阅读说明：本技术 基于数据增广的卷积神经网络地震层间多次波压制方法 (Data-amplification-based seismic interbed multiple suppression method for convolutional neural network ) 是由 胡天跃 刘小舟 刘韬 安圣培 肖彦君 于 2021-08-06 设计创作，主要内容包括：本发明公布了一种基于数据增广的卷积神经网络地震层间多次波压制方法,基于DnCNN和U-Net网络模型,搭建层间多次波压制的深层编解码网络模型,对三维陆地地震勘探资料采用搭建得卷积神经网络模型进行自适应层间多次波压制,实现三维陆地地震勘探资料的高效自适应层间多次波压制；本发明利用神经网络模型实现对三维地震数据的高效自适应层间多次波压制,无需人为调参,且具有良好的抗噪性和较高的计算效率。(The invention discloses a data augmentation-based convolutional neural network seismic interbed multiple suppression method, which is characterized in that a deep coding and decoding network model for interbed multiple suppression is built based on a DnCNN and U-Net network model, and the built convolutional neural network model is adopted for self-adaptive interbed multiple suppression on three-dimensional land seismic exploration data, so that high-efficiency self-adaptive interbed multiple suppression of the three-dimensional land seismic exploration data is realized; the method utilizes the neural network model to realize the high-efficiency self-adaptive interbed multiple suppression of the three-dimensional seismic data, does not need artificial parameter adjustment, and has good noise immunity and higher calculation efficiency.)

1. A convolution neural network seismic interbed multiple suppression method based on data augmentation is used for performing self-adaptive interbed multiple suppression on three-dimensional land seismic exploration data by adopting a convolution neural network, namely, the high-efficiency self-adaptive interbed multiple suppression of the three-dimensional land seismic exploration data is realized by adopting a neural network model method on different survey line data in the stacked three-dimensional seismic data; comprises the following steps:

A. preparing input data:

processing the original seismic data to obtain three-dimensional seismic data after stacking;

B. generating a tagged dataset comprising:

B1. selecting part of measuring lines in the three-dimensional seismic data after the stacking as training set samples;

B2. carrying out interlayer multiple suppression on original data d in the data set, and taking a denoising result as an estimated primary wave p₀Using the suppressed interbed multiples as estimated interbed multiples m₀；

B3. Taking the original data d as a data set sample, and obtaining an estimated primary wave p₀Generating a tagged data set as a tag;

C. establishing a wave field augmentation method for changing data augmentation of an interbed multiple wave field, comprising:

C1. for estimating interbed multiples m₀Changing amplitude, phase and arrival time to obtain interbedded multiples m with changed wave field₁；

C2. The interbed multiples m after changing the wave field₁And estimating the primary beam p₀Adding to obtain noise-containing data d' after the wave field is amplified;

C3. adding d' into the training set as a sample to enable the augmented training set to contain more interbed multiples characteristics, so that the interbed multiples suppression network learns enough interbed multiples characteristics under the condition of limited samples, and estimating the primary p₀The method is used as a label to complete the data amplification of the changed interlayer multiple wave field to obtain a wave field amplification data set;

D. data augmentation with gaussian noise injection:

D1. adding Gaussian noises with different levels to the data in the original training set to generate seismic data with different signal-to-noise ratios;

D2. adding seismic data with different signal-to-noise ratios into a training set as a sample to enable the augmented training set to containThe characteristics of two interferences, namely interlayer multiple waves and Gaussian noise, are obtained, so that an interlayer multiple wave suppression network learns the characteristics of the two interferences; will estimate the primary wave p₀As tag data, completing data amplification of Gaussian noise injection to obtain an amplified data set of Gaussian noise injection, avoiding multiple leakage when processing low signal-to-noise ratio data, and realizing noise-resistant neural network interlayer multiple suppression;

E. building, training and predicting a deep coding and decoding network:

E1. designing and building a deep coding and decoding network model for interlayer multiple suppression based on U-Net, and setting specific parameters of the network model by adopting batch normalization operation in DnCNN;

the deep coding and decoding network model for suppressing the interbed multiples comprises a coding network and a decoding network symmetrical to the coding network, abstract background information and specific detail information in the seismic data are extracted by the five convolutional layers and the five anti-convolutional layers respectively, so that the network model method can accurately suppress the interbed multiples and recover the effective waves; adding batch normalization operation after each convolution layer or each deconvolution layer to accelerate the convergence speed of the network, and activating by using a Relu function; connecting the same-layer feature maps in the coding network and the decoding network through jumping and connecting operations, so that the created network model can extract complex features in the seismic data;

E2.

inputting the augmentation data set obtained in the step D into the built interlayer multiple suppression deep coding and decoding network model for training, and determining the convergence condition of the hyper-parameters in the training process through a data verification set to obtain the trained interlayer multiple suppression deep coding and decoding network model;

E3. inputting other survey line data or a data set to be tested containing interbed multiples into the trained interbed multiple suppression deep coding and decoding network model, and obtaining a seismic interbed multiple suppression result based on the data-augmented convolutional neural network.

2. The data-based augmented convolutional neural network seismic interbed multiple suppression method of claim 1, wherein in step B1, a portion of the lines in the post-stack three-dimensional seismic data are selected as training set samples, and the remaining lines are selected as validation set and test set.

3. The data-based augmented convolutional neural network seismic interbed multiple suppression method of claim 2, wherein the selection of the training set is related to the degree of regional structural variation, and the training set is denser in regions with severe structural variation.

