La lecture à portée de main
Découvre YouScribe en t'inscrivant gratuitement
Je m'inscrisDécouvre YouScribe en t'inscrivant gratuitement
Je m'inscrisDescription
Sujets
Informations
Publié par | Thesee |
Nombre de lectures | 45 |
Langue | English |
Poids de l'ouvrage | 3 Mo |
Extrait
THÈSE
Présentée pour obtenir
LE GRADE DE DOCTEUR EN SCIENCES DE
L’UNIVERSITÉ PARIS-SUD 11
Spécialité : PHYSIQUE
École Doctorale « Sciences et Technologies de l’Information
des Télécommunications et des Systèmes »
par
Sándor Bilicz
ApplicationofDesign-of-ExperimentMethods
andSurrogateModels
inElectromagneticNondestructiveEvaluation
Soutenue le 30 mai 2011 devant la Commission d’examen:
M. B´iro´ József (Examinateur)
M. Dular Patrick (Examinateur & Rapporteur)
M. Idier Jérôme (Examinateur & Rapporteur)
M. Lesselier Dominique (Président)
Membres invités :
M. Gyimo´thy Szabolcs (Co-directeur de thèse)
M. Lambert Marc (Co-directeur de thèse)
M. Calmon Pierre
M. Vazquez Emmanuel
Rapporteurs :
M. Dular Patrick
M. Harsa´nyi Gábor
M. Idier Jérôme
tel-00601753, version 1 - 20 Jun 2011tel-00601753, version 1 - 20 Jun 2011ApplicationofDesign-of-ExperimentMethods
andSurrogateModels
inElectromagneticNondestructiveEvaluation
PhDDissertation
prepared under the “co-tutel” joint supervision scheme between the
UNIVERSITY OF PARIS-SUD 11
and the
BUDAPEST UNIVERSITY OF TECHNOLOGY AND ECONOMICS
Author:
SándorBilicz
Advisors:
MarcLambert
Laboratoire des Signaux et Systèmes UMR8506
(CNRS-SUPELEC-Univ Paris-Sud)
and
SzabolcsGyimothy´
Department of Broadband Infocommunications and Electromagnetic Theory
(Budapest University of Technology and Economics)
2011
tel-00601753, version 1 - 20 Jun 2011tel-00601753, version 1 - 20 Jun 2011Acknowledgement
I am grateful for the help I have got from my advisorsMarcLambert andSzabolcsGyimóthy during the past
several years. Not only this PhD Dissertation, but also my MSc Thesis was written under their supervision. I
really learned a lot from both of them. Marc Lambert showed me how an enthusiastic and tenacious research
looks like. Szabolcs Gyimóthy’s accuracy, judiciousness and unbreakable optimism was exemplary, without
doubt – not only in research, but also in my teaching activity at the Budapest University of Technology and
Economics. Both of them participated in the jury of the PhD defense as invited members.
I say thanks for the careful reviewing of this Dissertation to Patrick Dular from the University of Liège,
to Gábor Harsányi from the Budapest University of Technology and Economics and to Jérôme Idier from
the Institut de Recherche en Communications et Cybernétique de Nantes. Special thanks go to Patrick Dular
and Jérôme Idier for having accepted also to be member of the jury. I am grateful for József Bíró, who also
participated in the jury, representing the Budapest University of Technology and Economics.
Though I would like to express my gratitude for all people I have been in touch during my work, I
must highlight the most impressive ones herein. JózsefPávó taught me to “sight” the electromagnetic field,
whatever this means. Without his contribution, I think I would not have started to deal with electromagnetic
nondestructive evaluation –and I possibly would not have gotten in touch with the Laboratoire des Signaux
et Systèmes. A valuable technical help was also given by József, as he provided his code for the numerical
simulation of the illustrative eddy-current testing setup. An unforgettable experience is the cooperation with
Dominique Lesselier. I have learnt so many tiny things from him (e.g., to start with the bibliography when
reading a paper), –it is impossible to explain by words some of them–, and, what impressed me the most:
his exceptionally familiar manner. Dominique Lesselier accepted to be a member of the jury of the PhD
defense as well –moreover, he has been chosen as the president of the jury–, for what I am thankful for him.
