Chemometrics and

intelligent
laboratory systems
ELSEVIER Chemometrics and Intelligent Laboratory Systems 28 (1995) 3-21

Tutorial

Process analysis, monitoring and diagnosis, using multivariate

projection methods
Theodora Kourti *, John F. MacGregor
M&faster Adcanced Control Consortium, Department of Chemical Engineering, McMaster Unil’ersity, Hamilton, Ontario LSS 4L7, Canada

Received 5 August 1994; accepted 15 November 1994

Abstract

Multivariate statistical methods for the analysis, monitoring and diagnosis of process operating performance are
becoming more important because of the availability of on-line process computers which routinely collect measurements on
large numbers of process variables. Traditional univariate control charts have been extended to multivariate quality control
situations using the Hotelling T2 statistic. Recent approaches to multivariate statistical process control which utilize not only
product quality data (Y), but also all of the available process variable data (X) are based on multivariate statistical projection
methods (principal component analysis, (PCA), partial least squares, (PLS), multi-block PLS and multi-way PCA). An
overview of these methods and their use in the statistical process control of multivariate continuous and batch processes is
presented. Applications are provided on the analysis of historical data from the catalytic cracking section of a large
petroleum refinery, on the monitoring and diagnosis of a continuous polymerization process and on the monitoring of an
industrial batch process.

Contents

1. Introduction ...................................................... 4
2. Multivariate methods for monitoring product quality. ............................. 5
2.1. Traditional multivariate quality control charts .............................. 5
2.2. Quality control charts based on principal components ......................... 7
3. Multivariate methods for process monitoring .................................. 8
3.1. PLS - partial least squares. . . . . . . . . . . . . . . .. 10
3.2. Analysis of historical process data sets .. . . . . .. . . . . . . . . 10
3.3. Monitoring continuous processes. . . . .. . . . . . . . . . . . . 12
3.4. Diagnosing assignable causes. .. . . .. . . .. . . . . . . . . 13
3.5. Multi-block PLS . . . . . . . . . . . . . . . ., . . . . . . . . . 13
3.6. Monitoring batch processes . . . . . . . . . . . . .. . . . . . . . . . 16

* Corresponding author

0169.7439/95/$09.50 0 1995 Elsevier Science B.V. All rights reserved

SSDI 0169.7439(94)00079-4
 T. Kourti, J.F. MacGregor/ Chemometrics and Intelligent Laboratory Systems 28 (1995) 3-21

4. Summary ........................................................ 20
Acknowledgements .................................................... 20
References ......................................................... 20

1. Introduction data simultaneously can also extract information on

the directionality of the process variations, that is on
The objective of statistical process control @PC) how all the variables are behaving relative to one
is to monitor the performance of a process over time another. Furthermore, when important events occur
to verify that it is remaining in a ‘state of statistical in processes they are often difficult to detect because
control’. Such a state of control is said to exist if the signal to noise ratio is very low in each variable.
certain process or product variables remain close to But multivariate methods can extract confirming in-
their desired values and the only source of variation formation from observations on many variables and
is ‘common-cause’ variation, that is, variation which can reduce the noise levels through averaging.
affects the process all the time and is essentially The application of multivariate projection meth-
unavoidable within the current process. ods, such as principal component analysis (PCA),
Traditionally, SPC charts (Shewhart, CUSUM and partial least squares (PLS), multi-block PLS and
EWMA) are used to monitor a small number of key multi-way PCA to process monitoring and fault diag-
product variables (Y) in order to detect the occur- nosis is reported here. Similarities and differences
rence of any event having a ‘special’ or ‘assignable’ with the traditional methods are discussed. The use
cause. By finding assignable causes, long term im- of the projection methods for analyzing and inter-
provements in the process and in product quality can preting historical plant operating records available in
be achieved by eliminating the causes or improving computer data bases is illustrated with an example
the process or its operating procedures. However, from a large petroleum refinery. On-line monitoring
monitoring only a few quality variables is totally and diagnosis of process operating performance in
inadequate for most modern process industries. The continuous processes (using PLS and multi-block
traditional SPC approaches ignore the fact that with
computers hooked up to nearly every industrial pro-
-.
cess, massive amounts of data are being collected
routinely every few seconds on many process vari-
ables (X), such as temperatures, pressure, flow rates,
)
etc. Final product quality variables (Y), such as
polymer properties, gasoline octane numbers, etc.,
are available on a much less frequency basis, usually
from off-line laboratory analysis. All such data should Time --)

be used to extract information in any effective scheme

for monitoring and diagnosing operating perfor-
mance. However, all these variables are not indepen-
dent of one another. Only a few underlying events
are driving a process at any time, and all these
measurements are simply different reflections of these
same underlying events. Therefore, examining them
one variable at a time as though they were indepen-
dent, makes interpretation and diagnosis extremely
difficult. Such methods only look at the magnitude
of the deviation in each variable independently of all Fig. 1. Quality control of two variables - the misleading nature
others. Only multivariate methods that treat all the of univariate charts.
 T. Kourti,J.F. M acGregor/ Chemometricsand IntelligentLaboratorySy stems28 (1995) 3- 21 5

