data set includes large dip reflection formations, thrust cover constructions, relief topography, salt, scattering points, and the like. Wherein the sigbee data was used as validation data and the remaining 5 models were used as training data.
TABLE 1 attributes of training data sets
At each training session, 32 images of size 256x256 were randomly cropped from the training data set and subjected to data expansion prior to input to the generator. The training time was 12000 times, and several intermediate results of training are shown in fig. 6. As shown in fig. 6, the generator starts with learning the image structure and then progressively deepens the details. After about 300 iterations, the generator is able to describe the fine structure of the image. The error of the generator loss function gradually converges to a steady state, and after about 4000 cycles, the training process converges when the generator loss function becomes stable.
And A5, the noise-containing seismic imaging data needing denoising is wholly input into a generator, and the denoised noise-free seismic imaging data can be obtained at one time through the trained generator.
For noisy seismic imaging data input to the generator, the generator will work properly when the width and height of the image are multiples of 256, but if not, the filling of the image to meet the requirements of the generator is used in this example.
And A6, verifying the seismic imaging model. The Structural Similarity Index (SSIM) is a classical method for measuring perceptual quality, and in this example, a sigsebe model is used to verify the generalization performance of GAN, and is used as a means for measuring the similarity between the predicted image y and the real image x generated by the generator. SSIM evaluates image similarity from three perspectives, brightness, contrast, and structure. SSIM can be expressed as:
wherein, mu x And mu y Respectively image x and imageAverage value of pixel gray levels of image y, σ x 2 And σ y 2 Variance of pixel gray levels, σ, in images x and y, respectively xy Is the covariance of the pixel gray levels in image x and image y, c 1 And c 2 Two variables for stable division with weak denominator, respectively. The SSIM function is symmetrical, i.e., SSIM (x, y) = SSIM (y, x), and the value of the SSIM function ranges from 0 to 1 and is only reached when the two input images are identical.
As can be seen from fig. 7, in fig. 7, a) is the noisy seismic imaging data, and b) is the effect of denoising the noisy seismic imaging data, it can be found that the structure in the image is not deformed, and the noise is suppressed to a large extent, especially the noise near the shot point. In addition, SSIM is rising to be stable, and the condition generation countermeasure network has preliminary potential in denoising problems of high-resolution seismic imaging.
Example 2:
the method of embodiment 1 of the present invention can also be implemented by means of the architecture of a high-resolution seismic imaging system for generating a countermeasure network based on conditions as shown in fig. 8. As shown in fig. 8, the high resolution seismic imaging system may include: the system comprises a training data set construction module, a seismic imaging model construction module and a seismic imaging model training module; some modules may also have sub-units for implementing their functions, for example including a discriminator training unit and a generator training unit in a seismic imaging model training module. Of course, the architecture shown in FIG. 8 is merely exemplary, and in some embodiments, other elements may be added to some of the modules; in addition, when different functions are required, one or at least two components of the system shown in fig. 8 may be omitted according to actual needs.
Specific examples are used herein, but the foregoing description is only illustrative of the principles and embodiments of the present invention, and the description of the examples is only provided to assist understanding of the method and the core concept of the present invention; those skilled in the art will appreciate that the modules or steps of the invention described above can be implemented using general purpose computing apparatus, or alternatively, they can be implemented using program code executable by computing apparatus, such that it is executed by computing apparatus when stored in a storage device, or separately fabricated into integrated circuit modules, or multiple modules or steps thereof can be fabricated into a single integrated circuit module. The present invention is not limited to any specific combination of hardware and software.
Meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.
