Model-based slopping warning in the LD steel converter process

Model-based slopping warning in the LD steel converter processMagnus Evestedt *,Alexander MedvedevDepartment of Information Technology,Uppsala University,Box 337,75105Uppsala,Swedena r t i c l e i n f o Article history:Received 3December 2007Received in revised form 6January 2009Accepted 8January 2009Keywords:ModellingAdaptive filters Soft sensors Metals industry Steela b s t r a c tThe most prevalent steel-making process is the basic oxygen steel-making (BOS)process.Problems arise when the layer of foaming slag created on the surface of the molten metal exceeds the height of the vessel and overflows,causing metal loss,process disruption and environmental pollution.This phenomenon is commonly referred to as slopping.A method for automatic slopping detection is described in this contri-bution.The sound signal from a microphone located in the off-gas funnel is processed to obtain an esti-mate of the slag level in the converter.A model describing the relationship between off-gas flow rate,pressure and the slag level estimate is updated recursively in time.The output error is fed to a change detector yielding a warning system with three alarm levels indicating the persistence of slopping symp-toms.The algorithm was tested on data from 100heats at SSAB Oxelösund.Slopping was correctly detected in 80%of the blows.Ó2009Elsevier Ltd.All rights reserved.1.IntroductionBasic oxygen steelmaking (BOS)is the most common steelmak-ing process accounting for over 60%of the total steel production,[1].A basic oxygen furnace,also known as Linz-Donawitz (LD)con-verter,is normally run by an operator monitoring the blow pro-gress from a control room.Most of the available control variables (oxygen flow,lance height,bottom gas flow,etc.)are scheduled according to pre-defined programs that can be overrun by the operator in order to handle unexpected process deviations and dis-turbances,if needed.Automatic control of the BOS process is diffi-cult,due to the complexity of the underlying chemistry and physics.Besides,the key quality parameters such as carbon con-tent and temperature of the melt are not measured on-line,which altogether makes feedback control design very hard.The problem of on-line molten metal analysis estimation is addressed in [2,3].In [4]a study with the aim at model-based closed-loop control of the BOS process is presented.The purpose of developing automatic control systems is both to assure cost-efficient production with high quality and to minimize the environmental impact of the process.The foaming slag created on the surface of the molten metal during the blow is necessary for effective steel refinement but also poses the risk of uncontrolled overflow,so-called slopping.Preven-tion,prediction and mitigation of slopping is a long-standing prob-lem in steel technology,frequently addressed but yet unsolved in practice,[5].One reason to it is that slopping is widely agreed tobe a complex and even unpredictable phenomenon.Another one is the lack of reliable and relevant on-line measurements which is a general problem in converter processes.In contrast with most publications devoted to slopping predic-tion,this work suggests a slopping warning method that relies on conventional inexpensive instrumentation and takes advantage of several levels of advanced signal processing.Firstly,the acoustic emission signal from the furnace is processed by means of a mod-el-based method previously validated in a converter water model,[6,7].Secondly,this approach involves the fusion of different kinds of process information to achieve reliable and robust slopping detection,an idea originally suggested ten years ago in [4].Thirdly,the decision on whether or not a slopping warning should be is-sued is taken by a change detection algorithm based on adaptive filtering,[8].The paper is organized as follows.First,the LD converter pro-cess is described and the slopping detection problem is formulated.The employed algorithms for slopping warning are then summa-rized followed by a detailed analysis of the warning system perfor-mance on data from 100heats.2.Process descriptionThe LD converter,basically composed of a vessel for the hot me-tal and a water-cooled oxygen lance,is illustrated in Fig.1.Filling the vessel with material is called charging.The BOS vessel is charged to one-fifth with steel scrap and then molten iron from the blast furnace is added to obtain the proper heat balance.Slag forming agents,such as burnt lime or dolomite,are loaded into the converter to form slag which absorbs impurities of the melt.0959-1524/$-see front matter Ó2009Elsevier Ltd.All rights reserved.doi:10.1016/j.jprocont.2009.01.002*Corresponding author.E-mail address:Magnus.Evestedt@prevas.se (M.Evestedt).Journal of Process Control 19(2009)1000–1010Contents lists available at ScienceDirectJournal of Process Controlj o u r n a l ho m e p a