CN115630295A - Method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning - Google Patents
Method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning Download PDFInfo
- Publication number
- CN115630295A CN115630295A CN202211179180.5A CN202211179180A CN115630295A CN 115630295 A CN115630295 A CN 115630295A CN 202211179180 A CN202211179180 A CN 202211179180A CN 115630295 A CN115630295 A CN 115630295A
- Authority
- CN
- China
- Prior art keywords
- neural network
- data
- deep neural
- unsupervised
- residual
- 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
Images
Landscapes
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention discloses a method for separating multi-seismic source data by unsupervised learning with double-depth neural network constraint, which comprises the steps of constructing an unsupervised deep neural network structure with double-depth neural network constraint to carry out iterative inversion, and separating the collected multi-seismic source data by using trained neural network parameters through an unsupervised learning method based on the double-depth neural network constraint; belonging to the technical field of seismic coherent noise data processing. By adopting the technical scheme provided by the invention, the unsupervised deep neural network with double-deep neural network constraint is used, an expected separation result is not required to be used as the label data, the problem of label data loss can be solved, and the multi-seismic-source data can be further better separated, so that the field acquisition time is well shortened, the acquisition cost is saved, and the subsequent seismic imaging effect is improved.
Description
Technical Field
The invention belongs to the technical field of seismic coherent noise data processing, and particularly relates to a method for separating multi-seismic-source acquired data by using an unsupervised learning method based on double-depth neural network constraint.
Background
The simultaneous seismic source acquisition technique is also called a multi-seismic source acquisition technique, and can acquire seismic records of a plurality of cannons within the time of conventionally acquiring a single-cannon seismic record [1]. In recent years, with the development of simultaneous seismic source acquisition technology and wide application in the field of seismic exploration, wide azimuth and high density seismic acquisition becomes possible. Meanwhile, the application of the seismic source acquisition technology not only greatly improves the production efficiency and can obtain seismic data with high coverage times, but also greatly improves the quality of the seismic data [2]. Meanwhile, the seismic source acquisition technology adopts a plurality of groups of controllable seismic sources to independently excite. Only a short waiting time is required in each group, and thus there is no particular limitation on the number of vibroseis. The more the number of the seismic sources is, the higher the acquisition construction efficiency is. However, a plurality of seismic sources are excited in different excitation points within a short time, and strong adjacent shot interference, namely crosstalk noise, is necessarily generated on a single shot record. These crosstalk noises significantly reduce the signal-to-noise ratio of the seismic data, affecting the quality of subsequent seismic data processing [3]. Therefore, the seismic source data cannot be directly used in the traditional seismic data processing flow, and the aliasing wave field needs to be separated first to obtain the conventional single-shot data.
There are several features of the seismic source data that are acquired simultaneously. On a common shot gather, the energy has coherence; after pseudo separation, only the energy from the first shot has coherence, and the energy from the interference of the adjacent shots shows strong random amplitude or sharp pulse interference on the common offset gather, the common receiving point gather or the common central point gather. Separation methods can be generally divided into two broad categories, passive separation and active separation.
The passive separation is to convert the separation problem into the denoising problem, and regard the interference shot signals coded by delay time as noise. The filtering separation method is to separate the aliasing noises according to the distribution characteristic difference of the seismic signals and the aliasing noises in different domains. Document [4] uses a vector median filtering method to achieve aliasing noise separation in the non-shot domain. Document [5] according to the coherence difference of the reflected signal and the aliasing noise in different time domains, separating the aliasing noise based on an iterative method of interference noise estimation and adaptive denoising; document [6] proposes to solve the dip angle information of the formation by using plane wave decomposition, and then suppress the crosstalk noise by combining vector median filtering. Document [7] implements aliasing noise separation by estimating the slope of the in-phase axis in the local time window and performing median filtering by flattening the in-phase axis.
Another separation method is active separation, also called inversion method, which refines the separation problem into an inverse problem in mathematics and solves it by using the inversion method. Separating the mixed data using sparse Radon transforms is a good choice [8]. Document [9] proposes a separation method based on curvelet domain L1 constraint; document [10] compares curvelet transformation with a mixed mining separation algorithm of different regularization constraints; document [11] proposes the use of adaptive Wiener threshold in the Contourlet domain to boost the separation effect of simultaneous source data.
In recent years, deep neural networks have also gained much attention in the field of exploration of geophysical. When the deep neural network training is finished, the target can be directly predicted, and the prior experience is not excessively relied to participate in parameter adjustment. In view of the strong fitting and non-linear approximation ability of the deep learning method, a large number of deep neural network methods are applied to seismic data processing and interpretation, such as random noise suppression, velocity modeling, multiple suppression, wave impedance inversion, salt dome recognition, and the like.
Deep neural networks also have many applications in simultaneous source data separation. Document [12] trains a convolutional neural network by using aliased data and original unaliased data as input data and label data, respectively, and finally, directly inputting seismic source data which is not trained while directly inputting the seismic source data into a deep neural network to complete separation. Document [13] uses aliased data and raw unaliased data as training sets and preprocesses the training sets using amplitude normalization and size cutting. The trained deep neural network can perform iterative inversion separation on seismic source data at the same time, and a certain effect is achieved. The document [14] proposes that aliasing data and original unaliased data are used as a training set to train a U-net deep neural network [15], and the trained deep neural network obtains certain cross-work area crosstalk noise suppression capability after transfer learning. In the above 3 simultaneous seismic source data separation methods based on deep neural network, the training set uses the original unaliased data as the label data. In actual simultaneous source data acquisition, the original unaliased data is not available. Therefore, it is difficult to obtain a good simultaneous seismic source data separation effect in practical use by the above 3 simultaneous seismic source data separation methods based on the deep neural network.
The traditional passive separation method mainly separates the reflected signals and the aliasing noises in a common offset gather or a common detection point gather. The median filtering separation method is used for separating the seismic signals and the aliasing noise according to the coherence of the reflection signals among the seismic channels in the gather and the randomness of the distribution of the aliasing noise; the median filtering separation method has the problems of damaging the signal wave group characteristics, unclean separation and the like; the spectrum filtering separation method has the problem that the distribution range of the aliasing noise spectrum and the signal spectrum in a transform domain is overlapped, so that the aliasing noise cannot be completely removed or the reflected signal is cut off, and the phenomena of mixing, false frequency and the like are caused due to the limitation of the method. Therefore, the denoising method based on the time domain and the transform domain is simple to implement and high in calculation efficiency, but the suppression effect is poor. Whereas conventional active separation of multiple bases on a fixed transform domain uses fixed basis functions to identify simultaneous source data. Since the basis functions are fixed, the ability to identify and represent effectively coherent signals is weak. In addition, most of the existing deep neural networks in the field of exploration geophysical belong to supervised learning, the problem of label data loss cannot be solved, and practical application is difficult to achieve. In the prior art, the multi-seismic source data are difficult to be effectively separated, the multi-seismic source data are acquired in the field for a long time, the acquisition cost is high, and the subsequent seismic imaging effect is not facilitated.
Reference:
[1]A.G.Berkhout,"Changing the mindset in seismic data acquisition,"The Leading Edge,vol.27,no.7,pp.924-938,2008.
[2]K.Wapenaar,J.van der Neut,and J.Thorbecke,"Deblending by direct inversion,"Geophysics,vol.77,no.3,pp.A9-A12,2012.
[3]A.Mahdad,P.Doulgeris,and G.Blacquiere,"Separation of blended data by iterative estimation and subtraction of blending interference noise,"Geophysics,vol.76,no.3,pp.Q9-Q17,2011.
[4]S.Huo,Y.Luo,and P.G.Kelamis,"Simultaneous sources separation via multidirectional vector-median filtering,"GEOPHYSICS,vol.77,no.4,pp.V123-V131,2012/07/01 2012,doi:10.1190/geo2011-0254.1.
[5]A.Mahdad and G.Blacquière,"Iterative method for the seperation of blended encoded shot records,"in 72nd EAGE Conference and Exhibition incorporating SPE EUROPEC 2010,2010:European Association of Geoscientists&Engineers,pp.cp-161-00018.
