CN102968644A - Method for predicting smelting finishing point of argon-oxygen refined iron alloy - Google Patents

Abstract

The invention discloses a method for predicting a smelting finishing point of argon-oxygen refined iron alloy, which aims at solving the difficult problem that the finishing point is judged depending on manual fire observation operation in a low-carbon ferrochrome process in AOD (Argon Oxygen Decarburization) furnace smelting. The method adopts a machine vision technology to simulate the conventional manual fire observation process to judge the terminal. According to the method, the characteristics of flame at an AOD furnace inlet are extracted by virtue of gray levels; the training and the testing of image characteristics are implemented by a support vector, machine (SVM) algorithm; and the image characteristics of the flame are extracted by the machine vision technology, and the smelting finishing point is effectively identified by combining the machine vision technology with a support vector machine, and thus the method can achieve relatively high identification precision.

Description

A kind of argon oxygen refining ferroalloys smelting endpoint Forecasting Methodology

Technical field

The present invention relates to a kind of argon oxygen refining ferroalloys smelting endpoint Forecasting Methodology.

Background technology

Terminal point control is argon oxygen refining (Argon-Oxygen Decarburization, AOD) the middle-low-carbon ferrochrome alloy important operation in latter stage, its objective is aim carbon and the molten iron temperature of fire door are controlled, and makes both reach simultaneously the requirement of tapping a blast furnace.At present, China AOD method refining low-carbon ferrochromium alloy generally adopts empirical method, namely manually sees pyrogenic process.The method randomness is strong, and large to smelter's operant level dependence, the terminal point hit rate is unstable, often will melt down blowing in the smelting, causes production cost to improve the wasting of resources.At present, sublance detects and the mass spectrometric gentle terminal point hit rate of the raising smelting Automated water aspect that is introduced in has obtained sizable progress, but this method cost is high, the investment large, can't popularization and application on middle-size and small-size smelting equipment.After 1997, engender the method measurement composition of fumes that penetrates the fire door furnace gas with infrared laser abroad, and the optical sensor of terminal point control occurred being used for.This method is fit to low-carbon (LC) and detects, but since the optical sensor work place from fire door close to, cause the instrument life-span to be challenged.After entering 21 century, the domestic technology that has proposed to extract eigenwert by light intensity and the image information of furnace mouth radiation is as the foundation of judging terminal point, but the precision of prediction in the experiment is unsatisfactory.

Summary of the invention

The purpose of this invention is to provide a kind of argon oxygen refining ferroalloys smelting endpoint Forecasting Methodology, the method is a kind of AOD stove terminal point SVM prediction method based on machine vision, identify and extract smelting middle-low-carbon ferrochrome later stage AOD stove fire door flame characteristic by Vision Builder for Automated Inspection, utilize support vector machine (Support Vector Machine, SVM) set up the End-point Prediction model, identification is smelted and whether is finished, and finally determines smelting endpoint.The method can be in time, terminal point forecasts with unerring accuracy.

The present invention's method comprises following concrete steps:

(1), the flame characteristic based on machine vision extracts

(1), the formation of image capturing system

Image capturing system is comprised of optical image sensor ccd video camera, image pick-up card and industrial computer, the image information of ccd video camera is delivered to image pick-up card on the industrial computer by concentric cable, image pick-up card converts vision signal to digital signal, sends into industrial computer and carries out corresponding digital processing, image demonstration and forecast analysis.

Ccd video camera is converted into electric signal with the light signal of AOD stove fire door flame, image pick-up card is stored in the simulating signal of ccd video camera output in the computing machine through over-sampling, after discrete, and the gradation conversion of each sampled point become 0～255 gray level, deposit in order buffer register in, computing machine conducts interviews to each pixel through buffer register.

(2), feature selecting

The basic demand of terminal point control is when finishing blowing, and the chemical constitution in the molten iron and temperature reach the requirement of coming out of the stove simultaneously, and the AOD stove is smelted the middle-low-carbon ferrochrome later stage, because carbon will exhaust in the stove, reaction between carbon and oxygen tends towards stability, and fire door flame changes slowly, color jaundice deliquescing; By the flame available gray-scale (

), the high-temperature area gray scale (

), the flame useful area ( ), the high-temperature area area (

) can make fast flame characteristic and describing, describe smelting endpoint with [1,1] and whether reach.

