CN103487832B - Supervision waveform classification is had in a kind of 3-D seismics signal - Google Patents

Abstract

Have supervision waveform classification in 3-D seismics signal of the present invention, decayed tooth comprises number of steps Data preprocess, feature selecting and class indication.Beneficial effect is that the present invention is based on 3-D seismics signal data and log data information, by genetic algorithm, the attributive character extracted is optimized, utilize svm classifier algorithm, analyzed 3-D seismics objective interval data are carried out waveform separation division, identify different seismic facies, and then reliably support for latter earthquake data interpretation provides, improve the reliability to lithology prediction, Sand-body Prediction, fractured reservoirs prediction and subtle reservoir formation prediction etc.Relative to only comparing with SVM design category device, adding genetic algorithm and carrying out feature selecting, reducing the design complexities of SVM classifier, thus improve waveform separation treatment effeciency.

Description

Supervision waveform classification is had in a kind of 3-D seismics signal

Technical field

The invention belongs to the sorting technique field of waveform in seismic signal, relate to a kind of waveform classification of 3-D seismics signal, particularly wherein have supervision waveform sorting technique.

Background technology

Waveform separation technology based on seismic signal is the important means that seismic interpretation personnel carry out subsurface reservoir and stratal configuration analysis.Rationally and accurately seismic signal waveform classification results can reflect subsurface reservoir and stratal configuration structure truly, thus be conducive to seismic interpretation personnel structure elucidation is accurately carried out to underground structure, and then improve lithology prediction, Sand-body Prediction, fractured reservoirs prediction and subtle reservoir formation forecasting reliability, thus minimizing exploration risk, save prospecting prime cost, bring huge economic and social benefit.So, to the waveform separation research of seismic signal, there is important practical significance.

In the seismic prospecting of oil gas, the object of seismic data interpretation carries out the description of underground structure explanation and stratum and rock signature.Extraction and analysis and waveform separation technology that one of most effective method of these information is exactly seismic properties feature is obtained from seismic data.

Along with the development of scientific and technological level and improving constantly seismic data acquisition technology, make the earthquake information that comprises in seismic signal abundanter, and wherein many useful earthquake informations only can detect out by the observation of naked eyes, must be extracted it, be analyzed by seismic data processing technology and computer technology, and by certain mathematical method, the geologic feature of these earthquake informations is explained.For the geological data of existing collection, current waveform separation technology mainly based on unsupervised segmentation algorithm, particularly based on artificial neural network theories, as business software OpendTect, Petrel and Stratimgic etc.The method has higher holding capacity to containing noisy seismic signal data, also has the classification capacity without training sample data.But, also some shortcoming is there is: first without supervision waveform separation algorithm, the well logging information with important references meaning is ignored in geological data without supervision waveform separation algorithm, taxonomic structure is just based on internal distribution and the statistical nature of earthquake work area data, with actual conditions contact undertighten, classification results is not accurately and rationally; In addition, artificial neural network have that the computation complexity of network training is higher, some parameter arrange need we have some prior imformations and in some cases neural network may be absorbed in the problems such as local optimum.

Therefore, the difficulty based on the waveform separation problem of 3-D seismics signal is mainly manifested in following several aspect:

(1) along with the raising of seismic data acquisition technology, the earthquake information comprised in 3-D seismics signal is more and more abundanter, and how the effective information extracted in seismic signal carries out waveform separation process is exactly a large difficult point in geological data waveform separation problem.Rationally and effective seismic properties feature, the accuracy of waveform separation can be improved well, improve lithology prediction, Sand-body Prediction, fractured reservoirs prediction and subtle reservoir formation forecasting reliability.

(2) sorter model how building superior performance is another the large difficult point analyzed based on the waveform separation of 3-D seismics signal.Select suitable sorting algorithm, the sorter model that classification performance is superior can not only be set up, also can improve waveform separation analysis efficiency, so select suitable sorting algorithm to be emphasis in waveform separation.

(3) for the feature of geological data, data can be mixed into a certain amount of noise in gatherer process.Due to adding of random noise, serious impact can be caused on the waveform separation result of seismic signal, reduce the accuracy of classification results, and then lithology prediction, Sand-body Prediction, fractured reservoirs prediction and subtle reservoir formation forecasting reliability can be affected.

