CN1975462A - Coal seam thickness analyzing method based on earthquake attribute - Google Patents

Abstract

Because most coal seam belongs to typical seam, vertical resolution factor of traditional earthquake detector can not solve the thickness request of the coal seam, so, the analysis method of coal seam thickness based on earthquake attribute solves the problem of coal seam thickness in the exploration and exploitation of coal resources. Firstly, using the PAL module of the landmark company, choosing the appropriate time window from the three-dimensional excursion data, getting the earthquake attribute data such as swing, frequency, instantaneous, building the database of the earthquake attribute, doing self-correlation analysis of these attribute and the correlation analysis between attribute and the thickness, choosing the most significant earthquake attribute as the fundamental parameter of the coal seam forecast module,being combined to the known boring data, using multielement polynomial return and BP manpower neural network method builds multielement polynomial return model and BP manpower neural network model between every attribute and coal thickness, having developed the relevance algorithm procedure, doing error analysis of the model, having realized the forecast and valuation of the coal thickness in the prospecting area.

Description

Thickness of coal seam analytical approach based on seismic properties

Technical field is in the coal seismic prospecting field, also will provide the situation of change of thickness of coal seam except finding out in the exploiting field the structure.Along with the development of combining the technology of adopting, the situation of change that coal is thick has become urgent problem, because most of coal seam belongs to typical thin layer, vertical resolution does not reach and solves the thick requirement of coal.How utilizing earthquake information, accurately obtain thickness of coal seam information in conjunction with borehole data, is the problem that current domestic and international many scholars are studying.

How background technology comes evaluation of thin-bed thickness to be subjected to the attention of Chinese scholars by the thin bed reflections ripple, and carried out many theoretic discussions always, has delivered the correlative study paper.Ricker (1953) has proposed to differentiate standard---" the thunder gram standard " on stratum from the angle of resolution.This standard is, if this layer then can not be differentiated less than quarter-wave the time difference of the reflection wave of rock stratum upper and lower interface in time domain, only treats as a face.Widess (1973) proposes according to the relation of book layer thickness and seismic reflection response, when thickness of thin layer less than four of seismic event predominant wavelength/for the moment, earthquake wave amplitude and thickness of thin layer are approximated to direct ratio, break through pure method of geometry first and asked for the boundary of reflector thickness, from kinetic character, provided the concrete definition of thin layer quantification.Ruter and Schepers (1978), Koefoed and Voogd (1980) draw by synthetic earthquake model investigation, exist almost relation (quasi-Linerarity) between the compound wave amplitude of thickness of thin layer and seismic reflection.In China, the coalfield seismologist is according to the thin layer theory of Widess, adopt amplitude of vibration method to carry out the Coal Seam Thickness Change trend study, especially in the later stage eighties, coal seam on probation reflection wave comprehensive characteristics parameter (comprising amplitude, energy, energy ratio) is carried out the thickness of coal seam estimation, has obtained certain progress.Qi Jinghua (1996) has drawn the expression formula of utilizing spectral amplitude ratio and the direct inverting thickness of coal seam of spectral amplitude duplicate ratio by theoretical analysis and model test.Liu Tianfang, Chen Bin, Fu Jinsheng (1996) have proposed to use the thick spectral moment method of seismic inversion coal, have all obtained effect.But because amplitude, energy represent all is reflection wave intensity, thus usually be subjected to that the field excites, the influence of the thick factor of non-coal in reception and the Data Processing process.Cause the result of calculation dispersivity bigger, therefore, no matter home and overseas does not all also have the thick method of Inversion Calculation coal a kind of practicality, true than Huaihe River at present.

