CN102392095A - Termination point prediction method and system for converter steelmaking - Google Patents

Abstract

The embodiment of the present invention discloses a termination point prediction method and a system for converter steelmaking. The method includes the following steps: collecting current production parameter information in converter steelmaking and current flame information at the mouth of a converter as independent variables; creating a neural network, and training the created neural network with a training sample made up of the independent variables; and predicting the termination point of converter steelmaking by the neural network obtained after the training according to the independent variables, so as to obtain a termination point time prediction result, a carbon content prediction result and a molten steel temperature prediction result. Therefore, with the technical scheme provided by the embodiment of the present invention, by obtaining the production information of the converter and the flame information at the mouth of the converter in real time, the steelmaking termination point time, the molten steel temperature and the carbon content can be analyzed precisely online in real time, so as to precisely control the process to improve the automation and the production efficiency of converter steelmaking and lower the cost.

Description

BOF Steelmaking Endpoint Forecasting Methodology and system

Technical field

The present invention relates to make steel automation control area, more particularly, relate to a kind of BOF Steelmaking Endpoint Forecasting Methodology and system.

Background technology

The control of steel-making terminal point is one of gordian technique of converter steelmaking in the converter steelmaking, and the converter steelmaking steel output accounts for more than 80% of total steel output.Converter steel output is occupied an leading position in big-and-middle-sized emphasis iron and steel enterprise, therefore improves throughput and the controlled levels of improving converter steelmaking and receives people's attention always.Converter steelmaking is very complicated metallurgical reaction process, and influence factor is a lot.In order to realize the automatic control of converter steelmaking process, developed many detection techniques both at home and abroad, method commonly used mainly contains artificial experience method, chemical analysis, the control of static terminal point, sublance method, analysis of fumes method etc.

In actual production and theoretical investigation, find: artificial experience control steel-making terminal point, the experience and the working order of this and field worker have confidential relation, and it is low to have the terminal point hit rate, and labor strength is big, and the splash rate is high, problems such as production stability difference; Adopt static terminal point system, its hit rate does not still reach robotization steel-making demand, and control process can not be carried out on-line tracing and revised in real time; Though device such as combustion gas analyzer and sublance has better terminal point precision of prediction than artificial experience control and static control method; But because these equipment exist and involve great expense; Difficult in maintenance, big some problems that wait of size are only used in some big section steel works a few days ago; Chemical analysis, its Measuring Time can not satisfy the real-time requirement of smelting process control far away, and when sampling, have the accident of splash;

Therefore; Once to fall the qualification rate of stove be the converter terminal control model of target to improve converter in foundation; Study a kind of low cost, simple and reliable, can adapt to and make steel the BOF Steelmaking Endpoint Forecasting Methodology of on-the-spot severe environment, be the difficult problem that domestic and international field of steel-making technician needs to be resolved hurrily.

Summary of the invention

In view of this, the present invention provides a kind of BOF Steelmaking Endpoint Forecasting Methodology and system, to realize improving control accuracy and hit rate, improves the purpose of converter production efficiency, quality product.

For realizing above-mentioned purpose, the present invention provides following technical scheme:

A kind of BOF Steelmaking Endpoint Forecasting Methodology comprises:

Gather current manufacturing parameter information and the current flame information of converter mouth in the converter steelmaking; Said manufacturing parameter information comprises the weight of molten iron, temperature, carbon content, oxygen gun blowing time and the blowing oxygen quantity of molten iron, and said fire door flame information comprises light intensity characteristic value, flame temperature and flame image eigenwert;

With temperature, carbon content, oxygen gun blowing time, blowing oxygen quantity, light intensity characteristic value and the flame image eigenwert of the weight of said molten iron, molten iron as independent variable(s);

Utilize said independent variable(s) composing training sample;

Create neural network, utilize said learning sample that the neural network of creating is trained;

The neural network that obtains after utilizing training to stop; According to said independent variable(s) BOF Steelmaking Endpoint is predicted and to be predicted the outcome; Said BOF Steelmaking Endpoint comprises terminal time, carbon content and liquid steel temperature, and said predicting the outcome comprises that terminal time predicts the outcome, carbon content predicts the outcome and liquid steel temperature predicts the outcome.