In the waveform separation problem of seismic signal, have more implementation both at home and abroad at present, in these methods, major part all carries out waveform separation analysis based on unsupervised segmentation thought to seismic signal.Below three kinds of implementations wherein: (1) is based on the waveform classification of self-organizing map neural network: first the method carrys out Modling model track data to the seismic signal data of the objective interval extracted, namely train the output layer node of self organizing neural network to set up disaggregated model by seismic trace sample data, and then utilize model trace to carry out classifying and dividing to seismic signal.(2) based on the waveform classification of hierarchical cluster: first the method builds cluster spanning tree to the seismic signal data of the objective interval extracted, and then carries out classifying and dividing to cluster spanning tree to seismic signal.(3) based on the waveform classification of mixed Gaussian probability model: the method is first according to the seismic signature of the objective interval extracted, utilize probability statistical analysis theory to set up mixed Gaussian probability model, and then utilize mixed Gaussian probability model to carry out classifying and dividing to seismic signal.

Above prior art, be obtained in the seismic data analysis of reality and apply significantly, but these methods all process based on unsupervised segmentation thought, have ignored in earthquake data acquisition the log data information with important references meaning so to a great extent, cause very large difficulty to follow-up seismic interpretation; And the shortcomings such as it is high that these algorithms have computation complexity, and operational efficiency is lower.

Summary of the invention

The present invention is directed to the difficult point of seismic signal waveform sorting technique and the existing shortcoming without supervision waveform separation technology, proposing has supervision waveform classification in a kind of 3-D seismics signal, for solving the difficult point in the existing shortcoming without supervision waveform classification and waveform separation.

Technical scheme of the present invention is: have supervision waveform classification in a kind of 3-D seismics signal, it is characterized in that, comprises the following steps:

A, data prediction: comprise Noise reducing of data process, extract objective interval data and training sample and tag along sort are set up to well logging data analysis;

B, feature selecting: analyze the training sample and tag along sort data thereof that extract, in selection training sample, feature and the maximally related a certain amount of feature of tag along sort are as character subset, and the character subset after preferred in objective interval data and training sample is as the input amendment set of class indication;

C, class indication: the disaggregated model setting up support vector machine carries out class indication to seismic signal;

Further, in above-mentioned steps a, described denoise processing method is as follows:

A11, input geological data;

A12, judge whether the maximum iteration time reaching noise reduction process, if then terminate noise reduction process process, otherwise perform step a13;

A13, calculating earthquake number strong point Grad in three directions also calculate the structure tensor of this data point further;

A14, three-dimensional Gaussian low-pass filter is utilized to carry out filtering process to described structure tensor;

A15, structure diffusion matrix and discontinuous factor;

A16, according to diffusion equation to data filtering.

Further, in above-mentioned steps a, described extraction objective interval data method is as follows:

A21, single layer bit analyzing and processing is done to the data of step a16 gained;

A22, two-layer interdigit analyzing and processing is done to the data of step a16 gained.

Further, described step a21 is further comprising the steps of:

A211, really timing window and analysis layer position;

A212, extract along window data during layer;

A213, acquisition objective interval data.

Described step a22 is further comprising the steps of:

A221, determine to analyze two-layer position;

A222, extract along inter-layer data;

A223, acquisition objective interval data.

Further, in above-mentioned steps a, describedly to the concrete grammar that well logging data analysis sets up training sample and tag along sort be: according to log data in geological data work area, manual sort's expression is carried out in well logging in work area, set up the tag along sort of training sample, and the log data extracting in objective interval well logging and well lie position is as training sample, for feature selecting and class indication.

Further, in above-mentioned steps b, feature selecting adopts genetic algorithm, and the training sample set wherein extracting well logging and well lie is combined into T, and corresponding tag along sort is y, and feature selecting number is k and maximum iteration time is maxI, specifically comprises the following steps:

B1, according to gene code rule, obtain the initialization population of genetic algorithm;

B2, basis fitness function calculates the fitness of population at individual, wherein x _i(i=1,2 ..., xth in training sample set T k) selected by population at individual _idimensional feature component, y is the tag along sort of corresponding well logging; MIx _iy represents the xth of log data _ithe mutual information of dimensional feature component and tag along sort y; MIx _ix _jrepresent the x of log data _icharacteristic component and characteristic component x _jbetween mutual information;

B3, judge whether to reach maximum iteration time maxI or the continuous optimum solutions that keep for five times are constant, if so, then export current optimal feature subset, otherwise perform step b4;

B4, according to population at individual fitness perform select operation;

B5, execution interlace operation;

B6, execution mutation operation;

B7, return b2 operation;

B8, output characteristic subset.