Seismic properties has been reacted geometry, kinematics, dynamics and the statistics feature of seismic waveshape, the seismic properties technology be by applied research, algorithm development and integrated software system extract, storage, visual, analyze, checking and estimate the technology of seismic properties.The seismic properties technology is applied to seismic interpretation processing, seismotectonics drawing, seismic stratigraphic interpretation, seismic lithology and various aspects such as simulation, reservoir description and simulation.Since the nineties in 20th century, the seismic properties technology is calculated from the single track instantaneous attribute, develop into multiple tracks window when layer is got and calculate tens kinds of parameters, can comparatively accurately determine position and looks such as the equal feature of water-oil interface, lithological change, variation in thickness, Crack Detection and earthquake.Seismic properties is becoming the key component of reservoir geophysics, and has set up a kind of special contact between prospecting seismology and production seismology.Use seismic properties prediction thickness of thin layer and comprise two aspects: one is the extraction of thin layer seismic properties; One is the relation research of thickness of thin layer and these attributes.The method of research roughly has two classes: a class is to utilize in tuning thickness, and amplitude and thickness of thin layer are approximate linear; One class is to utilize spectral amplitude to come the forecasting coal layer thickness.These methods are owing to the use single parameter, and the effect on amplitude factor is a lot, shakes the multi-solution of information insuperably, and effect is unsatisfactory.Although what have has used the multiattribute prediction, just extract seismic properties with theoretical and model investigation achievement, mainly stress research at the pre-detection identifying method of oil and gas reservoir, with one or more computing method predicting oil reservoir information, not preferred in conjunction with the actual attribute that carries out of study area, the forecast model of being set up is not carried out error analysis yet, because the sensitivity of array mode between various seismic properties information and various attribute reflection thickness has very big uncertainty, in different regions, the seismic properties combination of different layers position exists than big difference, and the confidence level of prediction is reduced." based on the thickness of coal seam analytical approach of seismic properties " solved the thickness of coal seam problem in coal resources exploration and the exploitation.At first use the PAL module of Landmark company, in the three-D migration data volume, choose suitable time window, therefrom extract the amplitude class, the frequency class, seismic properties data such as instantaneous class, set up the seismic properties database, these attributes are done autocorrelation analysis to be analyzed with the thick relevant rank of attribute and coal, therefrom optimize the most significant seismic properties as the thick forecast model basic parameter of coal, in conjunction with known borehole data, utilize multinomial to return and the BP Artificial Neural Network, set up each attribute and coal multinomial regression model and the artificial nerve network model between thick, developed the related algorithm program, model has been carried out error analysis.Owing to considered the multiattribute parameter simultaneously, thereby the computation model that draws is more perfect and more approaching reality, reflection seismic properties forecasting coal layer thickness has good effect.

Summary of the invention is used the PAL module of Landmark company, chooses suitable time window in the three-D migration data volume, therefrom extracts seismic properties data such as amplitude class, frequency class, instantaneous class, sets up the seismic properties database.According to the reality of study area, use multiple mathematical method and set up correlationship and model between thickness of coal seam and the seismic properties.As multiple linear regression, polynary binomial returns, the BP neural network returns mathematical model have been set up, exploitation is based on seismic properties forecasting coal layer thickness software for calculation, analyze relative merits, usable range and the result of use of each algorithm predicts, realize purpose based on seismic properties forecasting coal layer thickness.

(1) seismic properties is extracted and method for optimizing

The sorting technique of seismic properties has a lot, mainly contains following 4 kinds: the one, in the comparatively popular sorting technique of China academia, promptly, seismic properties is divided into several big classes such as amplitude, frequency, phase place, energy, waveform and ratio from kinematics and dynamic (dynamical) angle; The 2nd, the method for picking up by attribute with seismic properties be divided into layer bit attribute and the time window attribute two classes sorting technique; The 3rd, by the sorting technique that seismic properties is divided into time, amplitude, frequency and decay 4 classes of Alistair R.Brown proposition in 1996; The 4th, by the sorting technique based on reservoir characteristic that Quincy Chen et al.1997 proposes, this method helps us according to the object primary election seismic properties that will study, to reduce the blindness and the randomness of property calculation.

Use the PAL module of Landmark company by studying us, in the three-D migration data volume, choose suitable time window, therefrom extract attributes such as amplitude class, frequency class, instantaneous class; And these attributes are done the thick relevant rank of autocorrelation analysis, attribute and coal analyze, therefrom optimize the most significant seismic properties, set up the seismic properties database.

(2) set up the thickness of coal seam forecast model

According to the reality of study area, use multiple mathematical method and set up correlationship and model between thickness of coal seam and the seismic properties.As set up the mathematical model that multiple linear regression, the recurrence of polynary binomial, BP neural network return, exploitation is based on the algorithm routine of seismic properties forecasting coal layer thickness.

