Journal of Applied Statistics, Vol. 28, Nos.

3& 4, 2001, 409± 425

The statistical monitoring of a complex

manufacturing process

M. WEIGHELL, E. B. MARTIN & A. J. MORRIS, Centre for Process

Analytics and Control Technology, University of Newcastle, UK

abstract This paper describes the development of a multivariate statistical process

performance monitoring scheme for a high-speed polyester ® lm production facility. The
objective for applying multivariate statistical process control (MSPC) was to improve
product consistency, detect process changes and disturbances and increase operator
awareness of the impact of both routine maintenance and unusual events. The background
to MSPC is brie¯ y described and the various stages in the development of an at-line
MSPC representation for the production line are described. A number of case studies are
used to illustrate the power of the methodology, highlighting its potential to assist in process
maintenance, the detection of changes in process operation and the potential for the
identi® cation of badly tuned controller loops.

1 Introduction
Process manufacturing is increasingly being driven by market forces and customer
needs and perceptions. This is resulting in the necessity for ¯ exible manufacturing
with many companies now being required to manufacture a wide variety of
products, some of which may only be produced in small quantities to meet the
needs of specialized markets. The development of process monitoring models in
this kind of manufacturing environment necessitates the use of empirical-based
techniques as opposed to phenomenological models since the development of ® rst
principles models is unrealistic in the time available and, due to the complexity of
many process operations, a detailed model of the process is not realizable.
The monitoring and control of the operational performance of process plants
and their associated instrumentation is of strategic importance to industry. Failures
can lead to increased costs, reduced product quality and consistency, a reduction

Correspondence: E. B. Martin, Centre for Process Analytics and Control Technology, University of
Newcastle, Newcastle upon Tyne NE1 7RU, UK. E-mail: [email protected]

ISSN 0266-4763 print; 1360-0532 online/01/030409-17 € 2001 Taylor & Francis Ltd
DOI: 10.1080/02664760120034144
410 M. Weighell et al.

in production due to an increase in re-work, give away, plant shutdowns, etc. All
these factors contribute to the process having a greater than desired impact on the
environment. The objective of monitoring plant performance is thus to reduce the
level of oþ -speci® cation production by providing early warning and identi® cation
of important process changes, disturbances, malfunctions or faults. Consequently,
online performance monitoring has become an integral and extremely important
part of plant supervision programs aimed at improvements in product quality and
consistency, and productivity.
The availability of cheaper and more robust sensor technology has resulted in
industrial processes becoming more extensively instrumented with the resulting
measurements being routinely recorded on data acquisition systems. A consequence
of this is that the application of standard univariate statistical process control
techniques to manufacturing plants that produce products by `processing’ is
limited. First, operators can only monitor and respond to signals from a few charts;
thus, the majority of process measurements recorded tend to be ignored. Linked
to this is the fact that the process measurements recorded are not independent.
Univariate SPC does not explicitly take into account the interrelationships that
occur between the variables and thus valuable information concerning the behav-
iour of the process can be lost. In practice, under normal operating conditions, a
few key process operations determine the performance of the process at any one
time, thus in¯ uencing product variability. In practice, such events can be described
by diþ erent combinations of the process variables.
One of the more recent approaches proposed for assessing and improving the
performance and operation of manufacturing processes, and the quality and
consistency of production, which takes into account the variable interactions, is
Multivariate Statistical Process Control (MSPC). MSPC is increasingly being
recognized as a valuable tool for providing early warning of process changes, the
identi® cation of potential plant faults, process malfunctions and process disturb-
ances and also for enabling a deeper understanding of the process to be achieved.
The methodology is equally applicable to continuous and batch processes and has
been investigated by a number of researchers, including Kaspar & Ray (1992),
Kresta et al. (1991), MacGregor & Kourti (1995), Nomikos & MacGregor (1994),
Simoglou et al. (1997), Weighell et al. (1997), Wise & Gallacher (1996); and Wise
& Ricker (1991).
The concepts and methods of SPC diþ er to those of engineering process
control. Process performance monitoring, using standard statistical process control
techniques, provides an ongoing check on the manufacturing stability of the process
and points to problems whose elimination can reduce variation and permanently
improve the system. Process control uses feedback control to compensate for those
sources of variation that cannot be eliminated in this way. Clearly, the two
approaches are complementary and considerable advantage can be gained by
augmenting feedback process control techniques with statistical process control
methods (Box & LucenÄo, 1997). Due to the potential confusion between the two
approaches, a more appropriate terminology for multivariate statistical process
control is that of multivariate statistical process performance monitoring, indicating
the potential of the tool to contribute to manufacturing excellence.
This paper describes the application of multivariate statistical process control
to a high-speed continuous ® lm production line. Initially, the methodology of
multivariate statistical process control is reviewed. The paper then focuses on a
number of diþ erent case studies where the application of multivariate statistical
 Statistical monitoring of a complex manufacturing process 411

