CN113962244A - Rayleigh wave seismic data noise removal method, storage medium and electronic device - Google Patents

Abstract

The invention discloses a method for removing Rayleigh wave seismic data noise, which comprises the following steps: constructing a deep convolution to generate a countermeasure network; preprocessing Rayleigh wave seismic data to obtain real noise-free data; adding Gaussian white noise based on the noise-free data to obtain noise-containing data; adding real noise-free data and noise-containing data serving as training samples into a deep convolution to generate a confrontation network for training, and performing iterative computation to obtain a confrontation network model generated by the deep convolution; and denoising other Rayleigh wave seismic data to be denoised. According to the method, the countermeasure network is generated based on the deep convolution in the deep learning field, random noise removal is carried out on the multi-channel surface wave data, high-quality denoised Rayleigh wave seismic data are obtained, parameters do not need to be adjusted manually, the labor cost is reduced, the denoising efficiency of the Rayleigh wave seismic data is improved, and the denoising effect is better.

Description

Rayleigh wave seismic data noise removal method, storage medium and electronic device

Technical Field

The invention relates to the technical field of engineering geophysical exploration, in particular to a Rayleigh wave seismic data noise removing method, a storage medium and electronic equipment.

Background

The Rayleigh wave is an elastic wave which propagates along a free surface and the amplitude of the elastic wave exponentially attenuates along the depth, and a surface wave multi-channel analysis Method (MASW) obtains a shallow earth surface shear wave velocity structure by inverting the Rayleigh wave phase velocity, and is widely applied to a series of shallow earth surface geophysical and geological problems.

Rayleigh surface wave exploration is a geophysical exploration method which is started in recent years, particularly in the aspect of engineering geophysical prospecting, the transverse wave velocity of a shallow surface underground medium can be accurately obtained by inverting a frequency dispersion curve of surface waves so as to obtain mechanical parameters of the shallow surface underground medium, and the Rayleigh surface wave exploration method is widely applied to multiple fields of roadbed detection, tunnel geological advanced prediction and the like. The field Rayleigh wave data acquisition is the first step of Rayleigh wave exploration, and actually acquired data often contain interference signals such as body waves or random environmental noise and the like, so that the mechanical parameters are inaccurate, and the denoising processing of the acquired data is very important.

The deep learning is used as a branch of machine learning and widely applied to the field of seismic exploration. The generation countermeasure network (GAN) is a new deep learning framework for estimating a generation model through the countermeasure network, a convolutional neural network is added on the basis, the advantages of the convolutional neural network and the convolutional neural network are combined, a deep convolution generation countermeasure network (DCGAN) is developed, and the generation countermeasure network (GAN) has the characteristics of higher resolution, more obvious generated sample characteristics and the like.

The existing Rayleigh wave signal denoising processing method comprises the following steps:

if a tau-p transformation method is used, the wave field separation is realized through the transformation of time domain signals to a tau-p domain, Rayleigh waves are highlighted, and various body waves and converted waves are eliminated; if an extended Prony method is used, Rayleigh waves are extracted according to the reasonability of the energy amplitude and the speed; if an f-k domain analysis method is adopted, a stronger energy group is reserved according to the energy characteristics of Rayleigh waves, and the rest energy group is removed as interference waves; if a wavelet analysis method is adopted, the Rayleigh wave with large energy is used for extracting a Rayleigh wave to calculate a frequency dispersion curve;

for example, a method for attenuating Rayleigh wave scattering noise in seismic data processing is characterized in that scattering noise attenuation processing is carried out on a frequency band mainly containing a scattering wave field, a linear wave field transformation, a wavelet transformation and an alpha-TRIMMED filtering technology are comprehensively utilized to extract a Rayleigh wave scattering incident wave field and calculate wavelets, then the optimal estimation of the scattering wave field is realized through a scattering wave field inversion technology, and the estimated scattering noise is subtracted from an original wave field, so that the signal-to-noise ratio of seismic data is effectively improved, the original wave field characteristics of other frequency bands are not influenced, and the denoising method is a relative fidelity denoising method;

the de-noising of Rayleigh wave signals is carried out on the basis of the multi-resolution characteristic of wavelet threshold analysis and the different characteristics and properties of the wavelet coefficients of useful signals and noise on the wavelet decomposition scale, and the de-noising effect of the wavelet threshold is obviously superior to the de-noising effect of repeated sampling and multiple superposition;

and denoising Rayleigh wave signals based on a wavelet analysis principle, and denoising by respectively using a wavelet threshold method and a wavelet modulus maximum method, wherein the wavelet threshold method can well remove high-frequency random noise, the wavelet modulus maximum method can effectively remove random noise, and the effect of filtering interference waves can be achieved according to the energy characteristics of Rayleigh waves.

The existing technology extracts the rayleigh wave according to the energy characteristics of the rayleigh wave and other interference waves, and the method only carries out denoising processing on single-channel data, which has certain application limitation. And when the interference wave energy is strong and is seriously coupled with the rayleigh wave, the effective signal loss of part of the rayleigh wave is often caused.

Most of the prior art combines a dispersion curve extraction method to remove interference waves, parameters need to be manually adjusted according to data characteristics in a data denoising process, manual participation is needed to perform data processing, and workload is increased.

Compared with forward modeling and inversion of Rayleigh waves, the research on interference wave removal and Rayleigh wave purification is relatively small at present, and a method for removing Rayleigh wave seismic data noise by generating a countermeasure network from a signal processing angle and combining depth convolution is needed, so that a Rayleigh wave dispersion curve with a better denoising effect is obtained, the transverse wave velocity is obtained by inversion, the mechanical parameters of an underground medium are obtained, and exploration personnel are helped to know the geophysical and geological conditions of a shallow earth surface.

Disclosure of Invention

The invention provides a Rayleigh wave seismic data noise removing method based on a deep convolution generation countermeasure network (DCGAN), which solves the technical problems of application limitation and manual parameter adjustment in the prior art, and can perform random noise removal on multi-channel surface wave data, so that the denoising result is more reliable.

The invention provides a method for removing Rayleigh wave seismic data noise, which comprises the following steps:

constructing a deep convolution generation countermeasure network, wherein the deep convolution generation countermeasure network comprises a generator and a discriminator;

preprocessing Rayleigh wave seismic data to obtain real noise-free data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

adding Gaussian white noise to the preprocessed noiseless data to obtain noised data;

adding the real noise-free data and the corresponding noise-containing data serving as training samples into the deep convolution to generate a confrontation network for training, performing iterative computation until the denoising precision meets a specified threshold or the iteration number meets a specified number, and then performing iterative computation to obtain a deep convolution generated confrontation network model;

and generating a confrontation network model based on the deep convolution to carry out denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised.

In an embodiment of the present invention, it is,

the method for constructing the deep convolution to generate the countermeasure network comprises constructing a discriminator and a generator, and establishing a first objective function of the discriminator and a second objective function of the generator.

In an embodiment of the present invention, it is,

the generator is composed of five convolutional layers, and the discriminator is composed of four convolutional layers and a full-connection layer.

In an embodiment of the present invention, it is,

the first objective function is calculated as:

wherein G is a generator, z is noisy Rayleigh data,

d is a discriminator, x is noiseless Rayleigh wave data,

θ^Dand theta^GRespectively representing the parameters to be optimized of the arbiter and the generator,

P_data(x) In order to be true of the noise-free data distribution,

P_z(z) is a noisy rayleigh wave data profile,

x～P_data(x) To obey the sampling under the distribution of the noise-free data,

z～P_z(z) is the sampling under the distribution of data subject to noise,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition,

d (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z.

