阅读说明：本技术 一种基于nar-tfpf压制地震勘探随机噪声的方法 (Method for suppressing random noise of seismic exploration based on NAR-TFPF ) 是由 李光辉 弓子卉 冯志强 李朝辉 于 2019-11-29 设计创作，主要内容包括：本发明属于地震勘探技术领域,具体涉及一种基于NAR-TFPF压制地震勘探随机噪声的方法。本发明基于NAR-TFPF压制地震勘探随机噪声的方法包括以下步骤：对地震含噪信号进行分段；根据地震资料中有效信号和随机噪声非线性特性有很大的差异,通过非线性自回归模型(NAR)区分信号段和噪声段；采用时频峰值滤波(TFPF)进行去噪处理,噪声段选取大窗长进行噪声压制,信号段选取小窗长进行保幅,本发明可以在压制强随机噪声的同时保留有效信号,提高地震资料的信噪比,能更准确地用于地质解释。(The invention belongs to the technical field of seismic exploration, and particularly relates to a method for suppressing random noise of seismic exploration based on NAR-TFPF. The method for suppressing random noise of seismic exploration based on NAR-TFPF comprises the following steps: segmenting the seismic noisy signals; according to the great difference between the nonlinear characteristics of effective signals and random noise in seismic data, distinguishing signal segments and noise segments through a nonlinear autoregressive model (NAR); the method adopts time-frequency peak filtering (TFPF) to perform denoising treatment, the noise section selects a large window length to perform noise suppression, and the signal section selects a small window length to perform amplitude preservation.)

1. A method for suppressing random noise of seismic exploration based on NAR-TFPF is characterized in that: the method comprises the following steps:

step 1, carrying out PCNN segmentation processing on seismic data;

step 2, NAR modeling is carried out, and the optimal order of the NAR is calculated to estimate the nonlinearity of the NAR;

and 3, selecting a proper window length for each time sequence to carry out TFPF filtering, and finishing the processing.

2. The method of claim 1, wherein the NAR-TFPF-based suppression of random noise in seismic exploration is performed by: the PCNN segmentation processing of the seismic data in the step 1 specifically comprises the following operations: and (3) selecting a pulse coupling neural network PCNN to segment the seismic data, and randomly dividing the time sequence into a plurality of segments.

3. The method of claim 1, wherein the NAR-TFPF-based suppression of random noise in seismic exploration is performed by: the NAR modeling in the step 2, and the specific operation of calculating the optimal order of the NAR and estimating the nonlinearity thereof are as follows:

y_n＝G[y_n-1,y_n-2,…,y_n-k]formula (1)

Where y is the system response, k is the memory factor, n is the discrete time, and G is a nonlinear equation;

when G is chaotic mapping, the system output is a chaotic sequence; when time series y_n1,2, … N is known, it can be modeled by equation G, a polynomial G that includes all the incremental orders used to obtain a generic nonlinear equation;

y_nis the historical state y_n-1+y_n-2+…+y_n-kThe sum of 0 to d powers, expressed as

Wherein

determining values of a memory factor k and an order d, and arranging all items according to ascending power after establishing a model according to a formula (2); intercepting the first r term of the built model to obtain a polynomial, coefficient a_mIs determined by least squares fitting and then the predicted time series is calculated

Obtaining a predicted value through a formula (2)

Prediction value

where epsilon represents the prediction error and,

Calculating a prediction error epsilon, changing the value of r, repeating the calculation processes of the formulas (2) to (3), and drawing an epsilon (r) -r curve;

when the prediction error epsilon is minimum, the value of d is the optimal order and can be recorded as d_optMThe larger the optimal order is, the higher the nonlinearity degree of the time series is, and the time series is a noise section; the smaller the optimal order number is, the lower the nonlinearity degree of the time sequence is, and the signal segment is represented;

the NAR model is used to model the time series.

4. The method of claim 1, wherein the NAR-TFPF-based suppression of random noise in seismic exploration is performed by: selecting a proper window length for each time sequence in the step 3 to carry out TFPF filtering, wherein the specific operation of finishing the processing is as follows: in general, the noisy signal is expressed as:

where s (t) is a noisy signal, x_g(t) is a valid signal, n_g(t) is additive noise, x_kg(t) is a component of the effective signal;

regarding the noisy signal s (t) as the instantaneous frequency of the frequency modulation signal, obtaining an analytic signal by modulating the frequency of s (t), as shown in the following formula:

wherein z is_s(t) is an analytic signal, exp is an exponential function with a natural constant e as a base, mu is a scale parameter, j is a pure imaginary number, pi is a circumference ratio, lambda is an integral variable, and s (lambda) is a noisy signal with the integral variable lambda;

by evaluating the analytic signal z_sThe WVD peak of (t) can recover the valid signal from the background noise,

wherein

the pseudo-wigner distribution PWVD is defined as follows:

where h (τ τ τ) is a window function, z_sFor the analytic signal in equation (5),

replacing the maximum value of the WVD with the maximum value of the pseudo-Wigner distribution PWVD;

the TFPF window function value for a seismic wavelet is:

wherein f is_sIs the signal sampling frequency, f_dIs the dominant frequency of the seismic wavelet, WL is the window length and can only take the odd number;

the nonlinear degree is high, and the noise is suppressed by selecting a longer TFPF window length;

the nonlinear degree is low, and the signal amplitude is preserved by selecting a shorter TFPF window length.