g e :w w w.e l s e v i e r.c o m /l o c a t e /j p r o c o ntPure oxygen is blown through the lance,at supersonic speed,hit-ting the metal bath from above and causing the temperature to rise to about1700degrC.The scrap melts and metal components such as iron(Fe),silicon(Si),manganese(Mn)and carbon(C)are oxi-dized during the blow,lowering the carbon content and removing contaminating chemical elements.The metal in the furnace forms an emulsion with the created slag giving rise to a foam.This facili-tates the refining process due to the large contact area between the chemical elements in the foaming slag and improves thermal con-ditions in the furnace.The slag formation can be controlled by the operator through chemical additives,changes of lance position,as well as manipulation of oxygenflow through the nozzle and bottom gasflow rate.Unfortunately,the operator lacks visual information on slag conditions in the vessel which makes it difficult to ade-quately react to swiftly developing foaming.After10–20min into the blow,the carbon content of the molten metal is expected to have reached its target percentage.The tem-perature is measured and samples are taken to analyze the chem-istry of the steel melt.If the analysis is approved,the steel is poured into a ladle and further refined later in the steel production process chain.2.1.MeasurementsKey process variables are stored in real-time to keep track of the blow progress.Besides,the data logging is used to track back pos-sible malfunctions when something goes wrong in the process.The information data-base is partially presented in Fig.1.Different types of sensors are used to provide the measurement signals and,thus,the data-base entries are not fully synchronized due to communication delays.In Table1some of the available measure-ments are characterized.Rich and diverse information about pro-cess conditions is represented in the data-base but none of the measurements is directly related to the key process parameters.2.2.The slopping problemTo make the LD converter process effective,a significant slag volume is needed in the vessel.However,the slag volume is bounded by the limited size of the converter vessel.If the slag level is too high,slopping will occur,resulting in severe dust emissions and reduction of the metal yield.Furthermore,the steel production must be stopped to clean the area below the vessel and the vessel mouth.The slopping phenomenon is complex,dynamic and dependent on many process variables.A list of contributing factors was pre-sented in[5]and includedSlag viscosity.Slag surface tension.Slag density.Size of the gas bubbles generated in the decarburization process. Vessel working lining height,volume and shape.Lance height above the bath.Oxygenflow rate through the lance.Lance hole wear.Chemistry of the hot metal and the scrap.Decarburization speed.The bare length of the above list explains the common belief that the slopping phenomenon is chaotic and unpredictable.The persistence of slopping problems in many steel mills have given rise to a search for ways to maintain a suitable foaming slag vol-ume while preventing slopping from occurring.Unfortunately,this has,over the last decades,proved to be a rather challenging task.The efforts to develop a system for slopping warning and miti-gation have been categorized into three major groups:(1)Modelling of slopping and its potential to occur.(2)Measurement devices for slag level that detect the onset ofslopping.(3)Mitigation measures undertaken in real-time to prevent thedevelopment of full-blown slopping events.An example of thefirst group is the results presented in[9], where an optimal blow profile was calculated based on the initial composition of the melt.A similar system based on a calculation of the slopping potential for each heat was developed in[10].Systems based on slopping potential(slopping factor)perform wellunderTable1A selection of the available process information data from the data-base.Measurement Location Obtained process information DelayOff-gas analysis In the off-gas system Percentage of CO and CO2plus the measuredoff-gasflow rate.15–25sSonic-meter In the hood above the converter mouth The sound level indicates changes in the slag level None Oxygenflow meter In the lance Gives the actual oxygenflow rate through the lance None Lance height In the lance control system Gives the lance position as calculated by the lancesystemNoneCamera surveillance/image analysis Below the LD converter and monitors theamount of falling slag to quantify sloppingNoneM.Evestedt,A.Medvedev/Journal of Process Control19(2009)1000–10101001steady production conditions with low variations in the quality of raw materials and product assortment.A microwave gauge was utilized in[11]to measure the slag surface level relative to the vessel mouth.Radio waves were em-ployed for slag depth measurements in[12].The main difficulty with this kind of approaches