CN103246801B - A kind of shaft furnace working of a furnace failure prediction method based on improving reasoning by cases - Google Patents

Abstract

Based on a shaft furnace working of a furnace failure prediction method of improving reasoning by cases, on the basis of traditional 4R cognitive model, increase the apportion model of attribute weight, and used the theoretical case calibration model that improves of GDM. Comprise: initialization of variable; Current variable is normalized, its numerical value is between 0～1; Case is represented, set up case library; Calculate based on the assign weight coefficient correlation of algorithm of water flood; Calculate the weight of case attribute; Calculate the similarity of target case and source case; Determine the number of coupling case according to similarity threshold. Effect is reused in judgement; Result to forecast is carried out GDM correction; Store corresponding case, and output function is instructed. Utilize online process data, realized the failure prediction based on improving reasoning by cases of the Approach for Shaft Furnace Roasting Process working of a furnace. Compared with the artificial judgment working of a furnace, reduce operating personnel's workload, reduce the uncertainty of artificial judgement, improve the ageing of failure prediction.

Description

A kind of shaft furnace working of a furnace failure prediction method based on improving reasoning by cases

Technical field

The invention belongs to technical field of automation, particularly a kind of working of a furnace of metallurgy industry shaft roasting production process eventThe intelligent forecasting method of barrier.

Technical background

In Approach for Shaft Furnace Roasting Process, shaft furnace working of a furnace more complicated, trouble point is a lot of, once break down, will affect lifeThe carrying out producing, and threaten the safety of personnel and equipment, therefore, need to find in time that early stage failure symptom also carry out correspondingProcess, to avoid fault to worsen and to bring unnecessary economic loss. But, the numerous and mechanism of the interference of Approach for Shaft Furnace Roasting ProcessModel is difficult to obtain, and makes conventional failure prediction method be difficult to play a role, and sets up shaft furnace therefore find a kind of suitable methodIt is necessary and urgent that the variation tendency forecasting model of the working of a furnace becomes.

Since last century, studied failure prediction the seventies, the method for failure prediction is a lot, mainly can be classified as two classes, i.e. mouldType drives method and data-driven method. Wherein, the common way of model-driven method obtains following by estimation model parameterFault signature. Because complex object often has change when slow, distributed constant, non-linear and close coupling characteristic, be difficult to accurateMathematical Modeling is described, and makes this method be difficult to use in practice. So, do not need to set up the accurate mechanism model of objectData-driven method is arisen at the historic moment, though these methods have obtained certain effect, and because method itself also exists some problems, such asBeing difficult to of the convergence problem of neutral net, rule obtains etc., so be difficult to separate in versatility, which is better and which is worse, failure predictionAccuracy is not high. The reasoning by cases (case-basedreasoning, CBR) of artificial intelligence field is a kind of new problem solvingWith machine learning method, solution procedure can be described with classical 4R cognitive model, i.e. the retrieval of case (Retrieve), heavyWith (Reuse), correction (Revise) and storage (Retain). CBR has obtained in the application in fault diagnosis and failure prediction fieldExtensive concern, demonstrates the development potentiality of CBR method. But whether the result that traditional CBR method provides rationally depends on experienceThe degree of enriching and learning ability thereof, in the process solving in CBR reasoning, also have some problems solve not yet completely, such as caseThe problem such as weight allocation, the correction of case solution of example attribute, these problems do not solve, and make equally the accuracy rate of failure prediction notHigh.

Based on above-mentioned analysis, urgently develop a kind of furnace condition prediction of improved CBR method research Approach for Shaft Furnace Roasting Process and askTopic, lays particular emphasis on the improvement of CBR method, to improve its reasoning precision, ensures the accuracy of failure prediction.

Summary of the invention

The object of the invention is to, by a kind of shaft furnace working of a furnace failure prediction method based on improving reasoning by cases is provided,For the Approach for Shaft Furnace Roasting Process that is difficult to set up mathematical models, by water-filling (water-fillingtheory, WFT) andGroup decision thought (groupdecision-making, GDM) is incorporated into reasoning by cases process, adopts data-driven method to set up stoveThe forecasting model of condition. Constructed the weight of Lagrangian calculating case characteristic according to water-filling; Further, by fixedThe confidence level of justice historical failure forecast result is carried out group decision correction to target case, thereby provides the forecast knot of the current working of a furnaceReally. Thereby reach the target that reduces rate of failing to report, rate of false alarm and rate of breakdown.