process control (MSPC) has enabled the detection of a number of unusual events
due to both maintenance procedures and process operating conditions.

2 Multivariate statistical process control and performance monitoring

Multivariate statistical process control methods address some of the limitations of
univariate monitoring techniques by considering all the data simultaneously and
extracting information on the `directionality’ of the process variations; that is, the
behaviour of one variable relative to the other variables. The basis of the Multi-
variate Statistical Process Control (MSPC) procedures used in this study is the
statistical projection technique of Principal Components Analysis (PCA) (Kresta
et al., 1991; MacGregor & Kourti, 1995; Nomikos & MacGregor, 1994). PCA is
a projection-based technique that summarizes the underlying sources of variability
in a data set comprising a single data matrix, X, through the de® nition of an
orthogonal set of latent variables (principal components). The ® rst principal
component, a linear combination of the original variables, de® nes the direction of
greatest variability in the data set and hence the largest amount of variability in the
data; the second principal component explains the second largest amount of
variation; and so on until the ® nal principal component, which explains the least
variation. Theoretically, as many principal components as original variables can be
calculated. However, in most industrial applications the major sources of variability
in the data set can be captured in the ® rst few principal components.
Only a brief summary of PCA is presented since the technique has been well
reported in the literature ( Jolliþ e, 1986; Jackson, 1991). Consider a data matrix,
X, representing n observations on each of k variables, x1, x2 , . . . , xk (e.g. tempera-
ture, pressure, ¯ ow rate etc). The ® rst step in applying PCA is to standardize the
data matrix, X, since PCA is scale variant. A number of alternative approaches
exist. The most common solutions are either to seek so-called natural units, i.e.
ensure that all variables measured are of the same type, such as temperatures or
¯ ows, or alternatively standardize the individual variables:
xi 2 xÅi
xis 5 i 5 1, 2, . . . , k (1)
si

where for variable i, x is the current measurement, xÅi is the mean of variable i and
si is its standard deviation. The principal components are then de® ned as:
T 5 Xs P (2)
where Xs is the normalized data matrix, P is the matrix of coeý cients (loadings)
that provide information as to which variables determine the directionality of the
individual principal components. T is the matrix of principal component scores
that act as surrogates for the observations. The loadings for the individual principal
components are the eigenvectors of the sample correlation matrix (XTs Xs ), whilst
the eigenvalues (k i ) of the sample correlation matrix are a measure of the amount
of variation explained by each individual principal component. The main attribute
of PCA is that it typically enables the dimensionality of the problem to be reduced.
If some of the process variables are collinear, or highly correlated, then a smaller
number of principal components (A) than the minimum of the number of original
variables (k) or number of samples (n) are required to explain the major sources
of variability in the data. In addition, the lower-order components are often
412 M. Weighell et al.

associated with instrument or process noise. Thus, the ® nal model can be
expressed as:
A
Xs 5 + ti pTi (3)
i5 1

