Multivariate Statistical Monitoring of Process Operating

Performance
JAMES V. KRESTA, JOHN F. MAcGREGOR and THOMAS E. MARLIN

Chemical Engineering Department, McMaster University, Hamilton, Ontario L8S 4L7

Process computers routinely collect hundreds to thousands of pieces of data from a multitude of plant sensors every
few seconds. This has caused a “data overload” and due to the lack of appropriate analyses very little is currently
being done to utilize this wealth of information. Operating personnel typically use only a few variables to monitor the
plant’s performance. However, multivariate statistical methods such as PLS (Partial Least Squares or Projection to
Latent Structures) and PCA (Principal Component Analysis) are capable of compressing the information down into
low dimensional spaces which retain most of the information. Using this method of statistical data compression a mul-
tivariate monitoring procedure analogous to the univariate Shewart Chart has been developed to efficiently monitor
the performance of large processes, and to rapidly detect and identify important process changes. This procedure is
demonstrated using simulations of two processes, a fluidized bed reactor and an extractive distillation column.

Les ordinateurs de procedes recueillent de manikre rdpktitive jusqu’k des milliers de donnees a partir d’une multitude
de capteurs dans I’usine en I’espace de quelques secondes. Cela a entrain6 une <<surcharge de donndew et en raison
de I’ahsence d’analyses appropriees, peu d’efforts sont faits pour exploiter cette masse d’informations. Le personnel
d’exploitationn’utilise gkndraleinent que quelques variables pour superviser la performance de I’installations. Cepcn-
dant, les methodes de statistiques multivarides telles que les methodes PLS (moindres carrEs partiels ou projection de
structure latentes) et PCA (analyse du composant principal) permettent de comprimer I’information dans un espace
minimum qui retient la majorite de I’information. En utilisant cette rnethode de compression de donndes statistiques.
o n a mis au point une procedure de contrhle multivarii. analogue au graphique de Shewart univarid afin de superviser
de facon efficace la performance de prockdds majeurs. et d’en detecter rapidement les fluctuations. On illustre cette
prockdure I’aide de simulations de deux prockdds. soit un reacteur i lit fluidisd et une colonne de distillation extractive.

Keywords: statistical process control. partial least squares. projection to latent structures. principal component
analysis, performance monitoring. fault detection.

M onitoring performance and detecting faults is an

integral part of the successful operation of any
procesc. Performance can be monitored by comparing the
actual results to the predictions from a mechanistic model,
or by using statistical process control (SPC) charts (e.g.
Shewan charts (Shewart, 1931), CUSUM charts (Woodward
and Goldsmith, 1964) or EWMA charts (Huntcr, 1986)) to
compare the current state of the process against “normal
operating conditions”. The challenge currently facing
process monitoring is the enormous amount of correlated data
being collected from a multitude of sensors every few
seconds. minutes or hours. This “data overload” and the
lack of appropriate analytical tools has meant that very little
is being done to utilize this wealth of information.
The biggest drawback to using a mechanistic approach is
the need for a detailed model. Even if a detailed mechanistic
model is available its parameters are uncertain, and often
they need to be updated in real-time. Using the established
statistical control charting methods has the advantage that
they need no model, but rather use the operational data
directly: they are also easily applied and interpreted. The
major drawback to these control charts is that they were
developed for monitoring univariate problems or sets of
independent variables, and the expansion to handle the case
of many highly correlated variables is difficult (Woodhall
and Ncube, 1985). These methods are still being used on
a one variable at a time basis on multivariate processes, either
formally using Shewart charts or informally by operators
monitoring key variables. Generally, this approach has been
adequate, although extremely inefficient; however, if the
variables are correlated this approach can lead to erroneous
results (Jackson, 1980).
This paper presents a different and more efficient approach
to process monitoring based on multivariable statistical
methods of Principal Component Analysis (PCA) and Par-
tial Least Squares or Projection to Latent Structures (PLS).
These methods are particularly suited to analyzing large sets
of correlated data. While, the analysis makes much more effi-
cient use of the data and the dimensionality of the problem
is greatly reduced, the results can still be interpreted and used
in the same way as originally outlined by Shewart (1931).
In the general problem process measurements are collected
from the plant and arranged in a ( IZ x k ) matrix X consisting
of n observations on k variables. while the product quality
variables are collected in a ( n x rn) matrix Y . The objec-
tives of a multivariate SPC scheme would be to monitor the
process and product quality using these observations and
detect process upsets, equipment malfunctions or other spe-
cial events. The final step in the procedure would be to find
and remove the assignable causes for these events, thus
improving process performance. In some situations one may
have only a single data block X which consists of variables
of the same type, either all process variables or all product
variables. In other situations one may have information on
both process, X , and product, Y , variables at the same time;
in this case another objective of any study is to develop a
predictive relationship for Y using X .
Given the current process control computer systems, on-
stream analyzers, and automated quality control labs, it is
not uncommon to measure hundreds of proce
line every few seconds or minutes, and tens o
ables every few minutes or hours. Although a large number
of variables may be measured, they are almost never indepen-
dent; rather, they are usually very highly correlated with one
another. The true dimension of the space in which the process
moves is almost always much lower than the number ofmea-
surements. In some situations this is due to underlying
fundamental relationships among the variables. For example,