[6]Q.Yang,W.Mao,H.Tang,X.Zhu,Z.Qian,and Y.Zhan,"Deblending with weak signal preserved by dip vector-median filter,"in SEG Technical Program Expanded Abstracts 2017:Society of Exploration Geophysicists,2017,pp.5044-5048.
[7]S.Gan,S.Wang,Y.Chen,and X.Chen,"Deblending using a Structural-Oriented Median Filter,"presented at the 2015SEG Annual Meeting,2015.
[8]P.Akerberg,G.Hampson,J.Rickett,H.Martin,and J.Cole,"Simultaneous source separation by sparse Radon transform,"in SEG Technical Program Expanded Abstracts2008,(SEG Technical Program Expanded Abstracts:Society of Exploration Geophysicists,2008,pp.2801-2805.
[9]T.T.Lin and F.Herrmann,"Designing simultaneous acquisitions with compressive sensing,"in 71st EAGE Conference and Exhibition incorporating SPE EUROPEC 2009,2009:European Association of Geoscientists&Engineers,pp.cp-127-00269.
[10]S.Qu et al.,"Deblending of Simultaneous-source Seismic Data using Fast Iterative Shrinkage-thresholding Algorithm with Firm-thresholding,"Acta Geophysica,vol.64,no.4,pp.1064-1092,2016/08/01 2016,doi:10.1515/acgeo-2016-0043.
[11] wangkun Happy and Viagra construction, "simultaneous seismic source wavefield separation based on Contourlet domain adaptive Wiener threshold," geophysical newspaper, vol.64, no.1, pp.263-278,2021.
[12]J.Sun,S.Slang,T.Elboth,T.Larsen Greiner,S.McDonald,and L.-J.Gelius,"A convolutional neural network approach to deblending seismic data,"GEOPHYSICS,vol.85,no.4,pp.WA13-WA26,2020/07/01 2020,doi:10.1190/geo2019-0173.1.
[13]S.Zu,J.Cao,S.Qu,and Y.Chen,"Iterative deblending for simultaneous source data using the deep neural network,"GEOPHYSICS,vol.85,no.2,pp.V131-V141,2020,doi:10.1190/geo2019-0319.1.
[14]B.Wang,J.Li,J.Luo,Y.Wang,and J.Geng,"Intelligent Deblending of Seismic Data Based on U-Net and Transfer Learning,"IEEE Transactions on Geoscience and Remote Sensing,vol.59,no.10,pp.8885-8894,2021,doi:10.1109/TGRS.2020.3048746.
[15]O.Ronneberger,P.Fischer,and T.Brox,"U-net:Convolutional networks for biomedical image segmentation,"in International Conference on Medical image computing and computer-assisted intervention,2015:Springer,pp.234-241.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a method for separating multi-seismic-source collected data by using an unsupervised learning method based on double-depth neural network constraint, an unsupervised deep neural network structure with double-depth neural network constraint is constructed to carry out iterative inversion, and trained neural network parameters are used for separating the multi-seismic-source collected data. The unsupervised deep neural network with double deep neural network constraints is used, an expected separation result is not needed to be used as the label data, the problem of label data loss can be solved, and the method has a good application prospect.
In the specific implementation of the invention, the first input data and the second input data are mapped into the separation result of the second shot (second seismic source data) and the separation result of the first shot (first seismic source data) through the first deep neural network and the second deep neural network respectively. The designed total loss function can enable the unsupervised deep neural network with the double-deep neural network constraint to be optimized towards the correct direction, and the problem of overfitting is avoided. According to the method, the first shot aliasing data and the second shot aliasing data are respectively used as the first input data and the second input data, so that the dependence of an unsupervised deep neural network on the label data can be reduced, and the application range and the effect of the deep neural network for separating multi-seismic source data are further improved. The method can further help to better separate multi-seismic-source data, thereby well shortening field acquisition time, saving acquisition cost and improving subsequent seismic imaging effect.
Meanwhile, the seismic source acquisition technology can greatly improve the seismic acquisition efficiency, but due to continuous excitation, serious adjacent shot interference noise exists in seismic source data, and the seismic source data cannot be directly used in a conventional data processing flow. Therefore, separation of aliased wavefields into conventionally acquired single shot records is required. Under a high-density acquisition observation system, the invention provides double-depth neural network constrained unsupervised learning separation multi-seismic source data. In the invention, the unsupervised deep neural network with double deep neural network constraints mainly comprises a residual deep neural network and a U-net deep neural network, and has excellent nonlinear optimization capability. The total loss function can enable the unsupervised deep neural network with the double-deep neural network constraint to be optimized towards the correct direction, and the problem of overfitting is avoided. By minimizing the loss function provided by the invention, the residual deep neural network branch and the U-net deep neural network branch of the supervision deep neural network can extract coherent effective signals of all cannons and suppress crosstalk noise. An unsupervised deep neural network method input data with a dual deep neural network constraint is composed of first aliased data and second aliased data. In actual use, the original unaliased data is not available. The unsupervised learning method based on the double-depth neural network constraint does not need original unaliased data as label data and training set data, so that the problem of training set loss is solved, and the unsupervised learning method based on the double-depth neural network constraint has good practical use value.
The technical scheme provided by the invention is as follows:
a method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning is designed. In the training phase, 2 different deep neural networks are combined into an unsupervised deep neural network with a dual deep neural network constraint. The unsupervised deep neural network can map the first and second aliased data into a separated result after suppressing crosstalk noise on the premise of minimizing a total loss function. Comprising the following steps (fig. 1): step 1, preprocessing multi-seismic source aliasing data acquired based on a high-density acquisition observation system to obtain preprocessed aliasing data, wherein the multi-seismic sources are two or more than two seismic sources; step 2, performing amplitude normalization and size segmentation on the preprocessed aliasing data to obtain an aliasing data block (an aliasing data block obtained by segmenting the first seismic source data and the second seismic source data) for model training; step 3, constructing a residual deep neural network and a U-net deep neural network; step 4, constructing an unsupervised deep neural network; step 5, training an unsupervised deep neural network; and 6, obtaining a separation result after crosstalk noise suppression.
The object of the invention can also be achieved by the following technical measures:
in step 1, the pretreatment comprises: and performing elevation static correction, converting to a common shot point domain gather, performing random noise suppression, surface wave suppression, surge noise suppression and guided wave suppression, and converting to a common detection point domain to obtain preprocessing aliasing data containing crosstalk noise. When the designed observation system is a high-density acquisition observation system, the invention takes the multiple seismic source data obtained by respectively exciting the main seismic source ship and the auxiliary seismic source ship within a small distance (such as 100 meters) and sequentially moving the main seismic source ship and the auxiliary seismic source ship in the same direction for the same distance as a precondition. And then, carrying out elevation static correction on the acquired actual multi-source data. And secondly, transforming the data subjected to elevation static correction to a common shot point domain gather to perform random noise suppression, surface wave suppression, surge noise suppression and guided wave suppression. Then, the pre-processed aliasing data containing crosstalk noise is obtained after conversion to the co-detection point domain. The pre-processing aliased data D may be represented as:
wherein s is i For the delay time operator at the excitation of the ith seismic source, d i The resulting effective coherent signal is excited for the ith source. n is the number of aliases, i.e. the number of seismic sources excited in a short time interval. Finally, the pre-processed aliased data is passed through a delay time operator s i By the inverse operator of (A) to obtain aliased data
Where T represents the matrix transpose.
In step 2, the aliased data is processed
The amplitude of (a) is subjected to amplitude normalization and data segmentation. Firstly, normalizing the amplitude of aliasing data to be between-1 and 1 through an operator R to obtain normalized data:
wherein
As delay time operator s i The inverse operator of (2). T represents a matrix transpose. D max The value at which the absolute value of the element in the preprocessed aliased data is the largest. Then the amplitude normalized data passes through a cutting operator K β Is cut intoData blocks of size β x β and obtaining aliased data blocks for final use in training unsupervised deep neural networks
In step 3, the invention comprises 2 deep neural networks of the residual deep neural network and the U-net deep neural network.