Recommendations for the selection of an appropriate number of principal compo-

nents for process performance monitoring are the same as for traditional PCA.
These include heuristic-based methods such as selecting the number of principal
components (A) so that R Ai5 1 k i > 0.9 R kj5 1 s2j , where k i is the eigenvalue for principal
component i and s2j is the variance for variable j ; or, alternatively, from a plot of
R ti 5 1 k i , t 5 1, 2, . . . , k versus j, where the number of principal components is selected
at the `knee’ of the curve. A more statistically based approach such as cross-
validation (Wold, 1978) can also be used. Cross-validation involves splitting the
data into a number of diþ erent groups, r say. PCA is ® rst performed on the data
from (r 2 1) of the groups and one principal component is retained. Based upon
this process representation, the prediction error sum of squares is then computed
for the group omitted from the initial nominal representation. This group is then
reintroduced and the procedure repeated leaving out a diþ erent group at each
iteration until every block of observations has been omitted once, and only once,
from the nominal data set. The r separate predicted sum of squares are then
summed to give the total predicted sum of squares for one principal component.
This procedure is repeated for two, three, up to k principal components. The
number of principal components on which to base the process representation is
then identi® ed as that which minimizes the total prediction error sum of squares.
In most industrial applications, this will be through measurements of the process
variables. The ® rst stage in applying PCA for process performance monitoring is
to collect a reference data set when the process is operating under `normal’
operating conditions and is producing `good’ quality product. PCA is then applied
to the sample correlation matrix of the standardized reference data set to construct
charts for the monitoring of future process operation. The principal component
scores of future process measurements are then calculated using the loadings
(eigenvectors) of the sample correlation matrix of the reference data set.
The original monitoring charts introduced by Hotelling (1947) (the T 2 chart)
plotted the sum of the principal component scores divided by the eigenvalue of the
corresponding principal component dimension in a univariate manner for the ® rst
A principal components:
A
t 2i
T2 5 + (4)
i5 1k i

where ti is the principal component score for the ith principal component and k i is
the eigenvalue for the corresponding ith principal component. Since the principal
components are orthogonal, and thus uncorrelated, it has become common practice
to plot the scores in a bivariate manner (ti versus tj , i ¹ j, i 5 1, 2 . . . , A) to monitor
the process. In a similar manner to univariate statistical process control, action and
warning limits need to be calculated to identify the occurrence of non-conforming
operation (Nomikos and MacGregor, 1994). An additional metric also frequently
used to monitor the behaviour of the process is the squared prediction error (SPE)
( Jackson and Mudholkar, 1979). The squared prediction error (SPE) is de® ned as
the squared diþ erence between the observed values and the predicted values from
 Statistical monitoring of a complex manufacturing process 413

the nominal or reference model. Attention in this paper focuses upon the bivariate
scores plot, the SPE plots, although the T 2 statistic is equally applicable.
When an unusual process `event’ is detected through the violation of the control
limits, by inspecting the contribution of the individual variables to the principal
component score or the squared prediction error, relevant diagnostic information
can be obtained about the variables responsible for, or indicative of, the out-of-
statistical-control signal. This is achieved through the use of the contribution plot
(Miller et al., 1998). Two types of contribution plots are utilized in the paper:
diþ erential contribution plots and absolute contribution plots. In the absolute
contribution plot, the sign of the contribution is ignored and it is those variables
whose contribution are large in magnitude that are identi® ed as being indicative of
a change in process operation. The diþ erential contribution plot, is a bar chart of
the diþ erence between the contribution for a point (or group of points) that was
in-control, and the observation (or group of observations) when the process had
moved outside the control limits. Once again the results are presented in terms of
absolute contributions.

