Submitted to Chemical Engineering Science

Inferential Sensors for Estimation of Polymer Quality Parameters: Industrial Application

of a PLS-based Soft Sensor for a LDPE plant

Rumana Sharmin, Uttandaraman Sundararaj, and Sirish Shah*

Department of Chemical and Materials Engineering
University of Alberta
Edmonton, Alberta T6G 2G6
CANADA
Larry Vande Griend, Yi-Jun Sun
AT Plastics
Edmonton, Alberta
CANADA
Abstract:
Low-density polyethylene (LDPE) and ethylene vinyl acetate (EVA) copolymers are produced
in free radical polymerization using reactors at extremely high-pressure. The reactors require
constant monitoring and control in order to minimize undesirable process excursions and meet
stringent product specifications. In industrial settings, polymer quality is mainly specified in
terms of melt flow index (MI) and density. These properties are difficult to measure and usually
unavailable in real time, which leads to major difficulty in controlling product quality in
polymerization processes. Researchers have attempted first principles modeling of
polymerization processes to estimate end use properties. However, development of detailed
first principles model for free radical polymerization is not a trivial task. The difficulties
involved are the large number of complex and simultaneous reactions and the need to estimate
a large number of kinetic parameters. To overcome these difficulties, some researchers
considered empirical neural network models as an alternative. However neural network models
provide no physical insight about the underlying process. We consider data-based multivariate
regression methods as alternative solution to the problem. In this paper, some recent
developments in modeling polymer quality parameters are reviewed, with emphasis given to
the free radical polymerization process. We present an application of PLS to build a soft-sensor
to predict melt flow index using routinely measured process variables. Issues of data
acquisition and preprocessing for real industrial data are discussed. The study was conducted
using data collected form an industrial autoclave reactor, which produces low-density
polyethylene (LDPE) and ethylene vinyl acetate (EVA) copolymer using free radical
polymerization. The results indicated that melt index can be successfully predicted using this
relatively straightforward statistical tool.
Keywords: polymerization, multivariate statistical analysis, melt index, viscosity, soft sensor,
statistical model
* Author to whom correspondence should be addressed

1.1 Introduction
Molecular and morphological properties of a polymer product strongly influence its physical,
chemical, rheological and mechanical properties as well as the end use properties. This makes
product quality a much more complicated issue in polymerization processes than in simple
chemical systems involving small molecules (Kiparissides, 1996, 2004). Polymer resins can be
sold only if they meet a set of specifications, which are defined in terms of end use properties
such as impact strength, melt index etc. In recent years increasing demand for product
diversification in the polymer industry has resulted in more stringent specifications
requirements in these properties. Due to the difficulty in measuring many fundamental
molecular properties, parameters which are conveniently measured and thought to be directly
related to various end uses are usually reported in the specifications rather than the fundamental
molecular properties (eg. melt index or intrinsic viscosity are reported in lieu of the number
average molecular weight, Mn, or weight average molecular weigh, Mw).
In the polyolefin industry, polymer grades are specified based on values of density and a
rheological parameter known as melt index. The melt index value (MI), which has a loose
inverse relation to weight average molecular weight, is the weight of a molten resin that flows
through a capillary die under a particular stress at a specified temperature (1900C for
polyethylene) for 10 minutes. MI indicates the viscosity of the polymer melt. Online
measurement of MI is difficult since it requires close human intervention. Even though online
hardware sensors have advanced, the use of new sensors has been largely limited to laboratory
reactors (Chien and Penlidis 1990, Kiparissides et al. 1998). As a result, in most plants, MI is
evaluated off-line and only infrequently, using a manual analytical test. Therefore, the process
has to operate without any real time quality indicator between successive measurements. The
production of multiple grades of polymer products from the same reactor, common for
polymerization reactors, is an additional complication, as it requires frequently changing
reactor-operating conditions. Due to the lack of online estimates of product quality, grade
changeover typically is a manual operation in most plants, and results in relatively large settling
time and/or overshoots. Consequently, a significant amount of off-specification polymer resins
are produced. From economic and environmental points of view, it is critical to reduce the
amount of off-grade material. This requires sophisticated control systems that are able to
provide optimal grade changeover trajectories. The most important factor in developing such a
1

system is the modeling and prediction of quality parameters. Consequently, a model for
estimating polymer properties such as melt index (MI), density etc. using routinely measured
process conditions would be extremely useful as an online sensor. In addition, it would allow
for optimal scheduling of grade sequence to avoid any mismatch of product quality during
grade transitions. However, a model derived from first principles, which is capable of
continuously predicting melt index from process operating conditions for any polymerization
process requires a significant effort and, to our best knowledge, a complete first principles
model does not exist in open literature for complex polymerization systems such as those
discussed here. Data-based multivariate regression methods are considered as a viable
alternative solution to the problem.

With the development of computer and data storage facilities, nearly every industrial process
now routinely collects and stores massive amounts of data on many process variables.
Variables such as temperature, pressure, flow rate, etc. are usually measured very frequently,
whereas product quality variables such as polymer molecular weight, density (), long chain
branching (LCB), short chain branching (SCB), conversion (x), etc are measured infrequently.
Yet, product quality variables are much more important to the polymer engineer and to the
customer. Efficient utilization of the large pool of process variable data can lead to significant
improvement in two areas. First, frequently measured process variables can be used to infer the
quality variables and an inferential control scheme can be developed. Second, the data can be
used to monitor the performance of the process over time for fault detection. In both situations,
multivariate statistics play a major role. Multivariate statistical techniques based on latent
variables such as PLS (partial least squares) have proven to be robust for handling large
number of ill-conditioned, highly correlated variables, which is the usual situation with
industrial data. Ordinary least square regression will usually fail in these cases. In addition,
process non-linearity can be incorporated in the model (if known from process knowledge) by
including appropriate non-linear terms in the model or by incorporating non-linear inner
relationships for PLS. Consequently, PLS has found numerous applications in areas of
chemometrics, process monitoring, and identification. Some examples include the use of PLS
to analyze and monitor an industrial ceramic melter (Wise et al., 1991), monitoring a batch
polymerization reactor (Kurti et al., 1996), and online estimation of quality variables in a
desulfurization process (Dayal et al., 1997, Miletic et al., 2004). MacGregor and co-workers
2

have done some excellent work with simulated data to show the potential of PLS in
identification and control of polymerization process (Duchesne et al., 2002, Jaeckle &
MacGregor, 1998, Skagerberg et al., 1992, etc.).