Further, in the process of above-mentioned genetic algorithm process, the gene code of step b1 adopts 0,1} binary string represents, and which position in binary string just to represent in data prediction extract seismic channel data u _i={ v _i1, v _i2..., v _iMand log data w _i={ x _i1, x _i2..., x _iMin which dimensional feature of feature space, 0 represents that this characteristic component is not selected, and 1 represents and selects this characteristic component, and M is u _ior w _iin data amount check.

Further, in above-mentioned steps c, use support vector machine to set up disaggregated model, and class indication is carried out to objective interval data, finally make waveform separation division figure.

The detailed process of above-mentioned steps c is:

C1, to the process of character subset data normalization;

C2,10 folding cross-validation methods are used to receive the best SVM parameters of rope;

C3, set up svm classifier model;

C4, class indication is carried out to the objective interval data after feature selecting;

C5, drawing waveforms classifying and dividing figure.

Beneficial effect of the present invention: of the present invention have supervision waveform classification based on 3-D seismics signal data and log data information, by genetic algorithm, the attributive character extracted is optimized, utilize svm classifier algorithm, analyzed 3-D seismics objective interval data are carried out waveform separation division, identify different seismic facies, and then reliably support for latter earthquake data interpretation provides, improve the reliability to lithology prediction, Sand-body Prediction, fractured reservoirs prediction and subtle reservoir formation prediction etc.Meanwhile, utilize well logging and well lie information to set up training sample, in waveform separation, be applied to the well logging information with important references meaning like this, improve nicety of grading, reduce the difficulty of later stage explanation; In feature selecting process, introduce genetic algorithm, reduce the redundance of feature, reduce the complexity of disaggregated model design, improve classification effectiveness, and reduce redundance feature to the impact of classification results simultaneously, improve classify accuracy to a certain extent; In class indication process, the mode identification method algorithm of support vector machine introducing the Corpus--based Method theories of learning carries out Classification and Identification.Compared with supervising waveform separation algorithm with nothing: owing to employing the well logging information that there is important references and be worth, improve the accuracy of waveform separation, improve lithology prediction, Sand-body Prediction, fractured reservoirs prediction and subtle reservoir formation forecasting reliability, for follow-up seismic interpretation provides great convenience; Relative to only comparing with SVM design category device, adding genetic algorithm and carrying out feature selecting, reducing the design complexities of SVM classifier, thus improve waveform separation treatment effeciency.

Accompanying drawing explanation

The evolution cycle schematic diagram of Fig. 1 genetic algorithm;

Fig. 2 is the method flow diagram of the embodiment of the present invention;

Fig. 3 is structure directing filtering process flow diagram in the scheme of the embodiment of the present invention;

Fig. 4 extracts objective interval data flowchart in the scheme of the embodiment of the present invention;

Fig. 5 is that in the scheme of the embodiment of the present invention, character subset extracts process flow diagram;

Fig. 6 sets up disaggregated model process flow diagram in the scheme of the embodiment of the present invention;

Fig. 7 is SOM waveform separation division figure;

Fig. 8 is SVM waveform separation division figure;

Fig. 9 is the waveform separation division figure of the embodiment of the present invention.

Embodiment

Specifically describe below in conjunction with accompanying drawing pair multiple specific embodiments identical with the principle of the invention, to promote the understanding to the principle of the invention.