(3) use seismic properties forecasting coal layer thickness

Analyze the thickness of coal seam error prediction model analysis of setting up, relative merits, usable range and the result of use of prediction, realize purpose based on seismic properties forecasting coal layer thickness.

Embodiment

1, the amplitude frequency characteristic of coal seam complex wave

The coal seam is as " thin layer " (H≤λ/4) that define usually in the seismic prospecting, and its reflection wave is a coal seam roof and floor boundary reflection, and there is tuning point in coefficient stack complex waves such as interformational multiples and transformed wave with the synthetic reflection wave of the variation of thickness of coal seam.In viscoelastic body, the amplitude frequency characteristic of coal seam complex wave is:

$\left|H\left(f\right)\right|=\frac{\left|r\right|\sqrt{1+e^{-2\mathrm{\β}}-2e^{-\mathrm{\β}}\cos (4\mathrm{\πd}/\mathrm{\λ})}}{\sqrt{1+r^4{e}^{-2\mathrm{\β}}-2r^2{e}^{-\mathrm{\β}}\cos (4\mathrm{\πd}/\mathrm{\λ})}}---\left(1\right)$

R in the formula: the upper and lower reflection coefficient in coal seam; D: thickness of coal seam; β=2 α d (α is the attenuation by absorption factor in coal seam).

Exist under the situation of thin layer, when ripple impinged perpendicularly on the thin layer surface, reflection coefficient is not only relevant with the wave impedance on both sides, interface, and was also relevant with the incident wave frequency.Thereby thin layer can regard a wave filter as, and incident wave on the thin layer surface reflex time takes place, and seems to have stood certain frequency filtering effect by a wave filter.The effect of thin bed reflections stack is that the composition to low frequency and high frequency has suppression, and the intermediate frequency composition of the reflection wave that receives is strengthened relatively.

3, seismic properties is preferred

(1) extraction of seismic properties

According to theoretical and model investigation achievement, extract amplitude class, complex seismic trace class, frequency spectrum statistics generic attribute, wherein the amplitude generic attribute is 15 kinds: RMS amplitude, average absolute amplitude, passages, average peak amplitude, maximum valley amplitude, average valley amplitude, maximum absolute amplitude, absolute amplitude total amount, amplitude total amount, average energy, energy are overall, average amplitude, amplitude variations, the asymmetry of amplitude variations, the kurtosis of amplitude.Seismic amplitude or energy properties have reflected that wave impedance is poor, zone thickness, rock composition, reservoir pressure, factor of porosity and contain the variation of fluid composition.Both can be used to discern amplitude anomaly or sequence feature, also can be used to follow the trail of stratigraphy feature such as delta watercourse or sandstone.In addition, also can be used for gathering of discerning lithological change, unconformability, gas and fluid etc.5 kinds of complex seismic trace statistics classes: average reflection intensity, average instantaneous frequency, average instantaneous phase, reflection strength slope, instantaneous frequency slope.Complex seismic trace is actual to be the Hilbert conversion of seismic signal.It can help feature, lithology, river course and delta sandstone, reefs, unconformity surface, bed succession, crack, tuning effect of analytical gas, fluid etc.Frequently (energy) spectrum statistics class is 6 kinds: effective bandwidth, arc length, average zero point of crossing frequency, dominant frequency sequence F1, F2, F3, dominant frequency peak value, dominant frequency peak value are to the slope of maximum frequency.It is to the frequency spectrum of seismic signal and energy spectrum, can disclose the wavelet that cranny development band, gassiness uptake zone, tuning effect, lithology or the absorption of stratum or oil gas effect cause and change.

(2) seismic properties is selected

Utilize other people experience or mathematical method, optimize finding the solution problem seismic properties the most responsive (or the most effective, the most representative), that the attribute number is minimum or seismic properties combination, improve earthquake reservoir prediction precision, improve the processing relevant and the effect of interpretation procedure with seismic properties.The process of picking out the attribute set that helps the earthquake reservoir prediction most from a property set is called attribute and selects.

(3) based on relevant seismic properties primary election

Other seismic properties of well and thickness of coal seam value are carried out normalized, and according to (2) formula, seismic properties and the thick related coefficient rank of coal behind the other normalizing of calculating well are selected and the bigger attribute in the thick related coefficient of coal rank, form and make the seismic properties collection that model is used.

