CN105608492A - Robust random-weight neural network-based molten-iron quality multi-dimensional soft measurement method - Google Patents

Robust random-weight neural network-based molten-iron quality multi-dimensional soft measurement method Download PDF

Info

Publication number
CN105608492A
CN105608492A CN201610118914.7A CN201610118914A CN105608492A CN 105608492 A CN105608492 A CN 105608492A CN 201610118914 A CN201610118914 A CN 201610118914A CN 105608492 A CN105608492 A CN 105608492A
Authority
CN
China
Prior art keywords
model
data
blast
molten steel
rvflns
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201610118914.7A
Other languages
Chinese (zh)
Other versions
CN105608492B (en
Inventor
周平
吕友彬
王宏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northeastern University China
Original Assignee
Northeastern University China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Northeastern University China filed Critical Northeastern University China
Priority to CN201610118914.7A priority Critical patent/CN105608492B/en
Publication of CN105608492A publication Critical patent/CN105608492A/en
Application granted granted Critical
Publication of CN105608492B publication Critical patent/CN105608492B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/10Interfaces, programming languages or software development kits, e.g. for simulating neural networks
    • CCHEMISTRY; METALLURGY
    • C21METALLURGY OF IRON
    • C21BMANUFACTURE OF IRON OR STEEL
    • C21B5/00Making pig-iron in the blast furnace
    • CCHEMISTRY; METALLURGY
    • C21METALLURGY OF IRON
    • C21BMANUFACTURE OF IRON OR STEEL
    • C21B7/00Blast furnaces
    • C21B7/24Test rods or other checking devices
    • CCHEMISTRY; METALLURGY
    • C21METALLURGY OF IRON
    • C21BMANUFACTURE OF IRON OR STEEL
    • C21B2300/00Process aspects
    • C21B2300/04Modeling of the process, e.g. for control purposes; CII

Landscapes

  • Engineering & Computer Science (AREA)
  • Chemical & Material Sciences (AREA)
  • Manufacturing & Machinery (AREA)
  • Materials Engineering (AREA)
  • Metallurgy (AREA)
  • Organic Chemistry (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Feedback Control In General (AREA)
  • General Factory Administration (AREA)

Abstract

The invention relates to a robust random-weight neural network-based molten-iron quality multi-dimensional soft measurement method which belongs to the blast-furnace smelting automatic control field, in particular to a Cauchy distribution weighted M-estimation random-weight neural network (M-RVFLNs) based method for multi-dimensional parameter-dynamic soft measurement of the molten-iron quality in the blast-furnace smelting process. According to the method of the invention, the principal component analysis (PCA) method is adopted to chose main parameters which affect the blast-furnace molten iron quality as model input variables, a molten-iron quality multi-dimensional dynamic prediction model which has an output self-feedback structure and takes into account input-output data at different moments is constructed, and it is possible to carry out multi-dimensional dynamic soft measurement of the main parameters Si content, P content, S content and molten iron temperature which represent the blast-furnace molten iron quality. The method of the invention comprises the following steps of (1) choosing auxiliary variables and determining model input variables and (2) training and using the M-RVFLNs soft measurement model.