The first deep neural network is a residual neural network f R . The residual neural network is a multi-layer neural network that solves the degradation problem and the gradient vanishing problem by convolutional layers and cross-layer connections for feature extraction. The convolution layer of the residual error neural network is composed of a plurality of characteristic graphs obtained by performing convolution operation on a previous layer of input image through convolution kernel, the convolution kernel is an important component of the convolution layer, each characteristic graph is composed of a plurality of neural nodes, all neurons on the characteristic graphs share the parameters of the same convolution kernel, the convolution kernel performs convolution operation on the previous layer of input image, feature extraction is realized through the characteristics of local connection and weight sharing in the convolution process, and the parameter quantity is greatly reduced while different features are extracted through the two characteristics of the convolution kernel. The convolution neural network gradually disappears with the deepening of the network, the shallow parameters cannot be updated, and the short circuit connection structure (short Connections) of the residual neural network ensures the updating of the back propagation parameters and avoids the problem of gradient disappearance caused by back propagation. Except the 1 st convolution layer and the last convolution layer, each 3 convolution layers form a residual block, and the 3 convolution layers are connected in a cross-layer mode through residual connection to realize identity mapping. The identity mapping ensures that the network performance is not reduced, so that the network learns new characteristics on the basis of input characteristics, and the identity mapping does not add extra parameters and calculated amount to the network but can accelerate the modelThe training speed is improved, and the training effect is optimized.
The 2 nd deep neural network is a U-net deep neural network f U . U-net is a typical encoding and decoding U-shaped structure, and a symmetrical structure is adopted to connect all layers. The U-net network is characterized in that convolution operation is carried out on a down-sampling part for multiple times, so that the whole network learns more key seismic data characteristics. U-net focuses on the fusion of contextual features. The fusion connection can enable the U-net to fully utilize the extracted features, so that the features of all scales of the original input are reflected in the final feature diagram. The U-net network architecture comprises an input layer, an output layer and a plurality of groups of intermediate layers. The whole network adopts a symmetrical coding and decoding structure, the coding of the left half part extracts data characteristics, and the decoding of the right half part restores data. In order to ensure that the sizes of input and output seismic data are the same, the invention uses the boundary filling method of equal-size filling to ensure that the sizes of input and output are the same. The coding section downsampling function is to extract deep features of the seismic data. The number of convolution kernels of the U-net deep neural network is marked on the convolution layer, and the size of each convolution kernel is 3 x 3. Except for the last layer, each convolution is followed by activation using the ReLU function. The size of the pooling window is 2 x 2 with a step size of 2. The upsampling multiple of the upsampled convolutional layer is 2 × 2, and the convolutional kernel size is 2 × 2. The seismic data are subjected to 3 times of downsampling and 3 times of upsampling, low-layer information and deep-layer information are fused by fusion connection every time, information loss caused by a pooling process in the downsampling process is made up, and detail information of the characteristic image is increased.
In step 4, a residual deep neural network f is used R And U-net deep neural network f U And jointly combining to construct an unsupervised deep neural network f. Unsupervised deep neural network f has 2 inputs x 1 ；x 2 ]And 2 outputs [ y ] 1 ；y 2 ]1 st input data x 1 Acting as 2 nd input data x from an aliased block of data from the first source 2 Acting as an alias data block from the second source, the 1 st input data x 1 And 2 nd input data x 2 Feeding in the residueA difference deep neural network branch and a U-net deep neural network branch.
Residual deep neural network f R Is the 1 st input data x 1 Output as the 1 st output data y 1 1 st output data y 1 Is the result of the separation of the second source. Residual deep neural network f R The purpose of this is to extract the same continuous valid signal (i.e. the separation result of the second source) in the 1 st input data and the 2 nd input data and output this continuous valid signal as the 1 st output data. Hypothesis residual deep neural network f R The learned deep neural network parameter is theta R Input residual deep neural network f R Is x 1 The resulting output is y 1 The whole process is expressed as follows:
y 1 ＝f R (x 1 ,Θ R ) (5)
taking the first iteration as an example, the residual deep neural network f R 1 st input data x 1 Aliased data for the first source
. In the ideal case of complete separation, the 1 st output data is the separation result of the second source, and in this case (4) can be rewritten as:
d 2 ＝f R (s 1 -1 D,Θ R ) (6)
residual deep neural network f R The 1 st output data of (a) is the separation result of the second source.
U-net deep neural network f U The input data of (1) is the 2 nd input data, and the output is the 2 nd output data. U-net deep neural network f U The purpose of (1) is to extract the same continuous effective signal in the 2 nd input data and the 1 st input data and output the continuous effective signal as the 2 nd output data. Hypothesis U-net deep neural network f U The learned deep neural network parameter is theta U When inputting U-net deep neural network f U For the second input data x 2 The obtained output isTwo output data y 2 And the second output data is the result of the first seismic source separation, and the whole process is expressed as follows:
y 2 ＝f U (x 2 ,Θ U ) (7)
using the first iteration as an example, U-net deep neural network f U The input data is second-shot aliasing data
In the ideal case of complete separation, the 2 nd output data is the effective coherent signal in the first aliasing data, and in this case (6) can be rewritten as:
u-net deep neural network f U The output data of (a) is an estimate of the original unaliased data of the first shot.
Residual deep neural network f R Branch and U-net deep neural network f U The branch can extract the valid coherent signals in the second shot and the first shot, respectively. When high density acquisition is performed, the original unaliased data of the first shot and the original unaliased data of the second shot are similar, and the difference Δ e between the two is close to 0, namely:
Δe＝d 1 -d 2 ≈0. (9)
unsupervised deep neural network f will residual deep neural f R Branch and U-net deep neural network f U The branches are combined into one entity and it is desirable to minimize Δ e. Unsupervised deep neural network f inherits residual deep neural f R Branch and U-net deep neural network f U All of (a). Suppose the data input into the unsupervised deep neural network is x and the output is y, that is
x＝[x 1 ；x 2 ], (10)
y＝[y 1 ；y 2 ]. (11)
When learning to obtain the unsupervised deep neural network parameter theta, the following parameters are obtained:
y＝f(x,Θ). (12)
wherein Θ = [ Θ R ；Θ U ]。
In step 5, the unsupervised deep neural network f can provide good nonlinear approximation capability. However, how to extract the same coherent signal in the first and second shot original unaliased data is the key to train the unsupervised deep neural network using a suitable loss function L.
The design of the loss function L is discussed with the first iteration separation as an example. The unsupervised deep neural network gets 5 fundamental losses using 2 input data and 2 output data. In residual depth neural network f R In the branch, the obtained output data y 1 And y 2 Respectively, an estimate of the second-shot unaliased data and crosstalk noise in the first-shot aliased data. At this time, the residual deep neural network f R Loss function L in branch 1 And L 2 Respectively, as follows:
L 1 (Θ R )＝||x 2 -f R (x 1 ,Θ R )|| 1 +μ 1 ||Θ R || 1 (13)
L 2 (Θ R )＝||x 1 -f R (x 1 ,Θ R )|| 1 +μ 2 ||Θ R || 1 (14)
wherein mu 1 And mu 2 Representing a regularization factor. By minimizing L 1 (Θ R ) Residual deep neural network f R Can extract x 2 Of the effective coherent signal. But if only L is present 1 (Θ R ) Is liable to cause overfitting, resulting in a residual deep neural network f R Is completely close to x 2 . Additional loss function L 2 (Θ R ) Deep neural network f capable of preventing residual errors R Overfitting to extract the correct coherent signal.
