See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3771206

MultiWay PCA Applied to an Industrial Batch Process

Conference Paper · January 1994

DOI: 10.1109/ACC.1994.752268 · Source: IEEE Xplore

CITATIONS READS

71 333

5 authors, including:

Karlene A Hoo Michael J Piovoso

Montana State University University of Delaware
124 PUBLICATIONS   2,043 CITATIONS    101 PUBLICATIONS   2,081 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Cognitive Style in Engineering Education View project

All content following this page was uploaded by Michael J Piovoso on 11 April 2014.

The user has requested enhancement of the downloaded file.

 TA4 = 11:40
Multi-Way PCA Applied to an Industrial Batch Process
K. A. Kosanovichl, M. J. Piovoso, and K. S. Dah1
E. I. du Pont de Nemours & Co.
Wilmington, Delaware
J. F. MacGregor and P. Nomikos
Department of Chemical Engineering
McMaster University
Hamilton, Ontario

Abstract The main characteristics of batch processes: flexibd-

ity, finite duration, and nonlinear behavior, are asso-
Batch and semi-batch processes are common in most ciated with both their success and their incompati-
chemical companies. These processes are character- bility with the usual techniques for monitoring and
ized by a prescribed processing of materials for a finite control. However, disturbances and the absence of
duration of time. Feedback control often cannot be on-line quality measurements often d e c t the repro-
applied to correct for disturbances in a timely man- ducibility of batch processes and the lack of on-line
ner during the batch. Techniques which can provide quality instrumentation means that statistical qual-
insights into correlations among variables and their ity control is possible. SPC methods in batch pr+
relationships to product quality will provide insights cesses usually are limited only to final product qual-
in the design of a control strategy that may improve ity measurements [l],[2] or to a single variable mea-
product quality and minimize batch to batch varia- sured throughout the batch [3]. The use of statisti-
tions. In this paper, we apply the statistical technique cal methods such as Principal Component Analysis
of Multi-way Principal Component Analysis to ana- (PCA) and Partial Least Squares (PLS) to the anal-
lyze the data from an industrial batch process. Using ysis of multivariate continuous chemical process data
this technique, we were able to associate several sig- has been well documented 141, [5]. Nomikos and Mac-
nificant causes of variability with the recipe imposed Gregor [SIproposed SPC schemes for batch processes
by the process. This information gave rise to a differ- based on Multi-way PCA (MPCA), which directly
ent control strategy which is presented. Subtle effects uses the information of the on-line measurements to
among batches were also uncovered and identified. systematically and scientifically recognize significant
deviations from normal operating behavior of a pro-
cess through simple SPC charts. An empirical model,
based on the MPCA analysis of data obtained when
1. Introduction the process is operating well, is used to character-
ize the “in-control” behavior of the process. The
Batch and semi-batch processes play an important evolution of the future batches is then monitored by
role in the chemical industry due to their low vol- comparing them against this MPCA model using the
ume - high value products. Examples include re- statistical control limits developed from the reference
actors,crystallization, distillation, injection molding database [7]. Piovoso and Kosanovich [SI suggest a
processes and the manufacture of polymers. Batch Score space approach to control which transforms the
processes are characterized by a prescribed processing control moves in the score space to equivalent moves
of materials for a finite duration. Successful operation in the state space. The flavor of this method is sim-
means tracking a prescribed recipe with high degree ilar to the model-based control concept, however the
of reproducibility from batch to batch. Feed rate, model is a steady state PCR model.
temperature, and pressure profiles are implemented
with servo-controllers, and precise sequencing opera- In this paper, we use MPCA to analyze an indus
tions are produced with tools such as programmable trial polymer batch process. The goals of the anal-
logic controllers. ysis are: firstly, to corroborate and to extend our
process knowledge about the correlations among vari-
‘author to whom all correspondence should be addressed,
ables, their effectson quality, and batch to batch re-
kosanoh@esvax.dnet.dupont.com producibility; and secondly, to improve or redesign

1294
our control strategy to minimize quality variations. and the scores for each batch can be plotted in the
There are other techniques that may provide equiv- space of the principal components. All batches ex-
alent information such as first principles modelling. hibiting similar time histories will have scores which
In this rcgard, both techniques are complementary in cluster in the same region of the principal component
that the information garnered from the MPCA anal- space. Batches which exhibit deviations from normal
ysis can be used to guide the development of a first behavior will havc scores falling outside the main clus-
principles model. ter, but batches will1 similar faults will again cluster
in the same region.
The paper is organized as follows: first we provide
a brief review of MPCA, second, we introduce the The explained variance is computed by comparing thc
proccxs iI.nd jnstification for the analysis, and finally, true process data with the reconstructed data gcncr-
we discuss the rcsults of the MPCA analysis and how ated by the MPCA model. The reconstructcd data
they suggest improvements in the existing corilrol de- are generated from the scores for a particular batch
sign. and the corresponding loadings
R
ji = t,P
2. MPCA Method r=l