T H E CANADIAN JOURNAL OF CHEMICAL ENGINEERING. VOLUME 69. FEBRUARY, 1991 35

in the hypothetical reaction of A B+ - C where A and B
arc fcd to the reactor in a specified ratio, although the con-
centrations of A , Band Care being measured (3 dimensional
measurement space), the actual problem is univariate (the
stoichionietric relationship and the fixed feed ratio each
eliminate a degrce of freedom). In other situations the place-
ment of the measurements and the nature of the process make
the measurements highly correlated. Consider, a distillation
column where only three variables change independently,
reflux, reboil and feed composition. Originally, measure-
ments are being made of the temperature profile using every
fourth tray temperature. If the number of measurements is
increased to every tray temperature to obtain a more detailed
temperature profile, the dimension of the measured variable
space has been greatly increased, but the actual dimension
of the problem has not changed. As a final illustration con-
sider another situation where measurements on many
different variables are made, but the nature of the process
and the disturbances are such that they only allow the vari-
ables t o move is a much lower dimensional space. For
example, in the manufacture of synthetic fibers it is not
uncomnion to measure more than ten quality variables such
as denier. elongation under differcnt loads, breaking strength,
dye depths, etc.. The physical meaning of these measure-
ments guarantees that the process is only capable of making
fibers with certain combinations of properties, and distur-
bances to the process will affect many of these variables in
a highly correlated manner. For example fibers with very
small deniers (weight I unit length) cannot be made with very
high breaking strengths, and disturbances which lead to a
reduction in denier lead to a reduction in breaking strength.
Principal Components Analysis (PCA), and Partial Least
Squares or Projection to Latent Structures (PLS) are mul-
tivariate statistical methods which consider all the noisy and
highly correlated measurements made on the process, but
project this information down onto low dimensional sub-
spaces which contain all the relevant information about the
process. Principal Component Analysis was originally devel-
oped by Pcarson (1901) and is a standard multivariate
statistical method described in many textbooks (e.g.
Anderson, 1984; Mardia et al., 1982; S. Wold et al., 1987).
The development of PLS is more recent (Horst, 1961);
some of the current interest in the method was sparked by
H. Wold (1982) when he applied this approach for model-
ling socioeconomic trends, particularly for situations which
were data rich and theory poor. PLS has found many appli-
cations in the physical sciences, especially in the chemo-
metrics field. A review of the development of PLS is given
by Geladi (1988) and will not be retraced here; however,
a brief overview of the current work is provided. The basic
ideas behind this method can be found in several recent
papers (S. Wold et al.. 1984a; Geladi and Kowalski, 1986a.b;
Bccbc and Kowalski, 1987); further, several papers have
been published concerning the mathematical foundations of
PLS (S. Wold et al.. 1987; Lorber et al., 1987; Manne, 1987;
Hoskuldsson, 1988) and their relationship to other statistical
methods (Nacs and Martens, 1985; Helland, 1988). PLS and
SIMCA, a cluster analysis technique based on PLS (S. Wold
et al., 1984a), have been applied in the analysis of multivar-
iate data sets in the chemometrics field (Otto and
Wegscheider, 1985; Moseholm, 1988; Kvalheim, 1988).
Interest in this method is also spreading to chemical
engineering; however, as yet very little has been published
(Wise and Ricker, 1989: Wise et al., 1989; Kresta et al.,
1 990).
36 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING. VOLUME 69. FEBRUARY. 1991
The next section of this paper describes the statistical
methods used in this work. The conceptual ideas of dimen-
sion reduction are presented using a geometric argument,
and the mathematical details of this method, including the
algorithm, are presented in the appendix. In the following
section a multivariate SPC procedure based on PLS is devel-
oped for monitoring the performance and detecting faults of
unusual events in high dimensional processes. This proce-
dure is demonstrated using two simulations; a fluidized bcd
reactor and an extractive distillation column. Finally there
is a discussion concerning the implementation of thc pro-
posed method and some concluding remarks are made.
Statistical methods
The challenges for any multivariate statistical monitoring
scheme are as follows:
I . The method must be able to deal with collinear data
of high dimension, in both the independent (X)and the
dependent (Y) variables.
2. The method must reduce the dimension of the problem
substantially and allow simple graphical interpretations
of the results.
3. If both process ( X ) and product quality ( Y ) variables
are present, it must be able to provide good predictions
of Y .
Methods for a single data block ( X )
Consider first the case where the data available is of only
one type, either all measurements are of process variables
or only the block of quality variables is of interest. Then this
data can be arranged in a single mean centered matrix, which
we will call X.
Principal Components Analysis (PCA) is a procedure
(Anderson, 1984; Mardia et al., 1982) used to explain the
variance in a single data matrix (X).PCA calculates a vector,
called the first principal component, which describes the
direction of greatest variability (Figure 1). It is calculated
as the orthogonal regression o f a line through the data in the
space spanned by X , or as that linear combination of the
columns of X ( X p l ) , given by the eigenvector of X T X ( p l )
associated with the largest eigenvalue. The second principal
component is orthogonal to the first principal component and
explains the greatest amount of the remaining variability. It
is obtained by fitting a line through the residuals left after
fitting the first principal component, or as the linear
combination of the columns of X ( X p z ) , given by the
eigenvector of X’X@,) associated with the next largest
eigenvalue.
One can proceed in this manner until k principal compo-
nents are obtained. In this case one has not reduced the
dimensionality of the problem, but only rotated the axes of
the X to a new orthogonal basis. Fortunately, with large data
sets one often finds that, after defining the first A principal
components where A < < k, most of the variation in the data
matrix X has been explained. By stopping at this point and
summarizing the information in X using new variables (ti’s)
defined as those linear combinations of the x ’ s ( t , = Xp,
( i = 1,2,. . . A ) ) given by the first A principal components,
one has reduced the dimensionality of the space from k to
A . In geometric terms this is equivalent to approximating the
k-dimensional observation space by the project ions of the
observations onto a much smaller A-dimensional hyperplane.
PCA is illustrated for a 3-dimensional system in Figure I :
 x2 I dimension A . One can proceed until the percent of the vari-
Figure I -- Geometric representation of the steps in PCA: a) obser-
vations plotted in the space of the measurements; b) the first prin-
cipal component; c) the plane defined by the first two principal
components.
PC2
x1
I T2
_ _ ..
x3
(8) (b)
Figure 2 - The loading (a) and the score (b) plots for the system
in Figure I .
in this system the bulk of the variation is in the plane defined
by the first two principal components ( A = 2 ) .
The principal components define the plane of greatest
variability, and the loading vectors associated with these prin-
cipal components define the location of the plane in terms
of the original variables, and each observation is located on
this plane via its scores. The score is the distance from the
origin of the plane (X) along each principal component and
is calculated as the product of the loading vector and obser-
vation. The perpendicular distance from each observation
to the plane is the residual for that observation. The infor-
mation contained in the original space can now be described
using two 2-D plots: (1) the loading plot, showing the rela-
tionship between the original variables and the principal com-
ponents (Figure 2a) and ( 2 ) the score plot, showing the
relationship between the observations and the principal com-
ponents (Figure 2b).
Algebraically the X matrix has been approximated by
X = tip; + t2p.T + . .. + tApi+ E ...... ..
where E is a residual matrix. Ideally the dimension A is
chosen such that there is not significant process information
left in E. Rather E should represent random error, and adding
an ( A + 1) principal component would only be fitting some
of this random error, and thereby increasing the prediction
error of the principal components model in Equation ( I ) .
There are several ways for selecting this maximum significant
THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING. VOLUME 69, FEBRUARY, 1991
ation explained by adding additional principal components
is small, although this often leads to overfitting. A better
procedure is to use cross-validation (S. Wold, 1978) whereby
one holds back a certain fraction of the observations (say
1/5), performs a PCA analysis o n the remaining data and
then computes the prediction error sum of squares (PRESS)
for those observations left out. This is repeated until every
observation has been left out once. The optimal order of the
principal components model, A , is taken as that order
minimizing the overall PRESS.
In PCA it is very important to scale the data properly sincc,
the variance contribution of a particular x to the total varia-
tion in X is dependent upon its units of measurement. In
general each variable should be scaled relative to the others
in terms of its relative importance. With product quality vari-
ables the specification ranges give a sensible unit of scaling.
while with process data the span of the sensors might pro-
vide a natural means of scaling. Alternatively. it is often
common to scale all the variables to have unit variance,
although care must be taken not to scale up the variance of
variables that are almost constant nor to disturb the natural
relationships among the variables of the same type.
A very interesting example of PCA applied to monitoring
a desulpherization process is presented by Denney et al.
(1985). That example provided some of the motivation for
this present work.
Methods for two data blocks ( X , Y )
In PCA we are concerned only with finding that latent
vector space which explains the greatest amount of varia-
bility in a single matrix of data ( X ) . Often, we can identify
a second group of variables ( Y ) which are of greater impor-
tance - eg. product quality or productivity variables, which
we would like to include in the monitoring procedure. Unfor-
tunately, these variables are often measured on a much less
frequent basis than the process variables; therefore. we would
like to use the information contained in the process variables
( X ) to predict, monitor and detect changes in the output
(process performance or product quality) variables ( Y ) .
Multiple linear regression (MLR) is the most common
method for developing multivariate statistical models. Unfor-
tunately, it is well established that MLR may have severe
problems dealing with large sets of collinear data (Draper
and Smith, 1981), leading to very imprecise parameter esti-
mates and poor predictions. In many applications, subsets
of the independent variables are chosen to eliminate problems
due to dimensionality and collinearity ; however, the aim of
the present work is to extract information from all measure-
ments available, and we will focus on methods that retain
the entire data set. One method which can be used to stabi-
lize the regression coefficients is Ridge Regression (Hoerl
and Kennard, 1970; Smith and Campbell, 1980; Draper and
Smith, 1981);unfortunately this method does not reduce the
dimension of the problem and becomes cumbersome for high
dimensional problems. Principal Component Regression
(PCR) regresses each of they 's on the principal components
of the X matrix. In this way it overcomes both the dimen-
sionality and the collinearity problems. PCR treats each y
variable individually, where in fact the Y space often con-
sists of many highly correlated variables. This can lead to
inconclusive results if they variables have little or no infor-
mation individually but are highly informative as a group.
Furthermore, the space defined by the principal components
31