Claims (10)
1. A high resolution seismic imaging method for generating a countermeasure network based on a condition, the high resolution seismic imaging method comprising:
constructing a training data set comprising a plurality of pieces of training data; the training data comprises real noise-free seismic imaging data and noise-containing seismic imaging data corresponding to the real noise-free seismic imaging data;
generating a confrontation network based on the condition, and constructing a seismic imaging model; the seismic imaging model comprises a generator and a discriminator; the generator is used for predicting noise-free seismic imaging data corresponding to the input noise-containing seismic imaging data to obtain predicted noise-free seismic imaging data; the discriminator is used for judging the probability that the noise-free seismic imaging data input into the discriminator comes from the training data set; the noise-free seismic imaging data input into the discriminator comprises real noise-free seismic imaging data in a training dataset and the predicted noise-free seismic imaging data;
performing iterative training on the seismic imaging model by using the training data set, and updating the parameters of the discriminator and the parameters of the generator until the discriminator judges that the probability of the predicted noise-free seismic imaging data from the training data set is higher than a preset threshold value, so as to obtain a trained generator;
and acquiring noise-containing seismic imaging data to be converted, and inputting the noise-containing seismic imaging data to be converted into the generator after training to obtain predicted noise-free seismic imaging data corresponding to the noise-containing seismic imaging data to be converted.
2. The high resolution seismic imaging method of claim 1, wherein the generator is a U-net network; the generator comprises an encoder and a decoder; the encoder comprises a plurality of convolution layers; the decoder includes the same number of transposed convolutional layers as those in the encoder; and jump connection is adopted between each convolution layer of the encoder and each transposition convolution layer of the decoder.
3. The high resolution seismic imaging method of claim 2, wherein in the encoder, a Leaky ReLU and normalization are used between each convolutional layer except the first convolutional layer; in the decoder, each transposed convolutional layer except the last layer of transposed convolutional layer adopts ReLU and normalization processing; the final layer of the transposition convolution layer of the decoder adopts a Tanh nonlinear activation function.
4. The high resolution seismic imaging method of claim 1, wherein iteratively training the seismic imaging model using the training dataset, updating the parameters of the discriminator and the parameters of the generator until the discriminator determines that the probability that the predicted noise-free seismic imaging data is from the training dataset is higher than a predetermined threshold, specifically comprising:
inputting a plurality of pieces of noise-containing seismic imaging data in a training data set into the generator respectively as conditions to obtain a plurality of pieces of predicted noise-free seismic imaging data;
respectively scoring the plurality of pieces of predicted noise-free seismic imaging data, the real noise-free seismic imaging data corresponding to the plurality of pieces of noise-containing seismic imaging data and the random noise-free seismic imaging data by using the discriminator, and updating parameters of the discriminator according to a loss function of the discriminator so that the probability that the discriminator judges that the real noise-free seismic imaging data comes from the training data set is higher than a threshold value and the probability that the predicted noise-free seismic imaging data comes from the training data set is lower than a preset threshold value, thereby obtaining the optimized discriminator;
scoring the plurality of pieces of predicted noise-free seismic imaging data by using the optimized discriminator, updating parameters of the generator according to a loss function of the generator, and enabling the generator to enable the predicted noise-free seismic imaging data obtained according to the noise-containing seismic imaging data input into the generator to be close to the real noise-free seismic imaging data corresponding to the noise-containing seismic imaging data input into the generator to obtain an optimized generator;
iteratively updating the parameters of the discriminator and the parameters of the generator until the objective function of the seismic imaging model converges; the objective function is that the penalty function of the arbiter is the largest and the penalty function of the generator is the smallest.
5. The high resolution seismic imaging method of claim 4, wherein the objective function of the seismic imaging noise pressure modeling is as follows:
wherein G is * Modeling an objective function, θ, for the seismic imaging noise G As a parameter of the generator, θ D As a parameter of the discriminator, L D (θ G ,θ D ) Is a loss function of the discriminator, L G (θ G ) For the loss function of the generator, μ is a hyperparameter for balancing the arbiter and the generator; the objective function convergence is achieved with the conditions that the penalty function of the arbiter is maximum and the penalty function of the generator is minimum.