The present invention adopts following technological means to realize:

In conjunction with WFT principle and GDM thought, design a kind of new reasoning by cases system, improvement has as shown in Figure 2 been proposedThe model structure of type CBR failure prediction. On the 4R cognitive model basis of retrieving, reuse, revise and storing in tradition, increase genusProperty weight apportion model, and use the theoretical case calibration model that improves of GDM. Comprise the following steps:

Step 1, initialization of variable; Case property value is normalized, its numerical value is positioned in interval [0,1];

Step 2, case is represented, set up case library;

If every source case C_k(k=1,2 ..., p) represent the two tuple forms that can be expressed as:

C_k:<X_k；Y_k>, k=1,2 ..., wherein, p is case sum to p (1); X_kAnd Y_kRespectively source case C_kProblemDescribe (claiming process variable collection here) and answer (claiming the probability of happening of 5 kinds of faults here), can be expressed as:

X_k＝(x_1,k,x_2,k,…,x_13,k)(2)

Y_k＝(y_1,k,y_2,k,…,y_5,k) (3) wherein, x_1,k-x_13,kRepresent respectively source case C_kIn heating gas streamAmount, heated air flow amount, reducing gas flow, heating gas pressure, add heat air pressure, reducing gas pressure, calorific value of gas,Left side chamber temperature, right side chamber temperature, heating tape temperature, reduction temperature, take out of in time and stove 13 of negative pressure etc.The normalization numerical value of case attribute; y₁,_k-y₅,_kRepresent respectively source case C_kIn get angry, burn with anger, blow out, furnace and cross reduction etc.The forecast result of five kinds of faults, value is all positioned in interval [0,1], represents 5 kinds of probability that fault occurs separately;

Step 3, calculate based on the water flood algorithm (water-fillingtheory-basedweight that assigns weightAllocation, WFA) coefficient correlation;

${\mathrm{\&alpha;}}_i=\frac{\underset{k=1}{\overset{p}{\&Sigma;}}(x_{k,i}-{\stackrel{\&OverBar;}{x}}_i)(y_k-\stackrel{\&OverBar;}{y})}{\sqrt{\underset{k=1}{\overset{p}{\&Sigma;}}{(x_{k,i}-{\stackrel{\&OverBar;}{x}}_i)}^2}\sqrt{\underset{k=1}{\overset{p}{\&Sigma;}}(y_k-\stackrel{\&OverBar;}{y})}},i=1,2,\mathrm{...},13---\left(4\right)$

Wherein,Represent the mean value of i property value, x_k,iExpression source case C_kIn the normalization numerical value of i attribute,Y represents the mean value of all failure prediction results in case library, y_kExpression source case C_kIn the mean value of failure prediction result;

The weight of step 4, calculating case attribute;

The description formula that case attribute weight distributes is:

$R=\underset{i=1}{\overset{13}{\&Sigma;}}{\log }_2(1+{\mathrm{\&alpha;}}_i^2{\mathrm{\&omega;}}_i)---\left(5\right)$

In formula, the reasonability index that R is weight allocation, α_iThe coefficient correlation of the defined water-filling algorithm of formula (4), ω_iBeThe weight of i attribute, meets following constraints:

$\left\{\begin{array}{c}\underset{i=1}{\overset{13}{\&Sigma;}}{\mathrm{\&omega;}}_i=1\\ {\mathrm{\&omega;}}_i\&GreaterEqual;0\end{array}\right.---\left(6\right)$

According to the resolution principle of lagrange's method of multipliers, the equality constraint in formula (5) and formula (6) is combined,Obtain following Lagrangian:

$L(\mathrm{\&omega;},\mathrm{\&lambda;})=\underset{i=1}{\overset{13}{\&Sigma;}}{\log }_2(1+{\mathrm{\&alpha;}}_i^2{\mathrm{\&omega;}}_i)+\mathrm{\&lambda;}(1-\underset{i}{\overset{13}{\&Sigma;}}{\mathrm{\&omega;}}_i)---\left(7\right)$

Wherein, λ is Lagrangian, and LagrangianL (ω, λ) is illustrated inConstraints under try to achieveWeight be rational;

To ω in formula (7)_iAfter asking local derviation, obtain:

$\frac{\&part;L}{\&part;{\mathrm{\&omega;}}_i}=\frac{1}{ln2}\frac{{\mathrm{\&alpha;}}_i^2}{(1+{\mathrm{\&alpha;}}_i^2{\mathrm{\&omega;}}_i)}-\mathrm{\&lambda;}=0,i=1,2,\mathrm{...},13---\left(8\right)$