However, in the polymer industry very few industrial applications of PLS have been published.
Few examples include the use of PCA and PLS for monitoring and product quality prediction
for a batch emulsion polymerization process (Neogi and Schlags, 1998), and monitoring of
polymerization and spray drying processes (Hergeth Wolf-Dieter et al. 2003). In this paper,
PLS was used to build a soft-sensor to predict melt flow index using routinely measured
process variables. The study was conducted using data collected from an industrial autoclave
reactor, which produces low-density polyethylene (LDPE) and ethylene vinyl acetate (EVA)
copolymer using free radical polymerization.

1.2 Process Description

An autoclave reactor is a constantly stirred long vessel with a length to diameter ratio close to
20. The reactor is usually subdivided into different zones. Reaction conditions can be
separately adjusted in each zone to produce polymers with a wide molecular weight
distribution. Autoclave reactors behave more like a continuous stirred tank reactor (CSTR)
whereas tubular reactors are modeled as a plug flow reactor (PFR).
In this work, an industrial autoclave reactor at AT Plastics (Edmonton, Canada) was studied.
Figure 1A shows photograph of the reactor before it was installed. The plant uses ICI highpressure technology for the polymerization of ethylene to LDPE and copolymerization of
ethylene and vinyl acetate to EVA copolymers. Figure 1B shows a simplified process flow
diagram.

The four-zone reactor has a nominal capacity of 750L. Ethylene gas enters the reactor through
four separate streams. Free radical initiators are injected continuously into each zone of the
reactor to control the exothermic polymerization reaction at selected temperature levels. The
reactor is equipped with an axial stirrer and thermocouples along the reactor length. The reactor
operates at a pressure in the range of 1000 to 3000 atm and achieves about 20% conversion in a
single pass. After separation from the unreacted gasses in the high pressure separator and

extrusion hopper, molten polymer is fed into a single screw extruder which forces the polymer
to an underwater pelletizing unit in which polymer is cut into solid pellets.

Ethylene gas separated from the high pressure separator is recycled to the suction of the
secondary compressor. Any entrapped ethylene separated from the extrusion hopper is
compressed in a booster compressor, and finally mixed with fresh ethylene at the suction of the
primary compressor. The same unit also produces ethylene vinyl acetate copolymers (EVA)
when vinyl acetate is added to the system.

A trickle sampler is used to direct a small amount of polymer pellets to the quality control lab
where quality parameters such as MI, VA content, film quality etc are measured continuously.
The plant uses an online rheometer that measures MI continuously and the value is stored in a
computer database called the historian. At steady state, this equipment provides an accurate
measure of MI, but during grade transition, the estimates are erroneous. For many grade
changes, the equipment is taken offline to change the capillary die. A manual analytical test is
also performed in this lab at an irregular frequency to experimentally measure MI. These spot
measurements are used to update the online measurements for any bias correction, and also
during grade transition to make decisions about changing process conditions.

The rest of this paper is organized as follows: a review on recent development in modeling of
polymer quality parameters is given in section 1.2. A brief theory of partial least square
regression is presented in section 2. A description of data acquisition and preprocessing steps is
given in section 3. Results are shown in section 4, and some conclusions follow in section 5.

1.3 Different Approaches for Calculating Polymer Quality Variables

Development of inferential system for polymer properties is a very active research area in
polymerization reactor control (Chan et al. 1993, Chan & Nascimento 1994, McAuley 1990,
McAuley & MacGregor 1991, Kiparissides et al. 1993, Zabisky et al. 1992, Skagerberg et al.
1992 etc). The models used to estimate polymer properties can be roughly categorized into
three groups: (1) mechanistic models developed from first principles (for example, Chan et al.
1993, McAuley et al., 1990, Kiparissides et al. 1993, Zabisky et al., 1992), (2) black box
models using neural networks (Bhat & McAvoy 1990, Chan & Nascimento 1994, Qin &
4

McAvoy 1992, Rallo et al. 2002, Zhang et al. 1997, etc.), and (3) statistical models using
multivariate statistical tools (Jaeckle & MacGregor 1998, Kiparissides et al. 1993, MacGregor
et al. 1994, Martin et al. 1999, Skagerberg et al. 1992). The following section reviews the
recent development in each of these areas.

Among first principles models, one developed by McAuley & MacGregor (1991) is a
theoretically based model for the production of HDPE in a fluidized bed reactor to relate Mw
and comonomer composition using online temperature and gas composition measurements. An
empirical relationship between Mw and MI was incorporated into this model:

1 1

[M 2 ] + k [M 3 ] + k [H 2 ] + k [R] + k [I ]
ln (MI ) = k 7 + 3.5 ln k 6 + k 2
[M 1 ] 2 [M 1 ] 3 [M 1 ] 4 [M 1 ] 5 [M 1 ]

T T0

(1)

Here [M1], [M2], [M3], [H2], [I], and [R] are concentration of monomer, co-monomers,
hydrogen, co-catalyst, and impurities in the gas phase in the reactor. Parameters k1 to k4 and k7
were estimated using steady state offline data. The remaining model parameters were updated
using offline MI measurements by recursive prediction error method (RPEM). The model was
able to capture melt index and density of HDPE resins during both steady state operation and
grade transitions. This model was later used (sometime with minor modifications) by many
researchers in their study of HDPE and LLDPE reactors (Ogawa et al., 1999, Ohshima et al.,
1995, Ohshima and Tanigaki, 2000, Sato et al., 2000). Direct application of this model for
LDPE production is not suitable for two reasons:

HDPE

is

produced

using

coordination

polymerization

process

where

copolymerization of ethylene and -olefins are carried out in a fluidized bed reactor at
low pressure using a heterogeneous Ziegler-Natta or supported metal oxide catalyst.
LDPE production is performed in a free radical high-pressure polymerization process
using initiators. The basic reaction mechanisms are quite different for LDPE.