is their low accuracy and reliability under top blowing with high bath agitation.A detection system involving oxygen lance vibration measurements was used in[13] and[5],where a single on-line secondary information source is utilized to obtain an estimate of the slag level in the converter. The quality of the estimate depends very much on the choice of the measured vibration frequencies which are known to vary depending on the process conditions.These are all examples of the second group.The third group includes a new and promising direction of re-search,namely to combine on-line measurement devices for early slopping detection and use it for initiating process interventions for slopping mitigation.Extensive studies of slag forming have been performed to this end.To gain insight,different kinds of empirical equations describing the change in foam height have been suggested in the past,[14,15].A model with a physical back-ground is derived in[16]using results of cold and hot model exper-iments.The area of dynamic modelling of slag foaming is takeneven further by the results in[17].On the basis of a physical model, a system for control of dynamic foaming is developed in[6].A water model of the LD converter process is used to validate the re-sults.The approach is further refined in[7].In[4],slopping is de-tected by a combination of the sonic-meter and gas analysis. Another system utilizing the idea of combining several measure-ment sources was presented in[8],where adaptivefiltering and change detection algorithms are utilized to construct an on-line alarm system providing warnings to the operator.This paper presents a slopping alarm system based on the prin-ciple of change detection and reports results of its performance evaluation infield tests at SSAB Oxelösund.The evaluation is based on data from a large number of regular production heats and also highlights some issues pertaining to system validation and opera-tor practice.2.3.Camera for slopping quantificationFor evaluation purposes,an objective way of quantifying slop-ping is preferable.In[4]a person with a stop watch noted the times for slopping observations during the blow.In[9],VCR cam-eras were employed while an attempt to use IR camera devices were unsuccessful due to software issues.For the experiments de-scribed in the following,a camera system was implemented on-site to monitor the process.The resolution of the monochrome CCD camera was640Â480pixels and the frame rate was set to 15Hz.The position of the camera is shown in Fig.1.When slop-ping occurs,molten metal will fall from the top of the converter onto thefloor below the vessel.The camera position makes it pos-sible to capture images of the falling slag,see Fig.3.A histogram depicting the distribution of image pixel values between0and 255is shown in Fig.2.The choice of brightness constant or thresh-old on the image gray level is facilitated by the large difference in brightness between the falling slag and the background.Therefore the image can be easily segmented into parts where the pixel val-ues are above the threshold and parts with brightness below it, effectively separating the molten metal from the darker back-ground.The segmentation result of the image in Fig.3is depicted in Fig.4.The ratio between bright and dark image pixels gives an indica-tion of how severe the slopping event is.This ratio is averaged over a sampling period of2s and stored in real-time along with other process data in the data-base.2.4.Acoustic signals for slopping detectionSince the beginning of the1970s,a device called sonic-meter has been employed in many steel plants for indirect monitoring of slag foam level,[18].The basic idea is that as the foam level increases, the sound emission from the vessel under blowing decreases at certain frequency bands.The main source of sound is at the lance nozzle where the oxygen emerges at a speed above the speed of sound and impinges onto the molten metal surface.The sonic-me-ter signal is usually used by the operator for monitoring slag level changes but it has also been employed as a controller input in[19–21].Note that all those systems are not really dynamic feedback control systems since they use very low control signal sampling rates in order to avoid stability problems.As an alternative to commercially available sonic-meters whose exact signal processing algorithms are proprietary,the raw signal from the sonic-meter microphone can be utilized for slag level esti-mation.According to[6],the relationship between the foam height h and the sound intensity I at a time instant t and frequency x is exponential and given byIðhðtÞ;xÞ¼I0eÀb FðxÞhðtÞð1Þwhere I0denotes the sound intensity when no slag is present.The frequency dependent parameter b FðxÞcharacterizes bubble sizeand material.By applying the logarithm to both sides of Eq.