Stimulation Initiative for European Neural Applications
Esprit Project 9811
On-line Quality Modelling in Polymer Production
Company background
REPSOL QUÍMICA is the wholly owned subsidiary of Repsol S.A., the most
important oil company in Spain and among the first one hundred in the
world. REPSOL QUÍMICA is the main Spanish manufacturer of petrochemical
products holding a prominent position in the home industrial market. In
order to maintain and inhance its competitiveness has as priority
technological development programmes in which the current project has been
carried out.
Description of the problem
The quality of a polymer made in a REPSOL plant is determined
using two independent test: the Melt Index (MI) and the density (DE).
The MI is a measure of the flow rate of extrusion of the material
through a die while the DE is a measure of the packing of the molecules
within the polymer crystals. Both qualities measure the properties
of the polymer and they are of great importance when characterising
the polymer for practical applications. However, these measurements
are costly as well as technically delicate and therefore they are
carried out off-line in the laboratory each 2 and 4 hours respectively.
Hence, during this period of time no quality measurement of the final
product is performed. This is why a model that estimates on-line the
value of these properties as a function of the state of the plant
would be useful for controlling the final characteristics of polymer.
The state of the plant is defined by the set of all variables in the
reactor. However, within the great number of variables that determine
this state, four of them were identify as those which are mainly
responsible of the changes in both the MI and DE values.
Neural Network Application
Neural networks have been proved to be extremely successful in
system modelling. However, in industrial environments there often
exists an important drawback that is the presence of time varying
conditions. Modelling of non-linear system under these conditions
leads to the absence of global models and therefore oblige to
built adaptive systems using local models. Local models define
functions that are applicable only in the neighbourhood of the
current state of the plant and they are built using only those
states of the plant that were closed to the current one.
Two neural network local models were built to estimate the MI
and DE values as a function of the four variables that define
the state of the plant. The models are updated each time
that a new measurement of MI and DE performed in the
laboratory was carried out. Meanwhile estimations of MI and DE are
obtained using the neural network models.
Benefits
These neural network models offer the possibility of estimating,
on-line during its production, the polymer DE and MI values.
Therefore, the state of the plant can be changed in order to
improve the polymer quality maintaining the MI and DE values
within a fixed range of variation.
Generalisation
The system modelling described here is a typical application
of modelling in industrial environments where there exists
a set of independent variables along with one or two
dependent variables. Within industrial environments, it is
often found variables that are
difficult to measure or there is no simple form of being
obtained, therefore a model is required for their
estimation.
Contact persons
Guzmán García - Repsol Química - El Morell,
Apdo 398 - 43080 Tarragona - Barcelona - Spain.
Alberto Pérez, Instituto de Ingenierí
a del Conocimiento - IIC Unversidad
Autónoma de Madrid - Módulo C-XVI planta 2 - 28049 Madrid - Spain -
Phone: +34 1 397 39 73; Fax: +34 1 397 39 72; E-mail:=
alberto@irene.iic.uam.es.
![[ ]](../logo_small.gif){kind=logo}
SIENA