CN105259572A - Seismic facies calculation method based on non-linear automatic classification of multiple attribute parameters of earthquake - Google Patents

Abstract

The invention discloses a seismic facies calculation method based on non-linear automatic classification of multiple attribute parameters of an earthquake. A proper initial weight value is found; and on the basis of a basic principle of a self-organizing characteristic mapping neural network, a precise seismic facies value is calculated. According to the invention, the self-organizing characteristic mapping neural network is combined with seismic facies calculation; great improvement is made when an initial weight value is selected, so that the learning efficiency is substantially improved. Besides, because the seismic attribute parameter are classified automatically by the self-organizing characteristic mapping neural network is, calculation precision of the seismic facies value is substantially improved.

Description

Based on the seismic facies computing method of seismic multi-attribute parametrical nonlinearity automatic classification

Technical field

The present invention relates to a kind of seismic facies computing method, particularly relate to a kind of seismic facies computing method based on seismic multi-attribute parametrical nonlinearity automatic classification.

Background technology

This noun of seismic facies derives from sedimentary facies, and Sloss (1962) thinks " System Domain when being the generation of certain rock stratum mutually and the summation of material exhibits thereof ".Therefore, seismic facies can be understood as the summation that sedimentary facies shows on seismic section.Mitchum (1977) thinks " seismic facies unit is the unit that can chart, and the 3-D seismics reflectance signature of this unit is adjacent unit difference ".Its conventional seismologic parameter (as reflection configuration, amplitude, continuity, frequency and interval velocity) is different from adjacent cells, and its representative produces its reflection sedimental certain lithological combination, bedding and deposition characteristics.

Represent formation beds according to the reflection horizon in the known seismic facies of Brown (1980) understanding to seismic facies, have the unconformity surface of STRATIGRAPHIC SIGNIFICANCE or possible fluid contact level.In a word, seismic facies is the seismic reflection unit in the three dimensions limited by specific seismic reflection parameter, and it is the seismic response of particular deposition phase or geologic body.Therefore, seismic facies is a concentrated expression of underground geologic bodies.Seismic facies analysis technique is then " according to seismic data interpretation environmental background and petrofacies ", and its object carries out regional stratum explanation exactly, determines sedimentary system, and lithofacies characteristics and explanation depositional history, finally predict favourable oil generation district and reservoir facies belt.

Seismic facies analysis is a kind of geological method utilizing seismic data to carry out geologic interpretation grown up late 1970s.Along with developing by leaps and bounds of seismic exploration technique and deepening continuously of seismic facies analysis work, research method gets more and more.Traditional seismic facies analysis method is described by naked-eye observation, is commonly called as " metoposcopy ", has very large subjectivity and uncertainty.Along with improving constantly of seismic data acquisition technology, make earthquake information that seismic section comprises abundanter, and wherein many useful earthquake informations are depended naked eyes alone and are observed on section and can not detecting out, must be extracted, analyze by seismic data processing technology and computer technology to it.Therefore there is earthquake many genus Parameter analysis method, different according to its main operational adopted, the multiple branch of principal component analysis again, clustering methodology, analysis of neural network method, genetic algorithm etc.The popular seismic facies analysis method of Abroad in Recent Years mainly contains the methods such as waveform separation method, seismic geomorphology imaging method, seismic properties Feature Mapping method, Time-frequency Analysis.

(1) waveform separation seismic facies analysis

Traditional seismic facies analysis method utilizes that seismic reflection amplitude is strong and weak, continuity is fine or not, frequency height, seismic facies map such as establishment such as attribute such as reflection geometric shape and inner structure etc., carries out the conversion of seismic facies to sedimentary facies between well.The artificial property of this method is comparatively large, has certain uncertainty, and its precision is difficult to meet Subtle reservoir exploration and develops the needs studied mutually deposition (micro-).And utilize seismic waveshape feature to carry out seismic facies classification can to eliminate the limitation utilizing single Seismic Attribute Parameters analysis to bring, can meet the exploration stage to sedimentary micro and subtle reservoir research demand, be the explanation aid of a kind of advanced person solving deposition (micro-) phase between well, have a good application prospect in three-dimensional work area.

Waveform separation seismic facies analysis often loses two class essential informations owing to applying traditional Discussion of Earthquake Attribute Technologies such as seismic properties calculating, inverting: the entire change of seismic signal and the rule of change thereof.Waveform separation seismic facies analysis adopts neural network and mode identification technology, to the actual seismic data track in a certain interval by its waveform character of trace comparison and seismic properties feature, carefully portray its horizontal change, thus obtain seismic anomaly planar distribution, i.e. Seismic waveform classification figure.

Seismic waveform classification analysis be mainly applicable to parallel construction, zone thickness change less, construct fairly simple area, because in these places along type formation down or when up opening window carry out waveform separation analyze time, do not pass through seismic trace in screen work when to remain on etc., effect is better.And for the area that zone thickness changes greatly, as the wedge-shaped strata of basin edge, due to not uniform thickness, unified when opening window usually cause wearing the diachronous phenomenon that axle causes, cause research precise decreasing.In addition, in Complicated structure area, because influence factor is many, preferably first carry out target preserved amplitude processing, to obtain good effect.Representational seismic facies analysis system, namely Stratimagic seismic facies analysis software carries out seismic facies analysis.