A kind of BOF Steelmaking Endpoint prognoses system comprises establishment of collecting unit and neural network and training unit, wherein:

Said collecting unit; Be used for gathering the current manufacturing parameter information and the current fire door flame information of converter steelmaking; Said manufacturing parameter information comprises the weight of molten iron, temperature, carbon content, oxygen gun blowing time and the blowing oxygen quantity of molten iron; Said fire door flame information comprises light intensity characteristic value, flame temperature and flame image eigenwert, with temperature, carbon content, oxygen gun blowing time, blowing oxygen quantity, light intensity characteristic value and the flame image eigenwert of the weight of said molten iron, molten iron as independent variable(s);

Said neural network is created and training unit, creates neural network, and the learning sample that utilizes said independent variable(s) to constitute is trained the neural network of creating, and obtains training the neural network that obtains after the termination;

The neural network that said training obtains after stopping; Be used for BOF Steelmaking Endpoint being predicted and predicted the outcome according to said independent variable(s); Said BOF Steelmaking Endpoint comprises terminal time, carbon content and liquid steel temperature, and said predicting the outcome comprises that terminal time predicts the outcome, carbon content predicts the outcome and liquid steel temperature predicts the outcome.

It is thus clear that the technical scheme that adopts the embodiment of the invention to provide through obtaining the flame information of converter production information and converter mouth in real time, is come the terminal point of real-time estimate converter steelmaking.The technical scheme that the embodiment of the invention provides is to be based upon on human brain neural network's the basic understanding basis; Can online in real time analyze the time of steel-making terminal point, the temperature and the carbon content of molten steel accurately; Therefore can avoid over that is sought the steelmaking process deep layer law is endless; Then factors such as simulation human brain are handled actual converting process, make basis with true and data, realize the accurate control to process; Thereby improve automatization level, the production efficiency of converter steelmaking, reduce cost.

Description of drawings

In order to be illustrated more clearly in the embodiment of the invention or technical scheme of the prior art; To do to introduce simply to the accompanying drawing of required use in embodiment or the description of the Prior Art below; Obviously, the accompanying drawing in describing below only is some embodiments of the present invention, for those of ordinary skills; Under the prerequisite of not paying creative work, can also obtain other accompanying drawing according to these accompanying drawings.

The BOF Steelmaking Endpoint Forecasting Methodology schema that Fig. 1 provides for the embodiment of the invention;

Another schema of BOF Steelmaking Endpoint Forecasting Methodology that Fig. 2 provides for the embodiment of the invention;

Fig. 3 provides the optimization schema for the embodiment of the invention;

The another schema of BOF Steelmaking Endpoint Forecasting Methodology that Fig. 4 provides for the embodiment of the invention;

The another schema of BOF Steelmaking Endpoint Forecasting Methodology that Fig. 5 provides for the embodiment of the invention;

The BOF Steelmaking Endpoint prognoses system structural representation that Fig. 6 provides for the embodiment of the invention.

Embodiment

To combine the accompanying drawing in the embodiment of the invention below, the technical scheme in the embodiment of the invention is carried out clear, intactly description, obviously, described embodiment only is the present invention's part embodiment, rather than whole embodiment.Based on the embodiment among the present invention, those of ordinary skills are not making the every other embodiment that is obtained under the creative work prerequisite, all belong to the scope of the present invention's protection.

The embodiment of the invention provides a kind of BOF Steelmaking Endpoint Forecasting Methodology, and referring to Fig. 1, this method comprises at least:

Current manufacturing parameter information and the current flame information of converter mouth in S1, the collection converter steelmaking; Above-mentioned manufacturing parameter information comprises the weight of molten iron, temperature, carbon content, oxygen gun blowing time and the blowing oxygen quantity of molten iron, and above-mentioned fire door flame information comprises light intensity characteristic value, flame temperature and flame image eigenwert;

It should be noted that:

The temperature of molten iron is that manufacturer provides;

Light intensity characteristic value: in convertor steelmaking process; Mainly carry out decarburizing reaction in the stove, the oxygen reaction through carbon and oxygen rifle provide generates CO and CO2; Carbon content reduces through reaction gradually; And the CO and the oxygen that generate burn at converter mouth, and the flame light intensity value is different along with the variation of carbon content, so the light intensity characteristic value has characterized the flame intensity variations;

The unit of flame temperature is degree centigrade;

The flame image eigenwert is meant; Gather the flame of converter mouth through the flame image collecting device; Be converted into numerary signal; To find in the fire door flame that through experiment yellow pixel, red pixel etc. and converter terminal exist related closely, so image feature value has characterized the variation of yellow pixel and red pixel in the flame at least.

