CN113885077B - Multi-source seismic data separation method based on deep learning - Google Patents
Multi-source seismic data separation method based on deep learning Download PDFInfo
- Publication number
- CN113885077B CN113885077B CN202111162037.0A CN202111162037A CN113885077B CN 113885077 B CN113885077 B CN 113885077B CN 202111162037 A CN202111162037 A CN 202111162037A CN 113885077 B CN113885077 B CN 113885077B
- Authority
- CN
- China
- Prior art keywords
- data
- seismic
- source
- convolution
- layer
- 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.)
- Active
Links
- 238000000926 separation method Methods 0.000 title claims abstract description 43
- 238000013135 deep learning Methods 0.000 title claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 40
- 238000000034 method Methods 0.000 claims abstract description 32
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 24
- 230000008569 process Effects 0.000 claims abstract description 10
- 238000002156 mixing Methods 0.000 claims description 17
- 230000005284 excitation Effects 0.000 claims description 15
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 238000010304 firing Methods 0.000 claims description 5
- 238000013507 mapping Methods 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 5
- 230000003111 delayed effect Effects 0.000 claims description 4
- 238000012937 correction Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 6
- 238000012545 processing Methods 0.000 abstract description 4
- 238000001914 filtration Methods 0.000 description 15
- 238000010586 diagram Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000000087 stabilizing effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
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
Landscapes
- Engineering & Computer Science (AREA)
- Remote Sensing (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Acoustics & Sound (AREA)
- Environmental & Geological Engineering (AREA)
- Geology (AREA)
- General Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Geophysics (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention discloses a multi-source seismic data separation method based on deep learning, and belongs to the technical field of seismic data processing; the method comprises two main steps of CNN deep learning and multi-focus mixed data separation; the convolutional neural network deep learning technology is introduced into random aliasing denoising of multi-source seismic data, more complex data features in the seismic data are extracted through the convolutional neural network, the damage to the seismic signals of a main seismic source is small, and the denoising process is intelligently realized; according to the characteristics of random aliasing noise in multi-source mixed data, all seismic channels of original mixed data are subjected to random time coding, and the input of training samples containing random noise distribution is constructed, wherein the training samples are derived from seismic data and contain wave field characteristics of actual seismic data, and the problem that deep learning lacks training samples in actual multi-source separation application is solved.
Description
Technical Field
The invention belongs to the technical field of seismic data processing, and particularly relates to a multi-source seismic mixed data separation method based on convolutional neural network deep learning.
Background
The multi-source acquisition mode is a great innovation of the current seismic data acquisition. At present, in oil-gas seismic exploration, a multi-focus acquisition method is combined with a high-density wide-azimuth acquisition method, so that the acquisition efficiency is effectively improved while the quality of acquired data is ensured. However, unlike the seismic data acquired by the traditional single-source excitation, in the mixed record acquired by multiple sources, besides the main source record, the interference caused by secondary sources (other sources in the mixed cannon) also exists, so that the subsequent processing and interpretation are influenced. Since the current mature seismic imaging inversion method (AVO, FWI, RTM) is based on a single source data volume, how to obtain high-precision separation data becomes one of key technologies for multi-source acquisition.
The separation of multi-source mixed data is mainly based on the idea of 'domain transformation', namely, mixed data acquired by adopting a random time delay excitation mode, after the delay time of a main source is removed, the same phase axes of the main source and the secondary source are all represented as correlations in a common shot point gather, main source wave fields are still represented as correlations in other gather (such as a common detection point, a common center point, a common offset gather and the like) outside the common shot domain, and secondary source wave fields (aliasing noise) are represented as uncorrelated and are randomly distributed. The separation of the mixed data takes advantage of this distinction. At present, two separation methods are mainly adopted, namely a filtering denoising separation method and a sparse inversion separation method. These methods often require some parameters to be manually given. The seismic data of different work areas have different characteristics, so that corresponding parameters are also required to be given according to the characteristics of the seismic data of the actual work areas, and the fidelity of the separated seismic signals is influenced by the artificial parameters.