Description

A kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net
Technical field
The invention belongs to blast furnace process automation control area, particularly a kind of based on Cauchy distribution weighting M estimation random weightThe polynary molten steel quality dynamic state of parameters of the blast furnace ironmaking process flexible measurement method of neutral net (M-RVFLNs).
Background technology
Blast furnace ironmaking, for iron is restored from iron containing compoundses such as iron ores, smelts up-to-standard molten iron, smelts ironJourney is a very complicated nonlinear dynamic process. Solid blast furnace ironmaking by occur in stove complicated gas-solid, solid-, solid-Liquid reaction restores iron from the iron containing compoundses such as iron ore, smelts up-to-standard molten iron. Meanwhile, molten steel quality refers toMark, as of paramount importance production target in blast furnace ironmaking process, has directly determined quality and the blast furnace process mistake of follow-up steel productsThe power consumption state of journey. In actual production, conventionally adopt molten iron temperature (physical thermal), molten iron silicon content [Si] (chemical heat),The parameters such as element sulphur content [S], phosphorus element content [P] are comprehensively weighed the height of molten steel quality. But, high temperature in blast furnace,High pressure, Duo Chang, multiphase coupled dynamic characteristic, a lot of inherent interference and while directly measuring harsh environment make high furnace interiorState and molten steel quality difficult parameters are to detect online. Therefore, just need to depend on other convenient parameters that detect sets up blast-meltedThe on-line predictive model of quality, adopts the detection analysis and prediction of molten iron element and temperature to set up the polynary blast-melted of data-drivenQuality soft-sensing model.
Patent publication No. " CN101211383A " discloses a kind of feature analysis and prediction method of blast furnace molten iron silicon content. With blast furnaceThe blast furnace technology parameter of molten iron silicon content forecasting model is input variable, adopts improved dynamic Independent Component Analysis to inputThe sample data of variable is carried out feature extraction, eliminates the correlation between processing parameter, uses Optimization Model of Genetic Algorithm ginsengThe least square method supporting vector machine algorithm of number is set up the dynamic recurrence model of blast furnace molten iron silicon content forecast.
Patent publication No. " CN103320559B " provides a kind of blast-melted sulfur content forecasting procedure, with sulfur content short-term average,Sulfur content average in mid-term, sulfur content long-term mean value, enter oven coke S content, enter stove coal dust S content etc. and forecast as molten iron sulfur-bearingInput variable, utilize blast furnace to form the chemical reaction process of molten iron, in conjunction with RBF neutral net, forecast containing of molten iron next timeSulfur content, has obtained good forecast sulfur content precision.
Patent publication No. CN103981317A discloses " the continuous detecting side of the blast furnace iron notch molten iron temperature based on temperature drop modelMethod ", utilize the temperature measurement data of the trough bottom thermocouple that buries, the finally molten iron temperature at identification tapping hole place. The method has solvedBlast-melted temperature detection needs artificial participation, is interrupted discontinuously, and consumptive material is many, the unsettled problem of thermometric value.
The relevant similar approach of the method for above-mentioned patent report and other pertinent literatures just for single molten steel quality element (asSi content, S content, molten iron temperature etc.) forecast and soft measurement, fail to characterizing the major parameter of blast-melted quality,Si (silicon) content, P (phosphorus) content, S (sulphur) content and molten iron temperature carry out polynary online forecasting simultaneously, thereby can not reflect comprehensivelyMolten steel quality level, practicality is poor. Meanwhile, because these methods are not considered the time lag in input and output sequential and processRelation, the static models of setting up can not reflect the intrinsic dynamic characteristic of blast furnace ironmaking process well. In addition, refine in realityIn iron production process, the fault of the device such as examined instrument and transmitter and other abnormal impacts of disturbing, warp in measurement dataOften there is outlier. These methods have mainly been considered the soft measurement of molten steel quality parameter under the desirable working of a furnace, and robustness is poor, when being subject toWhen outlier is disturbed, these methods can not suppress outlier and disturb and molten steel quality parameter is measured comparatively exactly. To sum upDescribed, at present both at home and abroad also not specially for blast furnace ironmaking process molten steel quality parameter (Si content, P content, S content and ironCoolant-temperature gage) method of carrying out the soft measurement of polynary dynamic robust.
Summary of the invention
In order to solve the deficiency of molten steel quality parameter online forecasting method in above-mentioned blast furnace ironmaking process, the present invention propose based onThe polynary molten steel quality parameter of blast furnace ironmaking of many output random weight neutral net (M-RVFLNs) that Cauchy distribution weighting M estimatesRobust Modeling, using principal component analysis (PCA) method to filter out affects the main parameter of blast-melted quality as mode inputVariable, constructs one and has output self feed back structure and consider that the molten steel quality of inputoutput data is polynary dynamically not pre-in the same timeSurvey model, can carry out many to the major parameter Si content, P content, S content and the molten iron temperature that characterize blast-melted quality simultaneouslyUnit's dynamic soft measuring. The present invention can utilize data that the existing general measure equipment of iron-smelter provides as mode input, provides currentWith the estimated value of the polynary molten steel quality parameter in following fixed time interval, for Optimum Operation and the operation of blast furnace production process provideKey Quality Indicator.
For achieving the above object, the present invention adopts following technical scheme, the present invention includes following steps:
(1) auxiliary variable is selected to determine with mode input variable
Select the auxiliary variable of soft measurement to comprise: gas flowrate in bosh u1(m3), hot blast temperature u2(DEG C), hot-blast pressure u3(KPa)、Oxygen enrichment percentage u4, blast humidity u5(RH), injecting coal quantity u6(m3/h);
Determine following 16 input variables that variable is soft-sensing model:
Current time gas flowrate in bosh u1(t)(m3);
Current time hot blast temperature u2(t)(℃);
Current time hot-blast pressure u3(t)(KPa);
Current time oxygen enrichment percentage u4(t);
Current time blast humidity u5(t)(RH);
Current time is set injecting coal quantity u6(t)(m3/h).
A upper moment gas flowrate in bosh u1(t-1)(m3);
A upper moment hot blast temperature u2(t-1)(℃);
A upper moment hot-blast pressure u3(t-1)(KPa);
A upper moment oxygen enrichment percentage u4(t-1);