6. The high resolution seismic imaging method of claim 4, wherein the loss function of the discriminator is given by:
wherein L is D (θ G ,θ D ) Is a function of the loss of the arbiter,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is a latent variable for generating the noise-free seismic imaging image,
noise-free seismic imaging data, P, generated for the generator g Distribution of noise-free seismic imaging data, P, generated for the generator r To train the distribution of noise-free seismic imaging data in the dataset,
for the distribution of random noise-free seismic imaging data,
for random noise-free seismic imaging data,
randomly chosen from the data generated by the generator and the data of the training data set,
ε is a uniform random number between 0 and 1, θ G As a parameter of the generator, θ D Lambda is a penalty factor for the parameters of the discriminator,
is a gradient penalty term.
7. The high resolution seismic imaging method of claim 4, wherein the loss function of the generator is as follows:
wherein L is G (θ G ) Is a loss function of the generator, L GAN Is a score of the discriminator, L 1 An L1-norm term for the noise-free seismic imaging data generated by the generator and corresponding noise-free seismic imaging data in the training dataset,
for expectation, x is the noise-free seismic imaging data in the training dataset, y is the noisy seismic imaging image, z is the latent variable for generating the noise-free seismic imaging image, θ G As a parameter of the generator, θ D Is a parameter of the discriminator.
8. The high resolution seismic imaging method of claim 1, wherein, prior to the iterative training of the seismic imaging model with the training dataset, updating parameters of the discriminator and parameters of the generator until the discriminator determines that the probability of the predicted noise-free seismic imaging data from the training dataset is above a preset threshold, the high resolution seismic imaging method further comprises:
and performing image resampling on the original noise-free seismic imaging data and the original noise-containing seismic imaging data according to the preset sizes of the generator and the discriminator aiming at any group of original noise-free seismic imaging data and the corresponding original noise-containing seismic imaging data to obtain the cutting noise-free seismic imaging data and the corresponding cutting noise-containing seismic imaging data which accord with the preset sizes.
9. The high resolution seismic imaging method of claim 8, wherein after the image resampling of the original noise-free seismic imaging data and the original noise-containing seismic imaging data according to the preset sizes of the generator and the discriminator to obtain cropped noise-free seismic imaging data and corresponding cropped noise-containing seismic imaging data that conform to the preset sizes, the high resolution seismic imaging method further comprises:
performing data expansion processing on any group of the cut noise-free seismic imaging data and the corresponding cut noise-containing seismic imaging data to obtain new training data; the data expansion process includes any one of random flipping, random rotation, and random noise.
10. A high resolution seismic imaging system for generating a countermeasure network based on conditions, wherein the high resolution seismic imaging system, when executed by a computer, performs the high resolution seismic imaging method of any of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211113440.9A CN115453619A (en) | 2022-09-14 | 2022-09-14 | High-resolution seismic imaging method and system for generating countermeasure network based on conditions |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211113440.9A CN115453619A (en) | 2022-09-14 | 2022-09-14 | High-resolution seismic imaging method and system for generating countermeasure network based on conditions |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115453619A true CN115453619A (en) | 2022-12-09 |
Family
ID=84302081
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211113440.9A Pending CN115453619A (en) | 2022-09-14 | 2022-09-14 | High-resolution seismic imaging method and system for generating countermeasure network based on conditions |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115453619A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108765319A (en) * | 2018-05-09 | 2018-11-06 | 大连理工大学 | A kind of image de-noising method based on generation confrontation network |
| US20190302290A1 (en) * | 2018-03-27 | 2019-10-03 | Westerngeco Llc | Generative adversarial network seismic data processor |
