Hybrid intelligent control of combustion process for ore-roasting furnace

J Control TheoryAppl 2008 6(1) 80-85DOI 10.1007/51176-008-7001-6Hybrid intelligent control of combustion processfor ore-roasting furnaceAjun YAN ", Tianyou CHAI2, Fenghua WU2, Pu WANG 1(.College of Elctronic Information and Control Engineering. Beijing Univesity of Technology Beijing 10022,2 China;2.Reserch Center of Automation, Northeasterm University, Shenyang Liaoning 10004, China)Abstract: Because of its synthetic and complex characteristics, the combustion process of the shaft ore-roastingfurace is very difficult to control stably. A hybrid intelligent control approach is developed which consists of two systems:one is a cascade fuzzy control system with a temperature soft-sensor, and the other is a ratio control system for air flow witha compensation model for heating gas flow and air-fuel ratio. This approach combined itelligent control, soft- sensing andfault diagnosis with conventional control. It can adjust both the heating gas flow and the air-fuel ratio in real time. By thisway, the difficulty of online measurement of the fumace temperature is solved, the fault ratios during combustion processis decreased, the steady control of the fumace temperature is achieved, and the gas consumption is reduced. The successfulapplicaion in shaft furmaces of a mineral processing plant in China indicates its efectivness.Keywords: Furnace temperature; Inelligent control; Soft- sensing; Fault diagnosis; Shaft furace1 Introductiontine control. However, the applications of similar methodsThe first shaft furnace was built in 1926, and it fulfls the on the combustion control of a shaft ore-roasting fumaceore- roasting process which comprises an ore feed system, a have rarely been reported by far.combustion system, a high temperature deoxidization sys- To solve the problem that it is dificult to realize the sta-tem and a cool-discharge system. Among these sytems, the ble control, fault detection, and online measurement of thecontrol effect of the combustion system is significant to en- temperature for the shaft ore-roasting furmace, this papersure that the shaft furmace works reliably and to make the presents a hybrid intelligent control approach which inte-whole mineral processing plant more economical [1]. Be- grates the respective superiorities of neural network, expertcause the combustion system is a complex process which control, case-based reasoning and routine PID control. Thecontains the properties of multi-distributed parameters, in- secure, economic and stable control of the combustion pro-tensive nonlinear, large time-delay, and some faults occur- cess is realized by the cooperation work of a cascade fuzzyring frequently, it is virtually impossible to realize the con- control system for fumace temperature, a ratio control sys-trol stabilization of the furmace temperature with any single tem for air flow with a soft sensing model, plus a fault di-routine control method such as PID control.agnosis model. Thus, an excellent basic system has beenIn recent years, the development of expert systems (ES) constructed for the sucessful implementing of the muli-[2], case -based reasoning (CBR) [3]. fuzzy control (FC)[4] variable itelligent optimizing control system [12] of theand neural networks (NN) [5] has brought new fields to the ore-roasting process in the shaft furmace.researchers in engineering applications. Impacting these in-tligent control methods on the objet of temperature isa 2 Combustion system for furnace tempera-focal point in engineering applications at all times. Exam-ples of such control methods are the intelligent control ofFig.1 ilustrates the combustion system of the shaft ore-exit temperature in a gas-fuel can-type combustor [6], and roasting furmace, where the heating gas spouts into the com-the superisory control of furnace temperature [7]. Also, bustion chamber afer pssing the master valve, the confu.some pplied cases of process control [8, 9], fault diagnosis ence gas pipe, and the dividing valve seqaentill, while the[10] and process variable sf-sensing [1] have exhibited heaing air is blown in by a tasducerdriving supply fan.the prominent aspects of both ielelient control and rou- The heating gas and heating air will be mixed up to combustReceived 3 Januany 2007; revised 29 August 2007.