Utilize Seismic waveform classification to carry out seismic phase analysis quantitatively and have following 3 outstanding features: one is without the need to well data; Two is can carry out rapid scanning to whole data rapidly, carries out thinner research work again after determining to have the object region of off-note; Three is compared with seismic phase analysis traditionally, enhances quantitative and objectivity.But there is multi-solution when utilizing seismic facies to study and predict sand body, just can correctly must explain in conjunction with drilling geology achievement.

Waveform separation method distinguishes different sedimentary facies belt with waveform change, be applicable in a big way and the division of large sedimentary facies belt, its major influence factors is the determination of number of categories and the quality of seismic data, lacks certain man-machine intervention, can bring certain difficulty to explanation.

(2) based on morphological automatically seismic phase analysis

Seismic geomorphology mainly utilizes seismic data to recover the research (Posamentier, 2001) of the morphologic characteristics under ancient depositional setting.Based in morphological automatically seismic phase analysis, seismic facies is defined as under normal earthquake sampling rate (as a 4ms) condition, be limited in the seismic description of sedimentary facies in earthquake thin layer, this makes seismic facies analysis be different from the seismic facies analysis of traditional more thick layer deposition sequence.Be actually based on morphological seismic facies analysis and carry out on seismic slice (rock stratum section).Normally utilize the method for neural network, using morphologic characteristics attribute as input node, each seismic facies is for exporting.Its key finds the attribute (as size, direction, linear, curvature etc.) can stating different shape respectively, and the seismic facies of output can utilize simple Digital ID.

An advantage based on the automatically seismic phase analysis of landforms is that it has the ability of following the trail of deposition (stratum) feature in 3-D seismics body very soon, very in detail, and it is mainly realized by monitoring learning.Monitoring learning comprises experiment sedimentary facies (sand smeller identifies according to visual inspection) and becomes by random numeral (1,2,3 etc.) the phase class marked, if result is satisfied, the criterion acquiring (acquisition) by experiment can be applicable to whole geological data group, realizes the sedimentary facies drawing of robotization.

Based on morphological seismic facies analysis method, the sedimentary facies close for Seismic reflection character can be distinguished preferably.

Seismic Geomorphology method is actually the method utilizing neural network on seismic slice (rock stratum section), and using morphologic characteristics attribute (as size, direction, linear, curvature etc.) as input node, each seismic facies is for exporting.Based on morphological seismic facies analysis method, the sedimentary facies close for Seismic reflection character can be distinguished preferably.

(3) seismic structural best property of attribute mapping method seismic facies analysis

Seismic structural attribute is proposed by Love and Simaan (1984) the earliest, and it mainly supposes that the seismic signature of interval reflects the geologic media of this section, also needs in addition to have higher signal to noise ratio (S/N ratio) and standardized stratigraphic model.Seismic reflection is the sonic expression that geologic feature is reflected by wavelet, is the concentrated expression of all wavelet reflections in a little region in three dimensions.Seismic structural attribute (seismictexture) is exactly the instruction to spatial variations such as amplitude, thickness and stratal configurations between neighboring track, and it characterizes sedimentary facies belt by seismic reflection equally.The spatial variations of seismic structural attribute needs to be realized by certain time window, time window size can select according to the size of dominant frequency and research purpose layer.

Seismic structural attribute method identifies different sedimentary facies belts by amplitude, frequency and the successional Seismic reflection character described in a little space, has higher rock physics basis, can obtain a three-dimensional seismic facies distribution by given pattern.

Utilize seismic structural attribute to draw and be better than single seismic attributes analysis mutually, because the Seismic reflection character that it can be comprehensively multiple, such as one weak amplitude bass reflex has the feature that smoothly (homogeneity) is high; The low even reflection of high frequency has the higher feature of otherness; The entropy feature of the weak amplitude of Low coherence and mixed and disorderly reflection is than strong amplitude Gao Liangao coherent reflection.The method extracts attribute by amplitude and waveform character, has higher physical basis, needs the geologic model that explanation personnel provide certain simultaneously, has good man-machine interaction than waveform separation.

(4) seismic facies analysis of time-frequency analysis technology is adopted

" three wink section " that obtain after being applied to Fourier transform in early days carry out the description of seismic facies.Due to time-frequency time the immobilizing of window, different geologic media can not be adapted to and cause seismic frequency to change, therefore occur again short time Fourier transform, Gabor transformation, Cohen class time frequency analysis etc., but these conversion again by get time window type impact.There is wavelet transformation for this reason.When wavelet transformation solves-frequently time window stationarity, window when can select large when low frequency, and select a hour window to analyze during high frequency.Have " time-frequency magnifier " effect, simultaneously again there is multiresolution features.But the method can not accurately adopt suitable " wavelet basis ", and calculate numerous and diverse.Development in recent years goes out the seismic facies analysis technique based on S-transformation.