4. The data-based augmented convolutional neural network seismic interbed multiple suppression method of claim 1, wherein step B2 performs interbed multiple suppression on the raw data d in the dataset using a virtual event method.

5. The data-based augmented convolutional neural network seismic interbed multiple suppression method of claim 1, wherein in step E1, the model parameter settings are obtained by retuning for the interbed multiple suppression task; the model comprises two parts, namely an encoding network and a decoding network which is symmetrical to the encoding network, wherein abstract background information and specific detail information in seismic data are extracted by five-layer convolution layers and five-layer deconvolution layers respectively, so that the network accurately suppresses interbed multiples and recovers effective waves; adding batch normalization operation after each convolution layer or each deconvolution layer to enable the network to have higher convergence speed, and activating by using a Relu function; the same-layer feature maps in the coding and decoding network are concatenated by hopping successive operations, enabling the network to extract complex features in the seismic data.

6. The data-augmentation-based seismic interbed multiples suppression method for convolutional neural networks as defined in claim 5, wherein the hyper-parameters during model training are determined by the convergence of different data validation sets, including training rounds, optimizers, and learning rates.

7. The data-based augmented convolutional neural network seismic interbed multiples compression method of claim 5, wherein the model parameters include number of convolutional layer layers, number of convolutional kernels, size, and step size.

Technical Field

The invention belongs to the technical field of exploration seismic data processing, relates to a land seismic data interbed multiple self-adaptive suppression method, and particularly relates to a data-amplification-based convolution neural network seismic interbed multiple suppression method.

Background

In the seismic exploration process, because the underground structure is complex, the seismic data usually contain interference waves, which seriously influences the development of the subsequent seismic data processing, interpretation and other work. The multiple waves between the seismic layers are seismic waves which are reflected for multiple times among underground strata, the generation mechanism is complex, the characteristics of the multiple waves are similar to those of primary wave groups, the residual time difference is small, the prediction and suppression difficulty is high, and adverse effects are brought to the structure and oil gas identification. Therefore, the effective method for researching interlayer multiple suppression has great significance in the field of seismic exploration.

The existing multiple suppression method mainly comprises two main categories of a filtering method and a prediction subtraction method [1 ]. The filtering method is mainly based on seismic wave kinematics and dynamics difference, seismic data are converted to other data domains from a time-space domain, and then are separated according to normal time difference of multiple waves and primary waves or periodicity of the multiple waves, so that multiple wave suppression is realized, such as a prediction deconvolution method [2], a Radon transformation method [3], a bunching filtering method [4-6] and the like. The method has higher calculation efficiency, but has poor effect on processing data with small speed gradient or complex structure. The prediction subtraction method is based on a wave equation, predicts multiples from seismic data, and subtracts the predicted multiples from original data, such as a wave field extension method [7], an SRME method (surface-related multiple attenuation, surface-related multiple compression method) [8], and the like, in which the SRME method for surface multiple compression has been a commercialized module. Because of the strong similarity of the amplitude and time of the interbed multiples and the primary, a plurality of interbed multiples compression methods are proposed in succession, such as a feedback iterative model-based multiples compression method [9], an inverse scattering series method [10], and an imaginary iso-axis method [11 ]. Compared with a filtering method, the prediction subtraction method can obtain a better multiple suppression result for data in a region with a complex structure, but the calculation efficiency is low, the method depends on manual parameter adjustment [12] to a great extent, and multiple leakage is usually caused when low signal-to-noise ratio data is processed.

With the improvement of computer computing power and the development of artificial intelligence technology, deep learning models are widely applied to the field of geophysical science in view of the characteristics of strong feature extraction capability and data driving. Seismic data denoising is one of hot spots, a complex mapping relation of noisy data and clean data is established by using a deep learning model, so that efficient self-adaptive denoising of a new data set is realized, and the purposes of no need of artificial parameter adjustment and remarkable improvement of denoising efficiency are achieved [13 ]. The Convolutional Neural Network (CNN) is prominent in the field of deep learning image denoising, and common networks comprise a U-shaped coding and decoding network (U-Net) [14], a denoising convolutional neural network (DnCNN) [15], a generation countermeasure network [16] and the like. In the field of seismic data denoising, networks such as DnCNN and U-Net [13] [17] and the like obtain better effects in synthesis and random noise suppression of actual data. Furthermore, DnCNN was successfully used to suppress linear noise in synthetic and field data [13 ]. However, there are few technical solutions to apply neural network methods to seismic multiples suppression, and they are currently used only to suppress surface multiples in synthetic seismic data [13 ]. For the interbed multiples, it is difficult to realize adaptive denoising by using a neural network method, mainly for the following reasons: (1) a mature interlayer multi-wave compression method is lacked, so that a primary wave label cannot be generated on actual data and a label data set cannot be established; (2) the interbed multiples are taken as typical regular interference, the characteristics of the waveform and the like of the interbed multiples are related to local structural change, enough label data need to be generated in the three-dimensional seismic data volume, otherwise, the neural network cannot learn the characteristics of the multiples of each measuring line; the problem that the neural network method is long in time consumption is caused by the fact that a large amount of label data are generated by the traditional method.

Reference documents:

[1]Weglein A B.Multiple attenuation：an overview of recent advances and the road ahead[J].The Leading Edge，1999，18(1)：40-44.