3 Application to a polymer ® lms production facility

3.1 Process description and background
The manufacture of polymer ® lm can be de® ned in terms of a series of unit
operations that are applied sequentially to convert a polymer feed chip stock to a
rolled ® lm product. Due to the nature of the ® lms industry, a range of products are
produced using the same processing equipment by varying operating parameters. In
the initial application of MSPC to the ® lms line, attention focused upon one
polyester ® lm product. Over 600 measurements are recorded across the diþ erent
processing units every minute, and are retained in a data storage system, providing
a wealth of data for analysis. The process for the production of polyester ® lm
comprises seven stages: Drying and Crystallization; Extrusion, Filtering and Poly-
mer Metering; Casting; Forward Draw (FD); Coating; Sideways Draw (SWD) and
Crystallization; and, ® nally, Edge Trim and the Transport Section. A schematic
diagram is shown in Fig. 1. One of the key objectives of the process is to ensure
that every roll of ® lm produced is of consistently high quality and has the same
desired physical properties. This is of paramount importance to the customer since
inconsistencies can lead to problems on their manufacturing processes.
A simpli® ed ¯ ow diagram of a multivariate process performance monitoring and
operator advice scheme is shown in Fig. 2. From an online data storage system, a
nominal data set can be extracted for the statistical modelling of normal production.

Drying and Crystallisation Extrusion Casting

Forward Coating Sideways Transport

Draw Draw Section

Fig 1. Schematic diagram of the ® lms manufacturing process.

414 M. Weighell et al.

Customer Feedback

Product Specification

Process
Modelling

Films Process Multivariate Statistical

Production Information Process Performance
Process System Monitoring

Corrective Process Performance

Action Information and Advice

Fig. 2. The multivariate statistical process performance monitoring and operator advice system.

The multivariate statistical process monitoring framework is then used to infer the
current state of the process, providing the user with relevant process information.
In the event of a change in process operation, a process disturbance or process
malfunction, the information derived can be used by the operator or production
engineer to identify corrective action. In addition, customer feedback may be used
to tailor the model for optimized production.
Speci® c variables can be manipulated such that the ® lm line may be used to
produce many diþ erent product grades. However, since this was the ® rst time
multivariate statistical techniques had been applied to the ® lms production line,
the study focused on a single polyester ® lm product. The plant was operated to
meet varying manufacturing schedules. This is achieved by de® ning set point
ranges for the major process variables. Of the 600 process variables, approximately
400 are checked between two and four times a day. Consequently, if a set point is
not at the desired value, then it is possible for the process to operate at this setting
for up to 12 hours prior to any inconsistencies being detected. Additionally, process
deviations that occur between checks may be missed by the operators unless they
have an impact on the limited number of continuously monitored variables or the
® nal quality measures. Although the existing software is used to monitor over 600
process variables, in reality, the process operators are only likely to spot process
deviations if they have some indication of when and where to look. Current process
monitoring procedures therefore involve observing a number of key variables with
periodic monitoring of all major process variables and oü ine quality control
measurements.
The overall aims for introducing multivariate statistical process control to the
® lms line was to improve the quality and consistency of the ® nished product and
to provide early warning of potential process and equipment malfunctions. This
can be achieved by identifying process deviations and subsequently determining
the cause of the malfunction or problem and eliminating the cause, resulting in
improved production consistency and ® nal quality. Retrospective multivariate
analysis of production data can also facilitate a deeper process understanding that
can assist in process improvements and preventative maintenance.
 Statistical monitoring of a complex manufacturing process 415