PLS) and batch processes (using multi-way PCA) is with their corresponding control limits. Note that by
presented and illustrated with a continuous polymer- inspection of each of the individual Shewhart charts
ization process and an industrial batch process. the process appears to be clearly in a state of statisti-
cal control, and none of the individual observations
give any indication of a problem. The only indication
2. Multivariate methods for monitoring product
of any difficulty is that a customer has complained
quality
about the performance of the product corresponding
Statistical process control charts such as the to the 8 in Fig. 1. If only univariate charts were
Shewhart chart [II, the CUSUM plot [2] and the used, one would clearly be confused. The same
EWMA chart [3], are well established statistical pro- customer apparently liked all the other lots of prod-
cedures for monitoring stable univariate processes. A uct sent to him, many of them with values of y, and
Shewhart chart consists of plotting the observations y, much further from target. The true situation is
sequentially on a graph which also contains the only revealed in the multivariate y, vs. y, plot
target value and upper and lower control limits. If an where it is seen that the lot of product indicated by
observation exceeds the control limits a statistically the @J is clearly outside the joint confidence region,
significant deviation from normal operation is and is clearly different from the normal ‘in-control’
deemed to have occurred, which triggers the search population of product.
for an assignable cause. The control limits are usu- In spite of the misleading nature of univariate
ally determined by analyzing the variability in a quality control charts they continue to be almost the
reference set of process data collected when only only form of monitoring used by industry. However,
normal or ‘common cause’ variability is present and several multivariate extensions of the Shewhart,
acceptable operation is achieved. The limits are then CUSUM and EWMA based on Hotelling’s T2 statis-
usually set at plus and minus three standard devia- tic have been proposed in the literature (see review
tions about the target. articles by Wierda [4] and Sparks 151).
In most industries, traditional univariate control
charts (Shewhart, CUSUM and EWMA) are used to 2.1. Traditional multivariare quality control charts
separately monitor key measurements on the final
Natural extensions of the Shewhart chart to situa-
product which in some way define the quality of that
tions where one observes a vector of k variables
product. The difficulty with this approach is that y,, 1 at each time period are the multivariate x2 and
these quality variables are not independent of one
T2 charts. The T2 chart has its origins in the work of
another, nor does any one of them adequately define Hotelling [6], and several references [7-121 discuss
product quality by itself. Product quality is only the charts in more detail.
defined by the correct simultaneous values of all the Given a (k x 1) vector of measurements y on k
measured properties, that is, it is a multivariate prop- normally distributed variables with an in-control co-
erty. variance matrix 2 one can test whether the mean /.L
The difficulty with using independent univariate of these variables is at its desired target 7 by
control charts can be illustrated by reference to Fig.
computing the statistic
1. Here only two quality variables ( yl, yZ) are
considered for ease of illustration. Suppose that, (1)
when the process is in a state of statistical control This statistic will be distributed as a central x2
where only common cause variation is present, y, distribution with k degrees of freedom if I_L= 7. A
and y2 follow a multivariate normal distribution and multivariate x2 control chart can be constructed by
are correlated ( pY,y2 = 0.8) as illustrated in the joint plotting x 2 vs. time with an upper control limit
plot of y, vs. yZ in Fig. 1. The ellipse represents a (UCL) given by xi(k) where (Y is an appropriate
contour for the in-control process, and the dots repre- level of significance for performing the test (e.g.
sent a set of observations from this distribution. The ff = 0.01).
same observations are also plotted in Fig. 1 as Note that this multivariate test overcomes the
individual Shewhart charts on y, and y, vs. time difficulty illustrated in the example of Fig. 1, where
6 T. Kourti, J.F. MacGregor/Chemometrics and Intelligent Laboratory Systems 28 (1995) 3-21

univariate charts were incapable of detecting the where F,(k, n - k) is the upper 100~1% critical
special event denoted by @J The x2 statistic in Eq. point of the F distribution with k and n - k degrees
(1) represents the directed or weighted distance of freedom [13].
(Mahalanobis distance) of any point from the target The above charts are for a single new multivariate
7. All points lying on the ellipse in Fig. 1 would observation vector at each time. If an average of m
have the same value of x2. (The ellipse is the new multivariate observations are to be used at each
solution to Eq. (1) for x * = x,(k), for two variables). time or if the estimate of the variance S is based on
Hence, a x2 chart would detect as a special event pooling estimates from rational subgroups, then the
any point lying outside of the ellipse. above definitions of the x2 and T’ charts and their
When the in-control covariance matrix Z is not UCLs must be correspondingly redefined [4]. Fur-
known, it must be estimated from a sample of n past thermore, if the charts are utilized to examine past
multivariate observations as data that are also used in computing S, then the
distributional properties of T2 are different from the
S=(n-1))’ 5 (yi-y)(y,-j)T (2) zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHG
above [4,13].
r=l Alternatively, other types of multivariate charts,
When new multivariate observations (y) are ob- such as multivariate CUSUM and multivariate
tained, then Hotelling’s T* statistic given by zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFE
EWMA charts may be used [4,5].
The above ideas are illustrated here by monitoring
T*=(y - ~)~S- ‘(y - 7) (3) the properties of low-density polyethylene produced
can be plotted against time. An upper control limit in a multi-zone tubular reactor (we consider here the
(UCL) on this chart is given by: first two zones>. Details on this simulated process
can be found in MacGregor et al. [14]. The operating
T&c (n-r)(n+l)kF,(+k) (4) conditions in the reactor influence the molecular
n(n-k) properties of the polymer produced (weight and

100
56 zyxwvutsrqponmlkjihgfedcbaZY

90 -

80 - .54 _

70 -

60 -

50 -
.53
40 -

30 -

20-
UCL {99
____________L ______
0 %) -----____--.....___-....______________________-_________ 22

IO- *; l
I
l
.
. * l: l . l
.*
l* . l* l . l l.* l***
. **. l** l * .* l. l

0 l l l
l .
0 10 20 30 40 50 60

OBSERVATION NUMBER

Fig. 2. T2 chart on five product properties of polyethylene.

 T. Kourti, J.F. MacGregor/ Chemometrics and Intelligent Laboratory Systems 28 (1995) 3-21 7

number average molecular weights, MW, and MW,, PCA is scale dependent, and so the Y matrix must
and long and short chain branching, LCB and SCB) be scaled in some meaningful way. The most usual
and these in turn affect the behaviour of the polymer form of scaling is to scale all variables to unit
in its final application. The productivity variable of variance and then perform PCA on the correlation
interest is the conversion per pass, CONV. The five matrix. Alternatively, in quality control situations,
product variables (Y) of interest (MW,, MW,, LCB, scaling the Ys inversely proportional to their specifi-
SCB and CONV) are monitored with a T2 chart cation limits or some other measure of relative im-
(Fig. 2). The unnumbered observations in this chart portance is usually more meaningful.
were obtained by simulating normal operating condi- In practice, one rarely needs to compute all the k
tions; the numbered observations correspond to sim- principal components, since most of the variability in
ulated problematic operation caused by increasing the data is captured in the first few principle compo-
levels of fouling in the first zone of the reactor. The nents. The NIPALS algorithm [16] is ideal for com-
dashed line (- - -> corresponds to a 99% limit and puting the principal components in a sequential man-
the dotted line (. . . ) indicates the 95% limit. Notice ner when the number of variables is large. The
that the onset of fouling had an effect on the product, number of PCs that provide an adequate description
and that this effect could be detected in all cases, by of the data can be assessed using a number of
following the T* calculated from the product quality methods [11] with cross-validation [17] being per-
properties. Had we used univariate charts, points 52 haps the most reliable. By retaining only the first A
and 53 would have been missed. This is because, for PCs the Y matrix is approximated by:
these points, the individual values for the five Y
variables are within the expected limits of the corre- += &$ (6)
sponding univariate charts; however, the values of r=l

these five variables relative to each other are not In practice 2 or 3 PCs are often sufficient to explain
justified by the correlation structure of the Y matrix most of the predictable variations in the process.
(determined under normal operating conditions), and Having established a PCA model based on histori-
this was detected by the multivariate charts. cal data collected when only common cause varia-
tion was present, future behaviour can be referenced
2.2. Quality control charts based on principal com- against this ‘in-control’ model. New multivariate
ponents observations can be projected onto the plane defined
by the PCA loading vectors to obtain their scores zyxwvuts
When the number of measured quality variables = pTynew >, and the residuals enew = y,,, -
(ti,new

