CN-117169954-B - Improved convolution self-coding network-based seismic data surface wave suppression method

CN117169954BCN 117169954 BCN117169954 BCN 117169954BCN-117169954-B

Abstract

The invention provides an improved convolution self-coding network-based seismic data surface wave suppression method which comprises the steps of S1, carrying out channel head information stripping on surface wave-containing seismic data and surface wave-free seismic data, carrying out data separation according to cannons, setting data size to obtain N pairs of seismic data samples, S2, carrying out corresponding matching on the obtained N pairs of seismic data samples to obtain data samples of a required improved convolution self-coding network model, wherein N1 is randomly selected for data as a verification set, the rest N2 is used for carrying out experiments on the data as a training set, S3, firstly selecting a self-coding network as a basic network frame, inputting the surface wave-containing seismic data into the self-coding network, outputting the surface wave-free seismic data, obtaining network model mapping from the surface wave-containing seismic data to the surface wave-free seismic data, and S4, and carrying out optimization operation on a network model after the basic network model is built. The self-adaptive intelligent denoising of the seismic data surface wave noise is realized.

Inventors

WANG JIAO
WANG LINLIN
LIU PEIXUE
CHEN YUJIE
WANG ZHONGXUN

Assignees

青岛黄海学院

Dates

Publication Date: 20260512
Application Date: 20230724

Claims (8)

1. The method for suppressing the surface wave of the seismic data based on the improved convolution self-coding network is characterized by comprising the following steps of: s1, stripping the header information of the surface wave-containing seismic data and the surface wave-free seismic data, separating the data according to cannons, and setting the data size to obtain N pairs of seismic data samples; S2, carrying out corresponding matching on the obtained N pairs of seismic data samples to obtain data samples of a required improved convolution self-coding network model, wherein N1 data are randomly selected as a verification set, and the rest N2 data are used as a training set for experiments; S3, firstly, selecting a self-coding network as a basic network frame, inputting surface wave-containing seismic data to the self-coding network, and outputting surface wave-removed seismic data to obtain a network model mapping from the surface wave-containing seismic data to the surface wave-free seismic data; s4, after the basic network model is built, optimizing the network model: 1) The convolution layer is adopted to improve the basic network model: The convolution layer and the activation function form a coding unit, the deconvolution layer and the activation function form a decoding unit, and a plurality of coding units and decoding units which are in one-to-one correspondence form a complete improved convolution self-coding network; 2) And improving an activation layer of the convolutional self-coding network model by using an activation function to ensure nonlinear transfer of data in the transfer process between different network layers: the improved convolutional self-coding network uses Swish functions, swish functions are smooth and non-monotonic functions, no upper bound has a lower bound, and the derivative is constantly greater than 0; 3) After improving the network layer of the network model, the loss function of the network model is then improved: use of binary cross entropy loss functions in improved convolutional self-coding networks: In the formula (1), x i represents predicted data obtained after network training, y i represents original data, and the average value of the sum of squares of errors of corresponding points of the predicted data and the original data is calculated to judge the network training effect; in the formula (2), x i represents a label value obtained by network training, p (x i ) represents the probability of occurrence of a label value of x i , and q (x i ) represents the probability of occurrence of a label value of x i under real data; 4) The network model updates the link weight of the network by using the loss function, so that the method for updating the link weight is improved: The improved convolutional self-coding network uses an adaptive time estimation method function that reduces the likelihood of getting stuck in local minima with momentum, while it uses a separate learning rate for each parameter that can be learned, and the learning rate varies with each parameter over the training period to promote the training effect of the network: In the formula (3), m t represents the average gradient value at the first moment, v t represents the gradient variance at the second moment, θ t represents the parameter at the previous moment, θ t+1 represents the parameter at the later moment, η and e are settable coefficient values; 5) The prepared training set sample n2 is used for leading the data into an improved convolution self-coding network model to carry out network training, and the network model trained each time is stored; 6) Testing the stored network models one by one, inputting surface wave-containing seismic data to obtain surface wave-removed seismic data, and performing multi-angle result analysis on the obtained surface wave-removed seismic data to comprehensively analyze to obtain an optimal one of the network models; 7) And performing migration application verification on the surface wave-containing seismic data in the verification set n1 by using the selected optimal completed training network model.
2. The method for suppressing a surface wave of seismic data based on an improved convolutional self-coding network as set forth in claim 1, wherein in step S1, the number of seismic data samples is 78 pairs, the size of the seismic data samples is 512×512×3px, and the seismic data samples are converted into npy format.
3. The method for surface wave suppression of seismic data based on an improved convolutional self-coding network as set forth in claim 2, wherein in step S2, 15 data are selected from the 78 pairs of seismic data books as a verification set, and the remaining 63 data are used as a training set for experiments.
4. The method for suppressing seismic data surface waves based on an improved convolutional self-coding network as set forth in claim 1, wherein in step S3, the self-coding network comprises an input layer, a hidden layer and an output layer in this order, wherein the input layer is input with x, the high-dimensional data is mapped into the low-dimensional data h through the coding process, and the low-dimensional data h is converted into y through the decoding process and output from the output layer.
5. The method for suppressing seismic data surface waves based on an improved convolutional self-coding network as set forth in claim 4, wherein in step S4), x is input into the coding unit one by one, y is output from the decoding unit, and the relation map between x and y is obtained through network layer feature extraction.
6. The method for suppressing seismic data surface waves based on an improved convolutional self-coding network as set forth in claim 1, wherein the multi-angle result analysis content in step S4) at least comprises peak signal-to-noise ratio and structural similarity.
7. The method for suppressing seismic data surface waves based on an improved convolutional self-coding network as recited in claim 6, wherein the peak signal-to-noise ratio P snr is used to evaluate the quality of the network training result by using the structural similarity, and the mathematical formula is: in formula (4), M MSE represents the root mean square error, and max (c) represents the maximum value of the pixel value of the noiseless data; in the formula (5), x and y represent the surface wave-containing data and the surface wave-removed data, respectively, wherein mu x , Represents the mean and variance of x, mu y , Represents the mean and variance of y, respectively, σ xy represents the covariance of x and y, and c 1 ,c 2 is a constant.
8. The method for suppressing surface waves of seismic data based on the improved convolutional self-coding network as set forth in claim 3, wherein in step S4), specifically, the prepared 15 surface wave-containing seismic data are input into the trained improved convolutional self-coding network model, and the corresponding 15 surface wave-removed seismic data are output, so that the surface wave noise in the seismic data is effectively suppressed.