For the existing conventional shortcoming of 3-D seismics signal waveform sorting technique and the feature of waveform separation itself, what the present invention proposed has supervision waveform separation scheme mainly to comprise data prediction, the characteristic optimization based on GA algorithm and the Classification and Identification three parts based on svm classifier algorithm.General flow chart is as shown in Figure 2: have supervision waveform classification in a kind of 3-D seismics signal of the present embodiment, comprises step: a, data prediction: comprise Noise reducing of data process, extract objective interval data and set up training sample and tag along sort to well logging data analysis; B, feature selecting: analyze the training sample and tag along sort data thereof that extract, to select in training sample feature and the maximally related a certain amount of feature of tag along sort as character subset, and using the input amendment set of the character subset after preferred in objective interval data and training sample as class indication; C, class indication: the disaggregated model setting up support vector machine carries out class indication to seismic signal.In step a, described denoise processing method is as follows: a11, input geological data; A12, judge whether the maximum iteration time reaching noise reduction process, if then terminate noise reduction process process, otherwise perform step a13; A13, calculating earthquake number strong point Grad in three directions also calculate the structure tensor of this data point further; A14, three-dimensional Gaussian low-pass filter is utilized to carry out filtering process to described structure tensor; A15, structure diffusion matrix and discontinuous factor; A16, according to diffusion equation to data filtering.Extract objective interval data method as follows: a21, single layer bit analyzing and processing is done to the data of step a16 gained; A22, two-layer interdigit analyzing and processing is done to the data of step a16 gained.Step a21 comprises the following steps: a211, really timing window and analysis layer position; A212, extract along window data during layer; A213, acquisition objective interval data.Step a22 comprises the following steps: a221, determine to analyze two-layer position; A222, extract along inter-layer data; A223, acquisition objective interval data.

In step a, concrete grammar well logging data analysis being set up to training sample and tag along sort is: according to log data in geological data work area, manual sort's expression is carried out in well logging in work area, set up the tag along sort of training sample, and the log data extracting in objective interval well logging and well lie position is as training sample, for feature selecting and class indication.

In step b, feature selecting adopts genetic algorithm, the training sample set wherein extracting well logging and well lie is combined into T, corresponding tag along sort is y, feature selecting number is k and maximum iteration time is maxI, specifically comprise the following steps: b1, according to gene code rule, obtain the initialization population of genetic algorithm; B2, basis fitness function calculates the fitness of population at individual, wherein x _i(i=1,2 ..., xth in training sample set T k) selected by population at individual _idimensional feature component, y is the tag along sort of corresponding well logging; MIx _iy represents the xth of log data _ithe mutual information of dimensional feature component and tag along sort y; MIx _ix _jrepresent the x of log data _icharacteristic component and characteristic component x _jbetween mutual information; B3, judge whether to reach maximum iteration time maxI or the continuous optimum solutions that keep for five times are constant, if so, then export current optimal feature subset, otherwise perform step b4; B4, according to population at individual fitness perform select operation; B5, execution interlace operation; B6, execution mutation operation; B7, return b2 operation; B8, output characteristic subset.Further, in the process of above-mentioned genetic algorithm process, the gene code of step b1 adopts 0,1} binary string represents, and which position in binary string just to represent in data prediction extract seismic channel data u _i={ v _i1, v _i2..., v _iMand log data w _i={ x _i1, x _i2..., x _iMin which dimensional feature of feature space, 0 represents that this characteristic component is not selected, and 1 represents and selects this characteristic component, and M is u _ior w _iin data amount check.

Above-mentioned step c uses support vector machine to set up disaggregated model, and carries out class indication to objective interval data, finally makes waveform separation division figure.Detailed process is: c1, to the process of character subset data normalization; C2,10 folding cross-validation methods are used to receive the best SVM parameters of rope; C3, set up svm classifier model; C4, class indication is carried out to the objective interval data after feature selecting; C5, drawing waveforms classifying and dividing figure.

Below in conjunction with principle of work and effect, the scheme to above-described embodiment is further described.

A, data prediction: carrying out pretreated fundamental purpose to 3D seismic data data is that its main working process mainly comprises in order to feature selecting and class indication are prepared: the extraction of Noise reducing of data process, objective interval data and training sample and label three part are set up to the analysis of log data.Due to the uncertainty of the physical limitation of system for acquiring seismic data equipment, the limitation of transmission medium and environmental baseline, gather the geological data obtained and can contain a certain amount of noise information, and due to can the detailed information such as tomography and geological objects boundary be contained in geologic structure.So should be noted that oriention analysis, rim detection and edge keep guiding smothing filtering three main points in the filtering of geological data, therefore choice structure Steerable filter method carries out noise reduction to data, improve signal to noise ratio (S/N ratio), and keep or enhance original geologic structure.Its work for the treatment of flow process as shown in Figure 3.Extract objective interval data: according to the layer position that will analyze, extract the objective interval data that we will carry out waveform separation analysis.Extract the overall procedure of objective interval data as shown in Figure 4.Analyze log data, set up training sample and label: according to log analysis data in geological data work area, well logging in work area is carried out manual sort's expression by us, set up training sample label, and the data extracting in objective interval well logging and well lie position are as training sample, with feature selecting later and class indication.