(k) is large, one often finds that they are highly 9”,,) where j,,, = PAtA,new, and tA,new is the (A X 1)
correlated with one another and their covariance vector of scores from the model and PA is the
matrix 2 is nearly singular. A common procedure (k X A) matrix of loadings. Multivariate control
for reducing the dimensionality of the quality vari- charts based on Hotelling’s T2 can be plotted based
able space is principal component analysis (PCA) on the first A PCs, where
[11,15,16]. The first principal component (PC) of y
Ti= 5;
is defined as that linear combination zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
t, =p :y that (7)
i=l I,
has maximum variance subject to 1p1 I= 1. The sec-
ond PC is that linear combination defined by t, = p: y and ~5 is the estimated variance of ti. If A = 2, a
which has next greatest variance subject to Ip 21 = 1, joint t, vs. t, plot can be used.
and subject to the condition that it be uncorrelated Note that the traditional Hotelling T* in Eq. (3) is
with (orthogonal to) the first PC (t,). Additional PCs equivalent [15,18] to
up to k are similarly defined. In effect PCA decom-
poses the observation matrix Y as: T+;+ 2 $
(8)
i=l f, i=A+ 1 ‘t,

Y = TPT = i tip: (5) By scaling each tf by the reciprocal of its vari-

i= zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
1 ance, each PC term plays an equal role in the
8 T. Kourti, J.F. MacCregor/Chemometrics and Intelligent Laboratory Systems 28 (1995) 3-21

computation of T2 irrespective of the amount of discussed. This use of only product quality data has
variance it explains in the Y matrix. This illustrates been the common approach to quality control meth-
some of the problems with using T2 when the ods developed throughout the statistical literature.
variables are highly correlated and z is very ill- However, in these approaches, all of the data on the
conditioned. When the number of variables (k) is process variables (X) are being ignored. If one truly
large, 2 is often singular and cannot be inverted. wants to do statistical process control (SPC), one
Even if it can, the last PCs (i =A + 1, . . , k) in Eq. must look at all of these process data as well. There
(8) explain very little of the variance of Y and are often hundreds of process variables, and they are
generally represent random noise. By dividing these zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJ
measured much more frequently and usually more
t,s by their very small variances, slight deviations in accurately than the product quality data (Y). Further-
these tjs which have almost no effect on Y will lead more, any special events which occur will also have
to an out-of-control signal in T*. Therefore, T,” their fingerprints in the process data (X). Sometimes
based on the first A (cross-validated) PCs provides a product quality is only determined by the perfor-
test for deviations in the product quality variables mance of the product later, in another process (i.e.,
that are of greatest importance to the variance of Y. catalyst conditioning); it would be useful to know if
However, monitoring product quality via T,’ based the product is good before using it; monitoring the
on the first A PCs is not sufficient. This will only process would help detect problems during produc-
detect whether or not the variation in the quality tion that may lead to a questionable product.
variables in the plane of the first A PCs is greater There are several other reasons why monitoring
than can be explained by common cause. If a totally the process is advantageous. Sometimes, only a few
new type of special event occurs which was not properties of the product are measured, but these are
present in the reference data used to develop the not sufficient to define entirely the product quality.
in-control PCA model, then new PCs will appear and For example, only the relative viscosity (RV) is
the new observation y,,, will move off the plane. charted in nylon production, although there are other
Such new events can be detected by computing the properties (amine end groups) that affect the dye
squared prediction error (SPE,) of the residuals of a properties of the product. If process problems that
new observation [ 191. zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
affect amine groups occur, they will not be detected
by following the RV only. In these cases the process
SE, -kv,r)2
= t (Y,,,,, (9) data may contain more information about events
i=l with special causes that may affect the product qual-
ity (product performance).
This is also often referred to as the Q statistic [ll] or
Finally, even if product quality measurements are
distance to the model. It represents the squared
frequently available, monitoring the process may
perpendicular distance of a new multivariate obser-
help in diagnosing assignable causes for an event.
vation from the projection space. When the process
When monitoring product quality, even if we deter-
is ‘in-control’, this value of SPE, or Q should be
mine which quality variable caused the multivariate
small. Upper control limits for this statistic can be
chart to go out of limits, it may still be difficult to
computed, from historical data, using approximate
determine what went wrong in the process. For
results for the distribution of quadratic forms [ 11,201.
example, in the LDPE process by following the
A very effective set of multivariate control charts is
product variables it was determined that for point 56,
therefore a T2 chart on the A dominant orthogonal
SCB is the major contributor to the out-of-control
PCS (t,, . ..) tA) plus a SPE, chart.
signal. However, there may be several reasons (com-
binations of process conditions) that might have
caused this property to change. Monitoring the pro-
3. Multivariate methods for process monitoring cess would bring us closer to the answer as will be
demonstrated later.
SO far, statistical quality control (SQC) methods Certainly one could apply the previously dis-
based only on product quality data (Y) have been cussed SQC charting methods directly to the process
 T. Kourti, J.F. MacGregor/Chemometrics and Intelligent Laboratory Systems 28 (1995) 3-21 9

variables (X) as well [18]. However, as discussed more, they offer no way of relating the X and Y
previously, with large numbers of highly correlated data, and least squares regression analysis is also
variables, these methods are impractical. Further- impractical in this situation. Another problem is that

t1

X-block scores

t3

.. .... .. . . . -

-1OL
-10 -5 0 5 10 15 20 2s

Fig. 3. PLS score plots for 15 days of operation of the catalytic cracking and fractionation section of a refinery (top: II vs. t,; bottom: t, VS.
t,).
IO T. Kourti, J.F. MacGregor/ Chemometrm and Intelligent Laboratory Systems 28 (199.5) 3-21