(2-1) the flame available gray-scale (

)

Determine by experiment the threshold value of flame available gray-scale, the part that pixel grey scale is higher than this threshold value is considered as effective flame region.

(2-2) the high-temperature area gray scale (

)

Be determined by experiment the high-temperature area gray threshold, the part that pixel grey scale is higher than this threshold value is considered as the flame high-temperature area.

(2-3) the flame useful area ( )

The computing formula of flame area is

$S_i=\underset{j=1}{\overset{G}{\mathrm{\Σ}}}L(g_i^j-g_d)$

In the formula, S _iBe the flame area of the i time sampling, G is the image pixel number,

The gray-scale value of pixel j in the image when representing the i time sampling; G _dBe predefined threshold value, rational threshold value is chosen the calculating of " flame pixels area " most important, can be by the observation definite threshold to on-the-spot flame image and furnace flame video recording, and L (.) is step function, characterizes the fluctuation of fire door flame fringe,

Get 1,

Get 0.

(2-4) flame high temperature area (

)

With respect to the flame area of flame high-temperature area gray scale, consistent with flame useful area acquisition methods.

(3), normalized

Because the proper vector implication of choosing is different, dimension is different, and it is larger that numerical value differs; If directly use these variablees as the sample of training network, just then learning process can be handled by the large variable fluctuation of eigenwert, can't embody the situation of change of little eigenwert, therefore, need to do normalized to above-mentioned characteristic variable; Processing Algorithm is

$x=\frac{\stackrel{\‾}{x}-{\stackrel{\‾}{x}}_{\min }}{{\stackrel{\‾}{x}}_{\max }-{\stackrel{\‾}{x}}_{\min }}$

In the formula,

, x is respectively the value before and after the reduction; ,

Be respectively maximal value and the minimum value of sample.

(2), identify based on the smelting endpoint of SVM

(1), algorithm of support vector machine

In two quasi-mode identification problems, given training sample (x _i, y _i), x _iThe input of i sample, y _iThe desired output of i sample, x _i∈ R ⁿ, y _i∈ (1 ,+1), i=1,2 ..., l, the target of SVM is exactly by discriminant function of training sample structure, and training sample is separated with largest interval, and will as much as possible correctly classification of sample be adjudicated.Classifying rules is

$f\left(x\right)=\mathrm{sgn}[\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{a}_i^{*}(x_i,x)+b^{*}]$

In the formula, sgn () is sign function;

The optimum solution of Lagrange coefficient, b ^*It is the threshold value of classification; Find the solution classification function and need to construct optimization problem suc as formula [1] and formula [2]

$\underset{w,\mathrm{\ξ},b}{\min }\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i---\left[1\right]$

$\mathrm{Sub}.\mathrm{to}\left\{\begin{array}{c}{y}_i[(w\·\mathrm{\φ}\left(x_i\right))+b]+{\mathrm{\ξ}}_i\≥1\\ {\mathrm{\ξ}}_i\≥0,i=\mathrm{1,2},...,l\end{array}\right.---\left[2\right]$

In the formula, w is adjustable weight vectors, and C is penalty factor, ξ _iBe slack variable, φ () is the Nonlinear Mapping from the sample space to the higher dimensional space.

Then the Wolfe antithesis of formula [1] and formula [2] is

$\underset{\mathrm{\α}}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{y}_j{\mathrm{\α}}_i{\mathrm{\α}}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i---\left[3\right]$

$\mathrm{Sub}.\mathrm{to}\left\{\begin{array}{c}{\underset{i=1}{\overset{l}{\mathrm{\Σ}}}y}_i{\mathrm{\α}}_i=0\\ 0\≤{\mathrm{\α}}_i\≤C,i=\mathrm{1,2},...,l\end{array}\right.---\left[4\right]$

If the optimization problem of formula [3] and formula [4] statement has optimum solution

