Copyright © IFAC Dynamics and Control of Process System s.

Corfu. Greece. 1998

MUL TIV ARIA TE STATISTICAL PROCESS CONTROL AND PROCESS

PERFORMANCE MONITORING

E.B. Martin', A.J. Morris' and C. Kiparissides 2

J Centre for Process Analysis, Chemometrics and Control

University of Newcastle, Newcastle upon Tyne, NE1 7RU, United Kingdom

2 Department of Chemical Engineering and
Chemical Process Engineering Research Institute,
Aristotle University of Thessaloniki, P.O. Box 472, Greece

Abstract: Multivariate Statistical Process Performance Monitoring (MSPPM) provides

a diagnostic tool for the monitoring and detection of process malfunctions for continuous
and batch manufacturing processes. This paper initially reviews the concept of process
performance monitoring through an industrial application to a fluidised bed-reactor and
a simulation of a batch methyl methacrylate polymerisation reactor, prior to describing
some of the more recent work being carried out. This includes the development of
performance monitoring schemes from minimal process data, the use of multi-block
techniques for plant-wide monitoring and the development of generic models for the
monitoring of multiple products, grades or recipes. Copyright © 1998 IFAC

Keywords : Multivariate Quality Control; Batch and Continuous Processes; Plant-wide

Monitoring; Generic Models; Minimal Data.

I. INTRODUCTION manufacturing processes, and the quality and

consistency of production, has been Multivariate
A major technological challenge facing the process Statistical Process Control (MSPC). MSPC is
industries is the requirement to ensure the increasingly being recognised as a valuable tool for
manufacture of consistently good products. This is providing early warning of potential plant faults,
particularly important in highly demanding situations process malfunctions and process disturbances and
where manufacturing processes are subject to varying also for enabling a deeper understanding of the
raw material properties, changing market needs and process to be achieved . The methodology is equally
specifications and fluctuating operating conditions applicable to continuous and batch processes and has
due to equipment or process degradation. In highly been investigated by a number of researchers
competitive global markets, there is a real need to including, Kaspar and Ray (1992); Kresta et at.
minimise operating costs, reduce levels of re-work (1991); MacGregor et at. (1991); MacGregor and
whilst ensuring that variability inherent to the process Kourti (1995); Nornikos and MacGregor (1995);
is minimised. This has led to the development of Simoglou et al. (1997); Wise and Gallacher (1995);
intelligent supervisory systems that not only provide and Wise and Ricker, (1991).
early warning of the onset of process upsets,
equipment malfunctions or other special events, but
which can also contribute to the identification of the 2. MULTIV ARIA TE PROCESS PERFORMANCE
location of the assignable causes of these events. MONITORING

One of the more recent approaches for assessing and Multivariate statistical process control (MSPC) is the
improving the performance and operation of multivariate extension of univariate statistical process