A upper moment blast humidity u5(t-1)(RH);
A upper moment is set injecting coal quantity u6(t-1)(m3/h);
A upper moment Si content estimated value
A upper moment P content estimated value
A upper moment S content estimated value
A upper moment molten iron temperature estimated value
(2) training of M-RVFLNs soft-sensing model and use
(A) start: all initialization of variable;
(B), if be chosen as M-RVFLNs model training, go to (C) and read the data set that need to carry out model training; If be chosen asThe soft measurement of molten steel quality parameter, goes to (I) and transfers the blast-melted mass parameter M-RVFLNs forecasting model of having trained;
(C) reading model training desired data collection: read from database or input model training study desired data collectionZ={(xi,ti)|xi∈Rn,ti=Rm, i=1 ... N}, N >=L carries out model learning initialization, and N is number of samples, and L is hidden layer jointThe number of point, xiFor input data set, tiFor output data set; N is input data set dimension, and m is output data set dimension;
(D) data pretreatment: data are carried out to missing values filling, after being normalized afterwards, as final forecast modelTraining data;
(E) the relevant undetermined parameter of model is determined: M-RVFLNs model needs predefined undetermined parameter to comprise:
Activation primitive type g, hidden layer node number L, condition of convergence E;
(F) training of M-RVFLNs model initialization and model parameter are determined:
Based on model training sample set and the relevant undetermined parameter of predefined model, carry out study and the training of model; ModelTraining and study comprise two stages, initial phase (initializationphase) and Robust Learning stage (robustlearningphase);
The concrete steps in model initialization stage are as follows:
Step (a): from given data set O={ (xi,ti)|xi∈Rn,ti∈Rm, i=1 ... in choose training datasetZ={(xi,ti)|xi∈Rn,ti=Rm, i=1 ... N}, N >=L, comprises input data With target data Y={yj(t) | j=1,2 ..., m}, initialization hidden node is L, and robust instruction is setPractice condition of convergence E;
Step (b): choose at random input weight vector aiWith Hidden unit deviation bi,i=1,…,L;
Step (c): calculate hidden layer output matrix H
Step (d): calculate output weight beta by following formula:
Step (e): model output when calculating hidden layer node number is L:
Y ~ = H · β
Step (f): computation modeling residual error r:
r = T - Y ~
The concrete steps in model Robust Learning stage are as follows:
Step (g): normalized residual vectorWherein median () is median calculation public affairsFormula;
Step (h): standardized residual is updated to Cauchy weighting functionIn obtain m dimension data pairThe weight matrix of answeringFurther try to achieve weight function square formation W(0), wherein
μ = m e d i a n ( r i ) , a = 1 / 1 N Σ i = 1 N ( r i - r ‾ i ) 2
Step (i): according to introducing the iterative formula of exporting weight after M estimatesWhereinK=1,2 ... for iterations, iterative computation is asked forIf each estimates of parametersAllBe less than and specify condition of convergence E, iterative computation stops, and obtains and finally exports weight
(G) modeling recruitment evaluation:
Introduce root-mean-square error RMSE:
R M S E = 1 N ( H ( a i , b i , x i ) β * - Y ) ( H ( a i , b i , x i ) β * - Y ) T
Modeling error is carried out to overall merit, if the realistic operating mode standard of modeling error finishes this M-RVFLNs modelTraining study process, turns (H); If error does not meet preassigned, training, turns (E) again;
(H) preserve M-RVFLNs model: model training study finishes, using the M-RVFLNs model that obtains as blast-melted matterAmount Parameters Forecasting model;
(I) read M-RVFLNs model: the good soft measurement of blast-melted mass parameter M-RVFLNs of initial training before recallingModel
(J) process data of reading model forecast;
(K) judge that whether data are abnormal or lack; Whether 16 input data of judgment models have shortage of data situation; If haveTurn (L) and carry out data processing, carry out molten steel quality Parameters Forecasting otherwise turn (M);
(L) data processing: if there is shortage of data situation, replace by previous moment relevant variable data;
(M) molten steel quality Parameters Forecasting: after input variable data normalization is processed, the M-RVFLNs training before callingModel carries out molten steel quality Parameters Forecasting;
(N) molten steel quality Parameters Forecasting result shows: in forecast system man-machine interface, show this molten steel quality Parameters ForecastingResult;
(O) data are preserved: the correlated inputs output data of this soft measurement are saved in to corresponding historical data base, for follow-up systemSystem is assessed, revises and inquire about used.
As a kind of preferred version, activation primitive of the present invention is Sigmoid function.
As another kind of preferred version, the present invention carries out statistical analysis and accuracy evaluation to the molten steel quality actual result of month,Model initialization training data N used is 200, hidden layer node number L=35, condition of convergence E=10-6
As another kind of preferred version, the present invention adopts 16 inputs 4 to export self feed back structure, and 4 outputs are respectively: when currentThe Si content estimated value of carvingP content estimated valueS content estimated valueAnd molten iron temperature is estimatedEvaluation
Secondly, the present invention is further comprising the steps of:
Regularly the molten steel quality forecast result in set time segment limit and actual value are carried out to error analysis, if estimation error statisticsRoot-mean-square error or variance exceed user specify accuracy rating, restart model training.
In addition, the root-mean-square error of estimation error statistics of the present invention or variance exceed user specify accuracy rating refer to: ironThe root-mean-square error that coolant-temperature gage is estimated > 30 DEG C.
Beneficial effect of the present invention.
The online process data that the present invention utilizes conventional instrumentation to provide, considers between blast furnace ironmaking process input/output variableSequential and time lag relation, the M-RVFLNs intelligent modeling technology based on data-driven, has realized blast furnace ironmaking process molten steel qualityThe soft measurement of polynary robust of index. Compared with existing manual measurement or chemical examination molten steel quality index, by improved M-RVFLNsAlgorithm can be realized the soft measurement of Si content, P content, S content and molten iron temperature four large molten steel quality indexs, measurement effect simultaneouslyMore accurate, generalization ability is stronger, is more of practical significance for practical operation than the prediction by single quality index only. Meanwhile,The present invention considers in actual iron-making production, and the fault of the devices such as examined instrument and transmitter and other disturb extremelyOn the impact of modeling, model is carried out to robustness improvement, the many output random weight god who estimates based on Cauchy distribution weighting M is proposedThrough network, improve the robustness of model. Method proposed by the invention contribute to realize blast-melted quality optimal control andOptimize operation.
Brief description of the drawings