Deep neural network f in U-net U In the branch, the obtained output data y 2 Is an estimate of the unaliased data of the first shot. At this time, the U-net depthNeural network f U Loss function L in branch 3 And L 4 Respectively, as follows:
L 3 (Θ U )＝||x 1 -f U (x 2 ,Θ U )|| 1 +μ 3 ||Θ U || 1 (15)
L 4 (Θ U )＝||x 2 -f U (x 2 ,Θ U )|| 1 +μ 4 ||Θ U || 1 (16)
wherein mu 3 And mu 4 Representing a regularization factor. By minimizing L 3 (Θ U ) U-net deep neural network f U Can extract x 1 Of the effective coherent signal. But if only L is present 3 (Θ U ) Easily causes overfitting, resulting in a U-net deep neural network f U Is completely close to x 1 . Additional loss function L 4 (Θ U ) Deep neural network f capable of preventing U-net U Overfitting to extract the correct coherent signal. The invention uses the residual error deep neural network f R And U-net deep neural network f U Combined into an unsupervised neural network and expecting Δ e to approach 0 and proposing a loss function L 5 :
L 5 (Θ)＝||f R (x 1 ,Θ R )-f U (x 2 ,Θ U )|| 1 +μ 5 ||Θ|| 1 (17)
Wherein mu 5 Representing a regularization factor. By minimizing a loss function L 5 (Θ), it is possible to further prevent the unsupervised deep neural network from overfitting and to obtain an estimate of the correct first shot and second original unaliased data. Thus, the final loss function L (Θ) for training the unsupervised deep neural network f is as follows:
where μ denotes a regularization factor, f 1 And f 2 Respectively representing the 1 st and 2 nd outputs of the unsupervised deep neural network f. By minimizing the loss function L (Θ), the unsupervised deep neural network is able to accurately extract coherent signals in the first and second shot aliased data, respectively.
In step 6, the deep neural network parameter Θ is obtained by minimizing the total loss function of equation (18). The invention will alias data blocks
And sending the data to a trained unsupervised deep neural network with the parameter theta, and obtaining separation data G after crosstalk noise suppression through an iterative inversion separation method. The iterative separation expression for G is as follows:
where P represents combining every 2 adjacent aliased data out of the total aliased data. B is represented as an inverse mapping operator that maps from the data space to the model space. And m is the iteration number. R -1 Is the inverse of R, representing the restoration of the amplitude to the original amplitude.
Is K β Represents the splicing of the divided sizes into the original sizes. Gamma is an aliasing operator, and the expression is as follows
And I is an identity matrix.
Through the steps, the crosstalk noise suppression is carried out on the data collected by the multiple seismic sources under the high-density acquisition observation system, and the separated data are obtained.
Compared with the prior art, the invention has the beneficial effects that:
the unsupervised deep neural network respectively and comprehensively utilizes the nonlinear expression advantages of the residual deep neural network and the U-net deep neural network. Under a high-density acquisition observation system, adjacent 2 seismic sources obtain the same coherent effective signal but have different crosstalk noise respectively. The invention designs an unsupervised deep neural network method with 2 deep neural networks to extract the same coherent effective signals, thereby iteratively inverting and separating simultaneous seismic source data. The unsupervised deep neural network structure of the invention is composed of a residual deep neural network and a U-net deep neural network respectively, and has good nonlinear approximation capability. The designed total loss function consists of 5 basic loss terms, and the overfitting phenomenon of the residual deep neural network and the U-net deep neural network can be avoided. The residual error deep neural network branch is mainly used for extracting an effective signal in the second shot by minimizing a total loss function; the U-net deep neural network branch is mainly used for extracting the validity signal in the first shot. Through fewer iterations, the crosstalk noise is suppressed completely, and simultaneously the seismic source data are separated. The unsupervised deep neural network method training set data is composed of aliasing data, and original unaliased data are not needed to serve as label data, so that the problem of training set loss is solved. The invention has the advantages that:
in the method, similar effective coherent signals can be obtained by adjacent seismic sources in the high-density acquisition process, and the effective coherent signals obtained by different seismic sources provide position information to improve the effect of separating multi-seismic-source data by an unsupervised deep neural network;
the method does not need label data or original unaliased data as training set data, and well solves the problem of training set loss;
in the method, parameters do not need to be adjusted manually, and the intelligent degree of seismic data processing software and processing instruments can be greatly improved.
Drawings
FIG. 1 is a block diagram of a process for separating multi-source data by an unsupervised learning method based on a double-depth neural network constraint provided by the invention.
FIG. 2 is a block diagram of the residual deep neural network architecture of the present invention.
FIG. 3 is a block diagram of the U-net deep neural network architecture of the present invention.
FIG. 4 is a block diagram of an unsupervised deep neural network structure obtained by combining two deep neural networks.
FIG. 5 is a schematic of the time variation of the original unaliased data and aliased data before and after aliasing in the domain of the common detector point for a complex SEAM model. The data are 747 in total, and the time distribution range in the delay time operator is-1.5 seconds. Wherein the track pitch is 25m. (a) And (b) the original unaliased data for the first and second shots, respectively. (c) And (d) aliased data for the first and second shots, respectively. Note that crosstalk noise is different in (c) and (d). Because of the high density acquisition, the original unaliased data for the first shot (a) and the second shot (b) are highly similar or even identical.
Fig. 6 shows the separation and residual of a complex SEAM model in the domain of the common detector point. These data were 747 tracks in total, with a track pitch of 25m. . (a) And (b) the separation of the first shot and the second shot, respectively. (c) A residual between the first-shot original unaliased data and the first-shot separation result. (d) Residual between the second-shot original unaliased data and the second-shot separation result.
FIG. 7 is a plot of actual raw unaliased and aliased data acquired in the North sea area in the domain of the common detector point. The data are 132 paths in total, and the time distribution range in the delay time operator is-0.8 second. Wherein the track pitch is 16.6m. . (a) And (b) the original unaliased data for the first and second shots, respectively. (c) And (d) aliased data for the first and second shots, respectively. Note that crosstalk noise is different in (c) and (d). Because of the high density acquisition, the original unaliased data of the first shot (a) and the second shot (b) are highly similar or even identical.
FIG. 8 shows the separation result and residual error of the actual data in the North sea area in the domain of common detector points. The data had a total of 132 tracks with a track pitch of 16.6m. (a) And (b) the separation of the first shot and the second shot, respectively. (c) A residual between the first-shot original unaliased data and the first-shot separation result. (d) Residual between the second-shot original unaliased data and the second-shot separation result.
Detailed Description
The invention will be further described by way of examples, without in any way limiting the scope of the invention, with reference to the accompanying drawings.
The method provided by the invention is a method for separating multi-seismic source data based on unsupervised learning of double-depth neural network constraint.
FIG. 1 is a block flow diagram of a method for separating multi-source data in unsupervised learning based on dual-depth neural network constraints according to the present invention;
(1) The method comprises the steps of preprocessing the aliasing data, and when the designed observation system is a high-density acquisition observation system, requiring that a main seismic source ship and an auxiliary seismic source ship are respectively excited at a small distance, and sequentially leading to the same moving distance to obtain all multi-seismic-source data. And then, carrying out elevation static correction on the acquired actual multi-source data. And secondly, transforming the static correction data to a common shot point domain gather to perform random noise suppression, surface wave suppression, surge noise suppression and guided wave suppression. And finally, converting to a coherent detection point domain to obtain preprocessing aliasing data containing crosstalk noise. Pre-processing aliased data D may be represented as
Wherein s is i For the delay time operator at the excitation of the ith seismic source, d i The resulting effective coherent signal is excited for the ith source. n is the number of aliases, i.e. the number of seismic sources excited in a short time interval. Passing the preprocessed aliased data through a delay time operator s i By the inverse operator of (A) to obtain aliased data
(2) Carrying out amplitude normalization and size segmentation on the aliasing data to obtain a first source aliasing data block and a second source aliasing data block, namely the aliasing data
Processing amplitude normalization and data segmentation to obtain aliasing data block
Firstly, normalizing the amplitude of aliasing data to be between-1 and 1 through an operator R to obtain normalized data R
Wherein
For delay time operator s i The inverse operator of (2). T represents matrix transposition. D max The element whose absolute value is the maximum value in the preprocessed aliasing data is used. Then, the data obtained after amplitude normalization is processed by a cutting operator K β Cutting the data into data blocks with the size of beta multiplied by beta, and obtaining aliasing data blocks finally used for training the unsupervised deep neural network
(3) And constructing a residual deep neural network and a U-net deep neural network, and enabling the 2 designed deep neural networks to serve as the core of the unsupervised deep neural network. The residual deep neural network and the U-net deep neural network have the following structures:
the first deep neural network is a residual neural network f R . The residual neural network is a multi-layer neural network that solves the degradation problem and the gradient vanishing problem by convolutional layers and cross-layer connections for feature extraction. The convolution layer of the residual error neural network is composed of a plurality of characteristic graphs obtained by performing convolution operation on a previous layer of input image through convolution kernel, the convolution kernel is an important component of the convolution layer, each characteristic graph is composed of a plurality of neural nodes, all neurons on the characteristic graphs share the parameters of the same convolution kernel, the convolution kernel performs convolution operation on the previous layer of input image, feature extraction is realized through the characteristics of local connection and weight sharing in the convolution process, and the parameter quantity is greatly reduced while different features are extracted through the two characteristics of the convolution kernel. The gradient of the convolutional neural network gradually disappears along with the deepening of the network, the shallow parameter cannot be updated, and the short Connections structure ensuresThe residual error neural network structure designed by the invention has 20 convolutional layers and 6 residual error blocks in total. Except the 1 st and the last convolution layers, each 3 convolution layers form a residual block, and the 3 convolution layers are connected in a cross-layer mode through residual connection to achieve identity mapping. The identity mapping ensures that the network performance cannot be reduced, so that the network learns new features on the basis of input features, and the identity mapping does not increase additional parameters and calculated amount for the network, but can accelerate the training speed of the model and optimize the training effect.