Description

Improved convolution self-coding network-based seismic data surface wave suppression method Technical Field The invention belongs to the technical field of seismic exploration, relates to a coherent noise elimination mode, and in particular relates to a method for suppressing seismic data surface waves based on an improved convolution self-coding network. Background Because of complex environment in actual seismic exploration, the obtained seismic data contains a large amount of noise, and the normal operation of links such as subsequent seismic data processing, seismic data interpretation and the like is seriously influenced. The surface wave is used as typical coherent noise in seismic exploration, is a low-frequency, low-speed and high-amplitude regular interference wave, widely exists in seismic records, and has obvious broom-like divergence distribution characteristics, so that the correct analysis of seismic data by seismic data interpreters is seriously affected. Therefore, suppressing the surface wave noise in the seismic noise is always a key means for improving the signal-to-noise ratio and the resolution of the seismic data. According to the characteristics of strong surface wave energy, low propagation speed and slower attenuation, the conventional surface wave suppression method commonly used at present comprises frequency domain filtering, wavelet transformation filtering, regional filtering, abnormal amplitude attenuation, three-dimensional FFK filtering and the like. The frequency domain filtering and the wavelet transform filtering consider the characteristic of low frequency of the surface wave, but when the effective signal and the surface wave interference have frequency overlap, the suppression effect is not ideal. The region filtering is a band-pass filtering for a target region, and the high-pass filtering is performed on the surface wave region by utilizing the characteristic of low frequency of the surface wave so as to achieve the purpose of suppressing the surface wave. The method has the advantages of simplicity, practicability, no influence on data outside the area by processing signals in the area of the surface wave, and the defect that only the frequency attribute is considered, and effective signals in the area are filtered while the surface wave is removed, so that the effective information intensity is weakened. Whereas the abnormal amplitude decay is effectively a median filter. Within a certain frequency band, tracks whose amplitudes deviate from a median specified threshold will be attenuated or interpolated from adjacent tracks. The method has the advantages that amplitude statistics can be carried out in different frequency bands, noise close to the amplitude of the effective value in the shallow layer is attenuated, and the method has the defect that surface wave residues exist in the deep layer, and the residual surface wave can influence the accuracy of the subsequent energy compensation coefficient. In addition, the three-dimensional FFK filtering is based on crisscross arrangement, and the two-dimensional FK filter is spatially expanded from a fan shape to a conical filtering. The method can obtain better effect from shallow to deep by pressing the surface wave, but the method processes the whole gather data and processes the signals outside the surface wave area, so that the effective signals are attenuated. The conventional surface wave suppression method generally transforms signals so that effective signals and surface waves have larger differences in a certain dimension, and further separation of the effective signals and surface wave noise signals is achieved by setting a reasonable threshold value. However, aliasing exists inevitably in part of the distinction between the effective signals and the surface waves, so that the traditional surface wave suppression method removes part of the effective waves while removing the surface waves, and a large amount of effective information in the seismic data is lost. Disclosure of Invention Aiming at the problems in the prior art, the invention provides a method for suppressing the surface wave of the seismic data based on an improved convolutional self-coding network, which is based on the fact that actual seismic data is taken as a sample training set, and a convolutional self-coding network is improved to be taken as a deep neural network method of a model frame for training, so that model mapping from the surface wave-containing seismic data to the surface wave-free seismic data is obtained, and the self-adaptive intelligent denoising of the surface wave noise of the seismic data is realized. The aim of the invention can be achieved by the following technical scheme: the method for suppressing the surface wave of the seismic data based on the improved convolution self-coding network comprises the following steps: s1, stripping the header information of the surface wave-containing seismic data and the surface wav