S2, with temperature, carbon content, oxygen gun blowing time, blowing oxygen quantity, light intensity characteristic value and the flame image eigenwert of the weight of above-mentioned molten iron, molten iron as independent variable(s);

S3, utilize above-mentioned independent variable(s) composing training sample;

S4, establishment neural network utilize above-mentioned learning sample that the neural network of creating is trained;

The neural network that S5, utilization training obtain after stopping; According to above-mentioned independent variable(s) BOF Steelmaking Endpoint is predicted and to be predicted the outcome; Above-mentioned BOF Steelmaking Endpoint comprises terminal time, carbon content and liquid steel temperature, and above-mentioned predicting the outcome comprises that terminal time predicts the outcome, carbon content predicts the outcome and liquid steel temperature predicts the outcome.

In the specific implementation, can be through following mode among the above-mentioned steps S1 to the collection of flame temperature:

The argument of conducting heat according to planck radiation: object all has the wavelength of a maximum radiant intensity under each temperature, and along with the rising of temperature, the wavelength of maximum radiant intensity shortens, and the color of object is white by red stain.So the color of flame can reflect the temperature height of flame.

The flame temperature that converter mouth ejects is mixed by two portions and is formed.A part is the temperature that the CO gas of from molten steel, overflowing is had, the actual liquid steel temperature that reflected of this temperature.To be CO gas carry out the chemical heat of emitting after the perfect combustion at fire door and oxygen to another part, and flame temperature is raise.

Under the cryogenic situation of high frequency, Planck blackbody radiation theorem can be approximate by the Wien theorem, and the radiation power of black matrix is:

$M_{\mathrm{B\&lambda;}}={\mathrm{<figure-callout\; id="1"\; label="C"\; filenames="HDA0000101005870000061.png"\; state="\{\{state\}\}">C</figure-callout>}}_1{\mathrm{\&lambda;}}^{-5}{e}^{-\frac{{\mathrm{<figure-callout\; id="2"\; label="C"\; filenames="HDA0000101005870000061.png"\; state="\{\{state\}\}">C</figure-callout>}}_2}{\mathrm{\&lambda;T}}}---\left(1\right)$

C wherein ₁=3.7418 * 10-16Wm2 is first radiation constant, C ₂=1.4388 * 10-2m is a second radiation constant.Under certain specific wavelength, the monochromatic irradiation degree is:

$E_{\mathrm{\&lambda;}}\left(T\right)={\mathrm{<figure-callout\; id="1"\; label="C"\; filenames="HDA0000101005870000061.png"\; state="\{\{state\}\}">C</figure-callout>}}_1{\mathrm{\&lambda;}}^{-5}{e}^{-\frac{{\mathrm{<figure-callout\; id="2"\; label="C"\; filenames="HDA0000101005870000061.png"\; state="\{\{state\}\}">C</figure-callout>}}_2}{\mathrm{\&lambda;T}}}{\mathrm{\&epsiv;}}_{\mathrm{\&lambda;}}\left(T\right)---\left(2\right)$

ε wherein _λBe monochromatic emissivity.Black matrix monochromatic emissivity under different wave length is all identical, if in wavelength X ₁And λ ₂Record the yield of radiation of same point down, then can obtain object temperature according to the ratio of yield of radiation:

$T=\frac{{\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="HDA0000101005870000061.png"\; state="\{\{state\}\}">c</figure-callout>}}_2}{\ln (E_{{\mathrm{\&lambda;}}_1}/E_{{\mathrm{\&lambda;}}_2})+5\ln ({\mathrm{\&lambda;}}_1/{\mathrm{\&lambda;}}_2)}(\frac{1}{{\mathrm{\&lambda;}}_2}-\frac{1}{{\mathrm{\&lambda;}}_1})---\left(3\right)$

Can collect the colour picture of flame image through CCD, store according to three Color Channels of red (R) green (G) and blue (B).Single Color Channel collection value direct ratio has in the secondary illumination of monochrome:

$R=K_R{E}_{{\mathrm{\&lambda;}}_R}\left(T\right)$

$G=K_G{E}_{{\mathrm{\&lambda;}}_G}\left(T\right)$

$B=K_B{E}_{{\mathrm{\&lambda;}}_B}\left(T\right)$

K wherein _R, K _G, K _BBe respectively the response coefficient of three Color Channels.(4) formula substitution (3) formula, can get like this:

$T=\frac{{\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="HDA0000101005870000061.png"\; state="\{\{state\}\}">c</figure-callout>}}_2}{\ln (G/R)+5\ln ({\mathrm{\&lambda;}}_G/{\mathrm{\&lambda;}}_R)+\mathrm{\&delta;}1}(\frac{1}{{\mathrm{\&lambda;}}_R}-\frac{1}{{\mathrm{\&lambda;}}_G})---\left(5\right)$

In the formula 5: the temperature value of T for calculating through R, two Color Channels of G, δ 1 is monochromatic emissivity ε _R, ε _GResponse coefficient K _R, K _GEtc. coefficient constant, wavelength adopts the CIE-RGB system specification, and R is that wavelength is 700.0nm, and the G wavelength is 546.1nm, and the B wavelength is 435.8nm.