MPCA [9] is an extension of PCA to handlc data in The skm of the explained variance for each variablc
three-dimensional arrays. The three dimensions arise can be used to identify variables whose variations are
from batch trajectories that consist of batch runs, not related to other variables, as in the case of vari-
variables, and time. These data are organized into ables that are well controlled. Likewise the explained
an array X of dimension (IXJXK) where I is the num- variance for each time interval over all batches can
ber of batches, J is the number of varisblcs, and K is be used to identify periods where there is greater ex-
the number of time samples over the duration of the plained variability. This (:a11 be very useful in the case
batch. MPCA is equivalent to performing ordinary where varying phenomena occur in different stages of
PCA on a two-dimensional matrix formed by unfold- the reaction. Thcsc regions arc also where the vari-
ing tlie array X so that each of its vertical slices con- ables are highly correlated. We will show in the 34
tain the observed variables for all batches at a given that these two types of explained variances assisted
time. In this approach, MPCA explains the variation in clarifying and classifying the data.
of variablcs about their mcan trajectorics.

MPCA decomposes X into a summation of lhe prod- 3. Process Description

uct o f t score vectors (t) and p-loading matrices (P),
plus a residual matrix (E) that is minimized in a least The chemical process from which data werc taken is a
squares sense single batch reactor. The critical properties that must
be controlled Lor the h a 1 product, a polymer, are re-
lated to the cxtent of thc reaction (e.g., molecular
r=l weight distribution). Failure to attain on-aim con-
whcre R is the number of principal components uscd trol of these properties leads to higher manufacturing
in thc analysis. costs either through blending with other batches of
off-aim product or lhrough downgrading the product
This dccornpositiori summarizes and compresses the to end uses that have a lower selling price. The reac-
data both with respect to variables and time into tor is equipped with sensors that measure relevant
low dimerisional score spaces. Tlicsc spaces rcprcseiit tetnperalrires, pressures, and material Rows. The
thc major variability over the batchcs at all points product’s critical propcrties are determilied by off-
in time. Each p-loading matrix suiiiiitarizt>stlie ma- line chemical ineasiirenients and not every balcli Iias
jor tiinc variations of the viwiables almrtstheir aver- its crilical properties measured. The property mea-
age trajectories ovcr all the batches. l3y doing this, surement rcsults are available 12 hours or inore after
MI’CA utilizes not jus1 the magnitude of l h e devia- each the completion of tlie batcli. These results can-
tion of each variable from its memi trajcctory but also not be uscd in a timely fasliion to compcnsatc for poor
the correlations among them. The appropriate num- product quality. Furthermore, it is often difficult t o
ber of principal componcnts may bc foiind by cross- establish what causcd the properties to deviate froin
validatiori [7]. aim in the manufacture of a bad batch.

To analyze tlie perforiiiance of a set of batch runs, an The total time for the batch cycle is less thaii 2 hours.
MPCA iitialysis can he performed 011 all IJic balcties For the purposcs of lliis work, we can consider the two
1295
primary phenomena that occur to produce the poly- source and (2) the first principal component explains
mer; one dominates the early part of the cycle while significant variations in the first 50 minutes of the
the other takes place in the latter part. During the batch. The first feature explains why some batches
first part, the solvent is vaporized and removed from are grouped on the right hand side of Figure 2; they
the reactor, which takes approximately one hour. In were processed at a higher than normal initial pres-
the second part, the polymerization is completed to sure setpoint on the heat source. The second feature
attain a clesired molecular weight distribution. The indicates that there is more correlated variability in
finished product is then expelled under pressure from the early part of the recipe and that this time pe-
the vcssd to complete the cycle. A sample of some riod coincides with the time spent in boiling off the
of the process measurements profiles is showri iri Fig- solvent. This suggests the division of the data into
ure 1. The time speiit processing in bolh slages is two parts, stage 1 and stage 2, that reflect the two
caretiilly monitored, with the main control knob be- major chemical phenomcna discussed in the previous
ing the external heat source. Therc is also statistical section.
control of the vessel temperature. Known sources of
variability from batch to batch within a product line
are mainly due to variations in the heal conled, of llie 40
e4
heat source, various levels of impurities in the ingredi-
ents, and residual polymer buildup over the operating n I 33
life of a reactor.