$r=\frac{\underset{i}{\mathrm{\Σ}}(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})}{\sqrt{\underset{i}{\mathrm{\Σ}}{(x_i-\stackrel{\‾}{x})}^2}\sqrt{\underset{i}{\mathrm{\Σ}}{(y_i-\stackrel{\‾}{y})}^2}}---\left(2\right)$

(4) based on the attributive analysis of simple crosscorrelation

In order to improve confidence level, to carrying out cross-correlation analysis with the bigger seismic properties of the thick related coefficient of coal, the seismic properties that correlation is bigger merges, and has relative independentability to guarantee the seismic properties that is used to predict.If the attribute that related coefficient is very big returns, the stability of meeting impact prediction algorithm.The computing formula of simple crosscorrelation is identical with (2) formula.

4, ask the related coefficient of equation of linear regression between thick and each seismic properties of coal

Utilize the least square square law to ask error between thickness of coal seam and each the seismic properties equation of linear regression.

$\hat{Y}=\mathrm{ax}+b---\left(3\right)$

Wherein Be the forecasting coal layer thickness, x is the seismic properties value, and a, b are regression coefficient.

(4) formula of utilization is calculated the thick and seismic properties coefficient R of coal ²

$R^2=1-\frac{\mathrm{\Σ}{(Y_i-{\hat{Y}}_i)}^2}{\left(\mathrm{\Σ}{Y}_i^2\right)-\frac{{\left(\mathrm{\Σ}{Y}_i\right)}^2}{n}}---\left(4\right)$

Wherein, Y _iFor the coal of known boring is thick, Be the thick value of coal that calculates according to (3) formula.

Comprehensive above-mentioned three kinds of methods select the most significant seismic properties forecasting coal thick.

5, the approximation of function method of thickness of coal seam

(1) multiple regression analysis

According to the thickness of coal seam and preferred seismic properties value of the other seismologic record of well, do normalized, its principle is: establishing sample data is x _p(p=1,2 ..., P), definition x _Max=max{x _p, x _Min=min{x _p, normalized is calculated the data that promptly by (5) formula sample data are converted into 0～1 interval.

$\frac{x_p-x_{\min }+a}{x_{\max }-x_{\min }+a}\→x_p---\left(5\right)$

With the property set after the top normalization, calculate other seismic properties of well and the thick higher polynomial regression model of coal, suppose to have p attribute, set up the m order polynomial regression equation of thick and p the attribute of coal, promptly

$\hat{y}=a_{00}+a_{11}{x}_1+a_{12}{x_1}^2+...+a_{1m}{x}_{1}m$

$+a_{21}{x}_2+a_{22}{x_2}^2+...+a_{2m}{x_2}^m+...+---\left(6\right)$

$a_{p1}{x}_p+a_{p2}{x_p}^2+...+a_{\mathrm{pm}}{x_p}^m$

Wherein: -forecasting coal is thick; x _i(i=1,2 ..., p)---the value of each amplitude attribute;

a _Ij(i=0,1 ..., p; J=1,2 ..., n, n are sample number)-regression coefficient.

According to the property value of borehole data and seismic trace near well, obtain a collection of test figure: A _1i, A _2i..., A _Pi, y _i(i=1,2 ..., m), make actual tests numerical value y _iGo up corresponding with formula (6)

Between residual sum of squares (RSS)

$Q=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{(y_i-{\hat{y}}_i)}^2---\left(7\right)$

Be minimum, ask each factor alpha with least square method _IjValue.

(2) BP neural network prediction coal is thick

The normalization of sample data: owing to select for use the Sigmoid function as neuronic excitation function in the network, therefore, in order to effectively utilize the characteristic of S type function, to guarantee the neuronic nonlinear interaction of network, carry out normalized for the learning sample and output data utilization (4-5) formula of numeric type.

Utilize back propagation learning to set up the neural network model of the thick prediction of coal: being provided with learning sample is (x _1p, x _2p..., x _Npt _p) (p=1,2 ..., P; P is a sample number).Provide W (w at random _Ij, θ _i, v _i) after, according to the output y of p sample of (8)～(10) formula computational grid _p

$I_i=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_j+{\mathrm{\θ}}_i,(i=\mathrm{1,2},\·\·\·,m)---\left(8\right)$

Wherein, n is the neuron number of input layer; M is the neuron number of hidden layer; w _IjConnection weight for hidden neuron i and input layer j; θ _iThreshold values for hidden neuron i.