It is thus clear that the method that the embodiment of the invention provides through obtaining the flame information of converter production information and converter mouth in real time, is come the terminal point of real-time estimate converter steelmaking.The method that the embodiment of the invention provides is to be based upon on human brain neural network's the basic understanding basis; The time of can online in real time accurate analyses and prediction steel-making terminal point, the temperature and the carbon content of molten steel; Therefore can avoid over that is sought the steelmaking process deep layer law is endless; Then factors such as simulation human brain are handled actual converting process, make basis with true and data, realize the accurate control to process; Thereby improve automatization level, the production efficiency of converter steelmaking, reduce cost.

Owing to need terminal time, carbon content and the liquid steel temperature of prediction BOF Steelmaking Endpoint, therefore, can adopt three neural sub-networks respectively above-mentioned terminal time, carbon content and liquid steel temperature to be predicted.Because the BP neural network is extensively adopted; Therefore; In other embodiment of the present invention, above-mentioned neural network specifically can comprise the neural sub-network of the BP that is used for the terminal time prediction, be used for the neural sub-network of BP of carbon content prediction, and the neural sub-network of BP that is used for the liquid steel temperature prediction.

Above-mentioned three child network can adopt similar mode to create and train, and therefore, the present invention will not distinguish it later, directly illustrate and create and the concrete mode of training.

In other embodiment of the present invention, referring to Fig. 2, above-mentioned steps S4 can specifically comprise the steps:

S41, confirm neural (son) topology of networks of BP;

Its topological framework generally comprises an input layer, an output layer, and at least one hidden layer;

The weights of S42, neural (son) network of initial above-mentioned BP and the length of threshold value;

Wherein, above-mentioned weights comprise the weights that are connected that are connected weights and hidden layer and output layer of input layer and hidden layer, and above-mentioned threshold value comprises hidden layer threshold value and output layer threshold value;

S43, utilize above-mentioned learning sample that the initial weight and the initial threshold of neural (son) network of BP are optimized;

In other embodiment of the present invention, can be optimized through genetic algorithm.

S44, neural (son) network of BP that utilizes above-mentioned learning sample that warp is optimized are trained; When predicting the outcome of neural (son) network output of housebroken BP reached preset requirement (such as the difference with actual result of predicting the outcome less than a certain threshold value) or reach predefined frequency of training, training stopped.

Based on genetic algorithm, referring to Fig. 3, the concrete implementation of above-mentioned steps S43 can comprise:

Population scale, iterations, hereditary strategy, the maximum evolution number of times of A, setting population;

Above-mentioned hereditary strategy comprises selection operation, interlace operation, mutation operation, crossover probability and variation probability, and each individuality in the population has all comprised all weights and the threshold value of neural (son) network of above-mentioned BP;

B, above-mentioned population is carried out initialize, make each individuality in the population all have corresponding assignment, the individuality in the population is carried out individuality coding (real coding);

Also promptly, when comprising 10 individuals in the population, respectively the initial weight and the initial threshold of neural (son) network of above-mentioned BP are carried out assignment, can obtain 10 structures, weights, definite neural (son) network of BP of threshold value with this 10 individuals;

C, population is carried out fitness calculate, obtain each individual ideal adaptation degree value in the population, above-mentioned ideal adaptation degree value is used to characterize individual fitness;

In other embodiment of the present invention; Obtain prediction output behind neural (son) network of available above-mentioned learning sample input BP; As ideal adaptation degree value F, its calculation formula is following the Error Absolute Value between prediction output and the desired output (desired output obtains through actual measurement) and E:

$F=\mathrm{\&xi;}\left(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{abs}(y_i-o_i)\right)$

Wherein, n is the network output node; y _iDesired output for i node of neural (son) network of BP; o _iBe the prediction output of i node; ξ is a coefficient, and 0＜ξ＜1.

D, utilize above-mentioned heredity strategy that the population among the step C is carried out heredity, form population of future generation;

Step D can specifically comprise again:

Select the good individuality of fitness to form new population in D1, the population from step C;

In other embodiment of the present invention, the selection Probability p of each individual m _mFor:

f _m＝ξ/F _m

$p_m=\frac{f_m}{\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_m}$

In the following formula, F _mBe the fitness value of individual m, because fitness value is more little good more, so before individual selection, ask inverse to obtain f to fitness value _m, ξ is a coefficient, N is population scale (also being the individual number of population).