4
-20

I
-30 -20 -10 0 10 20 30
Scorcs of principal component 1

Figure 2: Score plot of the first two principal compo-

nents for 50 batches

time 4 Redoing the analysis separately on the set of data

corresponding to only stage 1, shows more clearly
Figure I.: Process mcnsurement profiles those batches with the unusual initial pressure set-
point. In the score plol., there is a clear separation
of the batches with different pressure setpoints (see
Figure 3). I&”ving these batches will allow us to
4. Analysis & Results investigate the variability of the other batches pro-
cessed under the same controlled conditions.
Process variables for 50 batches made in the same
reactor, for Ihe same balch recipe, are sauipled at
one-minute intervals. These variables contain infor-
mation about thc stale of the reactor (temperatures,
pressures) and the state of the external lieat source.
This is no1 unusual as temperatures aiitl pressures re-
flect thc progress of the reaction in thc vessel. The
readings are scaled using the mean and standard de-
at
10

vinthn for each variable at each t h e interval prior

to analysis [6].

A plote of llie GrsL lwo principal cotttponwits, gen-

erated from applying MPCA to the data, is shown L I
in Figure 2 where the nunibers represent the indi- -30 -20 -10 0 IO 20 30
S a r a dpMc@.l ampncnt 1
vidual batches. These two components explain less
than 40% d the total variation in the dah. Other
features observed are: (1) the first loading vector is Figure 3: Score plot. of the first two principal compo-
dominated by the initial pressure setpoint on the lieat nents for 50 batches using stage 1 data

1296
4.1. MPCA of Stage 1 Data ticular temperature end point is achieved. Different
Applying MPCA to stage 1 data reveals that the first temperatures can occur during the increased heating
direction of variability is related to the reactor tem- period which will affect the reaction rates. Addition-
perature rise while the second corresponds to the heat ally, the time to the end state temperature will also
supplied by the heat source. These two principal com- be affected. Thus some batch cycles will be longer
ponents explain approximately 55% of the total vari- while others will be shorter which cause variations in
ance. A score plot of principal comporients 1 vs. 2 the polymer molecular weight and other properties.
is shown in Figure 4. It is not surprising that heat
effccts dominate stage 1 since boiling is the primary The last 60% of the time spent in this stage occurs
event that occurs there, and the boiling ratc and sub- after the recliict.ion in heat input. Careful inspection
sequenl kmperature rise is depended OIL lhe heat of thc data rcvcals a variation in the reactor temper-
tmnsfer from the heat source to [.he material in the aturc from batch to batch over these long trajcctorics
rcactor. What is learned howcvcr, is that for a con- during this reduced heat input period. The reactor
stant heat input and similar initial conditions, the pressure and temperature changes picked up by the
time history of the reactor temperature rise differs second principal component occur over the short pe-
From latch l o batch. 'Vllis implies lliat S O I I I ~batches riod of incrcascd heat input; yet wc believe that they
may boil more quickly than others and that prolonged are the key determinant of final product quality.
vigorous boiling may drive off' more than just the sol-
vent, hereby alfccting lhe final polymer composition. 11score plot of principal components 1vs. 2 for stage
A control strategy that forccs thc batch to follow a 2 analysis is shown in Figure 5. Comparing this figure
prescribed reactor temperature trajectory, in spite of to that of Figurc 4 we observe a pattern of clustering
heat transfcr limitations and pressure and tempera. among the batches. This is primarily due to the vari-
lure variations of the heat source should reduce lhe ations in the leiriperahire and pressure of the healirtg
batch to batch variability in the temperature rise. source. One possible control strategy to reduce the
variations is to more tightly control the heat input
as a function of the reactor tcmperaturc and p r c s
sure profiles as oppose to the current mode of timed
seqircnccs.