One major assumption used in McAuleys model is the uniform mixing of the reactants
in the gas phase of the reactor, which allows use of a uniform temperature and gas
composition throughout the reactor. Georgakis and Marini (1982) studied the mixing
pattern in LDPE autoclave reactors. The initiator tends to decompose near the feed
points and not in the bulk of the reactor resulting in non-uniform composition and a
temperature gradient along the reactor.
5

Neural networks appear to be quite an attractive method to infer polymer quality; however, it
suffers from some practical limitations. One disadvantage in using neural networks for data
mining is its slow learning process. Moreover, the model is implicit, and knowledge is hidden
in the network structure and optimized weights between the nodes. There is limited acceptance
of this method in industrial practice, particularly in polymerization processes, due to its black
box nature.

In polymerization processes, the use of multivariate statistical methods to develop inferential

models is limited mainly to simulation studies. A series of articles has been published by
MacGregor & co-workers and Kiparissides & co-workers in this area. Skagerberg et al., (1992)
and MacGregor et al. (1994) use PLS to model Mw, Mn, LCB, SCB, and concentration of vinyl
and vinylidene groups in a simulation study for a single grade of LDPE produced in a tubular
reactor using the reactor temperature profile and the solvent flow rates as the input. Although
the model was shown to perform well, the results were not validated with real plant data. In
addition, the issue of multiple grades and grade changes, which are very common to
polymerization reactors, was not covered in their work. Jaeckle & MacGregor (1998) used a
PLS based method to find a window of operating conditions to obtain a product with a desired
set of quality specifications. They developed a non linear PLS model for Mw, Mn, LCB, SCB
and conversion for the production of LDPE in a tubular reactor using reactor pressure, feed
temperature, initiator and solvent flow rates. Martin et al. (1999) reported a similar application
of inverse PLS to a pilot scale batch methyl methacrylate (MMA) suspension polymerization
reactor. In both cases, the data were generated using a simulation model developed by
Kiparissides et al. (1993) and there was no validation with industrial data. However, in a more
recent study, Jaeckle & MacGregor (2000) applied the inverse PLS method on historical data
taken from two industrial polymerization processes.

Although the method was not

implemented, the results could serve as a starting point for further experimentation.

A number of challenging issues are involved in direct application of PLS to model

polymerization processes. A major complexity arises from frequent grade changes required to
produce a large number of different products. Even for a single reactor system, operating
conditions and polymer properties can vary over a wide range. Polymer quality parameters such
as MI are believed to have non-linear relationships with operating conditions, but the explicit
6

nature of non-linearity is still unknown. Bremner and Rudin (1990) developed an empirical
relationship to relate MI with Mw for LLDPE and HDPE. Their model had some validity for
linear polymers of similar molecular polydispersivity and processing history. For branched
polymers (LDPE), they concluded that the situation is more complex, and no general
relationship can be obtained unless the rheology of the polymer can be characterized. The
process dynamics also change during grade transition. Sato et al. (2000) examined grade
transition through a simulation study and showed that the process gain can change greatly due
to product change. These issues suggest that a single PLS model will be insufficient to capture
the true process dynamics unless it is developed within an adaptive framework.

2. Partial Least Squares (PLS) Regression: Theory

Partial least squares regression (PLS) is a latent variable based method used for the linear
modeling of the relationship between a set of response variables Y (of size Nk) and a set of
predictor variables X (of size Nn). The model can be represented as follows:
Y = XB + E

(2)

The ordinary least square solution of B is given by

B = XTX

X TY

(3)

A common problem with this solution is that XTX can become ill-conditioned and the inverse
may not exist when X contains many correlated variables. Partial least squares regression (or
projection to latent structures) is used to overcome this problem (Lakshminarayanan et al.,
1997, Di Ruscio, 2000):
We can define

t1 = Xw1

(4)

where t1 is the latent variable in X-space and

u1 = Yl1

(5)

where u1 is the latent variable in Y-space. w1 and l1 are weight vectors of unit length that will
maximize the covariance between t1 and u1. It has been proven that w1 is the first left singular
vector and l1 is the first right singular vector of XTY. Obtaining score vectors t1 and u1 from X
and Y is known as the PLS outer model. Next an inner model between t1 and u1 is obtained
through a univariate regression of u1 on t1 as follows:

u1 = t1b1 + n1

(6)

( )

The least square solution for b1 is given by u1T t1 / t1T t1 . In order to get orthogonal X scores,
slightly rotated X loading vectors are defined as

( )

p1 = X T t1 / t1T t1

(7)

Y loadings q1 are the same as Y-weights l1. Successive PLS vectors can be obtained for the
deflated X and Y matrices:
E1 = X t1 p1T

(8)

F1 = Y b1t1 q1T

(9)

The procedure is repeated until all PLS dimensions (up to maximum of n dimensions) are
extracted. If r PLS dimensions are included in the final regression model, the coefficients B in
Eqn 2 are given by:
CPLS = W*BQT

(10)

Where,
B = diag{b1, b2, , br}
Q = [q1 q2 qr]
W* = [w1* w2* wr*]
i 1

w = I n w h p hT wi
*
i

h =1

P = [p1 p2 pr]

However, a simpler way to compute CPLS is: C PLS = W (PTW ) QT

1

The pioneering work in PLS was done by H. Wold in the late sixties. Since then, much
literature has emerged to describe PLS theory and many algorithms have been proposed
(Geladi & Kowalski 1986, Hoskuldsson 1988, Lindgren et al. 1993, Dayal & MacGregor
1997). The different PLS algorithms differ mainly in the iterative manner in which the
eigenvectors are calculated. In this work, NIPALS (Non-Linear Iterative Partial Least Squares)
algorithm (Geladi & Kowalski, 1986) was used to develop the basic PLS models. NIPALS is a
robust procedure for solving eigenvalue eigenvector related problems where the eigenvectors
are calculated in a partial fashion, one at a time, until all variance in the data structure is
explained. For each new dimension, the information explained by the last component is
subtracted from the data matrices X and Y to create residuals on which subsequent dimensions

are calculated by the same procedure (Lindgren et al. 1993). A summary of this algorithm is
given below.