The filtering denoising separation method is mainly used for designing a proper filter to remove aliasing noise which is randomly distributed. At present, multi-source data separation based on filtering denoising mainly comprises median filtering and curvelet transform filtering (Han Liguo and the like, 2013), multi-level median filtering (Zhou Li and the like, 2016), multi-directional vector median filtering (Guo Jianhong and the like, 2019) and S-transform adaptive filtering (Huang Dezhi and the like, 2020); and the separation method based on the sparse inversion theory is used for solving the separation of the mixed data as an inverse problem. Since the inverse problem is a significant underdetermined problem, corresponding regularization constraints need to be added in the solving process, for example, based on Seislet sparse transform (Zu Shaohuan, etc., 2016; king, etc., 2018), low rank constraint (Zu Shaohuan, etc., 2018), L1 norm-analog constraint (Zhu Lihua, etc., 2018), and the like.
In the existing two multi-source separation methods, parameters are manually given in the implementation process, whether the filtering denoising method or the sparse inversion method is based. As in the filtering denoising-based method, different filtering methods need to give corresponding filtering parameters: median filtering requires a given filter length; vector median filtering requires a given number of directions of the filter in addition to the length of the filter; the curvelet transform filtering requires a threshold size for a given sparse domain. Also in the separation method based on inversion theory, implementation of regularization constraint conditions also requires given parameters: constraints are implemented in various sparse transform domains, and to give a rule of threshold transform, a maximum threshold and a minimum threshold, a threshold for preserving a maximum matrix rank needs to be given in a low rank constraint condition. All the parameters need to be given manually, multiple tests are generally needed according to different seismic data, and the most suitable parameters are selected. The choice of these parameters directly affects the removal of aliasing noise and the fidelity of the primary source seismic signal.
Since the nature of both methods is to remove random aliased noise from the hypocenter, any method of removing noise can be used to separate multiple sources. The deep learning has obvious advantages in the aspect of signal denoising, can intelligently perform denoising, can apply the deep learning technology to the separation of multiple seismic sources, and performs artificial intelligent denoising through a training network. The deep learning technology is mature in denoising of theoretical synthesized data at present, and the main reason is that training samples are easy to construct. However, in denoising of actual seismic data, deep learning training samples, particularly tag data, are difficult to acquire, the acquisition of the current training samples is often constructed through theoretical synthetic data of different main frequencies and different apparent velocities and phase axes, and the constructed training samples hardly contain all wave field features of the actual mixed seismic data, so that the effectiveness of the method is affected.
Disclosure of Invention
The invention overcomes the defects of the prior art and provides a multi-source seismic data separation method based on deep learning. The method can be used for intelligently denoising mixed data containing aliasing noise, so that the influence of artificial parameter selection on separation results is avoided, and most of effective signals can be reserved for the separated seismic data; when the deep learning training sample is constructed, the mixed data seismic trace random coding technology is adopted to directly construct the label of the training data set, so that the problem that label data in CNN is difficult to acquire is solved.
In order to achieve the above purpose, the invention is realized by the following technical scheme:
a multi-source seismic data separation method based on deep learning comprises the following steps:
s1: CNN deep learning
1.1 Co-shot (CSP, common shot point) Multi-Source Mixed dataPerforming seismic trace random time delay coding and performing +.>Adding to obtain common shot record with random noise as input P of neural network training input 。
1.2 mixing data of multiple sources at common shotAs training tag, i.e. output of neural network training +.>Forms corresponding pairs of network training samples (P input ，P out )。
1.3 dividing the whole input and output seismic data obtained in the steps 1.1 and 1.2 into small data sets as an input layer and an output layer of the CNN.