The second objective function calculation formula is:

wherein D (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition.

In an embodiment of the present invention, it is,

the convolutional layers in the generator adopt a full convolutional network, the activation functions of 1 st to 4 th convolutional layers in the generator are ReLU functions, and the activation functions of 5 th convolutional layers are Tanh functions;

the discriminator adopts a step convolution layer, and the activation function of all layers in the discriminator is a LeakyReLU function.

In an embodiment of the present invention, it is,

the calculation formula adopted by the inter-channel equalization processing is as follows:

wherein, x'_iFor the data after the i-th track is equalized,

average value of absolute value of whole shot set data，

Taking the average value of the ith data after absolute value,

x_ithe ith trace of seismic data;

in an embodiment of the present invention, it is,

based on the data after the inter-channel equalization processing, performing energy normalization processing, wherein the energy normalization processing adopts a calculation formula as follows:

wherein, x ″)_i,jRepresents the value of the ith sampling point after normalization,

x′_i,jfor the value of the jth sampling point of the ith channel after the inter-channel equalization,

X′_minand X'_maxRespectively representing the minimum value and the maximum value of the gather after the equalization of the whole channel.

In an embodiment of the present invention, it is,

the step of adding the real noiseless data and the corresponding noisy data as training samples into the deep convolution to generate a countermeasure network for training comprises the following steps:

and adding the noisy data into the generator for training to obtain a generated noiseless data, adding the generated noiseless data and the real noiseless data which are simultaneously used as input data into the discriminator for training to obtain similarity degree values of 2 input data, and feeding back the similarity degree values to the generator to realize the mutual iterative optimization of the generator and the discriminator.

In an embodiment of the present invention, it is,

the mutual iterative optimization of the generator and the arbiter comprises the following steps:

fixing the noise-containing data and each layer of parameters of the generator, and performing iterative computation to optimize a first objective function of the discriminator until the discrimination accuracy of the discriminator reaches the maximum;

then fixing the real noiseless data and each layer parameter of the discriminator, and carrying out iterative computation to optimize a second objective function of the generator until the discrimination accuracy of the discriminator reaches the minimum;

and when the generated noiseless data is consistent with the real noiseless data, a global optimal solution is reached, and the mutual iterative optimization of the generator and the discriminator is exited.

In an embodiment of the present invention, it is,

the specified threshold is that the noise proportion is lower than 20%, and the specified times are more than or equal to 50.

In an embodiment of the present invention, it is,

upsampling using deconvolution in the generator;

constructing a nonlinear mapping from noisy data to residual data by adopting a residual learning algorithm in the generator;

before convolution operation is carried out on the convolution layers of the generator and the discriminator, the boundary is expanded for the input characteristic data, and zero values are filled until the sizes of the input characteristic data and the input data are consistent;

and adopting a batch normalization optimization algorithm and a discarding regularization layer in the first 4 convolutional layers of the generator and the discriminator.

The present invention also provides a storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the method for noise removal of rayleigh wave seismic data as described in any of the above.

The present invention also provides an electronic device, comprising: a memory having a computer program stored thereon; and a processor for executing the computer program in the memory to implement the steps of the rayleigh wave seismic data noise removal method of any of the above.

One or more embodiments of the present invention may have the following advantages over the prior art:

according to the method, a countermeasure network (DCGAN) is generated through deep convolution based on the deep learning field, and a part of noisy Rayleigh wave seismic data and noiseless Rayleigh wave seismic data are selected as samples and are respectively added into a generator and a discriminator for training. And applying the trained DCGAN to noise-containing Rayleigh wave seismic data in a test work area, and performing random noise removal on the multi-channel surface wave data to obtain high-quality denoised Rayleigh wave seismic data. After the DCGAN training is completed, parameters do not need to be adjusted manually, the labor cost is reduced, the denoising efficiency of the Rayleigh wave seismic data is improved, and the denoising effect of the Rayleigh wave data acquired under high background noise is better, so that the Rayleigh wave dispersion curve with better denoising effect is obtained, the transverse wave speed is obtained through inversion, the mechanical parameters of the underground medium are obtained, and exploration personnel are helped to know the geophysical and geological conditions of the shallow earth surface.

Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by the practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:

FIG. 1 is a schematic flow diagram of a method for denoising Rayleigh wave seismic data in accordance with an example of the invention;

FIG. 2 is a schematic diagram of an exemplary deep convolution generated countermeasure network architecture of the present invention;

FIG. 3 is Rayleigh wave seismic data with high random noise in accordance with an example of the present invention;

FIG. 4 illustrates DCGAN denoised Rayleigh wave seismic data in accordance with an embodiment of the present invention;

FIG. 5 is a graph of Rayleigh wave dispersion energy with high random noise according to an exemplary embodiment of the present invention;

FIG. 6 is a chart of the Rayleigh wave dispersion energy after de-noising by DCGAN according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the following detailed description of the present invention with reference to the accompanying drawings is provided to fully understand and implement the technical effects of the present invention by solving the technical problems through technical means. It should be noted that, as long as there is no conflict, the embodiments and the features of the embodiments of the present invention may be combined with each other, and the technical solutions formed are within the scope of the present invention.

First embodiment

Fig. 1 is a schematic flow chart of a rayleigh wave seismic data noise removal method according to the embodiment;

FIG. 2 is a schematic diagram of a deep convolution generation countermeasure network structure according to the present embodiment;

FIG. 3 is Rayleigh wave seismic data with high random noise for this embodiment;

FIG. 4 is the DCGAN denoised Rayleigh wave seismic data of the present embodiment;

FIG. 5 is a graph of Rayleigh wave dispersion energy with high random noise for the present embodiment;

fig. 6 is a graph of the rayleigh wave dispersion energy after being denoised by DCGAN according to the embodiment.

The embodiment provides a method for removing noise of Rayleigh wave seismic data, which comprises the following steps:

constructing a deep convolution generation countermeasure network, wherein the deep convolution generation countermeasure network comprises a generator and a discriminator;

preprocessing Rayleigh wave seismic data to obtain real noise-free data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

adding Gaussian white noise to the preprocessed noiseless data to obtain noised data;

adding the real noise-free data and the corresponding noise-containing data serving as training samples into the deep convolution to generate a confrontation network for training, performing iterative computation until the denoising precision meets a specified threshold or the iteration number meets a specified number, and then performing iterative computation to obtain a deep convolution generated confrontation network model;

generating a confrontation network model based on the deep convolution to carry out denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised;

and (4) obtaining the transverse wave speed and solving the mechanical parameters of the underground medium based on the denoised Rayleigh wave frequency dispersion curve inversion.

The basic idea of generating a countermeasure network (GAN) is derived from nash equilibrium theory in game theory, and a generative network model (Generator) and a discriminant network model (discriminant) in the GAN network can be considered as two parties participating in game play. The generator learns the real data distribution characteristics to generate generated data similar to the real data, with the ultimate purpose of enabling the generated data to deceive the discriminator; the final purpose of the discriminator is to correctly judge whether the input data is real data or generated data through learning; in order to win, both parties must continuously learn and optimize, improve the generation capability and discrimination capability of themselves, and finally achieve nash balance between the two.

In this embodiment, DCGAN is a deep convolution generation countermeasure network, GAN is a countermeasure network, FCN is a full convolution neural network, CNN is a convolution neural network, G is a generator, D is a discriminator, P is a convolutional code, and_Gto generate noise-free data.

Specifically, the steps of generating the countermeasure network DCGAN based on the deep convolution and performing random noise removal on the rayleigh wave signal are as follows:

and step 100, constructing a deep convolution generation countermeasure network DCGAN, wherein the deep convolution generation countermeasure network DCGAN comprises a generator G and a discriminator D.