D2, above-mentioned new population is carried out interlace operation and mutation operation, form population of future generation;

The concrete implementation of above-mentioned interlace operation comprises: select two individuals as two parent individualities, to above-mentioned two parent individualities by above-mentioned crossover probability intersect obtain two new individual, with above-mentioned two above-mentioned two parent individualities of new individual replacement;

When individuality adopted real coding, interlace operation also need be adopted the real number interior extrapolation, and its method is following:

Suppose k karyomit(e) a _KWith Z karyomit(e) a _ZInterlace operation method in the j position is following:

$\left.\begin{array}{c}{a}_{\mathrm{Kj}}=a_{\mathrm{Kj}}(1-b)+a_{\mathrm{Zj}}b\\ a_{\mathrm{Zj}}=a_{\mathrm{Zj}}(1-b)+a_{\mathrm{Kj}}b\end{array}\right\}$

In the formula, b is the randomized number between [0,1].

The concrete implementation of above-mentioned mutation operation comprises: select body one by one at random, obtain new individuality by above-mentioned variation probability variation, with the above-mentioned new individual above-mentioned individuality of selecting at random that replaces.

When individuality adopts real coding, j gene a of picked at random m individuals _MjMake a variation, its formula is following:

$a_{\mathrm{mj}}=\left\{\begin{array}{cc}{a}_{\mathrm{mj}}+(a_{\min }-a_{\mathrm{mj}})*f\left(g\right)& r\&le;0.5\\ a_{\mathrm{mj}}+(a_{\mathrm{mj}}-a_{\max })*f\left(g\right)& r>0.5\end{array}\right.$

In the formula, a _MjBe the new individuality that the back that makes a variation produces, a _MaxBe gene a _MjThe upper bound, a _MinLower bound for gene; F (g)=r ₂(1-g/G _Max) ², r ₂Be a randomized number, g is the current iteration number of times; G _MaxBe maximum evolution number of times, r is the randomized number between [0,1].

Whether E, the performance of judging above-mentioned colony of future generation satisfy preset index and/or have accomplished the iterations of above-mentioned setting, if not, then return step C, otherwise, get into step F;

F, the optimum individual in the population of future generation that forms among the step D is carried out assignment to the initial weight and the initial threshold of neural (son) network of above-mentioned BP, accomplish the initial weight of neural (son) network of above-mentioned BP and the optimization of initial threshold.

Accordingly, the concrete implementation of above-mentioned steps S44 can comprise:

One, hidden layer output is calculated: according to input parameter X, be connected weights ω between input layer and hidden layer _XyAnd hidden layer threshold value a, calculate hidden layer output H.

$H_y=f(\underset{x=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{xy}}{x}_x-a_y)$ y＝1，2，…，l

In the formula, n representes the input layer number, and l is the hidden layer node number, and y representes y hidden layer node, and f is the hidden layer excitation function.

Two, output layer output is calculated: according to hidden layer output H, the connection weights ω of output layer _YzWith threshold value b, calculate neural (son) network prediction of BP output O.

$O_k=\underset{y=1}{\overset{1}{\mathrm{\&Sigma;}}}{H}_y{\mathrm{\&omega;}}_{\mathrm{yz}}-b_z$ z＝1，2，…，m

In the following formula, k=z=1,2...m are output layer node number.

Three, Error Calculation: according to prediction output O and desired output Y, computational grid predicated error e.

e _k＝Y _k-O _k?k＝1，2，…，m

Four, the weights threshold value is upgraded: upgrade network according to network predicated error e and connect weights and threshold value.

With neural (son) network of three layers of (also promptly only having one deck hidden layer) BP is example, weights ω _Xy, ω _YzExpression formula following:

${\mathrm{\&omega;}}_{\mathrm{xy}}={\mathrm{\&omega;}}_{{\mathrm{xy}}^{\&prime;}}+\mathrm{\&eta;}{H}_y(1-H_y)X\left(x\right)\underset{z=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{yz}}{e}_z$ x＝1，2，…，n；y＝1，2，…，l

ω _yz＝ω _yz+ηH _ye _z?y＝1，2，…，l；z＝1，2，…，m

Threshold value a, the expression formula of b is following:

$a_y=a_y+\mathrm{\&eta;}{H}_y(1-H_y)\underset{z=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{yz}}{e}_z$ y＝1，2，…，l

b _z＝b _z+e _z?z＝1，2，…，m

In the formula, η is a learning rate, the input value of X () expression BP neural network, and n representes the input layer number.