1.4 deep learning convolutional neural networks with designed convolutional topologies: the convolution topological structure is divided into three parts, wherein the first part is C1, and only one convolution layer exists, the convolution layer is subjected to convolution to obtain a plurality of feature mappings, and the size of the convolved output is consistent with that of the input data; the second part is C2 and has a plurality of convolution layers; the third part is C3, which only comprises one convolution layer, and 1 characteristic mapping is obtained through convolution operation, namely the output of the network; substituting the input layer and the output layer obtained in the step 1.3 into a designed convolutional neural network, and training to obtain a training model phi.
S2: multi-source hybrid seismic data separation
2.1 mixing data for multiple sources on common shot DomainPerforming time correction according to the delayed excitation time operators of all the seismic sources in the mixed super gun to obtain pseudo-separation data P 'on a common shot point domain' csp 。
2.2 pseudo-split data P 'on common shot domain' csp Reproduction of the programming into other domains outside the common shot point, such as P 'in the common receiver point (CRP, common receiver point) domain' crp At this time, the main source is a continuous signal, and the secondary sources are random signals.
2.3 trace set P 'at common detector' crp In the process, a training model phi is adopted for denoising, and a denoising result P is obtained crp E.g. formula (1)
P crp ＝F predict (P’ crp ,Φ) (1)
2.4 Co-detector Point Domain P after denoising crp The trace sets are rearranged to a common shot trace set P csp The separation is completed.
Further, in step 1.1, the neural network trains the input P input As formula (2):
wherein N is the total number of receiving channels in the mixed cannon,for the j-th channel of the mixed gun, Γ' j Representing a random time encoding operator of the jth lane, the random delay time not being equal to the excitation delay operator of the actual hybrid acquisition; wherein, the random timeThe expression of the coding operator in the frequency domain is: Γ ' = [ exp (-iωδt ') ' 1 )、exp(-iωδt’ 2 )...exp(-iωδt’ N )]Omega is the angular frequency, δt' j For the random encoding time of each track.
Further, in step 1.3, the whole seismic data is divided into a small data set with the size of 40×40 pixels, and the small data set is used as an input layer and an output layer of the CNN; to ensure continuity of the segmented image, there is a 30 pixel overlap between adjacent datasets.
Further, in step 1.4, the first part C1 convolution layer uses 64 convolution kernels with a size of 3×3, and the 64 feature maps are obtained through convolution; the second part is C2, and 15 convolution layers are total, and each convolution layer consists of 64 convolution kernels with the size of 3 multiplied by 3; the third part is C3, the convolution layer adopts 1 convolution kernel with the size of 3 multiplied by 3, and 1 characteristic mapping is obtained through convolution operation, namely the output of the network.
Further, for the first part C1, in order to ensure that the size of the convolved output is consistent with that of the input data, before performing the convolution operation, each data is subjected to boundary expansion, and zero is directly added to the same size as the input data; the feature data is then input to an activation function ReLU of the activation layer for removing redundancy in the seismic data, preserving the data features as much as possible.
Still further, for the second portion C2, a batch normalization layer batch normalization is added between the convolutional layer and the active layer to speed up and stabilize the training process.
Further, in step 2.1, pseudo-split data P 'on the common shot domain' csp As formula (3):
wherein,,the time delay inverse operator representing the source i has the expression: Γ -shaped structure -1 ＝[exp(iωδt 1 )、exp(iωδt 2 )...exp(-iωδt K )]K is the mixing degree of multi-focus acquisition, omega is the angular frequency, δt i Is the delayed firing time of the ith source.
Compared with the prior art, the invention has the following beneficial effects:
1. according to the invention, a convolutional neural network deep learning technology is introduced into random aliasing denoising of multi-source seismic data, more complex data features in the seismic data are extracted through the convolutional neural network, the damage to the seismic signals of a main seismic source is small, the fidelity of the seismic signals of the main seismic source is high, the denoising process is intelligently realized, the trained network can be used for intelligently denoising mixed data containing aliasing noise, the influence of artificial parameter selection on separation results is avoided, and the separated seismic data can retain most of effective signals.