Can solve the assigned weight of each attribute, computing formula is:

${\mathrm{\&omega;}}_i=\frac{1+\underset{i=1}{\overset{13}{\&Sigma;}}\frac{1}{{\mathrm{\&alpha;}}_i}}{13}-\frac{1}{{\mathrm{\&alpha;}}_i^2}---\left(9\right)$

The weights omega that through type (9) calculates_i, do not meet ω in formula (6) at < 0 o'clock_i>=0 constraints, in this situationUnder, make ω_i=0, what represent this attribute act as zero; Formula (9) shows, the distribution of weight is affected by the significance level of attribute, noteThe coefficient correlation α of water algorithm_iLarger, the weight being assigned to is larger, thus, is met the weight allocation of formula (5) and formula (6)The algorithm rationalizing;

The similarity of step 5, calculating target case and source case;

Following formula is for calculating target case X and each source case X_kSimilarity:

$S(X,X_k)=\underset{i=1}{\overset{13}{\&Sigma;}}{\mathrm{\&omega;}}_i(1-\frac{|x_i-x_{i,k}|}{max(x_i,x_{i,k})}),k=1,2,\mathrm{...},p---\left(10\right)$

Wherein, ω_i(i=1,2 ..., 13) and be the weight of i attribute, represent that each attribute is right in failure prediction processThe active force of result;

Step 6, according to similarity threshold determine coupling case number. When Incomplete matching, go to step 8; While coupling completelyGo to step 7;

Effect is reused in step 7, judgement, if result is correct, goes to step 9; Otherwise go to step 8;

Step 8, the result of forecast is carried out to GDM correction, and go to step 7;

In Case Retrieval link, calculate after p similarity by formula (10), set the threshold value sim of a similarity_v∈(0,1], suppose to be greater than sim in p similarity_vCase number be m, the similarity of these cases of mark is respectively S again₁～S_m, corresponding historical failure forecast result m group altogether, is respectively Y₁～Y_m, can be expressed as:

$\left\{\begin{array}{c}{Y}_j=(y_{1,j},y_{2,j},...,y_{5,j})\\ s.t.S_j\&GreaterEqual;{\mathrm{sim}}_v\end{array},j=\mathrm{1,2},...,m\right.---\left(11\right)$

Wherein, y₁,_j-y₅,_jRepresent respectively to get angry, burn with anger, blow out, furnace and cross the j group history of five kinds of faults such as reductionForecast result, uses for reference GDM thought, above-mentioned m is organized to historical forecast result and be considered as m decision-making expert, and note test cases is concentratedForecast result is Y '=(y₁′,y₂′,…,y₅'), the overall error that m organizes between historical forecast result and test forecast result is:

$E=\underset{j=1}{\overset{m}{\&Sigma;}}\underset{i=1}{\overset{5}{\&Sigma;}}|y_{i,j}-y_i^{\&prime;}|---\left(12\right)$

Calculate the confidence level λ that each organizes historical forecast result_j(j=1,2 ..., m), definition confidence level is:

${\mathrm{\&lambda;}}_j=\frac{\underset{i=1}{\overset{5}{\mathrm{\&Sigma;}}}|y_{i,j}-y_i^{\&prime;}|}{E}---\left(13\right)$

λ_jLess, represent that j group operation result levels off to the error of test forecast result less, confidence level is just high; From GDMAngle, represents that this decision-making expert's authority is larger; Utilize the structure thought of group's cardinal utility function, can pass through eachOrganize the confidence level λ of historical forecast result_jAnd m group forecast result carries out GDM correction to the forecast of target case:

$y_i=\frac{1}{m}\underset{j=1}{\overset{m}{\&Sigma;}}\frac{y_{i,j}}{{\mathrm{\&lambda;}}_j},i=1,2,\mathrm{...},5---\left(14\right)$

Step 9, store corresponding case, and output function instructs, if furnace condition prediction does not finish, go to step 2.

The present invention compared with prior art, has following obvious advantage and useful effect:

The present invention utilizes online process data, has realized the fault based on improving reasoning by cases of the Approach for Shaft Furnace Roasting Process working of a furnaceForecast. Compared with the artificial judgment working of a furnace, reduce operating personnel's workload, reduce the uncertainty of artificial judgement, improveFailure prediction ageing. Add because case library constantly has the new knowledge that represents up-to-date operating mode, be not suitable with the old of operating mode and knowKnow and constantly deleted again replacement, so failure prediction method has very strong self adaptation and self-learning capability in the present invention. In addition,Because the present invention can realize the reasonable forecast of fault, the fault time of having reduced rate of failing to report, rate of false alarm and production, reach and fallenThe object of rate of breakdown in low production process.

