CN101975092A - Real-time prediction method of mine gas concentration in short and medium terms based on radial basis function neural network integration - Google Patents

Abstract

The invention discloses a real-time prediction method of mine gas concentration in short and medium terms based on radial basis function neural network integration. The method comprises the following steps of: taking mine gas concentration data as a chaotic time series to construct a plurality of prediction sub-models of radial basis function (RBF) neural networks, and taking a weighted mean of synchronous prediction results of all prediction sub-models as an integrated prediction value to realize prediction model initializtion of RBF neural network integration; then realizing prediction of the gas concentration in the range of from a short term to a medium term through setting an integrated capacity parameter (the integrated capacity parameter is also equal to an RBF network prediction step-length); and obtaining a new prediction sub-model by utilizing an incremental training mode aiming at the characteristics that gas concentration information is continuously collected, and realizing updating of the RBF neural network integration according to a first in first out queue sequence so as to improve real-time prediction precision of the gas concentration, therefore, a proper compromise can be obtained between prediction range and prediction precision requirements, and the technical requirement on a mine gas information management system is satisfied.

Description

Based on integrated mine gas concentration real-time predicting method a middle or short term of radial basis function neural network

Technical field

The invention belongs to mine gas concentration prediction field, at the real-time of mine gas concentration measurement and control system and the prediction requirement from the short-term to the mid-term, the invention particularly relates to a kind of mine gas concentration real-time predicting method a middle or short term based on the radial basis function neural network integrated technology.

Background technology

Country's " energy medium-term and long-term development planning outline (2004-2020) " determined Chinese will " adhere to coal be the theme, electric power is the center, the energy strategy of oil gas and new forms of energy development in an all-round way ", coal and coal industry remain the energy theme and the basic industry of China obviously.Along with the growth of China to energy demand, coal production also increases rapidly, already brings considerable economic for coal.Yet accident frequently takes place, and is shocking.According to statistics, the annual death by accident number of coal in China industry nearly 10000 people, direct economic loss is above 4,000,000,000 yuan.The gas disaster has directly hindered the ordinary production of colliery unit, lasting, healthy, the development stably of having blocked coal industry, so the diaster prevention and control of reinforcement gas is stable, the reliable supply of guaranteeing the coal energy, promotes the important leverage that national economy is comprehensive, develop in a healthy way.

At present, Chinese scholars has carried out going deep into and careful research to the forecasting problem of mine gas, has proposed a lot of effectively Forecasting Methodologies.According to the history order, these methods roughly can be divided into traditional Predicting Technique and modern Predicting Technique two big classes.The tradition Predicting Technique belongs to gas outstanding static state and the prediction of discontinuous contact, it is some quantizating index according to coal containing methane gas volume property and occurrence condition thereof, as gas index, coal seam character index, geostatic stress index or its overall target, prediction is exactly whether the single or multiple indexs of investigating wherein surpass threshold.Index commonly used has gush out initial velocity, coal powder quantity of bore, gas of drilling cuttings desorb index, boring gas to diffuse index, coal body Protodyakonov coefficient, gas pressure etc.Because gas is outstanding to be all overdetermined by the thick variation of structural performance, geological structure, coal, coal body structure and the country rock feature etc. of geostatic stress, high pressure gas, coal, various mechanical functions and these act on formed geologic body, great majority all are in complicated nonlinear state, so the precision of prediction of conventional art often is difficult to reach the requirement of Safety of Coal Mine Production.Modern prediction is the Predicting Technique that mainly is based on mathematics and physics, promptly utilize fuzzy theory, gray theory, neutral net, expert system, fractal, chaos and nonlinear theory, rheology and catastrophe theory etc. by prediction mine outburst amount as judging that gas gives prominence to the effective way of this noncontact prediction index.This shows that the prediction of gas emission (being gas density) is not only had theoretic value, great production practical significance is more arranged.

Along with the needs of country to improving constantly of requiring of Safety of Coal Mine Production and enterprise's self-growth, the monitoring mining monitoring system has all been equipped in each big-and-middle-sized colliery successively by China, monitor at toxic gases such as gas and visual plant, greatly improved the mine safety level of production and production safety management efficient.Yet, these monitor datas only are a kind of records to the work at present state, also lack the behave of the development trend of monitor data being made effective prediction in following a period of time, in fact hazard forecasting and the prevention that the accurate prediction work of gas density is given prominence to for gas all has significant meaning.At present, the prediction of gas density is also belonged to the off-line type prediction to a great extent, as Forecasting Methodologies such as artificial neural network and chaos time sequences.Gas density information is continual is gathered and is transmitted, if the off-line forecast model can not be corrected and updated timely, will certainly influence the precision of prediction of whole model, finally causes prediction to be lost efficacy.To sum up, the real-time predicting method of exploitation and development mine gas concentration is imperative.

