Intelligent Software Sensors and Process Prediction for Glass Container Forming Processes based on Multivariate Statistical Process Control Techniques

Transcription

1 UKACC Internatonal Conference on Control 20 Cardff, UK, 3-5 September 20 Intellgent Software Sensors and Process Predcton for Glass Contaner Formng Processes based on Multvarate Statstcal Process Control echnques Dean Butler Process Control Ardagh Group, Glass: Doncaster Unted Kngdom E-mal Hongwe Zhang Faculty of Arts, Computng, Engneerng and Scence Sheffeld Hallam Unversty Howard Street, Sheffeld, S1 1WB, UK E-mal Abstract Glass contaner formng processes have attracted more attenton over the past years due to the problem of lackng process nformaton and correlaton for key varables wthn the processes. In ths paper an approach to develop process modelng and ntellgent software sensng s presented for applcaton based on multvarate statstcal process control methods. he ntellgent software sensors are able to provde real tme estmaton of key varables, and Partal Least Squares (PLS) technques have allowed for forward predcton of fnal product qualty varables. An applcaton of software sensors used for contaner formng blank temperature s presented along wth PLS beng appled to predct the wall and base dmensons of glass contaner products. Intal results show that these methods are very promsng n provdng a sgnfcant mprovement wthn ths area whch s usually unmontored and s susceptble to long tme delays between formng and qualty nspecton. Keywords- Glass Contaner Formng; Batch Control, Software Sensors; Condton Montorng; MSPC, PLS, PCA I. INRODUCION he Glass contaner manufacturng ndustry s a huge global sector of package manufacturng whch produces products for varous ndustral sectors such as cosmetcs, pharmaceutcal and the food and beverage ndustry. Glass contaners are produced by the meltng of raw materals manly sand, soda ash and lmestone n a gas or electrc fred furnace at temperatures above 00 C. he glass once molten s then dstrbuted away from the furnace along channels known as fore-hearths whch are at set controlled temperatures to the relevant glass formng machnes, where upon arrval the molten glass s allowed to dran through an orfce. he molten glass s then cut as t leaves the orfce at a precse set rate to allow for enough molten materal for the partcular product beng formed. hs cut molten glass s then delvered to the glass formng machne, typcally n double or trple cavtes. At ths stage hgh capacty, hgh volume producton machnery s used to mould the molten glass nto the varous batches of contaners that one manufacturng faclty may produce. It s ths part of the Glass Contaner Formng Processes whch have been a dffcult challenge to engneers. Dffcultes wth the harsh nature of the envronment and obtanng accurate nformaton relatng to ts process make t etremely challengng. he lack of on-lne sensng for process varables has been a serous obstructon and has left the process to somewhat a black art over the years. Now wth hgher epectatons and demands from customers ths process has never been as mportant to montor and refne. At present almost all control methods and polces appled are based on off lne nformaton for the process operators and supervson, ths compromses qualty as the delays between product formaton and nspecton s so large that abnormaltes can go undetected for a whle. It s also relatvely unknown at ths stage as to the actual lmts and constrants that est on the varables wthn ths part of the process or as to the actual combned contrbuton that each of these varables actually has upon the fnal product beng made. Multvarate Statstcal process control methods based upon lnear projecton have attracted nterest and have been are proven method for producng emprcal models for Industral processes. Prncple Component Analyss (PCA) and Partal Least Squares (PLS) technques have been appled to many practcal regresson problems to estmate qualty related varables. Zhang and Lenno [1] appled these technques to batch fed fermentaton processes wth the focus on the adopton and applcaton of PLS and PCA technques for sensor falure detecton and predcton. Other promsng results were acheved wthn batch fermentaton processes wth varous publcatons made [2][3][4][5], applcatons have also been deployed wthn the steel ndustres and used for advanced montorng of plant functons to determne the relatonshp between process varables and producton qualty [6]. he applcatons of the above-mentoned PLS and PCA technques have proved successful n provdng soft-sensng technques and lnear regresson model predcton for process varables //$ IEEE 281