Brief description of the drawings

Fig. 1 is Approach for Shaft Furnace Roasting Process and fault schematic diagram;

Fig. 2 is the model structure schematic diagram of modified CBR failure prediction;

Fig. 3 the present invention is based on the schematic flow sheet of the shaft furnace working of a furnace failure prediction method of improving reasoning by cases.

Detailed description of the invention

Below in conjunction with Figure of description, the specific embodiment of the present invention is described in detail.

Referring to shown in Fig. 1, is Approach for Shaft Furnace Roasting Process and fault schematic diagram. Air blast V₁The air and the control valve V that carry₂The coal gas of carrying is in the mixed combustion of the combustion chamber of both sides, be heated to 700 by enter pending raw ore in stove from shaft furnace top～850 DEG C, ore falls to zone of reduction and is cooled to 570 DEG C of left and right, and with control valve V₃The coal gas generation reduction reaction of carrying is rawBecome ferromagnetic roasted ore, finally by motor V₄Take out of outside stove. In the production process of shaft furnace, there are 5 kinds of modal faults,Respectively the y that gets angry₁, y burns with anger₂, blow out y₃, furnace y₄With mistake reduction y₅, these several faults may occur also likely to distinguish simultaneouslyOccur. Once break down, the continuity of production will be broken, and may threaten the safety of personnel and equipment, obviously, stopThe generation tool significance of fault. In general, shaft furnace fault be not to accomplish in one move, have quantitative change accumulationProcess, therefore find in time early stage failure symptom, and notify operating personnel to take appropriate measures, and just can avoid faultWorsen and occur.

For the y that gets angry₁, y burns with anger₂, blow out y₃, furnace y₄With mistake reduction y₅The distribution of these 5 kinds of faults, gets angry and betidesFurnace roof, burns with anger and betides the combustion chamber of both sides, blows out and appears at heating tape or zone of reduction with furnace, crosses reduction and goes back at oreWhen former reaction, easily occur. By the Analysis on Mechanism of roasting, the reason that these 5 kinds of faults are initiated is a lot, is mainly manifested in as ShiShimonosekiBe in formula:

(y₁(k+1),y₂(k+1),…,y₅(k+1))＝f(x₁(k),x₂(k),…,x₁₃(k)；y₁(k),…,y₅(k)|Ω(k)) (15) wherein, x₁～x₃Represent respectively heating gas flow, heated air flow amount, the reducing gas flow in k moment; x₄～x₆Represent respectively heating gas pressure, add heat air pressure, reducing gas pressure; x₇It is calorific value of gas; x₈～x₁₁Respectively leftSide combusion chamber's temperature, right side chamber temperature, heating tape temperature, reduction temperature; x₁₂The time of taking out of, x₁₃It is negative pressure in stove; ΩThe design parameter set that represents shaft furnace, these parameters are all functions of time. Formula (16) is the non-line of a structural parameters the unknownProperty dynamical equation, has comprehensive complexity, shows: the design parameter Ω of shaft furnace is due to of the remote past, mostly off-design value,Be difficult to obtain mechanism model accurately; Between the variation of each parameter and every kind of fault, exist close coupling relation, make Fault PreThe rule of report is difficult to extract. The existence of these complicated factors, the failure prediction method based on model or rule is difficult to play a role.

If can utilize the rich experiences of long-term accumulation, formulate a set of complete Fault Pre syndicate to Approach for Shaft Furnace Roasting ProcessSystem, analyzes, provides in time fault pre-alarming the situation in Approach for Shaft Furnace Roasting Process based on this, and operator is graspedCoach, above-mentioned several typical faults just may obtain avoiding to a certain extent. In order to improve systematic function and to reduce operationMember's dependence, expert forecast and CBR that we have successively studied Approach for Shaft Furnace Roasting Process fault tendency forecast, but due to ruleBe difficult to obtain, the effect that expert is forecast is unsatisfactory, and failure prediction based on CBR is in the index such as rate of failing to report and rate of false alarmMake progress though upper, still have nearly 40% fail to report and 10% wrong report, when traditional CBR method application is described, also there is limitationProperty, be worth the improvement of further research method.