${\mathrm{\α}}^{*}=({\mathrm{\α}}_1^{*},{\mathrm{\α}}_2^{*},...,{\mathrm{\α}}_l^{*})$

According to optimality condition---KKT, the solution of this optimization problem must satisfy

α _i{[(x _i·w)+b]-y _i-1}=0 [5]

Finding the solution the optimal classification function that obtains after the problems referred to above is

$f\left(x\right)=\mathrm{sgn}[\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x)+b^{*}]---\left[6\right]$

$b^{*}=y_i-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x_j)---\left[7\right]$

$w^{*}=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}{x}_i---\left[8\right]$

In the formula [7] K () in formula [3]～formula [7] formula is the kernel function that satisfies the Mercer condition, its effect is exactly the linearly inseparable problem of the former input space is converted into the linear separability problem in higher-dimension even Infinite-dimensional Hibert space, then finds the solution the optimization problem of formula [3] and formula [4] at this higher-dimension or Infinite-Dimensional Space.

At present, the kernel function of most study mainly contains following three kinds of forms

1) polynomial kernel

K(x,x _i)=[(x,x _i)+1] ^q

2) radial basis nuclear

$K(x,x_i)=\exp [-\frac{{\left|\right|x-x_i\left|\right|}^2}{{\mathrm{\σ}}^2}]$

3) Sigmoid nuclear

K(x,x _i)=tanh(v(x·x _i)+C

In the formula, q, σ, v are the parameter of corresponding kernel function, and parameter is selected to obtain by experiment.

(2), flame identification

Utilize the concrete steps of SVM identification terminal point flame as follows:

1) obtains the training sample data;

2) inner product function (being kernel function) and the penalty factor of selection nonlinear transformation;

3) form protruding double optimization problem, find the solution this optimization problem to obtain the support vector machine of classification;

4) the sample to be tested data are sent into the support vector machine of classification, obtained recognition result.

Wherein, the 1st) step realizes that by the image characteristics extraction algorithm of step () four eigenwerts of every two field picture consist of an input vector of support vector machine; The 2nd) step is to choose kernel function, chooses by experiment; The 3rd) step is to set up the svm classifier device, finds the solution optimum solution by algorithm of support vector machine

Try to achieve weight vectors w according to formula [7] and formula [8] again ^*With side-play amount b ^*At last will

And b ^*Substitution formula [6] obtains the optimal classification function; The 4th) step is to utilize SVM identification flame, the sample to be tested data is sent into the SVM that trains, the recognition accuracy of the SVM that checking is trained.

Beneficial effect of the present invention: the present invention utilizes gray level to extract AOD fire door flame characteristic, adopts support vector machine (SVM) algorithm to realize the training and testing of characteristics of image.The method that adopts machine vision technique to extract Flame Image Characteristics and be combined with support vector machine can effectively be identified smelting endpoint, and has preferably accuracy of identification.

Description of drawings

Fig. 1 is the formation synoptic diagram of the used image capturing system of the present invention.

Embodiment

The present invention's method comprises following concrete steps:

(1), the flame characteristic based on machine vision extracts

(1), the formation of image capturing system

Image capturing system is comprised of optical image sensor ccd video camera 1, image pick-up card 2 and industrial computer 3, the image information of ccd video camera 1 is delivered to image pick-up card 2 on the industrial computer 3 by concentric cable, image pick-up card 2 converts vision signal to digital signal, send into industrial computer 3 and carry out corresponding digital processing, image demonstration and forecast analysis, image capturing system as shown in Figure 1.

Ccd video camera 1 is converted into electric signal with the light signal of AOD stove A fire door flame, image pick-up card 2 is stored in the simulating signal of ccd video camera 1 output in the computing machine through over-sampling, after discrete, and the gradation conversion of each sampled point become 0～255 gray level, deposit in order buffer register in, computing machine conducts interviews to each pixel through buffer register.