中国煤化工This work was supported by the National Key Basic Research and Development7, andnd posored by theScientific Research Foundation for the Doctor of Beijing University of TechnologMHC N M H Gg Projet for AcademieHuman Resources Development in Intitions of Higher Leamning Under the Junsdiction of Bejig Muncipality (nnovaive Research Team onthe Control Theory, Technology Research and Aplicaion).A. YAN et al.1J Control Theory Appl 2008 6(1) 80-8581and will provide heat for the required furnace temperatureFor a long time, the combustion process has been work-(normally within the range of 1050°C~1 150°C).ing in the pattemn of manual operation. Depending on theOres from ore-feeding systempatrolling detection to the condition of combustion, the op-erators make judgments and decisions on the varying infor-Heating gas Dividingmation with their experiences. Then, they adjust the flowMaster valve valverate of heating gas and heating air manually. However, thetX-operators can bardly discover the variety of information dueXto the complication of the working conditions including theSupplycalorific value of the heating gas fuctuating frequently, andfan:|the furnace temperature being hard to be measured online.Consequently, neither the safety and the economy nor theg XHaccuracy of the control of the combustion process can beX→x+guaranteed, which may lead to such situations that the faultsoccur frequently and inferior control performances of thefumnace temperature emerge with excessive gas consump-Confluence gas pipetionConfuence airpipe I3 Hybrid intelligent control for the combus-To high temperature deoxidization systemFig. 1 Combution system of the shaft ore roating furnace.tion processThe dominant variable, the furmace temperature, is infu-3.1 Structure and functionenced by various factors during the combustion process. ItThe hybrid itelligeat control strategy for the combustioncan be described approximately byprocess is presented in Fig.2, which combines intelligentcontrol method, soft- sensing and fault diagnosis with rou-T = f(u,u,p,h,n),(1)tine control. In Fig.2, Actuator 1 represents the adjustmentwhere T denotes the furmace temperature, u represents thevalve of heating gas, Actuator 2 indicates the regulationgas consumption, v means the air expenditure,pis the pres- transducer of heaing air, K is the arfrel ratio, Ti referssure of heaing gas, h stands for the calorific value of heat- to the set-point of fumace temperature, F is the fault duringing gas, and n refers to the negaive pessre inside the fur- the combustion pocess B means the boundary condition,Inace.is the information from production line, X* denotes the sta-Equation (1) rflects a nonlinear and complex process,tistical value of the fumace temperature measured manually,both the structure and the parameters of which are unknown.e1 is the error between set- point and feedback of the heat-The emboied complexiy of this system icudes the fol- ing gas ARow, e2 stands for the eror between stpoint andlowing aspects. First, the large time-delay and uncertaintyfeedback of the heating air fow, e3 is the error between set-hamper the temperature control. Second, the components ofpoint and feedback of the furmace temperature, Us denotesthe heating gas are unstable. The pressure and the calorificthe set-point of heating gas flow, Un denotes the compensa-value of the heating gas are always in random fluctuation.tion value of heating gas flow, Ug is the set-point of heatingThese all bring the uncertainty to the system. Third, the fur-air, u1 represents the input of the control of heating gas flow,nace temperature can hardly be measured on-line. Fourh, u2 is the input of lhe control of heating air Aow. The mean-the beat exchange is very complicated inside the furmace asings of other variables may be referred to (1). From Fig.2,the ores are under a flowing state and the hot gas flows ad-the proposed method involves two parts: one is of cascadeversely to the ores. Fifth, the combustion process is acom- fuzzy temperature control with a sftsensor, the other isma-panied by a series of faults such as the fre emitting fromtio control for air flow with a compensation model for boththe combustion chamber (FE), the fire ataching to the topheating gas flow and the air-fuel ratio. The main functionsof the furmace (FA) and the ores inside the furnace beingof these parts are as follows:melted and sticking to the inner wall of the furnace (MS).1) Ratio control for heating air fow Because of the ad-Sixth, the production information and boundary conditions justment with a fixed ratio of heating gas and air can rarelyinterfere with the system frequenty. The former includes met the tchnical requirement, the air-fuel ratio K shouldthe working status of the right and left crriers as well as be coreted in real time to improve the combustion eff-the belt conveyer that bring ores