![](../images/docto.jpg)
Doctoraatsverdediging
Sebastien Gourvenec
Chemometric methods for
batch process control using NIR spectroscopy
The thesis investigates chemometric methods that could
be applied to near-infrared (NIR) spectroscopic data to monitor and control
efficiently industrial batch processes. Monitoring and controlling processes
is a key activity since it guarantees that the manufactured products will
meet the required specifications. To
perform this control, several kinds of analyses exist. The in-line analysis
that makes the analysis directly in the process line without taking any
sample is the one that gives the most immediate information about the
process quality. This thesis focuses exclusively on methods that can be
applied in-line. NIR spectroscopy is described in chapter 1 and some candidate
methods are presented in chapter 2. The Orthogonal Projection Approach
(OPA) is especially studied because of its simplicity and because of the
interpretability of the results. Batch data present also an n-way structure
(typically samples x variables x batches) and methods able to deal with
them are also described in chapter 2. To control processes in-line, models
can be used. In the thesis, using curve resolution method such as OPA
does not consist, in the theoretical sense, in building models but in
finding simple meaningful variables that represent and describe as closely
as possible the original data. Practically, although this is not strictly
correct, this representation can also be considered as a model. However,
while modelling, problems can occur, especially about complexity. The
complexity problem is a general problem in chemometrics and means to deal
with it are described in chapter 3. Another important topic of this thesis
is the selection of appropriate data. Part B of chapter 3 is about the
selection of NIR wavelengths
able to reproduce similar OPA results as with complete spectra. This is
important from an industrial point of view to decrease the NIR spectra
acquisition time, and consequently also the time needed to obtain predictions
via models. Genetic algorithms are here used for this purpose. Chapter
4 is about the use of OPA to monitor batch processes in an industrial
context. It gives an overview of the industrial data, of the way the OPA
model is computed and of the obtained results. Some Multivariate Statistical
Process Control (MSPC) statistics based on Principal Components Analysis
(PCA) are also investigated in the OPA context in chapter 4. In chapter
5, a method called STATIS is described to monitor the evolution in time
of a batch. Another approach using the Hausdorff distance is also proposed
in the same chapter. Chapter 6 deals with some limitations of OPA for
batch process data, especially when rank deficiency occurs. The problem
is examined and some solutions are given.
![](photos/sDSC05646.jpg) ![](photos/blankimg.jpg) ![](photos/sDSC05647.jpg)
Terug naar doctoratenoverzicht
|