[2]Taner M T，O'Doherty R F，Koehler F.Long period multiple suppression by predictive deconvolution in the x-t domain[J].Geophysical Prospecting，2010，43(4)：433-468.

[3]Dan H.Inverse velocity stacking for multiple elimination[C].SEG Technical Program Expanded Abstracts，1986，5(1)：419-419，Houston，Texas.

[4] muster filtering method in seismic data processing [ J ] geophysical science, 2000, 43 (1): 105-115.

[5] The method comprises the following steps of Honghui, picnic jump, Zhang wenpo, and the like, eliminating multiple waves [ J ] in seismic data with low signal-to-noise ratio by using an optimized bunching filtering method, geophysical report 2004, 47 (6): 1106-1110.

[6] Honghe, Hutianjie, Wangrun autumn, utilize three-dimensional bunching filtering method to eliminate correlated noise [ J ]. oil geophysical prospecting, 2005, 40 (1): 42-47.

[7]Verschuur D J，Berkhout A J.Removal of internal multiples with the common-focus-point (CFP)approach：Part 2-Application strategies and data examples[J].Geophysics，2005，70 (3)：61-72.

[8]Baumstein A，Hadidi M T.3D surface-related multiple elimination：Data reconstruction and application to field data[J].Geophysics，2006，71(3)：25-33.

[9]Brookes D.Case studies in 3D interbed multiple attenuation[J].Leading Edge，2011，30(8)： 914-918.

[10]Zou Y，Weglein A B.A new method to eliminate first order internal multiples for a normal incidence plane wave on a 1D earth[C].SEG Technical Program Expanded Abstracts，2013， 4136-4140，Houston，Texas.

[11] Liujiahui, Hutianyun, Peng for updating, self-adaptive virtual event axis method for suppressing multiple [ J ] between earthquake layers, geophysical report, 2018, 61 (3): 1196-1210.

[12] Multiple wave identification and suppression method between layers in the QT region [ J ] oil and gas reservoir evaluation and development, 2020, 10 (6): 121-125.

[13]Yu S，Ma J，Wang W.Deep learning for denoising[J].Geophysics，2019，84(6)：V333-V350. [14]Mao X，Shen C，Yang Y B.Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections[C].Advances in neural information processing systems， 2016：2802-2810，Barcelona，Spain.

[15]Zhang K，Zuo W，Chen Y，et al.Beyond a gaussian denoiser：Residual learning of deep cnn for image denoising[J].IEEE transactions on image processing，2017，26(7)：3142-3155.

[16]Chen J，Chen J，Chao H，et al.Image blind denoising with generative adversarial network based noise modeling[C].Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition，2018：3155-3164，Salt Lake City，UT，USA.

[17] Wangyqing, Luwenka, Liujin forest, etc. earthquake random noise suppression based on data augmentation and CNN [ J ]. geophysical newspaper, 2019, 62 (1): 421-433.

Disclosure of Invention

Aiming at the technical problems, the invention provides a neural network interlayer multi-wave self-adaptive suppression method for three-dimensional land seismic data, and a better suppression effect is obtained in actual seismic data of a certain basin in China. The input data is post-stack three-dimensional land seismic data, and the neural network efficient self-adaptive suppression of the multiple between different survey line data layers in the three-dimensional seismic data is realized through tag data set generation, data augmentation, neural network training and neural network prediction.

The core of the invention is: a small amount of labeled data sets are generated by utilizing the conventional interlayer multiple suppression method, a training set is expanded by two data amplification methods aiming at interlayer multiple suppression designed by the invention, and the amplified data sets are input into a newly built deep coding and decoding network model suitable for interlayer multiple suppression for training, so that the efficient self-adaptive interlayer multiple suppression of the three-dimensional land seismic data without manual parameter adjustment and with good anti-noise performance is realized. The invention designs two data augmentation methods aiming at interlayer multiple suppression: by changing data amplification of an interlayer multiple wave field, the problem that a large number of primary wave labels are needed when a neural network method is used for multiple wave compression is solved, and self-adaptive interlayer multiple wave compression under a limited training label is realized; by means of data amplification of Gaussian noise injection, the problem of multiple leakage when the existing interlayer multiple pressing method is used for processing low signal to noise ratio data is solved, and interlayer multiple pressing with noise resistance is achieved. In addition, a deep coding and decoding network suitable for interlayer multiple compression is built by combining the characteristics of the DnCNN network and the U-Net network, and the problems that the manual parameter adjustment is required and the time consumption is too long in the conventional interlayer multiple compression method can be solved after network training. The invention realizes the high-efficiency self-adaptive interbed multiple suppression of the three-dimensional seismic data by utilizing the neural network without artificial parameter adjustment, and has good noise immunity and higher calculation efficiency.