The 2 nd deep neural network is a U-net deep neural network f U . U-net is a typical encoding and decoding U-shaped structure, and a symmetrical structure is adopted to connect all layers. The U-net network is characterized in that convolution operation is carried out on a down-sampling part for multiple times, so that the whole network learns more key seismic data characteristics. U-net focuses on the fusion of contextual features. The fusion connection can enable the U-net to fully utilize the extracted features, so that the features of all scales of the original input are reflected in the final feature diagram. The U-net network architecture comprises an input layer, an output layer and a plurality of groups of intermediate layers. The whole network adopts a symmetrical coding and decoding structure, the coding of the left half part extracts data extraction characteristics, and the decoding of the right half part restores data. In order to ensure that the sizes of input and output seismic data are the same, the invention uses the boundary filling method of equal-size filling to ensure that the sizes of input and output are the same. The coding part down-sampling function is used for extracting deep features of the seismic data. The number of convolution kernels of the U-net deep neural network is marked on the convolution layer, and the size of each convolution kernel is 3 x 3. Except for the last layer, a ReLU function is used for activation after each convolution. The size of the pooling window is 2 x 2 with a step size of 2. The upsampling multiple of the upsampled convolutional layer is 2 × 2, and the convolutional kernel size is 2 × 2. The seismic data is subjected to 3 times of downsampling and 3 times of upsampling, the fusion connection is used for fusing the low-layer information and the deep-layer information each time, the information loss caused by the pooling process in the downsampling process is made up, and the information loss is increasedDetail information of the feature image is displayed.
(4) Constructing an unsupervised deep neural network using a residual deep neural network f R And U-net deep neural network f U And jointly combining to construct an unsupervised deep neural network f. Unsupervised deep neural network f has 2 inputs x 1 ；x 2 ]And 2 outputs [ y ] 1 ；y 2 ]Respectively inputting the 1 st input data x 1 And 2 nd input data x 2 And feeding the residual deep neural network branch and the U-net deep neural network branch.
Residual deep neural network f R Is the 1 st input data x 1 The output is the 1 st output data. Residual deep neural network f R The purpose of (1) is to extract the same continuous effective signal in the 1 st input data and the 2 nd input data and output the continuous effective signal as the 1 st output data. Hypothesis residual deep neural network f R The learned deep neural network parameter is theta R Input residual deep neural network f R Is x 1 The resulting output is y 1 The whole process is expressed as y 1 ＝f R (x 1 ,Θ R )。
Taking the first iteration as an example, the residual deep neural network f R The input data of (1) is the aliasing data of the first shot
Under the ideal condition of complete separation, the 1 st output data is an effective coherent signal in aliasing data obtained by the second shot, and the output result obtained by the residual deep neural network can be rewritten into
Residual deep neural network f R Is an estimate of the original unaliased data of the second shot.
U-net deep neural network f U The input data of (1) is the 2 nd input data, and the output is the 2 nd output data. U-net deep neural network f U Is to extract the 2 nd input data and the 1 st input dataThe same consecutive valid signals in the input data are output as the 2 nd output data. Suppose U-net deep neural network f U The learned deep neural network parameter is theta U When inputting U-net deep neural network f U Is x 2 The resulting output is y 2 The whole process is expressed as y 2 ＝f U (x 2 ,Θ U )。
Using the first iteration as an example, the U-net deep neural network f U The input data is second-shot aliasing data
Under the ideal condition of complete separation, the 2 nd output data is an effective coherent signal in the first aliasing data, and the output data obtained by the U-net deep neural network can be rewritten into
U-net deep neural network f U The output data of (a) is an estimate of the original unaliased data of the first shot.
Residual deep neural network f R Branch and U-net deep neural network f U The branch is capable of extracting the effective coherent signals in the second shot and the first shot, respectively. When high-density acquisition is carried out, the original unaliased data of the first shot and the original unaliased data of the second shot are similar, and the difference delta e between the unaliased data of the first shot and the original unaliased data of the second shot approaches 0, namely delta e = d 1 -d 2 ≈0。
Unsupervised deep neural network f will residual deep neural f R Branch and U-net deep neural network f U The branches are combined as a whole and it is desirable to minimize Δ e. The unsupervised deep neural network f inherits the residual deep neural f R Branch and U-net deep neural network f U All of (a). Suppose that the data input into the unsupervised deep neural network is x and the output is y, i.e. x = [ x ] 1 ；x 2 ]And y = [ y = 1 ；y 2 ]. When the unsupervised deep neural network parameter is obtained as Θ through learning, y = f (x, Θ), wherein Θ = [ Θ = R ；Θ U ]。
(5) And (3) training an unsupervised deep neural network, wherein the unsupervised deep neural network f can provide good nonlinear approximation capability. However, how to extract the same coherent signal in the first shot original unaliased data and the second shot original unaliased data is achieved, the key is to train the unsupervised deep neural network by using a proper loss function.
The design of the loss function is discussed with the first iteration separation as an example. The unsupervised deep neural network gets 5 fundamental losses using 2 input data and 2 output data. In residual deep neural network f R In the branch, the obtained output data y 1 And y 2 Respectively, an estimate of the second-shot unaliased data and crosstalk noise in the first-shot aliased data. At this time, the residual deep neural network f R Loss function L in branch 1 And L 2 Are respectively represented as L 1 (Θ R )＝||x 2 -f R (x 1 ,Θ R )|| 1 +μ 1 ||Θ R || 1 And L 2 (Θ R )＝||x 1 -f R (x 1 ,Θ R )|| 1 +μ 2 ||Θ R || 1 In which μ 1 And mu 2 Representing a regularization factor. By minimizing L 1 (Θ R ) Residual deep neural network f R Can extract x 2 Of the effective coherent signal. But if only L is present 1 (Θ R ) Is liable to cause overfitting, resulting in a residual deep neural network f R Is completely close to x 2 . Additional loss function L 2 (Θ R ) Deep neural network f capable of preventing residual errors R Overfitting to extract the correct coherent signal.
Deep neural network f in U-net U In the branch, the obtained output data y 2 Is an estimate of the unaliased data of the first shot. At this time, the U-net deep neural network f U Loss function L in branch 3 And L 4 Are respectively represented as L 3 (Θ U )＝||x 1 -f U (x 2 ,Θ U )|| 1 +μ 3 ||Θ U || 1 And L 4 (Θ U )＝||x 2 -f U (x 2 ,Θ U )|| 1 +μ 4 ||Θ U || 1 In which μ 3 And mu 4 Representing a regularization factor. By minimizing L 3 (Θ U ) U-net deep neural network f U Can extract x 1 A medium effective coherent signal. But if only L is present 3 (Θ U ) Easily causes overfitting, resulting in a U-net deep neural network f U Is completely close to x 1 . Additional loss function L 4 (Θ U ) Deep neural network f capable of preventing U-net U Overfitting to extract the correct coherent signal. The invention uses the residual error deep neural network f R And U-net deep neural network f U Are combined into an unsupervised neural network and expect Δ e to approach 0 and propose a loss function L 5 ，L 5 Can be represented as L 5 (Θ)＝||f R (x 1 ,Θ R )-f U (x 2 ,Θ U )|| 1 +μ 5 ||Θ|| 1 In which μ 5 Representing a regularization factor. By minimizing the loss function L 5 (Θ), it is possible to further prevent the unsupervised deep neural network from overfitting and to obtain an estimate of the correct first shot and second original unaliased data. Thus, the final loss function L (Θ) for training the unsupervised deep neural network f is as follows:
where μ denotes a regularization factor, f 1 And f 2 Respectively representing the 1 st and 2 nd outputs of the unsupervised deep neural network f. By minimizing the loss function L (Θ), the unsupervised deep neural network is able to accurately extract coherent signals in the first and second shot aliased data, respectively.