347
control (SPC). The objective of SPC is to monitor 5. The calculation of summary information e.g.
the performance of a process over time to verify that mean, median, maximum, mInimUm, range,
it remains in a state-of-statistical-control, Wetherill skewness, kurtosis, correlation coefficients, etc.
and Brown (1991). Perhaps a more appropriate term 6. Test of the variance for single variables and the
for the MSPC concept is that of Multivariate possible elimination of non-significant
Statistical Process Performance Monitoring measurements.
(MSPPM). This can avoid the confusion which 7. The removal of unusual noise effects without
sometimes arises with the connotation of the words affecting the signal.
process colltroi. 8. Data alignment through time-shifting variables
which exhibit delays.
The primary requirement for the implementation of a 9. The identification and elimination of redundant
process performance monitoring scheme is the variables by utilising the results from the
acquisition of data that is representative of process correlation and principal component analysis.
behaviour when 'within specification product' is being 10. Mathematical transformations for variables
manufactured. The data can be obtained from either exhibiting non-linear relationships or non-normal
appropriate historical data bases or designed behaviour.
experiments. There are three key stages in the
development of a process performance monitoring Once the pre-screening phase has been completed, the
scheme - historical data collection, data pre-screening next stage is the development of a process
and finally the development of the process representation or model. The basis of multivariate
representation. Data pre-processing is an area which process performance monitoring on a continuous
has received little attention in the literature but which process are the statistical projection techniques of
in practice can be the key to success or failure of the Principal Components Analysis (PCA), lackson
final application. For an industrial data set, this stage (1991) and Projection to Latent Structures (PLS),
can consume up-to pproximately 70% of the Hoskuldsson (1988) with their multi-way extensions,
development time. A comprehensive pre-screening multi-way PCA (Kosanovich et ai, 1996) and multi-
procedure goes through a series of well-defined steps way PLS (Nomikos and MacGregor, 1996) being
in a structured manner, from the reading in of the raw applicable to the monitoring of batch processes.
data, the handling of data anomalies and finally These methods have been shown to be suitable for
through to the implementation of a statistical handling noisy and highly correlated data. PCA
representation/model for either retrospective analysis reduces the dimensionality of a single matrix of data,
or on-line implementation. The main tasks to be typically process data, laboratory assayed quality data
executed during pre-screening can be summarised as or reference values, by defining a series of latent
follows :- variables, principal components, which explain
decreasing amounts of variability in the data. In PLS
1. Input of raw data matrix (as collected by the data the dimensionality of both the process and quality
acquisition system(s», specification of the process space is compressed to a few pairs of factors . The
(input) and quality (output) variables including latent variables generated by both PCA and PLS are
identification of the measurement technique, linear combinations of the original process variables.
accuracy of the instrumentation, response time of
the measurement, whether the data acquisition The basis of the success of projection based
system has pre-filtered the variables, etc. techniques is the recognition that many of the
2. Detection and handling of incomplete/missing measurements are highly correlated and thus different
information. This stage involves the investigation combinations of the variables may define the same
of whether a pattern is exhibited by the missing underlying disturbances or events occurring on the
data, i.e. is it missing at random (MAR) or not process. Consequently, it can be assumed that when
missing at random, e.g. has the instrument reached the process is producing product within predefined
saturation or has a sensor failed . Depending upon specification limits, the dimensionality of the process
whether the data is MAR or otherwise, the next can be substantially reduced to a few latent variables
stage is to select the appropriate algorithm for which define the major sources of variability
handling the data. A number of alternative associated with the process (Kresta et aI, 1991). The
approaches can be tried at this stage such as data four most common forms of presenting the
deletion or linear interpolation. information to operators is through one and two
3. Graphical presentation of the data e.g. time series dimensional plots of the latent variable/principal
plots, histograms, scatter plots, principal component scores, Hotelling's T2 (Jackson, 1991)
components analysis, etc. can help identify and the squared prediction error, SPE.
whether the data is noisy, contains questionable
points, is normally distributed , etc. Squared Prediction Error:- Once a model has been
4. The resolution of questionable observations, developed from the nominal data set using k principal
identified from Step 3, through the conjunction of components or latent variables, the fitted values, can
engineering and statistical knowledge . be calculated for each new multivariate observation.

348
These values are then used to evaluate the squared 3. AN INDUSTRIAL APPLICATION
prediction error, SPE, or Q-Statistic for each new
observation, i.e . the squared difference between the A process performance monitoring scheme for a
observed values and the predicted values from the continuous process is illustrated by application to an
nominal or reference model. The squared prediction industrial large scale fluidised bed reactor. The
error plot provides the user with a facility to identify a primary aim was to provide early warning of process
previously unidentified event. abnormalities, cooling coil fouling ; catalyst
degradation and loss; and on-line analyser
Confidence Bounds:- By adopting an approach drift/malfunctions. The application does not lend
similar to that for univariate SPC, action and warning itself to mechanistic modelling approaches due to the
limits can be defined for each principal component complexity of the process. Thirty six process
scores plot based upon standard statistical variables and nine chemical quality variables were
distributional theory. One approach is to assume that measured on-line, the former at five minute intervals
the underlying k-dimensional data is normally and the latter by a high temperature gas
distributed and by calculating Hotelling's (1947) chromatograph with a frequency of forty minutes.
squared distance for the principal components of Further details of the process are withheld for reasons
interest, determine whether the process is in control :- of commercial confidentiality. A data-set was
initially extracted from the historical database which
(n - l)kFk n-k ex (I) was representative of desirable operation. The first
T2 > ..
n(n - k)
stage in the analysis was to pre-process the data and
in particular, eliminate periods of data collected
where
during undesirable process operation. This was
') - T
achieved through discussions with plant personnel
T- = n(J1o - x) S
-1 -
(J1o - x) (2) and the analysis of the chemical and process
information. Of particular interest was abnormal
This relationship can then be used to establish control events, for example when the chemical quality fell
limits where there is an 100a% chance of a false below the desired level. The data monitored during
alarm, where a is typically of the value 0.05 or 0.0 I . these periods was removed and retained in a separate
Rules similar to those adopted for univariate SPC for data base to enable the potential of a multivariate
identifying when a process is out of control, or statistical process performance monitoring scheme to
moving out of control, can be applied. Adopting a identify the onset of process problems to be assessed .
similar approach , nominal operating regions can be
defined for the squared prediction error, lackson and PCA Based Performance Monitoring:- A principal
Mudholkar (1979). component analysis was first carried out on the
nominal data set. Using cross-validation, six
Interpreting the 'Out of Control' Signal:- When a components were identified as the 'best' number to
process is recognised to be out-of-control, or moving describe the major sources of variability in the
out-of-control , operational personnel need a tool to process. Multivariate monitoring charts were then
identify the variables, or combination of variables, constructed based upon the principal component
that are indicative of a change in the process scores vectors, PCI and PC2, PC3 and PC4, PC5 and
operating conditions. Adjustments can then be made PC6 together with the squared prediction error (SPE),
to the process to avoid the continued manufacture of Figures I, 2, 3 and 4. In the subsequent plots '- - - '
non-conforming product. This stage is more easily defines the warning limits and ' - - ' the action
carried out by reverting back to the original process limits.
variables and examining their contribution to the
calculated scores and the squared prediction error. It
",
........