Time-frequency analysis technology mainly carries out time-frequency convert to analyze seismic waveshape characteristic sum dynamic characteristic value.The S-conversion of current use with time and frequency for variable describes the energy density of signal or the intensity of signal.The advantage that S-transformation combines short time Fourier transform and wavelet transformation provides the Copula of time and frequency, and with time and frequency for variable describes the energy density of signal or the intensity of signal, S-transformation has higher time frequency resolution.And for the less seismic facies body of scale, time section is be difficult to identify that seismic facies parameter (particularly frequency) in its sequence is with whilst on tour situation of change.Portraying and total volume description of local detail can be done to characteristics of seismic by single track and multiple tracks S-transformation time-frequency representation.

S-transformation has higher time frequency resolution.And for the less seismic facies body of scale, time section is be difficult to identify that seismic facies parameter (particularly frequency) in its sequence is with whilst on tour situation of change.Portraying and total volume description of local detail can be done to characteristics of seismic by single track and multiple tracks S-transformation time-frequency representation.

Professor KohonenT. of Helsinki university of Finland proposes a kind of Self-Organizing Feature Maps.Kohonen thinks, when a neural network accepts extraneous input pattern, will be divided into different corresponding regions, there is different response characteristics in each region to input pattern, and this process completes automatically, and self organizing neural network is without tutor's learning network.It is by the inherent law in Automatic-searching sample and essential attribute, self-organization, adaptively modifying network parameter and structure.

The great advantage of self organizing neural network is certain relevant information storage in certain area, suitably selects controling parameters that neural network can be made to play good function, obtains the resolution and effect expected.By the relevant parameters of regulating networks, make network responsive to input, the petrofacies class number that can divide is many and careful.Simultaneously due to self organizing neural network feature, unmanned supervision type ground learning process can be realized, thus reach the object of automatic petrofacies classification.

Summary of the invention

Object of the present invention is just to provide a kind of seismic facies computing method based on seismic multi-attribute parametrical nonlinearity automatic classification solved the problem, and overcomes following defect:

1. because of stratum not uniform thickness, unified when opening window cause wearing the diachronous phenomenon that axle causes;

2. do not provide detailed Geological Mode then cannot classify to seismic facies accurately, and form the drawing of accurate sedimentary facies;

3. because the geology multi-solution of earthquake information causes to distinguish preferably for the sedimentary facies that Seismic reflection character is close;

4., for the seismic facies body that scale is less, time section is the situation being difficult to identify that seismic facies parameter, particularly frequency in its sequence change with whilst on tour;

5. sedimentary facies data cannot be provided to retrain by drilling well for without well or few well area, thus seismic facies calculating accurately can not be carried out;

6. in neural computing, initial weight is difficult to accurately reasonably determine, easily causes the phenomenon that iterative process causes calculating not restrain.

To achieve these goals, the technical solution used in the present invention is such: a kind of seismic facies computing method based on seismic multi-attribute parametrical nonlinearity automatic classification, comprise the following steps,

(1) according to geologic information, sedimentary facies kind is set up;

(2) multiple Seismic Attribute Parameters X is extracted along in window during the zone of interest of seismic trace _ij, wherein j is earthquake Taoist monastic name, j=1,2 ... N, i are attribute number;

Described Seismic Attribute Parameters comprises: coefficient of kurtosis, the coefficient of skewness, standard deviation, RMS amplitude, average energy, gross energy, average amplitude, variance, the degree of bias, kurtosis, average reflection intensity, average instantaneous frequency, average instantaneous phase, reflection strength slope and instantaneous frequency slope;

(3) according to following formula by extract Seismic Attribute Parameters standardization to eliminate dimension difference;

$\begin{array}{cc}{X_{ij}}^{*}=\frac{X_{ij}-X_{imin}}{X_{imax}-X_{imin}}& j=1,2,...,N\end{array}$

Wherein, X ^*for the Seismic Attribute Parameters after standardization, X is original property value, X _minfor attribute minimum value, X _maxfor attribute maximal value;

(4) by the Seismic Attribute Parameters after standardization, carry out dimension-reduction treatment by Karhunen-Loeve transformation, obtain the Seismic Attribute Parameters after dimensionality reduction;

(5) according to the number of parameters after dimensionality reduction, adopt with standardization after Seismic Attribute Parameters X _ij ^*corresponding contribution margin P _mas initial weight W _j;

(6) calculate seismic facies, concrete grammar is:

The first step: set up seismic facies classification computing formula:

$Y_j=\underset{i=1}{\overset{N}{\Σ}}{X}_{ij}{W_j}^{*}$

In formula: j=1,2 ... N, represents sedimentary facies kind, X _ijfor the compression parameters after input, W _jfor self organizing neural network trains the rear final weights of classification automatically;

Second step: according to self-organizing feature map Competitive Learning Algorithm, obtains the final weights W of Seismic Attribute Parameters _j*;

3rd step, by weights W final in second step _j*, bring in first step formula, finally obtain seismic facies value Y _j.