2 hs paper ams to develop software sensors and also lnear regresson model predcton of glass contaner formng qualty related varables usng PLS technques. Also the paper wll demonstrate the ablty of PCA to provde abnormal condton detecton and solaton wthn glass contaner formng processes. II. SAISICAL MODELLING AND SOF-SENSING USING MULIVARIAE SAISICAL PROCESS CONROL ECHNIQUES A. Prncple Component Analyss Prncple Component analyss (PCA) s a multvarate statstcal method for dentfyng patterns wthn data sets by hghlghtng smlartes and dfferences wthn the data presented. PCA attempts to fnd combnatons of factors or varables that descrbe trends wthn the data, after data assembly PCA mathematcally t s a method of wrtng a matr of X varables of rank R as the sum of R matrces of rank 1 ntally assumng the data are mean cantered [7]. X = M1 + M2 + + M3 + + Mr (1) Each matr wth m rows and n columns, and each varable beng a column and each sample a row PCA decomposes the matr X as the sum of r t and p pars where r s the rank of the matr X X = t + (2) 1p1 + t 2p2 + + t kpk + t rpr he t, p pars are ordered by the amount of varance captured. he t vectors are known as scores and contan nformaton on how the samples relate to each other. he p vectors are known as loadngs and contan nformaton on how the varables relate to each other. Mathematcally, PCA reles upon an egenvector decomposton of the covarance or correlaton matr of the process varables. For a gven data matr X wth m rows and n columns, the covarance matr of X s X X cov(x) = m 1 provded that the columns of X have been mean-cantered" by subtractng off the orgnal mean of each column. In the PCA decomposton, the p vectors are egenvectors of the covarance matr; (3) cov(x)p = λ p (4) where λ s the egenvalue assocated wth the egenvector p. he score vector t s the lnear combnaton of the orgnal X varables defned by p. Another way to look at ths s that the t are the projectons of X onto the p. he t, p pars are arranged n descendng order accordng to the assocated λ. he λ are a measure of the amount of varance descrbed by the t, p par. he frst par captures the largest amount of nformaton n the decomposton and each subsequent par captures the greatest possble amount of varance remanng after subtractng t p from X. B. Partal Least Squares PLS s a system dentfcaton tool that s capable of dentfyng the relatonshps between nput (X) and output () varables. he advantage that ths approach offers over more tradtonal dentfcaton technques, such as ordnary least squares, s that t s able to etract robust models even n applcatons nvolvng large numbers of hghly correlated and nosy process varable measurements. he approach works by selectng factors of cause varables n a sequence that successvely mamzes the eplaned covarance between the cause and effect varables. Gven a matr of cause data, X, and effect data,, a factor of the cause data, t k, and effect data, u k, s evaluated, such that: and np n X = < tk pk + E k=1 np n = < k=1 u q k k + where E and F are resdual matrces, np s the number of nner components that are used n the model and n s the number of causal varables. p k and q k are referred to as loadng vectors. hese equatons are referred to as the outer relatonshps. he vectors t k are mutually orthogonal. hese vectors and u k are selected so as to mamse the covarance between each par, (t k, u k ). Lnear regresson s performed between the t k and the u k vectors to produce the nner relatonshp, such that: u k F k k k (5) (6) = b t + ε (7) where b k s a regresson coeffcent, and ε k refers to the predcton error. he PLS method provdes the potental for a regularsed model through selectng an approprate number of latent varables, u k n the model (np). he number of latent varables s typcally generated through the use of cross valdaton. III. ADVANCED MONIORING ECHNIQUES USING PCA A. PCA Arrangement PCA does not attempt to resolve any relatonshp between nput and output data but t dentfes patterns wthn data, as the glass contaner formng process s constant, by analysng the data present at one condton would gve a sutable set of data to tran a PCA model for dentfcaton of process changes. he data was constructed nto a Matr (X), where each column was a varable and each row a sample of the varables at a specfc pont n tme. 11 X =.... 1(86) (86) n 1 n2.... n(86) 282

3 B. Implementaton Of Advanced Montorng Usng PCA A PCA Model was generated whch conssted of 4 prncple components whch descrbed 95% of the entre varaton wthn the data sets, ths model was then used to test aganst known faulty condtons whch occurred wthn the glass contaner formng process. Fgure 1 shows the mpact of the erroneous data upon the frst two prncple components. It depcts the new loaded data to est well outsde that of known good lmts, ths breach alone could be used to detect process abnormalty. here are other components of further nterest when performng Prncple Component Analyss these are that of the Q resduals and Hotellng 2 charts. he Q resduals gve an ndcaton of the measure of dfference or resdual between a sample of data and ts projecton onto the components retaned n the orgnal model and the Ho tellng 2 contrbutons descrbe how much varaton s wthn each sample to that retaned wthn the model. Fgure 2 shows ths nformaton. Fgure 3 Q Resdual Contrbutons at pont of detecton From Fgure 2, t can be seen that at around sample 570 a dfference between the data n the model and the data now loaded s detected, ths dfference becomes even worse, at around sample 624 where the Q resdual value goes etremely large. What we are actually eperencng here s an ndcaton that somethng s wrong, to what we deem normal operatng condtons for blank temperatures. Further nvestgaton nto the data at sample 624 shows that varables 5 and 6 are the man cause of ths anomaly. Lookng at the data at ths pont ndcates a sensor falure on IS staton 3, varables 5 and 6. he above shows promsng results that PCA technques can be used for dentfyng process abnormaltes. IV. CONSRUCION OF INELLIGEN SOFWARE SENSORS Fgure 1 Frst two PC s loaded aganst Model Data A. PLS Model Development he frst stage n the development of a sutable PLS Model s to obtan relevant tranng and model testng data. In ths applcaton Data was collected from two man process areas, the actual process varables themselves and the fnal product qualty data measured and recorded by the operatves themselves. here was also a further problem ntroduced wthn ths process as there s huge transport lag between the process data recorded and the qualty data obtaned from the operatves, here an emprcal approach was undertaken n order to calculate the desred transport lag and construct the data matr for analyss to begn. Fnally a PLS Model was generated whch contaned 3 latent varables was developed usng ths data. In ths model the followng measurements were used as nput varables (X) 50 Ps I-S Staton operatng Ar, 35 Ps I-S Staton Operatng Ar, Plunger Coolng I-S Staton Ar and IS Staton Blank emperatures, wth Output Varables () of contaner wall thckness measurements and contaner base thckness Measurements. Fgure 2 Q resdual chart of new loaded PCA data 283