In addition, the Ek in the Ez in " renewal of weights threshold value " and " Error Calculation " represents same error, k=z=1,2...m.

Five, whether the training of judgement number of times finishes, if do not finish, returns hidden layer output and calculates link.

The network input data that comprised in the neural network algorithm of standard are that the investigator chooses according to professional standing and experience in advance; Yet in convertor steelmaking process; Owing to there is not theoretical foundation clearly; The independent variable(s) that neural network comprised is that the network input feature vector is difficult to confirm in advance, if some unessential independent variable(s) are also introduced the precision that neural network can reduce model, therefore selecting significant independent variable(s) is a very crucial step as input parameter.

Therefore, in other embodiment of the present invention, referring to Fig. 4, above-mentioned all embodiment can also screen independent variable(s).

Above-mentioned steps S5 specifically can comprise:

S51, the neural network (can be the neural sub-network of BP neural network or BP in certain embodiments) that obtains after utilizing above-mentioned training to stop are screened all independent variable(s) that collect among the step S1, filter out the independent variable(s) that the influence degree that predicts the outcome is reached preset standard;

That the neural network that S52, above-mentioned training obtain after stopping will filter out, as will to reach preset standard to the above-mentioned influence degree that predicts the outcome independent variable(s) is as input parameter, prediction of output result.

Comprise that terminal time predicts the outcome, carbon content predicts the outcome and liquid steel temperature predicts the outcome owing to predict the outcome; Therefore the above-mentioned independent variable(s) that the above-mentioned influence degree that predicts the outcome is reached preset standard can comprise that also the influence degree that above-mentioned terminal time is predicted the outcome reaches the independent variable(s) of preset standard; The influence degree that above-mentioned carbon content is predicted the outcome reaches the independent variable(s) of preset standard, and the influence degree that above-mentioned liquid steel temperature predicts the outcome is reached the independent variable(s) of preset standard.

Concrete, can carry out the independent variable(s) screening through the MIV algorithm.Also promptly, referring to Fig. 5, the concrete realization of above-mentioned steps S51 comprises:

S511, each independent variable(s) in the above-mentioned learning sample is carried out the MIV value calculate;

Concrete, the MIV value is calculated and is comprised again:

1, the initial value S that the independent variable(s) of MIV value to be calculated is corresponding adds/subtracts a%S respectively, constitutes new learning sample P1 and P2;

Also promptly, comprise the initial value of other independent variable(s) among the learning sample P1, and the corresponding value S-a%S of the independent variable(s) of MIV value to be calculated, and comprise the initial value that other independent variable(s) are corresponding among the learning sample P2, and the corresponding value S+a%S of the independent variable(s) of MIV value to be calculated.

2, utilize the neural network that obtains after the above-mentioned training termination to carry out simulation and prediction as simulation sample respectively P1 and P2, obtain two simulation and prediction A1 and A2 as a result;

3, obtain the difference of A1 and A2, as influencing changing value IV to what output produced behind the change independent variable(s);

4, IV is on average drawn the MIV value of the independent variable(s) correspondence of MIV value to be calculated by the routine number of observation.

S512, independent variable(s) is sorted (absolute value of MIV value is big more, representes that its influence degree is big more), choose the independent variable(s) that absolute value reaches preset value according to the size of the absolute value of MIV value.Above-mentioned preset value is corresponding with preset standard.

Through experiment, utilize the independent variable(s) of the above-mentioned terminal time of prediction that the MIV algorithm filters out to comprise: blowing oxygen quantity, oxygen blow duration, light intensity characteristic value, image feature value; The independent variable(s) of the prediction liquid steel temperature that filters out comprises: flame temperature, blowing oxygen quantity, oxygen blow duration and light intensity characteristic value; The independent variable(s) of the predict carbon content that filters out comprises: blowing oxygen quantity, oxygen blow duration and light intensity characteristic value.

Conversely, the independent variable(s) that filters out can be used for instructing acquisition step, simplifies the information category of being gathered.

In other embodiment of the present invention, aforesaid method also can comprise:

According to predicting the outcome the oxygen rifle is controlled.

For example: carbon content is by the decision of institute of steel mill steelmaking kind, when the predictor of carbon content is consistent with set(ting)value, stops oxygen gun blowing.

Adopt above-mentioned control method can online in real time to analyze whole change procedures of molten steel carbon, temperature, accurately control steel-making terminal point.Therefore alleviate steelmaker's labour intensity, guarantee the quality of steel; Can realize that blow end point controls automatically, improve the automatization level of converter steelmaking; Because the terminal point through flame image and light intensity judgement converter steelmaking improves the converter smelting endpoint hit rate, shortened the heat 3-4 minute, thereby enhances productivity, and reduces cost.