In the network construction, any differentiable function can be used to represent the generator and the arbiter of GAN, and differentiable functions D and G are used to represent the arbiter and the generator, respectively. And constructing the deep convolution to generate the countermeasure network DCGAN, wherein the construction of the discriminator D and the generator G is carried out, and a first objective function of the discriminator and a second objective function of the generator are established. The generator G is composed of a full convolution neural network (FCN) and used for learning feature mapping from noisy data to noiseless data, the generator G is composed of five convolution layers, the discriminator D is composed of four convolution layers and a full connection layer, and the discriminator D is composed of a Convolution Neural Network (CNN) and used for assisting the training of the generator G. The structure of the deep convolution generation countermeasure network DCGAN is shown in fig. 2, where the name of each convolution layer in the diagram is k, where k denotes the size of a convolution kernel, i.e., k3, i.e., the size of the convolution kernel is 3 × 3, n denotes the number of feature maps, i.e., n64 denotes 64 feature maps obtained after the convolution operation is completed, s denotes the step size of the convolution operation, i.e., s2 bits of the convolution operation is 2.

The purpose of the generator is to learn real data distribution as much as possible and perform denoising to obtain generated noise-free data, and the purpose of the discriminator is to judge the similarity degree of the real noise-free data in the input data and the generated noise-free data of the generator as correctly as possible.

In the embodiment, on the basis of generating the countermeasure network GAN, the convolutional neural network is introduced, and the effect of GAN is improved by utilizing the strong feature extraction capability of the convolutional layer. To improve convergence speed and sample quality, the DCGAN makes the following changes and settings to the convolutional neural network of generator G and discriminator D:

(1) canceling a pooling layer of an original convolutional neural network, performing up-sampling by using deconvolution in a generator G, and replacing the pooling layer with a stride convolutional layer (stranded constants) in a discriminator D;

(2) removing a full connection layer of the original convolutional neural network from a generator G to obtain a Full Convolutional Network (FCN), wherein convolutional layers are adopted;

(3) in the generator G network, a ReLU function is adopted as an activation function of 1 st to 4 th convolutional layers, and a Tanh function is adopted as an activation function of a 5 th convolutional layer;

(4) the LeakyReLU function is used for all layer activation functions in the arbiter D network.

In the denoising of the rayleigh wave signal in this embodiment, the generator is G, and the input data thereof is noisy rayleigh wave data z; the discriminator is D, and the input data is noiseless Rayleigh wave data x; during actual training, a training noiseless Rayleigh wave data set of the discriminator comes from two parts: preprocessed true noise-free data set P from an actual survey_data(x) (labeled 1) Generation of noiseless dataset P with Generator_G(x) (labeled 0).

For arbiter D, when generator G is fixed, the first objective function of trained arbiter D is calculated as:

wherein G is a generator, z is noisy Rayleigh data,

d is a discriminator, x is noiseless Rayleigh wave data,

θ^Dand theta^GRespectively representing the parameters to be optimized of the arbiter and the generator,

P_data(x) In order to be true of the noise-free data distribution,

P_z(z) is a noisy rayleigh wave data profile,

x～P_data(x) To obey the sampling under the distribution of the noise-free data,

z～P_z(z) is the sampling under the distribution of data subject to noise,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition,

d (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z.

When the input data is from a true noise-free data set P_data(x) In this case, the goal of the discriminator D is to approximate the output D (x) to 1, and when the input data is from the generated noise-free data set G (z), the goal of the discriminator D is to approximate the output D (G (z)) to 0, while the goal of the generator G is to approximate D (G (z)) to 1.

Thus, the process of generating the antagonistic network training is a minimization-maximization problem, and the generator G calculates the second objective function as:

wherein D (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition.

Step 200, preprocessing Rayleigh wave seismic data to obtain real noiseless data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

because the difference of the seismic recording energy between different tracks and between different cannons of Rayleigh surface wave data is large and sometimes differs by several orders of magnitude, if the Rayleigh surface wave data directly participates in training, the training process is unstable or not converged, and therefore preprocessing is needed.

Selecting a part of seismic data with high signal-to-noise ratio in a Rayleigh wave exploration work area as noise-free data, and preprocessing the part of data according to the following calculation formula, namely performing interchannel equalization and energy normalization processing.

The calculation formula adopted by the inter-channel equalization processing is as follows:

wherein, x'_iFor the data after the i-th track is equalized,

to be integratedThe individual shot gather data is averaged after taking the absolute value,

taking the average value of the ith data after absolute value,

x_ithe ith trace of seismic data;

and then, based on the data after the inter-channel equalization processing, performing energy normalization processing to obtain real noiseless data.

The energy normalization process adopts the following calculation formula:

wherein, x ″)_i,jRepresents the value of the ith sampling point after normalization,

x′_i,jfor the value of the jth sampling point of the ith channel after the inter-channel equalization,

X′_minand X'_maxRespectively representing the minimum value and the maximum value of the gather after the equalization of the whole channel.

Step 300, adding Gaussian white noise based on the preprocessed real noiseless data to obtain noisy data;

specifically, Gaussian white noise with the average value of 0 and the amplitude of different proportions is artificially added to each preprocessed single-shot real noise-free data to serve as noise-containing data containing random noise.

Step 400, adding real noise-free data and corresponding noise-containing data serving as training samples into a deep convolution to generate a confrontation network DCGAN for training, performing iterative computation until denoising precision meets a specified threshold or iteration times meet specified times, and then performing iterative computation to obtain a confrontation network model generated by deep convolution;

the training process of DCGAN is as follows: the method comprises the steps of adding noisy data into a generator to be trained and denoised, obtaining generated noiseless data through a series of convolution operations of convolution layers, simultaneously using the generated noiseless data and real noiseless data as input data to be trained by a discriminator to obtain similarity degree values of 2 input data, and feeding the similarity degree values back to the generator to achieve mutual iterative optimization of the generator and the discriminator.

The noisy data input by the generator can be regarded as a mxnx1 vector, the mxnx256 vector is generated after the first convolution layer in the generator, the mxnx128 vector is generated after the second convolution layer, and the like, the 250 × 90 × 1 vector is output after the last convolution layer, and the noiseless data is generated.

The input of the discriminator is an mxnx1 vector representing noiseless data, an mxnx64 vector is generated after the first convolution layer, and so on, and finally a real number between 0 and 1 is output through the full connection layer.

In this embodiment, the mutual iterative optimization of the generator and the discriminator specifically includes:

firstly, fixing the noisy data and parameters of each layer of the generator G, and performing iterative calculation to optimize a first objective function of the discriminator until the discrimination accuracy of the discriminator reaches the maximum;

then fixing the real noise-free data and each layer parameter of the discriminator, and carrying out iterative computation to optimize a second objective function of the generator until the discrimination accuracy of the discriminator reaches the minimum;

when generating the noise-free data P_GWith true noise-free data P_dataWhen coincident, i.e. P_data＝P_GAnd then, the global optimal solution is reached, and the mutual iterative optimization of the generator and the discriminator is exited.

In the same round of parameter updating of training, the parameters of the generator G are updated multiple times and then the parameters of the discriminator D are updated once.

In summary, the process of generating the training learning against the net not only trains arbiter D to maximize the accuracy of the data sources, but also trains generator G to minimize log (1-D (G (z))).

In this embodiment, in practical application, sometimes the optimal solution may not be achieved, so that according to the actual denoising precision requirement of exploration, or the training frequency reaches the specified frequency, the DCGAN training is considered to be completed, and the mutual iterative optimization of the generator and the discriminator may be exited.