PLS Algorithm:
1. Start with mean centered, and scaled X and Y; set E 0 = X ; F0 = Y
2. For each component: ustart = some Yj
w = u X / u u

= wold
/ wold

3. In the X block: w new

t = Xw

q = t Y / t t

= q old
/ q old

4. In the Y block: q new

u = Yq / q q

5. Check for convergence

6.

p = t X / t t

7.

= pold
/ pold

p new

8. t new = t old p old

= wold
p old

9. w new
10. Find regression coefficient b for inner relation: b = u t / t t

E h = E h 1 t h p h ; X = E h
11. Calculate deflated X and Y:

Fh = Fh 1 bh t h q h ; Y = Fh
12. Go to step 2 for the next component.

3. Data Acquisition and Preprocessing

Three basic steps in any data-based modeling or monitoring scheme are data collection, data
preprocessing and development of the model. Though it has received little attention in the
literature, data preprocessing can be the key to the success or failure of the final application. In
this work, offline process data for a total of 48 variables was collected from the plant site. The
data was queried from an OPC server located at the plant site using Structured Query language
(SQL) program files. Data for all 48 variables were accessed at every second and saved in a
database folder. It is worth mentioning that the appropriate length of the training data depends

on whether we intend to build composite model or a model for single grade product. In both the
cases, plant production history was studied to select the duration and length of the data. The
usual time length for producing any particular grade of polymer varies from 2~7 days at the
plant. Several batches of data, each having length ranging from 2~7 days, was collected to
build single grade models. For composite models, data over one month period was collected.
In this work fast rate of data storage placed an upper limit to the length of the data.

Process measurements are often corrupted with noise. Low signal to noise ratio leads to
reduced resolution and may cause false alarms during process monitoring. The same effect is
observed if bad data or outliers are present in the measurements. These effects were reduced
by filtering the data using an exponentially weighted moving average (EWMA) low pass filter
that removes high frequency fluctuations. The filter can be represented by a 1st order transfer
function:

y (t ) = x(t ) f

dy (t )
dt

Here x(t) is the raw measurement at time t, y(t) is the filtered value at the same instant, and f is
filter time constant. f was chosen such that it should be smaller than the dominant time
constant of the process. In this application, the residence time in the reactor, which was close to
42 seconds for most grades of product, was chosen as the dominant time constant. Accordingly,
a value of f = 4 was selected. Using this value of f, the differential equation was discretized
and used as the model equation for the EWMA filter:

y n = 0.85 x n + 0.15 y n 1
However, some fine tuning of the filter parameter was carried out based on the noise level of
the data. Manual removal of some bad data and outliers was accomplished using the plant daily
event log sheets where any special events, faults, etc. that occurred during the operation is
recorded.

It was expected that the total reactor feed should play an important role in affecting the product
quality. However, this quantity was not measured in the plant. A first principles model to
estimate the mass flow rate through the secondary compressor provided by the plant personnel
was used successfully in a previous work by Kumar et al. (2003). The same model was used in

10

this work to calculate the mass flow rate. This additional calculated variable was appended to
the original data matrix.

A quantitative measure of compression present in any data is given by compression factor (CF).
Several techniques have been proposed by researchers to estimate the compression factor. In
this work, an algorithm presented in Thornhill et al. (2004) was used. The results indicated very
high compression in the data, CFmin = 6 (CF<3 is recommended by Thornhill et al., 2004). A
thorough investigation of the historians data storage procedure revealed that the historian only
stores new data every 6 seconds and uses zero order hold for intermediate values. Process
variables are sampled in the plant using a sampling time of either 1 second (for flow, pressure,
and temperature) or 12 minutes (for composition measurements from GC). However, values are
displayed on the DCS at every second using a zero order hold device. Due to technical
limitation, the historian takes 6 seconds to collect all 48 tags. For example, in the first second
the first 8 tags are accessed; in the next second, next 8 tags are accessedhowever, the values
for the first eight tags remain constant at their previous levels. In this way, in 6 seconds, all 48
tags are accessed and stored in the historian. In the seventh second, the historian goes back to
measure the first set of 8 tags and the cycle repeats. Therefore, even for a variable that changes
rapidly (such as flowrate), there will be many straight-line segments where at least six values
will remain constant (see Figure 2), resulting in a minimum compression factor of six. For a
slowly changing variable (composition, temperature), the length of the straight-line segments
may be even larger. Any compression detection algorithm will falsely indicate the presence of
data compression if this particular feature of data storage is ignored. Therefore, it is concluded
that the data obtained was essentially compression free and can be used in any data based
analysis. However, since the data is updated at every 6 seconds, this should be the minimum
down sampling rate.

It is important in any data based modeling exercise to appropriately time shift the data
according to the time delay between different processing units to maximize the correlation in
the data matrices. Using mass flow rates, equipment dimensions, and piping dimensions, we
estimated the time delay from the feed points to the MI measurements taken at the lab. The
calculation of time delay is summarized in the Appendix. Some of the estimates are shown in
Table 1.
11

For PLS modeling, it is essential to use some scaling method. In this application, the data was
scaled by removing the mean and dividing by the standard deviation. Coefficients obtained
from models based on both scaled and unscaled data indicated that, when the data was not
scaled, variables with large numerical values heavily dominated the model. Variables with
large numerical values in most cases also exhibit larger standard variations than variables with
smaller numerical values of the same data set. Therefore, one should expect that PLS on
unscaled variables is dominated by those variables.