2. The random noise in the multi-source mixed data separation is different from the conventional random noise, which itself is derived from the seismic signals of the hypocenter. In order to enable deep learning to be applied to separation of actual multi-seismic hybrid data, according to the characteristics of random aliasing noise in the multi-seismic hybrid data, all seismic channels of original hybrid data are subjected to random time coding, and input of training samples containing random noise distribution is constructed, and at the moment, the hybrid data naturally becomes corresponding network output. The problem that the label data in the CNN is difficult to acquire is solved, the training samples are from the data rather than the theoretical synthesized data, and all wave field features of the mixed data can be contained to the greatest extent, so that the applicability of CNN deep learning in actual multi-source mixed data separation can be improved.
Drawings
FIG. 1 is a flow chart of multi-source hybrid data separation as employed by the present invention.
Fig. 2 is a schematic diagram of a Convolutional Neural Network (CNN) structure and input-output layer small data segmentation designed in the present invention, and the convolutional neural network structure is divided into three parts. The first part is C1, only one convolution layer is provided, the convolution layer adopts 64 convolution kernels with the size of 3 multiplied by 3, 64 feature maps are obtained through convolution, and then feature data are input into an activation function ReLU of the activation layer to remove redundancy in the seismic data, and the data features are reserved as far as possible; the second part is C2, and 15 convolution layers are totally arranged, each convolution layer consists of 64 convolution kernels with the size of 3 multiplied by 3, and a batch normalization layer batch normalization is added between the convolution layers and the activation layer and used for accelerating and stabilizing the training process; the third part is C3, which only contains one convolution layer, the convolution layer adopts 1 convolution kernel with the size of 3×3, and 1 feature map is obtained through convolution operation, namely the output of the network.
The CNN input layer and the output layer are small data volume sets for dividing the original data into 40×40 pixels, as shown in the small frames in fig. 2, and there are 30 pixels overlapping between adjacent small frames.
In fig. 3, a is a geological seismic model for acquiring a synthetic seismic record in the embodiment, b is a multi-source random excitation time with a mixing degree of 2 in the embodiment, and c is a multi-source acquisition observation system in the embodiment.
FIG. 4 is a graph of a conventional single source excitation and multiple source excitation co-shot seismic record. Wherein a is the seismic record of the 13 th seismic source 1 which is excited independently, b is the seismic record of the 13 th seismic source 2 which is excited independently, and c is the seismic record of the 13 th multi-seismic source mixed excitation.
Fig. 5 is four of the pairs of training sample inputs and outputs constructed using co-shot hybrid recordings in an embodiment. Wherein (a-d) is input layer data, and (e-h) is corresponding output layer data.
Fig. 6 is a pseudo-split record of a common shot domain obtained by performing time correction on a certain common shot mixed record according to random mixing time in the embodiment. The method comprises the steps of (a) carrying out 13 th mixed seismic source (b) carrying out 13 th common shot point record of a first seismic source ship after pseudo separation and (c) carrying out 13 th common shot point record of a second seismic source ship after pseudo separation.
Fig. 7 is a schematic diagram of the sorting of second source vessel pseudo-separation data from the common shot domain (first 100 shot domain data) to the common shot domain (201 shot domain).
FIG. 8 is a plot of a seismic record, a co-geophone seismic record denoising result, and an error accuracy analysis of a co-geophone domain, for separating pseudo-split seismic data from co-geophone domain, in an embodiment. (a) 60 th co-range seismic data (b) CNN denoising result (c) removed noise (d) error (f) error e of conventional single source excited co-range data (e) denoising result b and conventional single source excited co-range data d and local similarity analysis of data d.
FIG. 9 is a graph showing the results of multi-source separation for common shot domains using the present invention in the examples. (a) blending data (b) the separated first source vessel data (c) the separated second source vessel data.