is then possible to identify the most likely J -'--

i
I
.-----
combination of original variables whose contribution sI // /
.' /
has increased over that defined from the nominal i I
! ..
;' \
model and which are therefore possibly indicative of !

n
non-conforming behaviour. One possible approach is
, <'
that of the contribution plot, Miller (1993), which
describes the contribution of the original process "-
~~;

,-
./ /
/
variables relative to the average value calculated from
the nominal PCNPLS representation .
'" ~
-~
.----

... -+i-~-----;-~--"'-~--;-~--;-~-'
. " ~ 0
S(.,u •• 'riad•• 1 c .... _.. _, I
•
"
Fig. I. Bivariate Scores Plot for PC 1 versus PC 2.

349
 ·~I The next stage in the analysis was to reintroduce the
I
1
-.. .. data which had previously been identified as ' non-

-
I .. ..

/
/' -- conforming ' by the plant engineers, Figure 5. From
this plot it is apparent that non-conforming operation
( would have been identified. The next step was to
'l identify the variable. or combination of variables.
J \ indicative of the cause of non-conforming operation .
-. -!

., I 1
'" '" ---- -- '. ... ~-
- This was achieved by reverting back to the original 36
process variables through the contribution plot. From
Figure 6. it is clear that variables three. four and six
exhibit larger values than anticipated. From the time
, series plot for variable three. Figure 7. it is clear that a
Pr .• clp.IC o lII,uultl

problem had occurred on the process.

Fig. 2. Bivariate Scores Plot for PC 3 versus PC 4

..-._..........
....... ......

/
/
, ",,/"

/ -- ---- "-
--, ,
'\
I

~
v
c [ ? ... ;~
)
, \ /
j "---
'-
"
~~"
..........
-- ./
".-,,,,,,,,
/

Fig. 6. Contribution of the Process Variables to the

, Change in the Scores for PC I .
I'riui,_I<':",", ... utS

Fig. 3. Bivariate Scores Plot for PC 5 versus PC 6

.,

"
1.5
~
: 20 1.4

1.3

1.2
" 1.1 '---~--~-~--~--~---'
o 200 400 600 BOO 1000 1200

Fig . 7. Time Series Plot for Process Variable 3

l OO

Although from the retrospective analysis the source of

Fig. 4. Residuals (SPE) Plot for the Nominal Data Set the problem was clear. the question could be raised.
would the problem would have been resolved more
-- -'-. rapidly if multivariate process performance
monitoring had been on-line? The plant operator
//~
:/ / '" '\ \,\
\ ..
monitors and controls two interacting reactors and
their associated equipment. It is therefore
/ questionable whether he would have noticed the
I. 1 problem as rapidly as through the multivariate
"\" "\ ... /
monitoring techniques. This analyses has given
operators and engineers confidence to move to full
... \ /
" '--'"
"- "-
/
./
/
/
/ on-line implementation of MSPPM.

Sur" ••
. ,
r'.~i,.I t: •• , .... 11
4. BATCH PROCESS PERFORMANCE
MONITORING
Fig. 5. Bivariate Scores Plot of for PC 1 versus PC 2
Recent manufacturing trends in Europe have been
towards high value added batch and semi-batch