Referring to shown in Fig. 2, is the model structure schematic diagram of modified CBR of the present invention failure prediction. In conjunction with WFT principle withGDM thought, designs a kind of new reasoning by cases system, has proposed the model knot of modified CBR failure prediction as shown in Figure 2Structure. On the basis of traditional 4R cognitive model, increase the apportion model of attribute weight, and used the theoretical case school of improving of GDMPositive model. In figure, the major function of each several part is:

What in case library, store is the example forecasting in the past, is called source case, when process parameter x₁～x₁₃CompositionTarget case X inputs to after system, need to forecast 5 kinds of fault y shown in Fig. 1₁～y₅The probability Y occurring separately. Examine by caseRope, generally has two kinds of results to occur: the one, in case library, find the source case of mating completely with target case, in this caseThe predicting condition of fault directly can be reused; The 2nd, the source case retrieving and target case Incomplete matching, nowShould not directly use, need make suitable adjustment, amendment to the failure prediction result retrieving, after GDM proofreaies and correct, carry out againReuse. The effect process of reusing described in above two kinds of situations is assessed, if effect is undesirable, still needs forecast result to carry out repeatedlyAfter correction, go again multiplexing. So repeat, assess while requirement until the value of forecasting of fault reaches, thus the current state of obtainingThe probability Y that under X, various faults occur, exports the Operating Guideline of avoiding fault to occur, for operator's reference by man-machine interface. ?After enter case memory phase, this target case X and corresponding operation result Y are stored in case library, forFailure prediction next time. In a word, the function embodiment of failure prediction model is in the following aspects:

(1) can realize intelligent decision and the reasoning of the shaft furnace working of a furnace;

(2) in reasoning process, significance level that can objective analysis fault signature, and give suitable weight;

(3) can realize group decision to the forecast result of fault proofreaies and correct;

(4), in running, can provide friendly Operating Guideline and suggestion.

Refer to shown in Fig. 3, for the flow process that the present invention is based on the shaft furnace working of a furnace failure prediction method of improving reasoning by cases is shownIntention. As can be seen from the figure, concrete implementation step is as described in flow chart Fig. 3:

Step 1: initialization of variable;

Step 2: case is represented, set up case library;

Step 3: calculate coefficient correlation;

Step 4: the weight of calculating case attribute;

Step 5: the similarity of calculating target case and source case;

Step 6: the number of determining coupling case according to similarity threshold. When Incomplete matching, go to step 8; While coupling completelyGo to step 7;

Step 7: effect is reused in judgement, if result is correct, goes to step 9; Otherwise go to step 8;

Step 8: the result to forecast is carried out GDM correction, and goes to step 7;

Step 9: store corresponding case, and output function guidance, if furnace condition prediction does not finish, go to step 2.

Step 10: in order to verify validity of the present invention, carried out application study in certain ore dressing plant. Concrete scheme is as follows:

According to long-term production practices, sum up some the source cases for failure prediction, totally 150 records, are depositedBe stored in case library. Article 1 source case in case library is to form like this: it is X that the problem shown in formula (3) is described₁＝(x_1,1,x_2,1,…,x_13,1)=(3402,3191 ... ,-2.16), the answer shown in formula (4) is Y₁＝(y_1,1,y_2,1,…,y_5,1)=(0.73,0.04 ..., 0.01), the process of establishing of other source case is similarly. Develop the man-machine boundary of failure prediction modelFace and calculation procedure.

In order to eliminate the impact of different dimensions, first to initialize to the historical data in case library i.e. normalizationTo process. the characteristic value with i attribute in formula (2) is treated to example, finds out the minimum of a value min (x of its characteristic value_i) and maximummax(x_i), then calculate with following formula

$x_i^{\&prime;}=\frac{x_i-\min \left(x_i\right)}{\max \left(x_i\right)-\min \left(x_i\right)}---\left(16\right)$

Adopt improved CBR method, researched and developed the fault prediction system based on man-machine interaction. Main reasoning forecasting processDivide three steps to describe:

First, determine each parameter x by WFA method₁～x₁₃Corresponding weight is respectively: ω₁＝0.034，ω₂＝0.027，ω₃＝0.086，ω₄＝0.028，ω₅＝0.029，ω₆＝0.030，ω₇＝0.105，ω₈＝0.110，ω₉＝0.112，ω₁₀＝0.085，ω₁₁＝0.123，ω₁₂＝0.132，ω₁₃＝0.099。

Secondly, observe 60 seconds internal procedure parameter x₁～x₁₃Variation tendency, except heating gas pressure x₄And reduction temperaturex₁₁Outside playing pendulum, other parameter is among dynamic stability, and wherein, heating gas pressure is become from 2.5KPa3.1KPa, reduction temperature becomes 579 DEG C from 576 DEG C. To x₄And x₁₁Detected value can find out after being normalized, 60sInterior heating gas pressure x₄In the trend that continues to rise, zone of reduction temperature x₁₁There is rising trend. The detected value process of each parameterDCS Process Control System forms target case X after processing, calculate X and each source case history X_k(k＝1,2,…,150)Similarity, and according to set similarity threshold sim_v=0.65, find the number m=17 that mates case.