(6) Obtaining a separation result after crosstalk noise suppression;
the deep neural network parameter Θ is obtained by minimizing the total loss function. The invention will alias data blocks
And sending the data to a trained unsupervised deep neural network with the parameter theta, and combining an iterative inversion separation method to obtain separation data G after crosstalk noise is suppressed. The iterative separation expression of G is
Where P represents combining every 2 adjacent aliased data of the total aliased data. B represents the inverse mapping operator that maps from the data space to the model space. And m is the iteration number. R -1 Is the inverse of R, representing the restoration of the amplitude to the original amplitude.
Is K β Represents the splicing of the divided sizes into the original sizes. The expression of gamma is as follows
Where I is the identity matrix.
As shown in fig. 2, the structure diagram of the residual deep neural network designed by the present invention is shown. The convolution layer in the network structure is composed of a plurality of characteristic graphs obtained by convolution operation of a previous layer of input images through the convolution core, the convolution core is an important component of the convolution layer, each characteristic graph is composed of a plurality of nerve nodes, all nerve cells on the characteristic graphs share parameters of the same convolution core, the convolution core is obtained by convolution operation of the previous layer of input images, feature extraction is achieved through characteristics of local connection and weight sharing in the convolution process, and the parameter quantity is greatly reduced when different features are extracted through the two characteristics of the convolution core. The convolution neural network gradually disappears with the deepening of the network, the shallow parameter cannot be updated, the short Connections structure ensures the updating of the back propagation parameter, and the gradient disappearance problem caused by the back propagation is avoided. Except the 1 st convolution layer and the last convolution layer, each 3 convolution layers form a residual block, and the 3 convolution layers are connected in a cross-layer mode through residual connection to realize identity mapping. The identity mapping ensures that the network performance cannot be reduced, so that the network learns new features on the basis of input features, and the identity mapping can not increase additional parameters and calculated amount for the network, but can accelerate the training speed of the model and optimize the training effect.
As shown in FIG. 3, the structure of the U-net deep neural network designed by the present invention is shown. U-net is a typical encoding and decoding U-shaped structure, and a symmetrical structure is adopted to connect all layers. The U-net network is characterized in that convolution operation is carried out on a down-sampling part for multiple times, so that the whole network learns more key seismic data characteristics. U-net focuses on the fusion of contextual features. The fusion join enables the U-net to make full use of the extracted features, so that the features of each scale of the original input are reflected in the final feature map. The U-net network architecture comprises an input layer, an output layer and a plurality of groups of intermediate layers. The whole network adopts a symmetrical coding and decoding structure, the coding of the left half part extracts data extraction characteristics, and the decoding of the right half part restores data. In order to ensure that the sizes of input and output seismic data are the same, the invention uses the boundary filling method of equal-size filling to ensure that the sizes of input and output are the same. The coding part down-sampling function is used for extracting deep features of the seismic data. The number of convolution kernels of the U-net deep neural network is marked on the convolution layer, and the size of each convolution kernel is 3 x 3. Except for the last layer, a ReLU function is used for activation after each convolution. The size of the pooling window is 2 x 2 with a step size of 2. The upsampling multiple of the upsampled convolutional layer is 2 × 2, and the convolutional kernel size is 2 × 2. The seismic data are subjected to 3 times of downsampling and 3 times of upsampling, low-layer information and deep-layer information are fused by fusion connection every time, information loss caused by a pooling process in the downsampling process is made up, and detail information of the characteristic image is increased.
As shown in fig. 4, it is a structural diagram of an unsupervised deep neural network obtained by combining two deep neural networks. Using residual deep neural networks f R And U-net deep neural network f U And jointly combining to construct an unsupervised deep neural network f. Unsupervised deep neural network f has 2 inputs x 1 ；x 2 ]And 2 outputs [ y ] 1 ；y 2 ]Respectively inputting the 1 st input data x 1 And 2 nd input data x 2 And feeding the residual deep neural network branch and the U-net deep neural network branch. Residual deep neural network f R The input data of (1) th input data x 1 And the output is the 1 st output data. Residual deep neural network f R The purpose of (1) is to extract the same continuous effective signal in the 1 st input data and the 2 nd input data and output the continuous effective signal as the 1 st output data. Hypothesis residual deep neural network f R The learned deep neural network parameter is theta R Input residual deep neural network f R Is x 1 The resulting output is y 1 The whole process is expressed as y 1 ＝f R (x 1 ,Θ R ). Taking the first iteration as an example, the residual deep neural network f R The input data of (a) is the aliasing data of the first shot
Under the ideal condition of complete separation, the 1 st output data is an effective coherent signal in aliasing data obtained by the second shot, and the output result obtained by the residual deep neural network can be rewritten into
Residual deep neural network f R Is an estimate of the original unaliased data of the second shot. U-net deep neural network f U The input data of (1) is the 2 nd input data, and the output is the 2 nd output data. U-net deep neural network f U For the purpose of extracting the 2 nd input data and the 1 st inputAnd inputting the same continuous effective signals in the data and outputting the continuous effective signals as 2 nd output data. Suppose U-net deep neural network f U The learned deep neural network parameter is theta U When inputting U-net deep neural network f U Is x 2 The resulting output is y 2 The whole process is expressed as y 2 ＝f U (x 2 ,Θ U ). Using the first iteration as an example, the U-net deep neural network f U The input data is second-shot aliasing data
Under the ideal condition of complete separation, the 2 nd output data is an effective coherent signal in the first aliasing data, and the output data obtained by the U-net deep neural network can be rewritten into
U-net deep neural network f U The output data of (a) is an estimate of the original unaliased data of the first shot. Residual deep neural network f R Branch and U-net deep neural network f U The branch can extract the valid coherent signals in the second shot and the first shot, respectively. When high-density acquisition is carried out, the original unaliased data of the first shot and the original unaliased data of the second shot are similar, and the difference delta e between the unaliased data of the first shot and the original unaliased data of the second shot approaches 0, namely delta e = d 1 -d 2 0. Unsupervised deep neural network f will residual deep neural f R Branch and U-net deep neural network f U The branches are combined as a whole and it is desirable to minimize Δ e. Unsupervised deep neural network f inherits residual deep neural f R Branch and U-net deep neural network f U All of (a). Suppose the data input into the unsupervised deep neural network is x and the output is y, i.e. x = [ x ] 1 ；x 2 ]And y = [ y = 1 ；y 2 ]. When the unsupervised deep neural network parameter is learned to be Θ, y = f (x, Θ), wherein Θ = [ Θ = R ；Θ U ]。
In the specific implementation of the method, the data obtained by the SEAM complex model is used for measuring the multi-seismic source data separation effect of the method. The SEAM model comprises complex geological structures such as inclined stratum, salt dome and the like, and is a test model commonly used in the industry. FIGS. 5a and 5b show the original unaliased common probe gather data for the first and second shots, for a total of 747 shots, showing that the energy of the salt dome-generated reflector signal is very strong. Under the effect of the random delay time, the second shot field will be superimposed into the first shot field in the form of pseudo-random noise with an aliasing degree of 2, resulting in the first shot aliased data as shown in fig. 5 c. Similarly, the first shot field will be superimposed on the second shot field in the form of pseudo-random noise, resulting in second shot aliased data as in FIG. 5 d. The iterative inversion separation based on the unsupervised deep neural network is iterated for 5 times. FIG. 6a shows the first shot separation result with a converged SNR of 28.85dB, and FIG. 6b shows the second shot separation result with a converged SNR of 28.53dB. It can be seen that the seismic source data have higher signal-to-noise ratio while the SEAM model is separated by using the unsupervised deep neural network method, the effective signals are almost completely recovered, and the weak reflection signals positioned in the salt dome are well reserved. The experiment also proves that the unsupervised deep neural network method can well protect weak signals and has a good separation effect. Fig. 6c is a residual between fig. 5a and 6a, and fig. 6d is a residual between fig. 5b and 6b, and it can be seen that most of crosstalk noise is suppressed well and the residual is small, except for a small amount of noise remaining at a portion where the energy of the reflected wave is too large.