350
processing. Currently, the most common industrial consisting of score vectors and loading matrices plus
practice for achieving consistent and reproducible a residual which is as small as possible, in a least
results from a batch process is based upon precise squares sense. Multi-way PCA utilises not just the
sequencing and automation of the major stages in magnitude of the deviation of a variable from its
each batch run . Monitoring is confined to checking mean trajectory, but the correlation structure between
that these sequences are followed and certain process variables. Indeed, it is the correlation structure which
variables follow acceptable trajectories. Abnormal can be particularly important in the detection of
conditions that develop during these batch operations faults . Usually a small number of principal
can lead to the production of at least one batch or, components can express most of the variability in the
more seriously, a sequence of batches of poor quality data when the variables are highly correlated and
product if the problem is not detected and remedied. identify similarities, or dissimilarities, amongst the
The monitoring of batch performance is becoming batches.
essential to ensure that the process operates safely and
high quality products are produced consistently.
Recent research approaches to batch monitoring has 4.2 Performance Monitoring of a Batch Process
typically focused upon the use of mechanistic,
knowledge-based and empirical methodologies. The batch polymerisation reaction used to illustrate
the implementation of multi-way PCA is based upon a
detailed mathematical model and simulation of a pilot
4.1 Multi-way Principal Components Analysis scale methyl methacrylate (MMA) reactor installed at
the University of Thessaloniki in Greece. A
Batch data differs from continuous data in that the mathematical model which includes reaction kinetics,
problem becomes three way, the added dimension heat and mass balances has been developed to
being that of time. Unlike continuous systems, it is provide a rigorous simulation which has been
not simply the relationships between the variables validated against the pilot plant. Using this
which are important, it is the entire past history of the simulation, representative studies of reactor operation
trajectory which contributes to the overall and the effects of process malfunctions and faults can
performance of the process. The data reduction be realistically studied. In this study, 40 reactor
techniques of PCA and PLS can again be used to batches are obtained by Monte-Carlo simulation to
reduce the dimensionality of the problem by provide a nominal (or reference) data set. An
projecting the information down onto a lower additional nine batches were generated which
dimensional sub-space. This new hyper-space, comprised two forms of malfunction - initiator
extracts the wealth of information contained in impurity and reactor fouling .
previously hidden relationships between variables and
within their time histories during previous successful Multi-way PCA was first applied to the forty nominal
batches. This extension to PCA and PLS is called batches. Using cross-validation two latent variables
multi-way PCA and multi-way PLS respectively. were identified as being sufficient to describe the
predictable variation of all the variables about their
average trajectories.

Principal Component 2
100

+
50
+

'"
Q) x
§I +
co
ID

Variables (M) · 50
+

·10

+
· 150
· 500 -400 ·300 -200 · 100 0 100
Principal Component 1

Fig. 8. The unfolding of three-way data

Fig. 9. Bivariate Scores plot for PC 1 versus PC 2.
A simple way to view multi-way PCA is to consider
Figure 9 presents the scores for principal component
opening out the three dimensional matrix into a two-
one and two. The three batches (+) lying close to the
dimensional array, Figure 8, by placing each 2-D
nominal operating region are associated with the
block consecutively and performing a standard PCA,
onset of fouling within the reactor.
Nomikos and MacGregor (1995 ,1996). The objective
of multi-way PCA is thus to decompose the three-way
array X into a series of principal components