$o_i=\frac{1}{1+e^{-I_i}}---\left(9\right)$

Wherein, I _iIt is the input of i hidden neuron; O _iBe the output of i hidden neuron.

$y_p=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{v}_i{o}_i---\left(10\right)$

Wherein, v _iConnection weight for output layer neuron and hidden neuron i; y _pBe the output of p sample.

Definition is by the connection weight w of hidden neuron and input layer _Ij, hidden neuron threshold values θ _iConnection weight v with output layer neuron and hidden neuron _iThe vector of forming is the connection weight vector W of network.

For sample p, the output error of define grid is

$o_{\mathrm{ip}}=\frac{1}{1+e^{-I_{\mathrm{ip}}}}{d}_p=t_p-y_p---\left(11\right)$

And the definition error function is

$e_p=\frac{1}{2}{(t_p-y_p)}^2---\left(12\right)$

Along error function e _pWith the negative gradient direction that W changes W is revised.If the modified value of W is Δ W, get

$\mathrm{\Δw}=-\mathrm{\η}\frac{\∂e_p}{\∂w}---\left(13\right)$

In the formula: η is a learning rate, gets 0～1 number.

After trying to achieve Δ W, adopt iterative

W+ΔW→W (14)

Former W is carried out corrected Calculation, obtain new connection weight vector W.

For all learning samples, all put in order and carry out above-mentioned computation process, then the fixing value of W according to sample.P sample carried out forward calculating respectively, thereby obtain the energy function value of learning sample

$E=\underset{P=1}{\overset{P}{\mathrm{\Σ}}}{e}_p---\left(15\right)$

By iterating, W revises to the network connection weight, makes E satisfy a certain accuracy requirement.

(3) model error analysis

$e=\frac{\sqrt{\mathrm{\Σ}{(y-y_i)}^2}}{N}---\left(16\right)$

$R=1-\frac{{\mathrm{\Σ}(y-y_i)}^2}{\mathrm{\Σ}{(y_i-\stackrel{\‾}{y})}^2}---\left(17\right)$

In the following formula, E is a standard error of estimate, and its value heals the bright institute of novel established model better; R is the coefficient of determination, and the bigger explanation of its value institute established model better.Make E, R reach certain accuracy requirement.

6 application example analyses

Compose in the Huainan field Xie Qiao exploration area and have the dozens of coal seam, but thickness is less mostly, perhaps unstable, and fundamental purpose layer 13-1 coal, 11-2 coal and 8 thickness of coal seam are all bigger, and tax is deposited stable, and there is bigger physical difference based on tonstein, mud stone and sandstone in the roof and floor lithology in these coal seams with coal seam itself, therefore 13-1 coal, 11-2 coal and 8 coal seam roof and floors all are good reflecting interfaces, the energy reflection wave that produce power is strong, continuity is good.13-1 coal seam wherein, thickness 3～6m, and the seismic reflection wave energy is stronger, signal to noise ratio (S/N ratio) is higher, and waveform is outstanding, laterally can follow the trail of continuously, has good seismic geological codition.

(1) seismic properties is extracted

Adopt the PAL property extracting module of the Landmark Poststack of company, determine to open along the 13-1 coal seam 20ms the time window as the time window that extracts attributive analysis.In the window, extract 28 kinds of seismic properties altogether at this moment, wherein the amplitude generic attribute is 15 kinds, 5 kinds of complex seismic trace generic attributes, (energy) spectrum statistics 8 kinds of generic attributes (seeing Table 1) frequently.

At first, the thickness of coal seam of the other seismologic record of well is carried out normalized with preferred seismic properties data; Then, according to the data after the normalized, calculate related coefficient between the thick and seismic properties of coal according to (2) formula, r is a related coefficient in the formula; According to 7 the boring points in 13-1 coal seam, Xie Qiao colliery, Huainan and 28 tunnel point data, calculating is produced coal thick as shown in table 1 with the seismic properties related coefficient.

