Two practical approaches to the control of final product quality in semi-batch reactors are illustrated with an industrial experience. The first is an orthodox approach to the control of final quality which is achieved through a high degree of automation of all the reactor charging operations, good temperature and pressure control, and sequence of steps implemented throughout the course of the batch. The second approach to the control of final quality is to use mid-course correction policies. Both approaches to the control of final quality are illustrated with an industrial example. Through our experience with this example, the first approach, although very basic, is shown to be very important to reduce the variability in the final product quality. The second approach is useful to compensate for the new disturbances that are affecting the batch being run at the current time. Good practice is that mid-point control policies should be used after this basic automation step has been implemented as efficiently as possible.
The control of final product quality in an industrial
semi-batch reactor is a difficult problem for various
reasons. Foremost among them is the fact that robust
on-line sensors for monitoring the progress of product
quality development during the batch are almost never
available. Rather, quality variables are usually measured
only on the final product after the batch is
completed. Furthermore, batch reactors exhibit highly
nonlinear dynamic behavior, and the complex nonlinear
kinetic and dynamic models needed for model-based
inferential and nonlinear control are rarely available for
industrial processes. As a result, most batch and semibatch
reactors are operated in an open loop manner
with respect to product quality.The basic approach to control final quality is
achieved through a high degree of automation of all the
reactor charging operations, the start-up sequence, and
sequence of steps that must be implemented throughout
the course of the batch. This automation eliminates the
variability that might otherwise arise from the batch-to-batch variations in the implementation of these steps.
Good temperature and pressure controls are used to
minimize any effect of variations in these variables on
final quality.
In spite of using such advanced automation, some
batch-to-batch variations in quality will still be present.
If this quality variability exhibits predictable trends (i.e.
autocorrelation) due to persistent trends in raw material
properties, impurities, catalyst activities, etc., then
batch-to-batch feedback control schemes can be used to
eliminate it. They involve making small recipe adjustments
to each new batch based on the recent history of
quality measurements from past batches.
Although this batch-to-batch feedback control can
eliminate the predictable component of quality variation
between successive batches, it does nothing to
compensate for the new disturbances that are affecting
the batch being run at the current time. If adequate
on-line sensors are available to detect the presence of
disturbances arising from sources such as raw materials,
impurity and catalyst variations or from charging or
start-up variations, and if models can be developed to
predict the effects of these detected disturbances on the
final product quality, then model-based within-batch feedback control is possible. This within-batch control
will lead to a further reduction in the product quality
deviations from target for each batch.
Various approaches to this within-batch control
problem can be taken depending upon the type of
on-line information available and upon the complexity
of models available. Kozub and MacGregor (1992)
present an approach to the within-batch control of
multiple product quality variables in the semi-batch
emulsion polymerization of styrene-butadiene rubber
(SBR), using nonlinear control and state estimation.
However, this approach requires a good fundamental
dynamic model of the SBR emulsion polymerization
process, and a set of good on-line sensors (e.g. for
residual monomer concentrations and particle size) to
ensure observability of the state variables throughout
the time history of the batch. Neither of these is usually
available in an industrial setting.
Data-based quality prediction and control approaches
to the within-batch control problems have
been proposed. These approaches explore a more realistic
approach to within-batch product quality control
that is based on simple and readily available on-line
measurement and some off-line analyses.
Multivariate statistical process control (MSPC) has
been introduced for batch and semi-batch processes
based on the concept of multiway PCA or multiway
PLS (Nomikos & MacGregor, 1994, 1995). This statistical
monitoring is done using on-line process measurements
rather than off-line quality measurements. With
this technique, an abnormal batch can be detected
quickly, providing an opportunity to stop the batch or
make compensatory adjustments.
Mid-course control adjustment policies with an empirical
models (based on neural network or PLS) have been proposed for reducing predict deviations of final
product quality variables from their target values (Tsen,
Jamg, Wong and Joseph (1996), Yabuki and MacGregor
(1997)).
Two practical approaches to the control of final
product quality in semi-batch reactors are illustrated
with an industrial example. Through our experience
with this example, the first approach, although very
basic, is shown to be very important to reduce the
variability in the final product quality. The mid-point
correction approach is shown to be useful to occasionally
compensate for the new disturbances that are affecting
the batch being run at the current time. Good
practice is that mid-point control policies should be
used after this basic automation step has been implemented
as efficiently as possible.