The technical scheme provided by the invention is as follows:

a convolution neural network seismic interbed multiple suppression method based on data augmentation is a convolution neural network adaptive interbed multiple suppression method of three-dimensional land seismic exploration data, the method adopts a neural network method for different survey line data in the three-dimensional stacked seismic data to realize efficient adaptive interbed multiple suppression of the three-dimensional land seismic data, and the method comprises the following steps (as shown in figure 1):

A. input data preparation: processing the original seismic data to obtain three-dimensional seismic data after stacking;

B. generation of tagged data sets:

B1. selecting part of measuring lines in the three-dimensional seismic data after stacking as a training set sample, and using the rest measuring lines as a verification set and a test set, wherein the selection of the training set depends on the structural change degree of the area, and the training set is denser in the area with severe change;

B2. utilizing a virtual homomorphic axis method to carry out interlayer multiple suppression on original data d in the data set, and respectively taking a denoising result and the suppressed interlayer multiple as an estimated primary wave p₀And estimating the interbed multiples m₀；

B3. Will be originalUsing the initial data d as a data set sample, and obtaining an estimated primary wave p by a virtual homodyne axis method₀As a label, a labeled data set is generated, so that the problem that a training label is difficult to obtain when a neural network is used for interlayer multiple suppression is solved;

C. data augmentation to alter interbed multiples wavefields (known as wavefield augmentation method):

C1. interlayer multiple m estimated by virtual homophase axis method₀Changing amplitude, phase and arrival time to obtain interbedded multiples m after changing wave field₁The magnitude of the wavefield change needs to be selected after the trial based on the particular data;

C2. the interbed multiples m after changing the wave field₁And estimating the primary beam p₀Adding to obtain noise-containing data d' after the wave field is amplified;

C3. adding d' into the training set as a sample, enabling the augmented training set to contain more complex interbed multiples characteristics so as to ensure that the interbed multiples suppression network learns enough interbed multiples characteristics under the condition of limited samples, and estimating the primary p of the original data by using a virtual homodyne axis method₀The method is used as a label to complete data amplification for changing an interbed multiple wave field, obtain a wave field amplification data set, solve the problem that a large number of primary wave labels are needed when the neural network method is used for interbed multiple suppression to a certain extent, and realize the neural network self-adaptive interbed multiple suppression under a limited training label;

D. data augmentation for gaussian noise injection:

D1. adding Gaussian noises with different levels to the data in the original training set to generate seismic data with different signal-to-noise ratios;

D2. adding seismic data with different signal-to-noise ratios into a training set as a sample, enabling an augmented training set to contain characteristics of two interferences of interbed multiples and Gaussian noise, ensuring that an interbed multiples suppression network learns the characteristics of the two interferences, having the capability of suppressing the two interferences, and estimating a primary wave p of original data by using the existing virtual event-to-event method₀As tag data, completing data amplification of Gaussian noise injection to obtain Gaussian noise injectionThe data set is enlarged, so that the problem of multiple leakage when the existing interlayer multiple pressing method is used for processing low signal-to-noise ratio data is solved, and the interlayer multiple pressing of the neural network with noise resistance is realized;

E. building, training and predicting a deep coding and decoding network:

E1. building a deep coding and decoding network model suitable for interlayer multiple suppression, wherein the macro architecture refers to a U-Net design, batch normalization operation in DnCNN is added, and specific parameter setting is obtained by debugging the interlayer multiple suppression task again, such as the number of layers of convolution layers, the number of convolution kernels, the size and the step length; the model comprises two parts, namely an encoding network and a decoding network which is symmetrical to the encoding network, wherein abstract background information and specific detail information in seismic data are extracted by five-layer convolution layers and five-layer deconvolution layers respectively, so that the network accurately suppresses interbed multiples and recovers effective waves; adding batch normalization operation after each convolution layer or each deconvolution layer to enable the network to have higher convergence speed, and activating by using a Relu function; connecting the same-layer feature maps in the coding and decoding network through jumping successive operations, so that the network can extract complex features in the seismic data;

E2. inputting the augmented data set into a built interlayer multiple suppression deep coding and decoding network for training, wherein the hyper-parameters in the training process are determined by the convergence conditions of verification sets of different data, such as training rounds, optimizers, learning rates and the like;

E3. after the network training is finished, other line measurement data containing the interbed multiples are input into the deep coding and decoding network, and then the interbed multiples pressing result of the method can be obtained, the neural network interbed multiples pressing with strong noise resistance and adaptability is realized, and the problems of manual parameter adjustment requirement and long time consumption of the existing method are solved.

The invention has the beneficial effects that:

the existing commonly used interbed multiple pressing method is low in calculation efficiency, the pressing effect depends on artificial parameter adjustment, and interbed multiple leakage usually exists when low signal-to-noise ratio data is processed. The method provided by the invention utilizes the neural network to realize the efficient self-adaptive compression of the interbed multiples, takes a small quantity of interbed multiple compression results obtained by a virtual homophase axis method as labels, and utilizes the two data amplification methods aiming at the interbed multiple compression and the newly built deep coding and decoding convolution neural network suitable for the interbed multiple compression to train, thereby realizing the self-adaptive interbed multiple compression of the neural network, needing no artificial parameter adjustment, obviously improving the calculation efficiency and having better noise immunity. The technical advantages of the invention are as follows:

in the method, the problems of artificial parameter adjustment requirement and long time consumption of the conventional interbed multiple method are solved by utilizing the strong characteristic extraction capability of deep learning, and efficient and self-adaptive interbed multiple suppression on a new data set is realized;

the invention constructs a deep coding and decoding network suitable for interlayer multiple suppression, can effectively learn the characteristics of different signals in seismic data, more accurately suppress interlayer multiples and protect primary waves, and has smaller error of suppression results;

based on the characteristics of deep learning data driving, the invention designs two data amplification methods aiming at interlayer multiple suppression to ensure that the network has better performance: by changing the data amplification of an interlayer multiple wave field, the problem that a large number of primary wave labels are needed when a neural network method is used for multiple suppression is solved, and the neural network self-adaptive interlayer multiple suppression under a limited training label is realized; by means of data amplification of Gaussian noise injection, the problem of multiple leakage when the existing interlayer multiple suppression method processes low signal to noise ratio data is solved, and noise-resistant interlayer multiple suppression of the neural network is achieved.