3.2 Development of the nominal model

To realize an eþ ective performance monitoring tool, it was essential to develop a
nominal representation of the process. The data forming the basis of the analysis
must be representative of the entire desired operation region of the plant. The ® rst
step was to de® ne the scope of the model. The problem was initially simpli® ed by
focusing upon one particular product type to eliminate complications arising from
intentional changes to the process settings. From the operating log of the plant,
periods of production were identi® ed when the ® lm line produced product within
the speci® cation limits for the product concerned. Representative examples of these
periods of nominal operation were collected over a period of several months in
order to capture a complete picture of nominal plant operation.
The data collected gave a representative picture of the diþ erent ® lm production
schedules and associated modes of plant operation. For the particular grade of
polymer ® lm selected, the resulting data set was too large to provide any practical
information to the operational staþ in this ® rst application of multivariate statistical
monitoring. A number of diþ erent approaches were adopted to reduce the data
set. The ® rst approach was to identify those process variables that were known,
through process knowledge, to have a signi® cant eþ ect on the ® lm line. The
resulting data set for analysis was still extremely large and further reductions were
achieved by focusing upon speci® c subsections of the process that were believed to
have the greatest impact on product quality. This enabled production personnel
and management to assess the applicability and usefulness of the MSPC method-
ology and eþ ectively created a series of sub-processes with a manageable number
of process variables, between 40 and 150.
An important aspect of the modelling and performance monitoring of dynamic
systems is the impact of fast data sampling and its eþ ect on the false alarm rate
(Average Run LengthÐ ARL). It is common practice to ensure that, in order to
capture the true response of the system, the sampling frequency is set to be
appropriately fast relative to the important time constants of the system. This
results in the process variables being both serially correlated as well as correlation
existing between the variables themselves. This is not a straightforward problem to
solve since serial correlation can potentially have a signi® cant impact on the
Average Run Length, and can result in too many false alarms (Vasilopoulos &
Stamboulis, 1974). This issue is required to be addressed to ensure that the
results from the multivariate statistical performance monitoring scheme were not
misleading. The aim of the industrial study was to demonstrate that `steady’
production could be improved using MSPC technology. By averaging the data over
a longer time interval, the serial correlation resulting from the fast sampling rate
was signi® cantly reduced and could theoretically be ignored.

3.3 Issues with set-point changes

One of the challenging issues encountered during the development of the nominal
model was how to deal with set point changes. Set point changes are necessary to
implement production grade changes as well as to respond to routine process
requirements. The eþ ect of set-point changes is shown in Fig. 3. It is clear that
this particular unit process was operated in a number of diþ erent regions over the
period of time that the nominal data set was collected. If region one of Fig. 3(a),
i.e. the cluster with approximate principal component ordinates (2,5) is examined
416 M. Weighell et al.

Fig. 3. (a) Process set point clusters. (b) Region 1 clusters.

in detail, it can be seen that three additional zones of operation are revealed,
Fig. 3(b).
The contribution plots for these scores plots (not shown) revealed that each of
the individual clusters represented a unique combination of process set points, all
within the acceptable limits of plant operation. Focusing on one of the clusters
revealed that the data are normally distributed and that MSPC can be applied in
the normal way. It is clear that the ® rst few principal components are not monitoring
natural process variations but set point changes that are implemented for process
speci® c reasons. It is clearly impractical to have separate representations for every
set-point combination as this will result in a large number of possible con® gurations
for every product type. When performing PCA on the nominal data, variations due
to the set point changes appeared predominately in the ® rst three principal
components. This moves the variations due to other deviations in the process to
the lower principal components. However, in the following case studies, attention
focused on speci® c clusters, and thus the issue of multiple set-points ceased to be
an issue in this initial study.
 Statistical monitoring of a complex manufacturing process 417

4 Case studies
Four case studies are presented and discussed to illustrate the detection capabilities
of the MSPC methodology. The ® rst two relate to the detection of process
maintenance activities during ® lm production and the potential impact of such
activities on product quality and consistency. The third case study highlights the
at-line detection of an incorrect set-point setting during a production grade change.
The ® nal case study demonstrates the ability of the performance monitoring system
to identify a badly tuned temperature control loop during process start-up of a
new grade of ® lm production.

