The application of batch profile characterization tools to enhance process understanding by uncovering the signature of the primary disturbances on the profiles and its effect on the product quality is illustrated on a nylon-6,6 process. The historical profile data for the fixed recipe operation are systematically studied to understand the primary disturbances affecting the process, and it is shown that good online predictions of the final product quality are possible much before the completion of the batch from the available measurement profiles. A simple online recipe adjustment strategy based on the predicted quality deviation from the target is proposed. Results show that the recipe adjustments significantly reduce the variation in the final product quality. Issues in the use of empirical prediction models from recipe-based data are discussed.
One of the primary objectives of batch manufacturing
in the chemical industry is the consistent
production of on-target quality batches. This is
due to the premium on the quality of the valueadded
chemicals that are typically manufactured
using batch processes. Even though the final product
quality is a major concern, online quality measurements
are seldom available. The batch operation
is consequently based on a fixed recipe, which
may represent years of process experience, with
variables considered important, such as reactor
temperature, being controlled to a prespecified trajectory.
A completed batch is characterized as either
on-spec or poor quality from laborious analytical
measurements on a sample of the product.The fixed-recipe-based operation with tight profile
trajectory control helps in rejecting many disturbances
that can potentially affect the final product
quality. However, some significant common
cause sources of variation, such as impurities in
the raw material, are not compensated for, resulting
in high quality variation. The occurrence of
these disturbances is reflected, although indirectly,
in the various online measurements such as temperatures,
pressures, and flows. Measurement profiles
for several past batches are typically stored in
a historical database. The database is a rich source
of information, which can be systematically studied
to understand the sources of variation affecting
the process and to suggest strategies for improving
the process, especially for tighter quality control.
The development of batch data mining tools has
thus received much attention in recent years.
These profile characterization methodologies essentially
quantify the systematic variability along the time axis and the measurement axis. Tools
such as dynamic time warping @1# and template
matching @2# characterize the former, while multivariate
methods such as principal component
analysis @3# and partial least squares @4# are used
for characterizing the latter.
The reported applications use profile characterization
methodologies towards various ends. Process
monitoring with the past successful operation
as the baseline for comparison is the primary goal
in several articles @5–10#. Some applications focus
on the development of quality prediction models
from the profiles @11–13#. Quality predictions are
central to the implementation of online recipe adjustment
strategies, also referred to as inferential
control. Russell et al. demonstrate recursive databased
prediction and control of product quality for
a nylon-6,6 process @14#. Midcourse recipe correction
using empirical models has been demonstrated
on a semibatch process @15#. Recent reports
use the empirical predictions in model predictive
quality control @16,17#.A major criticism of the quality control applications
in the reported literature is the use of empirical
models that are not based on a fixed-recipe
operation but include batches spanning the control
moves in the training set. Such rich data sets are
seldom available, especially in the very conservative
industrial environment, limiting the applicability
of the methodologies. Also, the strategy for
quality control is implicitly assumed. The strategies
may not always be very obvious for a process.
Indeed, the real challenge for process improvement
lies in deciphering from the historical data,
the signature of the primary sources of variation
on the measurement profiles, and their effect on
the product quality. In cases when the occurrence
of the disturbances or significant quality deviations
from target can be inferred from the measurement
profiles well before the end of the batch,
corrective action can be taken to reduce the variability
in the product quality. This is referred to as
within-batch control and its demonstration using
empirical models forms the thrust of the work reported
here.
Data mining tools are applied to profiles from a
recipe-based nylon-6,6 process, and opportunities
for online recipe adjustments for tighter control of
the polymer molecular weight ~MW!, the primary
quality variable, are identified. A simple online
recipe adjustment scheme based on final product quality prediction is suggested. Results from
implementation of the within-batch control
scheme show significant improvement in the product
quality variability. Emphasis is laid on the
physical relevance of the various empirical model
parameters to understand the cause and effect relationships
governing the process. Such an understanding
is essential in proposing an effective control
strategy.
The article is organized as follows. The recipebased
nylon-6,6 process is briefly described in the
next section. The database of profiles generated
from the simulation is then subjected to data mining
to obtain time and magnitude scale parameters
that characterize the profiles. The correlation
structure of the scale parameters is studied to uncover
the signature of the primary disturbances affecting
the process. The scale parameters are also
used to build online product quality prediction
models. It is shown that good predictions are possible
much before the end of the batch opening the
possibility of online recipe adjustments. Withinbatch
control schemes involving the addition of
amine salt and reducing the jacket pressure are
proposed. The improvement in quality control due
to the implementation of the online recipe adjustments
is quantified. A discussion of the various
engineering issues and conclusions that can be
drawn from the work complete the article.
Online batch recipe adjustments for tighter
product quality control has been demonstrated on
a nylon-6,6 process. The methodology is general
and uses readily available profile measurements to
build online quality prediction models. Midcourse
adjustments are possible when good quality predictions
are obtained well before the completion
of the batch. Results show that the adjustments
lead to significant improvements in quality control.
The batch data mining tools help in enhancing
process understanding by providing insights
on the primary disturbances affecting the process.
Such understanding is necessary in proposing effective
quality control schemes, especially when
only recipe-based operation data are available.