Detailed Description
In order to make the technical problems, technical schemes and beneficial effects to be solved more clear, the invention is further described in detail by combining the embodiments and the drawings. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. The following describes the technical scheme of the present invention in detail with reference to examples and drawings, but the scope of protection is not limited thereto.
By the description of the present invention and the further description of the embodiments below, the separation of multi-source blended data may be achieved using the present invention. In this embodiment, we first simulate to obtain the seismic record under the excitation of the traditional single seismic source by using forward modeling method, and forward modeling is shown as a in fig. 3.
Firstly, sequentially exciting by adopting a single seismic source, wherein the gun spacing is 20m, and the total excitation is 201 guns. All shots receive seismic records from 201 detectors at fixed locations on the earth's surface, with a detector-to-detector spacing of 20m. The time length of the seismic record obtained by forward modeling is 3s, and the time step is 0.002s. The 201 single source seismic records are considered as the seismic records of the first source vessel.
And then keeping the positions of 201 wave-detecting points on the ground surface unchanged, moving the positions of the wave-detecting points of the 201 wave-detecting points to the right by 5m, and repeating forward to obtain another 201 wave-detecting single-source seismic record which is taken as the seismic record of the second source ship.
And carrying out random time mixing on the 201 seismic sources corresponding to the two seismic source ships to obtain a plurality of seismic source mixed seismic records with the mixing degree of 2, and obtaining 201 mixed records in total. Wherein the random mixing time is shown as b in fig. 3, and the mixing mode is shown as c in fig. 3. The firing time of the first source vessel is set to 0ms all and the random delay time range of the second source vessel is set to (0,500 ms). Fig. 4 shows the 13 th single source record (fig. 4a and 4 b) of the first source vessel and the second source vessel before mixing, and the 13 th mixed shot record (fig. 4 c) after mixing, respectively.
According to 1.1 in the step S1 deep learning, each record in each common shot domain mixed record is subjected to random time delay, and then seismic data containing random noise is obtained by adding the common shot domain mixed records, so that 201 common shot domain records with random noise can be obtained in total and can be used as input samples for neural network training. According to 1.2 in the step S1, the original common shot domain mixed record corresponds to an output sample of the neural network training, so that a total of 201 input/output pairs of training samples can be obtained. Fig. 5 shows the network inputs and outputs for pairs 1, 50, 100 and 150.
According to 1.3 in step S1 of the present invention, the input-output pairs of 201 training samples are partitioned into data blocks of 40×40 pixels with 30 pixel repetition between neighboring blocks.
The number of traces of each seismic record is 201, and the time sampling point of each trace is 1501, so the number of blocks after segmentation is 201× (201-30)/10× (1501-30)/10= 502299 in total. According to 1.4 in the step S1, the neural network training is carried out by adopting a Keras deep learning framework, the initial learning rate is set to be 1 multiplied by 10 < -3 >, and the learning target is optimized by using an Adam algorithm. The batch size (batch size) was set to 128 times and the number of training epoch was set to 50. The training process uses a Graphic Processing Unit (GPU) to increase the training speed, the model of the GPU is GeForce RTX2080Ti, and the training time is 38448 seconds. Finally obtaining the training network model phi.
According to 2.1 in the step S2, pseudo-separation is carried out on 201 common shot domain mixed records. Each of the mixed recordings was corrected in time according to the random excitation time in the mixed source (fig. 3 b), resulting in 2 pseudo-split recordings (see fig. 6). The first corresponding to a single source record for a first source vessel (fig. 6 b) and the second corresponding to a single source record for a second source vessel (fig. 6 c). Because the random firing times of the first source vessel are all 0, the hybrid record is corrected in time according to the firing time of the first source vessel without any change from the hybrid record (fig. 6 b); after the time alignment of the hybrid recordings according to the excitation time of the second source vessel, the signals belonging to the second source vessel in the hybrid recordings have been corrected to have no time delay, and the signals belonging to the first source vessel have also been corrected, and the whole is moved upwards (fig. 6 c).