For the topic of my PhD work, I must say thanks toEmmanuelVazquez, who also participated in the jury as
an invited member. He called our attention to the kriging approach, moreover, he also provided some codes
as well at the beginning. Though in this Dissertation, the work we have done in the domain of radar forest
observations is only slightly concerned, I would like to thank to Laetitia Thirion-Lefevre for this. I believe
that this cooperation –basically initiated by her– is fruitful.
In terms of organizations, I would like to thank the Department of Broadband Infocommunications and
Electromagnetic Theory at the Budapest University of Technology and Economics and the Laboratoire des
Signaux et Systèmes for the milieu I had the privilege to work in during the three years of this co-tutel PhD
study. Both contributed to the financial background of my work as well. I am grateful also to DIGITEO for
providing me a scholarship, and to the Commissariat à l’Énergie Atomique (CEA) for having taken part in
the establishment of the frame of my work. CEA was represented also at the defense by Pierre Calmon, to
whom I say thanks for having accepted the invitation for being an invited jury member.
The last but not least, I express my sincere gratitude to my family and my friends for their affection.
i
tel-00601753, version 1 - 20 Jun 2011tel-00601753, version 1 - 20 Jun 2011Abstract
Electromagnetic Nondestructive Evaluation (ENDE) is applied in various industrial domains for the ex-
ploration of hidden in-material defects of structural components. The methods rely on the fact that the
electromagnetic (EM) constitutive parameters of the material are locally changed in the presence of a de-
fect. Based on the measured EM field (generated by an external source and interacting with the examined
specimen), ENDE aims at the characterization of the defects, i.e., the solution of the relatedinverseproblem.
To this end, one has to be able to determine the EM field corresponding to a known defect, i.e., to solve the
forward problem. Practically, this is performed via the mathematical modeling (based on the Maxwell’s
equations) and the numerical simulation (e.g., by the Method of Moments) of the studied ENDE setup.
Such simulators can provide fine precision, but at a price of computational cost. However, the solution
of an inverse problem often requires several runs of these “expensive-to-evaluate” simulators, making the
inversion procedure firmly demanding in terms of runtime and computational resources. To overcome this
challenge, surrogate modeling (SM) is getting more and more widespread in electromagnetics. A surrogate
model imitates the true model, but as a rule, it is much less complex than the latter. A way to construct such
surrogates is to perform a couple of simulations and then to approximate the model based on the obtained
data. The choice of the “prototype” simulations is usually controlled by a sophisticated strategy, drawn from
the tools ofDesign-of-Experiments (DoE).
The goal of the research work presented in this Dissertation is the improvement of ENDE methods
by using SM and DoE techniques. The EM simulator –assuming a parametric defect model– is treated
as a black-box, i.e., accounting only for the input-output relationship. The input is the set of parameters
describing the defect geometry, the output is related to the measurable EM field in function of the receiver
position, respectively. Three self-sufficient approaches are then discussed in detail.
First, an optimization-based inversion algorithm is presented. By tuning the input parameters of the
simulator, the best similarity (in terms of an appropriate objective function) between the measured and
simulated data is to be achieved. Inspired by the need for reducing the number of simulation runs, the
objective function is minimized by using the “Efficient Global Optimization” (EGO) algorithm, which is
known to converge within a relatively small number of iterations. EGO chooses the inputs to be simulated
one-by-one, yielding a sequence converging to the global minimum, i.e., to the solution of the inverse
problem. The algorithm is based on the surrogate modeling of the objective function using kriging (modeling
by means of a Gaussian process).
We propose two generic surrogate modeling methods as well. Both are based on a sequential sampling
strategy for the generation of a problem-specific database of corresponding input parameter - output data
pairs. This database will subsequently be used as the support of a surrogate model. To improve the perfor-
mance of this model, the database generation is adapted both to the modeled forward problem and to the goal
of the surrogate being built. In the first approach, a functional kriging interpolator is fitted to the samples,
and the objective is the reduce the discrepancy between the true and the interpolated output data. To this end,
the sampling strategy is driven by the estimated uncertainty of the functional kriging prediction (via a jack-
knife variance-estimator). The second approach aims at uniformly filling the “output-space” (the domain of
the feasible output data) by samples. The main benefit of such output space-filling is the meta-information
on the modeled problem provided by the structure of the database. We present two inverse mappings to
exploit this meta-information for the quantitative characterization of the related inverse problem.
All approaches are illustrated by examples drawn from Eddy-Curr