Wherein, for the seismic properties in step (2), be explained as follows in detail: " seismic properties " one word start to introduce geophysics circle in 20 century 70s.Seismic properties reflects different geology characteristic component or subset in geological data, is to portray, describe stratal configuration, lithology with the seismic signature amount of the geological informations such as transitivity.To late 1990s, QuincyChen and SteveSidney is on the basis of above-mentioned sorting technique, propose a kind of than more complete sorting technique, poststack attribute and prestack attribute are regarded as 2 stages that attribute technology develops by them, in this sense, seismic properties is divided into the large class of geometry attribute, kinematics attribute, dynamic behavior and statistics attribute 4.The Wang Yonggang of China University Of Petroleum Beijing tends to following classification: be based upon the seismic properties type on kinematics, dynamical foundation: comprise amplitude, waveform, frequency, attenuation characteristic, phase place, correlation analysis, energy, ratio etc.; Seismic properties type based on characteristics of reservoirs: comprise and characterize that bright spot, dim spot, AVO characteristic, unconformity trap or block uplift exception, oily exception, thin layer reservoir, stratigraphic break, structure are discontinuous, the seismic properties of lithologic pinch out, particular lithologic body etc.According to the feature of this classification, consider that seismic facies mainly reflects lithologic character and lithofacies variation characteristic, the seismic properties be therefore based upon on kinematics, dynamical foundation is more suitable for the calculating of seismic facies, is also that this seismic facies calculates the Main Seismic Areas property parameters selected.

This research is mainly extracted the parameter totally 15 kinds of amplitude class, frequency class, phase place class, energy class, they respectively: coefficient of kurtosis, the coefficient of skewness, standard deviation, RMS amplitude (RMS), average energy (ME), gross energy (TE), average amplitude (MA), variance (VA), the degree of bias (SA), kurtosis (KA), average reflection intensity (ARS), average instantaneous frequency (AIF), average instantaneous phase (AIP), reflection strength slope (SRS) and instantaneous frequency slope (SIF).

For the class seismic properties value in step (3), be explained as follows in detail: seismic line are come, there is N number of seismic trace (j=1, 2, N), during the research layer position of each seismic trace in window, every corresponding attribute time window in have M sampling point, a class seismic properties value can be extracted according to this M sampling point, here a class seismic properties value, refer to by under the same sampling time, the summation of sampling point in window time each, therefore N number of value is had for each generic attribute, the seismologic parameter in this class seismic properties value can be obtained from this this N number of value, 15 seismologic parameters namely in step (2), certain the present invention only lists 15 kinds, but not only only have these 15 kinds.

According to the difference in sampling time, a class seismic properties value has a lot of, what we supposed the calculating of step (3) is wherein i-th attribute.

As preferably: in step (6), second step concrete grammar is:

(a) vectorial normalized:

The weight vector W corresponding to each neuron in the current input mode vector X in self-organizing network, competition layer _j(j=1,2,3 ... .., N), be all normalized, obtain with

$\hat{X}=\frac{X}{\left|\right|X\left|\right|},{\hat{W}}_j=\frac{W_j}{\left|\right|W_j\left|\right|}$

Wherein, current input mode vector X refers to the Seismic Attribute Parameters in step (4) after dimensionality reduction, the weight vector W that in competition layer, each neuron is corresponding _jrefer to the initial weight in step (5);

B () finds triumph neuron:

Will with (j=1,2,3 ... .., m) carry out similarity system design, the most similar neuron is won;

Concrete grammar is see following rational formula

$\left|\right|\hat{X}-{\hat{W}}_{j^{*}}\left|\right|=\underset{j\∈\{1,2,...,n\}}{\min }\left\{\right||\hat{X}-{\hat{W}}_j|\left|\right\}$

$\⇒\left|\right|\hat{X}-{\hat{W}}_{j^{*}}\left|\right|=\sqrt{(\hat{X}-W_{j^{*}}){(\hat{X}-W_{j^{*}})}^T}=\sqrt{\hat{X}{\hat{X}}^T-2{\hat{W}}_{j^{*}}{\hat{X}}^T+{\hat{W}}_{j^{*}}{{\hat{W}}_{j^{*}}}^T}=\sqrt{2(1-{\hat{W}}_{j^{*}}{\hat{X}}^T)}$

$\⇒{\hat{W}}_{j^{*}}{\hat{X}}^T=\underset{j}{\max }\left({\hat{W}}_j{\hat{X}}^T\right)$

: will star weight vector Wj in corresponding with all neurons of competition layer (j=1,2,3 ..., m) carry out similarity system design, the most similar neuron is won, and weight vector is Wj*:

C () network exports and adjusts with power

Triumph neuron adjusts its weight vector W _j*, its weight vector study adjustment is as follows:

$\left\{\begin{array}{c}{W}_{j^{*}}(t+1)={\hat{W}}_{j^{*}}\left(t\right)+{\mathrm{\ΔW}}_{j^{*}}={\hat{W}}_{j^{*}}\left(t\right)+\mathrm{\α}(\hat{X}-{\hat{W}}_{j^{*}})\\ \begin{array}{cc}{W}_j(t+1)={\hat{W}}_j\left(t\right)& j\≠j^{*}\end{array}\end{array}\right.$

Wherein, α is learning rate, 0< α≤1, and α reduces along with the process of study, and the degree namely adjusted is more and more less, is tending towards cluster centre;

D () be normalized again

Weight vector after normalization, after adjustment, repeats step (1) (2) (3) and re-starts normalization, until learning rate α decays to 0, obtain final weights W _j*.

As preferably: in step (4), the concrete grammar that dimension-reduction treatment is carried out in Karhunen-Loeve transformation is:

A. parameter X after standardization is asked for _ijcovariance matrix C, its Elements C _ijbe made up of following formula:

$C_{ij}=\frac{1}{N}\left(\underset{k=1}{\overset{N}{\&Sigma;}}\right(x_{ij}-\stackrel{\&OverBar;}{x})(x_{jk}-\stackrel{\&OverBar;}{x})$

In formula: i, j=1,2 ..., M property parameters number, K=1,2 ..., N is earthquake Taoist monastic name. for property value is to the mean value in each road;

B. due to the real symmetric tridiagonal matrices that C is M × M rank, adopt Jacobi method to solve and can obtain M latent vector Z _mmwith M eigenvalue λ _m, new property parameters value can be obtained by following formula:

$X_{ik}=\underset{m=1}{\overset{N}{\&Sigma;}}{a}_{mk}{Z}_{mi}$

In formula: Z _mifor eigenvalue λ _mcorresponding latent vector value, and a _mkfor undetermined coefficient, m represents that getting N number of original property value carries out linear combination (N≤M);

C. a is asked for _mk, in fact coefficient a _mkfor X _ikwith Z _mifor related coefficient:

$a_{mk}=\underset{i=1}{\overset{N}{\&Sigma;}}{X}_{ik}{Z}_{mi}$

D. the λ asked for above _mcarry out backward sequence, and ask for the contribution margin P of its new property value _mwith accumulative contribution margin Q _m:

$P_m={\mathrm{\&lambda;}}_m/\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m$

$Q_m=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m/\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m$

E: ask for the parameter after compression:

The result of combining step c and steps d, substitutes in step b the property parameters after can obtaining compression.

Compared with prior art, the invention has the advantages that: not only overcome other several seismic facies calculate in defect, do not exist because of stratum not uniform thickness, unified when opening window cause wearing the diachronous phenomenon that axle causes; Can classify accurately to seismic facies, and form the drawing of accurate sedimentary facies; Can distinguish preferably for the sedimentary facies that Seismic reflection character is close, the synthetically seismic phase calculated can determine sedimentary facies border exactly, defines obvious instruction to the division on sedimentary facies border and scope.

In the present invention, first according to geologic information, set up sedimentary facies kind, sedimentary facies kind comprises reef beach, Sha Ba, continental shelf etc. here, is different according to geologic structure, the classification carried out.

By Seismic Attribute Parameters aims of standardization be: owing to there is obvious dimension difference between multiple seismic properties, therefore first carry out the standardization of seismic properties.

The object of Karhunen-Loeve transformation dimensionality reduction is: owing to having correlativity between multiple Seismic Attribute Parameters, there is obvious data redundancy, thus must carry out dimension-reduction treatment before seismic facies calculates.Concrete data compression dimension-reduction algorithm selects the Karhune-Loeve mapping algorithm extensively adopted at present, is called for short Karhunen-Loeve transformation.

In conjunction with the parameter after dimensionality reduction, self-organizing feature map neural network is utilized to carry out seismic facies classification to data.In self-organizing feature map neural network, the little equally distributed random number of general employing is as initial weight, but in conjunction with actual conditions of the present invention, find through large component analysis and repetition test, directly adopt above-mentioned initial weight not reach the problem of setting classification capacity.Find by analysis: random initial weight vector likely differs all larger with all learning samples vector; When certain learning sample vector is minimum with the Euclidean distance of certain output contact weight vector, according to weights amendment rule, node weight vector moves to the direction close to learning sample vector; Because similarity between sample vector is large more than the similarity of sample vector and random weight vector, subsequent samples is minimum with the weight vector Euclidean distance of greater probability and this node, thus a lot of sample is gathered is a class.

So can not, directly by the initial weight the most of the parameter after dimensionality reduction, need to carry out principal component analysis (PCA), principal component analysis (PCA) is a kind of statistical analysis technique original multiple variable being divided into a few overall target.In Karhunen-Loeve transformation, each parameter character pair value can be obtained after solution secular equation, and this eigenwert represents major component variance yields, reflect the similar features between learning sample, if similarity is too high between different classes of learning sample vector, the neural network of high-class ability just cannot be obtained.Select the proper vector value corresponding to each parameter character pair value, or signature contributions rate (each eigenwert/total characteristic value) is used as the initial weight of neural network, discovery can improve learning efficiency and nicety of grading greatly for this reason.Therefore in the present invention, be exactly according to dimensionality reduction after number of parameters, adopt with standardization after Seismic Attribute Parameters X _ij ^*corresponding contribution margin P _mas initial weight W _j.Through the Seismic Attribute Parameters X adopted and after standardization _ij ^*corresponding contribution margin P _mas initial weight W _j, learning efficiency and nicety of grading are improved all greatly, and this has a very important role for asking for seismic phase subtly.

Have found suitable initial weight, according to the ultimate principle of self-organizing feature map neural network, calculate seismic facies.