Summary of the invention

Purpose of the present invention: overcome the defective of predicting mine gas concentration at present based on the chaos time sequence off-line, a kind of on-line prediction method based on the radial basis function neural network integrated technology is provided, realizes real-time estimate mine gas concentration from short-term to the scope in mid-term.

The present invention is a kind of based on integrated mine gas concentration real-time predicting method a middle or short term of radial basis function neural network, and the present invention thes contents are as follows:

A kind of based on integrated mine gas concentration real-time predicting method a middle or short term of radial basis function neural network, comprise the steps:

The first step, gather the gas density data, deposit the gas density historical data base in by firedamp sensor

X _lib＝{x _i|i＝1，2，...，l}(l＝n+2p)；

Second step, the gas density data that database is deposited are handled as chaos time sequence, utilize the delay time T of the C-C method sequence of calculation and embed dimension m;

The 3rd step, set the integrated capacity parameter p of radial basis function neural network and update time t;

The 4th the step, set up the integrated forecast model of radial primary function network, step is as follows:

(1) sets up phase space reconfiguration X respectively _iWith p step predicted vector Y _I+p(i=1+ (m-1) τ ..., n+p),

X _i＝[x _i-(m-1)τ，...，x _i-τ，x _i]，

Y _i+p＝[x _i+1，x _i+2，...，x _i+p]，

Wherein, x _iThe gas density data of gathering constantly for i;

(2) according to training sample set { (X _iY _I+p) | i=1+ (m-1) τ ..., n+1}, the prediction network of monolithic training radial primary function network (RBF) is f ¹

(3) again by incremental training sample set { (X _iY _I+p) | i=n+2} is at f ¹Increment type training RBF prediction network on the basis, or by { (X _iY _I+p) | i=1+ (m-1) τ ..., n+2} monolithic training RBF network is f ²

(4) by that analogy, up to by incremental training sample set { (X _iY _I+p) | i=n+p} is at f ^P-1Increment type training or by { (X on the basis _iY _I+p) | i=1+ (m-1) τ ..., n+p} monolithic training RBF network is f ^p:

(5) finish the integrated initial work of radial basis function neural network, obtain containing the integrated prediction model f of p RBF network ¹, f ²..., f ^p

The 5th the step, carry out real-time estimate according to RBF network integrated model, step is as follows:

(1) by the real-time image data X of firedamp sensor _Real={ x _i| i=l+1, l+2 ..., l+t} is in conjunction with gas density historical data base X _Lib, carry out phase space reconfiguration

X _i＝[x _i-(m-1)τ，...，x _i-τ，x _i](i＝l+1-p，...，l+t-1)；

(2) according to phase space X _iCarry out the prediction work of RBF network integrated model, respectively with X _i, X _I+1..., X _I+p-1(i=l+1-p ..., l+t-1) bring forecast model f into ¹, f ²..., f ^p, have

$\left\{\begin{array}{c}{\tilde{Y}}_{i+\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="HSA00000334446500021.png,HSA00000334446500031.png"\; state="\{\{state\}\}">p</figure-callout>}}^1=[{\tilde{x}}_{i+1}^1,{\tilde{x}}_{i+2}^1,...,{\tilde{x}}_{i+p}^1]=f^1\left(X_i\right)\\ {\tilde{Y}}_{i+p+1}^2=[{\tilde{x}}_{i+2}^2,...,{\tilde{x}}_{i+p}^2,{\tilde{x}}_{i+p+1}^2]=f^2\left(X_{i+1}\right)\\ ......\\ {\tilde{Y}}_{i+2p-1}^p=[{\tilde{x}}_{i+p}^p,{\tilde{x}}_{i+p+1}^p,...,{\tilde{x}}_{i+2p-1}^p]=f^p\left(X_{i+p-1}\right)\end{array}\right.,$

Obtain x _I+pP predict the outcome synchronously

Then with its weighted average as the integrated predicted value of RBF network:

${\hat{x}}_{i+p}=\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{w}_j{\tilde{x}}_{i+p}^j=w_1{\tilde{Y}}_{i+\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="HSA00000334446500021.png,HSA00000334446500031.png"\; state="\{\{state\}\}">p</figure-callout>}}^1\&CirclePlus;w_2{\tilde{Y}}_{i+p+1}^2\&CirclePlus;...\&CirclePlus;w_p{\tilde{Y}}_{i+2p-1}^p.$

(3) equal update time during t when the sampling interval, carry out the integrated prediction model modification: based on historical data and image data structure phase space reconfiguration X in real time _iWith p step predicted vector Y _I+p, form incremental training sample set { (X _iY _I+p) | i=l+1-p ..., l+t-p} is at f ^pCarry out the increment type training on the basis and be designated as f ^P+1, upgrade the contained predictor model of integrated prediction model by " FIFO " lining up mode again, deposit all t real-time image data in historical data base then, upgrade historical data base length l=l+t;

(4) judge whether to continue to gather, be then to return (1) to continue, otherwise finish.

Described real-time predicting method, the firedamp sensor of the described first step adopts gas wireless monitor sensor, be placed on one's body rib the place ahead and extractive equipment and the operating personnel, and the mobile base station is set outside 50～100 meters accepts gas density information, reach ground monitoring system by the underground communica tion network again.

Described real-time predicting method, the C-C method in described second step calculate the delay time T of chaos time sequence and embed dimension m and carry out according to following steps: at first, calculate following three parameters:

$\mathrm{\&Delta;}\stackrel{\&OverBar;}{S}\left(t\right)=\frac{1}{4}\underset{m=2}{\overset{5}{\mathrm{\&Sigma;}}}\mathrm{\&Delta;S}(m,t),$ $\stackrel{\&OverBar;}{S}\left(t\right)=\frac{1}{16}\underset{m=2}{\overset{5}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{4}{\mathrm{\&Sigma;}}}S(m,r_k,t),$ $S_{\mathrm{cor}}\left(t\right)=\mathrm{\&Delta;}\stackrel{\&OverBar;}{S}\left(t\right)+\stackrel{\&OverBar;}{S}\left(t\right),$

In the formula,

(m=2,3,4,5;

);

ΔS(m，t)＝max{S(m，r _j，t)}-min{S(m，r _j，t)}；

$C(m,r_k,t)=\underset{N\&RightArrow;\&infin;}{\lim }\frac{2}{m(m-1)}\underset{1\&le;i\&le;j\&le;m}{\mathrm{\&Sigma;}}\mathrm{\&delta;}(r_k-||X_i-X_j|\left|\right);$

Impulse function $\mathrm{\&delta;}\left(x\right)=\left\{\begin{array}{cc}1,& x\&GreaterEqual;0\\ 0,& x<0\end{array}\right.;$

Then, get

The t value of first minimum correspondence is for getting S time delays again _Cor(t) the t value of (0≤t≤200) global minimum correspondence is time window τ _w, calculate thus and embed dimension: m=τ _w/ τ+1.

Described real-time predicting method, integrated capacity parameter p lower limit equals 1 in described the 3rd step, and the upper limit equals chaos time sequence maximum Lyapunov exponent λ ₁Inverse 1/ λ ₁, i.e. 1≤p≤1/ λ ₁, maximum Lyapunov exponent λ wherein ₁Carry out according to following steps:

(1) to time series { x _i| i=1,2 ..., n} carries out Fourier transform FFT, and calculating average period is P;

(2) calculate embedding dimension m and time lag τ according to the C-C method, and phase space reconstruction:

{Y _i|i＝1，2，…，M＝n-(m-1)τ}；

(3) seek each some Y of phase space _jNearest neighbor point And limit of short duration separation:

$d_j\left(0\right)=\underset{j^{*}}{\min }\left|\right|Y_j-Y_{j^{*}}\left|\right|,\left|\right|j-j^{*}\left|\right|>P;$

(4) calculate phase space point Y _jDistance after right i the discrete step of adjoint point:

$d_j\left(i\right)=\left|\right|Y_{j+i}-Y_{j^{*}+i}\left|\right|,i=\mathrm{1,2},...,\min (M-j,M-j^{*});$

(5) to i=1,2 ..., min (M-j, M-j ^*), calculate Q is non-zero d _j(i) number,

Do regression straight line based on least square method, this straight slope is maximum Lyapunov exponent λ ₁

Described real-time predicting method, described the 3rd update time in step t setting, determine according to the time lag τ of chaos time sequence: t=k * τ, wherein k is an empirical coefficient, can select k=0.5～1.5.