B, feature selecting: the fundamental purpose of feature selecting is the feature in order to optimize in objective interval data, by analyzing the training sample and tag along sort data thereof that extract, select feature and the maximally related a certain amount of feature of tag along sort in sample, so not only can reduce the complexity of disaggregated model design below, efficiency is provided, also can improve nicety of grading to a certain extent, reduce the impact of irrelevant feature on classification results.The feature optimization algorithm that this programme is selected is genetic algorithm, and it is input as the well logging of extraction and the training sample set T of well lie, and corresponding tag along sort y, feature selecting number k and maximum iteration time maxI.And export with the character subset preferably.The process flow diagram of feature selecting as shown in Figure 5.

C, class indication: after the training sample set formed objective interval data and well logging carries out feature selecting, carry out most important work and namely set up disaggregated model class indication is carried out to seismic signal.Few for log data in 3-D seismics work area, there is the feature of little training sample, and seismic signal waveform classification is generally Nonlinear Classification problem, and combine support vector machine and have and better must solve the small sample in the past perplexing a lot of learning method, non-linear, the ability of crossing the practical problemss such as study, high dimension, local minimum point, this programme have selected support vector machine to set up disaggregated model, and class indication is carried out to objective interval data, finally make waveform separation division figure.The idiographic flow of this part process as shown in Figure 6.

In the solution of the present invention, the some algorithm principle related to is as follows:

Structure directing filtering

For the structure directing filtering technique of 3D seismic data filtering ^[7]be on the anisotropic diffusion equation basis of diffusion tensor matrices, structure reflects a kind of edge preserving smooth filter algorithm that the structure tensor of 3-D view partial structurtes realizes.In 3-D view process, we definition structure tensor S is:

$S=\left[\begin{array}{ccc}\frac{\∂u}{\∂x}\frac{\∂u}{\∂x}& \frac{\∂u}{\∂x}\frac{\∂u}{\∂y}& \frac{\∂u}{\∂x}\frac{\∂u}{\∂z}\\ \frac{\∂u}{\∂x}\frac{\∂u}{\∂y}& \frac{\∂u}{\∂y}\frac{\∂u}{\∂y}& \frac{\∂u}{\∂y}\frac{\∂u}{\∂u}\\ \frac{\∂u}{\∂x}\frac{\∂u}{\∂z}& \frac{\∂u}{\∂y}\frac{\∂u}{\∂z}& \frac{\∂u}{\∂z}\frac{\∂u}{\∂z}\end{array}\right]---(1-1)$

In above formula: u (x, y, z) is input original three-dimensional image; In structure tensor matrix represent the partial derivative of x direction, y direction and these three dimensions of z direction respectively.Can there is certain instability at the structure tensor of some marginal texture present positions of 3-D seismics image, so itself and small scale gaussian kernel can be carried out convolution, structure tensor new like this can have certain robustness, and this process can be expressed as follows:

S _σ＝S(G _σ*u(t))(1-2)

Structure tensor matrix S in above formula _σa positive semidefinite matrix, so by carrying out Eigenvalues Decomposition to this matrix, thus can obtain:

$S_{\mathrm{\σ}}=\left(\begin{array}{ccc}{v}_1& v_2& v_3\end{array}\right)\left[\begin{array}{ccc}{\mathrm{\λ}}_1& 0& 0\\ 0& {\mathrm{\λ}}_2& 0\\ 0& 0& {\mathrm{\λ}}_3\end{array}\right]{\left(\begin{array}{ccc}{v}_1& v_2& v_3\end{array}\right)}^T---(1-3)$