Drawings

FIG. 1 is a block diagram of a flow chart of a method for suppressing an interbed multiple of a convolutional neural network based on data augmentation according to the present invention.

FIG. 2 is a deep codec network structure constructed according to the present invention;

the network comprises two parts, namely an encoding network and a decoding network symmetrical to the encoding network. The coding network consists of five convolutional layers and abstract feature extraction is carried out through down-sampling; the decoding network consists of five deconvolution layers and recovers the detail characteristics by up-sampling; jumping connection operation is added between the second to the fifth layers of output characteristic diagrams of the decoding and coding network, so that the network fuses characteristics with different dimensionalities; the network input is original data containing interbed multiples, and the network output is a primary wave obtained by suppressing interbed multiples.

FIG. 3 is a schematic diagram of a tilted layered model for generating synthetic data containing interbed multiples;

wherein the shot point and the wave detection point are both positioned on the earth surface; the color scale in the figure indicates the magnitude of the velocity values.

FIG. 4 is an example of wave field augmented synthetic data using a neural network approach in accordance with an embodiment of the present invention;

wherein the synthetic data is obtained by forward modeling of the model of fig. 3; (a) full wavefield data containing interbed multiples; (b) data after wave field augmentation; (c) carrying out interlayer multiple pressing on the (a) by a virtual homomorphic axis method to obtain a primary wave label; (d) the result of interlayer multiple pressing of the data of (b) by the method of the invention; (e) the error of the multiple wave suppression between layers, i.e., (c) - (d), is applied to the data of (b) for the method of the present invention.

FIG. 5 is an example of Gaussian noise injection augmented synthesized data using a neural network approach in accordance with an embodiment of the present invention;

wherein, the dotted line arrow and the solid line arrow respectively mark the multiple residual and the primary loss; (a) injecting the amplified data with the signal-to-noise ratio of 5dB for Gaussian noise; (b) the result of interlayer multiple pressing of (a) by a virtual homophase axis method; (c) is the result of the inventive method of interlayer multiple pressing of (a); (d) performing interlayer multi-wave compression on the original data by a virtual homophase axis method to obtain label data; (e) the error of interlayer multiple suppression for (a) by the virtual homophase axis method, namely (d) - (b), has multiple leakage; (f) the errors in the interlayer multiple suppression of (a), i.e., (d) - (c), are performed for the method of the present invention.

FIG. 6 is a result of processing actual three-dimensional post-stack land seismic data for a basin in an embodiment of the invention;

wherein, the dashed ellipse marks the larger error generated by the neural network method; (a) is original data; (b) selecting 1/2 line as the neural network pressing result of the label; (c) to select 1/4 line as the pressing result of the label; (d) selecting 1/6 measuring line as label and carrying out data amplification pressing result; (e) to select 1/6 line as the pressing result of the label; (f) the result is a virtual co-axial method pressing result; (g) to select 1/2 side lines as the compression error for the label, i.e., (b) - (f); (h) to select 1/4 side lines as the compression error for the label, i.e., (c) - (f); (i) to select 1/6 gauge as a label and perform the data-augmented hold-down errors, i.e., (d) - (f); (j) is the compression error for selecting 1/6 side line as the label, i.e. (e) - (f).

FIG. 7 is a result of processing low signal-to-noise ratio land seismic data for a basin in an embodiment of the invention;

wherein, the dotted ellipse marks the larger error generated by the neural network method, and the arrow marks the weaker primary wave recovered by the virtual homophase axis method; (a) is original data; (b) the result of the pressing of (a) by a virtual co-axis method is shown; (c) low signal-to-noise ratio data after Gaussian noise injection; (d) is the result of the pressing of (c) by the process of the invention; (e) disturbances (c) - (d) suppressed for the method of the invention; (f) errors (b) - (d) are compressed for the method of the present invention.

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 three-dimensional land seismic data convolution neural network interbed multiple suppression method, can realize high-efficiency interbed multiple suppression of land seismic data, and obtains a better interbed multiple suppression effect in actual land seismic data of a certain basin in China. The input data is post-stack three-dimensional land seismic data, and adaptive suppression of multiple waves among different survey line data layers in the three-dimensional seismic data is realized through tag data set generation, data augmentation, neural network training and neural network prediction.

The invention establishes a deep coding and decoding network and provides two data augmentation methods which are designed aiming at an interlayer multiple suppression task. FIG. 1 is a block flow diagram of a convolutional neural network adaptive interbed multiple suppression method provided by the present invention; the method for realizing interlayer multiple suppression on the three-dimensional land seismic data by applying the neural network method comprises the following steps of:

A. input data preparation: processing original seismic data to obtain post-stack three-dimensional seismic data, wherein the processing flow mainly comprises static correction, pre-stack denoising, amplitude compensation, deconvolution, velocity analysis, dynamic correction, horizontal stacking and migration, the pre-stack denoising comprises random noise, linear interference, surface waves, surface multiple suppression and the like, the specific processing flow is designed according to the characteristics of different seismic data, and the post-stack three-dimensional seismic data is obtained after the processing is finished;