Below in conjunction with the drawings and specific embodiments, the present invention will be further described. Protection domain of the present invention is not only confined to followingThe statement of content.
The measuring instrument allocation plan of Fig. 1 blast furnace ironmaking process.
Fig. 2 is the FB(flow block) of the molten steel quality online forecasting software based on M-RVFLNs of the present invention.
The polynary molten steel quality parameter soft measurement effect figure of Fig. 3 based on M-RVFLNs.
In Fig. 1: 1 blast furnace, 2 hot-blast stoves, 3 flowmeters, 4 thermometers, 5 pressure gauges, 6 hygrometers, 7 bosh coal gasAmount analyzer, 8 oxygen enrichment percentage analyzers, 9 data acquisition units, the computer system of 10 operation soft-sensor softwares.
Fig. 1 label symbol used is as follows:
Gas flowrate in bosh---u1
Hot blast temperature---u2
Hot-blast pressure---u3
Oxygen enrichment percentage---u4
Blast humidity---u5
Injecting coal quantity---u6
Oxygen enrichment flow---v1
Cold flow---v2
Detailed description of the invention
As shown in the figure, the present invention is based on general measure system, data acquisition unit, M-RVFLNs soft-sensor software and operating softwareComputer system form, detailed structure is as shown in Figure 1. The general measure instrument such as flowmeter, pressure gauge and thermometer are installed onEach relevant position of blast furnace process system. Data acquisition unit connects general measure system, and is connected and operated in by communication busThe computer system of line software of forecasting. General measure system mainly comprises that following general measure instrument comprises:
Three flowmeters, are respectively used to on-line measurement Pulverized Coal Injection System with Fuzzy coal powder blowing amount, oxygen enrichment flow, cold flow;
A thermometer, for the hot blast temperature of on-line measurement blast-furnace hot-air system;
A pressure gauge, for the hot-blast pressure of on-line measurement blast-furnace hot-air system;
A hygrometer, for the blast humidity of on-line measurement blast-furnace hot-air system.
In addition, general measure system also comprises following two analyzers:
Cold flow, oxygen enrichment flow and a coal powder blowing amount that gas flowrate in bosh analyzer obtains by flowmeter survey,And the blast humidity that measures of hygrometer, analytical calculation goes out gas flowrate in bosh parameter;
Cold flow, oxygen enrichment flow that oxygen enrichment percentage analyzer obtains by flowmeter survey, and hygrometer is measuredThe blast humidity arriving, analysis meter calculates oxygen enrichment percentage parameter.
Implementation method of the present invention comprises, (1) auxiliary variable is selected to determine with mode input variable, (2) M-RVFLNs modelTraining and use.
(1) auxiliary variable is selected to determine with mode input variable
The blast-melted mass parameter that needs online forecasting is Si (silicon) content y1(%), P (phosphorus) content y2(%), S (sulphur) contenty3(%) with molten iron temperature y4(DEG C). According to the correlation between the surveying of process mechanism and variable, considerable and variable, selectThe auxiliary variable of soft measurement comprises: gas flowrate in bosh u1(m3), hot blast temperature u2(DEG C), hot-blast pressure u3(KPa), oxygen enrichment percentage u4、Blast humidity u5(RH), injecting coal quantity u6(m3/h)。
According to dynamic characteristic of the course, based on above-mentioned 6 auxiliary variables, determine that following 16 variablees are molten steel quality online forecasting mouldThe input variable of type:
Current time gas flowrate in bosh u1(t)(m3);
Current time hot blast temperature u2(t)(℃);
Current time hot-blast pressure u3(t)(KPa);
Current time oxygen enrichment percentage u4(t);
Current time blast humidity u5(t)(RH);
Current time is set injecting coal quantity u6(t)(m3/h).
A upper moment gas flowrate in bosh u1(t-1)(m3);
A upper moment hot blast temperature u2(t-1)(℃);
A upper moment hydro-thermal wind pressure u3(t-1)(KPa);
A upper moment oxygen enrichment percentage u4(t-1);
A upper moment blast humidity u5(t-1)(RH);
A upper moment is set injecting coal quantity u6(t-1)(m3/h);
A upper moment Si content estimated value
A upper moment P content estimated value
A upper moment S content estimated value
A upper moment molten iron temperature estimated value
(2) training of M-RVFLNs soft-sensing model and use
(A) start: all initialization of variable;
(B), if be chosen as M-RVFLNs model training, go to (C) and read the data set that need to carry out model training; If be chosen asThe soft measurement of molten steel quality parameter, goes to (I) and transfers the blast-melted mass parameter M-RVFLNs forecasting model of having trained;
(C) reading model training desired data collection: read from database or input model training study desired data collectionZ={(xi,ti)|xi∈Rn,ti=Rm, i=1 ... N}, N >=L carries out model learning initialization, and here, N is number of samples, and L isThe number of hidden layer node, xiFor input data set, tiFor output data set; N is input data set dimension, and m is output dataCollection dimension.
(D) data pretreatment: data are carried out to missing values filling, after being normalized afterwards, as final forecast modelTraining data;
(E) the relevant undetermined parameter of model is determined: M-RVFLNs model needs predefined undetermined parameter to comprise:
Activation primitive type g, hidden layer node number L, condition of convergence E.
(F) training of M-RVFLNs model initialization and model parameter are determined:
Based on model training sample set and the relevant undetermined parameter of predefined model, carry out study and the training of model; ModelTraining and study specifically comprise two stages, initial phase (initializationphase) and Robust Learning stage(robustlearningphase)。
The concrete steps of model initialization training are as follows:
Step (a): from given data set O={ (xi,ti)|xi∈Rn,ti∈Rm, i=1 ... in choose training datasetZ={(xi,ti)|xi∈Rn,ti=Rm, i=1 ... N}, N >=L, comprises input dataWith target data Y={yj(t) | j=1,2 ..., m}, initialChange hidden node is L, and robust training condition of convergence E is set.
Step (b): choose at random input weight vector aiWith Hidden unit deviation bi, i=1 ..., L.
Step (c): calculate hidden layer output matrix H
Step (d): calculate output weight beta by following formula:
Step (e): model output when calculating hidden layer node number is L:
Y ~ = H · β
Step (f): computation modeling residual error r:
r = T - Y ~
The concrete steps in model Robust Learning stage are as follows:
Step (g): normalized residual vectorWherein median () is median calculation public affairsFormula;
Step (h): standardized residual is updated to Cauchy weighting functionIn obtain m dimension data pairThe weight matrix of answeringFurther try to achieve weight function square formation W(0), wherein
μ = m e d i a n ( r i ) , a = 1 / 1 N Σ i = 1 N ( r i - r ‾ i ) 2
Step (i): according to introducing the iterative formula of exporting weight after M estimatesWhereink=1, 2 ... for iterations, iterative computation is asked forIf each estimates of parametersAllBe less than and specify condition of convergence E, iterative computation stops, and obtains and finally exports weight
(G) modeling recruitment evaluation:
Introduce root-mean-square error RMSE:
R M S E = 1 N ( H ( a i , b i , x i ) β * - Y ) ( H ( a i , b i , x i ) β * - Y ) T