500, generating a confrontation network model based on the depth convolution to perform denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised;

and when the DCGAN training is finished, inputting the Rayleigh wave signal seismic data to be denoised into the DCGAN training, namely outputting denoised data according to the DCGAN network, and finishing the random noise removal in the Rayleigh wave signal seismic data.

According to the method, after the DCGAN training is completed, parameters do not need to be adjusted manually, so that the labor cost is reduced, and the denoising efficiency of Rayleigh wave seismic data is improved.

And 600, obtaining the transverse wave speed and solving the mechanical parameters of the underground medium based on the denoised Rayleigh wave frequency dispersion curve inversion.

According to the steps of the Rayleigh wave seismic data noise removal method, the measured data of Rayleigh wave exploration in a certain work area is selected for denoising, and the denoising performance of the DCGAN method on random noise is tested.

The selected Rayleigh wave seismic data has 24 channels per shot, and each channel has 400 sampling points (2ms sampling, total time length 800 ms). Selecting 300 shots of data with high signal-to-noise ratio in a work area, performing interchannel equalization and energy normalization processing to obtain real noise-free data, artificially adding Gaussian white noise with the average value of 0 and the amplitude of 10% -30% to each shot to obtain noise-containing data with different intensity noises, training by using the DCGAN of the embodiment, and obtaining a Rayleigh wave signal denoising DCGAN model aiming at random noise through 100 epochs.

Note: epoch, when a complete data set passes through the neural network once and back once, this process is called once.

In the work area, random noise data of one shot alternative in the work area is preprocessed and then used as a test shot for carrying out denoising effect analysis, fig. 3 shows the Rayleigh wave seismic record data of the test shot containing high random noise, denoising is carried out by using trained DCGAN, and the denoising result is shown in fig. 4. From the rayleigh wave dispersion curve, fig. 5 is a dispersion energy diagram of a noisy experimental cannon, and fig. 6 is a dispersion energy diagram of denoised data. It can be seen that after the noise removal method for the rayleigh wave seismic data is used for denoising, random noise on a seismic record is well suppressed, and a dispersion curve is closer to a real dispersion characteristic, so that the transverse wave velocity can be obtained and the mechanical parameters of an underground medium can be solved based on re-inversion of the denoised rayleigh wave dispersion curve, and the method is also proved to have a better denoising effect on the rayleigh wave data acquired under high background noise.

In summary, the method generates the countermeasure network (DCGAN) by deep convolution in the deep learning field, and selects a part of noisy rayleigh seismic data and noiseless rayleigh seismic data as samples to be respectively added into the generator and the discriminator for training. And applying the trained DCGAN to noise-containing Rayleigh wave seismic data in a test work area, and performing random noise removal on the multi-channel surface wave data to obtain high-quality denoised Rayleigh wave seismic data. After the DCGAN training is completed, parameters do not need to be adjusted manually, the labor cost is reduced, the denoising efficiency of the Rayleigh wave seismic data is improved, and the denoising effect of the Rayleigh wave data acquired under high background noise is better, so that the Rayleigh wave dispersion curve with better denoising effect is obtained, the transverse wave speed is obtained through inversion, the mechanical parameters of the underground medium are obtained, and exploration personnel are helped to know the geophysical and geological conditions of the shallow earth surface.

Second embodiment

Fig. 1 is a schematic flow chart of a rayleigh wave seismic data noise removal method according to the embodiment;

FIG. 2 is a schematic diagram of a deep convolution generation countermeasure network structure according to the present embodiment;

FIG. 3 is Rayleigh wave seismic data with high random noise for this embodiment;

FIG. 4 is the DCGAN denoised Rayleigh wave seismic data of the present embodiment;

FIG. 5 is a graph of Rayleigh wave dispersion energy with high random noise for the present embodiment;

fig. 6 is a graph of the rayleigh wave dispersion energy after being denoised by DCGAN according to the embodiment.

The embodiment provides a method for removing noise of Rayleigh wave seismic data, which comprises the following steps:

constructing a deep convolution generation countermeasure network, wherein the deep convolution generation countermeasure network comprises a generator and a discriminator;

preprocessing Rayleigh wave seismic data to obtain real noise-free data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

adding Gaussian white noise to the preprocessed noiseless data to obtain noised data;

adding the real noise-free data and the corresponding noise-containing data serving as training samples into the deep convolution to generate a confrontation network for training, performing iterative computation until the denoising precision meets a specified threshold or the iteration number meets a specified number, and then performing iterative computation to obtain a deep convolution generated confrontation network model;

generating a confrontation network model based on the deep convolution to carry out denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised;

and (4) obtaining the transverse wave speed and solving the mechanical parameters of the underground medium based on the denoised Rayleigh wave frequency dispersion curve inversion.

The basic idea of generating a countermeasure network (GAN) is derived from nash equilibrium theory in game theory, and a generative network model (Generator) and a discriminant network model (discriminant) in the GAN network can be considered as two parties participating in game play. The generator learns the real data distribution characteristics to generate generated data similar to the real data, with the ultimate purpose of enabling the generated data to deceive the discriminator; the final purpose of the discriminator is to correctly judge whether the input data is real data or generated data through learning; in order to win, both parties must continuously learn and optimize, improve the generation capability and discrimination capability of themselves, and finally achieve nash balance between the two.

In this embodiment, DCGAN is a deep convolution generation countermeasure network, GAN is a countermeasure network, FCN is a full convolution neural network, CNN is a convolution neural network, G is a generator, D is a discriminator, P is a convolutional code, and_Gto generate noise-free data.

Specifically, the steps of generating the countermeasure network DCGAN based on the deep convolution and performing random noise removal on the rayleigh wave signal are as follows:

and step 100, constructing a deep convolution generation countermeasure network DCGAN, wherein the deep convolution generation countermeasure network DCGAN comprises a generator G and a discriminator D.

In the network construction, any differentiable function can be used to represent the generator and the arbiter of GAN, and differentiable functions D and G are used to represent the arbiter and the generator, respectively. And constructing the deep convolution to generate the countermeasure network DCGAN, wherein the construction of the discriminator D and the generator G is carried out, and a first objective function of the discriminator and a second objective function of the generator are established. The generator G is composed of a full convolution neural network (FCN) and used for learning feature mapping from noisy data to noiseless data, the generator G is composed of five convolution layers, the discriminator D is composed of four convolution layers and a full connection layer, and the discriminator D is composed of a Convolution Neural Network (CNN) and used for assisting the training of the generator G. The structure of the deep convolution generation countermeasure network DCGAN is shown in fig. 2, where the name of each convolution layer in the diagram is k, where k denotes the size of a convolution kernel, i.e., k3, i.e., the size of the convolution kernel is 3 × 3, n denotes the number of feature maps, i.e., n64 denotes 64 feature maps obtained after the convolution operation is completed, s denotes the step size of the convolution operation, i.e., s2 bits of the convolution operation is 2.

The purpose of the generator is to learn real data distribution as much as possible and perform denoising to obtain generated noise-free data, and the purpose of the discriminator is to judge the similarity degree of the real noise-free data in the input data and the generated noise-free data of the generator as correctly as possible.