Thick and the seismic properties related coefficient of table 1 coal

Table 1 Correlation coefficients between coal thickness and seismic attributes

Seismic properties	Correlation coefficient r	Seismic properties	Correlation coefficient r
				The kurtosis of the asymmetry amplitude of the total energy population mean of the absolute amplitude total amount RMS amplitude average absolute amplitude peak-peak amplitude maximum absolute amplitude average energy amplitude of the average valley amplitude of maximum valley amplitude amplitude amplitude variations amplitude variations	-0.201 -0.287 -0.201 -0.503 -0.308 -0.084 -0.514 -0.287 0.079 -0.287 0.079 -0.350 -0.051 -0.403	The average instantaneous phase reflected intensity of the average instantaneous frequency of average reflection intensity slope instantaneous frequency slope effective bandwidth arc length average zero crosspoint frequency dominant frequency F1 dominant frequency F2 dominant frequency F3 dominant frequency peak value dominant frequency peak value is to the slope average peak amplitude of peak frequency	-0.244 -0.306 0.195 -0.062 -0.400 0.015 -0.201 -0.026 0.054 0.281 0.103 -0.019 0.035 -0.340

(2) seismic properties is preferred

Based on the related coefficient between thickness of coal seam and the seismic properties, therefrom optimize related coefficient greater than-0.3 8 kinds of seismic properties (table 1), i.e. the kurtosis of passages, maximum valley amplitude, maximum absolute amplitude, amplitude variations, amplitude, average instantaneous frequency, instantaneous frequency slope and average peak amplitude.For the relative independentability that guarantees each attribute and the stability of algorithm, carry out the cross-correlation analysis of seismic properties, as shown in table 2.The computing formula of cross-correlation analysis is identical with (2) formula, according to the related coefficient between each seismic properties, rejects the bigger attribute of related coefficient.By the cross-correlation analysis of seismic properties, obtain the basic parameter of 4 useful seismic properties at last as forecast model, they are respectively: average peak amplitude, the kurtosis of amplitude, maximum absolute amplitude and instantaneous frequency slope.With the basic parameter of these seismic properties as regression model and BP neural network prediction model.

Cross-correlation analysis between table 2 seismic properties

Table 2 Results of cross-correlation analysis between seismic attributes

Attribute	Average instantaneous frequency	The average peak amplitude	The kurtosis of amplitude	Maximum absolute amplitude	Passages	Maximum valley amplitude	The instantaneous frequency slope	Amplitude variations	Coal is thick
										The maximum valley amplitude of the maximum absolute amplitude peak-peak of the kurtosis of average instantaneous frequency average peak amplitude amplitude amplitude instantaneous frequency slope amplitude variations coal is thick	1.000 -0.051 0.442 0.074 0.038 0.716 0.617 0.350 -0.306	-0.051 1.000 0.660 0.744 0.754 0.318 0.100 0.749 -0.340	0.442 0.660 1.000 0.608 0.589 0.785 0.569 0.946 -0.403	0.074 0.744 0.608 1.000 0.995 0.366 0.377 0.553 -0.514	0.038 0.754 0.589 0.995 1.000 0.331 0.332 0.541 -0.503	0.716 0.318 0.785 0.366 0.331 1.000 0.669 0.653 -0.308	0.617 0.100 0.569 0.377 0.332 0.669 1.000 0.403 -0.400	0.35 0.749 0.946 0.553 0.541 0.653 0.403 1.000 -0.35	-0.306 -0.340 -0.403 -0.514 -0.503 -0.308 -0.400 -0.350 1.000

(3) thickness of coal seam forecast model

(a) multivariate statistics forecast model

According to 13-1 coal seam, Xie Qiao colliery, Huainan actual observation point data,, set up seismic properties and the coal multinomial regression model between thick based on the seismic properties collection after the normalization.

Polynomial regression model of quaternary is passed through the average peak amplitude, the kurtosis of amplitude, and the correlation analysis between maximum absolute amplitude and instantaneous frequency slope attribute and the thickness of coal seam, polynomial regression model of quaternary of calculating acquisition is

y＝8.0790-1.9102x ₁-0.8189x ₂-0.7723x ₃-2.9346x ₄， (4)

Y is the thick value of forecasting coal (m) in the formula; x ₁Be the average peak amplitude; x ₂Kurtosis for amplitude; x ₃Be maximum absolute amplitude; x ₄Be the instantaneous frequency slope.