Through practice, terminal point forecast accuracy>=90% of the embodiment of the invention, forecast time error≤10s, forecast precision [C] ± 0.02%, T ± 12 ℃.

Corresponding with it, the present invention also provides a kind of BOF Steelmaking Endpoint prognoses system, and it comprises collecting unit 1 and neural network establishment and training unit 2, wherein:

Collecting unit 1 is used for gathering the current manufacturing parameter information and the current fire door flame information of converter steelmaking, obtains independent variable(s); Manufacturing parameter information and current fire door flame information can not given unnecessary details at this referring to the aforementioned record of this paper;

Neural network is created and training unit 2, is used to create neural network, utilizes learning sample that the neural network of creating is trained, and obtains training the neural network 3 that obtains after the termination;

The neural network 3 that above-mentioned training obtains after stopping then can be predicted BOF Steelmaking Endpoint according to above-mentioned independent variable(s) and predicted the outcome.BOF Steelmaking Endpoint with predict the outcome, comprise terminal time, carbon content and liquid steel temperature, said predicting the outcome comprises that terminal time predicts the outcome and can not give unnecessary details at this referring to the aforementioned record of this paper.

In addition, native system also leaves abundant software interface, can connect with the converter PLC of steel mill, to realize that utilizing predicts the outcome the oxygen rifle is controlled automatically.

To the above-mentioned explanation of the disclosed embodiments, make this area professional and technical personnel can realize or use the present invention.Multiple modification to these embodiment will be conspicuous concerning those skilled in the art, and defined General Principle can realize under the situation that does not break away from the spirit or scope of the present invention in other embodiments among this paper.Therefore, the present invention will can not be restricted to these embodiment shown in this paper, but will meet and principle disclosed herein and features of novelty the wideest corresponding to scope.

Claims (10)

1. a BOF Steelmaking Endpoint Forecasting Methodology is characterized in that, comprising:

Gather current manufacturing parameter information and the current flame information of converter mouth in the converter steelmaking; Said manufacturing parameter information comprises the weight of molten iron, temperature, carbon content, oxygen gun blowing time and the blowing oxygen quantity of molten iron, and said fire door flame information comprises light intensity characteristic value, flame temperature and flame image eigenwert;

With temperature, carbon content, oxygen gun blowing time, blowing oxygen quantity, light intensity characteristic value and the flame image eigenwert of the weight of said molten iron, molten iron as independent variable(s);

Utilize said independent variable(s) composing training sample;

Create neural network, utilize said learning sample that the neural network of creating is trained;

The neural network that obtains after utilizing training to stop; According to said independent variable(s) BOF Steelmaking Endpoint is predicted and to be predicted the outcome; Said BOF Steelmaking Endpoint comprises terminal time, carbon content and liquid steel temperature, and said predicting the outcome comprises that terminal time predicts the outcome, carbon content predicts the outcome and liquid steel temperature predicts the outcome.

2. method according to claim 1 is characterized in that, the neural network that obtains after said utilization training stops predicts to BOF Steelmaking Endpoint that according to said independent variable(s) the embodiment that obtains predicting the outcome comprises:

The neural network that obtains after utilizing said training to stop is screened all said independent variable(s), filters out the independent variable(s) that the said influence degree that predicts the outcome is reached preset standard;

That the neural network that said training obtains after stopping will filter out, as will to reach preset standard to the said influence degree that predicts the outcome independent variable(s) is as input parameter, dopes to predict the outcome;

The said independent variable(s) that the said influence degree that predicts the outcome is reached preset standard comprises that the influence degree that said terminal time is predicted the outcome reaches the independent variable(s) of preset standard; The influence degree that said carbon content is predicted the outcome reaches the independent variable(s) of preset standard, and the influence degree that said liquid steel temperature predicts the outcome is reached the independent variable(s) of preset standard.

3. method according to claim 2 is characterized in that, said all said independent variable(s) is screened, and filters out the embodiment that the said influence degree that predicts the outcome is reached the independent variable(s) of preset standard and comprises:

Each independent variable(s) in the said learning sample is carried out the MIV value to be calculated;

Size according to the absolute value of MIV value sorts to independent variable(s), chooses independent variable(s) that absolute value reaches preset standard as the independent variable(s) that the said influence degree that predicts the outcome is reached preset standard;

Said MIV value is calculated and is comprised:

The initial value S that the independent variable(s) of MIV value to be calculated is corresponding adds/subtracts a%S respectively, constitutes new learning sample P1 and P2;

Utilize the neural network that obtains after the said training termination to carry out simulation and prediction as simulation sample respectively P1 and P2, obtain two simulation and prediction A1 and A2 as a result;

Obtain the difference of A1 and A2, as influencing changing value IV to what output produced behind the change independent variable(s);

IV is on average drawn the corresponding MIV value of independent variable(s) of MIV value to be calculated by the routine number of observation.