In the embodiment, on the basis of generating the countermeasure network GAN, the convolutional neural network is introduced, and the effect of GAN is improved by utilizing the strong feature extraction capability of the convolutional layer. To improve convergence speed and sample quality, the DCGAN makes the following changes and settings to the convolutional neural network of generator G and discriminator D:

(1) canceling a pooling layer of an original convolutional neural network, performing up-sampling by using deconvolution in a generator G, and replacing the pooling layer with a stride convolutional layer (stranded constants) in a discriminator D;

(2) removing a full connection layer of the original convolutional neural network from a generator G to obtain a Full Convolutional Network (FCN), wherein convolutional layers are adopted;

(3) in the generator G network, a ReLU function is adopted as an activation function of 1 st to 4 th convolutional layers, and a Tanh function is adopted as an activation function of a 5 th convolutional layer;

(4) the LeakyReLU function is used for all layer activation functions in the arbiter D network.

In the denoising of the rayleigh wave signal in this embodiment, the generator is G, and the input data thereof is noisy rayleigh wave data z; the discriminator is D, and the input data is noiseless Rayleigh wave data x; during actual training, a training noiseless Rayleigh wave data set of the discriminator comes from two parts: preprocessed true noise-free data set P from an actual survey_data(x) (labeled 1) Generation of noiseless dataset P with Generator_G(x) (labeled 0).

For arbiter D, when generator G is fixed, the first objective function of trained arbiter D is calculated as:

wherein G is a generator, z is noisy Rayleigh data,

d is a discriminator, x is noiseless Rayleigh wave data,

θ^Dand theta^GRespectively representing the parameters to be optimized of the arbiter and the generator,

P_data(x) In order to be true of the noise-free data distribution,

P_z(z) is a noisy rayleigh wave data profile,

x～P_data(x) To obey the sampling under the distribution of the noise-free data,

z～P_z(z) is the sampling under the distribution of data subject to noise,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition,

d (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z.

When the input data is from a true noise-free data set P_data(x) In this case, the goal of the discriminator D is to approximate the output D (x) to 1, and when the input data is from the generated noise-free data set G (z), the goal of the discriminator D is to approximate the output D (G (z)) to 0, while the goal of the generator G is to approximate D (G (z)) to 1.

Thus, the process of generating the antagonistic network training is a minimization-maximization problem, and the generator G calculates the second objective function as:

wherein D (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition.

Step 200, preprocessing Rayleigh wave seismic data to obtain real noiseless data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

because the difference of the seismic recording energy between different tracks and between different cannons of Rayleigh surface wave data is large and sometimes differs by several orders of magnitude, if the Rayleigh surface wave data directly participates in training, the training process is unstable or not converged, and therefore preprocessing is needed.

Selecting a part of seismic data with high signal-to-noise ratio in a Rayleigh wave exploration work area as noise-free data, and preprocessing the part of data according to the following calculation formula, namely performing interchannel equalization and energy normalization processing.

The calculation formula adopted by the inter-channel equalization processing is as follows:

wherein, x'_iFor the data after the i-th track is equalized,

taking the average value of the absolute value of the whole shot gather data,

taking the average value of the ith data after absolute value,

x_ithe ith trace of seismic data;

and then, based on the data after the inter-channel equalization processing, performing energy normalization processing to obtain real noiseless data.

The energy normalization process adopts the following calculation formula:

wherein, x ″)_i,jRepresents the value of the ith sampling point after normalization,

x′_i,jfor the value of the jth sampling point of the ith channel after the inter-channel equalization,

X′_minand X'_maxRespectively representing the minimum value and the maximum value of the gather after the equalization of the whole channel.

Step 300, adding Gaussian white noise based on the preprocessed real noiseless data to obtain noisy data;

specifically, Gaussian white noise with the average value of 0 and the amplitude of different proportions is artificially added to each preprocessed single-shot real noise-free data to serve as noise-containing data containing random noise.

Step 400, adding real noise-free data and corresponding noise-containing data serving as training samples into a deep convolution to generate a confrontation network DCGAN for training, performing iterative computation until denoising precision meets a specified threshold or iteration times meet specified times, and then performing iterative computation to obtain a confrontation network model generated by deep convolution;

the training process of DCGAN is as follows: the method comprises the steps of adding noisy data into a generator to be trained and denoised, obtaining generated noiseless data through a series of convolution operations of convolution layers, simultaneously using the generated noiseless data and real noiseless data as input data to be trained by a discriminator to obtain similarity degree values of 2 input data, and feeding the similarity degree values back to the generator to achieve mutual iterative optimization of the generator and the discriminator.

The noisy data input by the generator can be regarded as a mxnx1 vector, the mxnx256 vector is generated after the first convolution layer in the generator, the mxnx128 vector is generated after the second convolution layer, and the like, the 250 × 90 × 1 vector is output after the last convolution layer, and the noiseless data is generated.

The input of the discriminator is an mxnx1 vector representing noiseless data, an mxnx64 vector is generated after the first convolution layer, and so on, and finally a real number between 0 and 1 is output through the full connection layer.

In this embodiment, the mutual iterative optimization of the generator and the discriminator specifically includes:

firstly, fixing the noisy data and parameters of each layer of the generator G, and performing iterative calculation to optimize a first objective function of the discriminator until the discrimination accuracy of the discriminator reaches the maximum;

then fixing the real noise-free data and each layer parameter of the discriminator, and carrying out iterative computation to optimize a second objective function of the generator until the discrimination accuracy of the discriminator reaches the minimum;

when generating the noise-free data P_GWith true noise-free data P_dataWhen coincident, i.e. P_data＝P_GAnd then, the global optimal solution is reached, and the mutual iterative optimization of the generator and the discriminator is exited.

In the same round of parameter updating of training, the parameters of the generator G are updated multiple times and then the parameters of the discriminator D are updated once.

In summary, the process of generating the training learning against the net not only trains arbiter D to maximize the accuracy of the data sources, but also trains generator G to minimize log (1-D (G (z))).

In this embodiment, in practical application, sometimes the optimal solution may not be achieved, so according to the actual denoising precision requirement of exploration, if the denoising precision specified threshold value meets the requirement that the noise ratio is lower than 20%, or the number of times of training reaches the specified number of times, if the specified number of times is greater than or equal to 50, the DCGAN training is considered to be completed, and the mutual iterative optimization of the generator and the discriminator may be exited.

500, generating a confrontation network model based on the depth convolution to perform denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised;

and when the DCGAN training is finished, inputting the Rayleigh wave signal seismic data to be denoised into the DCGAN training, namely outputting denoised data according to the DCGAN network, and finishing the random noise removal in the Rayleigh wave signal seismic data.

According to the method, after the DCGAN training is completed, parameters do not need to be adjusted manually, so that the labor cost is reduced, and the denoising efficiency of Rayleigh wave seismic data is improved.

And 600, obtaining the transverse wave speed and solving the mechanical parameters of the underground medium based on the denoised Rayleigh wave frequency dispersion curve inversion.

According to the steps of the Rayleigh wave seismic data noise removal method, the measured data of Rayleigh wave exploration in a certain work area is selected for denoising, and the denoising performance of the DCGAN method on random noise is tested.

The selected Rayleigh wave seismic data has 24 channels per shot, and each channel has 400 sampling points (2ms sampling, total time length 800 ms). Selecting 300 shots of data with high signal-to-noise ratio in a work area, performing interchannel equalization and energy normalization processing to obtain real noise-free data, artificially adding Gaussian white noise with the average value of 0 and the amplitude of 10% -30% to each shot to obtain noise-containing data with different intensity noises, training by using the DCGAN of the embodiment, and obtaining a Rayleigh wave signal denoising DCGAN model aiming at random noise through 100 epochs.

Note: epoch, when a complete data set passes through the neural network once and back once, this process is called once.