4. Results and Discussion

Because of the wide range of melt index covered by the various polymer grades, we initially
selected two intervals, one from the high MI range and one from the low MI range, and built
two composite models that would be valid for each interval. The final data matrices, after
completion of all preprocessing steps, were used for model building. The input to the PLS
model was all variables starting from the fresh feed up to the reactor. All setpoints and
variables measured downstream from the reactor were discarded. The model structure is as
follows:

MI (t ) = a1 x1 (t t d1 ) + a 2 x 2 (t t d 2 )...a n x n t t d n

(11)

where x1 to xn are process variables, t di is the time delay between sensor i and MI online
sensor, and ai are PLS regression coefficients.
One composite model (high MI) was developed using data from four grades of polymer having
melt index ranging from 6 to 110. Online steady state data was used for model building. Figure
3a shows the prediction from this model. The correlation coefficient between the actual and the
predicted values was 0.99 indicating the goodness of fit. Another PLS model was developed by
combining data for four polymer grades having low melt index (between 2 and 7). The model
contains one output and 22 input variables. Both steady state and transient data were used in
model building and the data was appropriately time shifted using the estimated time delay.
Figure 3b shows the prediction from this model. The correlation coefficient between the actual
and the predicted values was 0.82. The spikes in the true MI values in Fig 3-b represents large
overshoot during grade transition. Also in some instants, product quality deviated significantly

12

from target values due to some abrupt change in operating conditions. The model was able to
capture some of these trends.

It is important to note that although correlation coefficients between the actual and the
predicted values are quite high over the entire range of MI, the predictive capability of these
composite models are not as good for stationary data segments. For example, in Figure 3b, if
we focus on each grade separately, correlation coefficients are only 0.33, 0.88, 0.49, and 0.25
for grade 1, 2, 3, and 4 respectively. This implies that although the universal models capture
large variation in MI between grades, small variation within a grade are not well captured. In
order to overcome this deficiency, we built separate steady state models for each single grade.
The universal models would be used for grade transition only; and once steady state was
reached, the appropriate single-grade model would be used for estimation of MI.

A single grade PLS model was built using the data for grade-D (see Fig 3a). This EVA product
has a target melt index of 6. Figure 4 shows the model prediction. 14 PLS dimensions captured
93.7% variance in X and explained 73% variance in Y. The correlation coefficient between the
predicted and the true MI values for validation data set is 0.70 compared to 0.30 obtained from
the composite model.

Figure 5 shows model prediction for the PLS model developed using data from grade-C only
(see Fig 3a). The model captures the variation in melt index quite well, and results in a
correlation coefficient of 0.7 when applied to validation data set.

Another single grade PLS model was developed for one specialized grade of EVA product with
high MI (grade-E, not shown in Fig 3). The target melt index for the grade is 480. The plant
operators sometimes find it difficult to maintain the MI within specification. We developed the
softsensor to predict MI for this grade. Figure 6 shows the model prediction. The model with
25 PLS dimensions captured 99% variance in X and explained 94% variance in Y. It captures
the timely variation in melt index quite well, and, when applied to validation data set, results in
a correlation coefficient of 0.7, although there is a bias.

13

Although the single grade PLS models showed acceptable performance for a number of grades
of polymer, before implementing the models in real time, we needed to validate the models
with data other than the training data set. A model built with a limited amount of data from one
production campaign was used to predict properties during a different production campaign of
the same grade of polymer. The model developed for Grade-D (in Fig 4) was applied to three
data sets collected from three different production campaigns for the same grade. Note that the
elapsed time between when the model was first built and tested and the new data sets are 3, 4,
and 15 months respectively. Figure 7 shows the model prediction for these new data sets. It is
clear from this figure that model performance deteriorates over time. It was observed that the
variation in the data is significantly different at these different times. Example of change in
process variability with time is shown in Figure 8. Since PLS is a technique which is based on
covariance among variables, it is important that the training data and the validation data have
similar variability, i.e., the covariance structure of the data doesn't change over time. Therefore,
it is important to update the model periodically to account for variation in the data (eg. outside
temperature variations at different times of the year may affect the variability in process
conditions).

In all cases, validation data was scaled using mean and standard deviation of the training data.
Therefore, the model that was applied for prediction was:

Y Y1
X X1

= C PLS

S
S
Y

(12)

Here X and Y are validation data set, X 1 , Y1 are the mean of the training data set, and SY, SX
are the standard deviation of the training data set. However, if the validation data were scaled
using the mean of validation data and standard deviation of the training data, the following
model would result:

Y Y2
X X2

= C PLS

S
S
Y

(13)

Here X 2 and Y2 are means of X and Y blocks of the validation data set. An improved result
was obtained with this revised scaling. Figure 9 shows a comparison of two scaling techniques
for the first data set, which is taken after three months as showed in Figure 7.

14

It is evident that in order to successfully apply such models to a time-varying process, they
must have some adaptive capability. Here we suggest a simple bias update scheme using spot
values of melt index. Initially, the original PLS model is used for estimating MI. After a
specified time interval, we calculate a bias term defined as the difference between the true MI
value and the model prediction. This bias is added as a correction term during the next
specified time window. A feasible implementation strategy is as follows:
Let Y(t) represent the true MI value as measured using the rheometer. Y (t ) represents the model
prediction at time t. Therefore initially (at t=0),
Y (t ) = X (t ) b

(14)

At t=T,
bias (T ) = Y (t ) X (t ) b

(15)

For T<t<2T,
Y (t ) = X (t ) b + bias (T )

(16)

This procedure resulted in excellent prediction for all polymer grades. To illustrate an example,
five new sets of data were collected for the production of Grade D. The elapsed time between
when the model was first built and tested and the new data sets are 4, 5, 13, 14, and 15 months
respectively. Figure 10 shows the prediction with and without the bias update.

Analytical samples of MI collected from the lab can also be used to update the bias correction
on the output. However, irregular sampling in this case will result in less uniformity in the
correction term. Nevertheless, the results obtained by updating the model using the lab samples
are also promising as shown in Figure 11.

The major advantage of using this soft sensor is that it predicts the MI values using the reactor
conditions. The estimated delay between the reactor and the quality control lab is at least ten
minutes. Therefore, current MI measurements at the plant (both online rheometer and analytical
samples) are taken 10 minutes after the soft sensor predicts the value using data from the
reactor. Therefore, the soft sensor will give the operator reasonable time to take corrective
action in case of a mismatch between the desired and the true MI values. This provides a
powerful tool to perform product quality adjustments on-line at the moment there is an
15

excursion in the reactor, resulting in better product for the customer and decreased material
waste for the production plant.

5. Conclusions
In this article, we showed how PLS can be used to develop a soft-sensor to predict melt flow
index using routinely measured process variables from an LDPE-EVA plant. Issues of data
acquisition and preprocessing are discussed. The results indicated that melt index can be
successfully predicted using this relatively simple statistical tool.

It is important to reiterate difficulties associated with data based modeling of industrial data.
Real data is usually corrupted with noise and outliers. Therefore, the training set should have
sufficient excitation (high signal to noise ratio, without outliers). This may require collection of
a large pool of data and choosing the appropriate time segments to build the model. In addition,
the plant data usually contains of a large number of variables. Selection of appropriate variables
requires judgment and process knowledge. Statistical techniques such as step-wise regression
(Han, 2005) can also be employed in variable selection. Nevertheless, the model structure may
often become too complicated with too many variables and it is difficult to give a simple
physical interpretation.

In most chemical plants, there will be transport delay between different units. It is critical to
estimate the delay and thus lag the appropriate variables to maximize the correlation between
the X and Y block matrices. In this application the delays were estimated from physical
measurements in the plant. In all these aspects, the task is facilitated by frequent interactions
with the plant operators. The model prediction becomes poorer with time as operators use new
process conditions in the plant to produce the same resin with the same specification. Poor
prediction results when the new conditions deviate significantly from the data with which the
model was trained. In this paper, we suggested a simple bias update scheme as a remedy to this
problem. An alternative and more rigorous approach is to recursively update the model with
new data. A number of RPLS algorithms have been published in the literature (Dayal, 1996,
Dayal & MacGregor, 1997, Helland et al., 1991, Qin, 1993, 1998). A comparative description
of these theories and their application to this particular LDPE reactor is currently in progress.

16

Although we presented the method for modeling MI using PLS in this paper, the method is
general and can be applied to any other industrial data.

Acknowledgements:
The authors wish to thank the Alberta Ingenuity Fund, AT Plastics, NSERC (CRD) and the
NSERC-Matrikon-ASRA Industrial Research Chair program for the financial support for this
project.

References:
Bhat, N.V., McAvoy, T.J., (1990). Use of neural nets for dynamical modeling and control of
chemical process systems. Computer and Chemical Engineering, 14, 573-583.

Bremner, T., and A. Rudin, (1990). Melt flow index values and molecular weight distribution
of commercial thermoplastics. Journal of Applied Polymer Science, 41, 1617-1627.

Chan, W.M., Gloor, P.E., Hamielec, A.E., (1993). A kinetic model for olefin polymerization in
high pressure autoclave reactors. AIChE Journal, 39(1), 111-126.

Chan, W.M., Nascimento, C.A.O., (1994). Use of neural network for modeling of olefin
polymerization in high pressure tubular reactors. Journal of Applied Polymer Science, 53,
1277-1289.

Chien, D.C.H., Penlidis, A., (1990). Online sensors for polymerization reactor. JMS-Rev,
Macromol.Chem. Phys., C30(1), 1-42.

Dayal, B.S., (1996). Ph.D. Thesis, McMaster University, Hamilton, Ontario.

Dayal, B.S., MacGregor, J.F., (1997). Improved PLS algorithms. Journal of Chemometrics, 11,
73-85.

Dayal, B.S., MacGregor, J.F., (1997). Recursive exponentially weighted PLS and its
applications to adaptive control and prediction. Journal of Process Control, 7(3), 169-179.
17

Di Ruscio, D., (2000). A weighted view on the partial least-squares algorithm. Automatica,
(36), 831-850.

Duchesne, C., Kourti, T., MacGregor, J.F., (2002). Multivariate SPC for startup and grade
transitions. AIChE Journal, 48(12), 2810-2901.

Geladi, P., Kowalski, B.R., (1986). Partial least square regression: a tutorial. Anylitica Chimica
Atca, 185, 1-17.

Georgakis, C., Marini, L., (1982). The effect of mixing on steady state and stability
characteristics of low density polyethylene vessel reactors. ACS Symp. Ser., 196, Chemical
Reaction Engineering, Boston, 591-602.

Han, Z., (2005). Multivariate process monitoring and modeling. Ph.D. Thesis, University of
Alberta.

Helland, K., Bernsten, H.E., Borgen, O.S., Martens, H., (1991). Recursive algorithms for
partial least square regression. Chemometrics and Intelligent Laboratory System, 14, 129-137.

Hergeth, W.D., Jaeckle, C, Krell, M., (2003). Industrial process monitoring of polymerization
and spray drying processes. Polymer Reaction Engineering, 11(4), 663-714.

Hoskuldson, A., (1988). PLS regression methods. Journal of Chemometrics, (2), 211-228.

Jaeckle, C.M., MacGregor, J.F., (1998). Product design through multivariate statistical analysis
of process data. AIChE Journal, 44(5), 1105-1118.

Kiparissides, C., Verros, G, MacGregor, J.F., (1993). Mathematical modeling, optimization,

and quality control of high pressure ethylene polymerization reactors. J.M.S-Rev, Macromol.
Chem. Phys. C33(4), 437-527.

18

Kiparissides, C., (1996). Polymerization reactor modeling: a review of recent developments

and future directions, Chemical Engineering Science, 51(10), 1637-1659.

Kiparissides, C., Chatzi, E.G., Kammona, O., (1998). Recent hardware sensors developments
for monitoring polymerization reactions, DYCOPS-5, 72-79.

Kiparissides, C., (2004). Challenges in polymerization reactor modeling and optimization: a

population balance perspective. IFAC: DYCOPS meeting, Cambridge, MA, USA.

Kourti, T., J. Lee, J.F. MacGregor, (1996). Experiences with industrial applications of
projection methods for multivariate statistical process control, Computers & Chemical
Engineering, 20(supplement), s745-s750.

Krevelen, D.W.V, (1997). Properties of Polymer: their correlation with chemical structure,
their numerical estimation and prediction from additive group contributions, 3rd Edition,
Amsterdam, New York, Elsevier, pp. 71-107.

Kumar, V., Sundararaj, U., Shah, S.L., Hair, D., Vande Griend, L.J., (2003). Multivariate
statistical monitoring of a high pressure polymerization process. Polymer Reaction
Engineering, 11(4), 1017-1052.

Lakshminarayanan, S., Patwardhan, R.S., Shah, S.L., Nandakumar, K., (1997). A Dynamic
PLS Framework for Constrained Model Predictive Control. IFAC : ADCHEM Meeting, Banff,
Alberta, Canada.

Lakshminarayanan, S., Shah, S.L., Nandakumar, K., (1997). Modeling and Control of
Multivariable Processes: The Dynamic Projection to Latent Structures Approach. AIChE
Journal, (43), 2307-2323.

Lindgren, F., Geladi, P., Wold, S., (1993). The kernel algorithm for PLS, Journal of
Chemometrics. 7, 45-59.

19

MacGregor, J.F., Jaeckle, C., Kiparissides, C., Koutoudi, M., (1994). Process monitoring and
diagnosis by multiblock PLS methods. AIChE Journal, 40(5) pp 826-838.

Martin, E.B., Morris, A.J., Kiparissides, C., (1999). Manufacturing performance enhancement
through multivariate statistical process control. Annual Reviews in Control. 23, 35-44.

McAuley, K.B., Macgregor, J.F., Hamielec, A.E., (1990). A kinetic model for industrial gas
phase ethylene copolymerization. AIChE Journal, 36(6), 837-850.

McAuley, K.B., Macgregor, J.F., (1991). Online inference of polymer properties in an

industrial polyethylene reactor. AIChE Journal, 37(6): 825-835.

Miletic, I., Quinn, S., Dudzic, M., Vaculic, V., Champagne, M., (2004). An industrial
perspective on implementing online applications of multivariate statistics. Journal of Process
Control, 14, 821-836.

Neogi, D., Schlags, C.E., (1998). Multivariate statistical monitoring of an emulsion batch
process. Industrial and Engineering Chemistry Research, 7, 3971-3979.

Oshima, M., Koulouris, A., Tomita, S., Stephanopolous, G., (1995). Wave-net based online
quality inference system for polymerization processes. DYCORD, Copenhagen, Denmark, 275280.

Oshima, M., Tanigaki, M., (2000). Quality control of polymer production processes. Journal of
Process Control, 10, 135-148.

Ogawa, M., Oshima, M., Morinaga, K., Watanabe, F., (1999). Quality inferential control of an
industrial high density polyethylene process. Journal of Process Control, 9(1), 51-59.

Qin, S.J., (1993). A recursive PLS algorithm for system identification. AIChE annual meeting,
November 7-12, St. Louis.

20

Qin, S.J., (1998). Recursive PLS algorithm for adaptive data modeling. Computers & Chemical
Engineering, 20(4/5), 503-514.

Qin, S.J., McAvoy, T.J., (1992). Nonlinear PLS modeling using neural networks. Computer
and Chemical Engineering, 16, 379-391.

Rallo, R., Ferre-Gine, J., Arenas, A., (2002). Neural virtual sensor for the inferential prediction
of product quality from process variables. Computer & Chemical Engineering, 26(12), 17351754.

Sato C, Ohtani, T., Nishitani, H., (2000). Modeling, simulation and nonlinear control of a gasphase polymerization process. Computers & Chemical Engineering, 24(2-7), 945-951.

Skagerberg, B., MacGregor, J.F., Kiparissides, C., (1992). Chemometrics and Intelligent
Laboratory System, 14, 341.

Thornhill, N.F., Choudhury, M.A.A.S., Shah, S.L., (2004). The impact of compression on data
driven process analysis. Journal of Process Control, 14, 389-398.

Wise, B.M., Veltkamp, D.J., Ricker, N.L. Kowalski, B.R., Barnes, S.M., Arakali, V., (1991).
Application of multivariate statistical process control (MSPC) to the west valley slurry fed
ceramic melter process. Waste Management 91 Proceedings, Tuscon, AZ.

Zabisky, R.C.M., W.M. Chan, P.E. Gloor, A.E. Hamielec, (1992). A kinetic model for olefin
polymerization in high pressure tubular reactors: a review and update. Polymer, 33, 2243.

Zhang, J., Martin, E.B., Morris, A.J., Kiparissides, C., (1997). Inferential estimation of polymer
quality using stacked neural networks. Computers & Chemical Engineering, 21(Supplement-1),
S1025-S1030.

21

Appendix: Time delay estimation

The calculation is based on a typical mass balance for the entire plant obtained from AT
Plastics. Mass flow rate through the secondary compressor was taken as a basis. It was assumed
that the plant was operating at normal temperature and pressure. Note that transport delay
estimation based on this approach is not exact since it depends on the specific day of operation,
grade of polymer being produced, and specific temperature and pressure. Also, the calculation
was performed only for LDPE production. However, we assumed that although the mass flow
rate may vary from grade to grade, the volumetric flow rate remains approximately constant for
different product grades. Therefore, the estimates for LDPE are assumed to be valid also for
EVA copolymers.

Basis for delay estimation: Mass flow rates for ethylene and PE are known for each pipeline
from the mass balance. Pipe dimension (diameter and length) and equipment sizes were
collected from appropriate P&I diagrams. Using these information, transport delay was
calculated according to the following procedure:
Mass flow
rate of
ethylene and
PE
(Kg/hr)

Pipe length (m)

divide by
Density
3
(Kg/m )

Volumetric
Flow rate
3
(m /hr)

divide with stream

velocity (m/s)

divide by pipe
cross sectional
2
area (m )

Stream
velocity (m/s)

Transport delay (s)

Density Calculation for ethylene, polyethylene and their mixture: Assumptions

1. Density of ethylene is calculated using the equation of state: PV = ZRT.
Compressibility factor Z is taken as 1 for low pressure, and its value is obtained from
generalized compressibility factor chart at high temperature and pressure.
2. Density of polyethylene is calculated using relations for thermal expansivities of
polymer. Effect of pressure on density is ignored, only the effect of temperature is
considered.
3. At conditions where both ethylene and polyethylene are present in a single phase,
mixture density is calculated using the mixing rule:
22

4. mixture = i xi ; i = 1,2
i =1

Procedure for calculation of the density of polyethylene: (ref: Krevelen, D.W.V, 1997)
Structural unit of PE = [CH2CH2 ]
Vw = VanderWaals volume =

w ,i

= 2(10.13) = 20.46 cm3/mol

E1 = molar thermal expansivity of melt = 10-3Vw = 0.02046 cm3/mol

Eg = molar thermal expansivity in glassy state = 0.45 x 10-3Vw = 0.009207 cm3/mol
Vc = molar volume of crystalline polymer at 298K =

c ,i

=29.4 cm3/mol

Vg = molar volume of glassy polymer at 298K =

g ,i

= 32.74 cm3/mol

Assume, for PE, Tg = 195K, Tm = 408K, therefore, at 298K, PE exists in crystalline state.
Let, V1(T) = molar volume of melt at desired temperature T, where T>Tm

V1 (T ) = Vc (298) + E c (Tm 298) + Vm + E1 (T Tm )

= Vc (298) + E g (Tm 298) + 0.55 10 3VwTm + E1 (T Tm )
= 29.4 + 0.009207(408 298) + 0.55 10 3 (20.46)(408) + 0.0246(T 408)
V1 (T ) = 35 + 0.0246(T 408)

Let M = molar mass of structural unit (-C2H4-) = 28.05

density = (T ) =

M
28.05
=
, T > 408 K
V1 (T ) 35 + 0.0246(T 408)

23

Figure Captions:

Figure 1A: Autoclave reactor at AT Plastics (note the scale by comparing the van in front of the
reactor)

Figure 1B: Polymerization process flow diagram

Figure 2: Plot of fresh feed flow rate shows how the data is stored in the historian

Figure 3: Prediction from low MI PLS model (Left); high MI PLS model (Right)

Figure 4: PLS model prediction for grade-D

Figure 5: PLS model prediction for grade-C

Figure 6: PLS model prediction for grade-E

Figure 7: Model validation with three new data sets for Grade-D (top: after 3 months; middle:
after 4 months; bottom: after 15 months)

Figure 8: Variation in (a) reactor pressure and (b) reactor feed gas temperature

Figure 9: Model validation new data set for Grade-D: a) scaling factors taken from training
data, b) scaling factor taken from validation data

Figure 10: Model performance for Grade D after 4, 5, 13, 14, and 15 months (from top to
bottom) without update (Left); with bias update at every hour (Right); (solid line: true MI,
dashed line: model prediction)

Figure 11: Prediction versus Lab MI samples: a) without bias update and b) with bias update

24

Table 1: Estimated Transport Delay

Units

Estimeted delay (sec)

Reactor

42

Reactor to product cooler

Product cooler

27

Product cooler to inlet of high pressure separator

1.2

High pressure separator

176

High Pressure separator to low pressure hopper

45

Extruder

262

Pelletizer to trickle sample collector

10

Lab Extruder

180

25

Figure 1A

26

Comonomer/
Additive
Injection
Primary
Secondary
Compressor
Compressor
Ethylene
Supply

Primary
Stock
Tank

Low
Booster
Pressure
Compressor
Stock
Tank

Autoclave
Reactor
High
High
Pressure
Pressure
Separator

Separator

Purge
Extrusion
Extrusion
Hopper
Hopper

Extrusion&
Packaging

Polymer

Figure 1B

27

Initiator
Injection

7770

feed flow rate, Kg/hr

7765
7760
7755
7750
7745
7740
7735

2.265 2.2655 2.266 2.2665 2.267 2.2675 2.268 2.2685 2.269 2.2695 2.27

time, sec
Figure 2

28

x 10

120
actual MI
prediction

grade-A
100

Melt Index

80

60
grade-B
40
grade-C
20
grade-D
0

500

1000

1500

2000

sample

Figure 3(a)

29

2500

3000

3500

4000

45
actual MI
prediction

40
35

melt index

30
25
20
15

grade - 4

10
5
0

grade -1
0

2000

grade - 3

grade - 2
4000

6000

sample

Figure 3(b)

30

8000

10000

12000

6.8
actual MI
prediction

6.6

Melt Index

6.4
6.2
6
5.8
5.6
5.4
5.2

200

400

600

sample
Figure 4

31

800

1000

1200

27

26.5

melt index

26

25.5

25

actual MI

24.5

24
400

prediction
450

500

550

600

sample
Figure 5

32

650

700

750

800

445
actual MI
prediction

440
435

Melt Index

430
425
420
415
410
405
400

200

400

600

sample
Figure 6

33

800

1000

1200

25
actual MI
prediction

melt index

20

15

10

0
0

2000

sample

Figure 7(a)

34

4000

25
actual MI
prediction

melt index

20

15

10

1000

2000

3000

4000

sample

Figure 7(b)

35

5000

6000

7000

8000

25
actual MI
prediction

melt index

20

15

10

0
0

1000
sample
Figure 7(c)

36

2000

1760

Reactor pressure, atm

1750
1740
1730
1720
1710
1700

Nov02
Mar03
Apr03

1690
1680

2000

4000

time

6000

Figure 8(a)

37

8000

10000

24
Nov02
Mar03
Apr03

Feed gas temperature

22
20
18
16
14
12
10
8

2000

4000

time

6000

Figure 8(b)

38

8000

10000

25

25
actual MI
prediction

20

actual MI
20

xe 15
dni
tle
m10

xe 15
dni
tle
m10

2000

4000

6000

8000

sample

prediction

2000

4000

6000

sample
Figure 9

39

8000

Melt Index

8
6
4
2
8
6
4
2
8
6
4
2
8
6
4
2
8
6
4
2

5000

2000

5000

1000

10000

4000

2000

6000

4000

10000

2000

8
6
4
2

sample

3000

8
6
4
2

8000

8
6
4
2

6000

8
6
4
2

15000

8
6
4
2

4000

Figure 10

40

5000

2000

10000

4000

6000

8000

2000

4000

6000

5000

10000

15000

1000

2000

sample

3000

4000

9
actual MI
prediction

xe 7
dni
tle 6
m5

7
6
5

4
3

actual MI
prediction
lab

1000

2000

3000

4000

sample

1000

2000

sample

Figure 11

41

3000

4000