According to step S2.2 of the present invention, the pseudo-split recordings are separated from the common shot domain to the common detector domain, yielding a total of 201 common detector domain seismic recordings. The essence of sorting is to arrange the seismic data in different spatial locations (fig. 7). Fig. 7 is a schematic diagram of sorting the first 100 shot domain seismic data of the second source vessel after pseudo-separation, 201 seismic traces of each shot, 1501 points of each trace. It can be seen that in the common geophone domain, the primary source (second source vessel) signal is continuous and the aliased noise from the secondary source (first source vessel) is randomly distributed.
In accordance with step S2.3 of the present invention, the deep learning network model Φ obtained in step S1 is used to denoise each of the seismic recordings of the common-detector domains (fig. 8a is one of the common-detector domain recordings) (fig. 8b and 8 c). In order to analyze the accuracy of multi-source separation, we compare the common-detector-point-domain seismic data (only continuous signals, without random noise, see fig. 8 d) with the conventional single-source excitation (fig. 8 e), and further perform similarity analysis on the error (fig. 8 e) and the theoretical data (fig. 8 d), so that only a small part of the error signals have higher similarity with the theoretical data, and the denoising method provided by the invention is used for denoising, so that most of effective signals are reserved, and the damage to the continuous signals is small.
After denoising the 201 common-detector-domain seismic data, according to 2.4 in step S2 of the present invention, the denoised common-detector-domain seismic data is rearranged into the common shot domain, thereby completing multi-source separation (see fig. 9).
While the invention has been described in detail in connection with specific preferred embodiments thereof, it is not to be construed as limited thereto, but rather as a result of a simple deduction or substitution by a person having ordinary skill in the art to which the invention pertains without departing from the scope of the invention defined by the appended claims.
Claims (7)
1. The multi-source seismic data separation method based on deep learning is characterized by comprising the following steps of:
s1: CNN deep learning
1.1 mixing data of multiple seismic sources with common shot pointsPerforming seismic trace random time delay coding and performing +.>Adding to obtain common shot record with random noise as input P of neural network training input ；
1.2 mixing data of multiple sources at common shotAs training tag, i.e. output of neural network training +.>Forms corresponding pairs of network training samples (P input ，P out )；
1.3 dividing the whole input and output seismic data obtained in the steps 1.1 and 1.2 into small data sets as an input layer and an output layer of CNN;
1.4 deep learning convolutional neural networks with designed convolutional topologies: the convolution topological structure is divided into three parts, wherein the first part is C1, and only one convolution layer exists, the convolution layer is subjected to convolution to obtain a plurality of feature mappings, and the size of the convolved output is consistent with that of the input data; the second part is C2 and has a plurality of convolution layers; the third part is C3, which only comprises one convolution layer, and 1 characteristic mapping is obtained through convolution operation, namely the output of the network; substituting the input layer and the output layer obtained in the step 1.3 into a designed convolutional neural network, and training to obtain a training model phi;
s2: multi-source hybrid seismic data separation
2.1 mixing data for multiple sources on common shot DomainPerforming time correction according to the delayed excitation time operators of all the seismic sources in the mixed super gun to obtain pseudo-separation data P 'on a common shot point domain' csp ；
2.2 pseudo-split data P 'on common shot domain' csp Reproducing and arranging the main seismic source to other domains outside the common shot point, wherein the main seismic source is a continuous signal, and the secondary seismic source is a random signal;
2.3 trace set P 'at common detector' crp In the process, a training model phi is adopted for denoising, and a denoising result P is obtained crp E.g. formula (1)
P crp ＝F predict (P' crp ,Φ) (1)
2.4 Co-detector Point Domain P after denoising crp The trace sets are rearranged to a common shot trace set P csp The separation is completed.