Technical Field

The invention belongs to the technical field of seismic exploration, and particularly relates to a method for suppressing random noise of seismic exploration, in particular to a method for suppressing random noise of seismic exploration based on NAR-TFPF.

Background

The random noise of seismic exploration always appears in seismic data along with effective signals, and the random interference with large energy can directly influence dynamic and static correction analysis and final imaging effect, thereby bringing adverse effect to subsequent geological interpretation. For this reason, expert scholars have devised, improved a number of algorithms for dealing with random noise. The design concept can be roughly divided into two types of signal enhancement and noise suppression, and aims to improve the signal-to-noise ratio of seismic data as much as possible, for example, f-x deconvolution, polynomial fitting and the like belong to a signal enhancement algorithm, and a vector decomposition algorithm under median constraint belongs to a noise suppression method. These algorithms have been widely used and their application results are relatively satisfactory for production. Meanwhile, the algorithms are limited by some assumed conditions in the application process, and the filtering effect is directly influenced. For example, these algorithms are based on effective signal spatial correlation and random noise spatial uncorrelated, and random noise is a two-dimensional random process in time-space domain, and it is far from sufficient to study random noise in spatial direction.

Boashash and Mesbah proposed a new signal enhancement algorithm of one year in 2004, time-frequency peak filtering. Based on the time-frequency analysis theory, firstly, the analytic signal is modulated by the signal containing noise, the Wigner time-frequency distribution (WVD) of the analytic signal is solved, and the instantaneous frequency is obtained by estimating the maximum value of the WVD, so that the purpose of signal enhancement is achieved. The TFPF is firstly applied to electroencephalogram analysis of newborns, is firstly applied to suppression of random noise in seismic exploration in 2005, and according to the filtering theory of the TFPF, the TFPF can recover linear signals from strong white Gaussian noise but can distort the nonlinear signals, and the seismic signals, particularly signals with high frequency, have nonlinear characteristics. To avoid this problem, the WVD is replaced by a pseudo-Wigner distribution (PWVD) to ensure a linear condition of the instantaneous frequency within the window. The improved TFPF achieves effective results in recovering seismic signals on a background of strong random noise, but the selection of the window length is critical to suppressing random noise and preserving effective signal waveforms by the TFPF. In order to obtain better filtering effect, in the last years, the relation between the window length, the sampling frequency and the main frequency of the signal is utilized by people like forest red wave, and the variable window length TFPF is proposed. According to the method, a noisy signal is divided into time sequences with any length, different window lengths are selected in different time sequences, and therefore high-frequency component distortion caused by improper TFPF window lengths is solved. The Liuyanping et al uses Empirical Mode Decomposition (EMD) to improve the conventional TFPF. The decomposition characteristic of EMD is utilized to decompose the noise-containing signal into several modes from high frequency to low frequency, different TFPF window lengths are selected for different frequency modes, and the noise reduction and amplitude preservation are realized by identifying signal components. The time period segmentation processing is directly carried out on the noise-containing signals, so that the noise segments and the signal segments cannot be accurately distinguished, the noise segments and the signal segments are influenced by the length of a time sequence, and the most appropriate window length cannot be adopted; EMD decomposition is carried out on a noisy signal, and the problem of mode aliasing is easy to occur.

In recent years, nonlinearity is one of important criteria in stochastic signal analysis, and is widely used in the fields of clinical diagnosis, stock prediction, mechanical failure diagnosis, and the like. The non-linearity degree of the seismic signal is different from that of random noise, a signal part and a noise part in the seismic data are distinguished through the non-linearity degree, a long window length is selected in the noise part, a short window length is selected in the signal part, and meanwhile the signal-to-noise ratio and the signal amplitude are enhanced. When a noisy signal is filtered by using a fixed window length, the noise can be well inhibited by selecting the long window length, but the effective signal amplitude is also attenuated; the short window length is selected to have better amplitude protection for effective signal amplitude, but noise can not be suppressed more effectively, and the signal-to-noise ratio of the filtered seismic data is lower. How to keep the amplitude of effective signals while suppressing noise is an urgent problem to be solved.