351
4.3 Data Density Confidence Bounds regarded to be of the same length. When there is a
large database available. it may be possible to select
Conventional statistical bounds based upon batches that are of the same length. However, in most
Hotelling's r 2 statistic are known to be conservative industrial situations frequent operational changes
and can incorporate regions containing no mean that the database for any particular nominal
information from the nominal data set. Thus the operating region is seldom large enough to choose
operator is compelled, through statistical evidence, to only batches that are equal in length to build the
assume that any observation in the future falling MPCA process representation or MPLS model. In
within this region will result in the production of addition, batches that are longer. or shorter. than the
satisfactory product, i.e. extrapolation is valid. A selected length might provide valuable information on
recently developed approach to the generation of the nominal operating region and may be as important
in defining the nominal operating space as batches
confidence bounds, the M 2 statistic , Martin and
that are of the most commonly occurring length. In
Morris (1997), helps address this issue. The
some applications of MPCA and MPLS. it is assumed
M 2 statistic is an empirical density based approach that the selected batch length is taken to be that of the
where the bounds are calculated using the statistical shortest batch and for any batches longer than this the
techniques of the bootstrap and density estimation. data is ignored. In effect all batches are 'chopped' to
The approach is based upon the construction of a the minimum nominal batch length. In other
95 % and 99% likelihood confidence region for a situations, all short batches are filled in by the current
vector parameter e of length d, using the bootstrap deviation up to the maximum batch length. A study
and non-parametric density estimation , Hall (1987). of alternative approaches to deal with unequal length
batches and to compare their model structure and
Prin c ipal Comp onent 2
100 performance in differentiating between nominal and
fault batches, has been carried out by Rothwell et al.
+
+ (1998). Three different methods were compared:
50
cutting to a minimum length, dynamic time warping
+ (Gollmer and Posten. 1995) and the use of an
+
indicator variable instead of time.
·50
+
By switching from monitoring variables on a time
· 10 basis to using an indicator variable from the available
+
process variable set. such as composition or weight,
·150,L--~--~-~~--'---~---'
· 500 · 400 ·300 ·200 · 100 10 0
the resulting process representations were shown to
Prin cipal C omponent 1 generally either demonstrate an improvement or
match the performance of existing batch monitoring
Fig. 10. Bivariate Scores plot for PC I versus PC 2. methods of cutting batches to minimum length. The
second method involving a dynamic time warping
In Figure 10, the results based upon the M2 statistic algorithm, whilst providing adequate performance.
are presented for the batch polymerisation process. was observed to be less successful than the other
Compared with the results based upon the standard methods, in this study. The indicator variable method
distributional approach, Figure 9, the empirical based is process specific and requires the selected process
variable to be continuous, monotonic and span the
metric M 2 does not incorporate the non-conforming
entire range of all other process variables being used
batch within the nominal data region. This confirms
in the multivariate statistical model. Such a variable
the belief that the approach based upon Hotelling's
may not exist in every industrial process. in which
r2 is conservative and areas of no information on case one of the other methods described could be
batch outcome can be encompassed within the limits used.
as demonstrated in Figure 9. This does not occur
with the approach based upon data density. This
approach is currently being pursued for on-line 5. INVERSE PROJECTION TO LATENT
application and also for non-linear and dynamic STRUCTURES (lPLS)
process performance monitoring.
In processes where there is only a limited amount of
Unequal Length Batches:- One practical limitation of process data or a small number of complete batches.
the existing methods for monitoring batch processes techniques to generate representative process data
using Multi-way Principal Component Analysis would be particularly advantageous in order to set up
(MPCA) and for predicting final batch quality an initial process performance monitoring scheme. In
through Multi-way Partial Least Squares (MPLS) is order to examine this conjecture, consider the
that they only are applicable to batches of equal situation either where a new process is being
length. Batch responses described as nominal for the commissioned, or an existing process is being
purpose of building an MPCA representation are all expanded to manufacture different products.

352
Consequently there may only be a few production These are known to influence the final properties of
batches available from a designed experiment to the polymer product.
identify the nominal product quality and the
associated operational conditions. The assumption is The first step was to define a set of nominal
made that the data is representative of successful experimental initial conditions for the production of a
operation and therefore defines the nominal operating particular grade of polymer which represents realistic
range of desired production and thus forms the basis conditions of production. This set was defined using
of the regression model. Having identified the set of a factorial design and corresponded to the production
data to be used to build the empirical representation, a of nine products of the same grade. A set of nine
PLS regression model is calculated to model the batches were generated through monte-carlo
features of interest, such as the initial process perturbation of the pilot plant simulation covering the
conditions and/or the final product properties, from desired operating region and representing normal
the process measurements. process operations i.e. to mimic a real industrial
situation. A multi-way PLS (MPLS) regression
Based upon the derived the PLS model based upon model based upon the nine experimental batches was
minimal process data, an unknown new process then built. The number of latent variables required to
measurement vector ( X o ) can be predicted from the provide a good prediction of the Y -block was
identified, through cross-validation, to be seven. The
corresponding predefined values of the features ( Yo) '
MPLS regression model developed from the nine
by inverting the PLS regression model to obtain the experimental batches relating the process variables to
classical estimator, Papazoglou et al. (l998b). In this the initial process conditions was then inverted to
way , using the process measurements computed from obtain the inverse MPLS model. The ability of the
the IPLS model , a large number of trajectories from IPLS methodology to predict the estimated
within the nominal region of the regression model can trajectories of the process measurements was then
be computed . These IPLS estimates of the process investigated by presenting the set of initial conditions
data, along with the existing plant measurements, can ( Yo), resulting from the original nine experimental
then be used to develop an initial process
batches, as the inputs to the IPLS model. The IPLS
performance monitoring scheme. As new process
measurements become available from the production estimates, .:to, corresponding to the process
plant, itself, new and improved PLS and IPLS models measurements were then calculated.
can be built from the updated process database .
Finally, when sufficient data from the actual
manufacturing process becomes available, a
multivariate process performance monitoring scheme

~.,
based purely upon the plant data as is current
practice.
,~

5.1 Application to a Batch Polymerisation Reactor

"-:~'- , I ~---~ '