(1), the following equation is obtainedhðtÞ¼lnðI0ÞÀlnðIðhðtÞ;xÞÞb FðxÞð2ÞThis means that the foam height depends logarithmically on the sound intensity at any given frequency.Since the bubble size and material content are unknown in the converter,a model for the foam dynamics is difficult or impossible to obtain analytically.Instead,the sound attenuation coefficient b FðxÞis estimated from experiments and used to produce a slag le-vel estimate^hðtÞ¼lnðI0ÞÀlnðIðhðtÞ;xÞÞ^bFðxÞð3ÞThe estimate can be in principle evaluated at any frequency x,and estimates at different frequencies can be combined to achieve higher accuracy.Frequencies at which the sound intensity is not1002M.Evestedt,A.Medvedev/Journal of Process Control19(2009)1000–1010attenuated as well as frequencies that mainly contain noise should be neglected.This implies using a weighted least-squares algorithmfor the slag level estimate ^hðt Þ,with zero weight on frequencies with low sensitivity to foam height changes.This approach has shown high estimation accuracy in water model experiments [7].Notice here that the informative frequencies are plant-specific and depend e.g.on the microphone position,vessel and nozzle geometry and have to be evaluated individually for each particular installation.In this study the frequencies for the slag level estimate have been selected from a frequency-wise comparison of acoustic signal intensity in the beginning of the blow (no slag)and at slop-ping instances (slag overflow),when the slag level reaches its maximum.2.5.Operator practices regarding sloppingThe LD converter process is run by the human operator but with significant help of pre-defined and product-dependent programs for lance level,oxygen flow,bottom purge gas and so far.These programs can be overrun by the operator under process disrup-tions such as slopping.The operator rates slopping events on a scale from 1to 3where 1stands for no slopping,2for slopping and 3for severe slopping.However,the rating is subjective,often made under stress or before the event is full-blown,and without direct visual contact with the vessel.Practices regarding slopping prevention and mitigation vary with the operator and are very much dependent on his/herprevi-Fig.3.Example of image taken by the camera for slopping alarm validation below the LD convertervessel.Fig.4.Segmented image for slopping quantification.M.Evestedt,A.Medvedev /Journal of Process Control 19(2009)1000–10101003ous process experience.As a consequence,the data collected in this study reflect not only scheduled changes in process variables and those caused by slopping,but also changes initiated by the opera-tor to adjust the process state.This must be kept in mind when evaluating a warning system.3.System designModel-based signal processing is chosen as the main tool of extracting information about slopping from available process sig-nals.The resulting system is hierarchical with three subsequent levels.Level1:Processing of acoustic signal.The two presented earlier in Section2.4options are considered for implementation.One can either use the sonic-meter signal or directly access the raw microphone signal.In the latter case,the weighted least-squares estimate suggested in[7]and based on Eq.(3)is used for slag level estimation.The results of an experimental compar-ison between these two alternatives are summarized in Section3.4.Level2:Adaptivefiltering is utilized for the fusion of the acoustic signal with other sensor information on slopping symptoms.A significant number of different model structures for adaptivefil-tering have been considered but only two are described in detail in the next section to save space.One of them has been consid-ered before in the literature and another one has performed best in this study.The choice of recursive estimation technique for the adaptivefilter is also a design degree of freedom as dis-cussed in Section3.2.Level3:Change detection is employed to formalize the evalua-tion of slopping symptoms and alarm generation from the resid-uals of the adaptivefilter.3.1.Process modelDue to the complexity of the LD converter process,an analytic model relating acoustical emission from the vessel to the process chemistry is difficult to obtain.A system identification approach is used instead where a model isfitted to the recorded data by assigning suitable numerical values to its parameters.The choice of process model is highly affected by the available process signals and their quality.Two ways of choosing the process model have been reported and are summarized in Table2.Model A corresponds to the initial approach in[8],with the in-put to the process model chosen,as in[4],to be the off-gasflow rate,u off-gas,and the CO content in the gas,u CO.The output signal, m,is the natural logarithm of the sonic-meter signal.As a second approach,suggested in[22],the sonic-meter signal was exchanged for an estimate of the slag level in the converter,based on Eq.