to the furnace. The ltter ciency中国煤化工ation model for gasconsists of the ore's category, its performance in roasting flow anMH. CNM H Srealizaton begisheating air is basedand its grade in size. Fnally, the mechanism of this process on the Iis complex, which makes the model-based control methodwith the detection of the anteceding of the varying of com-invalid.bustion conditions, responding to the negative pressure n8A. YAN et al. /J Control TheoryAppl 2008 6(1) 80-85inside the furmace, the pressure of heating gas p, the soft-2) Cascade fuzzy temperature control It is composedsensing value of the fumace temperature T, the calorific of a temperature soft-sensing model, a fuzzy controllervalue of heating gas h along with some of their varying and a PID controller. The soft-sensing model analyzes andtrends. Then the neural network-based compensation model disposes the combustion condition to attain the predictionadopts the fault during the combustion process F, and in- value T of the furmace temperature. Comparing T with thetegrates the boundary conditions set B and the informationset-point T, the fuzzy controller absorbs their error ez andset I, after which it suggests a compensation value to heat- the temperature change speed dez/dt and presents the set-ing gas Un and a correction value to the air-fuel ratio K. point of heating gas ug. PID1 calculates the input of the con-Next, the controller PID2 based on the error e2 between the trol of adjusting valve U1 of heating gas, based on the set-set point Us and measured value U presents the input of the point of gas flow ug, compensation value un, and the errorcontrol of the supply fan's transducer u2 through the PID e1 of real feedback u, through the PID control arithmetic.arithmetic. As such, the stable tracking control of air flow Thus, the stable tracking control of heating gas fow may becan be realized.flilled.TemperatureCascade fuzzy temperature controlsof-sensingmodelx*. e,(k)u(k)Actuator|PIDIde,/ dr controllerCompensation model for gasflow and air-fuel ratioAir flow ratio controlWorking|FNeuralconditions B network baseddiagnosiscompensationPID2Actuator2! modele,(k)(k)|Fig. 2 Frame of hybrid neleligeat control for the fumace temperature.Based on the above analysis, the integrated control of the experts' experiences, the following rules for fault diagnosiscombustion process via the cascade fuzzy control for tem- are atained:perature and the ratio control for air flow can achieve therule 1: IFp < 3.8KPaandh > 4000KJ/m3 and n 1000°C,THEN F =“FE will hap-the temperature, and avoiding the faults.pen";3.2 Realization of the control algorithmrule2:IFp>3.8KPaandh<4000KJ/m3andn>1) Working conditions diagnosis model Taking advan--1.5KPaand T > 1000°C, THEN F =“FA will hap-tage of the existing itelligent fault diagnosis expert system pen ;for the ore-roasting process [13], the inelligent diagnosis ofrule3:IFp > 3.8KPa andh < 4000KJ/m3 andn 1200PC,THEN F =“MS will hap-The knowledge involved in it can be expressed with the gen- pen’ ;eral production rules given byrule4: IFp < 3.8KPa and h < 4000KJ/m3 and n >-1.5KPa and T < 1000°C,THEN F =“GE will hap-IF THEN ,2)pen”中国煤化工where F denotes the four typical faults during the combus- TheMHCNMH G'g conditions diag-tion process, including FE, FA, MS, and the gas exploding nosis minputs to the neu-(GE) inside the furmace. The meanings of other variables ral network -based compensation model. Also, the outputsmay be referred to (1). For instance, summarized from the of this model, the compensation values of gas fow and air-A. YAN et al.1J Control TheoryAppl 2008 6(1) 80-85fuel ratio, are employed to adjust the gas flow and the air ture and parameters of the neural network will be preservedflow respectively to inhibit the potential faults.so as to be applied on-line to produce the compensation val-2) Neural network-based compensation model The ues of both heating gas fow and ar-fuel ratio.major effect factors to the flow of both heating gas and3) Soft-sensing model for furnace temperature Be-heating air composed by combustion condition F, bound- cause of the inconvenience of fxing any instruments to theary condition B and information of production line I are proposed point, the fumace temperature cannot be measureddetermined using the method of principal components anal- continuously, which brings adverse effects to the timelyysis (PCA) [14] along with the accumulated technical ex- control of the furmace temperature. This problem, however,periences. Then, these three variables are applied to the can be solved by the soft- sensing model for furmace temper-project of generating the compensation value of gas fow ature. During its establishment, (1) is rewitten as follows:Un and that of air-fuel ratio K. As both chemical reactionsT = f(u,v,p,h,n,X*).