The present invention not only by self-organizing feature map neural network together with the calculations incorporated of seismic facies, made significant improvement when selecting initial weight simultaneously, learning efficiency is improved greatly, in addition, owing to passing through self-organizing feature map neural network automatic classification to Seismic Attribute Parameters, also the precision of seismic facies value is greatly improved.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the embodiment of the present invention 1;

Fig. 2 is self organizing neural network structural drawing;

Fig. 3 is organic reef beach matter illustraton of model;

Fig. 4 is organic reef beach geologic model and is just drilling section;

Fig. 5 is organic reef beach geologic model first sequence seismic facies calculating chart;

Fig. 6 is organic reef beach geologic model second sequence seismic facies calculating chart;

Fig. 7 is that organic reef beach geologic model third layer sequence seismic facies calculates;

Fig. 8 is that organic reef beach geologic model the 4th sequence seismic facies calculates;

Fig. 9 is layer position, somewhere, the South Sea synthetically seismic phase planimetric map.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

Embodiment 1: see Fig. 1, a kind of seismic facies computing method based on seismic multi-attribute parametrical nonlinearity automatic classification, comprise the following steps,

(1) according to geologic information, set up sedimentary facies kind, sedimentary facies kind here divides according to geologic information, comprises reef flat facies, Sha Baxiang, continental shelf facies etc.;

(2) multiple Seismic Attribute Parameters X is extracted along in window during the zone of interest of seismic trace _ij, wherein j is earthquake Taoist monastic name, j=1,2 ... N, i are attribute number; Described Seismic Attribute Parameters comprises: coefficient of kurtosis, the coefficient of skewness, standard deviation, RMS amplitude (RMS), average energy (ME), gross energy (TE), average amplitude (MA), variance (VA), the degree of bias (SA), kurtosis (KA), average reflection intensity (ARS), average instantaneous frequency (AIF), average instantaneous phase (AIP), reflection strength slope (SRS) and instantaneous frequency slope (SIF);

(3) according to following formula by extract Seismic Attribute Parameters standardization to eliminate dimension difference;

$\begin{array}{cc}{X_{ij}}^{*}=\frac{X_{ij}-X_{imin}}{X_{imax}-X_{imin}}& j=1,2,...,N\end{array}$

Wherein, wherein j is earthquake Taoist monastic name, and i is attribute number, X _ij ^*for the Seismic Attribute Parameters after standardization, X _ijfor original property value, X _iminfor attribute minimum value, X _imaxfor attribute maximal value;

(4) by the Seismic Attribute Parameters after standardization, carry out dimension-reduction treatment by Karhunen-Loeve transformation, obtain the parameter after dimensionality reduction;

The concrete mode of dimension-reduction treatment is:

A. parameter X after standardization is asked for _ijcovariance matrix C, its Elements C _ijbe made up of following formula:

$C_{ij}=\frac{1}{N}\left(\underset{k=1}{\overset{N}{\&Sigma;}}\right(x_{ij}-\stackrel{\&OverBar;}{x})(x_{jk}-\stackrel{\&OverBar;}{x})$

In formula: i, j=1,2 ..., M property parameters number, K=1,2 ..., N is earthquake Taoist monastic name. for property value is to the mean value in each road;

B. due to the real symmetric tridiagonal matrices that C is M × M rank, adopt Jacobi method to solve and can obtain M latent vector Z _mmwith M eigenvalue λ _m, new property parameters value can be obtained by following formula:

$X_{ik}=\underset{m=1}{\overset{N}{\&Sigma;}}{a}_{mk}{Z}_{mi}$

In formula: Z _mifor eigenvalue λ _mcorresponding latent vector value, and a _mkfor undetermined coefficient, m represents that getting N number of original property value carries out linear combination (N≤M);

C. a is asked for _mk, in fact coefficient a _mkfor X _ikwith Z _mifor related coefficient:

$a_{mk}=\underset{i=1}{\overset{N}{\&Sigma;}}{X}_{ik}{Z}_{mi};$

D. the λ asked for above _mcarry out backward sequence, and ask for the contribution margin P of its new property value _mwith accumulative contribution margin Q _m:

$P_m={\mathrm{\&lambda;}}_m/\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m$

$Q_m=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m/\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_m;$

E: ask for the parameter after compression:

The result of combining step c and steps d, substitutes in step b the property parameters after can obtaining compression.

(5) according to the number of parameters after dimensionality reduction, adopt with standardization after Seismic Attribute Parameters X _ij ^*corresponding contribution margin P _mas initial weight W _j;

(6) calculate seismic facies, concrete grammar is:

The first step: set up seismic facies classification computing formula:

$Y_j=\underset{i=1}{\overset{N}{\&Sigma;}}{X}_{ij}{W_j}^{*}$

In formula: j=1,2 ... N, represents sedimentary facies kind, X _ijfor the compression parameters after input, W _jfor self organizing neural network trains the rear final weights of classification automatically;

Second step: according to self-organizing feature map Competitive Learning Algorithm, obtains the final weights W of Seismic Attribute Parameters _j*, comprise the following steps:

(a) vectorial normalized:

The weight vector W corresponding to each neuron in the current input mode vector X in self-organizing network, competition layer _j(j=1,2,3 ... .., N), be all normalized, obtain with