(3). Due to difficulties stemming from time delays in the off-gas anal-ysis the CO content,input signal is changed in favor of a pressure measurement,u pressure.The resulting process model is denoted as Model B.According to the preliminary results in[4],the slopping phe-nomenon is captured to a large degree by a fourth order dynamic model.Two ARX structures with different input and output signals are shown to be favorable in[8]and in[22].Both models can be written as linear regression in the unknown parameter vector h yðtÞ¼u TðtÞhþeðtÞð4Þwhere yðtÞis the scalar output measured at discrete time instances t¼0;1;...and the regressor vector contains previous input and output values.The corresponding regressor vectors as well as out-put signals are specified in Table2.The scalar eðtÞis disturbance which is assumed to be white for the purpose of change detection. As described in Section 3.3,this assumption is validated in experiments.To be able to track time variations in the model parameters due to changing process conditions,a recursive parameter estimation algorithm will be used and described in the next section.3.2.AdaptivefilteringTo turn regressor model Eq.(4)into an adaptivefilter,its parameter vector h has to be estimated by a recursive identification algorithm.An attempt to employ the recursive least squares algo-rithm[23]for estimation of h from process data faced difficulties in discriminating between the effects of model uncertainty and actual process changes in the residual sequence.A more general approach involves the Kalmanfilter.Assume that eðtÞis white and the parameter vector is subject to the random walk model driven by a zero-mean white noise se-quence wðtÞhðtÞ¼hðtÀ1ÞþwðtÞð5ÞThen the optimal,in the sense of minimum of the a posteriori parameter error covariance matrix,estimate is yielded by^hðtÞ¼^hðtÀ1ÞþKðtÞðyðtÞÀu TðtÞ^hðtÀ1ÞÞð6Þwith the Kalman gainKðtÞ¼PðtÞuðtÞð7Þwhere PðtÞ;t¼1;2;...is the solution to the Riccati equationPðtÞ¼PðtÀ1ÞÀPðtÀ1ÞuðtÞu TðtÞPðtÀ1ÞrðtÞþu TðtÞPðtÀ1ÞuðtÞþQðtÞð8ÞThe initial condition Pð0Þ¼P Tð0Þ;Pð0ÞP0describes the covari-ance of the initial guess^hð0Þ.Optimality of the estimate in Eq.(6)is guaranteed(see[24])only whenQðtÞ¼cov wðtÞ;rðtÞ¼var eðtÞð9ÞSince the above quantities are seldom a priori known,they are usu-ally treated as design parameters of the estimation algorithm and chosen as some QðÁÞ2R nÂn;QðÁÞP0,rðÁÞ>0in order to achieve de-sired properties of thefilter.A well-known practical complication occurs in the Riccati equa-tion of the Kalmanfilter when the regressor vector uðtÞdoes not cover all the directions in Im QðtÞ,e.g.the persistent excitation condition is not fulfilled.The phenomenon is usually referred to as covariance windup(or blow-up)and causes some eigenvalues of P in Eq.(8)to grow linearly with time.In the LD converter pro-cess,data are collected during normal operation and cannot beTable2Models for adaptivefiltering.Model A Model BInputs u CO;u off-gas u off-gas;u pressureOutput y m^hRegressor u TðtÞ½yðtÀ1ÞyðtÀ2Þu off-gasðtÀ1Þu COðtÀ1Þ ½yðtÀ1ÞyðtÀ2Þu off-gasðtÀ1Þu pressureðtÀ1Þ 1004M.Evestedt,A.Medvedev/Journal of Process Control19(2009)1000–1010guaranteed to be persistently exciting.It is therefore motivated to employ parameter estimators with strong anti-windup properties.Most anti-windup schemes for Kalmanfilter parameter estima-tion are of ad hoc nature and are not proved to be non-divergent under lack of excitation,see e.g.[25,26].In the approach taken in [27],a special choice of QðtÞis used to control the stationary point of the Riccati equationQðtÞ¼P d uðtÞu TðtÞP du d uð10Þwhere P d2R nÂn;P d>0.Consequently,the optimality of the Kalmanfilter is lost.Riccati equation(8),with the free term chosen according to Eq.(10),is called the Stenlund–Gustafsson(or just SG)algorithm.Thematrix P d is a stationary point and as well the tuning parameter ofthe method.A formal proof of the fact that the algorithm is non-diverging,even for the case of lack of excitation,is given in[28].The stationary behavior is investigated in[29],where it is shownthat if PðtÞhas converged to P d,it will stay there even if the excita-tion disappears in some directions.The SG algorithm was studiedfurther in[30],where an acoustic echo cancelation applicationwas used to compare its performance to that of a number ofwell-known techniques such as the normalized least-squares algo-rithm and the Kalmanfilter.3.3.Slopping warning by change detectorDuring normal operation,the LD converter process model Eq.