(7)and physical movements may happen during the combus-With the help of [15], which suggests a method of estab-tion process, the correction of the heating gas and the air-lishing the hybrid itelligent hybrid prediction model, thefuel ratio is a complex process in which serious nonlinearsoft-sensing from input date set {u,U,p,h,n,X*} to theproblems exist. We describe the correction values with thetemperature T is realized with the combination of the neu-following nonlinear system equation:ral network, case-based reasoning and expert system. Also,y= g(F,B,I),(3) the satistical value X* of the furnace temperature measuredmanually contributes to the self adaptation adjustment of thewherey=| and 9() denotes a nonlinear function.soft-sensing model with the purpose of ensuring the accu-Equation (3) is approached with Radial Basis Functionracy of the sof-sensing model. To be detailed, the adjust-(RBF) neural network, taking its advantages of fast conver- ment process is as such: Let the original prediction value ofgence speed and strong approaching ability.the furmace temperature be t, and the statistical value of theThe input set of the RBF netwok is epessed asx = manual measure be x*, then the eror of the output of the[F, B, I]T, the number of the nodes in the hidden layer issofl-sensing model can be given asm, and the output valuables are the cormpensation values ofe=T-X*.8)heaing gas flow un and air-fuel ratio K, which are calcu- Thereby, the corrected temperature islated by the following equations:T=T-e.(9)Un =v1o+ g usi1.G(|]X -1lI),(4)4) Fuzzy control arithmetic The fuzzy control algo台rithm is established with some fuzzy control tables whichK =w20+ 2 wj2. G(|IX - tI),(5) were built on the basis of the number of input and outputvariables, the membership function and the control rules.where Wso∈R (i = 1,2) stands for the offset value,Therefore, the control process under manual operation canWji∈R(j = 1,2... ,m) represents the weight from the be conveted to the form that is available to be accepled andhidden layer to the output layer, G() is the radial basis func-processed by the computer. The control rules of the fuzzytion, |I . I| denotes the Euclidean distance,tj∈R3 meanscontoller in Fig.2 adopt the following forms:the center of the radial basis function.IF {E=A;and dE= B} THEN Us=Ci, (10) .The radial basis function is selected as the followingwhere E is the language variables orresponding to the er-Gauss function:ror e3, dE represents the language variables correspond-G(x) = exp(-(6)ing to the change rate of the error一, Us means the lan-where B is a real constant number that determines the shape guage variables corresponding to the control variable us,of the Gauss function.A(i= 1,2,... ,n) refers to the language values of E, B:The initial values for the RBF network's training is de- denotes language values of dE, C; represents the languagefined as: the centre of the hidden t;(0) = [0.3,0.3,0.5]; values of Us.output weight wj;(0) = 0.1; and the width of the hiddenThe error E, the varying ratio of the error dE and the con-function g = [0.5,0.5,0.5]. The number of the samples trol variable Us have the same universe of [-3, 3], where thefor taining is 130 in this papr, while the number of hdden elements from“-3" to“3" crrerond to the center pointsfunctions being m = 11. The procedure of the training will of seve中国煤化工), PS, PM, and PB.adjust the weight, the hidden function's center and width. Take tMHCNMH G:scription of all theAlso, the training algrithm suggested in [13] is adopted membehere. When the error of the output of the network becomesless than 10-4, the training will be ceased, where the struc-u(E)= exp1-(1.674)4],(11)84A. YAN et al.1J Control TheoryAppl 2008 6(1) 80-85where a represents the center point of each fuzzy set.programmed under the circumstance of configuration soft-Then, the fuzzy relationship of the control rules is cal- ware ControlLogix5000, while the senior algorithms suchculated by R; = (A; x B;) x C. Fnall, the output as those in the working conditions model and in soft- sensingUs = (A.x B) x R, crresponding to the input, is ob- model are implemented with language VB inseted in thetainedbyR= R1V R2 V R3V...V Rn.monitoring software RSView32.5) PID control algorithm The flow of heating gas is ad-The fault ratio (i.e.. the proportion of fault duration timejusted through the opening percentage of the valves, while to the total