In the work area, random noise data of one shot alternative in the work area is preprocessed and then used as a test shot for carrying out denoising effect analysis, fig. 3 shows the Rayleigh wave seismic record data of the test shot containing high random noise, denoising is carried out by using trained DCGAN, and the denoising result is shown in fig. 4. From the rayleigh wave dispersion curve, fig. 5 is a dispersion energy diagram of a noisy experimental cannon, and fig. 6 is a dispersion energy diagram of denoised data. It can be seen that after the noise removal method for the rayleigh wave seismic data is used for denoising, random noise on a seismic record is well suppressed, and a dispersion curve is closer to a real dispersion characteristic, so that the transverse wave velocity can be obtained and the mechanical parameters of an underground medium can be solved based on re-inversion of the denoised rayleigh wave dispersion curve, and the method is also proved to have a better denoising effect on the rayleigh wave data acquired under high background noise.

Compared with the first embodiment, the embodiment defines the denoising precision requirement, sets the specified threshold, defines the specified times of training, sets the specified times requirement, and can obtain high-quality denoised rayleigh wave seismic data.

In summary, the method generates the countermeasure network (DCGAN) by deep convolution in the deep learning field, and selects a part of noisy rayleigh seismic data and noiseless rayleigh seismic data as samples to be respectively added into the generator and the discriminator for training. And applying the trained DCGAN to noise-containing Rayleigh wave seismic data in a test work area, and performing random noise removal on the multi-channel surface wave data to obtain high-quality denoised Rayleigh wave seismic data. After the DCGAN training is completed, parameters do not need to be adjusted manually, the labor cost is reduced, the denoising efficiency of the Rayleigh wave seismic data is improved, and the denoising effect of the Rayleigh wave data acquired under high background noise is better, so that the Rayleigh wave dispersion curve with better denoising effect is obtained, the transverse wave speed is obtained through inversion, the mechanical parameters of the underground medium are obtained, and exploration personnel are helped to know the geophysical and geological conditions of the shallow earth surface.

Third embodiment

Fig. 1 is a schematic flow chart of a rayleigh wave seismic data noise removal method according to the embodiment;

FIG. 2 is a schematic diagram of a deep convolution generation countermeasure network structure according to the present embodiment;

FIG. 3 is Rayleigh wave seismic data with high random noise for this embodiment;

FIG. 4 is the DCGAN denoised Rayleigh wave seismic data of the present embodiment;

FIG. 5 is a graph of Rayleigh wave dispersion energy with high random noise for the present embodiment;

fig. 6 is a graph of the rayleigh wave dispersion energy after being denoised by DCGAN according to the embodiment.

The embodiment provides a method for removing noise of Rayleigh wave seismic data, which comprises the following steps:

constructing a deep convolution generation countermeasure network, wherein the deep convolution generation countermeasure network comprises a generator and a discriminator;

preprocessing Rayleigh wave seismic data to obtain real noise-free data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

adding Gaussian white noise to the preprocessed noiseless data to obtain noised data;

adding the real noise-free data and the corresponding noise-containing data serving as training samples into the deep convolution to generate a confrontation network for training, performing iterative computation until the denoising precision meets a specified threshold or the iteration number meets a specified number, and then performing iterative computation to obtain a deep convolution generated confrontation network model;

generating a confrontation network model based on the deep convolution to carry out denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised;

and (4) obtaining the transverse wave speed and solving the mechanical parameters of the underground medium based on the denoised Rayleigh wave frequency dispersion curve inversion.

The basic idea of generating a countermeasure network (GAN) is derived from nash equilibrium theory in game theory, and a generative network model (Generator) and a discriminant network model (discriminant) in the GAN network can be considered as two parties participating in game play. The generator learns the real data distribution characteristics to generate generated data similar to the real data, with the ultimate purpose of enabling the generated data to deceive the discriminator; the final purpose of the discriminator is to correctly judge whether the input data is real data or generated data through learning; in order to win, both parties must continuously learn and optimize, improve the generation capability and discrimination capability of themselves, and finally achieve nash balance between the two.

In this embodiment, DCGAN is a deep convolution generation countermeasure network, GAN is a countermeasure network, FCN is a full convolution neural network, CNN is a convolution neural network, G is a generator, D is a discriminator, P is a convolutional code, and_Gto generate noise-free data.

Specifically, the steps of generating the countermeasure network DCGAN based on the deep convolution and performing random noise removal on the rayleigh wave signal are as follows:

and step 100, constructing a deep convolution generation countermeasure network DCGAN, wherein the deep convolution generation countermeasure network DCGAN comprises a generator G and a discriminator D.

In the network construction, any differentiable function can be used to represent the generator and the arbiter of GAN, and differentiable functions D and G are used to represent the arbiter and the generator, respectively. And constructing the deep convolution to generate the countermeasure network DCGAN, wherein the construction of the discriminator D and the generator G is carried out, and a first objective function of the discriminator and a second objective function of the generator are established. The generator G is composed of a full convolution neural network (FCN) and used for learning feature mapping from noisy data to noiseless data, the generator G is composed of five convolution layers, the discriminator D is composed of four convolution layers and a full connection layer, and the discriminator D is composed of a Convolution Neural Network (CNN) and used for assisting the training of the generator G. The structure of the deep convolution generation countermeasure network DCGAN is shown in fig. 2, where the name of each convolution layer in the diagram is k, where k denotes the size of a convolution kernel, i.e., k3, i.e., the size of the convolution kernel is 3 × 3, n denotes the number of feature maps, i.e., n64 denotes 64 feature maps obtained after the convolution operation is completed, s denotes the step size of the convolution operation, i.e., s2 bits of the convolution operation is 2.

The purpose of the generator is to learn real data distribution as much as possible and perform denoising to obtain generated noise-free data, and the purpose of the discriminator is to judge the similarity degree of the real noise-free data in the input data and the generated noise-free data of the generator as correctly as possible.

In the embodiment, on the basis of generating the countermeasure network GAN, the convolutional neural network is introduced, and the effect of GAN is improved by utilizing the strong feature extraction capability of the convolutional layer. To improve convergence speed and sample quality, the DCGAN makes the following changes and settings to the convolutional neural network of generator G and discriminator D:

(1) canceling a pooling layer of an original convolutional neural network, performing up-sampling by using deconvolution in a generator G, and replacing the pooling layer with a stride convolutional layer (stranded constants) in a discriminator D;

(2) removing a full connection layer of the original convolutional neural network from a generator G to obtain a Full Convolutional Network (FCN), wherein convolutional layers are adopted;

(3) in the generator G network, a ReLU function is adopted as an activation function of 1 st to 4 th convolutional layers, and a Tanh function is adopted as an activation function of a 5 th convolutional layer;

(4) the LeakyReLU function is used for all layer activation functions in the arbiter D network.

In addition, in the embodiment, a residual learning algorithm is also adopted in the generator G, and a non-linear mapping from noisy data to residual data is constructed. The residual error network has good characteristic extraction capability in the network training process, the advantages of the residual error network are utilized, the complex relation of waveform characteristics among actual seismic traces is considered, the residual error network is applied to the generator G, a nonlinear mapping from noisy data to residual error data is constructed, and the error range is convenient to optimize and control.

In this embodiment, data volume optimization is also performed, and before convolution operations are performed on the convolution layers of the generator G and the discriminator D, a boundary is extended for the input feature data, and zero padding is performed until the size of the boundary is consistent with that of the input data, so that the size of the output data is consistent with that of the input data, and meanwhile, an artifact phenomenon of the boundary data is avoided.