The methodology of inverse PLS is applied to a

comprehensive simulation of a pilot-scale batch
methyl-methacrylate suspension polymerisation
reactor, developed and installed in the Department of Fig. 11 . Typical process trajectory from the IMPLS
Chemical Engineering, Aristotle University of estimate and the corresponding trajectory from
Thessaloniki , Greece. The process variables the simulation.
monitored include the reactor temperature (Treac) ,
the coolant inlet and outlet temperatures, Tc,in and Key :- --------- IMPLS estimate; - - Actual
TC,out, respectively, the coolant flowrate , F j , and the
A typical process trajectory for the coolant inlet
conversion of monomer, Conv. On-line process
temperature was selected from the original nine
measurements were collected on a one minute basis
experimental batches to be examined more closely.
with noise being added at a level typical of that of the
The behaviour of this trajectory was then compared
real process. The productivity variables of interest
with the corresponding trajectories computed from
are the final conversion of monomer MMA ( X MMA ) the IPLS estimates, Figure 11 .
and the molecular properties, weight average
molecular weight (M IV ) and number average The ability of the estimates of the process variables,
molecular weight ( M N) . The length of each batch calculated from the inverse PLS model, to simulate
was 120 minutes . The initial process conditions the real process behaviour was validated through two
application studies. The first relates to the
which are subject to variability initial initiator weight
development of an inferential multi-way PLS model
( 10 ) and overall heat transfer coefficient (U 0) .

353
to estimate the initial process conditions at an early Extensions of the projection based techniques of
stage in the polymerisation process. The second PCA and PLS have been developed to enable data
application relates to the development of a obtained from complex and inter-connected processes
multivariate process performance monitoring scheme to be analysed, multi-block PCA and multi-block
based upon a multi-way principal component analysis PLS. Multi-block data analysis techniques have their
(MPCA) representation. Both applications were origin in path analysis and path modelling in the field
implemented using both the estimates of the process of sociology and econometrics. However the
measurements derived from the IMPLS model and the multivariate projection algorithms for analysing
corresponding process measurements obtained from blocked data can be primarily attributed to Wold
the pilot plant simulation, which acts as surrogate (1982), Wo Id et al. (1987) and Wangen and Kowalski
manufacturing plant data, Papazoglou et al (1998). (1988).

The above approach offers an attractive and effective The techniques of multi-block PCA and multi-block
way to implement multivariate process performance PLS allow a number of inter-connected data blocks to
monitoring schemes even in cases where there is be analysed in a similar manner to that of classical
initially limited process data from the plant. One of PCA and PLS. However the multi-block approach
the key requirements of this approach is that a takes account of the structure of the process and thus
satisfactory PLS regression model can be built from places greater emphasis on within block correlations
the initial 'design' data. An additional assumption is as opposed to the between block relationships which
that the process 'design' data defines the nominal by their very nature will be weaker. Figure 12
region of the PLS and IPLS models and that they are illustrates a typical structure for inter-connected
representative of the desired operating region for process operations - data-blocks.
acceptable production. Thus, process data deemed to
be outside the operating region must be excluded
from the analysis.

6. MULTI-BLOCK TECHNIQUES

The multivariate statistical projection techniques of

PCA and PLS and their multi-way counterparts,
multi-way PCA and multi-way PLS, have been shown
to be appropriate for establishing process
performance monitoring schemes for continuous and
batch processes respectively. However, in practice Fig. 12. A typical Multi-Block PLS structure
industrial processes comprise a number of processing
units and in such situations, the traditional approach Process performance monitoring schemes based upon
of PCA and PLS may be inappropriate to handle, the multi-block PLS have been reported in the literature,
large volumes of correlated data, efficiently and (MacGregor et aI.1994). Multi-block PLS was found
effectively. This may result in mis-information being to be able to provide the necessary information for
passed to the operators or alternatively an over-load process operators to monitor the process, to detect
of information being reported which the operating abnormal events and furthermore, to efficiently
personnel are required to interpret and act upon . This diagnose the cause of non-conforming operation by
is undesirable and could result in the advantages of locating its origin more specifically within the system
the technology being questioned . which comprised a number of different zones.
However, there are two major issues which must be
Traditionally in cases when there is a large number of addressed when considering the implementation of a
data items (observations or variables), one of two multi-block process performance monitoring scheme.
approaches is adopted by which the data is analysed . The first issue is the approach that will be used to
The first approach, sampling , is where only a few apply blocking or grouping to the process variables.
data items are selected to be examined in detail and The choice of blocks depends upon engineering
the remaining observations are assumed to exhibit judgement and the objectives of the study and should
similar behaviour to those inspected. The second be considered very careful. Inappropriate blocking
approach is to divide the data into groups, blocks, may lead to models which are unable to perform their
categories or clusters and then to consider these task effectively. Variables can be blocked according
groups as "super-items". This approach is called to process units, e.g. fluidised spray dryer, grinder,
grouping or blocking . Sampling is usually applied to classifier, etc. or alternatively by process type, e.g.
cases when there are many observations while one block may comprise temperatures, a second,
blocking is applied when there are many variables. flows , a third, production costs, etc. Although not
reported in this paper, it has been shown that the
efficiency of a multi-block PLS monitoring scheme