working time) of the combustion process afterthe flow of heating air is adjused through the frequency of the working conditions diagnosis model being launched isthe transducer. The opening percentage of the valves and obviously decreased, with about 50% additional workingsthe frequency of the transducer are calculated by contoller of the devices. The proposed method has guaranteed thePID1 and controller PID2, respectively, according to the fol- working security to a great extent.lowing PI arithmetic:The efect of soft-sensing and temperature control for theu(k) = u4(k- 1)P(e:(k) -e(k- 1)) + le:(k),furmace may be seen in Fig.3 and Fig.4, which iluminatesi=1,2,(12)the response curves of temperature variation within an 8-hour working shift. Fig. 3 describes the temperature varia-where P and I are the proportional and itegral coficients tion coroled by the rouine method without the compen-of the control loop respectively, which are calculated by thesation model for gas and air-fuel ratio and does not adopt amethod of self-tuning of the regulator based on phase andfuzzy controller but merely uses a PID controller. Comparedampliunde margin specifications [16]. Combined with ex- with the set-point T1(1 100C), the soft- sensing value T andperiences, it may achieve the stable tracking control of the the manual measured value x*, the control eror is higherfow of heating gas wih sected R = 670 and I = 0.037，than 6% of the emperature stpint Crespondingly, theand to that of the heating air with chosen Pr = 590 andtemperature variation curve in Fig.4 is generated by the in-I2= 0.033.elligent control method. Considering Ti (1100°C), T, andSummarily, a typical realizaion of the hybrid inelligent x*, we can find that the control error is lower than 4% ofcontrol algorithm for the combustion process can be de- the temperature set-point.scribed as follows:Fig.3 and Fig.4 indicate that the soft- sensing value ofStep 1 The infomation of working conditions diagno- the furmace temperature can track the trend of the stisticalsis F is atained by (2);value of the manual and hourly measurement. The analysisStep2 The compensation values of heating gas un and results of these data indicate that the errors produced by thethe air-fuel ratio K are gained through the mapping rela- sof-sensing model are less than 1%.Therefore, the outputstionships determined in (4) and (5);of the soft-sensing model for the furmace temperature mayStep 3 The soft-sensing value of the furnace tempera- represent its real value.ture T is calculated by (7);Step4 Based on the error between the set point valueonltTT and the soft sensing value T of the furmace temperature,100the original set-point of heating gas flow ug is given by (10);Step 5 Afier the set- point of heating gas fow is ascer-1050tained as ug + un, the input of the flow control of the heatinggas u1 is calculated by (12) to realize the stabilization of the100023-45678flow control of the heating gas;t/hStep6 The setpoint of heating air flow issetby K Xu1.Fig. 3 Furmace temperature trend under routine control.followed by a necessary transformation. Then the input of1150the flow control of the heating air u2 is also calculated by(12) to realize the stabilization of the fow control of the100个heating air.1050 |4 Industrial applicationThe largest hematite mineral processing factory in China345678has 22 shaft ore-roasting furnaces. The practical effect of中国煤化工tlieant cntrol.the control method proposed in this paper is examined byits application in this factory. Within the approach's re-The IMYHC N M H Ghe proposed controlalization, its selected control system is the ControlLogix method has validated its eficiency: the accuracy ratio of theof Rockwell, the arithmetic of fuzzy control and PID are fault diagnosis is kept consistently around 90%; the error ofA. YANet al. /J Control Theory Appl 2008 6(1) 80-8585the soft-sensing model for the furmace temperature is lessdiagnosis in electric drives using machine learning[]. IEEE/ASMEthan 1%; the furmace temperature can be confined to a closerange around its set-point with a control rror below 4%;the [1] H. Zhang. X Wang. C. Zhang, eti al Soft sensor echipe usingcombustion efciency has been enhanced for the compensa-LS-SVM and standard SVM[CV/Proceedings of 2005 IntemationalConference on Information Acquisition. Piscataway, New Jersey:tion of heating gas and air-fuel ratio; the gas consumption isIEEE Press, 2005: 124- 127.decreased by about 6%, which can meet the requirement of[12] A. Yan, T. Chai, H. Yue. Multivariable nelligeat optimizing controlcontrol accuracy and the technical indices.approach for shaft furmace roasting process[J]. Acta AutomaticaSinica, 2006. 