*
..............................................................
*
I.. ............................................................................
Observations in time
Figure 3 - A typical Shewart Chart: * - indicate in control;
o - indicate out o f control.
of X is only that space exhibiting the greatest variation in
the x ‘ s and in not necessarily the space that is most predic-
tive of Y (S. Wold et al., 1984a; Geladi and Kowalski. 1986a;
Lorber et al.. 1987).
The Projcction to Latent Structures (PLS) method appears
to best address all of the above problems. A description of
the mathematical algorithm for PLS is given in the appendix
and can be found in several of the earlier cited references.
Conceptually, PLS is similar to PCA except that it simul-
taneously reduces the dimensions of the X and Y spaces to
find the latent vectors for the X and Y spaces which are most
highly correlated. In effect the plane in the X space is tilted
so that it is more predictive of Y. The latent vector spaces
in PLS are usually found by an iterative method (see
appendix), but it can also be shown that the loading vectors
for PLS are the eigenvectors of ( X ‘ Y ) (Y ‘ X ) where ( Y TX)
is the covariance matrix between X and Y (Hoskuldsson,
1988).
Monitoring via multivariate SPC plots
Statistical process control charts such as the Shewart chart,
CUSUM plot and EWMA chart are well established statistical
procedures for monitoring stable univariate processes. The
assumption behind them is that a process subject only to its
natural (“common cause”) variability will remain in a state
of statistical control unless a special event occurs. The con-
trol charts represent several statistical hypothesis testing
procedures aimed at detecting the occurrence of a special
event as quickly as possible. Upon detecting such an event
the action called for would usually be to diagnose the
problem, find an assignable cause for the deviation, and then
correct the process by removing this assignable cause. In
practice, such an approach of continually monitoring the
process, detecting events, and removing their causes leads
to long term improvements in the process.
A Shewart chart consists of plotting the observations
sequentially on a graph which also contains the target value
and upper and lower control limits as shown in Figure 3.
If an observation exceeds the control limits a statistically sig-
nificant deviation from normal operation is deemed to have
occurred, which triggers the search for an assignable cause.
The control limits are usually determined by analyzing the
variability in a reference set of process data collected when
only normal or “common cause” variability is present and
acceptable operation is achieved. The limits are then usually
set at plus and minus three standard deviations about the
target.
The philosophy applied here in developing multivariate
SPC procedures is the same as that used for the univariate
systems. As with the Shewart Charts, an appropriate refer-
ence set is chosen which defines the “normal operating con-
ditions” (NOC) for a particular process. The choice of this
3x
00 0
a..........0..UCL
LCL
T1 SPE
(b)
SPE I D
0
O
0
I * * * * *
TIME
Figure 4 - The multivariate monitoring chart. The axes are defined
by TI, T2 and SPE: a) 3-dimensional representation, b) 2-dinien-
sional representation, c) SPE plotted vs. time.
reference set is critical to the successful application of the
procedure and is discussed in more detail when describing
the case studie
rence in multivariate SPC is that we are
usually facing the challenges set out previously in the sec-
tion entitled “Statistical Methods”. The underlying premise
behind the proposed procedure is that many of the measure-
ments are highly correlated and simply represent different
manifestations of the same underlying disturbances or other
events occurring in the process. It is also assumed that the
underlying dimensionality of the process, when it is oper-
ating normally, is quite low, and in these circumstances we
can represent the most important elements of its behaviour
on low dimensional plots defined by the dominant latent
vectors obtained via PCA or PLS. In effect the low dimen-
sional planes defined by these latent vectors provide a low
dimensional window on the behaviour of the very high
dimensional process. For a plant with an underlying dimen-
sion of two, the axes of the score plot ( T , and T2)form the
first two axes of the monitoring chart and each observation
is located on this plot via its score. Using the regression equa-
tion developed for the two latent vectors (using PLS), the
predicted Y’s are calculated for each new observation. These
values are then used to?calculate the Squared Prediction Error
(SPE) (CTLl (yi - j+)-) for each observation and this vari-
able could form the third dimension of the monitoring chart
(Figure 4a), or could be plotted against time as in standard
Range charts (Figure 4c). If, as is often the case, the Y vari-
ables are measured only infrequently, and hence are not
usually available at each monitoring interval, then one should
plot the Squared Prediction Error for X , i.e. C:”,(xi - it)?.
This latter value represents the squared distance of the new
x vector from the prediction plane (TI, T?). Note that in the
latter case one might want to weight each x, element by its
modelling power for Y , i.e. C ~ = l w i ( ( x-i 2;)’.
It is worthwhile noting that the score plot (TI, T2)and its
control limit contour is very similar to the Shewart chart for
the mean of the variable, (Z),while the SPE plot is similar
to the range or standard deviation plot for the same variable.
The observations included in the reference set are plotted
on this chart; the limits of acceptable performance are set
such that they encompass a certain percentage of the varia-
tion in the reference set (analogous to the control limits of
THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 69. FEBRUARY, 1991
 the Shewart chart). It has been shown for PCA (Jackson,
1980; Jackson and Mudholdar, 1979), that under normality
assumptions and specific scaling, the control limits on the
TI - 7: plane form an ellipse and the proper size of the
control ellipse can be calculated: further, the X-statistic can
be used to establish the control limits for the residual or SPE.
Alternately, rather than basing the control limits on any
hypothesized underlying distribution; they can be based on
the reference distribution of the historical data contained in
the reference set (Box et al., 1978). This is the method used
in this work. The “normal operating conditions” should
reflect well behaved plant conditions and will need to be reset
if plant conditions change.
Once the chart has been set up it can be used to monitor
the process. Using latent variable relationships developed for
the reference set, the scores for each new observation vector
can be located on the plane (ti = Xws,f 2 = ( X - t , p :’)w4)
and its SPE calculated (CK I (yi - 9,)-or C:“; wi (xi - a,) .
The structure of these plots reflects the two ways in which
abnormalities can enter the system and provides powerful
diagnostic capability to determine the cause of the abnor-
mality. I f the abnormality is caused by a larger than normal
change in one or more of the process variables but the basic
relationship between the process and quality variables does
not change. then the abnormality will manifest itself as a shift
in the 7 ,- T2 plane, and the SPE will remain at an accept-
able level below its action limits (the * symbols in Figure 4a
and 4b). It’ on the other hand the abnormality enters through
a new event not captured in the reference set, it will change
the nature and possibly the dimension of the relationship
between the process and quality variables; then, the SPE will
increase ( o in Figure 4a and 4b). Acceptable performance,
i.e. perfbrmance which matches the reference set, would fall
within the envelope of normal operation ( A in Figure 4a
and 4b).
One difference between the standard univariate SPC charts
and the multivariate SPC chart is that there is no convenient
time axis to measure the age of the observations. The pas-
sage of time can be indicated in the T , - T? plane by the
markers used. This can be done in several ways including:
( I ) by marking the passage of time using shading intensity