In above formula: λ ₁, λ ₂, λ ₃for the structure tensor matrix S calculated _σeigenwert on three directions, and λ ₁>=λ ₂>=λ ₃>=0; v ₁, v ₂, v ₃, be respectively and three eigenvalue λ ₁, λ ₂, λ ₃corresponding proper vector.In the process of diffusing filter, in order to Protect edge information structure and compacting noise, we by with gradient towards orthogonal two proper vector v ₂, v ₃each component v ₂₁, v ₂₂, v ₃₁, v ₃₂construct diffusion tensor D, be expressed as:

$D=\left[\begin{array}{ccc}{v}_{21}{v}_{21}+v_{31}{v}_{31}& v_{21}{v}_{22}+v_{31}{v}_{32}& v_{21}{v}_{23}+v_{31}{v}_{33}\\ v_{21}{v}_{22}+v_{31}{v}_{32}& v_{22}{v}_{22}+v_{32}{v}_{32}& v_{22}{v}_{23}+v_{32}{v}_{33}\\ v_{21}{v}_{23}+v_{31}{v}_{33}& v_{22}{v}_{23}+v_{32}{v}_{33}& v_{23}{v}_{23}+v_{33}{v}_{33}\end{array}\right]---(1-4)$

We just can show that structure directing filtering equations is like this:

$\begin{array}{cc}u(t+1)=u\left(t\right)+c\·div\left(D\right(G_{\mathrm{\σ}}*u\left(t\right))\▿u)& t\≥0\end{array}---(1-5)$

In the waveform separation application of seismic signal, the marginal texture of underground medium and discontinuous information are very important, so we construct a discontinuous factor f at this, and this factor are applied in structure directing filtering.

$f=\frac{diag\left(S^{0}S\right)}{diag\left(S^0\right)diag\left(S\right)}---(1-6)$

In formula: diag () represents the element summation on the main diagonal angle of matrix; S ⁰represent the structure tensor of input original image; S represents the structure tensor of circulation at present; Wherein the span of discontinuous factor f is [0,1], and time near marginal texture and intermittent configuration, f value is tending towards 0; But during away from marginal texture, f value is tending towards 1.After introducing discontinuous factor, structure directing filtering algorithm can be expressed as:

$\begin{array}{cc}u(t+1)=u\left(t\right)+c\·div(f\·D(G_{\mathrm{\σ}}*u\left(t\right))\▿u)& t\≥0\end{array}---(1-7)$

Genetic algorithm

Genetic algorithm is a kind of based on Darwinian natural selection and theory of heredity, by the efficient global optimization approach that survival of the fittest rule in biological evolution process combines with the random exchanging mechanism of information chromosomal in group.Complete genetic algorithm mainly comprises the judgement etc. of gene code, initialization of population, selection operation, interlace operation, mutation operation and termination condition.Complete genetic algorithm evolution cycle is as shown in Figure 1:

Selection, crossover and mutation operation are referred to as genetic manipulation, and for the setting that the most important thing is fitness function in selection operation, and crossover and mutation is the main operation of Random Effect ideal adaptation angle value; Termination condition is the criterion judging that genetic algorithm could stop, general adopt maximum heredity time number of operations and maximum iteration time or the optimum solution that calculates after genetic manipulation several times constant.

Support vector machine

Support vector machine (SupportVectorMachine) is a kind of machine learning method of the Corpus--based Method theories of learning grown up middle 1990s.It improves learning machine generalization ability by seeking structuring least risk, realizes minimizing of empiric risk and fiducial range, thus reach when statistical sample amount is less, also can obtain the object of good statistical law.Be widely used in fields such as recognition of face, genetic test, text identification and network flow classification at present.

For non-linear SVM model, support vector machine be by a mapping phi () by original data-mapping in new space, then do Linear SVM sort operation.Just become by the objective function that lagrange's method of multipliers is proper like this:

$\underset{\mathrm{\&alpha;}}{\max }L\left(\mathrm{\&alpha;}\right)=\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\&Sigma;}}\underset{j=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_j<\mathrm{\&phi;}\left(x_i\right)\&CenterDot;\mathrm{\&phi;}\left(x_j\right)>---(1-8)$

$\begin{array}{cc}subject& to:0\&le;{\mathrm{\&alpha;}}_i\&le;C,i=1,\mathrm{...}n;\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i{y}_i=0\end{array}$