Modeling error is carried out to overall merit, if the realistic operating mode standard of modeling error finishes this M-RVFLNs modelTraining study process, turns (H); If error does not meet preassigned, training, turns (E) again;
(H) preserve M-RVFLNs model: model training study finishes, using the M-RVFLNs model that obtains as blast-melted matterAmount Parameters Forecasting model;
(I) read M-RVFLNs model: the good soft measurement of blast-melted mass parameter M-RVFLNs of initial training before recallingModel Y ~ i = H ( a i , b i , x i ) β * ;
(J) process data of reading model forecast;
(K) judge that whether data are abnormal or lack; Whether 16 input data of judgment models have shortage of data situation; If haveTurn (L) and carry out data processing, carry out molten steel quality Parameters Forecasting otherwise turn (M);
(L) data processing: if there is shortage of data situation, replace by previous moment relevant variable data;
(M) molten steel quality Parameters Forecasting: after input variable data normalization is processed, the M-RVFLNs training before callingModel carries out molten steel quality Parameters Forecasting;
(N) molten steel quality Parameters Forecasting result shows: in forecast system man-machine interface, show this molten steel quality Parameters ForecastingResult;
(O) data are preserved: the correlated inputs output data of this soft measurement are saved in to corresponding historical data base, for follow-up systemSystem is assessed, revises and inquire about used.
Does (P) online forecasting finish? if desired proceed molten steel quality online forecasting, be back to (K); Otherwise turn (R).
(Q) finish.
While having significant operating mode due to actual blast furnace system, become and non-linear dynamic characteristic, molten steel quality parameter soft-sensing model existsAfter use a period of time, may there is meeting the prediction error of working condition requirement. In order to keep molten steel quality on-line predictive modelAccuracy, need regularly to adopt new sample data again to train forecasting model, its concrete grammar is: regularly to solidMolten steel quality forecast result and the actual value of fixing time in segment limit are carried out error analysis, if the root-mean-square error of estimation error statisticsOr variance exceedes the accuracy rating that user specifies, the root-mean-square error of estimating as molten iron temperature > 30 DEG C, illustrate that operation of blast furnace operating mode goes outNow significantly drift, and prediction error is larger, must restart model training.
Embodiments of the invention are the blast furnace object that a volume is 2600m3. According to the requirement of this description, this blast furnace pairResemble following general measure system be installed, comprising:
Yokogawa DPharpEJA series of pressure transmitters is for measuring the hot-blast pressure of blast-furnace hot-air system;
HH-WLB differential pressure flowmeter is used for measuring cold flow;
A+K balance flow meter is used for measuring oxygen enrichment flow;
JWSK-6CWDA air humidity sensor is used for measuring blast humidity;
YHIT infrared radiation thermometer is used for measuring hot blast temperature;
HDLWG-06 coal power flowmeter is used for measuring coal powder blowing amount.
In addition, general measure system also comprises following two analyzers:
Cold flow, oxygen enrichment flow, a coal powder blowing amount that gas flowrate in bosh analyzer measures by conventional instrument,And the blast humidity that measures of hygrometer, analytical calculation goes out gas flowrate in bosh parameter;
Gas flowrate in bosh Measurement and analysis instrument parameter arranges as follows:
Gas flowrate in bosh=1.21* cold flow/60+ (2* oxygen enrichment flow/60)+(44.8* blast humidity * (cold flow/60+ (richnessOxygen flow/60))/18000)+(22.4* hour injecting coal quantity * 1000* coal dust hydrogen content/12000)
Oxygen enrichment flow, blast humidity and cold flow that one oxygen enrichment percentage analyzer measures by conventional instrument, analyzeCalculate rich-oxygen of blast furnace rate parameter;
Oxygen enrichment percentage Measurement and analysis instrument parameter arranges as follows:
Oxygen enrichment percentage=((oxygen enrichment flow * 0.98/60+ ((0.21+ (0.29* blast humidity/8/100)) * cold flow/60))/(cold windFlow/60+ (oxygen enrichment flow/60))-(0.21+ (0.29* blast humidity/8/100))) * 100
Online forecasting program of the present invention is moved on independent computer, adopts C# high-level language to carry out the tool of forecasting procedure of the present inventionBody software is realized. This software interface has been realized the functions such as data demonstration, inquiry, the demonstration of soft measurement result and inquiry, Ke YifangJust allow operating personnel obtain its needed information. In addition, OPC bitcom being housed on this soft-sensor software computer is responsible forCarry out data double-way communication with slave computer and data acquisition unit.
M-RVFLNs adopts 16 input 4 export structures of introducing output self feed back structure. 16 inputs are respectively: current timeGas flowrate in bosh u1(t)(m3), current time hot blast temperature u2(t) (DEG C), current time hot-blast pressure u3(t) (KPa), currentMoment oxygen enrichment percentage u4(t), current time blast humidity u5(t) (RH), current time is set injecting coal quantity u6(t)(m3/ h), a upper momentGas flowrate in bosh u1(t-1)(m3), a upper moment hot blast temperature u2(t-1) (DEG C), a upper moment hydro-thermal wind pressureu3(t-1) (KPa), a upper moment oxygen enrichment percentage u4(t-1), a upper moment blast humidity u5(t-1) (RH), a upper moment is set coal powder injectionAmount u6(t-1)(m3/ h), a upper moment Si content estimated valueA upper moment P content estimated valueA upper moment S content estimated valueA upper moment molten iron temperature estimated value4 outputs are respectively:Need the Si content estimated value of the current time of estimatingP content estimated valueS content estimated valueAnd molten iron temperature estimated value
The relevant undetermined parameter of M-RVFLNs model is determined as follows:
Hidden layer node is counted L=35;
It is 200 that setting model initializes training data N used;
Activation primitive adopts Sigmoid function;
Condition of convergence E=10-6
In hidden layer matrix, input weight vector aiWith deviation bi, i=1,2 ..., L produces at random by system.
Concrete training algorithm is as shown in determining in the training of (F) M-RVFLNs model initialization and model parameter.
Fig. 3 is the molten steel quality indices prediction effect of soft measuring system a period of time of molten steel quality parameter, can find out each molten ironQuality index predicted value and its actual value are basically identical, and error is smaller, and variation tendency is basically identical. The present invention is a kind of toolThere are very high practical value, low cost blast furnace ironmaking process molten steel quality multicomponent metering means.
Be understandable that, above about specific descriptions of the present invention, only for being described, the present invention is not limited to the present invention realExecute routine described technical scheme, those of ordinary skill in the art should be appreciated that still can to the present invention modify or etc.With replacing, to reach identical technique effect; Use needs as long as meet, all within protection scope of the present invention.