Described real-time predicting method, the method of blended learning strategy is adopted in described the 4th step radial primary function network monolithic training, from the input layer to the hidden layer, utilize the self-organizing clustering method to determine the suitable data center, and determine the expansion constant of hidden node according to the distance between each data center for the RBF of hidden node; From the hidden layer to the output layer, utilize the corresponding weight of gradient descent algorithm algorithm training.

Described real-time predicting method, the increment type training of described the 4th step radial primary function network, when the training sample amount hour, can carry out according to RBF monolithic training method; Otherwise carry out according to following steps:

(1) on existing RBF network foundation, increases a training sample (X; Y), find the nearest C of data center of corresponding input X earlier ^*, and calculate corresponding output error E;

(2) if E and ‖ X-C ^*‖ then increases a latent node respectively greater than corresponding threshold value on the RBF network, generate a long-term memory item simultaneously; Otherwise, generate a series of pseudo-training samples based on all long-term memory items, calculate its hidden layer output valve then and form matrix Φ, utilize singular value decomposition to calculate RBF network weight matrix, calculate the RBF network output error E of input X correspondence then, if E then increases a latent node greater than corresponding threshold value on the RBF network;

(3) during given another new training sample, return (1) and continue.

Described real-time predicting method, described the 5th step radial primary function network integrated prediction result's calculating according to the form below carries out:

Wherein, the integrated prediction of RBF network is output as

L is the data length of historical data base; T is update time; P is a capacity parameter; w _jBe j RBF network correspondence weight (j=1,2 ..., p), satisfy normalizing condition

Described real-time predicting method, described the 5th step RBF network integrated prediction model upgrades according to " FIFO " queue sequence.

The present invention is applicable to the real-time of mine gas concentration is accurately predicted, can realize the forecast from short-term to the medium-term forecast scope, meets the technical requirements of current mine gas management system.Compare with the gas Forecasting Methodology of present existence, this method has the following advantages:

One, determines the estimation range of mine gas from the short-term to the mid-term by setting integrated capacity parameter.Capacity parameter has characterized the integrated RBF network model number that can hold, and is known by the integrated theory of neutral net, and the integrated model that is fit to the submodel composition of number has more outstanding popularization performance, has so just guaranteed the integrated forecasting accuracy of RBF network; Simultaneously, the RBF network prediction step of this method design just equals integrated capacity parameter, and predicting the outcome synchronously of all submodels is weighted on average as RBF integrated prediction result, realizes the prediction of mine gas concentration from the short to medium term scope.So, this method obtains between estimation range and precision of prediction require and obtains suitable compromisely, satisfies the technical requirements of mine gas information management system aspect a middle or short term real-time estimate.

Two, adopt the integrated prediction model that comprises a plurality of radial basis function neural networks, can utilize real-time image data that single RBF network is carried out the increment type training, and the integrated model that upgrades in time of the queue sequence by FIFO, guarantee required precision to the mine gas real-time estimate.

Description of drawings

Fig. 1: based on the integrated Forecasting Methodology overview flow chart of radial basis function neural network;

Fig. 2: the integrated initial phase flow chart of radial basis function neural network;

Fig. 3: flow chart of integrated real-time estimate stage of radial basis function neural network.

The specific embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail (referring to Fig. 1).

The first step, layout gas wireless monitor sensor are on one's body rib the place ahead, extractive equipment and operating personnel, with the truth of reflection with work plane forward position gas emission in the continuous motion process of excavator, and the mobile base station is set outside 50～100 meters accepts gas density information, gas density is passed to the historical data base X of ground monitoring system by the underground communica tion network _Lib, have

X _lib＝{x _i|i＝1，2，...，l}(l＝n+2p) (1)

Second step, according to the Takens theorem, to suitable embedding dimension m and time lag τ, " path " of reconstruction attractor in embedded space is the kinetics equivalence with original system under the differomorphism meaning.The gas data of historical data base are considered as chaos time sequence, and the C-C method of introducing at " Chaotic Time Series Analysis and application thereof " book in 2002 according to Lv Jinhu etc. is calculated the reconstruction parameter of phase space then: embed dimension m and time lag τ.