2. The method of claim 1, wherein in step 1.1, the neural network trains the input P input As formula (2):
wherein N is common shot point multi-focus mixed dataTotal number of reception channels>For co-shot multi-source mixing dataAnd Γ 'of (1)' j Representing a random time delay encoding operator of the jth lane, the random delay time not being equal to the excitation delay operator of the actual hybrid acquisition; the expression of the random time delay coding operator in the frequency domain is as follows: Γ ' = [ exp (-iωδt ') ' 1 )、exp(-iωδt' 2 )...exp(-iωδt' N )]Omega is the angular frequency, δt' j For the random encoding time of each track.
3. The method for separating seismic data of multiple seismic sources based on deep learning according to claim 1, wherein in step 1.3, the whole seismic data is divided into small data sets with the size of 40×40 pixels, and the small data sets are used as an input layer and an output layer of CNN; to ensure continuity of the segmented image, there is a 30 pixel overlap between adjacent datasets.
4. The method for separating seismic data of multiple seismic sources based on deep learning as claimed in claim 1, wherein in step 1.4, the first part of the C1 convolution layer adopts 64 convolution kernels with the size of 3×3, and the 64 feature maps are obtained through convolution; the second part is C2, and 15 convolution layers are total, and each convolution layer consists of 64 convolution kernels with the size of 3 multiplied by 3; the third part is C3 and the convolutional layer uses 1 convolutional kernel of size 3 x 3.
5. The method of claim 4, wherein for the first portion C1, before performing the convolution operation, each data expansion boundary is directly zero-padded to the same size as the input data; the feature data is then input to an activation function ReLU of the activation layer for removing redundancy in the seismic data.
6. The method of claim 4, wherein for the second portion C2, a batch normalization layer is added between the convolution layer and the activation layer to speed up and stabilize the training process.
7. The method of claim 4, wherein in step 2.1, pseudo-split data P 'in common shot domain' csp As formula (3):
wherein,,the time delay inverse operator representing the source i has the expression: Γ -shaped structure -1 ＝[exp(iωδt 1 )、exp(iωδt 2 )...exp(-iωδt K )]K is the mixing degree of multi-focus acquisition, omega is the angular frequency, δt i Is the delayed firing time of the ith source.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111162037.0A CN113885077B (en) | 2021-09-30 | 2021-09-30 | Multi-source seismic data separation method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111162037.0A CN113885077B (en) | 2021-09-30 | 2021-09-30 | Multi-source seismic data separation method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113885077A CN113885077A (en) | 2022-01-04 |
| CN113885077B true CN113885077B (en) | 2023-07-18 |
Family
ID=79004902
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111162037.0A Active CN113885077B (en) | 2021-09-30 | 2021-09-30 | Multi-source seismic data separation method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113885077B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114839673B (en) * | 2022-07-01 | 2022-09-23 | 中国海洋大学 | Separation method, separation system and computer equipment for multi-seismic-source efficient acquisition wave field |
| CN115267899B (en) * | 2022-08-15 | 2024-01-12 | 河北地质大学 | DnCNN mixed source seismic data separation method and system based on boundary preservation |
| CN115577247B (en) * | 2022-12-09 | 2023-07-11 | 中海油田服务股份有限公司 | Seismic noise removing method and device based on stacked feedback residual error network |
| CN115620113B (en) * | 2022-12-20 | 2023-04-07 | 成都理工大学 | Elastic wave vector separation method for generating countermeasure network based on deep convolution |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019099974A1 (en) * | 2017-11-19 | 2019-05-23 | Westerngeco Llc | Noise attenuation of multiple source seismic data |
| CN110568485A (en) * | 2019-09-06 | 2019-12-13 | 广州海洋地质调查局 | neural network-based multi-channel seismic continuous recording and separating method |