Finally, according to GDM bearing calibration, the result of forecast is adjusted, and exported forecast result respectively: y gets angry₁=0.03, the y that burns with anger₂=0.87, blow out y₃=0.02, furnace y₄=0.93, cross reduction y₅=0.95, corresponding Operating Guideline is:Current zone of reduction temperature slowly raises, and heating gas pressure has obvious ascendant trend, and other parameters are in normal range of operation. ?Likely the fault of generation successively: cross reduction, furnace, burn with anger, the probability that other faults occur is less. Suggestion minimizing adds hot coalGas supply, takes out of soon a little the time or suitably reduces reducing gas supply. Prove the validity of institute's extracting method.

The CBR forecasting procedure of application enhancements (improvedCBR-basedfaultpredictionmethod,ICBRP), after, by the operation of a period of time, added up in 30 days rate of failing to report and the rate of false alarm of fault in Approach for Shaft Furnace Roasting Process, withThere is no failure prediction model (withoutfaultprediction, WFP), rule-based forecast (rule-basedPrediction, RBP), forecast (BP-basedprediction, BPP) based on BP neutral net, traditional CBR forecastCBR forecasting procedure (the CBRprediction of (traditionalCBRprediction, TCBRP) and application WFAWithWFA, CBRPW) result contrast.

The fault frequency of ICBRP method in 30 days is 5, and rate of failing to report is 16.7%; CBRPW method is in 30 daysFault frequency is 7, and rate of failing to report is 23.3%; And the fault frequency of TCBRP is 11, rate of failing to report is 36.7%; RBP'sFault frequency is 13, and rate of failing to report is 43.3%; The fault frequency of BPP is 12, and rate of failing to report is 40%; The fault of WFPFrequency is 20, and rate of failing to report is 66.7%. Visible, the rate of failing to report of WFP is the highest; TCBRP, RBP, the failing to report of tri-kinds of methods of BPPRate is equally matched, is 40% left and right; The rate of failing to report that it should be noted that CBRPW method has declined 13.4% than TCBRP, explanationWFA assigns weight and can reduce rate of failing to report than traditional equal method of weighting; Meanwhile, ICBRP (existing WFA also has GDM to proofread and correct) ratioThe rate of failing to report of CBRPW method has declined again 6.6%, and the participation that GDM correction link has been described can further improve rate of failing to report and refer toMark.

In 30 days, their fault handling times used are respectively 325min, 392min, and 590min, 567min,579min, 1090min, illustrates that the application of the ICBRP with GDM and WFA can significantly reduce the time of troubleshooting, is ensureing lifeThe continuity aspect of producing has advantage.

ICBRP, CBRPW, WFP, TCBRP, RBP, the rate of false alarm of BPP is respectively 3.3%, 3.3%, 6.67%, 10%,13.3%, 13.3%.CBRPW has reduced by 6.7% than the rate of false alarm of TCBRP, illustrates that the weight of WFA distributive property is compared with equal weight, the result that can obtain. On the whole, improved ICBRP method has reduced rate of false alarm than other method, ICBRP'sLearning ability comparative superiority.

The True Positive Rate (truepositiverate, TPR) being defined as follows and false positive rate (falsepositiverate,FPR)：

TPR(％)＝TP/(TP+FN)×100％

FPR(％)＝FP/(FP+TN)×100％

Wherein, kidney-Yang (truepositive, TP) represent the working of a furnace while in fact showing as malfunction model judge also and beFault; Kidney-Yin (truenegative, TN) represents that working of a furnace state and model judgement are malfunction; False sun (falsePositive, FP) represent that model is judged as fault, but the working of a furnace is actually normally; False cloudy (falsenegative, FN) representsModel judges that the working of a furnace is for normal, and in fact shows as fault. Calculate each forecasting procedure forecast by formula aboveTPR (%) after 30 times and FPR (%), as shown in table 1.