The 3rd step, set the integrated capacity parameter p of radial basis function neural network and update time t.

(1) capacity parameter p is set at 1≤p≤1/ λ ₁, λ ₁Be the maximum Lyapunov exponent of chaos time sequence, the maximum Lyapunov exponent of introducing at " Chaotic Time Series Analysis and application thereof " book in 2002 according to Lv Jinhu etc. is improved algorithm and is calculated;

(2) update time t setting, determine according to the time lag τ of chaos time sequence, i.e. t=k * τ, wherein k is an empirical coefficient, can select k=0.5～1.5.

The 4th step, the integrated forecast model (referring to Fig. 2) of initialization RBF (RBF) network, as follows:

(1) sets up phase space reconfiguration X respectively _iWith p step predicted vector Y _I+p(i=1+ (m-1) τ ..., n+p),

X _i＝[x _i-(m-1)τ，...，x _i-τ，x _i] (2)

Y _i+p＝[x _i+1，x _i+2，...，x _i+p] (3)

Wherein, x _iThe gas density data of gathering constantly for i.

(2) according to training sample set

{(X _i；Y _i+p)|i＝1+(m-1)τ，...，n+1} (4)

The forecast model of monolithic training radial primary function network (RBF) is f ¹Wherein the method for blended learning strategy is adopted in the monolithic training, from the input layer to the hidden layer, utilize the self-organizing clustering method to determine the suitable data center, and determine the expansion constant of hidden node according to the distance between each data center for the RBF of hidden node; From the hidden layer to the output layer, utilize the corresponding weight of gradient descent algorithm training.

(3) again according to training sample set

{(X _i；Y _i+p)|i＝1+(m-1)τ，...，n+2 (5)

Monolithic training RBF prediction network,

{(X _i；Y _i+p)|i＝n+2} (6)

Or by the incremental training sample set at f ¹Increment type training RBF prediction network is f on the basis ²The increment type training is carried out according to following steps:

1. on existing RBF network foundation, increase training sample (X; Y), find the nearest C of data center of corresponding input X earlier ^*, and calculate corresponding output error E;

2. if E and ‖ X-C ^*‖ then increases a latent node respectively greater than corresponding threshold value on the RBF network, generate a long-term memory item simultaneously; Otherwise, generate a series of pseudo-training samples based on all long-term memory items, calculate its hidden layer output valve then and form matrix Φ, utilize singular value decomposition to calculate RBF network weight matrix, calculate the RBF network output error E of input X correspondence then, if E then increases a latent node greater than corresponding threshold value on the RBF network;

During 3. given another new training sample, return 1. and continue.

(4) by that analogy, up to according to training sample set

{(X _i；Y _i+p)|i＝1+(m-1)τ，...，n+p (7)

Monolithic training RBF prediction network, or by the incremental training sample set

{(X _i；Y _i+p)|i＝n+p} (8)

At f ^P-1Increment type training RBF prediction network is f on the basis ^p

(5) finish the integrated initial work of radial basis function neural network, obtain containing the integrated prediction model f of p RBF network ¹, f ²..., f ^p

The 5th the step, carry out real-time estimate (referring to Fig. 3) according to RBF network integrated prediction model, as follows:

(1) by the real-time image data of firedamp sensor

X _real＝{x _i|i＝l+1，l+2，...，l+t} (9)

In conjunction with gas density historical data base X _Lib, carry out phase space reconfiguration

X _i＝[x _i-(m-1)τ，...，x _i-τ，x _i](i＝1+(m-1)τ，...，l+t-1) (10)

(2) according to phase space X _iCarry out the prediction work of RBF network integrated model, respectively with X _i, X _I+1..., X _I+p-1(i=l+1-p ..., l+t-1) bring forecast model f into ¹, f ²..., f ^p, have

$\left\{\begin{array}{c}{\tilde{Y}}_{i+\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="HSA00000334446500021.png,HSA00000334446500031.png"\; state="\{\{state\}\}">p</figure-callout>}}^1=[{\tilde{x}}_{i+1}^1,{\tilde{x}}_{i+2}^1,...,{\tilde{x}}_{i+p}^1]=f^1\left(X_i\right)\\ {\tilde{Y}}_{i+p+1}^2=[{\tilde{x}}_{i+2}^2,...,{\tilde{x}}_{i+p}^2,{\tilde{x}}_{i+p+1}^2]=f^2\left(X_{i+1}\right)\\ ......\\ {\tilde{Y}}_{i+2p-1}^p=[{\tilde{x}}_{i+p}^p,{\tilde{x}}_{i+p+1}^p,...,{\tilde{x}}_{i+2p-1}^p]=f^p\left(X_{i+p-1}\right)\end{array}\right.---\left(11\right)$