(2), feature selecting

The basic demand of terminal point control is when finishing blowing, and the chemical constitution in the molten iron and temperature reach the requirement of coming out of the stove simultaneously, and the AOD stove is smelted the middle-low-carbon ferrochrome later stage, because carbon will exhaust in the stove, reaction between carbon and oxygen tends towards stability, and fire door flame changes slowly, color jaundice deliquescing; By the flame available gray-scale (

), the high-temperature area gray scale (

), the flame useful area (

), the high-temperature area area (

) can make fast flame characteristic and describing, describe smelting endpoint with [1,1] and whether reach.

(2-1) the flame available gray-scale (

)

Determine by experiment the threshold value of flame available gray-scale, the part that pixel grey scale is higher than this threshold value is considered as effective flame region.

(2-2) the high-temperature area gray scale (

)

Be determined by experiment the high-temperature area gray threshold, the part that pixel grey scale is higher than this threshold value is considered as the flame high-temperature area.

(2-3) the flame useful area (

)

The computing formula of flame area is

$S_i=\underset{j=1}{\overset{G}{\mathrm{\Σ}}}L(g_i^j-g_d)$

In the formula, S _iBe the flame area of the i time sampling, G is the image pixel number,

The gray-scale value of pixel j in the image when representing the i time sampling; g _dBe predefined threshold value, rational threshold value is chosen the calculating of " flame pixels area " most important, can be by the observation definite threshold to on-the-spot flame image and furnace flame video recording, and L (.) is step function, characterizes the fluctuation of fire door flame fringe,

Get 1, Get 0.

(2-4) flame high temperature area (

)

With respect to the flame area of flame high-temperature area gray scale, consistent with flame useful area acquisition methods.

(3), normalized

Because the proper vector implication of choosing is different, dimension is different, and it is larger that numerical value differs; If directly use these variablees as the sample of training network, just then learning process can be handled by the large variable fluctuation of eigenwert, can't embody the situation of change of little eigenwert, therefore, need to do normalized to above-mentioned characteristic variable; Processing Algorithm is:

$x=\frac{\stackrel{\‾}{x}-{\stackrel{\‾}{x}}_{\min }}{{\stackrel{\‾}{x}}_{\max }-{\stackrel{\‾}{x}}_{\min }}$

In the formula, , x is respectively the value before and after the reduction;

,

Be respectively maximal value and the minimum value of sample.

(2), identify based on the smelting endpoint of SVM

(1), algorithm of support vector machine

In two quasi-mode identification problems, given training sample (x _i, y _i), x _iThe input of i sample, y _iThe desired output of i sample, x _i∈ R ⁿ, y _i∈ (1 ,+1), i=1,2 ..., l, the target of SVM is exactly by discriminant function of training sample structure, and training sample is separated with largest interval, and will as much as possible correctly classification of sample be adjudicated.Classifying rules is

$f\left(x\right)=\mathrm{sgn}[\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{a}_i^{*}(x_i,x)+b^{*}]$

In the formula, sgn () is sign function; The optimum solution of Lagrange coefficient, b ^*It is the threshold value of classification; Find the solution classification function and need to construct optimization problem suc as formula [1] and formula [2]

$\underset{w,\mathrm{\ξ},b}{\min }\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i---\left[1\right]$

$\mathrm{Sub}.\mathrm{to}\left\{\begin{array}{c}{y}_i[(w\·\mathrm{\φ}\left(x_i\right))+b]+{\mathrm{\ξ}}_i\≥1\\ {\mathrm{\ξ}}_i\≥0,i=\mathrm{1,2},...,l\end{array}\right.---\left[2\right]$

In the formula, w is adjustable weight vectors, and C is penalty factor, ξ _iBe slack variable, φ () is the Nonlinear Mapping from the sample space to the higher dimensional space.