In this embodiment, algorithm optimization is also performed, and a Batch Normalization optimization algorithm and a discard regularization layer (dropout) are used in the first 4 convolutional layers of the generator G and the discriminator D. The batch normalization optimization algorithm can normalize the input of each layer, so that the mean value and the variance of each layer are 0 and 1, the data are more concentrated without worrying about too small or too large data, the training problem caused by poor initialization is favorably solved, and the stability of the network is improved; discarding the regularization layer can discard some characteristics at random after each convolution layer outputs, avoid the whole network from deviating to a certain characteristic, and avoid the overfitting problem to a certain extent.

In the denoising of the rayleigh wave signal in this embodiment, the generator is G, and the input data thereof is noisy rayleigh wave data z; the discriminator is D, and the input data is noiseless Rayleigh wave data x; during actual training, a training noiseless Rayleigh wave data set of the discriminator comes from two parts: preprocessed true noise-free data set P from an actual survey_data(x) (labeled 1) Generation of noiseless dataset P with Generator_G(x) (labeled 0).

For arbiter D, when generator G is fixed, the first objective function of trained arbiter D is calculated as:

wherein G is a generator, z is noisy Rayleigh data,

d is a discriminator, x is noiseless Rayleigh wave data,

θ^Dand theta^GRespectively representing the parameters to be optimized of the arbiter and the generator,

P_data(x) In order to be true of the noise-free data distribution,

P_z(z) is a noisy rayleigh wave data profile,

x～P_data(x) To obey the sampling under the distribution of the noise-free data,

z～P_z(z) is the sampling under the distribution of data subject to noise,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition,

d (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z.

When the input data is from a true noise-free data set P_data(x) In this case, the goal of the discriminator D is to approximate the output D (x) to 1, and when the input data is from the generated noise-free data set G (z), the goal of the discriminator D is to approximate the output D (G (z)) to 0, while the goal of the generator G is to approximate D (G (z)) to 1.

Thus, the process of generating the antagonistic network training is a minimization-maximization problem, and the generator G calculates the second objective function as:

wherein D (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition.

Step 200, preprocessing Rayleigh wave seismic data to obtain real noiseless data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

because the difference of the seismic recording energy between different tracks and between different cannons of Rayleigh surface wave data is large and sometimes differs by several orders of magnitude, if the Rayleigh surface wave data directly participates in training, the training process is unstable or not converged, and therefore preprocessing is needed.

Selecting a part of seismic data with high signal-to-noise ratio in a Rayleigh wave exploration work area as noise-free data, and preprocessing the part of data according to the following calculation formula, namely performing interchannel equalization and energy normalization processing.

The calculation formula adopted by the inter-channel equalization processing is as follows:

wherein, x'_iFor the data after the i-th track is equalized,

taking the average value of the absolute value of the whole shot gather data,

taking the average value of the ith data after absolute value,

x_ithe ith trace of seismic data;

and then, based on the data after the inter-channel equalization processing, performing energy normalization processing to obtain real noiseless data.

The energy normalization process adopts the following calculation formula:

wherein, x ″)_i,jRepresents the value of the ith sampling point after normalization,

x′_i,jfor the value of the jth sampling point of the ith channel after the inter-channel equalization,

X′_minand X'_maxRespectively representing the minimum value and the maximum value of the gather after the equalization of the whole channel.

Step 300, adding Gaussian white noise based on the preprocessed real noiseless data to obtain noisy data;

specifically, Gaussian white noise with the average value of 0 and the amplitude of different proportions is artificially added to each preprocessed single-shot real noise-free data to serve as noise-containing data containing random noise.

Step 400, adding real noise-free data and corresponding noise-containing data serving as training samples into a deep convolution to generate a confrontation network DCGAN for training, performing iterative computation until denoising precision meets a specified threshold or iteration times meet specified times, and then performing iterative computation to obtain a confrontation network model generated by deep convolution;

the training process of DCGAN is as follows: the method comprises the steps of adding noisy data into a generator to be trained and denoised, obtaining generated noiseless data through a series of convolution operations of convolution layers, simultaneously using the generated noiseless data and real noiseless data as input data to be trained by a discriminator to obtain similarity degree values of 2 input data, and feeding the similarity degree values back to the generator to achieve mutual iterative optimization of the generator and the discriminator.

The noisy data input by the generator can be regarded as a mxnx1 vector, the mxnx256 vector is generated after the first convolution layer in the generator, the mxnx128 vector is generated after the second convolution layer, and the like, the 250 × 90 × 1 vector is output after the last convolution layer, and the noiseless data is generated.

The input of the discriminator is an mxnx1 vector representing noiseless data, an mxnx64 vector is generated after the first convolution layer, and so on, and finally a real number between 0 and 1 is output through the full connection layer.

In this embodiment, the mutual iterative optimization of the generator and the discriminator specifically includes:

firstly, fixing the noisy data and parameters of each layer of the generator G, and performing iterative calculation to optimize a first objective function of the discriminator until the discrimination accuracy of the discriminator reaches the maximum;

then fixing the real noise-free data and each layer parameter of the discriminator, and carrying out iterative computation to optimize a second objective function of the generator until the discrimination accuracy of the discriminator reaches the minimum;

when generating the noise-free data P_GWith true noise-free data P_dataWhen coincident, i.e. P_data＝P_GAnd then, the global optimal solution is reached, and the mutual iterative optimization of the generator and the discriminator is exited.

In the same round of parameter updating of training, the parameters of the generator G are updated multiple times and then the parameters of the discriminator D are updated once.

In summary, the process of generating the training learning against the net not only trains arbiter D to maximize the accuracy of the data sources, but also trains generator G to minimize log (1-D (G (z))).

In this embodiment, in practical application, sometimes the optimal solution may not be achieved, so according to the actual denoising precision requirement of exploration, if the denoising precision specified threshold value meets the requirement that the noise ratio is lower than 20%, or the number of times of training reaches the specified number of times, if the specified number of times is greater than or equal to 50, the DCGAN training is considered to be completed, and the mutual iterative optimization of the generator and the discriminator may be exited.

500, generating a confrontation network model based on the depth convolution to perform denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised;

and when the DCGAN training is finished, inputting the Rayleigh wave signal seismic data to be denoised into the DCGAN training, namely outputting denoised data according to the DCGAN network, and finishing the random noise removal in the Rayleigh wave signal seismic data.

According to the method, after the DCGAN training is completed, parameters do not need to be adjusted manually, so that the labor cost is reduced, and the denoising efficiency of Rayleigh wave seismic data is improved.

And 600, obtaining the transverse wave speed and solving the mechanical parameters of the underground medium based on the denoised Rayleigh wave frequency dispersion curve inversion.

According to the steps of the Rayleigh wave seismic data noise removal method, the measured data of Rayleigh wave exploration in a certain work area is selected for denoising, and the denoising performance of the DCGAN method on random noise is tested.

The selected Rayleigh wave seismic data has 24 channels per shot, and each channel has 400 sampling points (2ms sampling, total time length 800 ms). Selecting 300 shots of data with high signal-to-noise ratio in a work area, performing interchannel equalization and energy normalization processing to obtain real noise-free data, artificially adding Gaussian white noise with the average value of 0 and the amplitude of 10% -30% to each shot to obtain noise-containing data with different intensity noises, training by using the DCGAN of the embodiment, and obtaining a Rayleigh wave signal denoising DCGAN model aiming at random noise through 100 epochs.

Note: epoch, when a complete data set passes through the neural network once and back once, this process is called once.