Obtain x _I+pP predict the outcome synchronously

Then its weighted average (is defined as p predicted vector

Weighting type

Operator) as the integrated predicted value of RBF network,

${\hat{x}}_{i+p}=\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{w}_j{\tilde{x}}_{i+p}^j\&equiv;\underset{j=1}{\overset{p}{\&CirclePlus;}}{w}_j{\tilde{Y}}_{i+p+j-1}^j=w_1{\tilde{Y}}_{i+\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="HSA00000334446500021.png,HSA00000334446500031.png"\; state="\{\{state\}\}">p</figure-callout>}}^1\&CirclePlus;...\&CirclePlus;w_p{\tilde{Y}}_{i+2p-1}^p---\left(12\right)$

Wherein, w _jBe j RBF network correspondence weight (j=1,2 ..., p), satisfy normalizing condition:

$\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{w}_j=1---\left(13\right)$

A kind of simple case is:

w _j＝1/p(j＝1，2，...，p)， (14)

The integrated prediction output of RBF this moment is the simple average that all are predicted the outcome synchronously, promptly

${\hat{x}}_{i+p}=\frac{1}{p}\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{\tilde{x}}_{i+p}^j;---\left(15\right)$

Another kind of situation is, considers that p RBF network comprised maximum so far training sample information,

So can carry out the weighted average that all predict the outcome synchronously according to the weight of arithmetic progression, promptly

$w_j=\frac{1}{2p}+\frac{j}{p(p+1)}(j=\mathrm{1,2},...,p)---\left(16\right)$

(3) equal update time during t when the sampling interval, enter the integrated prediction model modification stage.Based on historical data and real-time image data structure phase space reconfiguration X _iWith p step predicted vector Y _I+p, form incremental training sample set { (X _iY _I+p) | i=l+1-p ..., l+t-p} is at f ^pCarry out the increment type training on the basis and be designated as f ^P+1, upgrade integrated contained predictor model by " FIFO " lining up mode again, promptly upgrade in proper order according to the described First Input First Output of following table, deposit t real-time image data in historical data base then, upgrade historical data base length l=l+t.

Wherein, known integrated prediction model comprises f ¹, f ²..., f ^pBe total to p RBF network; For the real-time sampling concentration value of mine gas, at f ^pTrain according to increment type training strategy on the basis, its result is designated as f ^P+1, constitute formation f ¹, f ²..., f ^p, f ^P+1Because the finite capacity (can only hold p model) of integrated model so upgrade according to the order of " FIFO ", promptly abandons f ¹And reservation f ^P+1, renumber in order again, finish RBF network integrated model and upgrade.

(4) judge whether to continue to gather, be then to return (1) to continue, otherwise finish.

Should be understood that, for those of ordinary skills, can be improved according to the above description or conversion, and all these improvement and conversion all should belong to the protection domain of claims of the present invention.

Claims (9)

1. one kind based on integrated mine gas concentration real-time predicting method a middle or short term of radial basis function neural network, it is characterized in that, comprises the steps:

The first step, gather the gas density data, deposit the gas density historical data base in by firedamp sensor

X _lib＝{x _i|i＝1，2，...，l}(l＝n+2p)；

Second step, the gas density data that database is deposited are handled as chaos time sequence, utilize the delay time T of the C-C method sequence of calculation and embed dimension m;

The 3rd step, set the integrated capacity parameter p of radial basis function neural network and update time t;

The 4th the step, set up the integrated forecast model of radial basis function neural network, step is as follows:

(1) sets up phase space reconfiguration X respectively _iWith p step predicted vector Y _I+p(i=1+ (m-1) τ ..., n+p),

X _i＝[x _i-(m-1)τ，...，x _i-τ，x _i]，

Y _i+p＝[x _i+1，x _i+2，...，x _i+p]，

Wherein, x _iThe gas density data of gathering constantly for i;

(2) according to training sample set { (X _iY _I+p) | i=1+ (m-1) τ ..., n+1}, the prediction network of monolithic training radial primary function network (RBF) is f ¹