Then the Wolfe antithesis of formula [1] and formula [2] is

$\underset{\mathrm{\α}}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{y}_j{\mathrm{\α}}_i{\mathrm{\α}}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i---\left[3\right]$

$\mathrm{Sub}.\mathrm{to}\left\{\begin{array}{c}{\underset{i=1}{\overset{l}{\mathrm{\Σ}}}y}_i{\mathrm{\α}}_i=0\\ 0\≤{\mathrm{\α}}_i\≤C,i=\mathrm{1,2},...,l\end{array}\right.---\left[4\right]$

If the optimization problem of formula [3] and formula [4] statement has optimum solution

${\mathrm{\α}}^{*}=({\mathrm{\α}}_1^{*},{\mathrm{\α}}_2^{*},...,{\mathrm{\α}}_l^{*})$

According to optimality condition---KKT, the solution of this optimization problem must satisfy

α _i{[(x _i·w)+b]-y _i-1}=0 [5]

Finding the solution the optimal classification function that obtains after the problems referred to above is

$f\left(x\right)=\mathrm{sgn}[\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x)+b^{*}]---\left[6\right]$

$b^{*}=y_i-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x_j)---\left[7\right]$

$w^{*}=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}{x}_i---\left[8\right]$

In the formula [7]

K () in formula [3]～formula [7] formula is the kernel function that satisfies the Mercer condition, its effect is exactly the linearly inseparable problem of the former input space is converted into the linear separability problem in higher-dimension even Infinite-dimensional Hibert space, then finds the solution the optimization problem of formula [3] and formula [4] at this higher-dimension or Infinite-Dimensional Space.

At present, the kernel function of most study mainly contains following three kinds of forms

1) polynomial kernel

K(x,x _i)=[(x,x _i)+1] ^q

2) radial basis nuclear

$K(x,x_i)=\exp [-\frac{{\left|\right|x-x_i\left|\right|}^2}{{\mathrm{\σ}}^2}]$

3) Sigmoid nuclear

K(x,x _i)=tanh(v(x·x _i)+C

In the formula, q, σ, v are the parameter of corresponding kernel function, and parameter is selected to obtain by experiment.

(2), flame identification

Utilize the concrete steps of SVM identification terminal point flame as follows:

1) obtains the training sample data;

2) inner product function (being kernel function) and the penalty factor of selection nonlinear transformation;

3) form protruding double optimization problem, find the solution this optimization problem to obtain the support vector machine of classification;

4) the sample to be tested data are sent into the support vector machine of classification, obtained recognition result.

Wherein, the 1st) step realizes that by the image characteristics extraction algorithm of step () four eigenwerts of every two field picture consist of an input vector of support vector machine; The 2nd) step is to choose kernel function, chooses by experiment; The 3rd) step is to set up the svm classifier device, finds the solution optimum solution by algorithm of support vector machine

Try to achieve weight vectors w according to formula [7] and formula [8] again ^*With side-play amount b ^*At last will

And b ^*Substitution formula [6] obtains the optimal classification function; The 4th) step is to utilize SVM identification flame, the sample to be tested data is sent into the SVM that trains, the recognition accuracy of the SVM that checking is trained.

Experimental analysis

In order to verify Algorithm Performance of the present invention, the present invention adopts the test figure that obtains in certain ferroalloy works' argon oxygen refining middle-low-carbon ferrochrome production process, construct 180 groups of 4 characteristics of image to smelting the corresponding relation data set that whether finishes, then trained the svm classifier device to carry out end-point prediction.

1, the training stage

In order to choose kernel function, the present invention adopts respectively the polynomial kernel function, and radial basis kernel function and Sigmoid kernel function training svm classifier device are determined C=200 according to experiment.The recognition result of 3 kinds of kernel functions when table 1 is C=200.

The recognition result of 3 kinds of kernel functions during table 1 C=200

As can be seen from Table 1, utilize the svm classifier device recognition accuracy of support vector machine kernel function training the highest, and preliminary definite when C=200, the σ value is [2.5,3.5] interval, the precision of prediction is higher, and the support vector number stabilizes to 6, so the kernel function type that the present invention determines is the radial basis kernel function.Table 2 is C=200, the support vector of σ=3.0 correspondences and Laplace coefficient thereof.

Table 2 support vector and Laplace coefficient

2, test phase

(1) memory capability test

With whole 180 groups of training datas as input, the substitution sorter, the result is all correct, has proved that SVM has memory function.

(2) generalization ability test

Outside training dataset, extract at random again the generalization ability that 100 stack features data are come test machine out, accuracy rate is more than 93%.