The invention uses actual data collected in the north sea to prove the effectiveness of the unsupervised deep neural network method provided by the invention in separating complex actual seismic data. Extracting 2 adjacent shots of actual data results in the original unaliased data of the first shot and the second shot shown in FIG. 7a and FIG. 7b, respectively. Because of the high density of acquisitions, the spacing between adjacent shots is small, so that fig. 7a and 7b are highly similar or even identical. The direct waves in the unaliased data of the first shot and the second shot are developed, and the direct waves with strong energy can interfere with the separation of the seismic source data at the same time. The aliased data of the first and second shots shown in figures 7c and 7d are obtained under the delay time. Crosstalk noise in the aliased data spreads out over the entire track set, and deep weak significant signals are annihilated. The separation results of the first and second shots using the unsupervised deep neural network approach presented herein are shown in fig. 8a and 8b, respectively, with final separation signal-to-noise ratios of 19.11dB and 18.75dB, respectively. The continuity of shallow strong reflection in-phase axes is good, deep weak reflection in-phase axes are clear, strong noise is completely suppressed, and the earthquake signal to noise ratio is remarkably improved. Fig. 8c is the residual between fig. 7a and 8a, and fig. 8d is the residual between fig. 7b and 8 b. It can be seen from the residual data (fig. 8 c-d) that partial signal leakage occurs mainly at the direct wave of strong energy in the shallow part, but there is almost no effective signal leakage in the deep part and the reflected wave coherent signal is almost not damaged. The test further proves the effectiveness of the unsupervised deep neural network method in separating the complex actual seismic source data at the same time.
It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various alternatives and modifications are possible without departing from the invention and scope of the appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.
Claims (10)
1. A method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning is characterized in that an unsupervised depth neural network model with double-depth neural network constraint is constructed to carry out iterative inversion, and the collected multi-seismic source data are separated through the double-depth neural network constraint-based unsupervised learning method; the method comprises the following steps:
step 1, preprocessing multi-seismic source aliasing data acquired based on a high-density acquisition observation system; the method comprises the following steps: performing elevation static correction on the acquired multi-seismic-source aliasing data; random noise suppression, surface wave suppression, surge noise suppression and guided wave suppression are carried out; obtaining preprocessing aliasing data containing crosstalk noise;
step 2, carrying out amplitude normalization and size division on the preprocessed aliasing data to obtain an aliasing data block for model training;
step 3, constructing a residual deep neural network and a U-net deep neural network;
the residual deep neural network structure comprises a plurality of convolution layers and a residual block; except the 1 st convolutional layer and the last convolutional layer, forming a residual block by every 3 convolutional layers, and connecting the 3 convolutional layers in a cross-layer mode through residual connection to realize identity mapping;
the U-net deep neural network architecture comprises an input layer, an output layer and a plurality of groups of intermediate layers; the network adopts a symmetrical coding and decoding structure, the coding structure is used for extracting data characteristics, and the decoding structure is used for recovering data; the input seismic data and the output seismic data have the same size;
step 4, constructing an unsupervised deep neural network by using the residual deep neural network and the U-net deep neural network;
the constructed unsupervised deep neural network comprises input data [ x ] 1 ；x 2 ]And output data [ y 1 ；y 2 ](ii) a Wherein the 1 st input data x 1 An aliased data block for the first source; 2 nd input data x 2 An aliased data block for the second source; respectively sending the 1 st input data and the 2 nd input data into a residual deep neural network branch and a U-net deep neural network branch;
the residual deep neural network is used for extracting the same continuous effective signals in the 1 st input data and the 2 nd input data and outputting the continuous effective signals as the 1 st output data; the input data of the residual deep neural network is the 1 st input data x 1 Output is the 1 st output data y 1 1 st output data y 1 Is the separation result of the second seismic source;
the U-net deep neural network is used for extracting the same continuous effective signals in the 2 nd input data and the 1 st input data and outputting the continuous effective signals as the 2 nd output data; the input data of the U-net deep neural network is the 2 nd input data, and the output data is the 2 nd output data; the 2 nd output data is a first seismic source separation result;
respectively extracting the effective coherent signal in the second seismic source and the effective coherent signal in the first seismic source by the residual deep neural network branch and the U-net deep neural network branch;
the difference Δ e between the first source raw unaliased data and the second source raw unaliased data approaches 0; the unsupervised deep neural network combines the residual deep neural branch and the U-net deep neural network branch, and minimizes delta e, so that the unsupervised deep neural network inherits all outputs of the residual deep neural branch and the U-net deep neural network;
step 5, training an unsupervised deep neural network: the unsupervised deep neural network obtains 5 basic losses by using 2 input data and 2 output data, designs and minimizes the total loss function of the unsupervised deep neural network model constrained by the double-deep neural network, and obtains the output data y in the residual deep neural network branch 1 And y 2 Respectively, the estimated value of the unaliased data of the second shot and the crosstalk noise in the aliased data of the first shot; output data y obtained in U-net deep neural network branch 2 Is an estimate of the unaliased data of the first shot; mapping the first shot aliasing data and the second shot aliasing data into a separation result after crosstalk noise suppression, and realizing the separation of the multi-seismic-source data;
designing a loss function L separated by the first iteration, which comprises the following steps:
51 ) loss function L in residual deep neural network branches 1 And L 2 Respectively expressed as:
L 1 (Θ R )＝||x 2 -f R (x 1 ,Θ R )|| 1 +μ 1 ||Θ R || 1 (13)
L 2 (Θ R )＝||x 1 -f R (x 1 ,Θ R )|| 1 +μ 2 ||Θ R || 1 (14)
wherein, mu 1 And mu 2 Representing a regularization factor; f. of R Is a residual deep neural network; x is a radical of a fluorine atom 1 、x 2 First input data and second input data, respectively; residual deep neural network f R Learning to obtainThe deep neural network parameter is theta R ；
By minimizing the loss function L in the residual deep neural network branches 1 (Θ R ) Extracting x 2 A medium effective coherent signal; loss function L 2 (Θ R ) For preventing over-fitting of the residual deep neural network;
52 ) loss function L in a U-net deep neural network branch 3 And L 4 Respectively expressed as:
L 3 (Θ U )＝||x 1 -f U (x 2 ,Θ U )|| 1 +μ 3 ||Θ U || 1 (15)
L 4 (Θ U )＝||x 2 -f U (x 2 ,Θ U )|| 1 +μ 4 ||Θ U || 1 (16)
wherein mu 3 And mu 4 Representing a regularization factor; f. of U Is a U-net deep neural network;
by minimizing L 3 (Θ U ) U-net deep neural network f U Extracting x 1 A medium effective coherent signal; loss function L 4 (Θ U ) Deep neural network f for preventing U-net U Overfitting;
53 Using residual deep neural network f R And U-net deep neural network f U Constructing an unsupervised neural network, enabling delta e to approach 0, and designing a loss function L 5 Expressed as:
L 5 (Θ)＝||f R (x 1 ,Θ R )-f U (x 2 ,Θ U )|| 1 +μ 5 ||Θ|| 1 (17)
wherein mu 5 Representing a regularization factor; by minimizing a loss function L 5 (Θ) further preventing over-fitting of the unsupervised deep neural network and obtaining an estimate of the correct first shot and second original unaliased data;
54 The final loss function L (Θ) for training the unsupervised deep neural network f is expressed as follows:
where μ denotes a regularization factor, f 1 And f 2 Respectively representing the 1 st output and the 2 nd output of the unsupervised deep neural network f; obtaining a deep neural network parameter theta by minimizing a loss function L (theta) of the unsupervised deep neural network, namely obtaining a trained unsupervised deep neural network model;
step 6, sending the aliasing data block into a trained unsupervised deep neural network, and respectively extracting coherent signals in the first shot aliasing data and the second shot aliasing data through an iterative inversion separation method; thereby obtaining separation data after crosstalk noise suppression;
performing an iterative inversion separation is represented as follows:
wherein G is the obtained separation data; p represents combining every 2 adjacent aliased data of the total aliased data; b represents an inverse mapping operator which is mapped to a model space from a data space; and m is the iteration number. R is -1 Is the inverse of R, representing the restoration of the amplitude to the original amplitude;
is K β The inverse transformation of (3); f is an aliasing operator; i is an identity matrix;
through the steps, the crosstalk noise suppression is carried out on the multi-source data under the high-density acquisition observation system, and the separated data are obtained.