(3) again by incremental training sample set { (X _iY _I+p) | i=n+2} is at f ¹Increment type training RBF prediction network on the basis, or by { (X _iY _I+p) | i=1+ (m-1) τ ..., n+2} monolithic training RBF network is f ²

(4) by that analogy, up to by incremental training sample set { (X _iY _I+p) | i=n+p} is at f ^P-1Increment type training or by { (X on the basis _iY _I+p) | i=1+ (m-1) τ ..., n+p} monolithic training RBF network is f ^p

(5) finish the integrated initial work of radial basis function neural network, obtain containing the integrated prediction model f of p RBF network ¹, f ²..., f ^p

The 5th the step, carry out real-time estimate according to RBF network integrated model, step is as follows:

(1) by the real-time image data X of firedamp sensor _Real={ x _i| i=l+1, l+2 ..., l+t} is in conjunction with gas density historical data base X _Lib, carry out phase space reconfiguration

X _i＝[x _i-(m-1)τ，...，x _i-τ，x _i](i＝l+1-p，...，l+t-1)；

(2) according to phase space X _iCarry out the prediction work of RBF network integrated model, respectively with X _i, X _I+1..., X _I+p-1(i=l+1-p ..., l+t-1) bring forecast model f into ¹, f ²..., f ^p, have

$\left\{\begin{array}{c}{\tilde{Y}}_{i+p}^1=[{\tilde{x}}_{i+1}^1,{\tilde{x}}_{i+2}^1,...,{\tilde{x}}_{i+p}^1]=f^1\left(X_i\right)\\ {\tilde{Y}}_{i+p+1}^2=[{\tilde{x}}_{i+2}^2,...,{\tilde{x}}_{i+p}^2,{\tilde{x}}_{i+p+1}^2]=f^2\left(X_{i+1}\right)\\ ......\\ {\tilde{Y}}_{i+2p-1}^p=[{\tilde{x}}_{i+p}^p,{\tilde{x}}_{i+p+1}^p,...,{\tilde{x}}_{i+2p-1}^p]=f^p\left(X_{i+p-1}\right)\end{array}\right.,$

Obtain x _I+pP predict the outcome synchronously

Then with its weighted average as the integrated predicted value of RBF network:

${\hat{x}}_{i+p}=w_1{\tilde{Y}}_{i+p}^1\&CirclePlus;w_2{\tilde{Y}}_{i+p+1}^2\&CirclePlus;...\&CirclePlus;w_p{\tilde{Y}}_{i+2p-1}^p=\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{w}_j{\tilde{x}}_{i+p}^j.$

(3) equal update time during t when the sampling interval, carry out the integrated prediction model modification: based on historical data and image data structure phase space reconfiguration X in real time _iWith p step predicted vector Y _I+p, form incremental training sample set { (X _iY _I+p) | i=l+1-p ..., l+t-p} is at f ^pCarry out the increment type training on the basis and be designated as f ^P+1, upgrade the contained predictor model of integrated prediction model by " FIFO " queue sequence again, deposit all t real-time image data in historical data base then, upgrade historical data base length l=l+t;

(4) judge whether to continue to gather, be then to return (1) to continue, otherwise finish.

2. real-time predicting method according to claim 1, it is characterized in that: the firedamp sensor of the described first step adopts gas wireless monitor sensor, be placed on one's body rib the place ahead and extractive equipment and the operating personnel, and the mobile base station is set outside 50～100 meters accepts gas density information, reach ground monitoring system by the underground communica tion network again.

3. real-time predicting method according to claim 1 is characterized in that: the C-C method in described second step is calculated the delay time T of chaos time sequence and is embedded dimension m and carries out according to following steps: at first, calculate following three parameters:

$\mathrm{\&Delta;}\stackrel{\&OverBar;}{S}\left(t\right)=\frac{1}{4}\underset{m=2}{\overset{5}{\mathrm{\&Sigma;}}}\mathrm{\&Delta;S}(m,t),$ $\stackrel{\&OverBar;}{S}\left(t\right)=\frac{1}{16}\underset{m=2}{\overset{5}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{4}{\mathrm{\&Sigma;}}}S(m,r_k,t),$ $S_{\mathrm{cor}}\left(t\right)=\mathrm{\&Delta;}\stackrel{\&OverBar;}{S}\left(t\right)+\stackrel{\&OverBar;}{S}\left(t\right),$