354
where process variables were inappropriately Three contrasting approaches can be considered to
grouped , was impaired and as a result the location of present the information from the three product grades
the origin of the fault was masked. to the operator: a separate MPPM representation for
each grade, a MPPM representation based upon a
The second issue is the number of control charts that global model comprising all three grades, or
process operators are required to monitor. Using a alternatively a generic MPPM model. The results for
process performance scheme based on the classical principal component 1 and principal component 2 for
PLS method, a minimum of two control charts (latent the combined model and the generic model are
variable plot and SPE plot) require to be monitored. presented in Figures 13 and 14, respectively .
In the case of an inter-connected process comprising a
number of distinct processing units, one approach is
to build a separate PLS model for each unit. As a
result, this increases the number of control charts that
.", 2
~.
.,:.
.... -.
~ .. ...... ~

operators are required to monitor correspondingly. ~o .t:.. ". ".....", "'_-

However, by applying multi-block procedures, the
n

3 '.,: . "-. ··~··t··

..
-g ·2
number of control charts required can be reduced . ii
a
~·4

Finally, the implementation of a multi-block scheme ·6L-__ ~ __ ~ ____ ~ __ ~ __ ~____ ~ __ ~ __- J

for performance monitoring will be specific to the ·4 .) ·2 ·1
PrinCipal component I
process being considered. Thus the challenge of the
Fig. 14. Process performance monitoring based upon
implementation of such a scheme will be team work
a generic model for three recipes .
involving process engIneers, operators and
statisticians.
It is clear that the generic model does not exhibit any
clustering of the individual grades in sharp contrast to
the combined model. The major advantage of the
7. GENERIC MODELS
generic model , compared with the approach of
developing a process representation for each
Traditionally, principal components analysis has been
individual grade, is that the number of models
viewed as a single group method and in particular
required for monitoring is considerably reduced. This
process performance monitoring has been used as a
is especially beneficial if the production facility is
monitoring and diagnostic tool for single product
running a large number of grades or recipes in
manufacture. An extension to PCA enables a number
response to market changes. By developing a global
of similar products to be monitored through a single
model which draws together all the grades, without
generic model. The method is based upon the
initially developing a common sub-space for the
existence of a common eigenvector sub-space for the
various recipes, an overall reduction in fault
sample variance-covariance matrix of the individual
sensitivity of the model can result. In addition
products. The pooled sample variance-covariance of
additional principal components are required to be
the individual products are then used to estimate the
included to account for the variability associated with
loadings for the generic model. The method is
the different polymer grades.
illustrated through application to a two zone high
pressure low density poly ethylene reactor
Overall the potential of such a development is far-
(MacGregor et aI.1994) . Details of the approach are
reaching and provides the major step forward
described in Lane et at (1998). Three different
required to ensure the wide application of multivariate
product grades were produced and these provided the
process performance monitoring within the process
reference data set.
industries.

8. FUTURE WORK

There still remain a number of areas of research

which require to be addressed before these new
approaches to monitoring can be widely adopted and
routinely rolled out by industry.
·6 L-----~----~----~----~----~--~
·6 -4 ·2
Pnncipal component 1 • The extension of multivariate process
Fig. 13. Process performance monitoring based upon performance monitoring (MPPM) techniques to
a combination of three recipes. complete process plants using multi-block
projection methods.
Key :-. Grade 1 '+' Grade 2 ,*' Grade 3 '0' • The sensitivity of the PCA and PLS models and
the appropriateness of the confidence bounds .