Claims (6)

1. the polynary molten steel quality flexible measurement method based on robust random weight neutral net, is characterized in that comprising the following steps:
(1) auxiliary variable is selected to determine with mode input variable
Select the auxiliary variable of soft measurement to comprise: gas flowrate in bosh u1(m3), hot blast temperature u2(DEG C), hot-blast pressure u3(KPa)、Oxygen enrichment percentage u4, blast humidity u5(RH), injecting coal quantity u6(m3/h);
Determine following 16 input variables that variable is soft-sensing model:
Current time gas flowrate in bosh u1(t)(m3);
Current time hot blast temperature u2(t)(℃);
Current time hot-blast pressure u3(t)(KPa);
Current time oxygen enrichment percentage u4(t);
Current time blast humidity u5(t)(RH);
Current time is set injecting coal quantity u6(t)(m3/h).
A upper moment gas flowrate in bosh u1(t-1)(m3);
A upper moment hot blast temperature u2(t-1)(℃);
A upper moment hot-blast pressure u3(t-1)(KPa);
OnOneMoment oxygen enrichment percentage u4(t-1);
A upper moment blast humidity u5(t-1)(RH);
A upper moment is set injecting coal quantity u6(t-1)(m3/h);
A upper moment Si content estimated value
A upper moment P content estimated value
A upper moment S content estimated value
A upper moment molten iron temperature estimated value
(2) training of M-RVFLNs soft-sensing model and use
(A) start: all initialization of variable;
(B), if be chosen as M-RVFLNs model training, go to (C) and read the data set that need to carry out model training; If be chosen asThe soft measurement of molten steel quality parameter, goes to (I) and transfers the blast-melted mass parameter M-RVFLNs forecasting model of having trained;
(C) reading model training desired data collection: read from database or input model training study desired data collectionZ={(xi,ti)|xi∈Rn,ti=Rm, i=1 ... N}, N >=L carries out model learning initialization, and N is number of samples, and L is hidden layer jointThe number of point, xiFor input data set, tiFor output data set; N is input data set dimension, and m is output data set dimension;
(D) data pretreatment: data are carried out to missing values filling, after being normalized afterwards, as final forecast modelTraining data;
(E) the relevant undetermined parameter of model is determined: M-RVFLNs model needs predefined undetermined parameter to comprise:
Activation primitive type g, hidden layer node number L, condition of convergence E;
(F) training of M-RVFLNs model initialization and model parameter are determined:
Based on model training sample set and the relevant undetermined parameter of predefined model, carry out study and the training of model; ModelTraining and study comprise two stages, initial phase (initializationphase) and Robust Learning stage (robustlearningphase);
The concrete steps in model initialization stage are as follows:
Step (a): from given data set O={ (xi,ti)|xi∈Rn,ti∈Rm, i=1 ... in choose training datasetZ={(xi,ti)|xi∈Rn,ti=Rm, i=1 ... N}, N >=L, comprises input data With target data Y={yj(t) | j=1,2 ..., m}, initialization hidden node is L, and robust instruction is setPractice condition of convergence E;
Step (b): choose at random input weight vector aiWith Hidden unit deviation bi,i=1,…,L;
Step (c): calculate hidden layer output matrix H
H = G ( a 1 , b 1 , x 1 ) ... G ( a L , b L , x 1 ) . . . ... . . . G ( a 1 , b 1 , x N ) ... G ( a L , b L , x N ) N × L
Step (d): calculate output weight beta by following formula:
Step (e): model output when calculating hidden layer node number is L:
Y ~ = H · β
Step (f): computation modeling residual error r:
r = T - Y ~
The concrete steps in model Robust Learning stage are as follows:
Step (g): normalized residual vectorWherein median () is median calculation public affairsFormula;
Step (h): standardized residual is updated to Cauchy weighting functionIn obtain m dimension data pairThe weight matrix of answeringFurther try to achieve weight function square formation W(0), wherein
μ=median(ri), a = 1 / 1 N Σ i = 1 N ( r i - r i ‾ ) 2
Step (i): according to introducing the iterative formula of exporting weight after M estimatesWhereinK=1,2 ... for iterations, iterative computation is asked forIf each estimates of parametersAllBe less than and specify condition of convergence E, iterative computation stops, and obtains and finally exports weight
(G) modeling recruitment evaluation:
Introduce root-mean-square error RMSE:
R M S E = 1 N ( H ( a i , b i , x i ) β * - Y ) ( H ( a i , b i , x i ) β * - Y ) T
Modeling error is carried out to overall merit, if the realistic operating mode standard of modeling error finishes this M-RVFLNs modelTraining study process, turns (H); If error does not meet preassigned, training, turns (E) again;
(H) preserve M-RVFLNs model: model training study finishes, using the M-RVFLNs model that obtains as blast-melted matterAmount Parameters Forecasting model;
(I) read M-RVFLNs model: the good soft measurement of blast-melted mass parameter M-RVFLNs of initial training before recallingModel Y ~ i = H ( a i , b i , x i ) β * ;
(J) process data of reading model forecast;
(K) judge that whether data are abnormal or lack; Whether 16 input data of judgment models have shortage of data situation; If haveTurn (L) and carry out data processing, carry out molten steel quality Parameters Forecasting otherwise turn (M);
(L) data processing: if there is shortage of data situation, replace by previous moment relevant variable data;
(M) molten steel quality Parameters Forecasting: after input variable data normalization is processed, the M-RVFLNs training before callingModel carries out molten steel quality Parameters Forecasting;
(N) molten steel quality Parameters Forecasting result shows: in forecast system man-machine interface, show this molten steel quality Parameters ForecastingResult;
(O) data are preserved: the correlated inputs output data of this soft measurement are saved in to corresponding historical data base, for follow-up systemSystem is assessed, revises and inquire about used.
2. a kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net according to claim 1, its spyLevy and be that described activation primitive is Sigmoid function.
3. a kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net according to claim 1, its spyLevy the molten steel quality actual result being month and carry out statistical analysis and accuracy evaluation, model initialization training data N usedBe 200, hidden layer node number L=35, condition of convergence E=10-6
4. a kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net according to claim 1, its spyLevy and be to adopt 16 inputs 4 to export self feed back structure, 4 outputs are respectively: the Si content estimated value of current timeP content estimated valueS content estimated valueAnd molten iron temperature estimated value
5. a kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net according to claim 1, its spyLevy be further comprising the steps of:
Regularly the molten steel quality forecast result in set time segment limit and actual value are carried out to error analysis, if estimation error statisticsRoot-mean-square error or variance exceed user specify accuracy rating, restart model training.
6. a kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net according to claim 5, its spyLevy and be that the root-mean-square error of described estimation error statistics or variance exceed the accuracy rating that user specifies and refer to: molten iron temperature is estimatedRoot-mean-square error > 30 DEG C.
CN201610118914.7A 2016-03-02 2016-03-02 A kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net Active CN105608492B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610118914.7A CN105608492B (en) 2016-03-02 2016-03-02 A kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610118914.7A CN105608492B (en) 2016-03-02 2016-03-02 A kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net