(4)is assumed to apply with slowly changing parameters,h.Thetime variation is reasonably well explained by the rising moltenmetal temperature in the converter during the blow.When slop-ping occurs,the model parameters change abruptly due to a signif-icant alteration in the chemical reactions.Since the sloppingphenomenon manifests itself as a drastic deviation from the pro-cess’regular dynamic pattern,a change detector[31]acting onthe model residual can recognize slopping events in the LD con-verter.A warning system for slopping detection can be designedby combining the process model,recursive parameter estimationand the change detector as shown in Fig.5.The detector in Fig.5is fed by the residual ðtÞ¼yðtÞÀu TðtÞ^hðtÀ1Þof the adaptivefil-ter.Based on the residual,a distance measure,sðtÞ,is calculated.Assuming that all noises are Gaussian,the true system is time-invariant and belongs to the linear regression model set,yieldsðtÞ2Nð0;SðtÞÞwhere SðtÞ¼rðtÞþu TðtÞPðtÞuðtÞand the matrixPðtÞis the solution of Eq.(8).A whiteness test based on the auto-correlation of the residual sequence for one of the studied chargeswithout slopping shows this to be a valid assumption.Only1.75%(14)of the autocorrelation coefficients at lags between one and400lie outside of the95%confidence interval.It also demonstratesthat the applied model is adequately chosen and the parameterestimation algorithm works well.Fig.7.The graphical user interface of the warning system contains information about the blowing parameters and generated alarms.The depicted situation corresponds to a completed heat.(1)Slag level estimate;(2)camera signal;(3)alarm history with specific alarm times;(4)oxygenflow rate;(5)lance height;(6)current alarm grade(color coded);(7)start button;(8)alarm acknowledgment button;and(9)stop button.M.Evestedt,A.Medvedev/Journal of Process Control19(2009)1000–10101005A suitable choice of distance measure enabling the detection of both variance and parameter changes is given by the normalized squared residual s ðt Þ¼S À1ðt Þ 2ðt Þ,[31].The distance measure is averaged to get a test statistic g ðt Þ,that is then thresholded to trig-ger an alarm when g ðt Þ>h .Such a test statistic is for example gi-ven by the one-sided CUmulative SUM (CUSUM)detector,[31].It is defined byg ðt Þ¼max ðg ðt À1Þþs ðt ÞÀm ;0Þif g ðt Þ>h ;then alarm and g ðt Þ¼0where h is the alarm threshold ,s ðt Þis an input distance measure and m is a drift parameter to prevent positive drift of g ðt Þ,eventually yielding a false alarm.The threshold and drift parameter in the CUSUM change detec-tor need to be adjusted to the data from each individual charge in order to obtain desired detection rate with a reasonable number of false alarms.Since this is not feasible in a real-time application,the threshold and drift parameters in this study were carefully selected by running the system on plant data from 50blows,trimming the parameters to achieve acceptable overall performance.This of course decreases the performance of the warning system.A bettersolution to this problem would be to relate the values of detector parameters to the process operation conditions and raw material properties,thereby attaining adaptive selectivity of the alarms.Such an effort is outside of the scope of this paper.The severity of a slopping event is classified by keeping track on how many warnings that are collected by the system subsequently.One warning leads to a green alarm,two to a yellow alarm and if the number of consecutive warnings within a specified time frame exceeds three,the alarm is set to red indicating that the slopping persists over a period of time.The coloration of alarms is mainly introduced for interaction with the operator and the time frame should be chosen so that all warnings issued can be considered to be inflicted by the same slopping occasion.3.4.Sonic-meter versus slag level estimationTo illustrate the advantage of utilizing the slag level estimate for slopping detection instead of the sonic-meter signal,the two model structures in Table 2were evaluated on data from the same heat with known occurrence of slopping.The distance measure s ðt Þis shown in Fig.6,for either case,to illustrate the feasibility of its use in the change detector.To make threshold and drift parameter choices easier,process deviations should be clearly visible in the filter residuals.As can be seen,the distance measure for the slag level estimation based alarm system is twice the distance measure for the sonic-meter based one under the same slopping conditions.This facilitates the choice of CUSUM user parameters and indicates that high resolution microphone data is a better alternative for slopping detection.Thus,Model B is adopted in the following.4.System evaluationThe slopping warning system is implemented at SSAB Oxelös-und for field trials.A graphical user interface (GUI)was designed to provide the operator with relevant information and slopping alarms.The layout of the GUI is shown in Fig.7.Notice that this is a research version of the software intended for evaluationofFig.8.The warning system relies on data communication of measurements from different parts of the process.The blocks within the dashed line encircled area belong to the system described in this paper.1)Off-gas flow rate and pressure,2)Slag level estimate and camera signal,3)System alarms.1006M.Evestedt,A.Medvedev /Journal of Process Control 19(2009)1000–1010。