B. generation of tagged data sets:

when seismic data are subjected to interbed multiple suppression by using a neural network method, training needs to be performed on the basis of a labeled data set. Therefore, a primary wave is estimated for original data by using a virtual homomorphic axis method to generate a labeled data set, and the specific implementation process is as follows:

B1. selecting part of measuring lines in the three-dimensional seismic data after stacking as a training set sample, and using the rest of measuring lines as a verification set and a test set, wherein the selection of the training set depends on the structural change degree of the area, and the area with severe change is more dense; the method is used for testing basin seismic data, the geological conditions of the area are complex, the structures such as a fracture body and a fault are relatively developed and have certain changes, so that 1 contact test line in every 2, 4 and 6 contact test lines is tried to be used as a training set, the number of the training sets is properly increased at the position with larger data change, and at most, 1 contact test line in every 2 test lines is not exceeded;

B2. utilizing a virtual homomorphic axis method to carry out interlayer multiple suppression on original data d in the data set, and respectively taking a denoising result and the suppressed interlayer multiple as an estimated primary wave p₀And estimating the interbed multiples m₀The raw data model can be represented by equation 1:

d(t)＝p₀(t)+m₀(t) (formula 1)

Wherein t is a time sampling point;

B3. taking original data d as a data set sample, and obtaining an estimated primary wave p by a virtual homodyne axis method₀As a label, a labeled data set is generated, so that the problem that a training label is difficult to obtain when a neural network is used for interlayer multiple suppression is solved;

C. data augmentation to change the interbed multiple wavefield:

the invention designs a data amplification method for changing an interlayer multiple wave field, which is called a wave field amplification method for short, aiming at an interlayer multiple suppression task, so that limited label data are utilized more fully, the problem that a large number of primary wave labels are needed when a neural network method is used for multiple suppression is solved, and the network can keep better interlayer multiple suppression capability on a new data set under the condition of the limited labels. The wave field augmentation method is realized by the following specific steps:

C1. interlayer multiple m estimated by virtual homophase axis method₀Changing amplitude, phase and arrival time to obtain interbedded multiples m after changing wave field₁The magnitude of the wavefield change needs to be chosen after trial based on the specific data, which is expressed by equation 2:

m₁(t)＝m₀(t + Deltat). gamma (formula 2)

Wherein t is a time sampling point, Δ t represents the number of sampling points of the interbed multiples floating in the time dimension, γ represents the interbed multiple amplitude and phase change coefficient, the phase reversal of the wave field is represented when a negative number is taken, the absolute value of the amplitude is reduced when the absolute value is less than 1, the absolute value of the amplitude is increased when the absolute value is greater than 1, and | γ | -1| represents the percentage of the amplitude floating;

C2. the interbed multiples m after changing the wave field₁And estimating the primary beam p₀And adding to obtain noise-containing data d' after wave field amplification, wherein the formula can be represented by formula 3:

d'(t)＝p₀(t)+m₁＝p₀(t)+m₀(t + Deltat). gamma (formula 3)

Wherein t is a time sampling point, delta t represents the number of sampling points of interbed multiples floating in a time dimension, and gamma represents interbed multiple amplitude and a phase change coefficient;

C3. adding d' into the training set as a sample, enabling the augmented training set to contain more complex interbed multiples characteristics so as to ensure that the interbed multiples suppression network learns enough interbed multiples characteristics under the condition of limited samples, and estimating the primary p of the original data by using a virtual homophasic axis method₀The method is used as a label to complete the data amplification of changing an interlayer multiple wave field, solves the problem that a large number of primary wave labels are needed when a neural network method is used for interlayer multiple pressing to a certain extent, and realizes the neural network self-adaptive interlayer multiple pressing under a limited training label;

D. data augmentation for gaussian noise injection:

the invention designs a data amplification method of Gaussian noise injection aiming at an interlayer multiple suppression task, so as to avoid the problem of multiple leakage when the existing interlayer multiple suppression method processes low signal-to-noise ratio data, and the interlayer multiple suppression method has noise immunity to random noise. The specific implementation process is as follows:

D1. adding Gaussian noises with different levels to the data in the original training set to generate seismic data with different signal-to-noise ratios;

D2. adding seismic data with different signal-to-noise ratios into a training set as a sample, enabling an augmented training set to contain characteristics of two interferences of interbed multiples and Gaussian noise, ensuring that an interbed multiple suppression network learns the characteristics of the two interferences, having the capability of suppressing the two interferences, and estimating a primary wave p of original data by using the existing virtual event-to-event method₀As label data, the data amplification of Gaussian noise injection is completed, so that the problem of multiple leakage when the existing interlayer multiple suppression method is used for processing low signal-to-noise ratio data is solved, and the interlayer multiple suppression of the neural network with noise resistance is realized.

E. Building, training and predicting a deep coding and decoding network:

the common interbed multiple pressing method is low in calculation efficiency, the effect depends on manual parameter adjustment, and multiple leakage exists when low signal-to-noise ratio data are processed. Therefore, the invention utilizes the characteristics of strong characteristic extraction capability and data driving of deep learning to realize high-efficiency self-adaptive interbedded multiple suppression with noise immunity and avoid the problems of manual parameter adjustment requirement and poor noise immunity. Based on the structure of U-Net, the invention adds batch normalization operation in the DnCNN model, and the establishment and parameters of the network model are obtained by designing and debugging the interbedded multiple suppression task. The principle method of interlayer multiple suppression using a convolutional neural network is described below:

seismic data containing interbed multiples and random noise can be represented by equation 4:

d as p + m + n (formula 4)

Wherein d is original data; p is primary wave data; m and n are the interbed multiples and random noise in the original data respectively.