4. method according to claim 3 is characterized in that:

Said neural network is the BP neural network;

Said establishment neural network utilizes said learning sample that the neural network of creating is trained, and obtains training the embodiment of the neural network that obtains after the termination to comprise:

Confirm the topological framework of BP neural network, said topological framework comprises an input layer, an output layer, and at least one hidden layer;

The weights of initial said BP neural network and the length of threshold value;

Utilize said learning sample that the initial weight and the initial threshold of said BP neural network are optimized;

Utilize said learning sample to training through the BP neural network of optimizing, when predicting the outcome of housebroken BP neural network output reached preset requirement or reach predefined frequency of training, training stopped.

5. method according to claim 4; It is characterized in that, saidly utilize said learning sample that the embodiment that the initial weight and the initial threshold of said BP neural network is optimized is optimized the initial weight and the initial threshold of said BP neural network for utilizing said learning sample and genetic algorithm.

6. method according to claim 5 is characterized in that, saidly utilizes said learning sample and genetic algorithm that the initial weight of said BP neural network and initial threshold are optimized specifically to comprise:

A, the population scale of setting population, iterations, heredity strategy; Said hereditary strategy comprises selection operation, interlace operation, mutation operation, crossover probability and variation probability, and each individuality in the population has all comprised all weights and the threshold value of said BP neural network;

B, said population is carried out initialize, make each individuality in the population all have corresponding assignment;

C, population is carried out fitness calculate, obtain each individual ideal adaptation degree value in the population, said ideal adaptation degree value is used to characterize individual fitness;

D, utilize said heredity strategy that the population among the step C is carried out heredity, form population of future generation;

Whether E, the performance of judging said colony of future generation satisfy preset index and/or have accomplished the iterations of said setting, if not, then return step C, otherwise, get into step F;

F: the optimum individual in the population said of future generation that forms among the step D is carried out assignment to the initial weight and the initial threshold of said BP neural network, accomplish to the initial weight of said BP neural network and the optimization of initial threshold.

7. method according to claim 6 is characterized in that, the embodiment of said step D comprises:

Select the good individuality of fitness to form new population in the population from step C;

Said new population is carried out interlace operation and mutation operation, form population of future generation;

The concrete implementation of said interlace operation comprises: select two individuals as two parent individualities, to said two parent individualities by said crossover probability intersect obtain two new individual, with said two said two parent individualities of new individual replacement;

The concrete implementation of said mutation operation comprises: select body one by one at random, obtain new individuality by said variation probability variation, with the said new individual said individuality of selecting at random that replaces.

8. according to the said method of claim 7, it is characterized in that, also comprise:

According to said predicting the outcome said oxygen rifle is controlled.

9. method according to claim 8 is characterized in that, said BP neural network comprises the neural sub-network of the BP that is used for the terminal time prediction, is used for the neural sub-network of BP of carbon content prediction, and the neural sub-network of BP that is used for the liquid steel temperature prediction.

10. a BOF Steelmaking Endpoint prognoses system is characterized in that, comprises establishment of collecting unit and neural network and training unit, wherein:

Said collecting unit; Be used for gathering the current manufacturing parameter information and the current fire door flame information of converter steelmaking; Said manufacturing parameter information comprises the weight of molten iron, temperature, carbon content, oxygen gun blowing time and the blowing oxygen quantity of molten iron; Said fire door flame information comprises light intensity characteristic value, flame temperature and flame image eigenwert, with temperature, carbon content, oxygen gun blowing time, blowing oxygen quantity, light intensity characteristic value and the flame image eigenwert of the weight of said molten iron, molten iron as independent variable(s);

Said neural network is created and training unit, creates neural network, and the learning sample that utilizes said independent variable(s) to constitute is trained the neural network of creating, and obtains training the neural network that obtains after the termination;

The neural network that said training obtains after stopping; Be used for BOF Steelmaking Endpoint being predicted and predicted the outcome according to said independent variable(s); Said BOF Steelmaking Endpoint comprises terminal time, carbon content and liquid steel temperature, and said predicting the outcome comprises that terminal time predicts the outcome, carbon content predicts the outcome and liquid steel temperature predicts the outcome.

Images