Publications (2)

Publication Number Publication Date
CN105608492A true CN105608492A (en) 2016-05-25
CN105608492B CN105608492B (en) 2018-03-20

Family

ID=55988416

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610118914.7A Active CN105608492B (en) 2016-03-02 2016-03-02 A kind of polynary molten steel quality flexible measurement method based on robust random weight neutral net

Country Status (1)

Country Link
CN (1) CN105608492B (en)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106249724A (en) * 2016-09-14 2016-12-21 东北大学 A kind of blast furnace polynary molten steel quality forecast Control Algorithm and system
CN106295797A (en) * 2016-08-10 2017-01-04 深圳大学 A kind of FNR based on random weight network analyzes training method and the training system of model
CN106301950A (en) * 2016-09-07 2017-01-04 中国联合网络通信集团有限公司 A kind of OD stream quantitative analysis method and analytical equipment
CN107256000A (en) * 2017-05-27 2017-10-17 四川用联信息技术有限公司 Algorithm of the improved Domain Volume than weighing manufacturing process multivariate quality ability
CN107299170A (en) * 2017-08-10 2017-10-27 东北大学 A kind of blast-melted quality robust flexible measurement method
CN108345284A (en) * 2018-03-06 2018-07-31 宁波大学 A kind of quality dependent failure detection method becoming gauge block based on two
CN109142167A (en) * 2018-06-26 2019-01-04 中国矿业大学 A kind of method for building up of the online robust canonical soft-sensing model of grinding particle size
CN109492790A (en) * 2018-09-18 2019-03-19 北京光耀电力科技股份有限公司 Wind turbines health control method based on neural network and data mining
CN109635337A (en) * 2018-11-13 2019-04-16 中国矿业大学 A kind of industrial process soft-measuring modeling method based on block incremental random arrangement network
CN109695174A (en) * 2018-12-28 2019-04-30 东北大学 The PDF Shape Prediction method and system of defibrination process fiber length distribution
CN109935280A (en) * 2019-03-05 2019-06-25 东北大学 A kind of blast-melted quality prediction system and method based on integrated study
CN110066895A (en) * 2019-04-10 2019-07-30 东北大学 A kind of blast-melted quality section prediction technique based on Stacking
CN116759012A (en) * 2023-06-29 2023-09-15 中冶华天工程技术有限公司 Molten iron quality prediction method based on batch tracking and space-time conversion