355
• The effect of plant dynamics and plant control MacGregor J.F., C. Jaeckle and C. Kiparissides and
loops on the process performance monitoring M. Koutoudi (1994), Monitoring and Diagnosis of
methodology and their potenti al to mask non- Process Operating Performance by Multi-block
conforming process behaviour. PLS Methods with an Application to Low Density
• The extension of the density-defined confidence Polyethylene Production'. AlChE Journal 40, 826-
bounds to monitor dynamic processes. 838.
• The assessment of the need for the development of MacGregor J.F. and T. Kourti (1995), Statistical
non-linear multivariate statistical process control. Process Control for Multivariate Processe'. Chem.
Eng. Prac. 3, 403 - 414.
Martin, E. B. and Morris, A. J. , (1996), Non-
9. ACKNOWLEDGEMENTS parametric Confidence Bounds for Process
Monitoring Charts, 1. of Proc. Cont., 6, 6, 349-358 .
The authors acknowledge the contributions, support Miller, P., Swanson, R. E. , & Heckler, C. F., (1993),
and commitment and lively discussions with the Contribution Plots: The Missing Link in
members of the EU ESPRIT Project Process Multivariate Quality Control, 37th Annual Fall
Diagnostics for Plant Performance Enhancement Con! ASQC, Rochester N.Y.
(PROGNOSIS No . 22281); the European Union for Nomikos P. and J. F. MacGregor (1995), Monitoring
their funding contributions; and their colleagues in of Batch Processes Using Multi-way Principal
CPACC, in particular Mr A. Simoglou, Dr M. Components Analysis', AIChE Journal, 1361 - 1375.
Papazoglou and Mr S. Lane . Nomikos P. and MacGregor, J. F. (1996), Multi-way
PLS in Monitoring Batch Processes, First
International Chemometrics Conference Internet
10. REFERENCES Conference, InCINC '94.
Papazoglou, M. c., Martin, E. B. Kiparissides, C. and
Gollmer, K. and C. Posten, (1995), Detection of Morris, A. J. (1998) Generating Additional Process
Distorted Pattern using Dynamic Time Warping Data Through Inverse Projection to Latent
Algorithm and Application for Supervision of Structures for Multivariate statistical Process
Bioprocesses, In: On-line FauLt Detection and Control, submitted to Technometrics.
Supervision in the Chemical Process Industries , Rothwell , S. G. , Martin, E. B. and Morris, A. J.
(Martin and Morris , (Ed.») ISBN 0 08 0423760, (1998), Comparison of Methods for Dealing with
101-106. Uneven Length Batches with Application to MPCA
Hall, P. (1987), "On The Bootstrap and Likelihood- Monitoring of a Batch Process accepted 7U1
based Confidence Regions" , Biometrika, 74, 481- International Conference on Computer Applications
493 . in Biotechnology (CAB7), Osaka, Japan.
Hoskuldsson, A. (1988), "PLS Regression Methods" . Simoglou, A., E. B. Martin, A. J. Morris, M . Wood
J. Chemometrics, 2, 211-228. and G. C. Jones (1997), Multivariate Statistical
Hotelling, H., (1947), "Multivariate Quality Control", Process Control In Chemical Manufacturing, IFAC
in Eisenhart, Hastay and Wallis (Eds) Techniques Symposium on Fault detection, Supervision and
of Statistical Analysis, (New York) . Safety for Technical Processes, Safeprocesses'97,
Jackson, J. E. (1991), A Users Guide to Principal Kingston upon Hull, UK.
Components, John Wiley and Sons, New York, Wangen L. E. and B. R. Kowalski, (1988) A
NY . Multiblock Partial Least Squares Algorithm for
Jackson, J. E. and Mudholkar, G. S., (1979), Control Investigating Complex Chemical Systems, J. of
Procedures for Residuals Associated with Principal Chemometrics, 3, 3.
Components Analysis, Technometrics , 21 , 3. Wetherill, W.B. and D. Brown (1991), Statistical
Kaspar M.H. and W. H. Ray (1992), Chemometric Process Control for the Process Industries, 3rd edn.
Methods for Process Monitoring and High- Chapman and Hall, London.
performance Controller Desig'. AIChe JournaL 38, Wise, B. and N. Gallacher (1996), The Process
1593-1607. Chemometrics Approach to Process Monitoring and
Kresta J. , J. F. MacGregor and T. E. Marlin (1991), Fault Detection' Journal of Process Control, 6, 329
'Multivariate Statistical Monitoring of Process - 348.
Operating Performance, Canadian JournaL of Wise B.M. and N. L. Ricker. (1991), 'Recent
Chemical Engineering . 69, 34 - 47 . Advances in Multivariate Statistical Process
Kosanovich, K.A. , S. Dahl and M.J. Piovoso (1996), Control'. IFAC Symposium, ADC HEM '91,
Improved Process Understanding Using Multi-way Proceedings, Toulouse, France.
Principal Component Analysi', Ind. Eng. Chem. Wo Id H., Soft modelling. The basic design and some
Res., 35, 138-146. extensions, Systems under Indirect Observations,
Lane, S. E. B. Martin, R. Koijmans, and A. J. Morris North Holland, Amsterdam (1982).
(1998). The Use of Generic Sub-space Models for Wold S., Geladi, P., Esbensen, K. and Ohman, J.
the Monitoring of a Semi-Discrete Batch Process, (1987), 'Multi-way Principal Components and PLS
submitted to Technometrics. Analysis'. J. ofChemometrics 1,41-56.

356