4.1 Case study 1Ð heat exchanger coolant addition

A nominal model for the melt subsection of the process was built using six principal
components. Data for this case study were collected over a 200 minute period
when no process disturbances were identi® ed from the operations log. The squared
prediction error plot for this 200 minute period is shown in Fig. 4.
A deviation from nominal running was observed at time point 77 from the SPE
plot. This deviation was detected and observed for the subsequent 70 minutes.
Using a diþ erential contribution plot that compared the period, time point 80 to
time point 50, see Fig. 5, three variables were identi® ed as re¯ ecting the cause of
the disturbanceÐ variables 46, 48 and 49. All three variables are related, and are
associated with the dowtherm boiler that generates heat for the melt system metal
work.
A time series plot of one of these variable excursions is shown in Fig. 6. Further
investigation showed that, in response to a low dowtherm level warning on the
control panel, a quantity of cold ¯ uid was manually pumped into the reservoir.
This cold shot was believed to be responsible for the sudden change in temperature
that resulted in the process deviation. On this occasion, no other disturbances were
identi® ed in any of the other process sub-sections and it appeared that the
® nal product quality was unaþ ected. However, it was agreed that such `running
maintenance’ procedures should, if possible, be performed outside production

Squared Pre diction Error

45

40

35

30

25
SPE

20

15

10

0
0 20 40 60 80 100 120 14 0 160 180 2 00
T ime (Minutes)

Fig. 4. SPE Plot for a 200 minute prediction period.

418 M. Weighell et al.

Fig. 5. Diþ erential contribution plot.

Fig. 6. Fluid boiler temperature (scaled value).

periods to eliminate unnecessary sources of variability being introduced. Finally, it

is interesting to observe the step-like response in the temperature response. This
was identi® ed to be a consequence of the ¯ oating-point measurement being
converted to an integer value within the data acquisition and logging system.

4.2 Case study 2Ð heat exchanger maintenance

The second case study focused on a model built for another of the process
subsections. Twenty-two principal components were used to monitor 150 process
measurements. Examining the squared prediction error plot in Fig. 7, a number of
disturbances were identi® ed over the 24 hour period examined. These occurred at
time point 174 followed by a second larger disturbance at sample time 192, which
lasted until time point 233. A third disruption is seen between time points 374 and
465. Focusing upon the disturbance at time point 388, the absolute contribution
plot is given in Fig. 8. It can be seen that variables 33 and 42 are the major
 Statistical monitoring of a complex manufacturing process 419

Squared Prediction Error Plot

60

50

40
SPE

30

20

10

0
0 50 100 150 200 250 300 350 400 450 500
Time (Minutes)

Fig. 7. Squared prediction error plot.

Contribution Plot
25

20
AbsoluteContribution

15

10

0
0 20 40 60 80 100 120 140 160
Process Variable

Fig. 8. Absolute contributions plot to the SPE.

contributors to the process disturbance. Both of these variables are associated with
the temperature of the hot water delivered to the transport roll.
The controller output of variable 32, the temperature of the water, during this
time period is shown in Fig. 9. It highlights a series of disturbances. Again the
quantization eþ ect of the data is observed. The cause identi® ed was that, during
this particular period of time, the engineers were investigating a leak in a heat
exchanger.
It turned out that this disturbance did have an impact on the ® nished ® lm
quality. This is illustrated in Figs 10(a) and (b). In terms of the ® rst quality
parameter, the objective is to keep the value as low as possible whilst, for the
second quality parameter, the aim is to ensure that is kept constant. Figure 10
highlights the need to monitor all of the process parameters using MSPC since, in
many cases, process inconsistencies can lead to product property inconsistencies.

4.3 Case study 3Ð a set-point setting error

The third case study demonstrates the power of the technique when used as a
supervisory strategy. In changing product grades, a large number of process variable
420 M. Weighell et al.

Fig. 9. Variation of process variable 32Ð output power.