The performance of the each failure prediction method of table 1

According to the result of calculation of table 1, the fine or not order that can draw method is ICBRP successively, CBRPW, and TCBRP, BPP,RBP and WFP. further analyze, and can see that CBRPW method that WFA assigns weight is than traditional equal weight TCBRP method performanceSuperior, compare the only CBRPW method with WFA with GDM and the ICBRP of WFA and promote to some extent again at aspect of performance.

Comprehensive above-mentioned experiment effect, can see: improved forecasting procedure can normally be worked, can reduce rate of failing to report,The fault time of rate of false alarm and production, more superior in aspect of performance performance.

Claims (2)

1. based on improving the shaft furnace working of a furnace failure prediction method of reasoning by cases, it is characterized in that: retrieve in tradition, reuse,On the 4R cognitive model basis of revising and store, increase the apportion model of attribute weight, and use Group Decision Theory to improve caseCalibration model; Comprise the following steps:

Step 1, initialization of variable; Case property value is normalized, its numerical value is positioned in interval [0,1];

Step 2, case is represented, set up case library;

If every source case C_k(k=1,2 ..., p) represent the two tuple forms that can be expressed as:

C_k:<X_k；Y_k>,k＝1,2,…,p(1)

Wherein, p is case sum; X_kAnd Y_kRespectively source case C_kProblem describe and answer, be expressed as:

X_k＝(x_1,k,x_2,k,…,x_13,k)(2)

Y_k＝(y_1,k,y_2,k,…,y_5,k)(3)

Wherein, x_1,k-x_13,kRepresent respectively source case C_kIn heating gas flow, heated air flow amount, reducing gas flow, addHeating gas pressure, add heat air pressure, reducing gas pressure, calorific value of gas, left side chamber temperature, right side chamber temperature,Heating tape temperature, reduction temperature, take out of the normalization numerical value of 13 case attributes such as negative pressure in time and stove; y_1,k-y_5,kPointDo not represent source case C_kIn get angry, burn with anger, blow out, furnace and cross the forecast result of five kinds of faults such as reduction, value is all positioned atIn interval [0,1], represent corresponding fault rate;

The coefficient correlation of step 3, calculating water-filling algorithm;

${\mathrm{\&alpha;}}_i=\frac{\underset{k=1}{\overset{p}{\&Sigma;}}(x_{k,i}-{\stackrel{\&OverBar;}{x}}_i)(y_k-\stackrel{\&OverBar;}{y})}{\sqrt{\underset{k=1}{\overset{p}{\&Sigma;}}{(x_{k,i}-{\stackrel{\&OverBar;}{x}}_i)}^2}\sqrt{\underset{k=1}{\overset{p}{\&Sigma;}}{(y_k-\stackrel{\&OverBar;}{y})}^2}},i=1,2,...,13---\left(4\right)$

Wherein,Represent the mean value of i property value, x_k,iExpression source case C_kIn the normalization numerical value of i attribute,TableShow the mean value of all failure prediction results in case library, y_kExpression source case C_kIn the mean value of failure prediction result;

The weight of step 4, calculating case attribute;

The description formula that case attribute weight distributes is:

$R=\underset{i=1}{\overset{13}{\&Sigma;}}{\log }_2(1+{\mathrm{\&alpha;}}_i^2{\mathrm{\&omega;}}_i)---\left(5\right)$

In formula, the reasonability index that R is weight allocation, α_iThe coefficient correlation of the defined water-filling algorithm of formula (4), ω_iIThe weight of individual attribute, meets following constraints:

$\{\begin{array}{c}\underset{i=1}{\overset{13}{\&Sigma;}}{\mathrm{\&omega;}}_i=1\\ {\mathrm{\&omega;}}_i\&GreaterEqual;0\end{array}---\left(6\right)$

According to the resolution principle of lagrange's method of multipliers, the equality constraint in formula (5) and formula (6) is combined, obtainFollowing Lagrangian:

$L(\mathrm{\&omega;},\mathrm{\&lambda;})=\underset{i=1}{\overset{13}{\&Sigma;}}{\log }_2(1+{\mathrm{\&alpha;}}_i^2{\mathrm{\&omega;}}_i)+\mathrm{\&lambda;}(1-\underset{i=1}{\overset{13}{\&Sigma;}}{\mathrm{\&omega;}}_i)---\left(7\right)$

Wherein, λ is Lagrangian, and LagrangianL (ω, λ) is illustrated inConstraints under the power of trying to achieveHeavily rational;

To the ω in formula (7)_iAfter asking local derviation, obtain:

$\frac{\&part;L}{\&part;{\mathrm{\&omega;}}_i}=\frac{1}{ln2}\frac{{\mathrm{\&alpha;}}_i^2}{(1+{\mathrm{\&alpha;}}_i^2{\mathrm{\&omega;}}_i)}-\mathrm{\&lambda;}=0,i=1,2,...,13---\left(8\right)$

Can solve the assigned weight of each attribute, computing formula is:

${\mathrm{\&omega;}}_i=\frac{1+\underset{i=1}{\overset{13}{\&Sigma;}}\frac{1}{{\mathrm{\&alpha;}}_i}}{13}-\frac{1}{{\mathrm{\&alpha;}}_i^2}---\left(9\right)$

The weights omega that through type (9) calculates_i, do not meet ω in formula (6) at < 0 o'clock_i>=0 constraints, in this case, orderω_i=0, what represent this attribute act as zero; Formula (9) shows, the distribution of weight is affected by the significance level of attribute, water-filling algorithmCoefficient correlation α_iLarger, the weight being assigned to is larger, and thus, the weight allocation that is met formula (5) and formula (6) is rationalizedAlgorithm;

The similarity of step 5, calculating target case and source case;

Following formula is for calculating target case X and each source case X_kSimilarity:

$S(X,X_k)=\underset{i=1}{\overset{13}{\&Sigma;}}{\mathrm{\&omega;}}_i(1-\frac{\left|x_i-x_{i,k}\right|}{Max(x_i,x_{i,k})}),k=1,2,...,p---\left(10\right)$

Wherein, ω_i(i=1,2 ..., 13) be the weight of i attribute, represent each attribute in failure prediction process to resultActive force;

Step 6, according to similarity threshold determine coupling case number; When Incomplete matching, go to step 8; While coupling completely, turn stepRapid 7;

Effect is reused in step 7, judgement, if result is correct, goes to step 9; Otherwise go to step 8;

Step 8, the result of forecast is carried out to group decision correction, and go to step 7;

In Case Retrieval link, calculate after p similarity by formula (10), set the threshold value sim of a similarity_v∈(0,1]，Suppose to be greater than sim in p similarity_vCase number be m, the similarity of these cases of mark is respectively S again₁～S_m, correspondingAltogether m group of historical failure forecast result, be respectively Y₁～Y_m, can be expressed as:

$\left\{\begin{array}{c}{Y}_j=(y_{1,j},y_{2,j},...,y_{5,j})\\ s.t.S_j\&GreaterEqual;{\mathrm{sim}}_v\end{array}\right.,j=1,2,\mathrm{...},m---\left(11\right)$

Wherein, y_1,j-y_5,jRepresent respectively to get angry, burn with anger, blow out, furnace and historical forecast of j group of crossing five kinds of faults such as reductionAs a result, use for reference group decision thought, above-mentioned m is organized to historical forecast result and be considered as m decision-making expert, it is pre-that note test cases is concentratedReport result be Y '=(y '₁,y′₂,…,y′₅), the overall error that m organizes between historical forecast result and test forecast result is:

$E=\underset{j=1}{\overset{m}{\&Sigma;}}\underset{i=1}{\overset{5}{\&Sigma;}}\left|y_{i,j}-y_i^{\&prime;}\right|---\left(12\right)$

Calculate the confidence level λ that each organizes historical forecast result_j(j=1,2 ..., m), definition confidence level is:

${\mathrm{\&lambda;}}_j=\frac{\underset{i=1}{\overset{5}{\&Sigma;}}\left|y_{i,j}-y_i^{\&prime;}\right|}{E}---\left(13\right)$

λ_jLess, represent that j group operation result levels off to the error of test forecast result less, confidence level is just high; From group decision angleDegree, represents that this decision-making expert's authority is larger; Utilize the structure thought of group's cardinal utility function, historical by each groupThe confidence level λ of forecast result_jAnd m group forecast result carries out group decision correction to the forecast of target case:

$y_i=\frac{1}{m}\underset{j=1}{\overset{m}{\&Sigma;}}\frac{y_{i,j}}{{\mathrm{\&lambda;}}_j},i=1,2,...,5---\left(14\right)$

Step 9, store corresponding case, and output function instructs, if furnace condition prediction does not finish, go to step 2.

2. a kind of shaft furnace working of a furnace failure prediction method based on improving reasoning by cases according to claim 1, its feature existsIn: described Y_k＝(y_1,k,y_2,k,…,y_5,k) in y₁Expression is got angry, y₂Expression is burned with anger, y₃Expression is blown out, y₄Represent furnace, y₅Represented reduction.