The regression model of quaternary quadratic polynomial by the regression model that calculates the quaternary quadratic polynomial that obtains is

$y=-40.1827x_1+29.9812x_1^2+3.0684x_2-6.5302x_2^2+67.2576x_3-$

$40.9466x_3^2-51.4177x_4+56.9763x_4^2-1.7485---\left(5\right)$

(b) BP neural network prediction model

The BP neural network model has self study, self-organization, strong fault tolerance, calculates advantages such as simple, that parallel processing speed is fast, and it can approach any Nonlinear Mapping in theory arbitrarily, therefore is most widely used.Set up the BP neural network model, at first select for use the Sigmoid function as neuronic excitation function in the network.In order to effectively utilize the characteristic of S type function, to guarantee the neuronic nonlinear interaction of network, carry out normalized for the learning sample and output data utilization (1) formula of numeric type, the output valve of each node is 0～1.

Utilize back propagation learning to set up the neural network model of the thick prediction of coal, according to 13-1 coal seam, Xie Qiao colliery, Huainan actual observation point data, filter out 35 measured datas as learning training and test sample book (seeing Table 4), wherein 7 borings point data and 28 tunnel point data,, as learning sample network is trained with boring point seismic properties.

The BP network is by the network output error is fed back network parameter to be revised, thereby realizes the non-linear mapping capability of network.Robet-Nielson has proved that 3 layers of BP network model with 1 hidden layer can approach any continuous function effectively, promptly comprise input layer, hidden layer and output layer.Based on the study area actual conditions, the network structure of the thickness of coal seam BP neural network prediction model of foundation.Model adopts 3 layer network structures, and with 4 nodes of preferred 4 kinds of seismic properties as the e-learning input layer, the middle layer of network is 2 nodes, and output layer is 1 node, sets up thickness of coal seam BP neural network prediction model.

Through iteration, weight coefficient W between input layer and hidden layer and the weight coefficient V between hidden layer and output layer are respectively

$W=\left[\begin{array}{ccccc}-11.8379& 23.1739& 15.3692& -22.6159& -4.7556\\ 13.6481& -22.3774& -10.6692& 34.5491& -5.5601\end{array}\right],---\left(6\right)$

V＝[-9.0616 -8.9075 9.7173]. (7)

(4) the thickness of coal seam error analysis that predicts the outcome

The thickness of coal seam error prediction model analysis result of setting up according to 13-1 coal seam, Xie Qiao colliery, Huainan actual observation point data shows, polynomial regression model of quaternary, and regression error is relatively large; And the regression model of quaternary quadratic polynomial relative with BP neural network prediction model regression error less (table 3)

Table 3 thickness of coal seam error prediction model is analyzed

Table 3 Error analysis of prediction models of coal thickness

Forecast model	Standard error of estimate e	Coefficient of determination R	Model evaluation
				The regression model BP neural network prediction model of a polynomial regression model quaternary of quaternary quadratic polynomial	0.472 0.427 0.068	0.244 0.379 0.141	The less relatively regression error of the relatively large regression error of regression error is less relatively

For the reliability that further testing model predicts the outcome, with BP artificial neural network and polynomial regression model Xie Qiao colliery, Huainan 13-1 thickness of coal seam is carried out forecast analysis and check, it is as shown in table 4 to predict the outcome.According to model predication value and measured value and error comparative analysis thereof as can be seen: it is relatively large to use quadratic polynomial forecast of regression model thickness of coal seam error, although it is fine that the quadratic polynomial regression model coincide in the known point data, it is thick to be not useable for whole study area forecasting coal; But BP artificial nerve network model forecasting coal layer thickness data can be applicable to whole study area, remove non-value point, and nearly all data are all available, and error is also less, and the precision height illustrates with Neural Network model predictive thickness of coal seam the most stable (as table 4).