4 V. MODEL PREDICION OF GLASS CONAINER WALL AND BASE HICKNESS A. PLS Model Development Usng the data Matrces X, 1 and 2 prevously constructed t s also possble to utlse the obtaned process data to predct product qualty varables. Fgure 4 PLS Model Data Constructon he nput Varables were constructed nto the nput varable matr X and the output varables nto the column vectors. 11 X =.... 1(86) (86) n 1 n2,.... n(86) y1 1 = y. y1(86) 2 y2 = y 22. y2(86) hese data matrces were then splt nto two, half utlsed for model tranng data and the remanng half utlsed for model verfcaton. B. Implementaton of Software Sensors Based on the PLS model generated, a software sensng approach was nvestgated as due to the harsh nature of the envronment t was common to eperence sensor falures, so wth the ablty to predct these varables we would be at an advantage f requred to adopt such soft sensng technques for further appled control loops around ths process. A software sensor was developed to estmate a contaner formng IS staton Blank temperature. he accuracy of ths software sensor s llustrated n Fgure 5, although there s some modellng error present ths model development s a bg step forward and as a bass to provdng feasblty towards such applcatons. he fgure gves good evdence that PLS based model predcton s feasble for 'soft sensng' applcatons on glass contaner formng processes. B. Implementaton of Lnear regresson Modelllng usng PLS fttng In the prevous secton t was demonstrated that PLS technques can be used for software sensng technques wthn glass contaner formng processes, n ths secton t wll be shown that the same technques are able to be appled to predct two mportant glass contaner qualty varables - contaner wall thckness and base thckness. Based upon the PLS model generated, real data measured for contaner base thckness and wall thckness are compared wth what are predcted usng the contaner formng process nput varables, ths s of great beneft as the ablty to predct such varables would remove the transport tmes nvolved before any feedback s generated on glass contaner varables. he PLS Model generated s found to be very good n ths applcaton and has the ablty to predct the contaner wall thckness, the model generated actually does so to wthn 0.5mm. Further tests gave smlar promsng results wth the ablty to predct contaner base thckness as shown n Fgures 6 and 7. Comparng wth the real data collected from producton (sold green lne), PLS Model s found to predct these product qualty varables and shows promsng evdence that these technques and methods are applcable to glass contaner formng processes. Fgure 6 PLS Model Wall hckness Predcton Fgure 5 PLS Blank emperature Predcton 284

5 ACKNOWLEDGMEN he authors of ths paper would lke to thank Ardagh Group for the support and the fundng requred for the hardware and software tools that were vtal for ths project. REFERENCES VI. Fgure 7 PLS Model Base hckness Predcton CONCLUSIONS In ths paper, statstcal models of a glass manufacturng IS machne processes have been produced. Comparsons between the real-world data and that produced from the developed models were produced and gave promsng results towards the applcaton of such technques. Intellgent software sensors were also developed and dscussed. All the ntal results show that these methods are very promsng n provdng a sgnfcant mprovement wthn ths area whch s usually unmontored and s susceptble to long tme delays between formng and qualty nspecton. [1] H. Zhang and B. Lenno, Integrated condton montorng and control of fed-batch fermentaton processes, Journal Of Process Control, vol. 14, pp , [2] P. Nomkos and J. F. MacGregor, Mult-way partal least squares n montorng batch processes, Chemometrcs and Intellgent Laboratory Systems, vol. 30, pp , [3] M. Ignova, J. Glassey, A. C. Ward and G.A. Montague, Multvarate statstcal methods n boprocess fault detecton and performance forecastng, ransactons of the Insttute of Measurement and Control, vol. 19, pp , [4] A. AlGhazzaw and B. Lenno, Model Predctve Control montorng usng multvarate statstcs, Journal of Process Control, vol. 19, pp , [5] H. Zhang and B. Lenno, Mult way optmal Control of a Benchmark fed-batch fermentaton process, ransactons of the Instttute of Measurement and Control, vol. 25, pp , [6] M. Kanoa and. Nakagawa, Data-based process montorng, process control, and qualty mprovement: Recent developments and applcatons n steel ndustry, Computers and Chemcal Engneerng, vol. 32, pp. -24, [7] P. Gelad and B. R. Kowalsk, Partal Least Squares Regresson: A utoral, Analytca Chmca Acta, vol. 185, pp.1-17,