The goal of suppressing noise using neural networks is to recover p from d without any assumptions. The denoising process can be represented by equation 5:

in the formula:the primary wave output after the multiple wave passes between layers is suppressed for the network, and is an approximation of the real primary wave p; theta is a parameter to be optimized in the network, and the convolutional neural network comprises convolutional kernel weight w and deviation b; net is a network structure. In the network training process, input data are transmitted forward at first, and output data are obtained after layer-by-layer calculation in the network; after that, the output data is back-propagated to update the network parameters, which are updated by the criterion of minimizing the loss function, as follows:

in the invention, the Mean Square Error (MSE) of output data and label data obtained by network calculation is used as a loss function to improve the generalization capability of the model, and the process of updating the parameters is equal to the process of reducing the mean square error of the output and the primary wave label to the minimum; alpha is a learning rate, represents the updating amplitude of network parameters, uses a larger learning rate at the initial training stage, can enable the model to quickly obtain a better solution, and then gradually reduces the learning rate to enable the model to stably converge to obtain the optimal solution.

E1. And (3) building a deep coding and decoding network model suitable for interlayer multiple suppression, wherein the macro architecture refers to a U-Net design, batch normalization operation in DnCNN is added, and specific parameter setting is obtained by debugging the interlayer multiple suppression task again, such as the number of layers of convolution layers, the number of convolution kernels, the size, the step length and the like.

In the task of suppressing the interbed multiples, the arrival times of the interbed multiples and the primary waves in the same group of seismic data are relatively fixed, so that a network needs to extract macroscopic abstract position information to distinguish and identify the interbed multiples and the primary waves; meanwhile, the network needs to accurately suppress the interbed multiples and restore the details of the primaries, so that more specific local information needs to be extracted. The invention refers to two classic convolutional neural network architectures and builds a network which can simultaneously extract abstract information and detail information and is suitable for interbedded multiple suppression. The DnCNN is remarkable in image denoising, and batch normalization operation used by the network can normalize the input value of each layer to a more standard distribution, so that a nonlinear function is more sensitive to input, the problem of gradient disappearance in the network training process is solved, and the network convergence process is accelerated. The U-Net is a classic network in the field of image segmentation, the network adopts a coding and decoding structure, so that the network can simultaneously extract abstract and detail features, and the network jumps to be connected with the same-layer output in the coding and decoding network, so that the network fuses multidimensional information and extracts complex features. Therefore, the invention combines the batch normalization operation of DnCNN and the macro architecture of U-Net to build the interbed multiple suppression network, and resets the specific parameters of the network aiming at the interbed multiple suppression task.

The network structure constructed by the invention is shown in fig. 2 and comprises two parts, namely an encoding network and a decoding network symmetrical to the encoding network. The coding network is composed of five convolutional layers, the size of the convolutional core is set to be 4 multiplied by 4, the step length of the convolutional core is 2, the size of a feature graph output by the convolutional layers is changed into 1/2 of input, so that the network can extract abstract background features through downsampling, and the interlayer multiples and the primary waves are identified and judged according to the macroscopic position relation between the two; doubling the number of convolution kernels layer by layer from the second convolution layer, so that the number of output feature maps is enlarged by 2 times to fully reserve effective features; the decoding network is composed of five layers of deconvolution layers, the network can accurately suppress the interbed multiples and recover the specific detail information of the primary wave by up-sampling and extracting the detail characteristics, the size, the quantity, the step length and the like of the convolution kernels are the same as the parameter setting of the same-layer convolution kernels in the coding network, the size of the output characteristic graph of each deconvolution layer is enlarged by 2 times, the quantity is changed into 1/2, and finally the output characteristic graph is recovered to the size of the input data; after the convolution layer and the deconvolution layer below the second layer are added with batch normalization operation, the network has higher convergence speed; the activation function selects a Relu function to accelerate network convergence; and (3) jumping the feature maps of the same layer in the successively encoded network by the feature maps output by the second to the five deconvolution layers of the decoding network, so that the network fuses multidimensional information and extracts complex features in the seismic data. Finally, a convolution layer containing 1 convolution kernel is added into the network, and the Tanh function is selected to be activated and then output, so that the number of output channels is recovered to the number of input channels. The network input is original data containing interbed multiples, and the network output is a primary wave obtained by suppressing interbed multiples.