In the formula,

(m=2,3,4,5; );

ΔS(m，t)＝max{S(m，r _j，t)}-min{S(m，r _j，t)}；

$C(m,r_k,t)=\underset{N\&RightArrow;\&infin;}{\lim }\frac{2}{m(m-1)}\underset{1\&le;i\&le;j\&le;m}{\mathrm{\&Sigma;}}\mathrm{\&delta;}(r_k-||X_i-X_j|\left|\right);$

Impulse function $\mathrm{\&delta;}\left(x\right)=\left\{\begin{array}{cc}1,& x\&GreaterEqual;0\\ 0,& x<0\end{array}\right.;$

Then, get

The t value of first minimum correspondence is a delay time T; Get S again _Cor(t) the t value of (0≤t≤200) global minimum correspondence is time window τ _w, calculate thus and embed dimension: m=τ _w/ τ+1.

4. real-time predicting method according to claim 1 is characterized in that: integrated capacity parameter p lower limit equals 1 in described the 3rd step, and the upper limit equals chaos time sequence maximum Lyapunov exponent λ ₁Inverse 1/ λ ₁, i.e. 1≤p≤1/ λ ₁, maximum Lyapunov exponent λ wherein ₁Carry out according to following steps:

(1) to time series { x _i| i=1,2 ..., n} carries out Fourier transform FFT, and calculating average period is P;

(2) calculate embedding dimension m and time lag τ according to the C-C method, and phase space reconstruction:

{Y _i|i＝1，2，…，M＝n-(m-1)τ}；

(3) seek each some Y of phase space _jNearest neighbor point

And limit of short duration separation:

$d_j\left(0\right)=\underset{j^{*}}{\min }\left|\right|Y_j-Y_{j^{*}}\left|\right|,\left|\right|j-j^{*}\left|\right|>P;$

(4) calculate phase space point Y _jDistance after right i the discrete step of adjoint point:

$d_j\left(i\right)=\left|\right|Y_{j+i}-Y_{j^{*}+i}\left|\right|,i=\mathrm{1,2},...,\min (M-j,M-j^{*});$

(5) to i=1,2 ..., min (M-j, M-j ^*), calculate

Q is non-zero d _j(i) number is done regression straight line based on least square method, and this straight slope is maximum Lyapunov exponent λ ₁

5. real-time predicting method according to claim 1 is characterized in that: described the 3rd update time in step t setting, determine according to the time lag τ of chaos time sequence: t=k * τ, wherein k is an empirical coefficient, can select k=0.5～1.5.

6. real-time predicting method according to claim 1, it is characterized in that: the method for blended learning strategy is adopted in described the 4th step radial primary function network monolithic training, from the input layer to the hidden layer, utilize the self-organizing clustering method to determine the suitable data center, and determine the expansion constant of hidden node according to the distance between each data center for the RBF of hidden node; From the hidden layer to the output layer, utilize the corresponding weight of gradient descent algorithm algorithm training.

7. real-time predicting method according to claim 1 is characterized in that: the increment type training of described the 4th step radial primary function network, when the training sample amount hour, can carry out according to RBF monolithic training method; Otherwise carry out according to following steps:

(1) on existing RBF network foundation, increases a training sample (X; Y), find the nearest C of data center of corresponding input X earlier ^*, and calculate corresponding output error E;

(2) if E and ‖ X-C ^*‖ then increases a latent node respectively greater than corresponding threshold value on the RBF network, generate a long-term memory item simultaneously; Otherwise, generate a series of pseudo-training samples based on all long-term memory items, calculate its hidden layer output valve then and form matrix Φ, utilize singular value decomposition to calculate RBF network weight matrix, calculate the RBF network output error E of input X correspondence then, if E then increases a latent node greater than corresponding threshold value on the RBF network;

(3) during given another new training sample, return (1) and continue.

8. real-time predicting method according to claim 1 is characterized in that: described the 5th step radial primary function network integrated prediction result's calculating according to the form below carries out:

Wherein, the integrated prediction of RBF network is output as

L is the data length of historical data base; T is update time; P is a capacity parameter; w _jBe j RBF network correspondence weight (j=1,2 ..., p), satisfy normalizing condition

9. real-time predicting method according to claim 1 is characterized in that: described the 5th step RBF network integrated prediction model, upgrade according to the queue sequence of FIFO.

Images