$\begin{array}{cc}\hat{X}=\frac{X}{\left|\right|X\left|\right|},& {\hat{W}}_j=\frac{W_j}{\left|\right|W_j\left|\right|}\end{array}$

Wherein, current input mode vector X refers to the parameter in step (4) after dimensionality reduction, the weight vector W that in competition layer, each neuron is corresponding _jrefer to the initial weight in step (5);

B () finds triumph neuron:

Will with (j=1,2,3 ... .., m) carry out similarity system design, the most similar neuron is won;

C () network exports and adjusts with power

Triumph neuron adjusts its weight vector W _j*, its weight vector study adjustment is as follows:

$\left\{\begin{array}{c}{W}_{j^{*}}(t+1)={\hat{W}}_{j^{*}}\left(t\right)+{\mathrm{\&Delta;W}}_{j^{*}}={\hat{W}}_{j^{*}}\left(t\right)+\mathrm{\&alpha;}(\hat{X}-{\hat{W}}_{j^{*}})\\ \begin{array}{cc}{W}_j(t+1)={\hat{W}}_j\left(t\right)& j\&NotEqual;j^{*}\end{array}\end{array}\right.$

Wherein, α is learning rate, 0< α≤1, and α reduces along with the process of study, and the degree namely adjusted is more and more less, is tending towards cluster centre;

D () be normalized again

Weight vector after normalization, after adjustment, repeats step (1) (2) (3) and re-starts normalization, until learning rate α decays to 0, obtain final weights W _j*.

3rd step, by weights W final in second step _j*, bring in first step formula, finally obtain seismic facies value Y _j.

Self organizing neural network structural drawing is see Fig. 2, when a neural network accepts extraneous input pattern, will be divided into different corresponding regions, there is different response characteristics in each region to input pattern, and this process completes automatically, self organizing neural network is without tutor's learning network.It is by the inherent law in Automatic-searching sample and essential attribute, and self-organization, adaptively modifying network parameter and structure, this is also that it carries out the core place of seismic facies automatic classification.Typical structure: input layer adds competition layer.

Input layer: accept external information, by input pattern to competition layer transmission, plays " observation " effect.

Competition layer: be responsible for carrying out " com-parison and analysis " input pattern, finds rule and the effect sorted out.The physiological Foundations of its competitive learning rule is the lateral inhibition phenomenon of neurocyte: when after a neurocyte excitement, can produce inhibiting effect to the neurocyte around it.The strongest inhibiting effect is that competition is won " only I am solely emerging ", and this way is called " the victor is a king " (Winner-Take-All, WTA).Competitive learning rule obtains from the lateral inhibition phenomenon of neurocyte.

In order to multiparameter nonlinear neural network calculates the applicability of seismic facies, we establish a complicated reef flat facies comprehensive geology model, calculate seismic facies, with the applicability of the method for inspection by just drilling section.

Reef beach matter forward modeling and seismic facies calculate compliance test result:

(1) reef beach geologic model foundation and just drilling

As shown in Figure 3, the reef beach geologic model of three levels is set up.Upper water flat bed is sand and mud interstratification, model Zhujiang River group deposited atop, and its medium sand mud layer speed is respectively 3500m/s and 2600m/s.The intermediate level simulates a large-scale reef-shoal complex, and left side is the sand and mud interstratification tilted, and right side is the sand and mud interstratification of level, is used for respectively reflecting the impact on seismic reflection under different occurrences.Consider in mud that the interval velocity that affects having grey matter composition brings up to 3300m/s.Lower level is a Thin Sandbody bottom Zhujiang River group and bottom shale layer, and speed is the same.

Consider that the organic reef in study area always mixes existence with organic bank, and many features existed in alternating layers shape, therefore intermediate level reef-shoal complex is designed to 6 thin layer alternating layers.Be respectively from top to bottom: top layer is bioherm, and speed gets 4500m/s, at top from left to right distribution density seam hole body from low to high, wherein stuff speed is 4000m/s, representative have fluid-filled after speed reduce.The argillaceous limestone band of the low speed of a thin layer is arranged at bottom, and speed is 4400m/s.Again down three layers be the close Reef & bank layer of speed, middle layer decreases due to local white clouds speed, be 4300m/s, and levels speed is 4400m/s.Bottom is skim argillaceous limestone layer, and speed is 4000m/s.Consider the cliff debris that to collapse at steeper position, slope, on the left of reef flat body, therefore devise the Slope Facies body that a speed is 4000m/s.

Convolution is carried out to the Ricker wavelet of above reef beach geologic model 30Hz dominant frequency, generates seismic response section as shown in Figure 4.

(2) the non-linearly seismic phase of reef beach geologic model calculates and effect analysis

As shown in Figure 4, this right section is followed the trail of by four sequences, then extract seismic multi-attribute parameter along window during layer position, carried out the self-organizing feature map lattice non-linearly seismic phase calculating based on principal component analysis (PCA).2. 3. 4. 1. four layers of label be in the drawings respectively, and it is as follows that each sequence seismic facies calculates effect:

First sequence: this layer is without phase transformation, and it is a straight line that seismic facies calculates basic result, and phase boundary is perfectly clear, and only having near reef top is necessarily having variation, see Fig. 5.