The present invention is directed to the AOD stove and smelt the difficult problem that the fire operation is manually seen in the long-term dependence of endpoint in the middle-low-carbon ferrochrome technological process, a kind of method of utilizing flame characteristic and SVM to carry out smelting endpoint identification has been proposed, provided AOD stove fire door Flame Image Characteristics extracting method, made up the svm classifier model, test result has been verified the validity of the method.The method also can be used for predicting converter, be the production run end-point prediction of main smelting equipment with the electric furnace of comburant oxygen.

Claims (2)

1. argon oxygen refining ferroalloys smelting endpoint Forecasting Methodology, the method is to be identified and extract smelting middle-low-carbon ferrochrome later stage AOD stove fire door flame characteristic by Vision Builder for Automated Inspection, utilize support vector machine to set up the End-point Prediction model, identification is smelted and whether is finished, and finally determines smelting endpoint.

2. a kind of argon oxygen refining ferroalloys smelting endpoint Forecasting Methodology according to claim 1, the method comprises following concrete steps:

(1), the flame characteristic based on machine vision extracts

(1), the formation of image capturing system

Image capturing system is comprised of optical image sensor ccd video camera, image pick-up card and industrial computer, the image information of ccd video camera is delivered to image pick-up card on the industrial computer by concentric cable, image pick-up card converts vision signal to digital signal, sends into industrial computer and carries out corresponding digital processing, image demonstration and forecast analysis; Ccd video camera is converted into electric signal with the light signal of AOD stove fire door flame, image pick-up card is stored in the simulating signal of ccd video camera output in the computing machine through over-sampling, after discrete, and the gradation conversion of each sampled point become 0～255 gray level, deposit in order buffer register in, computing machine conducts interviews to each pixel through buffer register;

(2), feature selecting

The basic demand of terminal point control is when finishing blowing, and the chemical constitution in the molten iron and temperature reach the requirement of coming out of the stove simultaneously, and the AOD stove is smelted the middle-low-carbon ferrochrome later stage, because carbon will exhaust in the stove, reaction between carbon and oxygen tends towards stability, and fire door flame changes slowly, color jaundice deliquescing; By the flame available gray-scale ( ), the high-temperature area gray scale (

), the flame useful area (

), the high-temperature area area (

) can make fast flame characteristic and describing, describe smelting endpoint with [1,1] and whether reach;

(2-1) the flame available gray-scale (

)

Determine by experiment the threshold value of flame available gray-scale, the part that pixel grey scale is higher than this threshold value is considered as effective flame region.

(2-2) the high-temperature area gray scale (

)

Be determined by experiment the high-temperature area gray threshold, the part that pixel grey scale is higher than this threshold value is considered as the flame high-temperature area;

(2-3) the flame useful area (

)

The computing formula of flame area is

$S_i=\underset{j=1}{\overset{G}{\mathrm{\Σ}}}L(g_i^j-g_d)$

In the formula, S _iBe the flame area of the i time sampling, G is the image pixel number,

The gray-scale value of pixel j in the image when representing the i time sampling; g _dBe predefined threshold value, rational threshold value is chosen the calculating of " flame pixels area " most important, can be by the observation definite threshold to on-the-spot flame image and furnace flame video recording, and L (.) is step function, characterizes the fluctuation of fire door flame fringe, Get 1,

Get 0;

(2-4) flame high temperature area (

)

With respect to the flame area of flame high-temperature area gray scale, consistent with flame useful area acquisition methods;

(3), normalized

Because the proper vector implication of choosing is different, dimension is different, and it is larger that numerical value differs; If directly use these variablees as the sample of training network, just then learning process can be handled by the large variable fluctuation of eigenwert, can't embody the situation of change of little eigenwert, therefore, need to do normalized to above-mentioned characteristic variable; Processing Algorithm is:

$x=\frac{\stackrel{\‾}{x}-{\stackrel{\‾}{x}}_{\min }}{{\stackrel{\‾}{x}}_{\max }-{\stackrel{\‾}{x}}_{\min }}$

In the formula, , x is respectively the value before and after the reduction;

,

Be respectively maximal value and the minimum value of sample;