2. The method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning as claimed in claim 1, wherein the high-density acquisition observation system is excited by a main seismic source ship and an auxiliary seismic source ship respectively in small spacing distances, and the main seismic source ship and the auxiliary seismic source ship sequentially move in the same direction for the same distance to obtain the multi-seismic source data.
3. The method for separating multi-source data in dual-depth neural network constrained unsupervised learning as claimed in claim 2, wherein the small separation distance between the main source ship and the auxiliary source ship is 100 meters.
4. The method for separating multi-seismic source data through unsupervised learning of double-depth neural network constraints as claimed in claim 1, wherein in step 1, random noise suppression, surface wave suppression, surge noise suppression and guided wave suppression are performed on data which are subjected to elevation static correction and are transformed to a common shot point domain gather; and transforming to a co-detection point domain to obtain pre-processing aliasing data containing crosstalk noise; the pre-processed aliased data D is represented as:
wherein s is i For the delay time operator at the excitation of the ith seismic source, d i An effective coherent signal obtained for the ith seismic source excitation; n is the number of overlapping, namely the number of seismic sources excited in a short time interval;
pre-processing aliased data through a delay time operator s i By the inverse operator of (A) to obtain aliased data
Expressed as:
where T represents the matrix transpose.
5. The method for separating multi-source data in the case of unsupervised learning with the constraint of dual deep neural network as claimed in claim 4, wherein in step 2, the aliased data are processed
Normalized to between-1 and 1 by the operator R, resulting in normalized data, expressed as:
wherein
As delay time operator s i The inverse operator of (2); d max The value with the maximum absolute value of the element in the preprocessed aliasing data;
then, cutting the data after amplitude normalization into data blocks with the size of beta multiplied by beta, and obtaining aliasing data blocks for training the unsupervised deep neural network
Expressed as:
wherein, K β Is a cut operator.
6. The method for separating multi-source data through double-depth neural network constrained unsupervised learning as claimed in claim 1, wherein in step 3, the number of convolution kernels of the U-net depth neural network is marked on a convolution layer; except the last convolution layer, activating by adopting a ReLU function after each convolution; the size of each convolution kernel of the U-net deep neural network is 3 multiplied by 3; the size of the pooling window is 2 multiplied by 2, and the step length is 2; the upsampling multiple of the upsampled convolutional layer is 2 × 2, and the convolutional kernel size is 2 × 2.
7. The method for separating multi-source data through double-deep neural network constrained unsupervised learning as claimed in claim 1, wherein in step 3, the residual deep neural network structure specifically comprises 20 convolutional layers and 6 residual blocks.
8. The method for separating multi-source data through unsupervised learning and constrained by dual deep neural network as claimed in claim 1, wherein in step 4, the residual deep neural network f R In the first iteration of (2), the residual deep neural network f R 1 st input data x 1 Aliased data for the first source
Complete separation is performed and the 1 st output data is the separation of the second source, expressed as:
u-net deep neural network f U In the first iteration of (2), the U-net deep neural network f U The input data is second-shot aliasing data
Complete separation is performed, and the 2 nd output data is an effective coherent signal in the first aliasing data, and is expressed as:
u-net deep neural network f U The output data of (a) is an estimate of the original unaliased data of the first shot.
9. The method for separating multi-source data through double-deep neural network constrained unsupervised learning as claimed in claim 8, wherein the constructed unsupervised deep neural network inherits all outputs of the residual deep neural branch and the U-net deep neural network and is expressed as:
x＝[x 1 ；x 2 ] (10)
y＝[y 1 ；y 2 ] (11)
x and y are input data and output data of the unsupervised deep neural network respectively;
when the unsupervised deep neural network parameter obtained by learning is theta, the following parameters are obtained:
y＝f(x,Θ) (12)
wherein Θ = [ Θ R ；Θ U ]。
10. The method for separating multi-source data in dual deep neural network constrained unsupervised learning of claim 1, wherein the multiple sources are two or more sources.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211179180.5A CN115630295A (en) | 2022-09-27 | 2022-09-27 | Method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211179180.5A CN115630295A (en) | 2022-09-27 | 2022-09-27 | Method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115630295A true CN115630295A (en) | 2023-01-20 |
Family
ID=84905195
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211179180.5A Pending CN115630295A (en) | 2022-09-27 | 2022-09-27 | Method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115630295A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116561481A (en) * | 2023-05-10 | 2023-08-08 | 广东海洋大学 | Speed analysis method for AVO effect |
-
2022
- 2022-09-27 CN CN202211179180.5A patent/CN115630295A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116561481A (en) * | 2023-05-10 | 2023-08-08 | 广东海洋大学 | Speed analysis method for AVO effect |
| CN116561481B (en) * | 2023-05-10 | 2023-12-19 | 广东海洋大学 | Speed analysis method for AVO effect |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103926622B (en) | Method for suppressing multiple waves based on L1 norm multichannel matched filtering | |
| Si et al. | Attenuation of random noise using denoising convolutional neural networks | |
| Huang et al. | Self-supervised deep learning to reconstruct seismic data with consecutively missing traces | |
| CN104024890A (en) | Suppressing 4d-noise by weighted stacking of simultaneously acquired wave-fields | |
| Li et al. | Super-resolution of seismic velocity model guided by seismic data | |
| Nakayama et al. | Machine-learning-based data recovery and its contribution to seismic acquisition: Simultaneous application of deblending, trace reconstruction, and low-frequency extrapolation | |
| CN115630295A (en) | Method for separating multi-seismic source data through double-depth neural network constrained unsupervised learning | |
| CN113885077A (en) | Multi-seismic-source seismic data separation method based on deep learning | |
| Wang et al. | Deblending of seismic data based on neural network trained in the CSG | |
| CN108828658A (en) | A kind of ocean bottom seismic data reconstructing method | |
| Wang et al. | Seismic multiple suppression based on a deep neural network method for marine data | |
| Wang et al. | Unsupervised learning for seismic internal multiple suppression based on adaptive virtual events | |
| CN117631028A (en) | Low-frequency reconstruction method for seismic data of multi-scale global information fusion neural network | |
| Wang et al. | Iterative deblending using unsupervised learning with double-deep neural networks | |
| Park et al. | Coarse-refine network with upsampling techniques and Fourier loss for the reconstruction of missing seismic data | |
| Wang et al. | A multi‐data training method for a deep neural network to improve the separation effect of simultaneous‐source data | |
| Cheng et al. | Deblending of simultaneous-source seismic data using Bregman iterative shaping | |
| Li et al. | Low-Frequency Magnetotelluric Data Denoising Using Improved Denoising Convolutional Neural Network and Gated Recurrent Unit | |
| Jin et al. | A robust learning method for low-frequency extrapolation in GPR full waveform inversion | |
| Huang et al. | A fast least-squares reverse time migration method using cycle-consistent generative adversarial network | |
| CN111856559B (en) | Multi-channel seismic spectrum inversion method and system based on sparse Bayes learning theory | |
| Wang et al. | Random noise attenuation using an unsupervised deep neural network method based on local orthogonalization and ensemble learning | |
| CN115061200B (en) | Method for suppressing interlayer multiple based on virtual homophase axis method and unsupervised neural network | |
| Wang et al. | An unsupervised deep learning method for direct seismic deblending in shot domain | |
| CN116990772B (en) | Ground penetrating radar double-parameter real-time inversion method based on multi-scale convolution network |
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 |