In the work area, random noise data of one shot alternative in the work area is preprocessed and then used as a test shot for carrying out denoising effect analysis, fig. 3 shows the Rayleigh wave seismic record data of the test shot containing high random noise, denoising is carried out by using trained DCGAN, and the denoising result is shown in fig. 4. From the rayleigh wave dispersion curve, fig. 5 is a dispersion energy diagram of a noisy experimental cannon, and fig. 6 is a dispersion energy diagram of denoised data. It can be seen that after the noise removal method for the rayleigh wave seismic data is used for denoising, random noise on a seismic record is well suppressed, and a dispersion curve is closer to a real dispersion characteristic, so that the transverse wave velocity can be obtained and the mechanical parameters of an underground medium can be solved based on re-inversion of the denoised rayleigh wave dispersion curve, and the method is also proved to have a better denoising effect on the rayleigh wave data acquired under high background noise.

Compared with the 2 previous embodiments, the embodiment has the advantages that the generator is optimized more, the residual learning algorithm is adopted, the data volume optimization is carried out, the algorithm optimization is carried out, the batch normalization optimization algorithm and the discarding regularization layer are adopted, and the denoising effect of Rayleigh wave data collected under high background noise is better.

In summary, the method generates the countermeasure network (DCGAN) by deep convolution in the deep learning field, and selects a part of noisy rayleigh seismic data and noiseless rayleigh seismic data as samples to be respectively added into the generator and the discriminator for training. And applying the trained DCGAN to noise-containing Rayleigh wave seismic data in a test work area, and performing random noise removal on the multi-channel surface wave data to obtain high-quality denoised Rayleigh wave seismic data. After the DCGAN training is completed, parameters do not need to be adjusted manually, the labor cost is reduced, the denoising efficiency of the Rayleigh wave seismic data is improved, and the denoising effect of the Rayleigh wave data acquired under high background noise is better, so that the Rayleigh wave dispersion curve with better denoising effect is obtained, the transverse wave speed is obtained through inversion, the mechanical parameters of the underground medium are obtained, and exploration personnel are helped to know the geophysical and geological conditions of the shallow earth surface.

Fourth embodiment

Fig. 1 is a schematic flow chart of a rayleigh wave seismic data noise removal method according to the embodiment;

the present invention also provides a storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the method for noise removal of rayleigh wave seismic data as described in any of the above.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

Fifth embodiment

Fig. 1 is a schematic flow chart of a rayleigh wave seismic data noise removal method according to the embodiment;

the present invention also provides an electronic device, comprising: a memory having a computer program stored thereon; and a processor for executing the computer program in the memory to implement the steps of the rayleigh wave seismic data noise removal method of any of the above.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

Although the embodiments of the present invention have been described above, the above description is only for the convenience of understanding the present invention, and is not intended to limit the present invention. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as disclosed, and that the scope of the invention is not to be limited to the particular embodiments disclosed herein but is to be accorded the full scope of the claims.

Claims (13)

1. A method for removing noise of Rayleigh wave seismic data is characterized by comprising the following steps:

constructing a deep convolution generation countermeasure network, wherein the deep convolution generation countermeasure network comprises a generator and a discriminator;

preprocessing Rayleigh wave seismic data to obtain real noise-free data, wherein the preprocessing comprises interchannel equalization and energy normalization processing;

adding Gaussian white noise to the preprocessed noiseless data to obtain noised data;

adding the real noise-free data and the corresponding noise-containing data serving as training samples into the deep convolution to generate a confrontation network for training, performing iterative computation until the denoising precision meets a specified threshold or the iteration number meets a specified number, and then performing iterative computation to obtain a deep convolution generated confrontation network model;

and generating a confrontation network model based on the deep convolution to carry out denoising processing on the preprocessed other Rayleigh wave seismic data to be denoised.

2. The method of claim 1,

the method for constructing the deep convolution to generate the countermeasure network comprises constructing a discriminator and a generator, and establishing a first objective function of the discriminator and a second objective function of the generator.

3. The method of claim 2,

the generator is composed of five convolutional layers, and the discriminator is composed of four convolutional layers and a full-connection layer.

4. The method of claim 3,

the first objective function is calculated as:

wherein G is a generator, z is noisy Rayleigh data,

d is a discriminator, x is noiseless Rayleigh wave data,

θ^Dand theta^GRespectively representing the parameters to be optimized of the arbiter and the generator,

P_data(x) In order to be true of the noise-free data distribution,

P_z(z) is a noisy rayleigh wave data profile,

x～P_data(x) To obey the sampling under the distribution of the noise-free data,

z～P_z(z) is the sampling under the distribution of data subject to noise,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition,

d (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z.

The second objective function calculation formula is:

wherein D (x) represents the probability that x is noise-free data,

g (z) represents the generator generated data when the input is z,

is represented by x to P_data(x) The expected value is calculated under the conditions that,

is represented by z to P_z(z) calculating the expected value under the condition.

5. The method of claim 4,

the convolutional layers in the generator adopt a full convolutional network, the activation functions of 1 st to 4 th convolutional layers in the generator are ReLU functions, and the activation functions of 5 th convolutional layers are Tanh functions;

the discriminator adopts a step convolution layer, and the activation function of all layers in the discriminator is a LeakyReLU function.

6. The method of claim 1,

the calculation formula adopted by the inter-channel equalization processing is as follows:

wherein, x'_iFor the data after the i-th track is equalized,

taking the average value of the absolute value of the whole shot gather data,

taking the average value of the ith data after absolute value,

x_iis the ith seismic data.

7. The method of claim 6,

based on the data after the inter-channel equalization processing, performing energy normalization processing, wherein the energy normalization processing adopts a calculation formula as follows:

wherein, x ″)_i，jRepresents the value of the ith sampling point after normalization,

x′_i,jfor the value of the jth sampling point of the ith channel after the inter-channel equalization,

X′_minand X'_maxRespectively representing the minimum value and the maximum value of the gather after the equalization of the whole channel.

8. The method of claim 1,

the step of adding the real noiseless data and the corresponding noisy data as training samples into the deep convolution to generate a countermeasure network for training comprises the following steps:

and adding the noisy data into the generator for training to obtain a generated noiseless data, adding the generated noiseless data and the real noiseless data which are simultaneously used as input data into the discriminator for training to obtain similarity degree values of 2 input data, and feeding back the similarity degree values to the generator to realize the mutual iterative optimization of the generator and the discriminator.

9. The method of claim 8,

the mutual iterative optimization of the generator and the arbiter comprises the following steps:

fixing the noise-containing data and each layer of parameters of the generator, and performing iterative computation to optimize a first objective function of the discriminator until the discrimination accuracy of the discriminator reaches the maximum;

then fixing the real noiseless data and each layer parameter of the discriminator, and carrying out iterative computation to optimize a second objective function of the generator until the discrimination accuracy of the discriminator reaches the minimum;

and when the generated noiseless data is consistent with the real noiseless data, a global optimal solution is reached, and the mutual iterative optimization of the generator and the discriminator is exited.

10. The method of claim 9,

the specified threshold is that the noise proportion is lower than 20%, and the specified times are more than or equal to 50.

11. The method of claim 10,

upsampling using deconvolution in the generator;

constructing a nonlinear mapping from noisy data to residual data by adopting a residual learning algorithm in the generator;

before convolution operation is carried out on the convolution layers of the generator and the discriminator, the boundary is expanded for the input characteristic data, and zero values are filled until the sizes of the input characteristic data and the input data are consistent;

and adopting a batch normalization optimization algorithm and a discarding regularization layer in the first 4 convolutional layers of the generator and the discriminator.

12. A storage medium having stored thereon a computer program, characterized in that the program, when executed by a processor, implements the steps of the rayleigh wave seismic data noise removal method according to any of claims 1 to 11.

13. An electronic device, comprising:

a memory having a computer program stored thereon; and

a processor for executing the computer program in the memory to implement the steps of the rayleigh wave seismic data noise-removal method of any of claims 1-11.

Images