| CN110473154A (en) * | 2019-07-31 | 2019-11-19 | 西安理工大学 | A kind of image de-noising method based on generation confrontation network |
| WO2021051050A1 (en) * | 2019-09-12 | 2021-03-18 | Schlumberger Technology Corporation | Generating geological facies models with fidelity to the diversity and statistics of training images using improved generative adversarial networks |
| WO2021253316A1 (en) * | 2020-06-18 | 2021-12-23 | 深圳先进技术研究院 | Method and apparatus for training image noise reduction model, electronic device, and storage medium |
-
2022
- 2022-09-14 CN CN202211113440.9A patent/CN115453619A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190302290A1 (en) * | 2018-03-27 | 2019-10-03 | Westerngeco Llc | Generative adversarial network seismic data processor |
| CN108765319A (en) * | 2018-05-09 | 2018-11-06 | 大连理工大学 | A kind of image de-noising method based on generation confrontation network |
| CN110473154A (en) * | 2019-07-31 | 2019-11-19 | 西安理工大学 | A kind of image de-noising method based on generation confrontation network |
| WO2021051050A1 (en) * | 2019-09-12 | 2021-03-18 | Schlumberger Technology Corporation | Generating geological facies models with fidelity to the diversity and statistics of training images using improved generative adversarial networks |
| WO2021253316A1 (en) * | 2020-06-18 | 2021-12-23 | 深圳先进技术研究院 | Method and apparatus for training image noise reduction model, electronic device, and storage medium |
Non-Patent Citations (2)
| Title |
|---|
| ZHANG H，ET AL: "Imaging Domain Seismic Denoising Based on Conditional Generative Adversarial Networks (CGANs)", ENERGIES, vol. 15, no. 18, 8 September 2022 (2022-09-08), pages 1 - 14 * |
| 焦李成: "计算智能导论", 30 September 2019, 西安电子科技大学出版社, pages: 303 - 310 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Generating geologically realistic 3D reservoir facies models using deep learning of sedimentary architecture with generative adversarial networks | |
| US11645835B2 (en) | Hypercomplex deep learning methods, architectures, and apparatus for multimodal small, medium, and large-scale data representation, analysis, and applications | |
| Gai et al. | New image denoising algorithm via improved deep convolutional neural network with perceptive loss | |
| CN111242071B (en) | Attention remote sensing image target detection method based on anchor frame | |
| Liu et al. | Random noise suppression in seismic data: What can deep learning do? | |
| CN107688201B (en) | RBM-based seismic prestack signal clustering method | |
| CN112946749B (en) | Method for suppressing seismic multiples based on data augmentation training deep neural network | |
| Fattahi et al. | Prediction of porosity and water saturation using pre-stack seismic attributes: a comparison of Bayesian inversion and computational intelligence methods | |
| CN112731522B (en) | Intelligent recognition method, device and equipment for seismic stratum and storage medium | |
| CN113554084B (en) | Vehicle re-identification model compression method and system based on pruning and light convolution | |
| CN110895348B (en) | Method, system and storage medium for extracting low-frequency information of seismic elastic impedance | |
| Civitarese et al. | Semantic segmentation of seismic images | |
| Wang et al. | Learning from noisy data: An unsupervised random denoising method for seismic data using model-based deep learning | |
| Samson | Mineral resource estimates with machine learning and geostatistics | |
| CN112882123B (en) | CNN well-seismic joint inversion method, system and application based on two-step method | |
| Gao et al. | Seismic random noise attenuation based on M-ResUNet | |
| CN116068644B (en) | Method for improving resolution and noise reduction of seismic data by using generation countermeasure network | |
| CN115453619A (en) | High-resolution seismic imaging method and system for generating countermeasure network based on conditions | |
| US12013508B2 (en) | Method and system for determining seismic processing parameters using machine learning | |
| CN115639605B (en) | Automatic identification method and device for high-resolution fault based on deep learning | |
| Ravasi | Deep preconditioners and their application to seismic wavefield processing | |
| She et al. | Incorporating momentum acceleration techniques applied in deep learning into traditional optimization algorithms | |
| Santos | Geostatistical Subsurface Modeling Using Generative Adversarial Networks | |
| US12032111B2 (en) | Method and system for faster seismic imaging using machine learning | |
| US20220283329A1 (en) | Method and system for faster seismic imaging using machine learning |
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 |