Quality Variable 1
(a) 1.7

1.6

1.5

1.4

1.3
Quality

1.2

1.1

0.9

0.8

0.7
160 260 360 460 560 660 760 860 960
Time (Minutes)

Quality Variable 2
(b) 293

292

291

290
Quality

289

288

287

286
160 260 360 460 560 660 760 860 960
Time (Minutes)

Fig. 10. (a) Time series plot of quality variable 1. (b) Time series plot of quality variable 2.
 Statistical monitoring of a complex manufacturing process 421

Squared Prediction Error Plot

200

180

160

140

120
SPE

100

80

60

40

20

0
0 20 40 60 80 100 120 140 160 180 200
Time (Minutes)

Fig. 11. Squared prediction error.

set-points need to be changed. A number of these are changed manually. After a

product changeover, a manually adjusted set-point was inadvertently overlooked
when the set-points were adjusted. Subsequent analysis showed a high SPE for one
of the process subsections, see Fig. 11.
The contribution plot for the SPE (not shown) highlighted a small group of
closely related variables as being the major contributors to the deviation. An
immediate investigation revealed that a temperature set point setting was 5ë C above
that desired for that particular grade of ® lm. This was detected roughly 100
minutes into the process. Shortly afterwards, the set-point was corrected. The
temperature change and eþ ect on the SPE can be seen in Fig. 12. It is noted that
although the temperature falls quickly, it takes considerably longer time for the
process to achieve a steady production state. This is due to the long period of time
required to re-establish the appropriate steady-state heating conditions in the ovens
where complex interactions exist.

Fig. 12. Monitoring of a set-point variable.

422 M. Weighell et al.

4.4 Case study 4Ð a process start-up

The ® nal case study arose from the consideration of a production line run-up to a
new polymer grade. The plots of principal component one versus principal compo-
nent two, together with the SPE plot, were used to monitor the performance of
the process in real-time. Figure 13 shows the progress of the scores in the PC1
versus PC2 plane as the control systems bring the process into steady operation.
There is clear evidence of some form of oscillatory behaviour. This can also be
seen in the SPE plot in Fig. 14. Investigation of the contribution plots in real-time
(Fig. 15) shows the progress of the 130 process variables monitored at 20 minute
intervals, indicated by the time points in the top right-hand corner of the plots.
There are two groups of variables that are interesting to observe, variables 20 to
28 and variables 90 to 95, which are related to the second zone of the extruder
and the extruder outlet ¯ ange, respectively.
Focusing on variables 20 to 28, it can be seen that the overall trend exhibits a
slow decay as the process progresses towards steady state operation. In contrast,

Fig. 13. PCA plot during run-up.

250

200
Squared Prediction Error

150

100

50

0
0 20 40 60 80 100 120 140 160 180 200
Obser vation

Fig. 14. SPE plot during run-up.

Statistical monitoring of a complex manufacturing process 423

Fig. 15. Dynamic contribution plots.

424 M. Weighell et al.

variables 90 to 95 are seen to exhibit oscillatory behaviour. This oscillation was

associated with a badly tuned temperature control system on the extruder. Even
though the performance monitoring system was not designed to take account of
process dynamics, in this particular case it gave the process engineers a visual
representation of both a slow heating control loop as well as pointing to a controller
tuning problem. Ongoing research is addressing the extension of the methodologies
described to include the impact of process dynamics and serially correlated data in
the application of multivariate statistical performance monitoring systems.

5 Conclusions
This paper has described the application of multivariate statistical process control
to a high-speed continuous polymer ® lms production line. The work has shown
that PCA is a powerful method for reducing the dimensionality of collinear and
highly correlated data sets from a polyester ® lms line and has enabled the plant
engineers and process operators to look at the process more closely, and obtain
a deeper understanding of its operation and production performance. It has
demonstrated a capability to provide early warning of potential deviations in ® lm
quality and consistency as well as to detect process maintenance problems, grade-
changing set point problems and equipment problems. The study also demon-
strated the potential to identify control loop problems. By monitoring the principal
components, a number of previously unidenti® ed events were isolated and the
causes identi® ed through contribution plots. This has resulted in a series of
improvements to production and maintenance planning. Work is progressing to
take the current models and use them online for production performance monitor-
ing. This preliminary analysis also raised a number of issues associated with the
monitoring of multiple products, dynamic process monitoring and the monitoring
of the complete process as one unit. However, these issues were signi® cantly
outweighed by its overall success.