By the < φ (x in formula _i) φ (x _j) a > inner product operation kernel function κ (x _i, x _j) replace, then the objective function that just can obtain non-linear SVM is:

$\underset{\mathrm{\&alpha;}}{\max }L\left(\mathrm{\&alpha;}\right)=\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\&Sigma;}}\underset{j=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_j\mathrm{\&kappa;}(x_i,x_j)$

$subjectto:0\&le;{\mathrm{\&alpha;}}_i\&le;C,i=1,\mathrm{...},n;\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i{y}_i=0---(1-9)$

And the optimization lineoid (classification function) of non-linear SVM is:

$f\left(x\right)=\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i{y}_i\mathrm{\&kappa;}(x_i,x)+b=\underset{i\&Element;sv}{\&Sigma;}{\mathrm{\&alpha;}}_i{y}_i\mathrm{\&kappa;}(x_i,x)+b---(1-10)$

The kernel function commonly used in non-linear SVM mainly contains polynomial (polynomial expression) kernel function, RBF (radial basis) kernel function and sigmoid kernel function, and functional form is such as formula 1-11,1-12,1-13.In order to obtain higher nicety of grading, the parameter of kernel function needs correct setting.

(1) Polynomial kernel function:

κ(x _i,x _j)＝(1+x _i·x _j) ^d(1-11)

(2) RBF kernel function:

$\mathrm{\&kappa;}(x_i,x_j)=\exp (-\frac{\left|\right|x_i-x_j|{|}^2}{2{\mathrm{\&sigma;}}^2})---(1-12)$

(3) Sigmoid kernel function:

κ(x _i,x _j)＝tanh(kx _i·x _j-δ)(1-13)

In order to evaluate the performance of this programme, we carry out application simulation to actual seismic work area data.In experiment simulation, we compare conventional self-organizing map neural network algorithm respectively and only use the classification performance of SVM algorithm process.

Emulation: have chosen the demonstration data F3 work area 3D seismic data of business software Opendtect as the real data that will analyze.F3 is the block that the North Sea is positioned at Holland's part.The F3 work area scope that we select is No. Inline: 104-696, No. Xline: 304-1246, and wherein log well and always have 20 mouthfuls.The layer position of carrying out seismic facies analysis is MFS4.The objective interval data set that table 1 extracts for us.

The objective interval data set that table 1 extracts

Utilize SOM algorithm, SVM algorithm and this programme GA+SVM algorithm to set up disaggregated model to log data, be wherein extracted 20 characteristic components of the lonely data centralization of object Cen in this programme by genetic algorithm process, result is as shown in table 2.

The results contrast of disaggregated model set up by table 2 log data collection

Carry out class indication with the disaggregated model that three kinds of algorithms are set up to work area earthquake data set respectively, the waveform separation division figure made is respectively as shown in Fig. 7, Fig. 8, Fig. 9.

In the present invention program, following word is the special of this area or universal word, and the meaning of its representative is as follows: layer position, refers to a certain ad-hoc location in succession of strata, the layer position on stratum can be the boundary line of stratigraphic unit, also can be the reference lamina etc. belonging to a certain specific epoch; Tomography, earth formation reaches some strength because of stressed and break, and has the structure of obvious relative movement to be called tomography along the plane of fracture; Interpolation, utilizes known point to calculate the process of unknown point; Objective interval, refers to the time window data or the inter-layer data of two layer interdigits that will analyze along single layer position will analyzed in 3-D seismics signal data.

Those of ordinary skill in the art will appreciate that, embodiment described here is to help reader understanding's implementation method of the present invention, should be understood to that protection scope of the present invention is not limited to so special statement and embodiment.Those of ordinary skill in the art can make various other various concrete distortion and combination of not departing from essence of the present invention according to these technology enlightenment disclosed by the invention, and these distortion and combination are still in protection scope of the present invention.