E2. Inputting the augmented data set into the built interlayer multiple suppression deep coding and decoding network for training, wherein the hyper-parameters in the training process are determined by the convergence conditions of verification sets of different data, such as training rounds, an optimizer and a learning rate; the optimizer selected in the algorithm of the present invention is Adam, initialLearning rate of 1 × 10^-3The network convergence continuously decreases, the number of network training rounds is 600 rounds, and the convergence is basically stable around 250 th round;

E3. after the network training is finished, inputting other line measurement data containing the interbed multiples into a deep coding and decoding network, so that an interbed multiple pressing result of the method can be obtained, the interbed multiple pressing of the neural network with strong noise resistance and adaptability is realized, and the problems of manual parameter adjustment requirement and long time consumption of the existing method are avoided;

in the specific implementation of the method, the effect of suppressing the interbed multiples is verified by using the synthetic seismic data. Using the geological model shown in fig. 3, forward modeling is performed through a controllable layer hierarchical interbed multiple simulation method to obtain synthetic seismic data, after a labeled data set is generated by using the existing virtual event-phase axis method, two types of data amplification aiming at interbed multiple suppression are performed, so that a training set is expanded, the sample ratio of the training set, a verification set and a test set is 10:1:1, and samples in different data sets are not overlapped at all. Then, the deep coding and decoding network training set up by the method is utilized, reasonable hyper-parameters are selected through the condition of convergence of a verification set, an optimizer is Adam, the number of network training rounds is 600, the convergence is basically stable around 250 th round, and a good interlayer multiple suppression effect is obtained after the training is finished.

FIG. 4 shows the result of interbed multiple suppression of wavefield amplification data according to the present invention. Because the existing virtual event method cannot use real primary waves to match and suppress and change the interbed multiples of the wave field, the method is not used for processing wave field augmentation data. Fig. 4a is original synthetic data, fig. 4b is wave field augmentation data, a wave field augmentation parameter Δ t belongs to [0, 5], γ belongs to [0.5, 1.5], an amplitude floating range is 50%, fig. 4c is a primary wave label obtained by performing interlayer multiple suppression on the original data of fig. 4a by a virtual homodyne axis method, and fig. 4d and e are results and errors of the interlayer multiple suppression performed on fig. 4b by the method of the present invention, respectively. As a data-driven algorithm, under the condition of only original data and a virtual homophase axis label, the diversity of multiple features in a training set is increased through wave field amplification, and the interlayer multiple suppression capability of a neural network method on a new data set can be improved; therefore, the interlaminar multi-wave pressing method can effectively press the multi-waves in the new data set and has stronger adaptability.

FIG. 5 shows the result of inter-layer multiple suppression of Gaussian noise injected augmented data according to the present invention with an augmented data SNR of 0-30 dB. FIG. 5a is 5dB synthetic data after noise injection augmentation, FIGS. 5b and c are the results of inter-layer multiple suppression of the augmented data of FIG. 5a by the virtual event method and the method of the present invention, respectively, with dashed and solid arrows indicating multiple residual and primary loss, respectively; FIG. 5d is a label data obtained by interlayer multiple suppression of original data by the virtual event method; fig. 5e, f are the errors of inter-layer multiple suppression of the augmented data of fig. 5a by the virtual event method and the inventive method, respectively. When the virtual event axis method is used for processing 5dB seismic data, obvious multiple leakage and primary loss exist, and are respectively shown as a dotted arrow and a solid arrow in fig. 5 e; by the data amplification method of Gaussian noise injection, the neural network can learn the characteristics of two noises contained in the training set, and has anti-noise performance while having interlayer multiple suppression capability; therefore, the method can adaptively suppress multiple waves and Gaussian noise in new low signal-to-noise ratio data, fully protect effective waves and suppress errors.

FIG. 6 shows the result of interbed multiple suppression applied to the three-dimensional seismic data after a basin is stacked. FIG. 6a shows the raw data of 1865 th crossline, and FIG. 6f shows the compression results of the dashed-iso-axis method. When the number of the labels is 1/2 or 1/4 of the total number of the survey lines, the interbed multiple pressing result is better, and no obvious error exists, as shown in fig. 6g and h. Fig. 6e and j show the suppression result and error of the present invention when 1 tag is selected for each 6 lines to make a tag, and the prediction result of the deep codec network has relatively continuous effective wave loss due to the small number of samples, as shown by the dashed ellipse in fig. 6 j. Fig. 6d, i is the result and error of inter-layer multiple suppression of deep layer coding and decoding network after wave field augmentation to 1/6 label, the wave field augmentation parameter is Δ t ∈ [0, 1], γ ∈ [0.9, 1.1], and the amplitude floating range is 10%. The increased pressing error is significantly smaller (fig. 6i) than the non-increased pressing error (fig. 6 j); and the increased pressing error of 1/6 label (FIG. 6i) was not significantly increased compared to the pressing error of 1/4 label (FIG. 6 h). The wave field augmentation method can relieve the problem of lack of training samples, improve the adaptability of the network to suppress the interbed multiples of the new data set, and ensure that the neural network still has a good interbed multiples suppression effect during small sample training.

FIG. 7 shows the result of interbed multiple suppression of seismic data of a basin with Gaussian noise added by the algorithm of the present invention. FIG. 7a is the original seismic data of line 1690, and FIG. 7b is the result of suppressing the original data by the virtual event method, where the arrows indicate the weaker primaries recovered by the virtual event method. FIG. 7c is 5dB seismic data after Gaussian noise is added, after the data is processed by the method of the invention, multiple interlaminar waves and Gaussian noise are fully suppressed, weak primary waves at the arrow are well recovered (FIG. 7d), and a suppression error indicated by a dotted ellipse in FIG. 7f is weak. Under the condition of low signal-to-noise ratio, the interlayer multiple pressing method can still fully press interlayer multiples and protect primary waves, and has better noise resistance. The data amplification method of Gaussian noise injection in the invention can lead the network to learn the characteristics of two kinds of noise, and realizes the simultaneous suppression of the new data concentration interlayer multiples and random noise.

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 substitutions and modifications are possible without departing from the spirit and scope of the invention and 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.