Disclosure of Invention

Aiming at the problem that noise and signal amplitude cannot be simultaneously and effectively suppressed in the prior art, the invention provides a method for suppressing random noise in seismic exploration based on NAR-TFPF.

In order to achieve the purpose, the invention adopts the following technical scheme:

a method for suppressing random noise of seismic exploration based on NAR-TFPF comprises the following steps:

step 1, carrying out PCNN segmentation processing on seismic data;

step 2, NAR modeling is carried out on each time sequence, and the optimal order of the NAR is calculated to estimate the nonlinearity of the NAR;

and 3, selecting a proper window length for each time sequence to carry out TFPF filtering, and finishing the processing.

Further, the specific operation of performing PCNN segmentation processing on the seismic data in the step 1 is as follows: and (3) selecting a Pulse Coupled Neural Network (PCNN) to segment the seismic data, and randomly dividing the time sequence into a plurality of segments. The selection of the TFPF filter window length is a key parameter that affects the filtering effect. It is not reasonable to process the signal portion and the noise portion using the same window length, so we can divide the noisy seismic signal into several time segments. Since conventional time series segmentation methods, such as fixed length segmentation, fast segmentation, fuzzy segmentation, etc., may divide a cycle or a complete series of wavelets into two parts, the result is affected by the length and amplitude of the series, and PCNN can avoid these problems when used for time series segmentation. As shown in fig. 5, fig. 5(a) shows a noisy signal with only one wavelet, fig. 5(b) shows a fixed-length division result, and a complete wavelet is divided into two parts, and fig. 5(c) shows a PCNN division result, which avoids the problem in fig. 5 (b).

Still further, the performing NAR modeling in step 2, calculating the optimal order of NAR and estimating its nonlinearity specifically comprises:

y_n＝G[y_n-1,y_n-2,…,y_n-k]formula (1)

Where y is the system response, k is the memory factor, n is the discrete time, and G is a nonlinear equation;

when G is chaotic mapping, the system output is a chaotic sequence; when time series y _n1,2, … N is known, it can be modeled by equation G, a polynomial G that includes all the incremental orders used to obtain a generic nonlinear equation;

y_nis the historical state y_n-1+y_n-2+…+y_n-kThe sum of 0 to d powers, expressed as

Wherein

Representing the predicted system response, k is the memory factor, d is the order, TL is the number of terms, a_mM is 0,1,2 … k, k +1, k +2 … TL-1 is a coefficient obtained by an arbitrary curve fitting method;

the formula 2 can also be

Determining values of a memory factor k and an order d, and arranging all items according to ascending power after establishing a model according to a formula (2); intercepting the first r term of the built model to obtain a polynomial, coefficient a_mIs determined by least squares fitting and then the predicted time series is calculated

Obtaining a predicted value through a formula (2)

Prediction value

And the original value { y ] shown in equation (1)_nThe mean square error between k +1, k +2 … n:

where epsilon represents the prediction error and,

is the mean and satisfies

Calculating a prediction error epsilon, changing the value of r, repeating the calculation processes of the formulas (2) to (3), and drawing an epsilon (r) -r curve;

when the prediction error epsilon is minimum, the value of d is the optimal order and can be recorded as d_optMThe larger the optimal order is, the higher the nonlinearity degree of the time series is;

the NAR model is used for modeling the segmented time series. Fig. 1 shows white gaussian noise, and the curve of ∈ (r) -r obtained by NAR modeling is shown in fig. 2, and it can be seen that the prediction error is minimum when the term r is 28, and when d is 3, that is, the optimal order d of the noise sequence shown in fig. 1 is shown_optM3. FIG. 3 shows a time series of seismic wavelets with a dominant frequency of 30Hz, which were NAR modeled, and the ε (r) -r curve is shown in FIG. 4. It can be seen that the prediction error is minimal when r is 6, when d is 2, i.e. the optimal order d of the wavelet series shown in fig. 3_optM2. The noise sequence is more non-linear than the effective signal, so that the signal part and the noise part in a noisy signal can be distinguished by NAR modeling.

Furthermore, in step 3, a suitable window length is selected for each time sequence to perform TFPF filtering, and the specific operation of completing the processing is: in general, the noisy signal is expressed as:

where s (t) is a noisy signal, x_g(t) is a valid signal, n_g(t) is additive noise, x_kg(t) is a component of the effective signal;

regarding the noisy signal s (t) as the instantaneous frequency of the frequency modulation signal, obtaining an analytic signal by modulating the frequency of s (t), as shown in the following formula:

wherein z is_s(t) is an analytic signal, exp is an exponential function with a natural constant e as a base, mu is a scale parameter, j is a pure imaginary number, pi is a circumference ratio, lambda is an integral variable, and s (lambda) is a noisy signal with the integral variable lambda;

by evaluating the analytic signal z_sThe WVD peak of (t) can recover the valid signal from the background noise,

wherein

As a result of the filtering, W_Zs(t, f) is an analytic signal z_s(t) WVD time-frequency distribution, argmax is to find W_Zs(t, f) a function of the maximum value;

according to the TFPF filtering theory, when the effective signal changes linearly along with time and the background noise is Gaussian white noise, unbiased estimation that the filtered signal is a signal s (t) containing noise is obtained through formulas (4) to (6), so that the key of the WVD-based TFPF filtering method is to ensure the linear characteristic of the effective signal; the seismic signals, particularly the seismic signals with higher main frequency, are generally nonlinear, so windowed WVD distribution, namely PWVD, is adopted to replace WVD, thereby ensuring the approximate linear characteristic of the signals in the window;

the pseudo-wigner distribution PWVD is defined as follows:

where h (τ τ τ) is a window function, Z_sFor the analytic signal in equation (5),

is Z_sJ is a pure imaginary number, f is a frequency variable;

replacing the maximum value of the WVD with the maximum value of the pseudo-Wigner distribution PWVD;

the TFPF window function value for a seismic wavelet is:

wherein f is_sIs the signal sampling frequency, f_dIs the dominant frequency of the seismic wavelet, WL is the window length and can only take the odd number; the window length can only be chosen to be odd and too long can cause signal distortion. The selection of the TFPF filter window length is a key parameter that affects the filtering effect. It is not reasonable to process the signal part and the noise part using the same window length.

The nonlinear degree is high, and the noise is suppressed by selecting a longer TFPF window length;

the nonlinear degree is low, and the signal amplitude is preserved by selecting a shorter TFPF window length.

Firstly, the nonlinear degree of noise is higher than that of a signal, so that a noise section and a signal section are judged according to the nonlinear degree; the larger the window length is, the more the time sequence amplitude attenuation is, the longer the noise section is selected to be the large window length, the more the noise attenuation is serious, the longer the signal section is selected to be the small window length, and the signal amplitude attenuation is reduced, so that the signal-to-noise ratio is improved.

Compared with the prior art, the invention has the following advantages:

aiming at the problems that the signal-to-noise ratio of seismic data is extremely low, and the TFPF window length is not variable, the signal amplitude can not be kept while noise is effectively suppressed, the invention provides the method for searching the optimal filter window length of the TFPF through NAR modeling, and improves the signal-to-noise ratio and the resolution ratio of the seismic data. In the whole data processing process, modeling is carried out on each section of data according to the NAR, so that effective signals are identified and detected without the need of conditional assumption on an algorithm; each section adopts different window lengths to carry out TFPF filtering, so that the TFPF window length self-adaptive filtering is realized, effective signals are not reduced while noise is suppressed, the effective signals can be effectively recovered from strong background noise, the usability of seismic data is increased, and more favorable information is provided for subsequent geological interpretation.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate embodiments of the invention and, together with the description, serve to explain the invention and not to limit the invention.

FIG. 1 is a seismic random noise time series;

FIG. 2 is a plot of ln ε -r of a seismic random noise sequence;

FIG. 3 a Rake wavelet of a seismic effective signal;

FIG. 4 is a plot of ln ε -r of a Rake wavelet of a seismic effective signal;

FIG. 5 is a schematic diagram of a single trace seismic noisy recording and its segmentation results, wherein (a) represents a noisy time series, (b) represents a fixed time segmentation result, and (c) represents a PCNN segmentation result;

FIG. 6 is a diagram of embodiment 1 of the present invention, applied to a seismic noisy signal processing, wherein (a) is a noisy sequence after PCNN segmentation, (b) represents a TFPF de-noising result, and (c) represents a NAR-TFPF de-noising result;

FIG. 7 is a diagram of a seismic synthetic record processing of embodiment 2 of the present invention, (a) clean records, (b) noisy records, (c) after TFPF noise suppression, (d) NAR-TFPF de-noising results;

FIG. 8 is a frequency-wavenumber spectrum from each record shown in FIG. 7, (a) for clean records, (b) for noisy records, (c) for after TFPF noise suppression, (d) for NAR-TFPF de-noising results;

FIG. 9 is a partial field data acquisition processing diagram, (a) is data containing noise, (b) is a result after TFPF noise suppression, and (c) is a NAR-TFPF de-noising result;

FIG. 10 is an enlarged view of the portion of FIG. 9 shown in the box, (a) is noisy data, (b) is the result of TFPF noise suppression, and (c) is the NAR-TFPF de-noising result.

Detailed Description

In order to make those skilled in the art better understand the technical solution of the present invention, the technical solution in the embodiment of the present invention will be clearly and completely described below with reference to the drawings in the embodiment.