Claims (9)

1. there is a supervision waveform classification in 3-D seismics signal, it is characterized in that, comprise the following steps:

A, data prediction: comprise Noise reducing of data process, extract objective interval data and training sample and tag along sort are set up to well logging data analysis;

B, feature selecting: analyze the training sample and tag along sort data thereof that extract, in selection training sample, feature and the maximally related a certain amount of feature of tag along sort are as character subset, and the character subset after preferred in objective interval data and training sample is as the input amendment set of class indication;

In described step b, feature selecting adopts genetic algorithm, and the training sample set wherein extracting well logging and well lie is combined into T, and corresponding tag along sort is y, and feature selecting number is k and maximum iteration time is maxI, specifically comprises the following steps:

B1, according to gene code rule, obtain the initialization population of genetic algorithm;

B2, basis fitness function calculates the fitness of population at individual, wherein x _i, i=1,2 ..., k, xth in the training sample set T selected by population at individual _idimensional feature component, y is the tag along sort of corresponding well logging; MIx _iy represents the xth of log data _ithe mutual information of dimensional feature component and tag along sort y; MIx _ix _jrepresent the x of log data _icharacteristic component and characteristic component x _jbetween mutual information;

B3, judge whether to reach maximum iteration time maxI or the continuous optimum solutions that keep for five times are constant, if so, then export current optimal feature subset, otherwise perform step b4;

B4, according to population at individual fitness perform select operation;

B5, execution interlace operation;

B6, execution mutation operation;

B7, return b2 operation;

B8, output characteristic subset;

C, class indication: the disaggregated model setting up support vector machine carries out class indication to seismic signal.

2. have supervision waveform classification in a kind of 3-D seismics signal according to claim 1, it is characterized in that, in described step a, described denoise processing method is as follows:

A11, input geological data;

A12, judge whether the maximum iteration time reaching noise reduction process, if then terminate noise reduction process process, otherwise perform step a13;

A13, calculating earthquake number strong point Grad in three directions also calculate the structure tensor of this data point further;

A14, three-dimensional Gaussian low-pass filter is utilized to carry out filtering process to described structure tensor;

A15, structure diffusion matrix and discontinuous factor;

A16, according to diffusion equation to data filtering.

3. have supervision waveform classification in a kind of 3-D seismics signal according to claim 2, it is characterized in that, in described step a, described extraction objective interval data method is as follows:

A21, single layer bit analyzing and processing is done to the data of step a16 gained;

A22, two-layer interdigit analyzing and processing is done to the data of step a16 gained.

4. have supervision waveform classification in a kind of 3-D seismics signal according to claim 3, it is characterized in that, described step a21 is further comprising the steps of:

A211, really timing window and analysis layer position;

A212, extract along window data during layer;

A213, acquisition objective interval data.

5. have supervision waveform classification in a kind of 3-D seismics signal according to claim 3, it is characterized in that, described step a22 is further comprising the steps of:

A221, determine to analyze two-layer position;

A222, extract along inter-layer data;

A223, acquisition objective interval data.

6. in a kind of 3-D seismics signal according to claim 1, there is supervision waveform classification, it is characterized in that, in described step a, describedly to the concrete grammar that well logging data analysis sets up training sample and tag along sort be: according to log data in geological data work area, manual sort's expression is carried out in well logging in work area, set up the tag along sort of training sample, and the log data extracting in objective interval well logging and well lie position is as training sample, for feature selecting and class indication.

7. in a kind of 3-D seismics signal according to claim 1, there is supervision waveform classification, it is characterized in that, in the process of described genetic algorithm process, the gene code of step b1 adopts { 0,1} binary string represents, which position in binary string just to represent in data prediction extract seismic channel data u _i={ v _i1, v _i2..., v _iMand log data w _i={ x _i1, x _i2..., x _iMin which dimensional feature of feature space, 0 represents that this characteristic component is not selected, and 1 represents and selects this characteristic component, and M is u _ior w _iin data amount check.

8. in a kind of 3-D seismics signal according to claim 1, there is supervision waveform classification, it is characterized in that, in described step c, use support vector machine to set up disaggregated model, and class indication is carried out to objective interval data, finally make waveform separation division figure.

9. have supervision waveform classification in a kind of 3-D seismics signal according to claim 1 or 8, it is characterized in that, the detailed process of described step c is:

C1, to the process of character subset data normalization;

C2,10 folding cross-validation methods are used to receive the best SVM parameters of rope;

C3, set up svm classifier model;

C4, class indication is carried out to the objective interval data after feature selecting;

C5, drawing waveforms classifying and dividing figure.