-a0 - 10 0 10 20
Saxesof princ'prl cmpaent 1
-15 .
Figure 4: Score plot of thc first two principal compo- -a, - 10 0 10 20
nents for 35 batches using stage 1 data S a r a o f p M a p l amponexi 1

4.2. MPCA of Stage 2 Data Figtire 5: Score plot of the first two principal compo-
The greatest direction of variability in stage 2 data is nents for 35 batches using stage 2 data
related to a temperature rise in the heat, soiirce while
llie second greatest directiori of variability is related
to the reactor pressure and temperature. These two 5. Discussion 8c Summary
priucipal components explain 46% or h e variability
in thc data. The second stage of the rccipc calls for We have shown that the technique of Multi-way Prin-
an increase in the heat to the reactor for a short time cipal Component Analysis applied to industrial data
followed by a turn down for the rest of the cycle. The lakcii froin a batch process can hc used to provide
illcrease is to provide enough activation eiiergy for the new insighls, corroborate existing process knowledge,
reaction to proceed. Once the reaction reaches a cer- and propose meaningful process improvements. The
tain point, a.~inferred by temperature measurements analysis indicates clearly that a division of the data
or time in the reactor, the heat source is Lurned down into sels that correspond to the two major chemical
while the temperature continues to rise iintil a par- phenomena will provide clarity and allow for inter-
1297
 pretation based on process understanding. Doing so tageous over having to develop different models for
led to the identification of the principal directions of each reactor for each product. There are other fac-
variability associated with each phenomenon and sug- tors which are not discussed but the insights so far
gested where to improve the existing control strategy. gleaned have lead to new process control strategies
and improved process understanding.
The score plot of stage 1 data shows the effects of dif-
fercnt initial pressure setpoint on thc heat, sourcc. It While MPCA focuses on the routine process measure-
is not unusual for batches to be processed undcr dif- ments, tying the analysis to the product quality data
ferent initial conditions; indeed changes are madc in is imperative. For example, we have found that stage
the setpoint of the pressure controller i n wsponsc to a 1 data is mow closely correlated with onc of the qiiitl-
statislical control limit violation on thc reactor tem- ity mcasurcnients. MacGrcgor ct, al. [LO] hnvc pto-
pcratura. The analysis of stage 1 data points out the posed Partial Least Squares as one method to accom-
imporlarice of adhering to a reactor teinperattve pro- plishing this.
file. If such a control strategy could he implemented,
more uniform boiling could be achieved thereby min-
imizing variations from batch to batch and loss of The authors would like to thank Mahmud Rahman
reactants. In stage 2, clustering among the batches for providing the data atid allowing us to usc it.
is evideiit. The clusters are found to be a function
ol’ tht: Icrriperature and pressurc of thc heat sourcc. References
Here too, maintaining a constant rate of haat input [l] G. J . Hahn and M. B. Cockrum, J . of Applied
a: a hnction of the reactor pressure a i d temperature Statistics, 14, 1987.
should minimise batch to batch va.ria.taionsand final [2] S. A. VandcrWiel, W. T . nicker, E’. W. Faltin,
polynicr proycrtics. and N. Doganaskoy, Technometrics, 34, 1992.
[3] C. E. Marsh and T. W. Tucker, ISA ‘hans., 30,
Equivalent results on a second rcactor producing the 1991.
mmc mstcrial arc obtaincd using MPCA (SCCFigurc
6). Obscrve that the data groups have a similar struc- [4] J. F. MacGregor and C. Kiparissides, Chem. &
inl. Lab Sys., 41, 1994.
tare (note the parallel nature) and that these two
principal c~rriponentsaccount lor the same arnounl [5] M. J . Piovoso, K. A. Kosanovich, and 3 . P. Yuk,
of variability in both groups. The segregation is due IEEE Trans. on Instr. & Meas., 41, 1992.
to a difference in the mcans from each reactor, which [G] P. Nomikov and J. F. MacGregor, io appear in
is rclatcd to the calibration of their rcspcctivc mc& AIChE J , 1993.
surement sensors. Redoing the analysis, after inde- 171 P. Nomikos, and J. F. MacGregor, submitted
pcndcntly scaling the two scto of data, removes the t o Tcchnometrics, 1004.
segregation. [8] M. .T. l’iovoso and K . A. Kosanovich, to appear
in Intl. J. of Control, Special Issue on Process Con-
trol, 1994.
[9] S. Wold, P. G c l d i , K. Esbensen, and J .
Ohman, J . of Chemometrics, 1, 1987.
[lo] J. E‘. MacGregor, P. Norrlikos, arid T . Kourli,
WAC ADCHEM’94, Kyoto, Japan, 1994.

-151
-3n -20 -10 0 in 20
Scores of principal component 1

Figure 6: Score plot of the first two principal compo-

nents for stage 1 data

The implication of this is that a single calibration

tnotlel tlevaloped for one reactor is suilabla I;)r olher
similarly designed reactors producing the same prod-
uct. For an on-line implementation, this is advan-
1298

View publication stats