(2), identify based on the smelting endpoint of SVM

(1), algorithm of support vector machine

In two quasi-mode identification problems, given training sample (x _i, y _i), x _iThe input of i sample, y _iThe desired output of i sample, x _i∈ R ⁿ, y _i∈ (1 ,+1), i=1,2 ..., l, the target of SVM is exactly by discriminant function of training sample structure, training sample is separated with largest interval, and sample to be adjudicated is correctly classified as much as possible; Classifying rules is:

$f\left(x\right)=\mathrm{sgn}[\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{a}_i^{*}(x_i,x)+b^{*}]$

In the formula, sgn () is sign function;

The optimum solution of Lagrange coefficient, b ^*It is the threshold value of classification; Find the solution classification function and need to construct optimization problem suc as formula [1] and formula [2]:

$\underset{w,\mathrm{\ξ},b}{\min }\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i---\left[1\right]$

$\mathrm{Sub}.\mathrm{to}\left\{\begin{array}{c}{y}_i[(w\·\mathrm{\φ}\left(x_i\right))+b]+{\mathrm{\ξ}}_i\≥1\\ {\mathrm{\ξ}}_i\≥0,i=\mathrm{1,2},...,l\end{array}\right.---\left[2\right]$

In the formula, w is adjustable weight vectors, and C is penalty factor, ξ _iBe slack variable, φ () is the Nonlinear Mapping from the sample space to the higher dimensional space;

Then the Wolfe antithesis of formula [1] and formula [2] is:

$\underset{\mathrm{\α}}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{y}_j{\mathrm{\α}}_i{\mathrm{\α}}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i---\left[3\right]$

$\mathrm{Sub}.\mathrm{to}\left\{\begin{array}{c}{\underset{i=1}{\overset{l}{\mathrm{\Σ}}}y}_i{\mathrm{\α}}_i=0\\ 0\≤{\mathrm{\α}}_i\≤C,i=\mathrm{1,2},...,l\end{array}\right.---\left[4\right]$

If the optimization problem of formula [3] and formula [4] statement has optimum solution:

${\mathrm{\α}}^{*}=({\mathrm{\α}}_1^{*},{\mathrm{\α}}_2^{*},...,{\mathrm{\α}}_l^{*})$

According to optimality condition---KKT, the solution of this optimization problem must satisfy:

α _i{[(x _i·w)+b]-y _i-1}=0 [5]

Finding the solution the optimal classification function that obtains after the problems referred to above is:

$f\left(x\right)=\mathrm{sgn}[\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x)+b^{*}]---\left[6\right]$

$b^{*}=y_i-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x_j)---\left[7\right]$

$w^{*}=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}{x}_i---\left[8\right]$

In the formula [7]

K () in formula [3]～formula [7] formula is the kernel function that satisfies the Mercer condition, its effect is exactly the linearly inseparable problem of the former input space is converted into the linear separability problem in higher-dimension even Infinite-dimensional Hibert space, then finds the solution the optimization problem of formula [3] and formula [4] at this higher-dimension or Infinite-Dimensional Space;

(2), flame identification

Utilize the concrete steps of SVM identification terminal point flame as follows:

1) obtains the training sample data;

2) inner product function (being kernel function) and the penalty factor of selection nonlinear transformation;

3) form protruding double optimization problem, find the solution this optimization problem to obtain the support vector machine of classification;

4) the sample to be tested data are sent into the support vector machine of classification, obtained recognition result;

Wherein, the 1st) step realizes that by the image characteristics extraction algorithm of step () four eigenwerts of every two field picture consist of an input vector of support vector machine; The 2nd) step is to choose kernel function, chooses by experiment; The 3rd) step is to set up the svm classifier device, finds the solution optimum solution by algorithm of support vector machine

Try to achieve weight vectors w according to formula [7] and formula [8] again ^*With side-play amount b ^*At last will

And b ^*Substitution formula [6] obtains the optimal classification function; The 4th) step is to utilize SVM identification flame, the sample to be tested data is sent into the SVM that trains, the recognition accuracy of the SVM that checking is trained.

Images