Table 4 Huainan Xie Qiaoxi 1 exploiting field 13-1 coal thickness prediction error statistics table

Table 4 Error statistics of thickness prediction of coal seam 13-1 in the mining area west No.1 of the Xieqiao coal mine，Huainan

Pound sign	Actual value (m)	Neural network prediction			A polynomial regression			Quadratic polynomial returns
											Predicted value (m)	Absolute error	Relative error (%)	Predicted value (m)	Absolute error	Relative error (%)	Predicted value (m)	Absolute error	Relative error (%)
		L3 adds 2 D3 benefit v4 and examines 1 8-7 1703 * H1 H2 H3 H4 H5 H6 H7 H8 H10 h11 H12 H13 H17 H18 H19 H20 H21 H22 H23 H24 H25 H27 H28 H29 H30 H31 H32 H34	5.29 4.78 5.02 4.99 5.84 4.34 3.84 4.00 4.00 4.60 5.00 4.90 5.10 5.00 5.00 5.20 5.10 4.90 4.80 3.00 4.10 5.00 5.50 4.80 4.50 4.20 4.30 5.30 4.00 5.40 5.30 4.80 5.00 5.00 5.00	5.20 4.85 4.91 4.78 5.70 4.97 4.20 4.43 4.01 4.68 4.89 4.90 4.92 4.83 4.86 4.96 4.98 4.99 4.80 4.96 4.12 5.00 5.50 4.84 4.64 4.06 4.39 4.83 3.99 5.07 4.95 4.88 4.88 4.87 4.75	0.09 0.07 0.11 0.21 0.14 0.63 0.36 0.43 0.01 0.08 0.11 0.00 0.18 0.17 0.14 0.24 0.12 0.09 0.00 1.96 0.02 0.00 0.00 0.04 0.06 0.14 0.09 0.47 0.01 0.33 0.35 0.08 0.12 0.13 0.25	1.70 1.46 2.19 4.21 2.40 14.52 9.38 10.79 0.28 1.67 2.11 0.00 3.54 3.40 2.82 4.57 2.45 2.03 0.00 65.33 0.45 0.09 0.00 0.76 3.12 3.40 2.07 8.86 0.30 6.04 6.53 1.65 2.44 2.50 4.99	4.82 4.95 4.77 4.61 5.31 5.00 4.40 4.50 4.17 4.68 5.04 4.90 4.79 4.69 4.92 4.66 4.72 4.72 4.64 5.03 4.56 5.13 5.08 4.84 4.49 4.23 4.34 4.86 4.52 5.01 5.12 5.15 5.01 5.08 4.70	0.47 0.17 0.25 0.38 0.53 0.66 0.56 0.50 0.17 0.08 0.04 0.00 0.31 0.31 0.08 0.54 0.38 0.18 0.16 2.03 0.46 0.13 0.42 0.04 0.01 0.03 0.04 0.44 0.52 0.39 0.18 0.35 0.01 0.08 0.30	8.88 3.56 4.98 7.62 9.08 15.21 14.58 12.55 4.24 1.64 0.76 0.00 6.03 6.20 1.58 10.33 7.38 3.65 3.40 67.80 11.16 2.68 7.61 0.86 0.33 0.80 0.97 8.29 13.11 7.16 3.40 7.27 0.17 1.66 5.98	5.14 4.69 4.79 4.69 5.62 5.09 5.50 4.68 3.82 4.59 4.93 4.81 4.81 4.61 4.77 5.18 4.90 4.78 4.69 4.87 4.25 5.12 5.12 4.88 4.55 4.23 4.29 5.37 4.58 5.01 4.91 4.95 4.84 4.87 4.62										0.15 0.09 0.23 0.30 0.22 0.75 1.66 0.68 0.18 0.01 0.07 0.09 0.29 0.39 0.23 0.02 0.20 0.12 0.11 1.87 0.15 0.12 0.38 0.08 0.05 0.03 0.01 0.07 0.58 0.39 0.39 0.15 0.16 0.13 0.38	2.84 1.88 4.58 6.01 3.77 17.28 43.23 17.11 4.43 0.15 1.42 1.82 5.65 7.74 4.51 0.39 3.93 2.51 2.34 62.35 3.72 2.35 7.00 1.56 1.06 0.66 0.33 1.41 14.61 7.31 7.31 3.19 3.24 2.63 7.68

Annotate: band * person is the tunnel data for the checking hole with h beginning person.

Claims (4)

1. seismic properties extraction and method for optimizing and seismic properties database.

2. based on the thickness of coal seam multiple linear regression of seismic properties, polynary quadratic polynomial regression model, BP neural network model

3. based on the thickness of coal seam software for calculation of seismic properties.

4. use the method for seismic properties forecasting coal layer thickness, comprise thickness of coal seam forecast model and the error analysis method and the test block achievement data of foundation.