32(4);: 636- 640.5 Conclusions{13] A. Yan, F. Wu, T. Chai. Fault diagnosis expert system using neuralnetworks for roasting procssCVIProceedings of the 1l6th IFACThis paper presents a hybrid intelligent control methodWorld Congress. Prague: Elsevier Press, 2005.for the combustion process of the shaft ore-roasting fumace.[14] D. Lin. Facial expession casication using PCA and hierarchicalThis method integrates a temperature sofi-sensor imbeddedradial basis function network[] Jourmal of Information Science andcascade fuzzy control system with an air flow ratio controlEnineering, 2002(): 1033 - 1046.system, which contains a compensation model for heating [15] A. Yan, T. Chai. Ineligeat hybrid prediction method of magnetic tubegas fow and the air-fuel ratio. It has solved the problem ofrecovery rate[J]. Information and Control, 2005, 34(6): 759- 764.how the furnace temperature can be steadily controlled dur- [16] T. Chai, G Zhang. Sef-uning of PID rgulaors bascd on phase anding the combustion process. As its components, the workingamplitude margin specifications[J]. Acta Automatica Sinica, 1997,23(2): 167- 172.conditions diagnosis model has reduced the fault ratios andguaranteed the safety production of the combustion process,whereas the neural network based compensation model hasAjun YAN was bom in 1970. He received thePh.D. degree from Northeastem University inrealized the timely correction of the air-fuel ratio, which im-2006. He is now a lecturer in the College ofproved the combustion efficiency and economized the en-Electronic Informnation and Control Engineer-ergy consumption. With respect to the reliability and robust-ing, Beijing University of Technology. His他ness, the proposed method is superior to the conventionalsearch interests include modeling and ielelones. It has improved the control accuracy of the fumnacegent control of complex industrial processes,temperature and created a new way for the secure, stablesoft sensing and fault diagnosis. E-mai: yanai-jun@bjut.cdu.cn.and economical working of the combustion process. Theproposed method has a high potential of being further ap-plied to complex industries.Tianyou CHAI was bom in 1947. He receivedthe Ph.D. degree from Northeastem Univer-Referencessity in 1985. He is now a professor in the{1] X. Zhang. Z. Chai. Development and production practice of 100m3Research Center of Automation, Northeastemshaft furnacs in Jiuquan Steel Col[]. Metal Mine, 200(3):0) 32 -33.University, Shenyang, and is an academician[2] s.C. K. Shiu,J. N. K. Liu, D. s. Yeung. Formal description andof Chinese Academy of Engineering. His他e-verification of hybrid rul/frame-based expert systems[]. Expertsearch interests include mulivariable ieleli.entSystems with Applications, 1997, 13(3): 215 - 230.decoupling control, optimizing control and intc-[3] Y. Lei, Y. Peng, x. Ruan. Applying case based reasoning tograted automation of process industry. E-mail:cold forging process planning[J]. Joumal of Matrials Processing tychai@mail.neu.edu.cn.Technology, 2001, 112(1): 12- 16.[4] P. J. Costa Branco, J. A. Dente. Fuzzy systems modeling inFenghua WU was borm in 1977. She is a Ph.D.practicel[]. Fuzy Sets and Systems, 2001, 121(1): 73-93.candidate at the Research Center of Automa-[5] G. Gnanam, s. R. Habibi, R. T Burton, et al. Neural networktion, Northeasterm University. Her resarch in-control of air-to-fuel ratio in a bi-fuel engine[J]. IEEE Transactionsterests include modeling and itelligent con-of Systems, Man and Cybermetics-Par C, 2006. 36(5): 656- 667.trol of complex industrial processes. E-mail:6] C. A. Hsuan, R. Chen. Inelligeat control of exit temperature in awfh. neu@ 126.com.gas-fuel can-type combustor{J]. Engineering Applications ofArijficialInelligence, 2002, 15(5): 391 - 400.[7] W. Wang, H. Li, J. Zhang. A hybrid approach for superisorycontrol of furnace temperature[J]. Control Engineering Practice,2003, 11(11): 1325- 1334.Pu WANG was bom in 1962. He received the[8] C. Yang, M. Wu, D. Shen, et al. Hybrid nelligeant control of gasPh.D. degree from University of Science andcollectors of coke ovens[]. Control Engineering Practice, 2001, 9(7):中国煤化工、1988. He is now a pro-725 -733.liversity of Technology.[9] A. Andrasik, A. Meszaros, s. F. De Azeveine. On-line tuning of aMHC N M H Glude automatic cntro,ncural PID cotrloler based on plant hybrid modelingU]. ComputersCuupuucicm and field-bus controland Chemical Engineering, 2004, 28(8): 1499 - 1509.system. E-mail: wangpu@bjut.edu.cn.[10] Y. Murphey, M. A. Masrur, 乙Chen, et al. Model-based fault