these methods cannot handle missing data; sensor being collected and stored in data bases. Very little
failure is a common problem in the process indus- analysis and interpretation of these data are being
tries. The only practical approaches to multivariate performed because of the overwhelming size of the
SPC appear to be those based on multivariate statisti- data bases and because of the very ill-conditioned
cal projection methods such as PCA and PLS (pro- nature of the routine operating data being collected.
jection to latent structures or partial least squares). Furthermore, the signal to noise ratio is often poor in
The methods are ideal for handling the large number these data, and there are often significant amounts of
of highly correlated and noisy process variable mea- missing data. However, all these problems are well
surements that are being collected by process com- addressed by the multivariate statistical projection
puters on a routine basis; these methods can also methods of PCA and PLS. By examining the be-
handle missing data. PCA has already been de- haviour of the process data in the projection spaces
scribed, and a brief overview of PLS follows. defined by the small number of latent variables (t,,
t 2, ..., t,), and interpreting process movements in
3.1. PLS - partial least squares this reduced space by examining the corresponding
space defined by the loading vectors (p,, pz, . . . ,
p,), or (w,, w2, . . . . wA) in the case of PLS, it is
Given two matrices, an (n X ml process variable
often possible to extract very useful information
data matrix X, and an (n X k) matrix of correspond-
from these data bases, and to use this information to
ing product quality data Y, one would like to extract
improve the process. Some early notable attempts at
latent variables that not only explain the variation in
using these approaches for the analysis and interpre-
the process data (X), but that variation in X which is
tation of data bases are the works of Denney et al.
most predictive of the product quality data (Y). PLS
[23] on a sulphur recovery unit, and Moteki and Arai
is a method (or really a class of methods) which
[24] on a low-density polyethylene process. The
accomplishes this by working on the sample covari-
latter work is particularly notable in that the analysis
ante matrix (XTY)(Y TX). In the most common
was able to lead them quickly to process conditions
version of PLS [21,22], the first PLS latent variable
that yielded desired new lamination grades and injec-
t, = wTx is that linear combination of the x vari-
tion grades of polyethylene.
ables that maximizes the covariance between it and
Slama [2.5] used PCA and PLS to analyze data on
the Y space. The first PLS loading vector w, is the
more than 300 process variables and 11 product
first eigenvector of the sample covariance matrix
grades from the fluidized bed catalytic cracking and
XTYYTX. Once the scores t, = Xw, for the first
fractionation section of a refinery. The difficulty
component have been computed the columns of X
with such massive data sets is first to find out where
are regressed on t, to give a regression vector
in the data there is useful information. The projec-
p, = Xt,/tTt, and the X matrix is deflated to give
tions of hourly average data from fifteen days of
residuals X2 = X - t, PT. The second latent variable
continuous operation into the planes t,-t, and t,-t,
is then computed as t, = wzx where w? is the first
defined by the first three latent variables are shown
eigenvector of Xl YY TX 2 and so on. As in PCA the
in Fig. 3. The data appears to cluster into about five
new latent vectors or scores (t,, t2, ) and the
distinct regions, operating in a stable manner for
loading vectors (w,, w2, . . > are orthogonal. For
several days at each condition before shifting to
large ill-conditioned data sets, it is usually conve-
another region. There is very little information about
nient to calculate the PLS latent variables sequen-
the process within each stable data cluster. However,
tially via the NIPALS algorithm [21] and to stop
by focusing attention on the transitions between the
based on cross-validation criteria.
regions at time periods 58-59, 76-77, 110, 197 and
212-213 we can probably learn most of what there
3.2. Analysis of historical process data sets is to know about the 15day period of operation. To
help diagnose the reasons for these shifts in process
With process computers hooked up to most indus- operation, one can interrogate the underlying multi-
trial processes, massive amounts of process data are variate model (as discussed below in Section 3.4)
 T. Kourti, J.F. MacGregor/ Chrmometrics and Intelligent Laboratory Systems 28 (1995) 3-21 11

and display the process variable contributions to these methods to analyze process data. Wise et al.
these shifts. [26] applied PCA to analyze and diagnose systematic
There are several interesting examples of using variations in the behaviour of a slurry-fed ceramic zyxwvutsr

FOULING DETECTION IN LDPE REACTOR

I61
14-

12-
_ UCL, (99.0 %)

3
8- -
10 --------------------------- j---

0
0 10 20 30 40 50 60

OBSERVATION NUMBER
450

0
0 10 20 30 40 50 60

OBSERVATION NUMBER

Fig. 4. T’ chart on three PLS scores and squared prediction error (SPE,) chart for monitoring the LDPE process. Points 52-56 denote a
period where fouling occurred in the first zone of the reactor.
12 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
T. Km&, J.F. M acGregor/ Chemometrics and Intelligent Laboratory Sysmns 28 (1995) 3- 21

melter process. Skagerberg et al. [27] applied PLS to priate reference set is chosen which defines the
predict polymer properties from measured tempera- normal operating conditions for a particular process.
ture profiles in a tubular low-density polyethylene In other words, a PCA or PLS model must be built
reactor and to interpret the behaviour of this process. based on data collected from various periods of plant
Hodouin et al. [28] used PCA and PLS to analyze operation when performance was good. Any periods
and interpret the behaviour of mineral flotation and containing variations arising from special events that
grinding circuits in a large mineral processing plant. one would like to detect in the future are omitted at
Dayal et al. [29] used PLS to model the dynamic this stage. The choice of the reference set is critical
behaviour of a continuous Kamyr digester in a pulp to the successful application of the procedure as
mill, and diagnosed the reasons for poor control of discussed in Kresta et al. [19].
Kappa number by examining the loading plots (w,, When the data are serially autocorrelated the X
w*). and Y matrices can be augmented with time-lagged
values, in order to account for the dynamics of the
3.3. Monitoring continuous processes
process and the disturbances. Multivariate time series
Although the analysis of historical data bases is analysis is discussed for PCA by Jollife [30], and
an important first step towards process improvement, Jackson [l l] and for PLS by Wold et al. [xl] and
establishing multivariate control charts to detect spe- MacGregor et al. [32]. Dead times between variables
cial events as they occur, and to diagnose possible are accounted for by time shifting. An industrial
causes for them while the information is fresh, is an example where plant data had been both time shifted
essential part of SPC. The philosophy applied in to account for dead times between X and Y, and
developing multivariate SPC procedures based on lagged to account for autocorrelations in Y, is de-
projection methods is the same as that used for the scribed in Dayal et al. [29].
univariate or multivariate Shewhart charts. An appro- The multivariate control chart is now a T’ chart

;I.-
zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

ci- -
1 in-1
5 )-

I
4 T ut_l

-1
u
-2.-
u 2. 4 6 8 10 12 14 16
PROCESS VARIABLES
Fig. 5. Contribution plot showing the process variable contributions to the SPE,, for point 52
 T. Kourti, J.F. M acGrqpr/ Chemometric.s und Intelligent Laboratory Sy.wrns 28 (1995) 3-21 13

on the first A latent variables (Eq. (7)). Added to alarmed an out-of-control situation, on-line, before
this, is a chart on SPE, where laboratory data on product quality became available.
The T’ plot signalled later. As already discussed, the
m
SPE,= c (~“W,,--LJ
I= I
(10)
two plots are complementary in detecting special
events; both of them are required for proper monitor-
ing.
where f,,, is computed from the reference PLS or
PCA model. This latter plot will detect the occur-
3.4. Diagnosing assignuhle causes
rence of any new events which cause the process to
move away from the hyperplane defined by the
Both univariate and multivariate SPC charts are
reference model. Control limits for the T’ charts are
based on statistical tests to detect any deviations
chosen in the same manner as previously discussed,
from the in-control reference distribution upon which
and the UCL on SPE., is based on the x2 approxi-
the models and charts have been built. In classical
mation (Q statistic [ 11 JO]).
quality control approaches which chart only quality
The main concepts behind the development and
variables (Y), once an out-of-control signal has been
use of these multivariate SPC charts for monitoring
given, it is then left up to the process operators and
continuous processes were laid out by Kresta et al.
engineers to try to diagnose an assignable cause
[19], Wise et al. [26], Wise and Ricker [33], and
using their process knowledge and a one-at-a-time
MacGregor et al. [34,35]. Several illustrations of the
inspection of process variables. However, multivari-
methods were also presented in those papers along
ate charts based on PLS or PCA provide a much
with the algorithms and details on estimating control
greater capability for diagnosing assignable causes.
limits.
By interrogating the underlying PLS or PCA model
To illustrate the basic approach, consider the
at the point where an event has been detected, one
monitoring of the simulated multi-section high-pres-
can extract diagnostic or contribution plots which
sure tubular reactor process for the manufacture of
reveal the group of process variables making the
low-density polyethylene (LDPE) [14]. The reaction
greatest contributions to the deviations in the SPE,
kinetics and the fundamental modelling of this LDPE
and the scores [33,34,37]. Although these plots will
process can be found in a review by Kiparissides et
not unequivocally diagnose the cause, they will pro-
al. [36]. Measurements are available on a frequent
vide much greater insight into possible causes and
basis on all process variables (X) - reactor temper-
thereby greatly narrow the search.
ature profiles in each section, feed rates on all com-
Consider the out-of-control alarms shown in Fig.
ponent streams, cooling system flows and tempera-
4 for the LDPE process. Diagnostic plots showing
tures, and pressures in each reactor section. Every
the contribution of the process variables to the SPE.,
hour or so, measurements are available on product
at point 52 are shown in Fig. 5. These contribution
quality and productivity (Y) - polymer molecular
plots point to the temperature of the reaction mixture
weights and branching properties, and conversion of
at the exit from zone 1 and the temperature of the
monomer to polymer. Using data collected (X, Y)
cooling agent into the jacket of the first zone as
when the process was operating well, and no special
being the main process variables that are showing
events were present, a PLS model using only three
inconsistency (by contributing significantly to the
latent variables (A = 3) was able to explain 90.0%
large values of SPE i 1. This combination of variables
of the variation in the Y data. Fig. 4 illustrates the
would imply heat transfer problems and could lead
use of a T’ chart (on three PLS scores) and an SPE,
the operator to suspect fouling.
chart to monitor the behaviour of the reactor when
there is an increasing level of fouling in the first
section of the reactor. Unnumbered points indicate 3.5. Multi-block PLS
past conditions of normal operation. Fouling starts at
point 52. Notice that the squared prediction error plot When a large number of variables is included in
quickly detected the onset of this special event and the X space, the monitoring and diagnosing charts
14 T. Kourti, J.F. MucGregor/ C‘hemomrtric,s and Intelli~rnt Lohorator~ $y.stems 2X (199.5) S-21 zyxwvutsrqponmlkjihgfed

FOULING DETECTION: ZONE 1

(a’ 351
56

25 - zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

0 60

OBSERVATION NUMBER
CL
450 lr
r
5

r’
400

4
350

300

250

200

150
13

100

50

0
10 20 30 40 50 60

OBSERVATION NUMBER
Fig. 6. Monitoring the LDPE process using multi-block PLS. The disturbance is fouling in zone 1. Monitoring of individual zones detects
the problematic zone. (a) T’ chart on three MB-PLS scores and SPE, for the process variables of zone 1. (b) T’ chart on three MB-PLS
scores and SPE I for the process variables of zone 2.
 T. Kourti, J.F. M acGregor / Chemomrtrics und Intelligent Laboratory Systems 28 (I 995) 3-21 zyxwvutsrqponmlkjihgf
15

b) zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
FOULING DETECTION: ZONE 2

lo- zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

a- zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA

6 t

OBSERVATION NUMBER

125
lo-
99.0 %
_.--____.--____---___

a-

6-

4-

0
0 10 20 30 40 50 60

OBSERVATION NUMBER

Fig. 6 (continued).

discussed in the previous sections may be difficult to task. In the MB-PLS approach, large sets of process
interpret. The combined use of multi-block PLS variables (X) are broken into meaningful blocks;
(MB-PLS) and contribution plots may facilitate this usually each block corresponds to a process unit or a
section of a unit. Multivariate monitoring charts for problems were detected in zone 2 for observations
important subsections of the plant as well as for the 52-56. MB-PLS successfully detected that the proh-
entire process can be constructed. The principles lem is in zone 1 and that zone 2 operates normally.
behind multi-block data analysis methods and their Utilizing the contribution plots for fault diagnosis
algorithms can be found in Wold [38] and Wangen (on the scores and SPE, of zone 1) has revealed that
and Kowalski [39]. MacGregor et al. [ 141 discuss an process variables with unusual values were the tem-
application of MB-PLS to process monitoring and perature of the reacting mixture at the exit of zone 1
diagnosing for the LDPE reactor. Each block corre- and the temperature of the cooling agent in zone 1.
sponds to one zone. Plots of t, vs. t2 and SPE, Although the monitoring and diagnosis procedures
obtained for each block of the process were utilized based on MB-PLS and PLS gave comparable results
to detect an abnormal event in the zone it occurred; for this system with only 14 process variables, MB-
then contribution plots were successfully used to PLS offers an advantage when larger systems with
assign causes for it. When the number of latent tens or hundreds of variables are involved.
variables used for modelling is more than two, one
should combine the information of the scores in a 3.6. Mmitorir~g batch proccwrs
statistic, rather than plotting scores pair-wise. The
following example demonstrates a monitoring proce- Recent trends in most industrialized countries have
dure utilizing a chart of T’ calculated from the been towards the manufacture of higher value added
scores used for the MB-PLS model and a SPE., specialty chemicals (specialty polymers, pharmaceu-
chart. ticals and biochemicals) that are produced mainly in
Fig. 6a gives the zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
T’ chart on three scores (calcu- batch reactors. There are also many other batch type
lated from scores t,, t, and t3 of block 1) and the operations, such as crystallization and injection
SPE, chart for block 1 (corresponding to zone 1) for molding, which are very important to the chemical
the same simulated process conditions of Fig. 4 and manufacturing industries. Monitoring these batch
(fouling in zone 1). Notice that by monitoring block processes is very important to ensure their safe oper-
1 (zone 1, only) problems are detected for observa- ation and to assure that they produce consistent
tions 52-56. Fig. 6b gives the T’ chart on three high-quality products. The use of the multivariate
scores and the SPE, chart for block 2 (zone 2), for statistical projection methods has been extended to
the same simulated fouling case. Notice that no the analysis and the on-line monitoring and diagnosis

mode space operational space quality space

initial conditions On-line measurements quality measurements

Fig. 7. Nature of batch data. The batch process is described by X, quality variables by Y and feed propcrtics by Z
 T. Kourti, J.F. MacGregor/ Chemometrics and Intelligent Laboratory Systems 28 (19%) 3-21 17

of batch processes by MacGregor and Nomikos [40] ing the trajectories of all the variables from a single
and Nomikos and MacGregor [20,41,42]. Typical batch. Each of its vertical slices is a (I X .I) matrix
data from batch processes include time-varying tra- representing the values of all the variables for all the
jectories of all the measured process variables batches at a common time interval (k). The final
throughout the duration of each batch (X), product product quality measurements are taken at the end of
quality measurements (Y) at the end of each batch, each batch, for a few variables, L. These are summa-
and batch recipe and charge conditions (Z) at the rized in the (I XL) matrix Y. For each batch, mea-
start of each batch. If such data are available in a sured feed-stock properties and other variable initial
historical data base on many past batches, multivari- conditions may be available; these are summarized
ate PCA and PLS models can be developed for in a matrix Z.
analyzing these historical batches and for establish- Since the process data (X) are now a three-dimen-
ing on-line SPC charts for monitoring the progress of sional array (batch run X variable X time), Nomikos
each new batch. and MacGregor used three-dimensional or multi-way
The nature of the data available in a batch moni- PCA (MPCA) and PLS (MPLS) methods. Multi-way
toring problem is illustrated in Fig. 7. The X matrix PCA and PLS methods have been discussed in a
is a (I zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
X J X K) array, where I is the number of series of articles [43-461. Nomikos and MacGregor
batch runs, J is the number of variables and K is the proposed approaches for handling the fact that one
time intervals throughout the batch. Each horizontal dimension (time) is evolving during the progress of a
slice through this array is a (J X K) matrix contain- new batch, and for establishing control limits on the zyxwvuts

1
P

Fig. 8. Plots of t,-t2 and f2-t, for 30 hatches. Batch No. 6 is abnormal.
18 T. Kourti, J.F. MacGregor / C‘hemometrics and Intelligent Laboratory Systems 28 (1995) 3-21

multivariate SPE and score plots. They transformed

the three-dimensional array X to a two-dimensional
array by unfolding X in such a way as to put each of
its vertical slices (I zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
X J) side by side to the right,
starting with the one corresponding to the first time
interval. The resulting two-dimensional matrix has
size (I X JK ). This unfolding allows for analyzing
the variability among the batches in X by summariz-
ing the information in the data with respect both to
variables and their time variation. With this particu-
lar unfolding, by subtracting the mean of each col-
umn prior to performing the MPCA, one is decom-
posing the variation about the mean trajectories of all
the variables.
The MPCA approach classifies batches as good or
bad based on their similarity to a group of previous
batches that produced an acceptable product. Infor-
mation from quality measurements is not utilized
directly. MPLS may be used to utilize information
from the product quality. Once the X matrix has
been unfolded into a two-dimensional matrix, PLS
can be performed between Y and this new matrix, to
relate the quality characteristics to the process condi-
tions. By utilizing the quality measurements the
batches may be classified in a way that they are
more predictive of Y - in this case variables that
exhibit high variability but do not affect the quality TlME

of the product are weighted less heavily; as a result, Fig. 9 Monitoring a good batch. SPE, and D statistic
disturbances in these variables will be flagged but
not cause unnecessary alarms [42]. When extra infor-
mation relevant to the batch process is available (in and ‘bad’ batches with the available process data; in
the form of matrix Z in Fig. 7) this information may other words to assess if the system was obserllable.
also be utilized, by performing multi-way multi-block Fig. 8 shows the projections of these 30 batches on
PLS. Matrix Z and the unfolded X matrix may be the score planes (t,-t2 and tz-t,> defined by the
treated as two blocks, weighted appropriately. three first principal components. It can be seen that
The use of MPCA to monitoring batch processes batch No. 6 (the one characterized by the company
is illustrated here with an example. Data from 30 as ‘bad’) is out of the main cluster (normal operating
batches from an industrial process were provided. region) formed by the rest of the batches.
There were no product quality measurements; the Having established the observability of faults with
quality of the batch (‘good’ or ‘bad’) was assessed the analysis of past data, a model was built to
from the performance of the batch product in another summarize the information contained in the 29 good
process, later. For each batch, the trajectories of 4 batches about the normal operating region of the
variables for 375 time intervals were provided. One process. This model was then used as statistical
of the batches was characterized as ‘bad’ by the reference to classify new batches as normal (‘good’)
company. In a preliminary analysis, MPCA was or abnormal (‘bad’). The model was used for the
performed on all the batches (i.e., on the three-way classification of new batches in the way described in
array X with dimensions 30 X 4 X 375), to test if the Nomikos and MacGregor [41]. New batches are clas-
method would be able to discriminate between ‘good’ sified by monitoring a statistic, D (essentially a
 T. Kourti, J.F. MrrcCregor/Chemometrics and Intelligent Luhoratory Systems 28 (1995) 3-21 19

I
40 20 0 20 40

. ..
G

40 . . .

30 f$jg
, .

l3
_,..- :-._ . .
20 - . :
,;’ .. .i.
. . _ *A-. ‘..,
IO- ,:’ _ . .. . . ,,
;.q .. :;
t o- .:y.. :

-10

-2 0
I
t&y .
;~-----; -3 ol-

-40
1
3 50 100 150 200 250 300 350 400 -60 40 zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPON
zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
-20 0 zyxwvutsrqponmlkjihgfedcb
20 40 a
T lM E t3

Fig. 10. Monitoring batch No. 6 (a bad batch). (a) SPE, and D statistic. (b) Plots of f,-/? and f?-t, for the duration of the batch.

Hotteling zyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA
T2 calculated from the A latent variables points out of 375 are out of the 95% limit in the
at each time interval k), and by monitoring SPE at SPE, while the D statistic goes out of limits around
each time interval k. 100 min into the batch run. Plots of t, vs. t2 and t,
Fig. 9 shows the SPE response as a function of vs. t, (Fig. 1Ob) reveal that mainly t, and t, scores
time, and the D statistic of a batch that was eventu- show abnormalities. Indeed, individual plots of t3
ally classified as ‘good’. Notice that both of these and t, (not shown here) revealed that these latent
quantities remain well within the control limits variables were out of limits after 100 min into the
throughout the batch. (The solid line corresponds to run.
a 99% limit and the dashed line to a 95% limit for The proposed monitoring charts are in accordance
SPE; at the top of the figure it is indicated that for with the SPC requirements in that they can be easily
the current batch, 15 points, out of 375, were out the displayed and interpreted, and they can quickly de-
95% limit and one point out the 99% limit.) Fig. 10 tect a fault. Furthermore, it is also possible to pro-
shows how batch No. 6 would have behaved, had the vide the operators with diagnostic information by
model been in use on-line, when the data for this interrogating the underlying MPCA, MPLS or multi-
batch were becoming available. Fig. 10a shows the way multi-block PLS model. Other industrial appli-
SPE behaviour and the D statistic. Notice that 215 cations of these methods have been reported for the
20 T. Korrrfi, J.F. M acGregor / Chemometrics and lntelligtmt Laboratory Systems 28 (1995) 3- 21

analysis of historical batch data bases by Kosanovich [3] J.S. Hunter, Exponentially weighted moving average. Jour-
nal Qualiry Technology, 18 (1986) 203-210.
et al. [47], and for the monitoring of a batch poly-
[4] S.J. Wierda, Multivariate statistical process control - recent
merization by Nomikos and MacGregor [20].
results and directions for future rcscarch. Sfatistica Neer-
landica, 4X ( 19941.
[5] R.S. Sparks, Quality control with multivariatc data, Aus-
4. Summary tralian Journal of Slarisrics. 34 (lYY2) 37%3YO.
[6] H. Hotelling, Multivariate quality control, illustrated by the
air testing of sample bombsights, in C. Eisenhart, M.W.
This paper has provided an overview of the con- Hastay and W.A. Wallis (Editors), Technique.r of Stnrisfical
cepts behind multivariate statistical process control. Analy .xis, McGraw-Hill. New York. 1947, pp. 1133184.
Justifications for treating the data in a truly multi- [7] F.B. Alt, Economic Design of Control Churtv for Correlated,
variate manner are given. To genuinely do multivari- M ulril,ariare Ohsrr~~arions,Ph.D. Dissertation, Georgia Insti-
tute of Technology, lY77.
ate statistical process control (SPC) one must utilize
[8] F.B. Alt, Multivariate quality control, in S. Kotz and N.L.
not just the final product quality data (Y), but all the Johnson (Editors), Ency clopedia of Statistical Sciencec, Wi-
data on process variables (X) being collected rou- ley, New York, 1985, Vol. 6. pp. 110-122.
tinely by process computers. SPC approaches based [Y] F.B. Alt and N.D. Smith, Multivariate process control, in
on multivariate statistical projection methods (PCA P.R. Krishnaiah and C.R. Rao (Editors), Handbook ofSratis-
/ic.s, North-Holland, Amsterdam, 1088, Vol. 7. pp. 3333351.
and PLS) have been developed for this purpose. The
[lo] T.P. Ryan. Sratlsrical M ethods for Qua/iv Improt rment,
ideas behind these new approaches and the literature Wiley, New York, 1989.
on them is reviewed. Multivariate control charts in [ll] J.E. Jackson, A User’s Guidr lo Principal Components,
the projection spaces provide powerful methods for Wiley, New York, 1991.
both detecting out-of-control situations, and for diag- [12] C. Fuchs and Y. Bcnjamini, Multivariate profile charts for
statistical process control, Technometrics. 36 (1994) 182-195.
nosing assignable causes, and they are applicable
[13] N.D. Tracy, J.C. Young and R.L. Mason, Multivariate con-
both to continuous and batch processes. The only trol charts for individual observations. Journal of Qualify
requirement for applying these methods is the exis- Technology , 24 (1992) 8X- 95.
tence of a good data base on past operations. For this [14] J.F. MacGregor, C. Jacckle, C. Kiparissides and M. Koutoudi,
reason, they have attracted wide interest, and are Monitoring and diagnosis of process operating performance
by multi-block PLS methods with an application to low
rapidly being applied in many industries. Recent
density polyethylene production, AfChtl Journal. 40 (1994)
advances in the traditional multivariate SPC methods
826-838.
for monitoring and diagnosing process operating per- [ 151 K.V. Mardia, J.T. Kent and J.M. Bibby, M ulli~~ariarrAnaly-
formance are reported and compared to projection sis, Academic Press, London, 1989.
method approaches in Kourti and MacGregor [48]. [16] S. Wold, K. Esbcnsen and P. Gcladi, Principal component
analysis, Chemometrics and Intelligent Laboratory Systems,
2 (1987) 37- 52.
[I71 S. Wold, Cross-validatory estimation of the number of com-
Acknowledgements ponents in factor and principal components model, Techno-
metric.r, 20 (1978) 397- 405.
The authors wish to thank the 15 companies of [1X] T. Kourti and J.F. MacGregor. Multivariate SPC methods for
monitoring and diagnosing process performance, in En Sup
the McMaster Advanced Control Consortium for their
Yoon (Editor), PSE ‘94, Fifth Internationul Symposium on
financial support. Also, for their active involvement Process Systems Engineering, Preprints, lY94, Vol. II, pp.
in joint research projects which provide the opportu- 739-746.
nity to test new research ideas in industry. [lY] J. Kresta, J.F. MacGregor and T.E. Marlin, Multivariate
statistical monitoring of process operating performance,
Canadian Journal of Chemical Engineering, 69 ( 190 1) 3S-
47.
References [2O] P. Nomikos and J.F. MacGregor, Multivariate SPC charts for
monitoring batch processes, Technometrics, 37( 1) (199.5).
[l] W.A. Shewhart, Economic Control of Quality of M anufac- [211 P. Geladi and B.R. Kowalski. Partial least-squares regres-
tured Product, Van Nostrand, Princeton, NJ, 1931. sion: a tutorial, Analytrca Chimica Acfa, 185 (1986) 1-l 7.
[2] R.H. Woodward and P.L. Goldsmith, Cumulatrr,e Sum Tech- [22] A. Hoskuldsson, PLS regression methods, Journal of
niques, Oliver and Boyd, London, 1964. Chemomrtric~, 2 (1988) 21 l- 228.
 T. Kourti, J.F. MacGregor/ Chrmometrics and Intelligent Laboratory Systems 2X ~lUU_‘il 3-21 21

[23] D.W. Dcnncy, J. MacKay, T. MacHattic, C. Flora and E. plied to low density polyethylene reactors. in K. Najim and
Mastracci. Application of pattern recognition techniques to J.P. Barbary (Editors). IFAC Internutionul Symposium, AD-
process unit data. Cunudu~n Socwt) of Chemud Engineering CHEM’91 Proceedings, Paul Sabatier University, Toulouse,
Confcrow, Sarnia, Ontario, Canada, 1985. 1991, pp. 131-13s.
[24] Y. Moteki and Y. Arai, Operation planning and quality [36] C. Kiparissidcs, G. Vcrros and J.F. MacGregor, Mathemati-
design of a polymer process, Proceedings, IFAC Symposium, cal modelling, optimization and control of high prcssurc
DYCORD-86, Bournemouth, UK, 1986, pp. lS9- 166. cthylcnc polymerization reactors, Journal of Mucromolecu-
[2S] CF. Slama, Multirariute rtatistrcul unalysis of dutu obtained lar Science, Rer,iew.s in Macromoleculur Chemrstry, C33
from on industriul fluid&d catalytic process u.ping PCA and (1993) 437.
PLS, M.Eng. Thesis, Department of Chemical Engineering, [37] P. Miller, R.E. Swanson and C.F. Heckler, Contribution
McMaster University, Hamilton, Ontario, Canada. 1Y9 1. plots: the missing link in multivariate quality control, 37th
[26] B.M. Wise, D.J. Veltkamp, N.L. Ricker, B.R. Kowalski, S. Annual Fall Conference ASQC, Rochester, NY. 1993.
Barnes and V. Arakali, Application of multivariatc statistical [3X] H. Wold, Soft modelling. The basic design and some cxtcn-
process control (MSPC) to the West Valley slurry-red cc- sions, in K.G. Jorcskog and H. Wold (Editors), Systems
ramic melter process, in Post and Wacks (Editors), Waste under Indirect Obserwtion. Vol. 2. North-Holland, Amster-
Manugement ‘YI Proceedings, University of Arizona Press, dam, lY82, Chap. 1.
Tucson, AZ, 1991. [39] L.E. Wangcn and B. Kowalski, A multiblock partial least
[27] B. Skagerberg, J.F. MacGregor and C. Kipariasidcs, Multi- squares algorithm for investigating complex chemical sys-
variate data analysis applied to low-density polyethylene tems, Journal of Chcmomrtric.s, 3 (1988) 3-20.
reactors, Chemometrics and Intelligent Laboratory Systems, [40] J.F. MacGregor and P. Nomikos, Monitoring batch pro-
I4 (1992) 341-356. cesses, in Reklaitis. Rippin, Hortacsu and Sunol (Editors),
[28] D. Hodouin, J.F. MacGregor. M. Hou and M. Franklin, NATO Adwnced Study Institute (ASI) for Batch Processing
Multivariate statistical analysis of mineral processing plant Systems Engineering, May 29-June 7, 1992, Antalya,
data. CIM Bulletin of Mmerul Processing. 86 (1993) 23-34. Turkey, Springer.
New York.
[29] B. Dayal, J.F. MacGregor, P.A. Taylor, R. Kildaw and S. [41] P. Nomikos and J.F. MacGregor, Monitoring and batch
Marcikic, Application of fccdforward and neural networks processes using multi-way principal component analysis,
and partial least squares regression for modelling kappa AlChE .Jourrml, 40 (1004) 1361-1375.
number in a continuous Kamyr digester, Pulp and Paper [42] P. Nomikos and J.F. MacGregor, Multi-way partial least
Cunadu, 9.5 (1994) 26-32. squares in monitoring batch proccsscs, First international
[30] I.T. Jollife, Principul Component Analysts, Springer, New Chemometric~ Internet Conference, Scptcmber 1904, to be
York, 1986. published in Chemometrics and Intelligent Laborutory Sys-
[31] S. Wold, C. Albano, W.J. Dunn III, U. Edlund, K. Esbensen. tems.
P. Geladi, S. Hellberg, E. Johanxson, W. Lindberg and M. [43] J. LBhmoller and H. Wold, presented at the European Meet-
Sjlintrlim, Multivariate data analysis in chemistry, in B. ing of P.sychometrics Society. Groningen, The Netherlands,
Kowalski (Editor), Chemometrics. Muthemuticc and Statis- 1Y80.
tics in Chemistry, Reidel, Dordrecht. 1984, pp. 17-95. [44] S. Wold, P. Geladi, K. Esbcnscn and J. Ohman, Multi-way
[32] J.F. MacGregor, T. Kourti and J. Kresta. Multivariate identi- principal components and PLS analysis, Journal of Chemo-
fication: a study of several methods, in K. Najim and J.P. metricc. 1 (1987) 41-56.
Babary (Editors), IFAC International Symposium, AD- [4S] P. Geladi, Analysis of multi-way (multi-mode) data, Chemo-
CHEMYZ, Proceedings, Paul Sabatier University, Toulouse, metrics and Intelligent Laboratory Systetns. 7 (1989) I l-30.
1991, pp. 369-375. [46] A.K. Smilde and D.A. Doornbos, Three-way methods for the
[33] B.M. Wise and N.L. Ricker, Recent advances in multivariate calibration of chromatographic systems: comparing
statistical process control: improving robustness and sensitiv- PARAFAC and three-way PLS. Journal of Chemometrics, 5
ity, in K. Najim and J.P. Barbary (Editors), IFAC Intrrna- (1991) 345-360.
tionul Sympo.sium, ADCHEMYI, Proceedings, Paul Sabaticr [47] K.A. Kosanovich, M.J. Piovoso, KS. Dahl, J.F. MacGregor
University, Toulouse, 1991, pp. 12.5-130. and P. Nomikos, Multi-way PCA applied to an industrial
[34] J.F. MacGregor, T.E. Marlin, J. Krcsta and B. Skagerberg, batch process, Proceedings of the American Control Confer-
Multivariate statistical methods in process analysis and con- ence, Baltimore, MD, 1994.
trol, AIChE Symposium Proceedings of the Fourth Interna- [48] T. Kourti and J.F. MacGregor, Recent developments in mul-
tional Conference on Chemical Process Control, AIChE tivariate SPC methods for monitoring and diagnosing process
Pub]. No. P-67, New York, 1991, pp. 79-99. and product performance, submitted to the Journal of Qual-
[35] J.F. MacGregor, 8. Skagerberg and C. Kiparissides, Multi- ity Technology, 1994.
variate statistical process control and property inference ap-