- the most recent being the brightest and past values fading
so that only the pastp observations are visible; or (2) by using
time indexed markers (such as sl ,s2,s3,. . .), but still
showing only the past p observations on the current display.
Alternately, separate univariate charts for Ti and T2 could
be plotted against time in a manner similar to a Shewart chart.
This approach is better than plotting the original variables
individually, since there are many fewer T’s and they are
orthogonal; however, there is still a danger that an observa-
tion will lie within the individual limits but be clearly
abnormal when viewed on a bivariate chart (see Figure 7,
Case I ) . The SPE would usually be plotted against time
similar t o the standard Range or S-chart (Figure 4c).
The procedure described above and demonstrated in the
simulations is for the situation where both X and Y mea-
surements are available at regular intervals; or where the Y
measurements are available at least in the reference set
defining the NOC used in the PLS analysis. If, however, the
quality measurements ( Y ) were too infrequent or totally
unavailable, then the procedure could be reformulated. Two
possibilities exists: ( I ) if certain variables in X could be
considered good inferential variables for Y, then X could be
partitioned into XI and X,, where X , would contain the
inferential variables and act as the Y in the previous
T H E CANADIAN JOURNAL OF CHEMICAL ENGINEERING. VOLUME 69, FEBRUARY, 1991
discussion; or (2) if no inferential variables were available,
then the procedure would be based on PCA instead of PLS
and the principal component model, Equation ( I ) , for X
would be used to form an SPE (CyLlp, (xi - a;)’) in order
to detect changes which disturb the relationships among the
X variables. Both of these methods have limitations. The
method using PCA is less sensitive because it may include
variation which is not directly related to quality variables.
This method is also unable to detect disturbances which affect
the relationship between the process and quality variables
but are not included in X . The method using inferential vari-
ables is more sensitive to the specific disturbances which
affect the particular inferential variables used. It can detect
changes in the relationship between X and the inferential vari-
ables but may miss disturbances which do not affect these
but affect the other unmeasured quality variables.
In order to retain the ease of interpretation the system must
be chosen such that the bulk of the variation can be easily
presented to the user. This means that normally the
monitoring procedure would be restricted to three latent
vectors which can be presented on a 3-dimensional plot and
a separate SPE plot. For very complex systems, where three
latent vectors may not explain the bulk of the variation then,
the system should be divided into logical modular sections
which can be monitored separately. If this is impractical then
two further possibilities exist: (1) use only three latent vectors
and an SPE vs time plot. Although the three latent vectors
will not explain all the major variation, the variation not cap-
tured by them will still show up in the SPE plot; ( 2 ) if it
is felt that more than three latent vectors would be useful,
then the number of plots can be increased, although this
would quickly become cumbersome.
Monitoring a fluidized bed reactor
The multivariate monitoring procedure outlined in the last
section was applied to a fluidized bed reactor (FBR).
Although only simulations were used in this study, the non-
linear model used was one recently identified from
experimental runs a on pilot plant (Kclly, 1989). The reac-
tion in the FBR is the hydrogenolysis of n-butane on 10%
nickel on silica catalyst (Shaw et al., 1972; Kelly, 1989).
Due to the much faster concentration dynamics in the reactor
the concentrations can be considered at steady state with
respect to temperature. Although this process is non-linear.
especially when the extremes of operating conditions are
included, this did not create significant problems for the
monitoring method, since under normal operating conditions
the process is roughly linear. This process proved to be ideal
for a preliminary study of the technique because, while
numerous measurements were available, the underlying
process proved to be essentially two dimensional. A simpli-
fied schematic to this process is presented in Figure 5 along
with the measured variables.
The first step in the procedure is to choose which variables
will be considered as process variables and which will be con-
sidered as indicators of product quality. An on-line gas chro-
matograph (GC) was assumed to be available for the analysis
of the product stream, and the concentrations of the reaction
products (transformed to product selectivities and reactant
conversions) were chosen as the product space ( Y ) . The
remaining measurements were included in X . If the GC mea-
surements were not available, one alternative would be to
include the outlet reactor temperature in the product space,
as an inferential variable for the concentrations. The division
39
 Methane
Ethane
Propane
Hydrogen
Butane
Reactor Temperature
Cooling Oil Temperature
Hydrosen
Butane
Total Volumeuic Flowrate
Inlet Temperature
Conc. Hydrogen
Conc. Butane
Figure 5 - A schematic of the fluidized bed reactor (FBR)
including measured variables.
of measurements into process variable (X)and product vari-
able ( Y ) spaces is highly dependent on the measurements
available, and the objectives of the monitoring scheme.
The choice of the reference set and the proper scaling are
critical to the successful application of the procedure. The
reference set used to develop the monitoring chart will deter-
mine the variations considered to be part of normal opera-
tion and ideally includes all variation leading to acceptable
performance. If the reference set variation is too small the
procedure will alarm too often, and if too great the sensi-
tivity of the procedure for indicating abnormal operation will
be impaired. In an industrial application the reference set
would be chosen from past data. In the simulated example,
the reference set was generated from a normal distribution
about a nominal point. The variations and the nominal point
chosen for this simulation correspond to the conditions used
in the experimental work used to develop the model. The
four case studies were used to test the procedure: (1) a change
in the hydrogen to butane ration in the feed; ( 2 ) a ramped
increase in the cooling oil temperature; (3) a baseline drift
in the chromatograph reading, and (4) a change in the catalyst
activity.
Once a reference set has been established, the scaling of
the variables must be chosen. Scaling affects the relation-
ships between variables. Any variable with large amounts
of variation will dominate the first latent vector. Unless eco-
nomic or physical considerations indicate that certain vari-
ables are more important than others, the variance for each
measurement should be approximately the same. This has
led to the application of “autoscaling” (mean centering fol-
lowed by scaling all variables to unit variance) and although
this approach was successfully applied to the FBR, it is not
always the preferred choice due to physical considerations,
as described in the discussion of the distillation column
simulation.
Once the reference set and the scaling were defined, the
latent vectors of the PLS model were calculated. Table 1
presents the loading vectors (which define the latent vectors
in terms on the original variables) and the cumulative per-
cent of the variance explained for each variable. Only the first
two latent vectors were used in the monitoring procedure since
they explained 86.5% of the variation in Y , and because the
reduction in PRESS due to the third latent vector was small
and did not justify the added complexity (also it explained only
an additional 3.8% of the variance of Y). The scores plotted
in the monitoring chart are calculated as the product of the
measurement and the loading vector (tl = X(,,,, w I ;t z =
r)
- t I p w 2 ) ; the prediction of y is calculated as
9 = T B Q ~. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2)
or
9 = C tjb,qT .............................
, = I (3)

TABLE1
Results of the PLS Calculation for the Fluidized Bed Reactor Showing the Loadings for Each Latent Vector (LV)
and the Cumulative Percent of Variation Explained

First LV Second LV
Variables Loading Cum. % Expl. Loading Cum. % Expl.
Butane concentration 0.3002 29.5 -0.461 80.5
Hydrogen concentration -0.2957 39.2 0.4696 81.36
Ratio -0.3054 32.4 0.4951 88.34
Volumetric Flowrate -0.1008 0 0.085 4.5
T - Inlet 0.1099 9.0 -0.1893 12.0
T - Cooling Oil 0.586 56.0 0.4385 99.05
T - Reactor 0.6416 76.5 0.3131 98.0
Total % Explained of X 23.7 52.7

Methane selectivity 0.4607 83.7 0.4353 91.0

Ethane selectivity - 0.4046 65.3 -0.6862 82.6
Propane selectivity - 0.4594 83.2 -0.4127 89.8
Butane conversion 0.4471 78.14 0.255 81.0
Hydrogen conversion 0.4617 84.0 -0.3229 88.0
Total % Explained of Y 78.8 86.5

Regression Coefficient ( b ) 1.265 0.4188

40 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 69. FEBRUARY. 1991

 ”,
L

g40
w

w‘ 5
n
cno
-4 -2 0 2 4

z 5
8 4
cn
2i3
3 2
>
- 1
c
sa o
J -1
a
5 -2
8 -3
’ -4
cu
I- -5

T1 - First Latent Vector Scores SPE - Square of the Prediction Error

Figure 6 -- The monitoring chart for the fluidized bed reactor showing the normal operating region (the cross-hatched area) and the points
in the reference set ( 0).

In Figure 6 the points from the reference set are plotted on
the monitoring chart and the regions of acceptable perfor-
mance are defined (the shaded areas). The four case studies
described earlier were used to evaluate the ability of the
procedure to detect upsets or faults; the results are presented
on the monitoring chart in Figure 7.
The results clearly show the different types of abnormali-
ties which can arise. In the first case, the shift in the hydrogen
to butane ratio, the disturbance is measured by the process
variables and is indicated by a shift of the points on the
T, - T, plane. Since small variations in this ratio are
included in the reference set, the PLS model includes this
type of variation and the acceptable SPE shows that the
prediction obtained is still accurate. The second case, the
increase in the cooling oil temperature, indicates the same
type of abnormality. Further, if the process were truly linear,
the SPE would remain within the limits set by the reference
set; however, due to the non-linearities in the system the SPE
increases with increasing temperature until finally it exceeds
the acceptable limit. The last two cases, the base line drift
in the GC and the change in the catalyst activity, have no
effect on the measured process variables, but they have an
extreme effect on the prediction equation. These effects cause
drastic increases in the SPE indicating that the relationship
between X and Y has been altered. The data presented here
shows the method working on a system with variation but
no sensor noise. When sensor noise was added to the system
(white noise equivalent to one-half of the standard deviation
of the signal from each variable was added to each variable),
the same changes were detectable but less prominently.
Monitoring an extractive distillation column
The second simulation used to demonstrate the proposed
monitoring procedure is an Extractive Distillation Column
(EDC) separating an azeotropic mixture of acetone, methanol
and water (Figure 7). The results presented are from simu-
lation runs using a fundamental tray-by-tray model developed
at McMaster University (Chin, 1989). Only steady-state
monitoring was performed; however, the possibilities of
dynamic monitoring are discussed below.
As with the FBR the measurements around the distillation
column must be divided into the process and product vari-
ables. The configuration chosen is described in Table 2 . One
question that immediately comes to mind is whether the con-
centrations of all components in the product streams should
be included in the product variable set or only those which
will ultimately be used for control. The answer to this ques-
tion is case specific. However, the following rule should be
followed: use only those variables in Y which are of greatest
interest from a monitoring point of view. Extraneous vari-
ables in the product space do not normally affect the PLS
analysis; they do, however, impair the resolution of the
monitoring procedure by increasing the limits on the SPE.
Significant changes in the residuals of the important variables
can be hidden by the variance of the extraneous variables

THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 69, FEBRUARY, 1991 41

 xX
A

m 5
??
8 4
* 3
ti
v 2
2 ’
s o
c
0
-I -1
0

fi
-*
-3
8 1 - 4
(u
I- -5

T1 - First Latent Vector Scores SPE - Square of the Prediction Error

Figure 7 - The monitoring chart for the fluidized bed reactor showing the results of the simulated case studies used to evaluate the method.
The cross-hatched area indicates the normal operating region; Case 1 ( o ) indicates a change in the hydrogen to butane ratio; Case 2 ( + )
indicates an increase in the oil temperature; Case 3 (X) indicates a baseline drift on the GC; and Case 4 ( A ) indicates a change in the
catalyst activity

TABLE2 scaling approaches were investigated, ranging from unit

The Process and Product Variables Measured for the Extractive scaling all variables, except those in the temperature pro-
Distillation Column file, to no scaling at all, to specifically increasing the vari-
Process Measurements Product Measurements ance of certain variables. The following observations were
made: if the inter-relationships between a group of variables
Solvent Flow Flowrate of Distillate must be maintained (and they are measured in the same units)
Solvent Temperature* Acetone Concentration in Distillate then the same scaling factor must be used for the entire group.
Feed Flowrate Methanol Concentration in Distillate When the variances are almost the same size, then small
Feed Temperature* Water Concentration in Distillate adjustments in the scaling have an insignificant effect on the
Pressure* Flowrate of Bottoms
15 Tray Temperatures Acetone Concentration in Bottoms results.
Condenser Duty* Methanol Concentration in Bottoms By scaling selected variables to high variances compared
Reboiler Duty Water Concentration in Bottoms to the other variables, the latent vectors used by PLS model
were altered. When the scaled variable belongs to X,then
Note: * - Indicates variables held constant during simulation. this variable usually dominates the first latent vector. When
the scaled variable belongs to Y , then the variables which
which are included in the model. For this work all the infor- were highly predictive of this measurement were prominent
mation about the product streams was included in Y. When in the first latent vectors. This demonstrates the compromise
this was done nearly all the variation was still explained in being calculated by PLS between describing the X space and
three latent vectors and therefore in no way inhibited the per- predicting the Y space.
formance of the monitoring procedure. As with the FBR, the reference set for the EDC was gener-
Autoscaling was not used in this problem, as was done ated from a normal distribution around a nominal point. Five
for the FBR, because the relationships between the tray tem- case studies were used with this example, including: ( I ) an
peratures in the temperature profile should be kept intact. increase in the acetone to methanol ratio with a decrease in
This is done not because the final regression equation changes the water content in the feed; (2) no change in the acetone
but because more physical interpretation can be given to the to methanol ratio, but an increase in the water content in
first latent vectors if the data is unscaled. Several different the feed; (3) a decrease in the acetone to methanol ratio
42 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 69, F E B R U A R Y , 1991
 Cooling water temperature to those done for the FBR were performed and the results
n
plotted on the monitoring charts. The monitoring chart had
to be modified from the one used for the FBR because three
latent vectors were necessary to model the relevant infor-
mation. The scores from the three latent vectors form the
Distillate Flowrate axes of the monitoring chart (Figures 9 and 10) and the SPE
Acetone Conc. is shown on a separate chart (Figure 11). The SPE would
Solvent 10 Methanol Conc. generally be plotted sequentially in time, but can be plotted
__c
( Water ) Temp Water Conc. against any of the t 's, if this is more convenient for the appli-
cation. Since the data generated for this example was not
time dependent, SPE was plotted against T2, which gave the
Flow Tray Temperatures (15)
best resolution of the events.
Feed Te- The loading vectors for the three latent vectors used in the
5
(Acetone, 0.66; EDC monitoring chart are presented in Table 3 . Also
Methanol, 0.30; 2 presented in this table are the cumulative percent of vari-
ance explained for each variable. The loading vectors are
useful in interpreting the monitoring chart, and help assign
causes to abnormalities. The first loading vector indicates
that the first latent vector is dominated by changes in the
reboiler duty and the accompanying shifts in the tempera-
ture profile. Further evidence of this and the value of these
loadings in the interpretation of the monitoring chart is seen
in Figure 10, where the change in the steam temperature to
the reboiler are clearly seen as a shift in the Ti direction.
The second and third vectors are concerned with changes
in solvent and feedstock flowrates; for the second vector these
changes are in opposite directions, for the third they are in
the same direction. Again this interpretation is confirmed by

5-
8
v)

4 -
a
'0 3 -
c

8
> 2-
c
5
iii
1-

-I 0-
e
.-
iE -1-
I

p -2-

u) 4- 4
$j 3 - 3-
2 2- 2-
b 1 - 1 -
H
> 0- 0 -
4-
C
$ l -1 - -1 -
lu
-I -2 - -2 -
z -3- -3 -
1 - 4 - -4- rn
' -5 - -5 -
P -6 ' I ' ' ' ' I ' ' ' ' '

Figure 9 - The monitoring chart for the extractive distillation column showing the normal operating region (the cross-hatched area) and
the points used in the reference set ( m ) .

THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING. VOLUME 69, FEBRUARY, 1991 43

 -3" ' ' ' ' ' ' ' ' ' ' '
-8 -4 0 4 8 12
A
41
I
3 Y

'
P ILL--
-5
-6
-8
-4 0 4 8
T1 - First Latent Vector Scores
12
-5
-4 1
-6I '
-3 -1 1
T3 - Third Latent Vector Scores
A

3
A

5
I
I

Figure 10 - The monitoring chart for the extractive distillation column showing the results of the simulated case studies used to evaluate
the method. The cross-hatched area indicates the normal operating region; Case 1 ( v ) indicates an acetone to methanol ratio of 2.2-2.92;
Case 2 ( * ) a ratio of 1.96-2.07; Case 3 (X) a ratio of 1.71-1.64; in Case 4 ( A ) the solvent flow was increased by 33% and in Case
5 ( v ) the steam temperature was increased by 10 K.

events plotted in Figure 10; a positive change in the solvent
tlowrate is seen as a shift in the plane of T2 and T3, in
the positive direction for T3 and in the negative for T2.
One interesting thing should be noted in this simulation;
although feed composition variations were included in the
simulations, they were unmeasured disturbances. Since their
effect on the temperature profile is similar to that caused by
changes in the reboiler duty, it is difficult to detect feed com-
position changes solely from the information in the process
measurements. However, they are easily detected once the
plot of SPE is included (Figure 11).
Discussion
To obtain the best performance when using the monitoring
charts the following guidelines should be observed. The
product space should be restricted to those variables of
interest in the monitoring procedure. Extraneous variables
in the process measurement space do not inhibit the proce-
dure, and all possibly relevant measurements can be included.
Scaling should be performed in such a way that the variances
of the measurements reflect their relative importance (the
use of instrument ranges and engineering knowledge of the
process should be very effective). Care should be exercised
44 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 69. FEBRUARY. 1991
when using unit scaling not to disturb inter-relationships
among the variables. Reference sets must be designed to
reflect the purpose of the monitoring procedure. This means
that the procedure is case specific, and each application should
be tested to see if it performs according to the specifications.
The loading vectors contain information relating the
original variables to the latent vectors. This information used
in conjunction with past experience and pattern recognition
can be instrumental in assigning possible causes to the
abnormalities.
The work presented here has been restricted to steady-state
monitoring. By incorporating time lags (in the time series
sense) it should be possible to extend this procedure to
dynamic systems.
Finally, the work presented here was for the situation
where regular quality measurements are available; if these
are not available, then one of the previously mentioned
modifications needs to be implemented. Although these
methods have some additional limitations they can still be
effective monitoring tools; however, the details of their
implementation must still be investigated.
A major area for future research pertains to the appropriate
procedure for maintaining and updating the monitoring proce-
dure under changing operating conditions.
 TABLE3
Results of the PLS Calculation Using the Reference Set for the Extractive Distillation Column Showing the Loadings
for the Latent Vectors (LV) and the Cumulative Percent Variation Explained After Each Vector
First LV Second LV Third LV
Variables Loading Cum. % Expl. Loading Cum. % Expl. Loading Cum. % Expl.
Solvent Flow 0.06 0.0 -0.68 44.0 0.68 99.8
Feed Flow 0.06 0.0 0.52 50.0 0.65 99.1
Tray Temp 1 0.37 86.4 -0.31 95.0 -0.09 99.6
Tray Temp 2 0.15 86.4 -0.11 94.0 0.03 96.1
Tray Temp 3 0.004 6.0 -0.397 23.0 0.002 43.7
Tray Temp 4 0.06 68.0 -0. I06 92 0.02 92.4
Tray Temp 5 0.03 43.0 -0.1 I 86 0.03 90.3
Tray Temp 6 0.15 73.0 0. I6 89 0.03 90.7
Tray Temp 7 0. I5 64.7 0.2 I 90 0.03 92.0
Tray Temp 8 0.11 60.5 0.16 90 0.02 91.8
Tray Temp 9 0.06 59.5 0.09 89 0.02 91.5
Tray Temp 10 0.03 60.9 0.039 84 0.03 91.3
Tray Temp 11 0.03 62.8 0.028 84 0.03 97.8
Tray Temp 12 0.02 60.4 0.18 75 0.03 96.6
Tray Temp 13 0.01 54.6 0.008 63 0.03 94.6
Tray Temp 14 0.006 30.5 -0.004 29.6 0.03 86.3
Tray Temp 15 0.144 57.9 0.19 86 0.136 96.3
Steam Temp 0.874 99.8 -0.04 99.8 0.08 99.9
Total % Explained of X 74.9 87.1 98.9

Distillate flow 0.42 57.7 0.31 86. I 0.26 96.2

X , Acetone -0.4 54.5 -0.38 97.3 -0.06 97.8
X , Methanol 0.3 30.1 0.43 86.2 -0.24 95.1
X , Water 0.41 51.3 0.15 56.6 0.51 92.3
Bottoms Flow -0.17 8.0 -0.26 26.0 0.72 99. I
X , Acetone -0.51 84.5 0.2 96.0 0.16 99.5
X , Methanol 0.06 0.2 0.5 1 77.8 -0.24 86.2
X , Water 0.33 37.6 -0.425 94.5 0.02 94.8
Total % Explained of Y 40.4 77.0 95.1
Regression coefficient ( b ) 0.6325 1.485 1.128

6r I

Conclusions
i ,t
e A multivariate SPC procedure has been proposed for
-
G o
L
.-0
-5‘
8 1
-3 -1
T2 - Second latent vector scores.
1 3 handling large numbers of process and quality
variables. Multivariate statistical procedures are used to
t i . reduce the dimensionality of these large and highly

0
X

X
xx X 1I correlated data sets down to a few latent variables which
contain most of the information about the process
behaviour under normal operating conditions. By plot-
m
3
ting the projections (rows) of new process observations
fx over time on this low dimensional plane one is able to
detect larger than normal process variations, and by also
plotting the squared prediction errors, (ie. the perpen-
dicular distances from the plane) one is also able to
detect major changes in the behaviour of the process
caused by new events.
By compressing all the information on the process
down to low dimensional spaces, and using simple plots
-1 1 3 of the data in these spaces, together with meaningful
72 - Second latent vector scores. control limits, the essential idea and philosophy of
Shewart’s ( I 93 1) SPC methods have been preserved and
Figure \ 1 - The SPE chart for the extractive distillation column extended to handle the large number of variables
showing both the plot of the reference set (a) and the results of the
simulated case studies used to evaluate the method (b). The cross-
collected in most process industries today. The tools
hatched area indicates the normal operating region; Case 1 ( 0 ) necessary to establish the multivariate charts (PCA and
indicates an acetone to methanol ratio of 2.2-2.92; Case 2 ( + ) a PLS) may be more complex than usually used in univar-
ratio of 1.96-2.07; Case 3 (X) a ratio of 1.71-1.64; in Case 4 ( A ) iate SPC, but from the user’s point of view, the presen-
the solvent flow was increased by 33 % and in Case 5 ( v ) the steam tation of the data, and the interpretation of the results
temperature was increased by 10 K . is almost as simple.

THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING. VOLUME 69. FEBRUARY, 1991 45

 APPENDIX
Some mathematical details of P I S
The mathematical and statistical properties of PLS have been dis-
cussed in detail in by a number of workers (S. Wold et al., 1987;
Lorber et al., 1987; Naes and Martens, 1985; Hoskuldsson, 1988;
Hclland, 1988). In these papers many variations of the basic
algorithm used to do the calculation appear. The following is not
review of their work but rather as a summary, high-
se aspects which are particularly useful in obtaining a
general understanding of the procedure.
PLS is a regression technique, regressing Y onto X,where X is
and II x k matrix and Y is an n X rn matrix. This method is espe-
cially useful when the variables within X and within Yare correlated.
As with the geometric argument already presented, it is easier to
start with PCA and then introduce PLS.
PCA works on a sin& data matrix and attempts to explain the
structure of the variation. Generally the X matrix is mean centered
and scaled before PCA is applied. The methodology of PCA is to
decompose the data matrix into the following bilinear form:
X = C't,,p: t E,, . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
'
Mathematically TP is indeterminant; to obtain uniqueness one
can impose various conditions on t , and p a . One way to do this
is to force orthogonality a g the t o ' s and among the pa 's. If this
is done and it is further incd that they are orthogonal to the
rows/columns of E then it can be shown that the 1,'s are the eigen-
'
vectors of X X and pa 's are the eigenvectors of X 'X (Helland,
1988). This points to the equivalency between PCA and SVD
(Hoskuldsson, 1988). In SVD the eigenvalues are calculated in
descending order starting with the largest; in both methods the cor-
responding eigenvectors are also in descending order. For a detailed
discussion of PCA see Chartfield and Collins (1980) or Jackson
(1980). For correlated data sets A < < k using SVD to calculate
all the principal components is inefficient. The NIPALS algorithm
(S. Wold et al., 1987) calculates the PC's one at a time, the cor-
rect number of PCs can be determined from a number of stopping
criterion, one of the more popular techniques being cross valida-
tion (S. Wold, 1978).
The Y matrix can be similarly decomposed into
Performing the regression of u onto t leads to Principal Compo-
nent Regression. PLS follows a similar procedure except that it per-
forms both of the decompositions simultaneously in order to get
a better prediction of Y . The method can be described by the fol-
lowing algorithm (S. Wold et al., 1987).
1. Start: set u equal to a column of Y
2. w ' = u ' X / U 'U (regress columns of on u )
3. Normalize w to unit length
4. t = Xw / w ' y (calculate the scores)
'
S. q = t ' y i t ' t (regress columns of Y on t )
4
6. Normalize to unit length
7. u = Y q i q q (calculate new u vector)
8. Check convergence: if YES to 9, if NO to 2
9. X loadings: p = X ' t / t ' 1
10. Regression: 6 = u ' t i t ' t
1 1. Calculate residual matrices: E = X - tp' and F = Y - 6tq7
12. To calculate the next set of latent vectors replace X & Y by
E and F and repeat.
Several interesting things come from this algorithm, first, the
algorithm presented is for linear regression (step 10). Non-linearities
can be incorporated into the model by either transforming the
original variables in X and Y or by incorporating the known non-
linearity in step 10 (S. Wold et al., 1989). Secondly if the q vector
is not normalized then B = 1. Finally, if Y has only one variable then
46 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 69, FEBRUARY. 19'91
NIPALS is not iterative, some interesting properties of this case
are discussed by Lorbzr et al. (1987).
It is of interest to consider what objective function is being
minimized by PLS. Hoskuldsson (1988) addressed this from a geo-
metric point of view. He showed that the objective of PLS is find
orthogonal rotations of the original matrices X & Y such that thc
angle between the rotated matrices is the smallest. This objective
can be conversely stated that the covariance between u and t should
be maximized.
Hoskuldsson (1988) showed some interesting relationships
between PCA and PLS. These are of interest because PCA is con-
ceptually easier to grasp than PLS. PCA works on the X matrix.
and the loading vectors calculated using NIPALS are the eigen-
vectors of the covariance matrix (X ' X ) . PLS on the other hand,
works on both X and Y , the loading vectors calculated in this case
are the eigenvectors of the matrix X 'YY 'X. This last matrix can
be thought of in two ways: ( I ) the covariance of X has been scaled
using the "size" of the Y matrix (W'),or (2) PLS is simply PCA
preformed on the covariance matrix of X and Y (Y 7X).
Nomenclature
A = the optimal number of latent vectors
(A]) ba = the regression coefficient for the ath latent vectors of
X and Y measurements
B = the diagonal matrix of regression coefficients
E = the residual matrix of X
F = the residual matrix of Y
I = the identity matrix
k = the number of processhperationallinput measurements
m = the number of qualityioutput measurements
I1 = the number of observations
Pa = the loading vector for the uth latent vector of X cal-
culated bv PCA or PLS
P the load& matrix for X calculated by PCA or PLS
P, the loading for the ith X variable
4, loading vector for ath latent vector of Y
Q loading matrix for Y
fa scores associated with the ath latent vector of X
T the matrix of scores associated with X
T the axis of the monitoring chart
ua scores associated with the ath latent vector of Y
U the matrix of scores associated with Y
"'a the loading vector used for prediction. calculated
from PLS, for the uth latent vector
(A21
W the matrix of prediction loadings for X
w, the prediction loading for the ith X variable
X (n x k ) matrix of the processlopcrationaliinput
measurements
the mean of a variable of X and the vector of means
the value of the ith X variable for one measurenient
the predicted value of the ith X variable for one
measurement
x the composition of the bottoms product from the
EDC (Table 3 )
the composition of the distillate product from EDC
(Table 3)
( n X m) matrix of the process performanceiproduct
qualityioutput measurements
the value of the ith Y variable for one measurenicnt
the predicted value of the ith Y variable for one
measurement
Superscript
T = the transpose of the matrix
Abbreviations
CUSUM = Cummulative Sum
EDC = Extractive Distillation Column
EWMA = Exponentially Weighted Moving Average
FBR = Fluidized Bed Reactor
GC = Gas Chromatograph
LCL = Lower Control Limit
MLR = Multiple Linear Regression
NOC = Normal Operating Conditions
PCA = Principle Component Analysis
PCR = Principle Component Regression
PLS = Partial Least Squares or Projection to Latent
Structures
PRESS = Prediction Error Sum of Squares
SPC = Statistical Process Control
SPE = Squared Prediction Error
UCL = Upper Control Limit
References
Anderson. T. W., Introduction to Multivariate Statistical Analysis
2nd etl., Wiley, New York (1984).
Beebe, K . R. and B. R. Kowalski, “An Introduction to Multivar-
iate (’alibration and Analysis”, Anal. Chem. 59, 1007a-1017a
( 1987).
Box G € . P.. W. Hunter and J. S. Hunter, Statistics of
Experimenters, Wiley, New York (1978).
Chatfield. C. and A. J. Collins, Introduction to Multivariate Anal-
ysis, Chapman Hall, London (1986).
Chin, A , . Private Communication, Department of Chemical
Engineering, McMaster University (1989).
Denney, D. W., J . McKay, T. MacHattie, C. Flora and E. Mas-
tracci, “Application of Pattern Recognition Techniques to Process
Unit Data, Presented at CSChE Conference, Sarnia, Ont., Oct.
(1985).
Draper, N . R. and H. Smith, Applied Regression Analysis, Wiley,
New York (1981).
Geladi. P. and B. Kowalski, “Partial Least-Squares Regression:
A Tutorial”, Anal. Chim. Acta 185, 1-17 (1986a).
Geladi. P. and B. Kowalski, “An Example of 2-Block Predictive
Partial Least Squares Regression with Simulated Data”, Anal.
Chim. Acta 185, 19-32 (1886b).
Geladi, P., “Notes on the History and Nature of Partial Least
Squares (PLS) Modelling”, J. Chemometrics 2, 23 1-246 (1988).
Helland. I.. “On the Structure of Partial Least Squares Regres-
sion”, Cornrnun. Statist. - Simul. Comput. 17, 581-607 (1988).
Hoerl, A E. and R. W. Kennard, “Ridge Regression: Biased Esti-
mation for Non-orthogonal Problems”, Technometrics 12, 55-67
(1970).
Horst, P., ”Relations Among m Sets of Measures”, Psychometrica
26, 129-149 (1961).
Hoskuldsson, A . , “PLS Regression Methods”, J . Chemometrics
2, 21 1-228 (1988).
Hunter. J . S., “Exponentially Weighted Moving Average”, J. Qual.
Technol. 18. 203-210 (1986).
Jackson, J . E., “Principal Components and Factor Analysis: Part I
- Principal Analysis”, J. Qual. Technol. 12, 201-213 (1980).
Jackson. J. E., “Control Procedures for Residuals Associated with
Principal Component Analysis”, Technometrics 21, 341-349
(1970).
Kelly, J . , Private Communication, Department of Chemical
Engineering. McMaster University (1989).
Kresta, J . , T. E. Marlin and J. F. MacGregor, “Choosing Inferen-
tial Variables Using Projection to Latent Structures (PLS) with
Application to Multicomponent Distillation”, Presented at Annual
AIChE Meeting, Chicago, IL., Nov. (1990).
Kvalheim, 0. M., “A Partial-Least-SquaresApproach to Interpreta-
tive Analysis of Multivariate Data”, Chem. Intell. Lab. Sys. 3.
189-197 (1988).
THE CANADIAN JOURNAL O F CHEMICAL ENGINEERING, VOLUME 69, FEBRUARY, 1991
Lorber, A., L. A. Wagner and B. R. Kowalski, “ A Theoretical
Foundation for the PLS Algorithm”, J . Chemometrics 1 , 19-31
(1987).
Manne, R., “Analysis of Two Partial Least Squares Algorithms
for Multivariate Calibration”, Chem. Intell. Lab. Sys. 2, 187-194
(1987).
Mardia, K. V., J. T. Kent and J. M . Bibby, Multivariate Anal-
ysis, Academic Press, London (1982).
Moseholm, L., “Analysis of Air Pollution Plant Exposure Data:
The Soft Independent Modelling of Class Analogy (SIMCA) and
Partial Least Squares Modelling with Latent Variable (PLS)
Approaches”, Environ. Pollut. 53, 313-331 (1988).
Naes, T. and H. Martens, “Comparison of Prediction Methods for
Multicollinear Data”, Commun. Statist. - Simul. Comput. 14,
545-576 (1 985).
Otto, M. and W. Wegscheider, “Spectrophotometric Multicom-
ponent Analysis Applied to Trace Metal Determinations”, Anal.
Chem. 57, 63-67 (1985).
Pearson, K . , “On Lines and Planes of Closest Fit to Systems of
Points in Space”, Philos. Mag. 2, 559-572 (1901).
Shaw, 1. D., T. W. Hoffman, A. Orlickas and P. M. Reilly. “The
Hydrogenolysis of ()-Butane on Nickel on Silica Catalyst: I1 Flui-
dized Bed Studies”, Can. J. Chem. Eng. 50, 637-643 (1972).
Shewart, W. A., “Economic Control of Quality of Manufactured
Product, Van Nostrand, Princeton, NJ (1931).
Smith, G. and F. Campbell, “ A Critique of Some Ridge Regres-
sion Methods”, J . Am. Stat. Assoc. 75, 74-81 (1980).
Wise, B. M. and N. L. Ricker, “Feedback Strategies in Multiple
Sensor Systems”, AIChE Symp. Ser. 85, 19-23 (1989).
Wise, B. M., N. L. Ricker and D. J. Veltkamp, “Upset and Sensor
Detection in Multivariate Process”, paper 164b, AIChE Annual
Meeting, San Francisco, CA., Nov. (1989).
Wold, H., “Soft Modeling. The Basic Design and Some Exten-
sions”. in Systems Under Indirect Observation, K. Joreskog and
H. Wold, Ed., North Holland, Amsterdam (1982).
Wold, S . , “Cross-Validatory Estimation of the Number of Com-
ponents in Factor and Principal Component Models”, Techno-
metrics 20, 397-405 (1978).
Wold, S., C. Albano, J. Dunn 111, U . Edlund, K. Esbensen, P.
Geladi, S. Hellberg, E. Johansson, W. Lindberg and M. Sjos-
trom, “Multivariate Data Analysis in Chemistry”, in Chemo-
metrics: Mathematics and Statistics in Chemistry, B. Kowalski,
Ed., Reidel Publishing Co., Dordrecht. NL (1984a).
Wold, S . , A. Ruhe, H. Wold and W. Dunn 111, “The Collinearity
Problem in Linear Regression. The Partial Least Squares (PLS)
Approach to Generalized Inverses”, SIAM J. Sci. Stat. Comput.
5, 735-743 (1984b).
Wold, S., S. Hellberg, T. Lundstedt, M. Sjostrom and H. Wold,
“PLS Modelling with Latent Variables in two or more Dimen-
sions”, Frankfurt PLS Meeting, Frankfurt, FRG, Sept. (1987).
Wold, S . . K. Esbensen and P. Geladi, “Principal Component Anal-
ysis”, Chem. Intell. Lab. Sys. 2, 37-52 (1987).
Wold, S., N. Kettaneh-Wold and B. Skagerberg, “Non-Linear PLS
Modelling”, Chem. Intell. Lab. Sys. 7, 53-65 (1989).
Woodhall, W. H. and M. M. Ncube, “Multivariate CUSUM
Quality-Control Procedures”, Technornetrics 27, 285-292
(1985).
Woodward, R. H. and P. L. Goldsmith, Cumulative Sum Tech-
niques. Oliver and Boyd, London (1964).
Manuscript received February 5 , 1990; revised manuscript
received September 19, 1990; accepted for publication September
26, 1990.
47