Acknowledgements
The authors gratefully acknowledge the support of the Centre for Process Analytics
and Control Technology (CPACT) and the EPSRC Project MENTOR
(GR /M09971). In particular, the authors are very appreciative of the support and
contributions from Dr M. Bachmann and Ms J. Friend of DuPont FilmsÐ
Hopewell Site, Virginia, USA, without which this study would not have taken
place, and the business convinced of the technological and commercial bene® ts to
be gained from Multivariate Statistical Process Performance Monitoring. Finally,
Michael Weighell gratefully acknowledges the ® nancial support from the EPSRC
CASE Award Scheme, ICI and Du Pont.

REFERENCES
Box, G. & Lucen Ä o, a. (1997) Statistical Control by Monitoring and Feedback Adjustment (Wiley).
Hotelling, H. (1947) Multivariate quality control. In Eisenhart, C., Hastay, M. & Wallis, W. A. (eds)
Techniques of Statistical Analysis (New York, McGraw-Hill), 111± 184.
Jackson, J. E. (1991) A Users Guide to Principal Components (New York, Wiley).
Jackson, J. E. & Mudholkar, G. S. (1979) Control procedures for residuals associated with principal
components analysis, Technometrics, 21, 341± 349.
Jolliffe, I. T. (1986) Principal Component Analysis (New York, Springer-Verlag).
 Statistical monitoring of a complex manufacturing process 425

Kaspar, M. H., & Ray, W. H. (1992) Chemometric methods for process monitoring and high-
performance controller design, AIChE Jour nal, 38, pp. 1593± 1607.
Kresta, J., MacGregor, J. F. & Marlin, T. E. (1991) Multivariate statistical monitoring of process
operating performance, Canadian Jour nal of Chemical Engineering. 69, pp. 34± 47.
MacGregor, J. F. & Kourti, T. (1995) Statistical process control of multivariate processes, Control
Engineering Practice, 3(3), pp. 403± 414.
Miller, P., Swanson, R. E. & Heckler, C. F. (1998) Contribution plots: a missing link in multivariate
quality control. International Journal of Applied Mathematics and Computer Science, 8, pp. 775± 792.
Nomikos, P. & MacGregor, J. F. (1994) Monitoring of batch processes using multi-way principal
components analysis, AIChE Jour nal, vol. 40, pp. 1361± 1375.
Simoglou, A., Martin, E. B., Morris, A. J., Wood, M. & Jones, G. C. (1997) Multivariate statistical
process control in chemical manufacturing, IFAC Symposium, Safeprocesses ’ 97, Hull, UK, pp. 21± 28.
Vasilopoulos, A. V. & Stamboulis, A. P. (1974) Modi® cation of control chart limits in the presence
of data correlation, Jour nal of Quality Technology, 10(1), pp. 20± 30.
Weighell, M., Martin, E. B. & Morris, A. J. (1997) Multivariate statistical process control applied
to an industrial production facility, IFAC Conference ADCEM ’ 97, Banþ , Canada.
Wise, B. M. & Ricker, N. L. (1991) Recent advances in multivariate statistical process control. IFAC
Symposium, ADCHEM ’ 91, Proceedings, Toulouse, France.
Wise, B. & Gallacher, N. (1996) The process chemometrics approach to process monitoring and
fault detection, Journal of Process Control, 6, 329± 348.
Wold, S. (1978) Cross-validatory estimation of the number of components in factor and principal
components models, Technometrics, 20, pp. 397± 405.