Second sequence: this layer of phase transformation is large, transversely from left to right can be divided into three regions, see Fig. 6.Delta Area, Jiao Tan district and level deposition district.Seismic facies obviously can be divided into three sections, corresponding three phase regions.West section pendage change, stratum, eastern section level, but the same mutually, seismic facies is still nearly straight line, and therefore attitude of stratum calculates not impact to seismic facies.Result of calculation accurately can determine the boundary position of phase.

3rd sequence: this layer of phase transformation is large, transversely from left to right can be divided into four regions: Delta Area, slope, reef beach and level deposition, see Fig. 7.The seismic facies calculated also can be divided into four sections substantially, corresponding four phase regions.Simultaneously variant with or without the reflectance signature of margin slope phase, have margin slope phase seismic facies value on the low side, and non-flanged Slope Facies seismic facies value is higher.Therefore synthetically seismic phase can determine the boundary position of phase.

4th sequence: this layer is without phase transformation.Seismic facies totally should be a straight line.But stratum is too thin, affect comparatively large by upper strata reef flat body, earthquake phase line has fluctuation, see Fig. 8.Illustrate that the seismic facies Detection results of the method to thick-layer rock mass is better.

From reef flat facies geologic model positive evolution row seismic facies result of calculation, the synthetically seismic phase that the method calculates can determine sedimentary facies (petrofacies) border exactly, defines obvious indicative function to the division on sedimentary facies border and scope.

(3) concrete geologic province calculates:

The seismic facies result of calculation of layer position, somewhere, the South Sea: use the method technology, calculate the seismic facies on a stratum, somewhere, the South Sea, result as shown in Figure 9.By analyses such as drilling well well loggings, this stratum major developmental reef flat facies, sand dam phase and continental shelf facies.As can be seen from Figure 9, representated by each seismic facies, sedimentary facies border is fully aware of, and the Sedimentary facies of reflection is very accurate.Illustrate that the seismic facies that this method calculates is reliably accurate.

Claims (2)

1., based on seismic facies computing method for seismic multi-attribute parametrical nonlinearity automatic classification, it is characterized in that: comprise the following steps,

(1) according to geologic information, sedimentary facies kind is set up;

(2) multiple Seismic Attribute Parameters X is extracted along in window during the zone of interest of seismic trace _ij, wherein j is earthquake Taoist monastic name, j=1,2 ... N, i are attribute number;

Described Seismic Attribute Parameters comprises: coefficient of kurtosis, the coefficient of skewness, standard deviation, RMS amplitude, average energy, gross energy, average amplitude, variance, the degree of bias, kurtosis, average reflection intensity, average instantaneous frequency, average instantaneous phase, reflection strength slope and instantaneous frequency slope;

(3) according to following formula by extract Seismic Attribute Parameters standardization to eliminate dimension difference;

Wherein, wherein j is earthquake Taoist monastic name, and i is attribute number, X _ij ^*for the Seismic Attribute Parameters after standardization, X _ijfor original property value, X _iminfor attribute minimum value, X _imaxfor attribute maximal value;

(4) by the Seismic Attribute Parameters after standardization, carry out dimension-reduction treatment by Karhunen-Loeve transformation, obtain the parameter after dimensionality reduction;

(5) according to the number of parameters after dimensionality reduction, adopt with standardization after Seismic Attribute Parameters X _ij ^*corresponding contribution margin P _mas initial weight W _j;

(6) calculate seismic facies, concrete grammar is:

The first step: set up seismic facies classification computing formula:

In formula: j=1,2 ... N, represents sedimentary facies kind, X _ijfor the compression parameters after input, W _jfor self organizing neural network trains the rear final weights of classification automatically;

Second step: according to self-organizing feature map Competitive Learning Algorithm, obtains the final weights W of Seismic Attribute Parameters _j*;

3rd step, by weights W final in second step _j*, bring in first step formula, finally obtain seismic facies value Y _j.

2. the seismic facies computing method based on seismic multi-attribute parametrical nonlinearity automatic classification according to claim 1, is characterized in that: in step (6), second step concrete grammar is:

(a) vectorial normalized:

The weight vector W corresponding to each neuron in the current input mode vector X in self-organizing network, competition layer _j(j=1,2,3 ... .., N), be all normalized, obtain with

Wherein, current input mode vector X refers to the Seismic Attribute Parameters in step (4) after dimensionality reduction, the weight vector W that in competition layer, each neuron is corresponding _jrefer to the initial weight in step (5);

B () finds triumph neuron:

Will with (j=1,2,3 ... .., m) carry out similarity system design, the most similar neuron is won;

C () network exports and adjusts with power

Triumph neuron adjusts its weight vector W _j*, its weight vector study adjustment is as follows:

Wherein, α is learning rate, 0< α≤1, and α reduces along with the process of study, and the degree namely adjusted is more and more less, is tending towards cluster centre;

D () be normalized again

Weight vector after normalization, after adjustment, repeats step (1) (2) (3), re-starts normalization until learning rate α decays to 0, obtain final weights W _j*.