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101211383A (en) * 2007-12-21 2008-07-02 浙江大学 Blast furnace molten iron silicon content feature analysis and prediction method
CN101291251B (en) * 2008-05-09 2011-04-06 国网信息通信有限公司 Synchronized control method and system for multicomputer
CN103320559B (en) * 2013-07-10 2014-12-03 鞍钢股份有限公司 Blast furnace molten iron sulfur content forecasting method

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106295797A (en) * 2016-08-10 2017-01-04 深圳大学 A kind of FNR based on random weight network analyzes training method and the training system of model
CN106301950A (en) * 2016-09-07 2017-01-04 中国联合网络通信集团有限公司 A kind of OD stream quantitative analysis method and analytical equipment
CN106301950B (en) * 2016-09-07 2019-08-09 中国联合网络通信集团有限公司 A kind of analysis method and analytical equipment of OD flow
CN106249724A (en) * 2016-09-14 2016-12-21 东北大学 A kind of blast furnace polynary molten steel quality forecast Control Algorithm and system
CN107256000A (en) * 2017-05-27 2017-10-17 四川用联信息技术有限公司 Algorithm of the improved Domain Volume than weighing manufacturing process multivariate quality ability
CN107299170A (en) * 2017-08-10 2017-10-27 东北大学 A kind of blast-melted quality robust flexible measurement method
CN107299170B (en) * 2017-08-10 2019-02-05 东北大学 A kind of blast-melted quality robust flexible measurement method
CN108345284A (en) * 2018-03-06 2018-07-31 宁波大学 A kind of quality dependent failure detection method becoming gauge block based on two
CN108345284B (en) * 2018-03-06 2020-06-16 宁波大学 Quality-related fault detection method based on two variable blocks
CN109142167A (en) * 2018-06-26 2019-01-04 中国矿业大学 A kind of method for building up of the online robust canonical soft-sensing model of grinding particle size
CN109142167B (en) * 2018-06-26 2021-02-26 中国矿业大学 Method for establishing online robust regular soft measurement model of ore grinding granularity
CN109492790A (en) * 2018-09-18 2019-03-19 北京光耀电力科技股份有限公司 Wind turbines health control method based on neural network and data mining
CN109635337A (en) * 2018-11-13 2019-04-16 中国矿业大学 A kind of industrial process soft-measuring modeling method based on block incremental random arrangement network
CN109635337B (en) * 2018-11-13 2021-04-27 中国矿业大学 Industrial process soft measurement modeling method based on block increment random configuration network
CN109695174A (en) * 2018-12-28 2019-04-30 东北大学 The PDF Shape Prediction method and system of defibrination process fiber length distribution
CN109935280A (en) * 2019-03-05 2019-06-25 东北大学 A kind of blast-melted quality prediction system and method based on integrated study
CN109935280B (en) * 2019-03-05 2022-11-15 东北大学 Blast furnace molten iron quality prediction system and method based on ensemble learning
CN110066895B (en) * 2019-04-10 2021-01-12 东北大学 Stacking-based blast furnace molten iron quality interval prediction method
CN110066895A (en) * 2019-04-10 2019-07-30 东北大学 A kind of blast-melted quality section prediction technique based on Stacking
CN116759012A (en) * 2023-06-29 2023-09-15 中冶华天工程技术有限公司 Molten iron quality prediction method based on batch tracking and space-time conversion

Also Published As

Publication number Publication date
CN105608492B (en) 2018-03-20

Similar Documents

Publication Publication Date Title
CN105608492A (en) Robust random-weight neural network-based molten-iron quality multi-dimensional soft measurement method
CN104651559B (en) Blast furnace liquid iron quality online forecasting system and method based on multivariable online sequential extreme learning machine
CN107526927B (en) Blast furnace molten iron quality online robust soft measurement method
CN107299170B (en) A kind of blast-melted quality robust flexible measurement method
CN105886680B (en) A kind of blast furnace ironmaking process molten iron silicon content dynamic soft measuring system and method
CN104750902B (en) Molten steel quality dynamic soft-measuring method based on multi output support vector regression
CN109935280B (en) Blast furnace molten iron quality prediction system and method based on ensemble learning
US10090678B2 (en) Production energy management system and computer program
CN110066895B (en) Stacking-based blast furnace molten iron quality interval prediction method
Coakley et al. A review of methods to match building energy simulation models to measured data
CN100440081C (en) Generalized predictable control system and method of air separating tower
CN105821170A (en) Soft measuring system and method for quality indexes of multielement molten iron of blast furnace
CN106909705A (en) A kind of blast-melted quality prediction method and its system
CN106709197A (en) Molten iron silicon content predicting method based on slide window T-S fuzzy neural network model
CN101845531A (en) Converter smelting endpoint molten steel carbon and temperature control system and method thereof
CN101211383A (en) Blast furnace molten iron silicon content feature analysis and prediction method
CN105177199A (en) Blast furnace gas generation amount soft measurement method
CN111915080B (en) Raw fuel cost optimal proportioning method based on molten iron quality constraint
CN102925602B (en) Furnace profile maintenance method for blast furnace operation
CN102357934A (en) Quality monitor soft sensing method based on rubber mixing process
CN106018730A (en) Coal water content measurement device and method based on coal mill inlet primary air correction
JP6206368B2 (en) Blast furnace state estimation apparatus and blast furnace state estimation method
CN103279030A (en) Bayesian framework-based dynamic soft measurement modeling method and device
CN113657037A (en) Molten iron silicon content prediction method based on time series interpolation-attention mechanism
CN102999028B (en) Gas medium continuous data early warning system digital-to-analogue constructing method

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant