CN104392119A - Multiphase support vector regression-based seismic wave crest and trough modeling method - Google Patents

Abstract

The invention discloses a multiphase support vector regression-based seismic wave crest and trough modeling method, which comprises the following steps of performing attribute preprocessing on seismic data acquired by a seismic apparatus, wherein the attribute preprocessing refers to de-noising and filtering processing over seismic waves by virtue of seismic inversion software; establishing a multiphase support vector regression model E[y/x]=fd(x; Thetad), optimizing a parameter of the established model by adopting a genetic algorithm, selecting proper kernel functions to construct an optimal prediction model, and performing wave crest and trough prediction on a test sample which is data acquired by the seismic inversion software by utilizing the prediction model. According to the method, the wave crest and trough positions of the seismic waves are better predicted, so that an imminent waveform image is more approximate to original graphics; meanwhile, a plurality of kernel functions are effectively combined, so that the prediction capability of a multiphase support vector regression can be improved, and prediction errors are greatly reduced; the genetic algorithm is adopted for optimizing the model, so that higher practicability and robustness are achieved.

Description

A kind of scheitel trough modeling method based on heterogeneous support vector regression

Technical field

The present invention relates to machine learning modeling method, specifically, relating to a kind of scheitel trough modeling method based on returning heterogeneous support vector machine.

Background technology

Support vector regression (Support Vector Regression, SVR), has been widely used in the fields such as System Discrimination, forecast, modeling control as a kind of senior machine learning method.It is that Corpus--based Method theory and empirical risk minimization are set up, and has extremely strong generalization ability.Its core concept is to solve a quadratic programming problem, not only can ensure to obtain globally optimal solution, and kernel function can be adopted to replace the dot-product operation of high-dimensional feature space, thus greatly reduces complicated computation process.

Breakaway poing a kind ofly causes the point of exporting change because of input variable flip-flop, often there will be the point affecting global change like this, such as flex point, maximum point, minimum point etc. at function back warp.At present, SVR is adopted to can be good at solving the function approximation problem not having breakaway poing.Due to the existence of breakaway poing, traditional SVR method can not obtain extraordinary Approximation effect, now returns heterogeneous support vector machine (mp-SVR) and arises at the historic moment.

The Wave crest and wave trough of seismic event is the key component of geological data, and its importance is self-evident.Because it is present in maximum or minimum point, traditional inversion algorithm is difficult to the accurate location simulating Wave crest and wave trough accurately.The present invention can be good at predictably shaking the position of Wave crest and wave trough by the Forecasting Methodology introducing a kind of scheitel trough based on heterogeneous support vector regression.

Genetic algorithm is the searching algorithm with robustness that a class can be used for complex systems optimization, because the advantages such as it is easy to operate, robustness is good are widely used in Nonlinear Multiobjective optimization problem, it can walk abreast to objective function, the adaptive global search of Stochastic sum, is of great benefit to the precision and arithmetic speed improving forecast model.

Summary of the invention

In order to solve the problem of above-mentioned prior art, the invention provides a kind ofly can to facilitate, the modeling method of accurate, practical scheitel trough.

To achieve these goals, the technical solution used in the present invention is as follows:

(1) attribute pre-service is carried out to input data, carry out denoising, filtering, rectification, skew and overlap-add procedure comprising to seismic event;

(2) using original earthquake data as training sample, set up multinomial support vector regression model E [y/x]=f _d(x; θ _d);

(3) adopt genetic algorithm to be optimized institute's established model parameter, and choose suitable kernel function structure optimum prediction model;

(4) using utilize seismic inversion software to obtain data as test sample book, utilize described forecast model to carry out the prediction of Wave crest and wave trough.

The present invention has following beneficial effect:

(1) the present invention adopts different kernel functions to obtain different effects, and combine multiple kernel function the predictive ability that can improve and return heterogeneous support vector machine simultaneously effectively, greatly reduces predicated error;

(2) by recurrence of the present invention heterogeneous support vector machine scheitel trough modeling method, the matching performance to training sample can be improved, there is stronger generalization ability, crest and the wave trough position of seismic event can be doped preferably, make the waveform image that approaches closer to original figure;

(3) adopt genetic algorithm to carry out effectively optimizing to parameter, be of value to and obtain best modeler model.

Embodiment

In order to outstanding objects and advantages of the present invention, below in conjunction with accompanying drawing, the invention will be further described.

Fig. 1 is the process flow diagram of a kind of scheitel trough modeling method based on the heterogeneous support vector machine of recurrence of the present invention.Mainly comprise:

(S10) attribute pre-service is carried out to seismic event;

(S11) forecast model returning heterogeneous support vector machine is built;

(S12) parameter and the kernel function of Optimization Model of Genetic Algorithm is applied;

(S13) described forecast model is utilized to predict.

Part I: process of data preprocessing.

First, obtain original earthquake data by seismic instrument, then attribute pre-service is carried out to original earthquake data.Due to the noise that some factors such as instrument and environment cause may be comprised in original earthquake data, be therefore necessary to carry out denoising to original earthquake data.Meanwhile, may contain the wavelength of some redundancies in original earthquake data, these wavelength will affect precision of prediction and the speed of model, are therefore necessary to carry out filtering process.

Part II: modeling process.

Structure returns heterogeneous supporting vector machine model (mp-SVR): E [y/x]=f _d(x; θ _d), γ _d-1< x < γ _d, d=1,2 ..., D, (1) here, f _d() is regression function, θ _dbe parameter vector, D-1 is unique breakaway poing γ _dnumber, γ ₀and γ _dbe respectively the upper and lower bound that the input space interrupts knick point.

The model f of each phase in the model _d(x; θ _d), d=1,2 ..., D can be expressed as weight vectors w _dwith input vector φ _dthe inner product of (x): f _d(x; θ _d)=<w _d, φ _d(x) >+b _d, (2) here, <, > are inner product, b _dfor off-set value.

Suppose that one group of orderly data is for { (x ₁, y ₁), (x ₂, y ₂) ..., (x _n, y _n) as training set, γ=(γ ₁, γ ₂, γ _d-1) ^tfor breakaway poing wherein, now D part of being divided into of the whole input space, that is: $\{(x_1,y_1),(x_2,y_2),...,(x_{p_1},y_{p_1})\},\{(x_{p_1+1},y_{p_1+1}),(x_{p_1+2},y_{p_1+2}),...,(x_{p_2},y_{p_2})\},...,\{(x_{p_{D-1}+1},y_{p_{D-1}+1}),(x_{p_{D-1}+2},y_{p_{D-1}+2}),...,(x_{P_D},y_{p_D})\}$ Here d=1,2 ..., D-1.

By the upper original function that we obtain returning heterogeneous support vector machine be:

$\underset{d=1}{\overset{D}{\mathrm{\&Sigma;}}}\{\frac{1}{2}{\left|\right|w_d\left|\right|}^2+C_d\underset{i=p_d+1}{\overset{p_d}{\mathrm{\&Sigma;}}}({\mathrm{\&xi;}}^{+}+{\mathrm{\&xi;}}^{-})\}$

$s.t\left\{\begin{array}{c}{y}_i-<w_d,\mathrm{\&phi;}\left(x_i\right)>-b_d\&le;{\mathrm{\&epsiv;}}_d+{\mathrm{\&xi;}}^{+},i=(p_{d-1}+1),...,p_d,d=\mathrm{1,2},...,D,\\ <w_d,\mathrm{\&phi;}\left(x_i\right)>+b_d-y_i\&le;{\mathrm{\&epsiv;}}_d+{\mathrm{\&xi;}}^{-},i=(p_{d-1}+1),...,p_d,d=\mathrm{1,2},...,D,\\ <w_d,\mathrm{\&phi;}\left(x_i\right)>+b_d=<w_d,\mathrm{\&phi;}\left(x_i\right)>+b_{d+1},d=\mathrm{1,2},...,D-1,\\ {\mathrm{\&xi;}}^{+},{\mathrm{\&xi;}}^{-}\&GreaterEqual;0,i=\mathrm{1,2},...,n,\end{array}\right.---\left(3\right)$

Adopt KKT (Karush-Kuhn-Tucker) method and introduce Lagrangian function and obtain final objective function L _f:

$\begin{array}{c}{L}_F=-\underset{d-1}{\overset{D}{\mathrm{\&Sigma;}}}\{\frac{1}{2}\underset{i=p_d+1}{\overset{p_d}{\mathrm{\&Sigma;}}}\underset{j=p_d+1}{\overset{p_d}{\mathrm{\&Sigma;}}}({{\mathrm{\&lambda;}}_i}^{+}-{{\mathrm{\&lambda;}}_i}^{-})({{\mathrm{\&lambda;}}_j}^{+}-{{\mathrm{\&lambda;}}_j}^{-})H_d(x_i,x_j)+{\mathrm{\&epsiv;}}_d\underset{i=1}{\overset{p_d}{\mathrm{\&Sigma;}}}({{\mathrm{\&lambda;}}_i}^{+}+{{\mathrm{\&lambda;}}_i}^{-})\}+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i({{\mathrm{\&lambda;}}_i}^{+}-{{\mathrm{\&lambda;}}_i}^{-})-\\ \underset{2\&le;d\&le;D-1(D>2)}{\mathrm{\&Sigma;}}\{\frac{1}{2}\underset{i=1}{\overset{p_{d-1}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{p_{d-1}}{\mathrm{\&Sigma;}}}({{\mathrm{\&lambda;}}_i}^{+}-{{\mathrm{\&lambda;}}_i}^{-})({{\mathrm{\&lambda;}}_j}^{+}-{{\mathrm{\&lambda;}}_j}^{-})H_d({\mathrm{\&gamma;}}_{d-1},{\mathrm{\&gamma;}}_{d-1})+\underset{i=p_{d-1}+1}{\overset{p_d}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{p_{d-1}}{\mathrm{\&Sigma;}}}({{\mathrm{\&lambda;}}_i}^{+}-{{\mathrm{\&lambda;}}_i}^{-})({{\mathrm{\&lambda;}}_j}^{+}-{{\mathrm{\&lambda;}}_j}^{-})H_d(x_i,{\mathrm{\&gamma;}}_{d-1})\end{array}---\left(4\right)$

Here $H_d(x_1,x_2)=\left\{\begin{array}{cc}{K}_d(x_1,x_2)-K_d(x_1,{\mathrm{\&gamma;}}_d)-K_d(x_2,{\mathrm{\&gamma;}}_d)+K_d({\mathrm{\&gamma;}}_d,{\mathrm{\&gamma;}}_d),& d\&NotEqual;D\\ K_d(x_1,x_2)-2K_d(x_1,{\mathrm{\&gamma;}}_{d-1})+({\mathrm{\&gamma;}}_{d-1},{\mathrm{\&gamma;}}_{d-1}),& d=D\end{array}\right.---\left(5\right)$

And kernel function K _d(x ₁, x ₂) replace < φ _d(x ₁), φ _d(x ₂) >.

Part III: parameter optimization and Core Choice process.

The optimization problem of the model obtained by the modelling phase depends primarily on parameter γ, choosing of C, ε and kernel function type, choose that different IPs function obtains effect by different, therefore, parameter optimization and Core Choice seem particularly important.Traditional prioritization scheme often there is a large amount of calculating and efficiency is lower.For obtaining optimum mode, taking genetic algorithm and model is optimized, and adopt mixed base to encode to chromosome and kernel function respectively because of encoding scheme.

Carry out coding to gene to comprise: chromosome adopts real coding scheme, kernel function adopts integer coding scheme.Initialization is carried out to genetic algorithm and mainly comprises the initial value that initial population, cycle index, crossover probability, mutation probability, parameter area and core type are set.

By adopting selection, crossover and mutation three kinds of modes of operation successively as the system of selection of genetic manipulation.

Set up the database of training data and test data, in mp-SVR (as Fig. 2) model, training is when the sample of former generation, and use test sample book to do Fitness analysis analysis, if reach end condition, export best mp-SVR model, otherwise continue the next generation's just circulation optimizing process.

By the parameter γ that Using Genetic Optimization Algorithm obtains, C, ε and kernel function type, as the parameter value of mp-SVR model and kernel function type, set up the scheitel trough forecast model based on returning heterogeneous support vector machine.

Part IV: scheitel trough forecasting process.

Training sample in regression function is done the calculating of monokaryon and multi-kernel function.

Calculate the predicted value of geological data sample to be tested.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of a kind of scheitel trough modeling method based on the heterogeneous support vector machine of recurrence of the present invention.

Fig. 2 is that the present invention adopts genetic algorithm to carry out the process flow diagram of modeling optimization to mp-SVR model.

Claims (5)

1., based on the scheitel trough modeling method returning heterogeneous support vector machine, it is characterized in that, attribute pre-service is carried out to original earthquake data, mainly comprise denoising and filtering process;

Set up and return heterogeneous supporting vector machine model E [y/x]=f _d(x; θ _d);

Adopt genetic algorithm to be optimized institute's established model parameter, and choose suitable kernel function structure optimum prediction model;

Using utilize seismic inversion software to obtain data as test sample book, utilize described forecast model to carry out the prediction of Wave crest and wave trough.

2. a kind of scheitel trough modeling method returning heterogeneous support vector machine according to claim 1, it is characterized in that: the heterogeneous supporting vector machine model of the recurrence of setting up mainly considers the situation of the existence of breakaway poing, namely introduce parameter γ (breakaway poing), this is the diacritical point with traditional SVR.

3. a kind of scheitel trough modeling method returning heterogeneous support vector machine according to claim 1, it is characterized in that: mainly for parameter γ (breakaway poing), C (weight vectors), ε (parameter vector) are optimized.

4. a kind of scheitel trough modeling method returning heterogeneous support vector machine according to claim 1, is characterized in that: the basic kernel function of employing comprises linear function, polynomial function, radial basis function and splines.

5. a kind of scheitel trough Forecasting Methodology returning heterogeneous support vector machine according to claim 1, is characterized in that: what genetic algorithm adopted is hybrid coding mode, and namely chromosome adopts real coding scheme, and kernel function adopts integer coding scheme.