| CN110967750A (en) * | 2019-12-16 | 2020-04-07 | 中国海洋石油集团有限公司 | Multi-source seismic mixed wave field separation method and device |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120215453A1 (en) * | 2011-02-22 | 2012-08-23 | Cggveritas Services Sa | Device and method for multi-dimensional coherency driven denoising data |
| US11169293B2 (en) * | 2016-12-13 | 2021-11-09 | Cgg Services Sas | Device and method for model-based deblending |
| US11175424B2 (en) * | 2019-02-25 | 2021-11-16 | Saudi Arabian Oil Company | Seismic data de-blending |
-
2021
- 2021-09-30 CN CN202111162037.0A patent/CN113885077B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019099974A1 (en) * | 2017-11-19 | 2019-05-23 | Westerngeco Llc | Noise attenuation of multiple source seismic data |
| CN110568485A (en) * | 2019-09-06 | 2019-12-13 | 广州海洋地质调查局 | neural network-based multi-channel seismic continuous recording and separating method |
| CN110967750A (en) * | 2019-12-16 | 2020-04-07 | 中国海洋石油集团有限公司 | Multi-source seismic mixed wave field separation method and device |
Non-Patent Citations (3)
| Title |
|---|
| Amplitude-preserving iterative deblending of simultaneous source seismic data using high-order Radon transform;Yaru Xue 等;《Journal of Applied Geophysics》;第79-90页 * |
| 基于改进迭代去噪法的多震源地震数据分离;郭建宏 等;《物探与化探》;第43卷(第5期);第1054-1063页 * |
| 基于迭代去噪的多源地震混合采集数据分离;韩立国 等;地球物理学报;第56卷(第7期);第2402-2412页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113885077A (en) | 2022-01-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113885077B (en) | Multi-source seismic data separation method based on deep learning | |
| Gao et al. | Irregular seismic data reconstruction based on exponential threshold model of POCS method | |
| US11740375B2 (en) | Methods for simultaneous source separation | |
| WO2008112036A1 (en) | Imaging of multishot seismic data | |
| Ovcharenko et al. | Multi-task learning for low-frequency extrapolation and elastic model building from seismic data | |
| CN114839673B (en) | Separation method, separation system and computer equipment for multi-seismic-source efficient acquisition wave field | |
| Huang et al. | Self-supervised deep learning to reconstruct seismic data with consecutively missing traces | |
| US20230117321A1 (en) | Separation of Blended Seismic Survey Data | |
| Zhang et al. | 3D simultaneous seismic data reconstruction and noise suppression based on the curvelet transform | |
| Trad | Five-dimensional interpolation: New directions and challenges | |
| CN110716231B (en) | Offshore multi-seismic source wave field separation method and system based on confocal domain sparse inversion | |
| EA038811B1 (en) | Method and system for generating geophysical data | |
| WO2021055152A1 (en) | Noise attenuation methods applied during simultaneous source deblending and separation | |
| Liu et al. | A dictionary learning method with atom splitting for seismic footprint suppression | |
| CN116719086B (en) | Sparse seabed four-component data high-resolution imaging method based on point spread function | |
| Yan et al. | A method for denoising seismic signals with a CNN based on an attention mechanism | |
| CN114296134B (en) | Deep convolutional network seismic data unmixing method and system | |
| CN106950597A (en) | The mixing source data separation method filtered based on three sides | |
| Wang et al. | Multicomponent seismic noise attenuation with multivariate order statistic filters | |
| CN107255833B (en) | The determination method and apparatus of Seismic Stacked Section | |
| Wang et al. | A physics-augmented deep learning method for seismic data deblending | |
| Huang et al. | A fast least-squares reverse time migration method using cycle-consistent generative adversarial network | |
| CN112327356A (en) | Aliasing record separation method based on inphase axis iterative tracking extraction | |
| Nakayama et al. | Machine-learning based data recovery and its benefit to seismic acquisition: Deblending, data reconstruction, and low-frequency extrapolation in a simultaneous fashion | |
| CN110895346A (en) | Method for separating seismic diffracted waves by common offset distance domain SVD filtering |
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 |