Multi-model Soft Sensor For F-CaO Content In Cement Clinker

Free lime(f-CaO)content in ceme nt c linker is one of t he most important qua lity parameters for ceme nt.However,due to the comp lexit y of t he cement product ion process,suc h as distr ibuted parameters and large time lag,f-CaO conte nt is mostly obtained offline by ma nua l sa mpling and analys is in laboratory.The long samp ling time interva l(typ ica lly one ho ur)and delayed reporting of measured value of f-CaO conte nt ma y result in unt ime ly control of ceme nt clinker qua lit y and large fluct uat ions in cement product ion.Such a sit uat ion cannot meet the urge nt demand of ce ment industr y for high c linker qualit y and product ivit y.It is there fore necessary to deve lop methods for automatic estimation of f-CaO content.Soft sensor technique is an effective approach for rea l-t ime estimat ion of process parameters and is potent ial to solve the proble m of f-CaO content measure me nt.At present,most soft sensors for f-CaO content are based on single models.In fact,ce ment c linker productio n is a comp le x process characterized by diverse working cond it ions and non-linearity,strong coupling,and especia lly large time lag.It would be difficult to achieve predict ion accuracy for var ious process cond it ions when a single model is used.Therefore,in the fra mework of a NSFC project,a mult i-model soft sensor modeling met hod is proposed for estimat ion o f the f-CaO content for a local ce ment productio n line in J iangxi province.The proposed met hod is of theoretica l and practica l importance,and it is a basis for t ime ly control o f ce ment clinker quality.The main contributions of the paper are as follows:(1)To select inputs o f t he f-CaO content soft se nsor model,t he mechanis m o f ce ment c linker product ion process was analyzed.Based on the mechanis m a na lys is and exper ience of the kiln operators,five important process var iables(the feed rate of the raw materia l,the temperature in precalc iner,the electr ical current of the drive motor of the rotary kiln,the secondary air pressure at k iln head and the air pressure in the second chamber of grate cooler)were selected as inputs of the model.(2)Data preprocessing is per formed to obta in t he training sets and test ing sets for the soft sensor model for f-CaO content.The raw process data are processed first ly by mean filter to reduce noise and then by 3? criter io n to delete contaminated data caused by sensor fa ilure.Consider ing t he fact that t here is muc h t ime lag between the process variab les and f-CaO content as we ll as the fact that the clinker process is cont inuo us,the procedures of t ime series matching and we ight ing between inputs and output were taken.After data preprocessing,660 sets of inp uts-output pairs wit h wider range were obtained.(3)To so lve the proble m that a single model cannot well copy with t he varying cond it ions of ceme nt clinker productio n,a mult i-model soft sensor modeling met hod for estimat io n of f-CaO content is proposed.Cons ider ing the fact that t he work ing cond it ions o f ceme nt c linker product ion process are comple x and d iverse,the process data were clustered into several clusters by using fuzzy C-means cluster ing algor ithm.After that,data under s imilar working condit io ns were clustered into the same cluster.Then,several support vector machine(SVM)models were trained by several clusters and the n comb ined based on switching strategy to predict f-CaO content.The first 560 data were used for model tra ining and the last 100 data(collected over 4 successive days dur ing cont inuo us operation o f the rotary kiln)were used for model test ing.The results demonstrate that the proposed model achieves good predict ion perfor ma nce wit h s maller mean square error,The il's ineq uality coefficie nt and larger correlat ion coeffic ient.In addit ion,the proposed model can predict we ll the trend of f-CaO content change.(4)To furt her evaluate the performa nce of t he proposed model,the proposed model was compared wit h ava ilable models that were used for f-CaO predict io n in literature,inc luding single SVM model,Partia l Least Square Regress ion,Radia l Basis Funct ion Neural Network and Takagi-Suge no Fuzzy Neural Network.The results show that the proposed mode l achieves the minimum mean square error,The il's inequa lity coeffic ient a nd the maximum correlat ion coeffic ient and has obvious advantages in three model performance indices.